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 * We do not care about task placement until a task runs on a node
905 * other than the first one used by the address space. This is
906 * largely because migrations are driven by what CPU the task
907 * is running on. If it's never scheduled on another node, it'll
908 * not migrate so why bother trapping the fault.
910 if (mm->first_nid == NUMA_PTE_SCAN_INIT)
911 mm->first_nid = numa_node_id();
912 if (mm->first_nid != NUMA_PTE_SCAN_ACTIVE) {
913 /* Are we running on a new node yet? */
914 if (numa_node_id() == mm->first_nid &&
915 !sched_feat_numa(NUMA_FORCE))
918 mm->first_nid = NUMA_PTE_SCAN_ACTIVE;
922 * Reset the scan period if enough time has gone by. Objective is that
923 * scanning will be reduced if pages are properly placed. As tasks
924 * can enter different phases this needs to be re-examined. Lacking
925 * proper tracking of reference behaviour, this blunt hammer is used.
927 migrate = mm->numa_next_reset;
928 if (time_after(now, migrate)) {
929 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
930 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
931 xchg(&mm->numa_next_reset, next_scan);
935 * Enforce maximal scan/migration frequency..
937 migrate = mm->numa_next_scan;
938 if (time_before(now, migrate))
941 if (p->numa_scan_period == 0)
942 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
944 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
945 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
949 * Delay this task enough that another task of this mm will likely win
950 * the next time around.
952 p->node_stamp += 2 * TICK_NSEC;
955 * Do not set pte_numa if the current running node is rate-limited.
956 * This loses statistics on the fault but if we are unwilling to
957 * migrate to this node, it is less likely we can do useful work
959 if (migrate_ratelimited(numa_node_id()))
962 start = mm->numa_scan_offset;
963 pages = sysctl_numa_balancing_scan_size;
964 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
968 down_read(&mm->mmap_sem);
969 vma = find_vma(mm, start);
971 reset_ptenuma_scan(p);
975 for (; vma; vma = vma->vm_next) {
976 if (!vma_migratable(vma))
979 /* Skip small VMAs. They are not likely to be of relevance */
980 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
984 start = max(start, vma->vm_start);
985 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
986 end = min(end, vma->vm_end);
987 pages -= change_prot_numa(vma, start, end);
992 } while (end != vma->vm_end);
997 * It is possible to reach the end of the VMA list but the last few
998 * VMAs are not guaranteed to the vma_migratable. If they are not, we
999 * would find the !migratable VMA on the next scan but not reset the
1000 * scanner to the start so check it now.
1003 mm->numa_scan_offset = start;
1005 reset_ptenuma_scan(p);
1006 up_read(&mm->mmap_sem);
1010 * Drive the periodic memory faults..
1012 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1014 struct callback_head *work = &curr->numa_work;
1018 * We don't care about NUMA placement if we don't have memory.
1020 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1024 * Using runtime rather than walltime has the dual advantage that
1025 * we (mostly) drive the selection from busy threads and that the
1026 * task needs to have done some actual work before we bother with
1029 now = curr->se.sum_exec_runtime;
1030 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1032 if (now - curr->node_stamp > period) {
1033 if (!curr->node_stamp)
1034 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
1035 curr->node_stamp += period;
1037 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1038 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1039 task_work_add(curr, work, true);
1044 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1047 #endif /* CONFIG_NUMA_BALANCING */
1050 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1052 update_load_add(&cfs_rq->load, se->load.weight);
1053 if (!parent_entity(se))
1054 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1056 if (entity_is_task(se))
1057 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1059 cfs_rq->nr_running++;
1063 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1065 update_load_sub(&cfs_rq->load, se->load.weight);
1066 if (!parent_entity(se))
1067 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1068 if (entity_is_task(se))
1069 list_del_init(&se->group_node);
1070 cfs_rq->nr_running--;
1073 #ifdef CONFIG_FAIR_GROUP_SCHED
1075 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1080 * Use this CPU's actual weight instead of the last load_contribution
1081 * to gain a more accurate current total weight. See
1082 * update_cfs_rq_load_contribution().
1084 tg_weight = atomic_long_read(&tg->load_avg);
1085 tg_weight -= cfs_rq->tg_load_contrib;
1086 tg_weight += cfs_rq->load.weight;
1091 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1093 long tg_weight, load, shares;
1095 tg_weight = calc_tg_weight(tg, cfs_rq);
1096 load = cfs_rq->load.weight;
1098 shares = (tg->shares * load);
1100 shares /= tg_weight;
1102 if (shares < MIN_SHARES)
1103 shares = MIN_SHARES;
1104 if (shares > tg->shares)
1105 shares = tg->shares;
1109 # else /* CONFIG_SMP */
1110 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1114 # endif /* CONFIG_SMP */
1115 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1116 unsigned long weight)
1119 /* commit outstanding execution time */
1120 if (cfs_rq->curr == se)
1121 update_curr(cfs_rq);
1122 account_entity_dequeue(cfs_rq, se);
1125 update_load_set(&se->load, weight);
1128 account_entity_enqueue(cfs_rq, se);
1131 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1133 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1135 struct task_group *tg;
1136 struct sched_entity *se;
1140 se = tg->se[cpu_of(rq_of(cfs_rq))];
1141 if (!se || throttled_hierarchy(cfs_rq))
1144 if (likely(se->load.weight == tg->shares))
1147 shares = calc_cfs_shares(cfs_rq, tg);
1149 reweight_entity(cfs_rq_of(se), se, shares);
1151 #else /* CONFIG_FAIR_GROUP_SCHED */
1152 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1155 #endif /* CONFIG_FAIR_GROUP_SCHED */
1159 * We choose a half-life close to 1 scheduling period.
1160 * Note: The tables below are dependent on this value.
1162 #define LOAD_AVG_PERIOD 32
1163 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1164 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1166 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1167 static const u32 runnable_avg_yN_inv[] = {
1168 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1169 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1170 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1171 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1172 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1173 0x85aac367, 0x82cd8698,
1177 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1178 * over-estimates when re-combining.
1180 static const u32 runnable_avg_yN_sum[] = {
1181 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1182 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1183 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1188 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1190 static __always_inline u64 decay_load(u64 val, u64 n)
1192 unsigned int local_n;
1196 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1199 /* after bounds checking we can collapse to 32-bit */
1203 * As y^PERIOD = 1/2, we can combine
1204 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1205 * With a look-up table which covers k^n (n<PERIOD)
1207 * To achieve constant time decay_load.
1209 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1210 val >>= local_n / LOAD_AVG_PERIOD;
1211 local_n %= LOAD_AVG_PERIOD;
1214 val *= runnable_avg_yN_inv[local_n];
1215 /* We don't use SRR here since we always want to round down. */
1220 * For updates fully spanning n periods, the contribution to runnable
1221 * average will be: \Sum 1024*y^n
1223 * We can compute this reasonably efficiently by combining:
1224 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1226 static u32 __compute_runnable_contrib(u64 n)
1230 if (likely(n <= LOAD_AVG_PERIOD))
1231 return runnable_avg_yN_sum[n];
1232 else if (unlikely(n >= LOAD_AVG_MAX_N))
1233 return LOAD_AVG_MAX;
1235 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1237 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1238 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1240 n -= LOAD_AVG_PERIOD;
1241 } while (n > LOAD_AVG_PERIOD);
1243 contrib = decay_load(contrib, n);
1244 return contrib + runnable_avg_yN_sum[n];
1248 * We can represent the historical contribution to runnable average as the
1249 * coefficients of a geometric series. To do this we sub-divide our runnable
1250 * history into segments of approximately 1ms (1024us); label the segment that
1251 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1253 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1255 * (now) (~1ms ago) (~2ms ago)
1257 * Let u_i denote the fraction of p_i that the entity was runnable.
1259 * We then designate the fractions u_i as our co-efficients, yielding the
1260 * following representation of historical load:
1261 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1263 * We choose y based on the with of a reasonably scheduling period, fixing:
1266 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1267 * approximately half as much as the contribution to load within the last ms
1270 * When a period "rolls over" and we have new u_0`, multiplying the previous
1271 * sum again by y is sufficient to update:
1272 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1273 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1275 static __always_inline int __update_entity_runnable_avg(u64 now,
1276 struct sched_avg *sa,
1280 u32 runnable_contrib;
1281 int delta_w, decayed = 0;
1283 delta = now - sa->last_runnable_update;
1285 * This should only happen when time goes backwards, which it
1286 * unfortunately does during sched clock init when we swap over to TSC.
1288 if ((s64)delta < 0) {
1289 sa->last_runnable_update = now;
1294 * Use 1024ns as the unit of measurement since it's a reasonable
1295 * approximation of 1us and fast to compute.
1300 sa->last_runnable_update = now;
1302 /* delta_w is the amount already accumulated against our next period */
1303 delta_w = sa->runnable_avg_period % 1024;
1304 if (delta + delta_w >= 1024) {
1305 /* period roll-over */
1309 * Now that we know we're crossing a period boundary, figure
1310 * out how much from delta we need to complete the current
1311 * period and accrue it.
1313 delta_w = 1024 - delta_w;
1315 sa->runnable_avg_sum += delta_w;
1316 sa->runnable_avg_period += delta_w;
1320 /* Figure out how many additional periods this update spans */
1321 periods = delta / 1024;
1324 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1326 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1329 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1330 runnable_contrib = __compute_runnable_contrib(periods);
1332 sa->runnable_avg_sum += runnable_contrib;
1333 sa->runnable_avg_period += runnable_contrib;
1336 /* Remainder of delta accrued against u_0` */
1338 sa->runnable_avg_sum += delta;
1339 sa->runnable_avg_period += delta;
1344 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1345 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1347 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1348 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1350 decays -= se->avg.decay_count;
1354 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1355 se->avg.decay_count = 0;
1360 #ifdef CONFIG_FAIR_GROUP_SCHED
1361 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1364 struct task_group *tg = cfs_rq->tg;
1367 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1368 tg_contrib -= cfs_rq->tg_load_contrib;
1370 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1371 atomic_long_add(tg_contrib, &tg->load_avg);
1372 cfs_rq->tg_load_contrib += tg_contrib;
1377 * Aggregate cfs_rq runnable averages into an equivalent task_group
1378 * representation for computing load contributions.
1380 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1381 struct cfs_rq *cfs_rq)
1383 struct task_group *tg = cfs_rq->tg;
1386 /* The fraction of a cpu used by this cfs_rq */
1387 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1388 sa->runnable_avg_period + 1);
1389 contrib -= cfs_rq->tg_runnable_contrib;
1391 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1392 atomic_add(contrib, &tg->runnable_avg);
1393 cfs_rq->tg_runnable_contrib += contrib;
1397 static inline void __update_group_entity_contrib(struct sched_entity *se)
1399 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1400 struct task_group *tg = cfs_rq->tg;
1405 contrib = cfs_rq->tg_load_contrib * tg->shares;
1406 se->avg.load_avg_contrib = div_u64(contrib,
1407 atomic_long_read(&tg->load_avg) + 1);
1410 * For group entities we need to compute a correction term in the case
1411 * that they are consuming <1 cpu so that we would contribute the same
1412 * load as a task of equal weight.
1414 * Explicitly co-ordinating this measurement would be expensive, but
1415 * fortunately the sum of each cpus contribution forms a usable
1416 * lower-bound on the true value.
1418 * Consider the aggregate of 2 contributions. Either they are disjoint
1419 * (and the sum represents true value) or they are disjoint and we are
1420 * understating by the aggregate of their overlap.
1422 * Extending this to N cpus, for a given overlap, the maximum amount we
1423 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1424 * cpus that overlap for this interval and w_i is the interval width.
1426 * On a small machine; the first term is well-bounded which bounds the
1427 * total error since w_i is a subset of the period. Whereas on a
1428 * larger machine, while this first term can be larger, if w_i is the
1429 * of consequential size guaranteed to see n_i*w_i quickly converge to
1430 * our upper bound of 1-cpu.
1432 runnable_avg = atomic_read(&tg->runnable_avg);
1433 if (runnable_avg < NICE_0_LOAD) {
1434 se->avg.load_avg_contrib *= runnable_avg;
1435 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1439 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1440 int force_update) {}
1441 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1442 struct cfs_rq *cfs_rq) {}
1443 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1446 static inline void __update_task_entity_contrib(struct sched_entity *se)
1450 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1451 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1452 contrib /= (se->avg.runnable_avg_period + 1);
1453 se->avg.load_avg_contrib = scale_load(contrib);
1456 /* Compute the current contribution to load_avg by se, return any delta */
1457 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1459 long old_contrib = se->avg.load_avg_contrib;
1461 if (entity_is_task(se)) {
1462 __update_task_entity_contrib(se);
1464 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1465 __update_group_entity_contrib(se);
1468 return se->avg.load_avg_contrib - old_contrib;
1471 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1474 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1475 cfs_rq->blocked_load_avg -= load_contrib;
1477 cfs_rq->blocked_load_avg = 0;
1480 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1482 /* Update a sched_entity's runnable average */
1483 static inline void update_entity_load_avg(struct sched_entity *se,
1486 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1491 * For a group entity we need to use their owned cfs_rq_clock_task() in
1492 * case they are the parent of a throttled hierarchy.
1494 if (entity_is_task(se))
1495 now = cfs_rq_clock_task(cfs_rq);
1497 now = cfs_rq_clock_task(group_cfs_rq(se));
1499 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1502 contrib_delta = __update_entity_load_avg_contrib(se);
1508 cfs_rq->runnable_load_avg += contrib_delta;
1510 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1514 * Decay the load contributed by all blocked children and account this so that
1515 * their contribution may appropriately discounted when they wake up.
1517 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1519 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1522 decays = now - cfs_rq->last_decay;
1523 if (!decays && !force_update)
1526 if (atomic_long_read(&cfs_rq->removed_load)) {
1527 unsigned long removed_load;
1528 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1529 subtract_blocked_load_contrib(cfs_rq, removed_load);
1533 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1535 atomic64_add(decays, &cfs_rq->decay_counter);
1536 cfs_rq->last_decay = now;
1539 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1542 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1544 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1545 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1548 /* Add the load generated by se into cfs_rq's child load-average */
1549 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1550 struct sched_entity *se,
1554 * We track migrations using entity decay_count <= 0, on a wake-up
1555 * migration we use a negative decay count to track the remote decays
1556 * accumulated while sleeping.
1558 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1559 * are seen by enqueue_entity_load_avg() as a migration with an already
1560 * constructed load_avg_contrib.
1562 if (unlikely(se->avg.decay_count <= 0)) {
1563 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1564 if (se->avg.decay_count) {
1566 * In a wake-up migration we have to approximate the
1567 * time sleeping. This is because we can't synchronize
1568 * clock_task between the two cpus, and it is not
1569 * guaranteed to be read-safe. Instead, we can
1570 * approximate this using our carried decays, which are
1571 * explicitly atomically readable.
1573 se->avg.last_runnable_update -= (-se->avg.decay_count)
1575 update_entity_load_avg(se, 0);
1576 /* Indicate that we're now synchronized and on-rq */
1577 se->avg.decay_count = 0;
1582 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1583 * would have made count negative); we must be careful to avoid
1584 * double-accounting blocked time after synchronizing decays.
