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 * Do not set pte_numa if the current running node is rate-limited.
950 * This loses statistics on the fault but if we are unwilling to
951 * migrate to this node, it is less likely we can do useful work
953 if (migrate_ratelimited(numa_node_id()))
956 start = mm->numa_scan_offset;
957 pages = sysctl_numa_balancing_scan_size;
958 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
962 down_read(&mm->mmap_sem);
963 vma = find_vma(mm, start);
965 reset_ptenuma_scan(p);
969 for (; vma; vma = vma->vm_next) {
970 if (!vma_migratable(vma))
973 /* Skip small VMAs. They are not likely to be of relevance */
974 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
978 start = max(start, vma->vm_start);
979 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
980 end = min(end, vma->vm_end);
981 pages -= change_prot_numa(vma, start, end);
986 } while (end != vma->vm_end);
991 * It is possible to reach the end of the VMA list but the last few VMAs are
992 * not guaranteed to the vma_migratable. If they are not, we would find the
993 * !migratable VMA on the next scan but not reset the scanner to the start
997 mm->numa_scan_offset = start;
999 reset_ptenuma_scan(p);
1000 up_read(&mm->mmap_sem);
1004 * Drive the periodic memory faults..
1006 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1008 struct callback_head *work = &curr->numa_work;
1012 * We don't care about NUMA placement if we don't have memory.
1014 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1018 * Using runtime rather than walltime has the dual advantage that
1019 * we (mostly) drive the selection from busy threads and that the
1020 * task needs to have done some actual work before we bother with
1023 now = curr->se.sum_exec_runtime;
1024 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1026 if (now - curr->node_stamp > period) {
1027 if (!curr->node_stamp)
1028 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
1029 curr->node_stamp = now;
1031 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1032 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1033 task_work_add(curr, work, true);
1038 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1041 #endif /* CONFIG_NUMA_BALANCING */
1044 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1046 update_load_add(&cfs_rq->load, se->load.weight);
1047 if (!parent_entity(se))
1048 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1050 if (entity_is_task(se))
1051 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1053 cfs_rq->nr_running++;
1057 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1059 update_load_sub(&cfs_rq->load, se->load.weight);
1060 if (!parent_entity(se))
1061 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1062 if (entity_is_task(se))
1063 list_del_init(&se->group_node);
1064 cfs_rq->nr_running--;
1067 #ifdef CONFIG_FAIR_GROUP_SCHED
1069 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1074 * Use this CPU's actual weight instead of the last load_contribution
1075 * to gain a more accurate current total weight. See
1076 * update_cfs_rq_load_contribution().
1078 tg_weight = atomic_long_read(&tg->load_avg);
1079 tg_weight -= cfs_rq->tg_load_contrib;
1080 tg_weight += cfs_rq->load.weight;
1085 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1087 long tg_weight, load, shares;
1089 tg_weight = calc_tg_weight(tg, cfs_rq);
1090 load = cfs_rq->load.weight;
1092 shares = (tg->shares * load);
1094 shares /= tg_weight;
1096 if (shares < MIN_SHARES)
1097 shares = MIN_SHARES;
1098 if (shares > tg->shares)
1099 shares = tg->shares;
1103 # else /* CONFIG_SMP */
1104 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1108 # endif /* CONFIG_SMP */
1109 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1110 unsigned long weight)
1113 /* commit outstanding execution time */
1114 if (cfs_rq->curr == se)
1115 update_curr(cfs_rq);
1116 account_entity_dequeue(cfs_rq, se);
1119 update_load_set(&se->load, weight);
1122 account_entity_enqueue(cfs_rq, se);
1125 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1127 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1129 struct task_group *tg;
1130 struct sched_entity *se;
1134 se = tg->se[cpu_of(rq_of(cfs_rq))];
1135 if (!se || throttled_hierarchy(cfs_rq))
1138 if (likely(se->load.weight == tg->shares))
1141 shares = calc_cfs_shares(cfs_rq, tg);
1143 reweight_entity(cfs_rq_of(se), se, shares);
1145 #else /* CONFIG_FAIR_GROUP_SCHED */
1146 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1149 #endif /* CONFIG_FAIR_GROUP_SCHED */
1153 * We choose a half-life close to 1 scheduling period.