1586 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1590 /* migrated tasks did not contribute to our blocked load */
1592 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1593 update_entity_load_avg(se, 0);
1596 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1597 /* we force update consideration on load-balancer moves */
1598 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1602 * Remove se's load from this cfs_rq child load-average, if the entity is
1603 * transitioning to a blocked state we track its projected decay using
1606 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1607 struct sched_entity *se,
1610 update_entity_load_avg(se, 1);
1611 /* we force update consideration on load-balancer moves */
1612 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1614 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1616 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1617 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1618 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1622 * Update the rq's load with the elapsed running time before entering
1623 * idle. if the last scheduled task is not a CFS task, idle_enter will
1624 * be the only way to update the runnable statistic.
1626 void idle_enter_fair(struct rq *this_rq)
1628 update_rq_runnable_avg(this_rq, 1);
1632 * Update the rq's load with the elapsed idle time before a task is
1633 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1634 * be the only way to update the runnable statistic.
1636 void idle_exit_fair(struct rq *this_rq)
1638 update_rq_runnable_avg(this_rq, 0);
1642 static inline void update_entity_load_avg(struct sched_entity *se,
1643 int update_cfs_rq) {}
1644 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1645 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1646 struct sched_entity *se,
1648 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1649 struct sched_entity *se,
1651 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1652 int force_update) {}
1655 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1657 #ifdef CONFIG_SCHEDSTATS
1658 struct task_struct *tsk = NULL;
1660 if (entity_is_task(se))
1663 if (se->statistics.sleep_start) {
1664 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1669 if (unlikely(delta > se->statistics.sleep_max))
1670 se->statistics.sleep_max = delta;
1672 se->statistics.sleep_start = 0;
1673 se->statistics.sum_sleep_runtime += delta;
1676 account_scheduler_latency(tsk, delta >> 10, 1);
1677 trace_sched_stat_sleep(tsk, delta);
1680 if (se->statistics.block_start) {
1681 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1686 if (unlikely(delta > se->statistics.block_max))
1687 se->statistics.block_max = delta;
1689 se->statistics.block_start = 0;
1690 se->statistics.sum_sleep_runtime += delta;
1693 if (tsk->in_iowait) {
1694 se->statistics.iowait_sum += delta;
1695 se->statistics.iowait_count++;
1696 trace_sched_stat_iowait(tsk, delta);
1699 trace_sched_stat_blocked(tsk, delta);
1702 * Blocking time is in units of nanosecs, so shift by
1703 * 20 to get a milliseconds-range estimation of the
1704 * amount of time that the task spent sleeping:
1706 if (unlikely(prof_on == SLEEP_PROFILING)) {
1707 profile_hits(SLEEP_PROFILING,
1708 (void *)get_wchan(tsk),
1711 account_scheduler_latency(tsk, delta >> 10, 0);
1717 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1719 #ifdef CONFIG_SCHED_DEBUG
1720 s64 d = se->vruntime - cfs_rq->min_vruntime;
1725 if (d > 3*sysctl_sched_latency)
1726 schedstat_inc(cfs_rq, nr_spread_over);
1731 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1733 u64 vruntime = cfs_rq->min_vruntime;
1736 * The 'current' period is already promised to the current tasks,
1737 * however the extra weight of the new task will slow them down a
1738 * little, place the new task so that it fits in the slot that
1739 * stays open at the end.
1741 if (initial && sched_feat(START_DEBIT))
1742 vruntime += sched_vslice(cfs_rq, se);
1744 /* sleeps up to a single latency don't count. */
1746 unsigned long thresh = sysctl_sched_latency;
1749 * Halve their sleep time's effect, to allow
1750 * for a gentler effect of sleepers:
1752 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1758 /* ensure we never gain time by being placed backwards. */
1759 se->vruntime = max_vruntime(se->vruntime, vruntime);
1762 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1765 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1768 * Update the normalized vruntime before updating min_vruntime
1769 * through calling update_curr().
1771 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1772 se->vruntime += cfs_rq->min_vruntime;
1775 * Update run-time statistics of the 'current'.
1777 update_curr(cfs_rq);
1778 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1779 account_entity_enqueue(cfs_rq, se);
1780 update_cfs_shares(cfs_rq);
1782 if (flags & ENQUEUE_WAKEUP) {
1783 place_entity(cfs_rq, se, 0);
1784 enqueue_sleeper(cfs_rq, se);
1787 update_stats_enqueue(cfs_rq, se);
1788 check_spread(cfs_rq, se);
1789 if (se != cfs_rq->curr)
1790 __enqueue_entity(cfs_rq, se);
1793 if (cfs_rq->nr_running == 1) {
1794 list_add_leaf_cfs_rq(cfs_rq);
1795 check_enqueue_throttle(cfs_rq);
1799 static void __clear_buddies_last(struct sched_entity *se)
1801 for_each_sched_entity(se) {
1802 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1803 if (cfs_rq->last == se)
1804 cfs_rq->last = NULL;
1810 static void __clear_buddies_next(struct sched_entity *se)
1812 for_each_sched_entity(se) {
1813 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1814 if (cfs_rq->next == se)
1815 cfs_rq->next = NULL;
1821 static void __clear_buddies_skip(struct sched_entity *se)
1823 for_each_sched_entity(se) {
1824 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1825 if (cfs_rq->skip == se)
1826 cfs_rq->skip = NULL;
1832 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1834 if (cfs_rq->last == se)
1835 __clear_buddies_last(se);
1837 if (cfs_rq->next == se)
1838 __clear_buddies_next(se);
1840 if (cfs_rq->skip == se)
1841 __clear_buddies_skip(se);
1844 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1847 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1850 * Update run-time statistics of the 'current'.
1852 update_curr(cfs_rq);
1853 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1855 update_stats_dequeue(cfs_rq, se);
1856 if (flags & DEQUEUE_SLEEP) {
1857 #ifdef CONFIG_SCHEDSTATS
1858 if (entity_is_task(se)) {
1859 struct task_struct *tsk = task_of(se);
1861 if (tsk->state & TASK_INTERRUPTIBLE)
1862 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
1863 if (tsk->state & TASK_UNINTERRUPTIBLE)
1864 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
1869 clear_buddies(cfs_rq, se);
1871 if (se != cfs_rq->curr)
1872 __dequeue_entity(cfs_rq, se);
1874 account_entity_dequeue(cfs_rq, se);
1877 * Normalize the entity after updating the min_vruntime because the
1878 * update can refer to the ->curr item and we need to reflect this
1879 * movement in our normalized position.
1881 if (!(flags & DEQUEUE_SLEEP))
1882 se->vruntime -= cfs_rq->min_vruntime;
1884 /* return excess runtime on last dequeue */
1885 return_cfs_rq_runtime(cfs_rq);
1887 update_min_vruntime(cfs_rq);
1888 update_cfs_shares(cfs_rq);
1892 * Preempt the current task with a newly woken task if needed:
1895 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1897 unsigned long ideal_runtime, delta_exec;
1898 struct sched_entity *se;
1901 ideal_runtime = sched_slice(cfs_rq, curr);
1902 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1903 if (delta_exec > ideal_runtime) {
1904 resched_task(rq_of(cfs_rq)->curr);
1906 * The current task ran long enough, ensure it doesn't get
1907 * re-elected due to buddy favours.
1909 clear_buddies(cfs_rq, curr);
1914 * Ensure that a task that missed wakeup preemption by a
1915 * narrow margin doesn't have to wait for a full slice.
1916 * This also mitigates buddy induced latencies under load.
1918 if (delta_exec < sysctl_sched_min_granularity)
1921 se = __pick_first_entity(cfs_rq);
1922 delta = curr->vruntime - se->vruntime;
1927 if (delta > ideal_runtime)
1928 resched_task(rq_of(cfs_rq)->curr);
1932 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1934 /* 'current' is not kept within the tree. */
1937 * Any task has to be enqueued before it get to execute on
1938 * a CPU. So account for the time it spent waiting on the
1941 update_stats_wait_end(cfs_rq, se);
1942 __dequeue_entity(cfs_rq, se);
1945 update_stats_curr_start(cfs_rq, se);
1947 #ifdef CONFIG_SCHEDSTATS
1949 * Track our maximum slice length, if the CPU's load is at
1950 * least twice that of our own weight (i.e. dont track it
1951 * when there are only lesser-weight tasks around):
1953 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1954 se->statistics.slice_max = max(se->statistics.slice_max,
1955 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1958 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1962 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1965 * Pick the next process, keeping these things in mind, in this order:
1966 * 1) keep things fair between processes/task groups
1967 * 2) pick the "next" process, since someone really wants that to run
1968 * 3) pick the "last" process, for cache locality
1969 * 4) do not run the "skip" process, if something else is available
1971 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1973 struct sched_entity *se = __pick_first_entity(cfs_rq);
1974 struct sched_entity *left = se;
1977 * Avoid running the skip buddy, if running something else can
1978 * be done without getting too unfair.
1980 if (cfs_rq->skip == se) {
1981 struct sched_entity *second = __pick_next_entity(se);
1982 if (second && wakeup_preempt_entity(second, left) < 1)
1987 * Prefer last buddy, try to return the CPU to a preempted task.
1989 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1993 * Someone really wants this to run. If it's not unfair, run it.
1995 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1998 clear_buddies(cfs_rq, se);
2003 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2005 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2008 * If still on the runqueue then deactivate_task()
2009 * was not called and update_curr() has to be done:
2012 update_curr(cfs_rq);
2014 /* throttle cfs_rqs exceeding runtime */
2015 check_cfs_rq_runtime(cfs_rq);
2017 check_spread(cfs_rq, prev);
2019 update_stats_wait_start(cfs_rq, prev);
2020 /* Put 'current' back into the tree. */
2021 __enqueue_entity(cfs_rq, prev);
2022 /* in !on_rq case, update occurred at dequeue */
2023 update_entity_load_avg(prev, 1);
2025 cfs_rq->curr = NULL;
2029 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2032 * Update run-time statistics of the 'current'.
2034 update_curr(cfs_rq);
2037 * Ensure that runnable average is periodically updated.
2039 update_entity_load_avg(curr, 1);
2040 update_cfs_rq_blocked_load(cfs_rq, 1);
2041 update_cfs_shares(cfs_rq);
2043 #ifdef CONFIG_SCHED_HRTICK
2045 * queued ticks are scheduled to match the slice, so don't bother
2046 * validating it and just reschedule.
2049 resched_task(rq_of(cfs_rq)->curr);
2053 * don't let the period tick interfere with the hrtick preemption
2055 if (!sched_feat(DOUBLE_TICK) &&
2056 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2060 if (cfs_rq->nr_running > 1)
2061 check_preempt_tick(cfs_rq, curr);
2065 /**************************************************
2066 * CFS bandwidth control machinery
2069 #ifdef CONFIG_CFS_BANDWIDTH
2071 #ifdef HAVE_JUMP_LABEL
2072 static struct static_key __cfs_bandwidth_used;
2074 static inline bool cfs_bandwidth_used(void)
2076 return static_key_false(&__cfs_bandwidth_used);
2079 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2081 /* only need to count groups transitioning between enabled/!enabled */
2082 if (enabled && !was_enabled)
2083 static_key_slow_inc(&__cfs_bandwidth_used);
2084 else if (!enabled && was_enabled)
2085 static_key_slow_dec(&__cfs_bandwidth_used);
2087 #else /* HAVE_JUMP_LABEL */
2088 static bool cfs_bandwidth_used(void)
2093 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2094 #endif /* HAVE_JUMP_LABEL */
2097 * default period for cfs group bandwidth.
2098 * default: 0.1s, units: nanoseconds
2100 static inline u64 default_cfs_period(void)
2102 return 100000000ULL;
2105 static inline u64 sched_cfs_bandwidth_slice(void)
2107 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2111 * Replenish runtime according to assigned quota and update expiration time.
2112 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2113 * additional synchronization around rq->lock.
2115 * requires cfs_b->lock
2117 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2121 if (cfs_b->quota == RUNTIME_INF)
2124 now = sched_clock_cpu(smp_processor_id());
2125 cfs_b->runtime = cfs_b->quota;
2126 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2129 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2131 return &tg->cfs_bandwidth;
2134 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2135 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2137 if (unlikely(cfs_rq->throttle_count))
2138 return cfs_rq->throttled_clock_task;
2140 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2143 /* returns 0 on failure to allocate runtime */
2144 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2146 struct task_group *tg = cfs_rq->tg;
2147 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2148 u64 amount = 0, min_amount, expires;
2150 /* note: this is a positive sum as runtime_remaining <= 0 */
2151 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2153 raw_spin_lock(&cfs_b->lock);
2154 if (cfs_b->quota == RUNTIME_INF)
2155 amount = min_amount;
2158 * If the bandwidth pool has become inactive, then at least one
2159 * period must have elapsed since the last consumption.
2160 * Refresh the global state and ensure bandwidth timer becomes
2163 if (!cfs_b->timer_active) {
2164 __refill_cfs_bandwidth_runtime(cfs_b);
2165 __start_cfs_bandwidth(cfs_b);
2168 if (cfs_b->runtime > 0) {
2169 amount = min(cfs_b->runtime, min_amount);
2170 cfs_b->runtime -= amount;
2174 expires = cfs_b->runtime_expires;
2175 raw_spin_unlock(&cfs_b->lock);
2177 cfs_rq->runtime_remaining += amount;
2179 * we may have advanced our local expiration to account for allowed
2180 * spread between our sched_clock and the one on which runtime was
2183 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2184 cfs_rq->runtime_expires = expires;
2186 return cfs_rq->runtime_remaining > 0;
2190 * Note: This depends on the synchronization provided by sched_clock and the
2191 * fact that rq->clock snapshots this value.
2193 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2195 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2197 /* if the deadline is ahead of our clock, nothing to do */
2198 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2201 if (cfs_rq->runtime_remaining < 0)
2205 * If the local deadline has passed we have to consider the
2206 * possibility that our sched_clock is 'fast' and the global deadline
2207 * has not truly expired.
2209 * Fortunately we can check determine whether this the case by checking
2210 * whether the global deadline has advanced.
2213 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2214 /* extend local deadline, drift is bounded above by 2 ticks */
2215 cfs_rq->runtime_expires += TICK_NSEC;
2217 /* global deadline is ahead, expiration has passed */
2218 cfs_rq->runtime_remaining = 0;
2222 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2223 unsigned long delta_exec)
2225 /* dock delta_exec before expiring quota (as it could span periods) */
2226 cfs_rq->runtime_remaining -= delta_exec;
2227 expire_cfs_rq_runtime(cfs_rq);
2229 if (likely(cfs_rq->runtime_remaining > 0))
2233 * if we're unable to extend our runtime we resched so that the active
2234 * hierarchy can be throttled
2236 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2237 resched_task(rq_of(cfs_rq)->curr);
2240 static __always_inline
2241 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2243 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2246 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2249 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2251 return cfs_bandwidth_used() && cfs_rq->throttled;
2254 /* check whether cfs_rq, or any parent, is throttled */
2255 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2257 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2261 * Ensure that neither of the group entities corresponding to src_cpu or
2262 * dest_cpu are members of a throttled hierarchy when performing group
2263 * load-balance operations.