1154 * Note: The tables below are dependent on this value.
1156 #define LOAD_AVG_PERIOD 32
1157 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1158 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1160 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1161 static const u32 runnable_avg_yN_inv[] = {
1162 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1163 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1164 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1165 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1166 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1167 0x85aac367, 0x82cd8698,
1171 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1172 * over-estimates when re-combining.
1174 static const u32 runnable_avg_yN_sum[] = {
1175 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1176 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1177 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1182 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1184 static __always_inline u64 decay_load(u64 val, u64 n)
1186 unsigned int local_n;
1190 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1193 /* after bounds checking we can collapse to 32-bit */
1197 * As y^PERIOD = 1/2, we can combine
1198 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1199 * With a look-up table which covers k^n (n<PERIOD)
1201 * To achieve constant time decay_load.
1203 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1204 val >>= local_n / LOAD_AVG_PERIOD;
1205 local_n %= LOAD_AVG_PERIOD;
1208 val *= runnable_avg_yN_inv[local_n];
1209 /* We don't use SRR here since we always want to round down. */
1214 * For updates fully spanning n periods, the contribution to runnable
1215 * average will be: \Sum 1024*y^n
1217 * We can compute this reasonably efficiently by combining:
1218 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1220 static u32 __compute_runnable_contrib(u64 n)
1224 if (likely(n <= LOAD_AVG_PERIOD))
1225 return runnable_avg_yN_sum[n];
1226 else if (unlikely(n >= LOAD_AVG_MAX_N))
1227 return LOAD_AVG_MAX;
1229 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1231 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1232 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1234 n -= LOAD_AVG_PERIOD;
1235 } while (n > LOAD_AVG_PERIOD);
1237 contrib = decay_load(contrib, n);
1238 return contrib + runnable_avg_yN_sum[n];
1242 * We can represent the historical contribution to runnable average as the
1243 * coefficients of a geometric series. To do this we sub-divide our runnable
1244 * history into segments of approximately 1ms (1024us); label the segment that
1245 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1247 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1249 * (now) (~1ms ago) (~2ms ago)
1251 * Let u_i denote the fraction of p_i that the entity was runnable.
1253 * We then designate the fractions u_i as our co-efficients, yielding the
1254 * following representation of historical load:
1255 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1257 * We choose y based on the with of a reasonably scheduling period, fixing:
1260 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1261 * approximately half as much as the contribution to load within the last ms
1264 * When a period "rolls over" and we have new u_0`, multiplying the previous
1265 * sum again by y is sufficient to update:
1266 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1267 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1269 static __always_inline int __update_entity_runnable_avg(u64 now,
1270 struct sched_avg *sa,
1274 u32 runnable_contrib;
1275 int delta_w, decayed = 0;
1277 delta = now - sa->last_runnable_update;
1279 * This should only happen when time goes backwards, which it
1280 * unfortunately does during sched clock init when we swap over to TSC.
1282 if ((s64)delta < 0) {
1283 sa->last_runnable_update = now;
1288 * Use 1024ns as the unit of measurement since it's a reasonable
1289 * approximation of 1us and fast to compute.
1294 sa->last_runnable_update = now;
1296 /* delta_w is the amount already accumulated against our next period */
1297 delta_w = sa->runnable_avg_period % 1024;
1298 if (delta + delta_w >= 1024) {
1299 /* period roll-over */
1303 * Now that we know we're crossing a period boundary, figure
1304 * out how much from delta we need to complete the current
1305 * period and accrue it.