2265 static inline int throttled_lb_pair(struct task_group *tg,
2266 int src_cpu, int dest_cpu)
2268 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2270 src_cfs_rq = tg->cfs_rq[src_cpu];
2271 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2273 return throttled_hierarchy(src_cfs_rq) ||
2274 throttled_hierarchy(dest_cfs_rq);
2277 /* updated child weight may affect parent so we have to do this bottom up */
2278 static int tg_unthrottle_up(struct task_group *tg, void *data)
2280 struct rq *rq = data;
2281 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2283 cfs_rq->throttle_count--;
2285 if (!cfs_rq->throttle_count) {
2286 /* adjust cfs_rq_clock_task() */
2287 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2288 cfs_rq->throttled_clock_task;
2295 static int tg_throttle_down(struct task_group *tg, void *data)
2297 struct rq *rq = data;
2298 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2300 /* group is entering throttled state, stop time */
2301 if (!cfs_rq->throttle_count)
2302 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2303 cfs_rq->throttle_count++;
2308 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2310 struct rq *rq = rq_of(cfs_rq);
2311 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2312 struct sched_entity *se;
2313 long task_delta, dequeue = 1;
2315 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2317 /* freeze hierarchy runnable averages while throttled */
2319 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2322 task_delta = cfs_rq->h_nr_running;
2323 for_each_sched_entity(se) {
2324 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2325 /* throttled entity or throttle-on-deactivate */
2330 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2331 qcfs_rq->h_nr_running -= task_delta;
2333 if (qcfs_rq->load.weight)
2338 rq->nr_running -= task_delta;
2340 cfs_rq->throttled = 1;
2341 cfs_rq->throttled_clock = rq_clock(rq);
2342 raw_spin_lock(&cfs_b->lock);
2343 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2344 raw_spin_unlock(&cfs_b->lock);
2347 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2349 struct rq *rq = rq_of(cfs_rq);
2350 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2351 struct sched_entity *se;
2355 se = cfs_rq->tg->se[cpu_of(rq)];
2357 cfs_rq->throttled = 0;
2359 update_rq_clock(rq);
2361 raw_spin_lock(&cfs_b->lock);
2362 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2363 list_del_rcu(&cfs_rq->throttled_list);
2364 raw_spin_unlock(&cfs_b->lock);
2366 /* update hierarchical throttle state */
2367 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2369 if (!cfs_rq->load.weight)
2372 task_delta = cfs_rq->h_nr_running;
2373 for_each_sched_entity(se) {
2377 cfs_rq = cfs_rq_of(se);
2379 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2380 cfs_rq->h_nr_running += task_delta;
2382 if (cfs_rq_throttled(cfs_rq))
2387 rq->nr_running += task_delta;
2389 /* determine whether we need to wake up potentially idle cpu */
2390 if (rq->curr == rq->idle && rq->cfs.nr_running)
2391 resched_task(rq->curr);
2394 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2395 u64 remaining, u64 expires)
2397 struct cfs_rq *cfs_rq;
2398 u64 runtime = remaining;
2401 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2403 struct rq *rq = rq_of(cfs_rq);
2405 raw_spin_lock(&rq->lock);
2406 if (!cfs_rq_throttled(cfs_rq))
2409 runtime = -cfs_rq->runtime_remaining + 1;
2410 if (runtime > remaining)
2411 runtime = remaining;
2412 remaining -= runtime;
2414 cfs_rq->runtime_remaining += runtime;
2415 cfs_rq->runtime_expires = expires;
2417 /* we check whether we're throttled above */
2418 if (cfs_rq->runtime_remaining > 0)
2419 unthrottle_cfs_rq(cfs_rq);
2422 raw_spin_unlock(&rq->lock);
2433 * Responsible for refilling a task_group's bandwidth and unthrottling its
2434 * cfs_rqs as appropriate. If there has been no activity within the last
2435 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2436 * used to track this state.
2438 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2440 u64 runtime, runtime_expires;
2441 int idle = 1, throttled;
2443 raw_spin_lock(&cfs_b->lock);
2444 /* no need to continue the timer with no bandwidth constraint */
2445 if (cfs_b->quota == RUNTIME_INF)
2448 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2449 /* idle depends on !throttled (for the case of a large deficit) */
2450 idle = cfs_b->idle && !throttled;
2451 cfs_b->nr_periods += overrun;
2453 /* if we're going inactive then everything else can be deferred */
2457 __refill_cfs_bandwidth_runtime(cfs_b);
2460 /* mark as potentially idle for the upcoming period */
2465 /* account preceding periods in which throttling occurred */
2466 cfs_b->nr_throttled += overrun;
2469 * There are throttled entities so we must first use the new bandwidth
2470 * to unthrottle them before making it generally available. This
2471 * ensures that all existing debts will be paid before a new cfs_rq is
2474 runtime = cfs_b->runtime;
2475 runtime_expires = cfs_b->runtime_expires;
2479 * This check is repeated as we are holding onto the new bandwidth
2480 * while we unthrottle. This can potentially race with an unthrottled
2481 * group trying to acquire new bandwidth from the global pool.
2483 while (throttled && runtime > 0) {
2484 raw_spin_unlock(&cfs_b->lock);
2485 /* we can't nest cfs_b->lock while distributing bandwidth */
2486 runtime = distribute_cfs_runtime(cfs_b, runtime,
2488 raw_spin_lock(&cfs_b->lock);
2490 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2493 /* return (any) remaining runtime */
2494 cfs_b->runtime = runtime;
2496 * While we are ensured activity in the period following an
2497 * unthrottle, this also covers the case in which the new bandwidth is
2498 * insufficient to cover the existing bandwidth deficit. (Forcing the
2499 * timer to remain active while there are any throttled entities.)
2504 cfs_b->timer_active = 0;
2505 raw_spin_unlock(&cfs_b->lock);
2510 /* a cfs_rq won't donate quota below this amount */
2511 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2512 /* minimum remaining period time to redistribute slack quota */
2513 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2514 /* how long we wait to gather additional slack before distributing */
2515 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2517 /* are we near the end of the current quota period? */
2518 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2520 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2523 /* if the call-back is running a quota refresh is already occurring */
2524 if (hrtimer_callback_running(refresh_timer))
2527 /* is a quota refresh about to occur? */
2528 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2529 if (remaining < min_expire)
2535 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2537 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2539 /* if there's a quota refresh soon don't bother with slack */
2540 if (runtime_refresh_within(cfs_b, min_left))
2543 start_bandwidth_timer(&cfs_b->slack_timer,
2544 ns_to_ktime(cfs_bandwidth_slack_period));
2547 /* we know any runtime found here is valid as update_curr() precedes return */
2548 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2550 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2551 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2553 if (slack_runtime <= 0)
2556 raw_spin_lock(&cfs_b->lock);
2557 if (cfs_b->quota != RUNTIME_INF &&
2558 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2559 cfs_b->runtime += slack_runtime;
2561 /* we are under rq->lock, defer unthrottling using a timer */
2562 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2563 !list_empty(&cfs_b->throttled_cfs_rq))
2564 start_cfs_slack_bandwidth(cfs_b);
2566 raw_spin_unlock(&cfs_b->lock);
2568 /* even if it's not valid for return we don't want to try again */
2569 cfs_rq->runtime_remaining -= slack_runtime;
2572 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2574 if (!cfs_bandwidth_used())
2577 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2580 __return_cfs_rq_runtime(cfs_rq);
2584 * This is done with a timer (instead of inline with bandwidth return) since
2585 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2587 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2589 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2592 /* confirm we're still not at a refresh boundary */
2593 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2596 raw_spin_lock(&cfs_b->lock);
2597 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2598 runtime = cfs_b->runtime;
2601 expires = cfs_b->runtime_expires;
2602 raw_spin_unlock(&cfs_b->lock);
2607 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2609 raw_spin_lock(&cfs_b->lock);
2610 if (expires == cfs_b->runtime_expires)
2611 cfs_b->runtime = runtime;
2612 raw_spin_unlock(&cfs_b->lock);
2616 * When a group wakes up we want to make sure that its quota is not already
2617 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2618 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2620 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2622 if (!cfs_bandwidth_used())
2625 /* an active group must be handled by the update_curr()->put() path */
2626 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2629 /* ensure the group is not already throttled */
2630 if (cfs_rq_throttled(cfs_rq))
2633 /* update runtime allocation */
2634 account_cfs_rq_runtime(cfs_rq, 0);
2635 if (cfs_rq->runtime_remaining <= 0)
2636 throttle_cfs_rq(cfs_rq);
2639 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2640 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2642 if (!cfs_bandwidth_used())
2645 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2649 * it's possible for a throttled entity to be forced into a running
2650 * state (e.g. set_curr_task), in this case we're finished.
2652 if (cfs_rq_throttled(cfs_rq))
2655 throttle_cfs_rq(cfs_rq);
2658 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2660 struct cfs_bandwidth *cfs_b =
2661 container_of(timer, struct cfs_bandwidth, slack_timer);
2662 do_sched_cfs_slack_timer(cfs_b);
2664 return HRTIMER_NORESTART;
2667 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2669 struct cfs_bandwidth *cfs_b =
2670 container_of(timer, struct cfs_bandwidth, period_timer);
2676 now = hrtimer_cb_get_time(timer);
2677 overrun = hrtimer_forward(timer, now, cfs_b->period);
2682 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2685 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2688 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2690 raw_spin_lock_init(&cfs_b->lock);
2692 cfs_b->quota = RUNTIME_INF;
2693 cfs_b->period = ns_to_ktime(default_cfs_period());
2695 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2696 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2697 cfs_b->period_timer.function = sched_cfs_period_timer;
2698 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2699 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2702 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2704 cfs_rq->runtime_enabled = 0;
2705 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2708 /* requires cfs_b->lock, may release to reprogram timer */
2709 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2712 * The timer may be active because we're trying to set a new bandwidth
2713 * period or because we're racing with the tear-down path
2714 * (timer_active==0 becomes visible before the hrtimer call-back
2715 * terminates). In either case we ensure that it's re-programmed
2717 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2718 raw_spin_unlock(&cfs_b->lock);
2719 /* ensure cfs_b->lock is available while we wait */
2720 hrtimer_cancel(&cfs_b->period_timer);
2722 raw_spin_lock(&cfs_b->lock);
2723 /* if someone else restarted the timer then we're done */
2724 if (cfs_b->timer_active)
2728 cfs_b->timer_active = 1;
2729 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2732 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2734 hrtimer_cancel(&cfs_b->period_timer);
2735 hrtimer_cancel(&cfs_b->slack_timer);
2738 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2740 struct cfs_rq *cfs_rq;
2742 for_each_leaf_cfs_rq(rq, cfs_rq) {
2743 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2745 if (!cfs_rq->runtime_enabled)
2749 * clock_task is not advancing so we just need to make sure
2750 * there's some valid quota amount
2752 cfs_rq->runtime_remaining = cfs_b->quota;
2753 if (cfs_rq_throttled(cfs_rq))
2754 unthrottle_cfs_rq(cfs_rq);
2758 #else /* CONFIG_CFS_BANDWIDTH */
2759 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2761 return rq_clock_task(rq_of(cfs_rq));
2764 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2765 unsigned long delta_exec) {}
2766 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2767 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2768 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2770 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2775 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2780 static inline int throttled_lb_pair(struct task_group *tg,
2781 int src_cpu, int dest_cpu)
2786 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2788 #ifdef CONFIG_FAIR_GROUP_SCHED
2789 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2792 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2796 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2797 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2799 #endif /* CONFIG_CFS_BANDWIDTH */
2801 /**************************************************
2802 * CFS operations on tasks:
2805 #ifdef CONFIG_SCHED_HRTICK
2806 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2808 struct sched_entity *se = &p->se;
2809 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2811 WARN_ON(task_rq(p) != rq);
2813 if (cfs_rq->nr_running > 1) {
2814 u64 slice = sched_slice(cfs_rq, se);
2815 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2816 s64 delta = slice - ran;
2825 * Don't schedule slices shorter than 10000ns, that just
2826 * doesn't make sense. Rely on vruntime for fairness.
2829 delta = max_t(s64, 10000LL, delta);
2831 hrtick_start(rq, delta);
2836 * called from enqueue/dequeue and updates the hrtick when the
2837 * current task is from our class and nr_running is low enough
2840 static void hrtick_update(struct rq *rq)
2842 struct task_struct *curr = rq->curr;
2844 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2847 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2848 hrtick_start_fair(rq, curr);
2850 #else /* !CONFIG_SCHED_HRTICK */
2852 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2856 static inline void hrtick_update(struct rq *rq)
2862 * The enqueue_task method is called before nr_running is
2863 * increased. Here we update the fair scheduling stats and
2864 * then put the task into the rbtree:
2867 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2869 struct cfs_rq *cfs_rq;
2870 struct sched_entity *se = &p->se;
2872 for_each_sched_entity(se) {
2875 cfs_rq = cfs_rq_of(se);
2876 enqueue_entity(cfs_rq, se, flags);
2879 * end evaluation on encountering a throttled cfs_rq
2881 * note: in the case of encountering a throttled cfs_rq we will
2882 * post the final h_nr_running increment below.
2884 if (cfs_rq_throttled(cfs_rq))
2886 cfs_rq->h_nr_running++;
2888 flags = ENQUEUE_WAKEUP;
2891 for_each_sched_entity(se) {
2892 cfs_rq = cfs_rq_of(se);
2893 cfs_rq->h_nr_running++;
2895 if (cfs_rq_throttled(cfs_rq))
2898 update_cfs_shares(cfs_rq);
2899 update_entity_load_avg(se, 1);
2903 update_rq_runnable_avg(rq, rq->nr_running);
2909 static void set_next_buddy(struct sched_entity *se);
2912 * The dequeue_task method is called before nr_running is
2913 * decreased. We remove the task from the rbtree and
2914 * update the fair scheduling stats:
2916 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2918 struct cfs_rq *cfs_rq;
2919 struct sched_entity *se = &p->se;
2920 int task_sleep = flags & DEQUEUE_SLEEP;
2922 for_each_sched_entity(se) {
2923 cfs_rq = cfs_rq_of(se);
2924 dequeue_entity(cfs_rq, se, flags);
2927 * end evaluation on encountering a throttled cfs_rq
2929 * note: in the case of encountering a throttled cfs_rq we will
2930 * post the final h_nr_running decrement below.
2932 if (cfs_rq_throttled(cfs_rq))
2934 cfs_rq->h_nr_running--;
2936 /* Don't dequeue parent if it has other entities besides us */
2937 if (cfs_rq->load.weight) {
2939 * Bias pick_next to pick a task from this cfs_rq, as
2940 * p is sleeping when it is within its sched_slice.
2942 if (task_sleep && parent_entity(se))
2943 set_next_buddy(parent_entity(se));
2945 /* avoid re-evaluating load for this entity */
2946 se = parent_entity(se);
2949 flags |= DEQUEUE_SLEEP;
2952 for_each_sched_entity(se) {
2953 cfs_rq = cfs_rq_of(se);
2954 cfs_rq->h_nr_running--;
2956 if (cfs_rq_throttled(cfs_rq))
2959 update_cfs_shares(cfs_rq);
2960 update_entity_load_avg(se, 1);
2965 update_rq_runnable_avg(rq, 1);
2971 /* Used instead of source_load when we know the type == 0 */
2972 static unsigned long weighted_cpuload(const int cpu)
2974 return cpu_rq(cpu)->cfs.runnable_load_avg;
2978 * Return a low guess at the load of a migration-source cpu weighted
2979 * according to the scheduling class and "nice" value.
2981 * We want to under-estimate the load of migration sources, to
2982 * balance conservatively.