1307 delta_w = 1024 - delta_w;
1309 sa->runnable_avg_sum += delta_w;
1310 sa->runnable_avg_period += delta_w;
1314 /* Figure out how many additional periods this update spans */
1315 periods = delta / 1024;
1318 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1320 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1323 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1324 runnable_contrib = __compute_runnable_contrib(periods);
1326 sa->runnable_avg_sum += runnable_contrib;
1327 sa->runnable_avg_period += runnable_contrib;
1330 /* Remainder of delta accrued against u_0` */
1332 sa->runnable_avg_sum += delta;
1333 sa->runnable_avg_period += delta;
1338 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1339 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1341 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1342 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1344 decays -= se->avg.decay_count;
1348 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1349 se->avg.decay_count = 0;
1354 #ifdef CONFIG_FAIR_GROUP_SCHED
1355 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1358 struct task_group *tg = cfs_rq->tg;
1361 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1362 tg_contrib -= cfs_rq->tg_load_contrib;
1364 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1365 atomic_long_add(tg_contrib, &tg->load_avg);
1366 cfs_rq->tg_load_contrib += tg_contrib;
1371 * Aggregate cfs_rq runnable averages into an equivalent task_group
1372 * representation for computing load contributions.
1374 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1375 struct cfs_rq *cfs_rq)
1377 struct task_group *tg = cfs_rq->tg;
1380 /* The fraction of a cpu used by this cfs_rq */
1381 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1382 sa->runnable_avg_period + 1);
1383 contrib -= cfs_rq->tg_runnable_contrib;
1385 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1386 atomic_add(contrib, &tg->runnable_avg);
1387 cfs_rq->tg_runnable_contrib += contrib;
1391 static inline void __update_group_entity_contrib(struct sched_entity *se)
1393 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1394 struct task_group *tg = cfs_rq->tg;
1399 contrib = cfs_rq->tg_load_contrib * tg->shares;
1400 se->avg.load_avg_contrib = div_u64(contrib,
1401 atomic_long_read(&tg->load_avg) + 1);
1404 * For group entities we need to compute a correction term in the case
1405 * that they are consuming <1 cpu so that we would contribute the same
1406 * load as a task of equal weight.
1408 * Explicitly co-ordinating this measurement would be expensive, but
1409 * fortunately the sum of each cpus contribution forms a usable
1410 * lower-bound on the true value.
1412 * Consider the aggregate of 2 contributions. Either they are disjoint
1413 * (and the sum represents true value) or they are disjoint and we are
1414 * understating by the aggregate of their overlap.
1416 * Extending this to N cpus, for a given overlap, the maximum amount we
1417 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1418 * cpus that overlap for this interval and w_i is the interval width.
1420 * On a small machine; the first term is well-bounded which bounds the
1421 * total error since w_i is a subset of the period. Whereas on a
1422 * larger machine, while this first term can be larger, if w_i is the
1423 * of consequential size guaranteed to see n_i*w_i quickly converge to
1424 * our upper bound of 1-cpu.
1426 runnable_avg = atomic_read(&tg->runnable_avg);
1427 if (runnable_avg < NICE_0_LOAD) {
1428 se->avg.load_avg_contrib *= runnable_avg;
1429 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1433 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1434 int force_update) {}
1435 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1436 struct cfs_rq *cfs_rq) {}
1437 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1440 static inline void __update_task_entity_contrib(struct sched_entity *se)
1444 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1445 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1446 contrib /= (se->avg.runnable_avg_period + 1);
1447 se->avg.load_avg_contrib = scale_load(contrib);
1450 /* Compute the current contribution to load_avg by se, return any delta */
1451 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1453 long old_contrib = se->avg.load_avg_contrib;
1455 if (entity_is_task(se)) {
1456 __update_task_entity_contrib(se);
1458 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1459 __update_group_entity_contrib(se);
1462 return se->avg.load_avg_contrib - old_contrib;
1465 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1468 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1469 cfs_rq->blocked_load_avg -= load_contrib;
1471 cfs_rq->blocked_load_avg = 0;
1474 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1476 /* Update a sched_entity's runnable average */
1477 static inline void update_entity_load_avg(struct sched_entity *se,
1480 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1485 * For a group entity we need to use their owned cfs_rq_clock_task() in
1486 * case they are the parent of a throttled hierarchy.
1488 if (entity_is_task(se))
1489 now = cfs_rq_clock_task(cfs_rq);
1491 now = cfs_rq_clock_task(group_cfs_rq(se));
1493 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1496 contrib_delta = __update_entity_load_avg_contrib(se);
1502 cfs_rq->runnable_load_avg += contrib_delta;
1504 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1508 * Decay the load contributed by all blocked children and account this so that
1509 * their contribution may appropriately discounted when they wake up.
1511 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1513 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1516 decays = now - cfs_rq->last_decay;
1517 if (!decays && !force_update)
1520 if (atomic_long_read(&cfs_rq->removed_load)) {
1521 unsigned long removed_load;
1522 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1523 subtract_blocked_load_contrib(cfs_rq, removed_load);
1527 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1529 atomic64_add(decays, &cfs_rq->decay_counter);
1530 cfs_rq->last_decay = now;
1533 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1536 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1538 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1539 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1542 /* Add the load generated by se into cfs_rq's child load-average */
1543 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1544 struct sched_entity *se,
1548 * We track migrations using entity decay_count <= 0, on a wake-up
1549 * migration we use a negative decay count to track the remote decays
1550 * accumulated while sleeping.
1552 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1553 * are seen by enqueue_entity_load_avg() as a migration with an already
1554 * constructed load_avg_contrib.
1556 if (unlikely(se->avg.decay_count <= 0)) {
1557 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1558 if (se->avg.decay_count) {
1560 * In a wake-up migration we have to approximate the
1561 * time sleeping. This is because we can't synchronize
1562 * clock_task between the two cpus, and it is not
1563 * guaranteed to be read-safe. Instead, we can
1564 * approximate this using our carried decays, which are
1565 * explicitly atomically readable.
1567 se->avg.last_runnable_update -= (-se->avg.decay_count)
1569 update_entity_load_avg(se, 0);
1570 /* Indicate that we're now synchronized and on-rq */
1571 se->avg.decay_count = 0;
1576 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1577 * would have made count negative); we must be careful to avoid
1578 * double-accounting blocked time after synchronizing decays.
1580 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1584 /* migrated tasks did not contribute to our blocked load */
1586 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1587 update_entity_load_avg(se, 0);
1590 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1591 /* we force update consideration on load-balancer moves */
1592 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1596 * Remove se's load from this cfs_rq child load-average, if the entity is
1597 * transitioning to a blocked state we track its projected decay using
1600 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1601 struct sched_entity *se,
1604 update_entity_load_avg(se, 1);
1605 /* we force update consideration on load-balancer moves */
1606 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1608 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1610 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1611 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1612 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1616 * Update the rq's load with the elapsed running time before entering
1617 * idle. if the last scheduled task is not a CFS task, idle_enter will
1618 * be the only way to update the runnable statistic.
1620 void idle_enter_fair(struct rq *this_rq)
1622 update_rq_runnable_avg(this_rq, 1);
1626 * Update the rq's load with the elapsed idle time before a task is
1627 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1628 * be the only way to update the runnable statistic.