2984 static unsigned long source_load(int cpu, int type)
2986 struct rq *rq = cpu_rq(cpu);
2987 unsigned long total = weighted_cpuload(cpu);
2989 if (type == 0 || !sched_feat(LB_BIAS))
2992 return min(rq->cpu_load[type-1], total);
2996 * Return a high guess at the load of a migration-target cpu weighted
2997 * according to the scheduling class and "nice" value.
2999 static unsigned long target_load(int cpu, int type)
3001 struct rq *rq = cpu_rq(cpu);
3002 unsigned long total = weighted_cpuload(cpu);
3004 if (type == 0 || !sched_feat(LB_BIAS))
3007 return max(rq->cpu_load[type-1], total);
3010 static unsigned long power_of(int cpu)
3012 return cpu_rq(cpu)->cpu_power;
3015 static unsigned long cpu_avg_load_per_task(int cpu)
3017 struct rq *rq = cpu_rq(cpu);
3018 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3019 unsigned long load_avg = rq->cfs.runnable_load_avg;
3022 return load_avg / nr_running;
3027 static void record_wakee(struct task_struct *p)
3030 * Rough decay (wiping) for cost saving, don't worry
3031 * about the boundary, really active task won't care
3034 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3035 current->wakee_flips = 0;
3036 current->wakee_flip_decay_ts = jiffies;
3039 if (current->last_wakee != p) {
3040 current->last_wakee = p;
3041 current->wakee_flips++;
3045 static void task_waking_fair(struct task_struct *p)
3047 struct sched_entity *se = &p->se;
3048 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3051 #ifndef CONFIG_64BIT
3052 u64 min_vruntime_copy;
3055 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3057 min_vruntime = cfs_rq->min_vruntime;
3058 } while (min_vruntime != min_vruntime_copy);
3060 min_vruntime = cfs_rq->min_vruntime;
3063 se->vruntime -= min_vruntime;
3067 #ifdef CONFIG_FAIR_GROUP_SCHED
3069 * effective_load() calculates the load change as seen from the root_task_group
3071 * Adding load to a group doesn't make a group heavier, but can cause movement
3072 * of group shares between cpus. Assuming the shares were perfectly aligned one
3073 * can calculate the shift in shares.
3075 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3076 * on this @cpu and results in a total addition (subtraction) of @wg to the
3077 * total group weight.
3079 * Given a runqueue weight distribution (rw_i) we can compute a shares
3080 * distribution (s_i) using:
3082 * s_i = rw_i / \Sum rw_j (1)
3084 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3085 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3086 * shares distribution (s_i):
3088 * rw_i = { 2, 4, 1, 0 }
3089 * s_i = { 2/7, 4/7, 1/7, 0 }
3091 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3092 * task used to run on and the CPU the waker is running on), we need to
3093 * compute the effect of waking a task on either CPU and, in case of a sync
3094 * wakeup, compute the effect of the current task going to sleep.
3096 * So for a change of @wl to the local @cpu with an overall group weight change
3097 * of @wl we can compute the new shares distribution (s'_i) using:
3099 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3101 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3102 * differences in waking a task to CPU 0. The additional task changes the
3103 * weight and shares distributions like:
3105 * rw'_i = { 3, 4, 1, 0 }
3106 * s'_i = { 3/8, 4/8, 1/8, 0 }
3108 * We can then compute the difference in effective weight by using:
3110 * dw_i = S * (s'_i - s_i) (3)
3112 * Where 'S' is the group weight as seen by its parent.
3114 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3115 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3116 * 4/7) times the weight of the group.
3118 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3120 struct sched_entity *se = tg->se[cpu];
3122 if (!tg->parent) /* the trivial, non-cgroup case */
3125 for_each_sched_entity(se) {
3131 * W = @wg + \Sum rw_j
3133 W = wg + calc_tg_weight(tg, se->my_q);
3138 w = se->my_q->load.weight + wl;
3141 * wl = S * s'_i; see (2)
3144 wl = (w * tg->shares) / W;
3149 * Per the above, wl is the new se->load.weight value; since
3150 * those are clipped to [MIN_SHARES, ...) do so now. See
3151 * calc_cfs_shares().
3153 if (wl < MIN_SHARES)
3157 * wl = dw_i = S * (s'_i - s_i); see (3)
3159 wl -= se->load.weight;
3162 * Recursively apply this logic to all parent groups to compute
3163 * the final effective load change on the root group. Since
3164 * only the @tg group gets extra weight, all parent groups can
3165 * only redistribute existing shares. @wl is the shift in shares
3166 * resulting from this level per the above.
3175 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3176 unsigned long wl, unsigned long wg)
3183 static int wake_wide(struct task_struct *p)
3185 int factor = this_cpu_read(sd_llc_size);
3188 * Yeah, it's the switching-frequency, could means many wakee or
3189 * rapidly switch, use factor here will just help to automatically
3190 * adjust the loose-degree, so bigger node will lead to more pull.
3192 if (p->wakee_flips > factor) {
3194 * wakee is somewhat hot, it needs certain amount of cpu
3195 * resource, so if waker is far more hot, prefer to leave
3198 if (current->wakee_flips > (factor * p->wakee_flips))
3205 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3207 s64 this_load, load;
3208 int idx, this_cpu, prev_cpu;
3209 unsigned long tl_per_task;
3210 struct task_group *tg;
3211 unsigned long weight;
3215 * If we wake multiple tasks be careful to not bounce
3216 * ourselves around too much.
3222 this_cpu = smp_processor_id();
3223 prev_cpu = task_cpu(p);
3224 load = source_load(prev_cpu, idx);
3225 this_load = target_load(this_cpu, idx);
3228 * If sync wakeup then subtract the (maximum possible)
3229 * effect of the currently running task from the load
3230 * of the current CPU:
3233 tg = task_group(current);
3234 weight = current->se.load.weight;
3236 this_load += effective_load(tg, this_cpu, -weight, -weight);
3237 load += effective_load(tg, prev_cpu, 0, -weight);
3241 weight = p->se.load.weight;
3244 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3245 * due to the sync cause above having dropped this_load to 0, we'll
3246 * always have an imbalance, but there's really nothing you can do
3247 * about that, so that's good too.
3249 * Otherwise check if either cpus are near enough in load to allow this
3250 * task to be woken on this_cpu.
3252 if (this_load > 0) {
3253 s64 this_eff_load, prev_eff_load;
3255 this_eff_load = 100;
3256 this_eff_load *= power_of(prev_cpu);
3257 this_eff_load *= this_load +
3258 effective_load(tg, this_cpu, weight, weight);
3260 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3261 prev_eff_load *= power_of(this_cpu);
3262 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3264 balanced = this_eff_load <= prev_eff_load;
3269 * If the currently running task will sleep within
3270 * a reasonable amount of time then attract this newly
3273 if (sync && balanced)
3276 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3277 tl_per_task = cpu_avg_load_per_task(this_cpu);
3280 (this_load <= load &&
3281 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3283 * This domain has SD_WAKE_AFFINE and
3284 * p is cache cold in this domain, and
3285 * there is no bad imbalance.
3287 schedstat_inc(sd, ttwu_move_affine);
3288 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3296 * find_idlest_group finds and returns the least busy CPU group within the
3299 static struct sched_group *
3300 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3301 int this_cpu, int load_idx)
3303 struct sched_group *idlest = NULL, *group = sd->groups;
3304 unsigned long min_load = ULONG_MAX, this_load = 0;
3305 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3308 unsigned long load, avg_load;
3312 /* Skip over this group if it has no CPUs allowed */
3313 if (!cpumask_intersects(sched_group_cpus(group),
3314 tsk_cpus_allowed(p)))
3317 local_group = cpumask_test_cpu(this_cpu,
3318 sched_group_cpus(group));
3320 /* Tally up the load of all CPUs in the group */
3323 for_each_cpu(i, sched_group_cpus(group)) {
3324 /* Bias balancing toward cpus of our domain */
3326 load = source_load(i, load_idx);
3328 load = target_load(i, load_idx);
3333 /* Adjust by relative CPU power of the group */
3334 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3337 this_load = avg_load;
3338 } else if (avg_load < min_load) {
3339 min_load = avg_load;
3342 } while (group = group->next, group != sd->groups);
3344 if (!idlest || 100*this_load < imbalance*min_load)
3350 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3353 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3355 unsigned long load, min_load = ULONG_MAX;
3359 /* Traverse only the allowed CPUs */
3360 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3361 load = weighted_cpuload(i);
3363 if (load < min_load || (load == min_load && i == this_cpu)) {
3373 * Try and locate an idle CPU in the sched_domain.
3375 static int select_idle_sibling(struct task_struct *p, int target)
3377 struct sched_domain *sd;
3378 struct sched_group *sg;
3379 int i = task_cpu(p);
3381 if (idle_cpu(target))
3385 * If the prevous cpu is cache affine and idle, don't be stupid.
3387 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3391 * Otherwise, iterate the domains and find an elegible idle cpu.
3393 sd = rcu_dereference(per_cpu(sd_llc, target));
3394 for_each_lower_domain(sd) {
3397 if (!cpumask_intersects(sched_group_cpus(sg),
3398 tsk_cpus_allowed(p)))
3401 for_each_cpu(i, sched_group_cpus(sg)) {
3402 if (i == target || !idle_cpu(i))
3406 target = cpumask_first_and(sched_group_cpus(sg),
3407 tsk_cpus_allowed(p));
3411 } while (sg != sd->groups);
3418 * sched_balance_self: balance the current task (running on cpu) in domains
3419 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3422 * Balance, ie. select the least loaded group.
3424 * Returns the target CPU number, or the same CPU if no balancing is needed.
3426 * preempt must be disabled.
3429 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3431 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3432 int cpu = smp_processor_id();
3433 int prev_cpu = task_cpu(p);
3435 int want_affine = 0;
3436 int sync = wake_flags & WF_SYNC;
3438 if (p->nr_cpus_allowed == 1)
3441 if (sd_flag & SD_BALANCE_WAKE) {
3442 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3448 for_each_domain(cpu, tmp) {
3449 if (!(tmp->flags & SD_LOAD_BALANCE))
3453 * If both cpu and prev_cpu are part of this domain,
3454 * cpu is a valid SD_WAKE_AFFINE target.
3456 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3457 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3462 if (tmp->flags & sd_flag)
3467 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3470 new_cpu = select_idle_sibling(p, prev_cpu);
3475 int load_idx = sd->forkexec_idx;
3476 struct sched_group *group;
3479 if (!(sd->flags & sd_flag)) {
3484 if (sd_flag & SD_BALANCE_WAKE)
3485 load_idx = sd->wake_idx;
3487 group = find_idlest_group(sd, p, cpu, load_idx);
3493 new_cpu = find_idlest_cpu(group, p, cpu);
3494 if (new_cpu == -1 || new_cpu == cpu) {
3495 /* Now try balancing at a lower domain level of cpu */
3500 /* Now try balancing at a lower domain level of new_cpu */
3502 weight = sd->span_weight;
3504 for_each_domain(cpu, tmp) {
3505 if (weight <= tmp->span_weight)
3507 if (tmp->flags & sd_flag)
3510 /* while loop will break here if sd == NULL */
3519 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3520 * cfs_rq_of(p) references at time of call are still valid and identify the
3521 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3522 * other assumptions, including the state of rq->lock, should be made.
3525 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3527 struct sched_entity *se = &p->se;
3528 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3531 * Load tracking: accumulate removed load so that it can be processed
3532 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3533 * to blocked load iff they have a positive decay-count. It can never
3534 * be negative here since on-rq tasks have decay-count == 0.
3536 if (se->avg.decay_count) {
3537 se->avg.decay_count = -__synchronize_entity_decay(se);
3538 atomic_long_add(se->avg.load_avg_contrib,
3539 &cfs_rq->removed_load);
3542 #endif /* CONFIG_SMP */
3544 static unsigned long
3545 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3547 unsigned long gran = sysctl_sched_wakeup_granularity;
3550 * Since its curr running now, convert the gran from real-time
3551 * to virtual-time in his units.
3553 * By using 'se' instead of 'curr' we penalize light tasks, so
3554 * they get preempted easier. That is, if 'se' < 'curr' then
3555 * the resulting gran will be larger, therefore penalizing the
3556 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3557 * be smaller, again penalizing the lighter task.
3559 * This is especially important for buddies when the leftmost
3560 * task is higher priority than the buddy.
3562 return calc_delta_fair(gran, se);
3566 * Should 'se' preempt 'curr'.
3580 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3582 s64 gran, vdiff = curr->vruntime - se->vruntime;
3587 gran = wakeup_gran(curr, se);
3594 static void set_last_buddy(struct sched_entity *se)
3596 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3599 for_each_sched_entity(se)
3600 cfs_rq_of(se)->last = se;
3603 static void set_next_buddy(struct sched_entity *se)
3605 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3608 for_each_sched_entity(se)
3609 cfs_rq_of(se)->next = se;
3612 static void set_skip_buddy(struct sched_entity *se)
3614 for_each_sched_entity(se)
3615 cfs_rq_of(se)->skip = se;
3619 * Preempt the current task with a newly woken task if needed:
3621 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3623 struct task_struct *curr = rq->curr;
3624 struct sched_entity *se = &curr->se, *pse = &p->se;
3625 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3626 int scale = cfs_rq->nr_running >= sched_nr_latency;
3627 int next_buddy_marked = 0;
3629 if (unlikely(se == pse))
3633 * This is possible from callers such as move_task(), in which we
3634 * unconditionally check_prempt_curr() after an enqueue (which may have
3635 * lead to a throttle). This both saves work and prevents false
3636 * next-buddy nomination below.
3638 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3641 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3642 set_next_buddy(pse);
3643 next_buddy_marked = 1;
3647 * We can come here with TIF_NEED_RESCHED already set from new task
3650 * Note: this also catches the edge-case of curr being in a throttled
3651 * group (e.g. via set_curr_task), since update_curr() (in the
3652 * enqueue of curr) will have resulted in resched being set. This
3653 * prevents us from potentially nominating it as a false LAST_BUDDY
3656 if (test_tsk_need_resched(curr))
3659 /* Idle tasks are by definition preempted by non-idle tasks. */
3660 if (unlikely(curr->policy == SCHED_IDLE) &&
3661 likely(p->policy != SCHED_IDLE))
3665 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3666 * is driven by the tick):
3668 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3671 find_matching_se(&se, &pse);
3672 update_curr(cfs_rq_of(se));
3674 if (wakeup_preempt_entity(se, pse) == 1) {
3676 * Bias pick_next to pick the sched entity that is
3677 * triggering this preemption.
3679 if (!next_buddy_marked)
3680 set_next_buddy(pse);
3689 * Only set the backward buddy when the current task is still
3690 * on the rq. This can happen when a wakeup gets interleaved
3691 * with schedule on the ->pre_schedule() or idle_balance()
3692 * point, either of which can * drop the rq lock.
3694 * Also, during early boot the idle thread is in the fair class,
3695 * for obvious reasons its a bad idea to schedule back to it.
3697 if (unlikely(!se->on_rq || curr == rq->idle))
3700 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3704 static struct task_struct *pick_next_task_fair(struct rq *rq)
3706 struct task_struct *p;
3707 struct cfs_rq *cfs_rq = &rq->cfs;
3708 struct sched_entity *se;
3710 if (!cfs_rq->nr_running)
3714 se = pick_next_entity(cfs_rq);
3715 set_next_entity(cfs_rq, se);
3716 cfs_rq = group_cfs_rq(se);
3720 if (hrtick_enabled(rq))
3721 hrtick_start_fair(rq, p);
3727 * Account for a descheduled task:
3729 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3731 struct sched_entity *se = &prev->se;
3732 struct cfs_rq *cfs_rq;
3734 for_each_sched_entity(se) {
3735 cfs_rq = cfs_rq_of(se);
3736 put_prev_entity(cfs_rq, se);
3741 * sched_yield() is very simple
3743 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3745 static void yield_task_fair(struct rq *rq)
3747 struct task_struct *curr = rq->curr;
3748 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3749 struct sched_entity *se = &curr->se;
3752 * Are we the only task in the tree?
3754 if (unlikely(rq->nr_running == 1))
3757 clear_buddies(cfs_rq, se);
3759 if (curr->policy != SCHED_BATCH) {
3760 update_rq_clock(rq);
3762 * Update run-time statistics of the 'current'.
3764 update_curr(cfs_rq);
3766 * Tell update_rq_clock() that we've just updated,
3767 * so we don't do microscopic update in schedule()
3768 * and double the fastpath cost.
3770 rq->skip_clock_update = 1;
3776 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3778 struct sched_entity *se = &p->se;
3780 /* throttled hierarchies are not runnable */
3781 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3784 /* Tell the scheduler that we'd really like pse to run next. */
3787 yield_task_fair(rq);
3793 /**************************************************
3794 * Fair scheduling class load-balancing methods.
3798 * The purpose of load-balancing is to achieve the same basic fairness the
3799 * per-cpu scheduler provides, namely provide a proportional amount of compute
3800 * time to each task. This is expressed in the following equation:
3802 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3804 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3805 * W_i,0 is defined as:
3807 * W_i,0 = \Sum_j w_i,j (2)
3809 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3810 * is derived from the nice value as per prio_to_weight[].
3812 * The weight average is an exponential decay average of the instantaneous
3815 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3817 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3818 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3819 * can also include other factors [XXX].
3821 * To achieve this balance we define a measure of imbalance which follows
3822 * directly from (1):
3824 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3826 * We them move tasks around to minimize the imbalance. In the continuous
3827 * function space it is obvious this converges, in the discrete case we get
3828 * a few fun cases generally called infeasible weight scenarios.
3831 * - infeasible weights;
3832 * - local vs global optima in the discrete case. ]
3837 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3838 * for all i,j solution, we create a tree of cpus that follows the hardware
3839 * topology where each level pairs two lower groups (or better). This results
3840 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3841 * tree to only the first of the previous level and we decrease the frequency
3842 * of load-balance at each level inv. proportional to the number of cpus in
3848 * \Sum { --- * --- * 2^i } = O(n) (5)
3850 * `- size of each group
3851 * | | `- number of cpus doing load-balance
3853 * `- sum over all levels
3855 * Coupled with a limit on how many tasks we can migrate every balance pass,
3856 * this makes (5) the runtime complexity of the balancer.
3858 * An important property here is that each CPU is still (indirectly) connected
3859 * to every other cpu in at most O(log n) steps:
3861 * The adjacency matrix of the resulting graph is given by:
3864 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3867 * And you'll find that:
3869 * A^(log_2 n)_i,j != 0 for all i,j (7)
3871 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3872 * The task movement gives a factor of O(m), giving a convergence complexity
3875 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3880 * In order to avoid CPUs going idle while there's still work to do, new idle
3881 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3882 * tree itself instead of relying on other CPUs to bring it work.
3884 * This adds some complexity to both (5) and (8) but it reduces the total idle
3892 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3895 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3900 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3902 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3904 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3907 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3908 * rewrite all of this once again.]
3911 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3913 #define LBF_ALL_PINNED 0x01
3914 #define LBF_NEED_BREAK 0x02
3915 #define LBF_DST_PINNED 0x04
3916 #define LBF_SOME_PINNED 0x08
3919 struct sched_domain *sd;
3927 struct cpumask *dst_grpmask;
3929 enum cpu_idle_type idle;
3931 /* The set of CPUs under consideration for load-balancing */
3932 struct cpumask *cpus;
3937 unsigned int loop_break;
3938 unsigned int loop_max;
3942 * move_task - move a task from one runqueue to another runqueue.
3943 * Both runqueues must be locked.
3945 static void move_task(struct task_struct *p, struct lb_env *env)
3947 deactivate_task(env->src_rq, p, 0);
3948 set_task_cpu(p, env->dst_cpu);
3949 activate_task(env->dst_rq, p, 0);
3950 check_preempt_curr(env->dst_rq, p, 0);
3954 * Is this task likely cache-hot:
3957 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3961 if (p->sched_class != &fair_sched_class)
3964 if (unlikely(p->policy == SCHED_IDLE))
3968 * Buddy candidates are cache hot:
3970 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3971 (&p->se == cfs_rq_of(&p->se)->next ||
3972 &p->se == cfs_rq_of(&p->se)->last))
3975 if (sysctl_sched_migration_cost == -1)
3977 if (sysctl_sched_migration_cost == 0)
3980 delta = now - p->se.exec_start;
3982 return delta < (s64)sysctl_sched_migration_cost;
3986 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3989 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3991 int tsk_cache_hot = 0;
3993 * We do not migrate tasks that are:
3994 * 1) throttled_lb_pair, or
3995 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3996 * 3) running (obviously), or
3997 * 4) are cache-hot on their current CPU.
3999 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4002 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4005 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4007 env->flags |= LBF_SOME_PINNED;
4010 * Remember if this task can be migrated to any other cpu in
4011 * our sched_group. We may want to revisit it if we couldn't
4012 * meet load balance goals by pulling other tasks on src_cpu.
4014 * Also avoid computing new_dst_cpu if we have already computed
4015 * one in current iteration.
4017 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4020 /* Prevent to re-select dst_cpu via env's cpus */
4021 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4022 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4023 env->flags |= LBF_DST_PINNED;
4024 env->new_dst_cpu = cpu;
4032 /* Record that we found atleast one task that could run on dst_cpu */
4033 env->flags &= ~LBF_ALL_PINNED;
4035 if (task_running(env->src_rq, p)) {
4036 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4041 * Aggressive migration if:
4042 * 1) task is cache cold, or
4043 * 2) too many balance attempts have failed.
4046 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4047 if (!tsk_cache_hot ||
4048 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4050 if (tsk_cache_hot) {
4051 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4052 schedstat_inc(p, se.statistics.nr_forced_migrations);
4058 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4063 * move_one_task tries to move exactly one task from busiest to this_rq, as
4064 * part of active balancing operations within "domain".
4065 * Returns 1 if successful and 0 otherwise.
4067 * Called with both runqueues locked.
4069 static int move_one_task(struct lb_env *env)
4071 struct task_struct *p, *n;
4073 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4074 if (!can_migrate_task(p, env))
4079 * Right now, this is only the second place move_task()
4080 * is called, so we can safely collect move_task()
4081 * stats here rather than inside move_task().
4083 schedstat_inc(env->sd, lb_gained[env->idle]);
4089 static unsigned long task_h_load(struct task_struct *p);
4091 static const unsigned int sched_nr_migrate_break = 32;
4094 * move_tasks tries to move up to imbalance weighted load from busiest to
4095 * this_rq, as part of a balancing operation within domain "sd".
4096 * Returns 1 if successful and 0 otherwise.
4098 * Called with both runqueues locked.
4100 static int move_tasks(struct lb_env *env)
4102 struct list_head *tasks = &env->src_rq->cfs_tasks;
4103 struct task_struct *p;
4107 if (env->imbalance <= 0)
4110 while (!list_empty(tasks)) {
4111 p = list_first_entry(tasks, struct task_struct, se.group_node);
4114 /* We've more or less seen every task there is, call it quits */
4115 if (env->loop > env->loop_max)
4118 /* take a breather every nr_migrate tasks */
4119 if (env->loop > env->loop_break) {
4120 env->loop_break += sched_nr_migrate_break;
4121 env->flags |= LBF_NEED_BREAK;
4125 if (!can_migrate_task(p, env))
4128 load = task_h_load(p);
4130 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4133 if ((load / 2) > env->imbalance)
4138 env->imbalance -= load;
4140 #ifdef CONFIG_PREEMPT
4142 * NEWIDLE balancing is a source of latency, so preemptible
4143 * kernels will stop after the first task is pulled to minimize
4144 * the critical section.
4146 if (env->idle == CPU_NEWLY_IDLE)
4151 * We only want to steal up to the prescribed amount of
4154 if (env->imbalance <= 0)
4159 list_move_tail(&p->se.group_node, tasks);
4163 * Right now, this is one of only two places move_task() is called,
4164 * so we can safely collect move_task() stats here rather than
4165 * inside move_task().
4167 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4172 #ifdef CONFIG_FAIR_GROUP_SCHED
4174 * update tg->load_weight by folding this cpu's load_avg
4176 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4178 struct sched_entity *se = tg->se[cpu];
4179 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4181 /* throttled entities do not contribute to load */
4182 if (throttled_hierarchy(cfs_rq))
4185 update_cfs_rq_blocked_load(cfs_rq, 1);
4188 update_entity_load_avg(se, 1);
4190 * We pivot on our runnable average having decayed to zero for
4191 * list removal. This generally implies that all our children
4192 * have also been removed (modulo rounding error or bandwidth
4193 * control); however, such cases are rare and we can fix these
4196 * TODO: fix up out-of-order children on enqueue.
4198 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4199 list_del_leaf_cfs_rq(cfs_rq);
4201 struct rq *rq = rq_of(cfs_rq);
4202 update_rq_runnable_avg(rq, rq->nr_running);
4206 static void update_blocked_averages(int cpu)
4208 struct rq *rq = cpu_rq(cpu);
4209 struct cfs_rq *cfs_rq;
4210 unsigned long flags;
4212 raw_spin_lock_irqsave(&rq->lock, flags);
4213 update_rq_clock(rq);
4215 * Iterates the task_group tree in a bottom up fashion, see
4216 * list_add_leaf_cfs_rq() for details.
4218 for_each_leaf_cfs_rq(rq, cfs_rq) {
4220 * Note: We may want to consider periodically releasing
4221 * rq->lock about these updates so that creating many task
4222 * groups does not result in continually extending hold time.
4224 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4227 raw_spin_unlock_irqrestore(&rq->lock, flags);
4231 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4232 * This needs to be done in a top-down fashion because the load of a child
4233 * group is a fraction of its parents load.
4235 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4237 struct rq *rq = rq_of(cfs_rq);
4238 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4239 unsigned long now = jiffies;
4242 if (cfs_rq->last_h_load_update == now)
4245 cfs_rq->h_load_next = NULL;
4246 for_each_sched_entity(se) {
4247 cfs_rq = cfs_rq_of(se);
4248 cfs_rq->h_load_next = se;
4249 if (cfs_rq->last_h_load_update == now)
4254 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4255 cfs_rq->last_h_load_update = now;
4258 while ((se = cfs_rq->h_load_next) != NULL) {
4259 load = cfs_rq->h_load;
4260 load = div64_ul(load * se->avg.load_avg_contrib,
4261 cfs_rq->runnable_load_avg + 1);
4262 cfs_rq = group_cfs_rq(se);
4263 cfs_rq->h_load = load;
4264 cfs_rq->last_h_load_update = now;
4268 static unsigned long task_h_load(struct task_struct *p)
4270 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4272 update_cfs_rq_h_load(cfs_rq);
4273 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4274 cfs_rq->runnable_load_avg + 1);
4277 static inline void update_blocked_averages(int cpu)
4281 static unsigned long task_h_load(struct task_struct *p)
4283 return p->se.avg.load_avg_contrib;
4287 /********** Helpers for find_busiest_group ************************/
4289 * sg_lb_stats - stats of a sched_group required for load_balancing
4291 struct sg_lb_stats {
4292 unsigned long avg_load; /*Avg load across the CPUs of the group */
4293 unsigned long group_load; /* Total load over the CPUs of the group */
4294 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4295 unsigned long load_per_task;
4296 unsigned long group_power;
4297 unsigned int sum_nr_running; /* Nr tasks running in the group */
4298 unsigned int group_capacity;
4299 unsigned int idle_cpus;
4300 unsigned int group_weight;
4301 int group_imb; /* Is there an imbalance in the group ? */
4302 int group_has_capacity; /* Is there extra capacity in the group? */
4306 * sd_lb_stats - Structure to store the statistics of a sched_domain
4307 * during load balancing.
4309 struct sd_lb_stats {
4310 struct sched_group *busiest; /* Busiest group in this sd */
4311 struct sched_group *local; /* Local group in this sd */
4312 unsigned long total_load; /* Total load of all groups in sd */
4313 unsigned long total_pwr; /* Total power of all groups in sd */
4314 unsigned long avg_load; /* Average load across all groups in sd */
4316 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4317 struct sg_lb_stats local_stat; /* Statistics of the local group */
4320 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4323 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4324 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4325 * We must however clear busiest_stat::avg_load because
4326 * update_sd_pick_busiest() reads this before assignment.
4328 *sds = (struct sd_lb_stats){
4340 * get_sd_load_idx - Obtain the load index for a given sched domain.
4341 * @sd: The sched_domain whose load_idx is to be obtained.
4342 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4344 * Return: The load index.
4346 static inline int get_sd_load_idx(struct sched_domain *sd,
4347 enum cpu_idle_type idle)
4353 load_idx = sd->busy_idx;
4356 case CPU_NEWLY_IDLE:
4357 load_idx = sd->newidle_idx;
4360 load_idx = sd->idle_idx;
4367 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4369 return SCHED_POWER_SCALE;
4372 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4374 return default_scale_freq_power(sd, cpu);
4377 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4379 unsigned long weight = sd->span_weight;
4380 unsigned long smt_gain = sd->smt_gain;
4387 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4389 return default_scale_smt_power(sd, cpu);
4392 static unsigned long scale_rt_power(int cpu)
4394 struct rq *rq = cpu_rq(cpu);
4395 u64 total, available, age_stamp, avg;
4398 * Since we're reading these variables without serialization make sure
4399 * we read them once before doing sanity checks on them.
4401 age_stamp = ACCESS_ONCE(rq->age_stamp);
4402 avg = ACCESS_ONCE(rq->rt_avg);
4404 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4406 if (unlikely(total < avg)) {
4407 /* Ensures that power won't end up being negative */
4410 available = total - avg;
4413 if (unlikely((s64)total < SCHED_POWER_SCALE))
4414 total = SCHED_POWER_SCALE;
4416 total >>= SCHED_POWER_SHIFT;
4418 return div_u64(available, total);
4421 static void update_cpu_power(struct sched_domain *sd, int cpu)
4423 unsigned long weight = sd->span_weight;
4424 unsigned long power = SCHED_POWER_SCALE;
4425 struct sched_group *sdg = sd->groups;
4427 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4428 if (sched_feat(ARCH_POWER))
4429 power *= arch_scale_smt_power(sd, cpu);
4431 power *= default_scale_smt_power(sd, cpu);
4433 power >>= SCHED_POWER_SHIFT;
4436 sdg->sgp->power_orig = power;
4438 if (sched_feat(ARCH_POWER))
4439 power *= arch_scale_freq_power(sd, cpu);
4441 power *= default_scale_freq_power(sd, cpu);
4443 power >>= SCHED_POWER_SHIFT;
4445 power *= scale_rt_power(cpu);
4446 power >>= SCHED_POWER_SHIFT;
4451 cpu_rq(cpu)->cpu_power = power;
4452 sdg->sgp->power = power;
4455 void update_group_power(struct sched_domain *sd, int cpu)
4457 struct sched_domain *child = sd->child;
4458 struct sched_group *group, *sdg = sd->groups;
4459 unsigned long power, power_orig;
4460 unsigned long interval;
4462 interval = msecs_to_jiffies(sd->balance_interval);
4463 interval = clamp(interval, 1UL, max_load_balance_interval);
4464 sdg->sgp->next_update = jiffies + interval;
4467 update_cpu_power(sd, cpu);
4471 power_orig = power = 0;
4473 if (child->flags & SD_OVERLAP) {
4475 * SD_OVERLAP domains cannot assume that child groups
4476 * span the current group.