1630 void idle_exit_fair(struct rq *this_rq)
1632 update_rq_runnable_avg(this_rq, 0);
1636 static inline void update_entity_load_avg(struct sched_entity *se,
1637 int update_cfs_rq) {}
1638 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1639 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1640 struct sched_entity *se,
1642 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1643 struct sched_entity *se,
1645 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1646 int force_update) {}
1649 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1651 #ifdef CONFIG_SCHEDSTATS
1652 struct task_struct *tsk = NULL;
1654 if (entity_is_task(se))
1657 if (se->statistics.sleep_start) {
1658 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1663 if (unlikely(delta > se->statistics.sleep_max))
1664 se->statistics.sleep_max = delta;
1666 se->statistics.sleep_start = 0;
1667 se->statistics.sum_sleep_runtime += delta;
1670 account_scheduler_latency(tsk, delta >> 10, 1);
1671 trace_sched_stat_sleep(tsk, delta);
1674 if (se->statistics.block_start) {
1675 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1680 if (unlikely(delta > se->statistics.block_max))
1681 se->statistics.block_max = delta;
1683 se->statistics.block_start = 0;
1684 se->statistics.sum_sleep_runtime += delta;
1687 if (tsk->in_iowait) {
1688 se->statistics.iowait_sum += delta;
1689 se->statistics.iowait_count++;
1690 trace_sched_stat_iowait(tsk, delta);
1693 trace_sched_stat_blocked(tsk, delta);
1696 * Blocking time is in units of nanosecs, so shift by
1697 * 20 to get a milliseconds-range estimation of the
1698 * amount of time that the task spent sleeping:
1700 if (unlikely(prof_on == SLEEP_PROFILING)) {
1701 profile_hits(SLEEP_PROFILING,
1702 (void *)get_wchan(tsk),
1705 account_scheduler_latency(tsk, delta >> 10, 0);
1711 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1713 #ifdef CONFIG_SCHED_DEBUG
1714 s64 d = se->vruntime - cfs_rq->min_vruntime;
1719 if (d > 3*sysctl_sched_latency)
1720 schedstat_inc(cfs_rq, nr_spread_over);
1725 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1727 u64 vruntime = cfs_rq->min_vruntime;
1730 * The 'current' period is already promised to the current tasks,
1731 * however the extra weight of the new task will slow them down a
1732 * little, place the new task so that it fits in the slot that
1733 * stays open at the end.
1735 if (initial && sched_feat(START_DEBIT))
1736 vruntime += sched_vslice(cfs_rq, se);
1738 /* sleeps up to a single latency don't count. */
1740 unsigned long thresh = sysctl_sched_latency;
1743 * Halve their sleep time's effect, to allow
1744 * for a gentler effect of sleepers:
1746 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1752 /* ensure we never gain time by being placed backwards. */
1753 se->vruntime = max_vruntime(se->vruntime, vruntime);
1756 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1759 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1762 * Update the normalized vruntime before updating min_vruntime
1763 * through calling update_curr().
1765 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1766 se->vruntime += cfs_rq->min_vruntime;
1769 * Update run-time statistics of the 'current'.
1771 update_curr(cfs_rq);
1772 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1773 account_entity_enqueue(cfs_rq, se);
1774 update_cfs_shares(cfs_rq);
1776 if (flags & ENQUEUE_WAKEUP) {
1777 place_entity(cfs_rq, se, 0);
1778 enqueue_sleeper(cfs_rq, se);
1781 update_stats_enqueue(cfs_rq, se);
1782 check_spread(cfs_rq, se);
1783 if (se != cfs_rq->curr)
1784 __enqueue_entity(cfs_rq, se);
1787 if (cfs_rq->nr_running == 1) {
1788 list_add_leaf_cfs_rq(cfs_rq);
1789 check_enqueue_throttle(cfs_rq);
1793 static void __clear_buddies_last(struct sched_entity *se)
1795 for_each_sched_entity(se) {
1796 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1797 if (cfs_rq->last == se)
1798 cfs_rq->last = NULL;
1804 static void __clear_buddies_next(struct sched_entity *se)
1806 for_each_sched_entity(se) {
1807 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1808 if (cfs_rq->next == se)
1809 cfs_rq->next = NULL;
1815 static void __clear_buddies_skip(struct sched_entity *se)
1817 for_each_sched_entity(se) {
1818 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1819 if (cfs_rq->skip == se)
1820 cfs_rq->skip = NULL;
1826 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1828 if (cfs_rq->last == se)
1829 __clear_buddies_last(se);
1831 if (cfs_rq->next == se)
1832 __clear_buddies_next(se);
1834 if (cfs_rq->skip == se)
1835 __clear_buddies_skip(se);
1838 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1841 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1844 * Update run-time statistics of the 'current'.