4479 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4480 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4482 power_orig += sg->sgp->power_orig;
4483 power += sg->sgp->power;
4487 * !SD_OVERLAP domains can assume that child groups
4488 * span the current group.
4491 group = child->groups;
4493 power_orig += group->sgp->power_orig;
4494 power += group->sgp->power;
4495 group = group->next;
4496 } while (group != child->groups);
4499 sdg->sgp->power_orig = power_orig;
4500 sdg->sgp->power = power;
4504 * Try and fix up capacity for tiny siblings, this is needed when
4505 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4506 * which on its own isn't powerful enough.
4508 * See update_sd_pick_busiest() and check_asym_packing().
4511 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4514 * Only siblings can have significantly less than SCHED_POWER_SCALE
4516 if (!(sd->flags & SD_SHARE_CPUPOWER))
4520 * If ~90% of the cpu_power is still there, we're good.
4522 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4529 * Group imbalance indicates (and tries to solve) the problem where balancing
4530 * groups is inadequate due to tsk_cpus_allowed() constraints.
4532 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4533 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4536 * { 0 1 2 3 } { 4 5 6 7 }
4539 * If we were to balance group-wise we'd place two tasks in the first group and
4540 * two tasks in the second group. Clearly this is undesired as it will overload
4541 * cpu 3 and leave one of the cpus in the second group unused.
4543 * The current solution to this issue is detecting the skew in the first group
4544 * by noticing the lower domain failed to reach balance and had difficulty
4545 * moving tasks due to affinity constraints.
4547 * When this is so detected; this group becomes a candidate for busiest; see
4548 * update_sd_pick_busiest(). And calculcate_imbalance() and
4549 * find_busiest_group() avoid some of the usual balance conditions to allow it
4550 * to create an effective group imbalance.
4552 * This is a somewhat tricky proposition since the next run might not find the
4553 * group imbalance and decide the groups need to be balanced again. A most
4554 * subtle and fragile situation.
4557 static inline int sg_imbalanced(struct sched_group *group)
4559 return group->sgp->imbalance;
4563 * Compute the group capacity.
4565 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4566 * first dividing out the smt factor and computing the actual number of cores
4567 * and limit power unit capacity with that.
4569 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4571 unsigned int capacity, smt, cpus;
4572 unsigned int power, power_orig;
4574 power = group->sgp->power;
4575 power_orig = group->sgp->power_orig;
4576 cpus = group->group_weight;
4578 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4579 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4580 capacity = cpus / smt; /* cores */
4582 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4584 capacity = fix_small_capacity(env->sd, group);
4590 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4591 * @env: The load balancing environment.
4592 * @group: sched_group whose statistics are to be updated.
4593 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4594 * @local_group: Does group contain this_cpu.
4595 * @sgs: variable to hold the statistics for this group.
4597 static inline void update_sg_lb_stats(struct lb_env *env,
4598 struct sched_group *group, int load_idx,
4599 int local_group, struct sg_lb_stats *sgs)
4601 unsigned long nr_running;
4605 memset(sgs, 0, sizeof(*sgs));
4607 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4608 struct rq *rq = cpu_rq(i);
4610 nr_running = rq->nr_running;
4612 /* Bias balancing toward cpus of our domain */
4614 load = target_load(i, load_idx);
4616 load = source_load(i, load_idx);
4618 sgs->group_load += load;
4619 sgs->sum_nr_running += nr_running;
4620 sgs->sum_weighted_load += weighted_cpuload(i);
4625 /* Adjust by relative CPU power of the group */
4626 sgs->group_power = group->sgp->power;
4627 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4629 if (sgs->sum_nr_running)
4630 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4632 sgs->group_weight = group->group_weight;
4634 sgs->group_imb = sg_imbalanced(group);
4635 sgs->group_capacity = sg_capacity(env, group);
4637 if (sgs->group_capacity > sgs->sum_nr_running)
4638 sgs->group_has_capacity = 1;
4642 * update_sd_pick_busiest - return 1 on busiest group
4643 * @env: The load balancing environment.
4644 * @sds: sched_domain statistics
4645 * @sg: sched_group candidate to be checked for being the busiest
4646 * @sgs: sched_group statistics
4648 * Determine if @sg is a busier group than the previously selected
4651 * Return: %true if @sg is a busier group than the previously selected
4652 * busiest group. %false otherwise.
4654 static bool update_sd_pick_busiest(struct lb_env *env,
4655 struct sd_lb_stats *sds,
4656 struct sched_group *sg,
4657 struct sg_lb_stats *sgs)
4659 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4662 if (sgs->sum_nr_running > sgs->group_capacity)
4669 * ASYM_PACKING needs to move all the work to the lowest
4670 * numbered CPUs in the group, therefore mark all groups
4671 * higher than ourself as busy.
4673 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4674 env->dst_cpu < group_first_cpu(sg)) {
4678 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4686 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4687 * @env: The load balancing environment.
4688 * @balance: Should we balance.
4689 * @sds: variable to hold the statistics for this sched_domain.
4691 static inline void update_sd_lb_stats(struct lb_env *env,
4692 struct sd_lb_stats *sds)
4694 struct sched_domain *child = env->sd->child;
4695 struct sched_group *sg = env->sd->groups;
4696 struct sg_lb_stats tmp_sgs;
4697 int load_idx, prefer_sibling = 0;
4699 if (child && child->flags & SD_PREFER_SIBLING)
4702 load_idx = get_sd_load_idx(env->sd, env->idle);
4705 struct sg_lb_stats *sgs = &tmp_sgs;
4708 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4711 sgs = &sds->local_stat;
4713 if (env->idle != CPU_NEWLY_IDLE ||
4714 time_after_eq(jiffies, sg->sgp->next_update))
4715 update_group_power(env->sd, env->dst_cpu);
4718 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4724 * In case the child domain prefers tasks go to siblings
4725 * first, lower the sg capacity to one so that we'll try
4726 * and move all the excess tasks away. We lower the capacity
4727 * of a group only if the local group has the capacity to fit
4728 * these excess tasks, i.e. nr_running < group_capacity. The
4729 * extra check prevents the case where you always pull from the
4730 * heaviest group when it is already under-utilized (possible
4731 * with a large weight task outweighs the tasks on the system).
4733 if (prefer_sibling && sds->local &&
4734 sds->local_stat.group_has_capacity)
4735 sgs->group_capacity = min(sgs->group_capacity, 1U);
4737 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4739 sds->busiest_stat = *sgs;
4743 /* Now, start updating sd_lb_stats */
4744 sds->total_load += sgs->group_load;
4745 sds->total_pwr += sgs->group_power;
4748 } while (sg != env->sd->groups);
4752 * check_asym_packing - Check to see if the group is packed into the
4755 * This is primarily intended to used at the sibling level. Some
4756 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4757 * case of POWER7, it can move to lower SMT modes only when higher
4758 * threads are idle. When in lower SMT modes, the threads will
4759 * perform better since they share less core resources. Hence when we
4760 * have idle threads, we want them to be the higher ones.
4762 * This packing function is run on idle threads. It checks to see if
4763 * the busiest CPU in this domain (core in the P7 case) has a higher
4764 * CPU number than the packing function is being run on. Here we are
4765 * assuming lower CPU number will be equivalent to lower a SMT thread
4768 * Return: 1 when packing is required and a task should be moved to
4769 * this CPU. The amount of the imbalance is returned in *imbalance.
4771 * @env: The load balancing environment.
4772 * @sds: Statistics of the sched_domain which is to be packed
4774 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4778 if (!(env->sd->flags & SD_ASYM_PACKING))
4784 busiest_cpu = group_first_cpu(sds->busiest);
4785 if (env->dst_cpu > busiest_cpu)
4788 env->imbalance = DIV_ROUND_CLOSEST(
4789 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
4796 * fix_small_imbalance - Calculate the minor imbalance that exists
4797 * amongst the groups of a sched_domain, during
4799 * @env: The load balancing environment.
4800 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4803 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4805 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4806 unsigned int imbn = 2;
4807 unsigned long scaled_busy_load_per_task;
4808 struct sg_lb_stats *local, *busiest;
4810 local = &sds->local_stat;
4811 busiest = &sds->busiest_stat;
4813 if (!local->sum_nr_running)
4814 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
4815 else if (busiest->load_per_task > local->load_per_task)
4818 scaled_busy_load_per_task =
4819 (busiest->load_per_task * SCHED_POWER_SCALE) /
4820 busiest->group_power;
4822 if (busiest->avg_load + scaled_busy_load_per_task >=
4823 local->avg_load + (scaled_busy_load_per_task * imbn)) {
4824 env->imbalance = busiest->load_per_task;
4829 * OK, we don't have enough imbalance to justify moving tasks,
4830 * however we may be able to increase total CPU power used by
4834 pwr_now += busiest->group_power *
4835 min(busiest->load_per_task, busiest->avg_load);
4836 pwr_now += local->group_power *
4837 min(local->load_per_task, local->avg_load);
4838 pwr_now /= SCHED_POWER_SCALE;
4840 /* Amount of load we'd subtract */
4841 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4842 busiest->group_power;
4843 if (busiest->avg_load > tmp) {
4844 pwr_move += busiest->group_power *
4845 min(busiest->load_per_task,
4846 busiest->avg_load - tmp);
4849 /* Amount of load we'd add */
4850 if (busiest->avg_load * busiest->group_power <
4851 busiest->load_per_task * SCHED_POWER_SCALE) {
4852 tmp = (busiest->avg_load * busiest->group_power) /
4855 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4858 pwr_move += local->group_power *
4859 min(local->load_per_task, local->avg_load + tmp);
4860 pwr_move /= SCHED_POWER_SCALE;
4862 /* Move if we gain throughput */
4863 if (pwr_move > pwr_now)
4864 env->imbalance = busiest->load_per_task;
4868 * calculate_imbalance - Calculate the amount of imbalance present within the
4869 * groups of a given sched_domain during load balance.
4870 * @env: load balance environment
4871 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4873 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4875 unsigned long max_pull, load_above_capacity = ~0UL;
4876 struct sg_lb_stats *local, *busiest;
4878 local = &sds->local_stat;
4879 busiest = &sds->busiest_stat;
4881 if (busiest->group_imb) {
4883 * In the group_imb case we cannot rely on group-wide averages
4884 * to ensure cpu-load equilibrium, look at wider averages. XXX
4886 busiest->load_per_task =
4887 min(busiest->load_per_task, sds->avg_load);
4891 * In the presence of smp nice balancing, certain scenarios can have
4892 * max load less than avg load(as we skip the groups at or below
4893 * its cpu_power, while calculating max_load..)
4895 if (busiest->avg_load <= sds->avg_load ||
4896 local->avg_load >= sds->avg_load) {
4898 return fix_small_imbalance(env, sds);
4901 if (!busiest->group_imb) {
4903 * Don't want to pull so many tasks that a group would go idle.
4904 * Except of course for the group_imb case, since then we might
4905 * have to drop below capacity to reach cpu-load equilibrium.
4907 load_above_capacity =
4908 (busiest->sum_nr_running - busiest->group_capacity);
4910 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4911 load_above_capacity /= busiest->group_power;
4915 * We're trying to get all the cpus to the average_load, so we don't
4916 * want to push ourselves above the average load, nor do we wish to
4917 * reduce the max loaded cpu below the average load. At the same time,
4918 * we also don't want to reduce the group load below the group capacity
4919 * (so that we can implement power-savings policies etc). Thus we look
4920 * for the minimum possible imbalance.
4922 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
4924 /* How much load to actually move to equalise the imbalance */
4925 env->imbalance = min(
4926 max_pull * busiest->group_power,
4927 (sds->avg_load - local->avg_load) * local->group_power
4928 ) / SCHED_POWER_SCALE;
4931 * if *imbalance is less than the average load per runnable task
4932 * there is no guarantee that any tasks will be moved so we'll have
4933 * a think about bumping its value to force at least one task to be
4936 if (env->imbalance < busiest->load_per_task)
4937 return fix_small_imbalance(env, sds);
4940 /******* find_busiest_group() helpers end here *********************/
4943 * find_busiest_group - Returns the busiest group within the sched_domain
4944 * if there is an imbalance. If there isn't an imbalance, and
4945 * the user has opted for power-savings, it returns a group whose
4946 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4947 * such a group exists.
4949 * Also calculates the amount of weighted load which should be moved
4950 * to restore balance.
4952 * @env: The load balancing environment.
4954 * Return: - The busiest group if imbalance exists.
4955 * - If no imbalance and user has opted for power-savings balance,
4956 * return the least loaded group whose CPUs can be
4957 * put to idle by rebalancing its tasks onto our group.
4959 static struct sched_group *find_busiest_group(struct lb_env *env)
4961 struct sg_lb_stats *local, *busiest;
4962 struct sd_lb_stats sds;
4964 init_sd_lb_stats(&sds);
4967 * Compute the various statistics relavent for load balancing at
4970 update_sd_lb_stats(env, &sds);
4971 local = &sds.local_stat;
4972 busiest = &sds.busiest_stat;
4974 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4975 check_asym_packing(env, &sds))
4978 /* There is no busy sibling group to pull tasks from */
4979 if (!sds.busiest || busiest->sum_nr_running == 0)
4982 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4985 * If the busiest group is imbalanced the below checks don't
4986 * work because they assume all things are equal, which typically
4987 * isn't true due to cpus_allowed constraints and the like.
4989 if (busiest->group_imb)
4992 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4993 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
4994 !busiest->group_has_capacity)
4998 * If the local group is more busy than the selected busiest group
4999 * don't try and pull any tasks.
5001 if (local->avg_load >= busiest->avg_load)
5005 * Don't pull any tasks if this group is already above the domain
5008 if (local->avg_load >= sds.avg_load)
5011 if (env->idle == CPU_IDLE) {
5013 * This cpu is idle. If the busiest group load doesn't
5014 * have more tasks than the number of available cpu's and
5015 * there is no imbalance between this and busiest group
5016 * wrt to idle cpu's, it is balanced.
5018 if ((local->idle_cpus < busiest->idle_cpus) &&
5019 busiest->sum_nr_running <= busiest->group_weight)
5023 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5024 * imbalance_pct to be conservative.
5026 if (100 * busiest->avg_load <=
5027 env->sd->imbalance_pct * local->avg_load)
5032 /* Looks like there is an imbalance. Compute it */
5033 calculate_imbalance(env, &sds);
5042 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5044 static struct rq *find_busiest_queue(struct lb_env *env,
5045 struct sched_group *group)
5047 struct rq *busiest = NULL, *rq;
5048 unsigned long busiest_load = 0, busiest_power = 1;
5051 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5052 unsigned long power = power_of(i);
5053 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5058 capacity = fix_small_capacity(env->sd, group);
5061 wl = weighted_cpuload(i);
5064 * When comparing with imbalance, use weighted_cpuload()
5065 * which is not scaled with the cpu power.
5067 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5071 * For the load comparisons with the other cpu's, consider
5072 * the weighted_cpuload() scaled with the cpu power, so that
5073 * the load can be moved away from the cpu that is potentially
5074 * running at a lower capacity.
5076 * Thus we're looking for max(wl_i / power_i), crosswise
5077 * multiplication to rid ourselves of the division works out
5078 * to: wl_i * power_j > wl_j * power_i; where j is our
5081 if (wl * busiest_power > busiest_load * power) {
5083 busiest_power = power;
5092 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5093 * so long as it is large enough.
5095 #define MAX_PINNED_INTERVAL 512
5097 /* Working cpumask for load_balance and load_balance_newidle. */
5098 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5100 static int need_active_balance(struct lb_env *env)
5102 struct sched_domain *sd = env->sd;
5104 if (env->idle == CPU_NEWLY_IDLE) {
5107 * ASYM_PACKING needs to force migrate tasks from busy but
5108 * higher numbered CPUs in order to pack all tasks in the
5109 * lowest numbered CPUs.
5111 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5115 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5118 static int active_load_balance_cpu_stop(void *data);
5120 static int should_we_balance(struct lb_env *env)
5122 struct sched_group *sg = env->sd->groups;
5123 struct cpumask *sg_cpus, *sg_mask;
5124 int cpu, balance_cpu = -1;
5127 * In the newly idle case, we will allow all the cpu's
5128 * to do the newly idle load balance.
5130 if (env->idle == CPU_NEWLY_IDLE)
5133 sg_cpus = sched_group_cpus(sg);
5134 sg_mask = sched_group_mask(sg);
5135 /* Try to find first idle cpu */
5136 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5137 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5144 if (balance_cpu == -1)
5145 balance_cpu = group_balance_cpu(sg);
5148 * First idle cpu or the first cpu(busiest) in this sched group
5149 * is eligible for doing load balancing at this and above domains.
5151 return balance_cpu == env->dst_cpu;
5155 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5156 * tasks if there is an imbalance.
5158 static int load_balance(int this_cpu, struct rq *this_rq,
5159 struct sched_domain *sd, enum cpu_idle_type idle,
5160 int *continue_balancing)
5162 int ld_moved, cur_ld_moved, active_balance = 0;
5163 struct sched_domain *sd_parent = sd->parent;
5164 struct sched_group *group;
5166 unsigned long flags;
5167 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5169 struct lb_env env = {
5171 .dst_cpu = this_cpu,
5173 .dst_grpmask = sched_group_cpus(sd->groups),
5175 .loop_break = sched_nr_migrate_break,
5180 * For NEWLY_IDLE load_balancing, we don't need to consider
5181 * other cpus in our group
5183 if (idle == CPU_NEWLY_IDLE)
5184 env.dst_grpmask = NULL;
5186 cpumask_copy(cpus, cpu_active_mask);
5188 schedstat_inc(sd, lb_count[idle]);
5191 if (!should_we_balance(&env)) {
5192 *continue_balancing = 0;
5196 group = find_busiest_group(&env);
5198 schedstat_inc(sd, lb_nobusyg[idle]);
5202 busiest = find_busiest_queue(&env, group);
5204 schedstat_inc(sd, lb_nobusyq[idle]);
5208 BUG_ON(busiest == env.dst_rq);
5210 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5213 if (busiest->nr_running > 1) {
5215 * Attempt to move tasks. If find_busiest_group has found
5216 * an imbalance but busiest->nr_running <= 1, the group is
5217 * still unbalanced. ld_moved simply stays zero, so it is
5218 * correctly treated as an imbalance.
5220 env.flags |= LBF_ALL_PINNED;
5221 env.src_cpu = busiest->cpu;
5222 env.src_rq = busiest;
5223 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5226 local_irq_save(flags);
5227 double_rq_lock(env.dst_rq, busiest);
5230 * cur_ld_moved - load moved in current iteration
5231 * ld_moved - cumulative load moved across iterations
5233 cur_ld_moved = move_tasks(&env);
5234 ld_moved += cur_ld_moved;
5235 double_rq_unlock(env.dst_rq, busiest);
5236 local_irq_restore(flags);
5239 * some other cpu did the load balance for us.
5241 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5242 resched_cpu(env.dst_cpu);
5244 if (env.flags & LBF_NEED_BREAK) {
5245 env.flags &= ~LBF_NEED_BREAK;
5250 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5251 * us and move them to an alternate dst_cpu in our sched_group
5252 * where they can run. The upper limit on how many times we
5253 * iterate on same src_cpu is dependent on number of cpus in our
5256 * This changes load balance semantics a bit on who can move
5257 * load to a given_cpu. In addition to the given_cpu itself
5258 * (or a ilb_cpu acting on its behalf where given_cpu is
5259 * nohz-idle), we now have balance_cpu in a position to move
5260 * load to given_cpu. In rare situations, this may cause
5261 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5262 * _independently_ and at _same_ time to move some load to
5263 * given_cpu) causing exceess load to be moved to given_cpu.
5264 * This however should not happen so much in practice and
5265 * moreover subsequent load balance cycles should correct the
5266 * excess load moved.
5268 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5270 /* Prevent to re-select dst_cpu via env's cpus */
5271 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5273 env.dst_rq = cpu_rq(env.new_dst_cpu);
5274 env.dst_cpu = env.new_dst_cpu;
5275 env.flags &= ~LBF_DST_PINNED;
5277 env.loop_break = sched_nr_migrate_break;
5280 * Go back to "more_balance" rather than "redo" since we
5281 * need to continue with same src_cpu.
5287 * We failed to reach balance because of affinity.
5290 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5292 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5293 *group_imbalance = 1;
5294 } else if (*group_imbalance)
5295 *group_imbalance = 0;
5298 /* All tasks on this runqueue were pinned by CPU affinity */
5299 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5300 cpumask_clear_cpu(cpu_of(busiest), cpus);
5301 if (!cpumask_empty(cpus)) {
5303 env.loop_break = sched_nr_migrate_break;
5311 schedstat_inc(sd, lb_failed[idle]);
5313 * Increment the failure counter only on periodic balance.
5314 * We do not want newidle balance, which can be very
5315 * frequent, pollute the failure counter causing
5316 * excessive cache_hot migrations and active balances.
5318 if (idle != CPU_NEWLY_IDLE)
5319 sd->nr_balance_failed++;
5321 if (need_active_balance(&env)) {
5322 raw_spin_lock_irqsave(&busiest->lock, flags);
5324 /* don't kick the active_load_balance_cpu_stop,
5325 * if the curr task on busiest cpu can't be
5328 if (!cpumask_test_cpu(this_cpu,
5329 tsk_cpus_allowed(busiest->curr))) {
5330 raw_spin_unlock_irqrestore(&busiest->lock,
5332 env.flags |= LBF_ALL_PINNED;
5333 goto out_one_pinned;
5337 * ->active_balance synchronizes accesses to
5338 * ->active_balance_work. Once set, it's cleared
5339 * only after active load balance is finished.
5341 if (!busiest->active_balance) {
5342 busiest->active_balance = 1;
5343 busiest->push_cpu = this_cpu;
5346 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5348 if (active_balance) {
5349 stop_one_cpu_nowait(cpu_of(busiest),
5350 active_load_balance_cpu_stop, busiest,
5351 &busiest->active_balance_work);
5355 * We've kicked active balancing, reset the failure
5358 sd->nr_balance_failed = sd->cache_nice_tries+1;
5361 sd->nr_balance_failed = 0;
5363 if (likely(!active_balance)) {
5364 /* We were unbalanced, so reset the balancing interval */
5365 sd->balance_interval = sd->min_interval;
5368 * If we've begun active balancing, start to back off. This
5369 * case may not be covered by the all_pinned logic if there
5370 * is only 1 task on the busy runqueue (because we don't call
5373 if (sd->balance_interval < sd->max_interval)
5374 sd->balance_interval *= 2;
5380 schedstat_inc(sd, lb_balanced[idle]);
5382 sd->nr_balance_failed = 0;
5385 /* tune up the balancing interval */
5386 if (((env.flags & LBF_ALL_PINNED) &&
5387 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5388 (sd->balance_interval < sd->max_interval))
5389 sd->balance_interval *= 2;
5397 * idle_balance is called by schedule() if this_cpu is about to become
5398 * idle. Attempts to pull tasks from other CPUs.
5400 void idle_balance(int this_cpu, struct rq *this_rq)
5402 struct sched_domain *sd;
5403 int pulled_task = 0;
5404 unsigned long next_balance = jiffies + HZ;
5407 this_rq->idle_stamp = rq_clock(this_rq);
5409 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5413 * Drop the rq->lock, but keep IRQ/preempt disabled.
5415 raw_spin_unlock(&this_rq->lock);
5417 update_blocked_averages(this_cpu);
5419 for_each_domain(this_cpu, sd) {
5420 unsigned long interval;
5421 int continue_balancing = 1;
5422 u64 t0, domain_cost;
5424 if (!(sd->flags & SD_LOAD_BALANCE))
5427 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5430 if (sd->flags & SD_BALANCE_NEWIDLE) {
5431 t0 = sched_clock_cpu(this_cpu);
5433 /* If we've pulled tasks over stop searching: */
5434 pulled_task = load_balance(this_cpu, this_rq,
5436 &continue_balancing);
5438 domain_cost = sched_clock_cpu(this_cpu) - t0;
5439 if (domain_cost > sd->max_newidle_lb_cost)
5440 sd->max_newidle_lb_cost = domain_cost;
5442 curr_cost += domain_cost;
5445 interval = msecs_to_jiffies(sd->balance_interval);
5446 if (time_after(next_balance, sd->last_balance + interval))
5447 next_balance = sd->last_balance + interval;
5449 this_rq->idle_stamp = 0;
5455 raw_spin_lock(&this_rq->lock);
5457 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5459 * We are going idle. next_balance may be set based on
5460 * a busy processor. So reset next_balance.
5462 this_rq->next_balance = next_balance;
5465 if (curr_cost > this_rq->max_idle_balance_cost)
5466 this_rq->max_idle_balance_cost = curr_cost;
5470 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5471 * running tasks off the busiest CPU onto idle CPUs. It requires at
5472 * least 1 task to be running on each physical CPU where possible, and
5473 * avoids physical / logical imbalances.
5475 static int active_load_balance_cpu_stop(void *data)
5477 struct rq *busiest_rq = data;
5478 int busiest_cpu = cpu_of(busiest_rq);
5479 int target_cpu = busiest_rq->push_cpu;
5480 struct rq *target_rq = cpu_rq(target_cpu);
5481 struct sched_domain *sd;
5483 raw_spin_lock_irq(&busiest_rq->lock);
5485 /* make sure the requested cpu hasn't gone down in the meantime */
5486 if (unlikely(busiest_cpu != smp_processor_id() ||
5487 !busiest_rq->active_balance))
5490 /* Is there any task to move? */
5491 if (busiest_rq->nr_running <= 1)
5495 * This condition is "impossible", if it occurs
5496 * we need to fix it. Originally reported by
5497 * Bjorn Helgaas on a 128-cpu setup.
5499 BUG_ON(busiest_rq == target_rq);
5501 /* move a task from busiest_rq to target_rq */
5502 double_lock_balance(busiest_rq, target_rq);
5504 /* Search for an sd spanning us and the target CPU. */
5506 for_each_domain(target_cpu, sd) {
5507 if ((sd->flags & SD_LOAD_BALANCE) &&
5508 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5513 struct lb_env env = {
5515 .dst_cpu = target_cpu,
5516 .dst_rq = target_rq,
5517 .src_cpu = busiest_rq->cpu,
5518 .src_rq = busiest_rq,
5522 schedstat_inc(sd, alb_count);
5524 if (move_one_task(&env))
5525 schedstat_inc(sd, alb_pushed);
5527 schedstat_inc(sd, alb_failed);
5530 double_unlock_balance(busiest_rq, target_rq);
5532 busiest_rq->active_balance = 0;
5533 raw_spin_unlock_irq(&busiest_rq->lock);
5537 #ifdef CONFIG_NO_HZ_COMMON
5539 * idle load balancing details
5540 * - When one of the busy CPUs notice that there may be an idle rebalancing
5541 * needed, they will kick the idle load balancer, which then does idle
5542 * load balancing for all the idle CPUs.
5545 cpumask_var_t idle_cpus_mask;
5547 unsigned long next_balance; /* in jiffy units */
5548 } nohz ____cacheline_aligned;
5550 static inline int find_new_ilb(int call_cpu)
5552 int ilb = cpumask_first(nohz.idle_cpus_mask);
5554 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5561 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5562 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5563 * CPU (if there is one).
5565 static void nohz_balancer_kick(int cpu)
5569 nohz.next_balance++;
5571 ilb_cpu = find_new_ilb(cpu);
5573 if (ilb_cpu >= nr_cpu_ids)
5576 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5579 * Use smp_send_reschedule() instead of resched_cpu().
5580 * This way we generate a sched IPI on the target cpu which
5581 * is idle. And the softirq performing nohz idle load balance
5582 * will be run before returning from the IPI.
5584 smp_send_reschedule(ilb_cpu);
5588 static inline void nohz_balance_exit_idle(int cpu)
5590 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5591 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5592 atomic_dec(&nohz.nr_cpus);
5593 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5597 static inline void set_cpu_sd_state_busy(void)
5599 struct sched_domain *sd;
5602 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5604 if (!sd || !sd->nohz_idle)
5608 for (; sd; sd = sd->parent)
5609 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5614 void set_cpu_sd_state_idle(void)
5616 struct sched_domain *sd;
5619 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5621 if (!sd || sd->nohz_idle)
5625 for (; sd; sd = sd->parent)
5626 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5632 * This routine will record that the cpu is going idle with tick stopped.
5633 * This info will be used in performing idle load balancing in the future.
5635 void nohz_balance_enter_idle(int cpu)
5638 * If this cpu is going down, then nothing needs to be done.
5640 if (!cpu_active(cpu))
5643 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5646 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5647 atomic_inc(&nohz.nr_cpus);
5648 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5651 static int sched_ilb_notifier(struct notifier_block *nfb,
5652 unsigned long action, void *hcpu)
5654 switch (action & ~CPU_TASKS_FROZEN) {
5656 nohz_balance_exit_idle(smp_processor_id());
5664 static DEFINE_SPINLOCK(balancing);
5667 * Scale the max load_balance interval with the number of CPUs in the system.
5668 * This trades load-balance latency on larger machines for less cross talk.
5670 void update_max_interval(void)
5672 max_load_balance_interval = HZ*num_online_cpus()/10;
5676 * It checks each scheduling domain to see if it is due to be balanced,
5677 * and initiates a balancing operation if so.
5679 * Balancing parameters are set up in init_sched_domains.
5681 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5683 int continue_balancing = 1;
5684 struct rq *rq = cpu_rq(cpu);
5685 unsigned long interval;
5686 struct sched_domain *sd;
5687 /* Earliest time when we have to do rebalance again */
5688 unsigned long next_balance = jiffies + 60*HZ;
5689 int update_next_balance = 0;
5690 int need_serialize, need_decay = 0;
5693 update_blocked_averages(cpu);
5696 for_each_domain(cpu, sd) {
5698 * Decay the newidle max times here because this is a regular
5699 * visit to all the domains. Decay ~1% per second.
5701 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5702 sd->max_newidle_lb_cost =
5703 (sd->max_newidle_lb_cost * 253) / 256;
5704 sd->next_decay_max_lb_cost = jiffies + HZ;
5707 max_cost += sd->max_newidle_lb_cost;
5709 if (!(sd->flags & SD_LOAD_BALANCE))
5713 * Stop the load balance at this level. There is another
5714 * CPU in our sched group which is doing load balancing more
5717 if (!continue_balancing) {
5723 interval = sd->balance_interval;
5724 if (idle != CPU_IDLE)
5725 interval *= sd->busy_factor;
5727 /* scale ms to jiffies */
5728 interval = msecs_to_jiffies(interval);
5729 interval = clamp(interval, 1UL, max_load_balance_interval);
5731 need_serialize = sd->flags & SD_SERIALIZE;
5733 if (need_serialize) {
5734 if (!spin_trylock(&balancing))
5738 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5739 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5741 * The LBF_DST_PINNED logic could have changed
5742 * env->dst_cpu, so we can't know our idle
5743 * state even if we migrated tasks. Update it.
5745 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5747 sd->last_balance = jiffies;
5750 spin_unlock(&balancing);
5752 if (time_after(next_balance, sd->last_balance + interval)) {
5753 next_balance = sd->last_balance + interval;
5754 update_next_balance = 1;
5759 * Ensure the rq-wide value also decays but keep it at a
5760 * reasonable floor to avoid funnies with rq->avg_idle.
5762 rq->max_idle_balance_cost =
5763 max((u64)sysctl_sched_migration_cost, max_cost);
5768 * next_balance will be updated only when there is a need.
5769 * When the cpu is attached to null domain for ex, it will not be
5772 if (likely(update_next_balance))
5773 rq->next_balance = next_balance;
5776 #ifdef CONFIG_NO_HZ_COMMON
5778 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5779 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5781 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5783 struct rq *this_rq = cpu_rq(this_cpu);
5787 if (idle != CPU_IDLE ||
5788 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5791 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5792 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5796 * If this cpu gets work to do, stop the load balancing
5797 * work being done for other cpus. Next load
5798 * balancing owner will pick it up.
5803 rq = cpu_rq(balance_cpu);
5805 raw_spin_lock_irq(&rq->lock);
5806 update_rq_clock(rq);
5807 update_idle_cpu_load(rq);
5808 raw_spin_unlock_irq(&rq->lock);
5810 rebalance_domains(balance_cpu, CPU_IDLE);
5812 if (time_after(this_rq->next_balance, rq->next_balance))
5813 this_rq->next_balance = rq->next_balance;
5815 nohz.next_balance = this_rq->next_balance;
5817 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5821 * Current heuristic for kicking the idle load balancer in the presence
5822 * of an idle cpu is the system.
5823 * - This rq has more than one task.
5824 * - At any scheduler domain level, this cpu's scheduler group has multiple
5825 * busy cpu's exceeding the group's power.
5826 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5827 * domain span are idle.
5829 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5831 unsigned long now = jiffies;
5832 struct sched_domain *sd;
5834 if (unlikely(idle_cpu(cpu)))
5838 * We may be recently in ticked or tickless idle mode. At the first
5839 * busy tick after returning from idle, we will update the busy stats.
5841 set_cpu_sd_state_busy();
5842 nohz_balance_exit_idle(cpu);
5845 * None are in tickless mode and hence no need for NOHZ idle load
5848 if (likely(!atomic_read(&nohz.nr_cpus)))
5851 if (time_before(now, nohz.next_balance))
5854 if (rq->nr_running >= 2)
5858 for_each_domain(cpu, sd) {
5859 struct sched_group *sg = sd->groups;
5860 struct sched_group_power *sgp = sg->sgp;
5861 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5863 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5864 goto need_kick_unlock;
5866 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5867 && (cpumask_first_and(nohz.idle_cpus_mask,
5868 sched_domain_span(sd)) < cpu))
5869 goto need_kick_unlock;
5871 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5883 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5887 * run_rebalance_domains is triggered when needed from the scheduler tick.
5888 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5890 static void run_rebalance_domains(struct softirq_action *h)
5892 int this_cpu = smp_processor_id();
5893 struct rq *this_rq = cpu_rq(this_cpu);
5894 enum cpu_idle_type idle = this_rq->idle_balance ?
5895 CPU_IDLE : CPU_NOT_IDLE;
5897 rebalance_domains(this_cpu, idle);
5900 * If this cpu has a pending nohz_balance_kick, then do the
5901 * balancing on behalf of the other idle cpus whose ticks are
5904 nohz_idle_balance(this_cpu, idle);
5907 static inline int on_null_domain(int cpu)
5909 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5913 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5915 void trigger_load_balance(struct rq *rq, int cpu)
5917 /* Don't need to rebalance while attached to NULL domain */
5918 if (time_after_eq(jiffies, rq->next_balance) &&
5919 likely(!on_null_domain(cpu)))
5920 raise_softirq(SCHED_SOFTIRQ);
5921 #ifdef CONFIG_NO_HZ_COMMON
5922 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5923 nohz_balancer_kick(cpu);
5927 static void rq_online_fair(struct rq *rq)
5932 static void rq_offline_fair(struct rq *rq)
5936 /* Ensure any throttled groups are reachable by pick_next_task */
5937 unthrottle_offline_cfs_rqs(rq);
5940 #endif /* CONFIG_SMP */
5943 * scheduler tick hitting a task of our scheduling class:
5945 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5947 struct cfs_rq *cfs_rq;
5948 struct sched_entity *se = &curr->se;
5950 for_each_sched_entity(se) {
5951 cfs_rq = cfs_rq_of(se);
5952 entity_tick(cfs_rq, se, queued);
5955 if (numabalancing_enabled)
5956 task_tick_numa(rq, curr);
5958 update_rq_runnable_avg(rq, 1);
5962 * called on fork with the child task as argument from the parent's context
5963 * - child not yet on the tasklist
5964 * - preemption disabled
5966 static void task_fork_fair(struct task_struct *p)
5968 struct cfs_rq *cfs_rq;
5969 struct sched_entity *se = &p->se, *curr;
5970 int this_cpu = smp_processor_id();
5971 struct rq *rq = this_rq();
5972 unsigned long flags;
5974 raw_spin_lock_irqsave(&rq->lock, flags);
5976 update_rq_clock(rq);
5978 cfs_rq = task_cfs_rq(current);
5979 curr = cfs_rq->curr;
5982 * Not only the cpu but also the task_group of the parent might have
5983 * been changed after parent->se.parent,cfs_rq were copied to
5984 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
5985 * of child point to valid ones.
5988 __set_task_cpu(p, this_cpu);
5991 update_curr(cfs_rq);
5994 se->vruntime = curr->vruntime;
5995 place_entity(cfs_rq, se, 1);
5997 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5999 * Upon rescheduling, sched_class::put_prev_task() will place
6000 * 'current' within the tree based on its new key value.
6002 swap(curr->vruntime, se->vruntime);
6003 resched_task(rq->curr);
6006 se->vruntime -= cfs_rq->min_vruntime;
6008 raw_spin_unlock_irqrestore(&rq->lock, flags);
6012 * Priority of the task has changed. Check to see if we preempt
6016 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6022 * Reschedule if we are currently running on this runqueue and
6023 * our priority decreased, or if we are not currently running on
6024 * this runqueue and our priority is higher than the current's
6026 if (rq->curr == p) {
6027 if (p->prio > oldprio)
6028 resched_task(rq->curr);
6030 check_preempt_curr(rq, p, 0);
6033 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6035 struct sched_entity *se = &p->se;
6036 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6039 * Ensure the task's vruntime is normalized, so that when its
6040 * switched back to the fair class the enqueue_entity(.flags=0) will
6041 * do the right thing.
6043 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6044 * have normalized the vruntime, if it was !on_rq, then only when
6045 * the task is sleeping will it still have non-normalized vruntime.
6047 if (!se->on_rq && p->state != TASK_RUNNING) {
6049 * Fix up our vruntime so that the current sleep doesn't
6050 * cause 'unlimited' sleep bonus.
6052 place_entity(cfs_rq, se, 0);
6053 se->vruntime -= cfs_rq->min_vruntime;
6058 * Remove our load from contribution when we leave sched_fair
6059 * and ensure we don't carry in an old decay_count if we
6062 if (se->avg.decay_count) {
6063 __synchronize_entity_decay(se);
6064 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6070 * We switched to the sched_fair class.
6072 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6078 * We were most likely switched from sched_rt, so
6079 * kick off the schedule if running, otherwise just see
6080 * if we can still preempt the current task.
6083 resched_task(rq->curr);
6085 check_preempt_curr(rq, p, 0);
6088 /* Account for a task changing its policy or group.
6090 * This routine is mostly called to set cfs_rq->curr field when a task
6091 * migrates between groups/classes.
6093 static void set_curr_task_fair(struct rq *rq)
6095 struct sched_entity *se = &rq->curr->se;
6097 for_each_sched_entity(se) {
6098 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6100 set_next_entity(cfs_rq, se);
6101 /* ensure bandwidth has been allocated on our new cfs_rq */
6102 account_cfs_rq_runtime(cfs_rq, 0);
6106 void init_cfs_rq(struct cfs_rq *cfs_rq)
6108 cfs_rq->tasks_timeline = RB_ROOT;
6109 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6110 #ifndef CONFIG_64BIT
6111 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6114 atomic64_set(&cfs_rq->decay_counter, 1);
6115 atomic_long_set(&cfs_rq->removed_load, 0);
6119 #ifdef CONFIG_FAIR_GROUP_SCHED
6120 static void task_move_group_fair(struct task_struct *p, int on_rq)
6122 struct cfs_rq *cfs_rq;
6124 * If the task was not on the rq at the time of this cgroup movement
6125 * it must have been asleep, sleeping tasks keep their ->vruntime
6126 * absolute on their old rq until wakeup (needed for the fair sleeper
6127 * bonus in place_entity()).
6129 * If it was on the rq, we've just 'preempted' it, which does convert
6130 * ->vruntime to a relative base.
6132 * Make sure both cases convert their relative position when migrating
6133 * to another cgroup's rq. This does somewhat interfere with the
6134 * fair sleeper stuff for the first placement, but who cares.
6137 * When !on_rq, vruntime of the task has usually NOT been normalized.
6138 * But there are some cases where it has already been normalized:
6140 * - Moving a forked child which is waiting for being woken up by
6141 * wake_up_new_task().
6142 * - Moving a task which has been woken up by try_to_wake_up() and
6143 * waiting for actually being woken up by sched_ttwu_pending().
6145 * To prevent boost or penalty in the new cfs_rq caused by delta
6146 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6148 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6152 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6153 set_task_rq(p, task_cpu(p));
6155 cfs_rq = cfs_rq_of(&p->se);
6156 p->se.vruntime += cfs_rq->min_vruntime;
6159 * migrate_task_rq_fair() will have removed our previous
6160 * contribution, but we must synchronize for ongoing future
6163 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6164 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6169 void free_fair_sched_group(struct task_group *tg)
6173 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6175 for_each_possible_cpu(i) {
6177 kfree(tg->cfs_rq[i]);
6186 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6188 struct cfs_rq *cfs_rq;
6189 struct sched_entity *se;
6192 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6195 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6199 tg->shares = NICE_0_LOAD;
6201 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6203 for_each_possible_cpu(i) {
6204 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6205 GFP_KERNEL, cpu_to_node(i));
6209 se = kzalloc_node(sizeof(struct sched_entity),
6210 GFP_KERNEL, cpu_to_node(i));
6214 init_cfs_rq(cfs_rq);
6215 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6226 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6228 struct rq *rq = cpu_rq(cpu);
6229 unsigned long flags;
6232 * Only empty task groups can be destroyed; so we can speculatively
6233 * check on_list without danger of it being re-added.
6235 if (!tg->cfs_rq[cpu]->on_list)
6238 raw_spin_lock_irqsave(&rq->lock, flags);
6239 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6240 raw_spin_unlock_irqrestore(&rq->lock, flags);
6243 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6244 struct sched_entity *se, int cpu,
6245 struct sched_entity *parent)
6247 struct rq *rq = cpu_rq(cpu);
6251 init_cfs_rq_runtime(cfs_rq);
6253 tg->cfs_rq[cpu] = cfs_rq;
6256 /* se could be NULL for root_task_group */
6261 se->cfs_rq = &rq->cfs;
6263 se->cfs_rq = parent->my_q;
6266 update_load_set(&se->load, 0);
6267 se->parent = parent;
6270 static DEFINE_MUTEX(shares_mutex);
6272 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6275 unsigned long flags;
6278 * We can't change the weight of the root cgroup.
6283 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6285 mutex_lock(&shares_mutex);
6286 if (tg->shares == shares)
6289 tg->shares = shares;
6290 for_each_possible_cpu(i) {
6291 struct rq *rq = cpu_rq(i);
6292 struct sched_entity *se;
6295 /* Propagate contribution to hierarchy */
6296 raw_spin_lock_irqsave(&rq->lock, flags);
6298 /* Possible calls to update_curr() need rq clock */
6299 update_rq_clock(rq);
6300 for_each_sched_entity(se)
6301 update_cfs_shares(group_cfs_rq(se));
6302 raw_spin_unlock_irqrestore(&rq->lock, flags);
6306 mutex_unlock(&shares_mutex);
6309 #else /* CONFIG_FAIR_GROUP_SCHED */
6311 void free_fair_sched_group(struct task_group *tg) { }
6313 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6318 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6320 #endif /* CONFIG_FAIR_GROUP_SCHED */
6323 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6325 struct sched_entity *se = &task->se;
6326 unsigned int rr_interval = 0;
6329 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6332 if (rq->cfs.load.weight)
6333 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6339 * All the scheduling class methods:
6341 const struct sched_class fair_sched_class = {
6342 .next = &idle_sched_class,
6343 .enqueue_task = enqueue_task_fair,
6344 .dequeue_task = dequeue_task_fair,
6345 .yield_task = yield_task_fair,
6346 .yield_to_task = yield_to_task_fair,
6348 .check_preempt_curr = check_preempt_wakeup,
6350 .pick_next_task = pick_next_task_fair,
6351 .put_prev_task = put_prev_task_fair,
6354 .select_task_rq = select_task_rq_fair,
6355 .migrate_task_rq = migrate_task_rq_fair,
6357 .rq_online = rq_online_fair,
6358 .rq_offline = rq_offline_fair,
6360 .task_waking = task_waking_fair,
6363 .set_curr_task = set_curr_task_fair,
6364 .task_tick = task_tick_fair,
6365 .task_fork = task_fork_fair,
6367 .prio_changed = prio_changed_fair,
6368 .switched_from = switched_from_fair,
6369 .switched_to = switched_to_fair,
6371 .get_rr_interval = get_rr_interval_fair,
6373 #ifdef CONFIG_FAIR_GROUP_SCHED
6374 .task_move_group = task_move_group_fair,
6378 #ifdef CONFIG_SCHED_DEBUG
6379 void print_cfs_stats(struct seq_file *m, int cpu)
6381 struct cfs_rq *cfs_rq;
6384 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6385 print_cfs_rq(m, cpu, cfs_rq);
6390 __init void init_sched_fair_class(void)
6393 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6395 #ifdef CONFIG_NO_HZ_COMMON
6396 nohz.next_balance = jiffies;
6397 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6398 cpu_notifier(sched_ilb_notifier, 0);