1846 update_curr(cfs_rq);
1847 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1849 update_stats_dequeue(cfs_rq, se);
1850 if (flags & DEQUEUE_SLEEP) {
1851 #ifdef CONFIG_SCHEDSTATS
1852 if (entity_is_task(se)) {
1853 struct task_struct *tsk = task_of(se);
1855 if (tsk->state & TASK_INTERRUPTIBLE)
1856 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
1857 if (tsk->state & TASK_UNINTERRUPTIBLE)
1858 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
1863 clear_buddies(cfs_rq, se);
1865 if (se != cfs_rq->curr)
1866 __dequeue_entity(cfs_rq, se);
1868 account_entity_dequeue(cfs_rq, se);
1871 * Normalize the entity after updating the min_vruntime because the
1872 * update can refer to the ->curr item and we need to reflect this
1873 * movement in our normalized position.
1875 if (!(flags & DEQUEUE_SLEEP))
1876 se->vruntime -= cfs_rq->min_vruntime;
1878 /* return excess runtime on last dequeue */
1879 return_cfs_rq_runtime(cfs_rq);
1881 update_min_vruntime(cfs_rq);
1882 update_cfs_shares(cfs_rq);
1886 * Preempt the current task with a newly woken task if needed:
1889 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1891 unsigned long ideal_runtime, delta_exec;
1892 struct sched_entity *se;
1895 ideal_runtime = sched_slice(cfs_rq, curr);
1896 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1897 if (delta_exec > ideal_runtime) {
1898 resched_task(rq_of(cfs_rq)->curr);
1900 * The current task ran long enough, ensure it doesn't get
1901 * re-elected due to buddy favours.
1903 clear_buddies(cfs_rq, curr);
1908 * Ensure that a task that missed wakeup preemption by a
1909 * narrow margin doesn't have to wait for a full slice.
1910 * This also mitigates buddy induced latencies under load.
1912 if (delta_exec < sysctl_sched_min_granularity)
1915 se = __pick_first_entity(cfs_rq);
1916 delta = curr->vruntime - se->vruntime;
1921 if (delta > ideal_runtime)
1922 resched_task(rq_of(cfs_rq)->curr);
1926 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1928 /* 'current' is not kept within the tree. */
1931 * Any task has to be enqueued before it get to execute on
1932 * a CPU. So account for the time it spent waiting on the
1935 update_stats_wait_end(cfs_rq, se);
1936 __dequeue_entity(cfs_rq, se);
1939 update_stats_curr_start(cfs_rq, se);
1941 #ifdef CONFIG_SCHEDSTATS
1943 * Track our maximum slice length, if the CPU's load is at
1944 * least twice that of our own weight (i.e. dont track it
1945 * when there are only lesser-weight tasks around):
1947 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1948 se->statistics.slice_max = max(se->statistics.slice_max,
1949 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1952 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1956 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1959 * Pick the next process, keeping these things in mind, in this order:
1960 * 1) keep things fair between processes/task groups
1961 * 2) pick the "next" process, since someone really wants that to run
1962 * 3) pick the "last" process, for cache locality
1963 * 4) do not run the "skip" process, if something else is available
1965 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1967 struct sched_entity *se = __pick_first_entity(cfs_rq);
1968 struct sched_entity *left = se;
1971 * Avoid running the skip buddy, if running something else can
1972 * be done without getting too unfair.
1974 if (cfs_rq->skip == se) {
1975 struct sched_entity *second = __pick_next_entity(se);
1976 if (second && wakeup_preempt_entity(second, left) < 1)
1981 * Prefer last buddy, try to return the CPU to a preempted task.
1983 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1987 * Someone really wants this to run. If it's not unfair, run it.
1989 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1992 clear_buddies(cfs_rq, se);
1997 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1999 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2002 * If still on the runqueue then deactivate_task()
2003 * was not called and update_curr() has to be done:
2006 update_curr(cfs_rq);
2008 /* throttle cfs_rqs exceeding runtime */
2009 check_cfs_rq_runtime(cfs_rq);
2011 check_spread(cfs_rq, prev);
2013 update_stats_wait_start(cfs_rq, prev);
2014 /* Put 'current' back into the tree. */
2015 __enqueue_entity(cfs_rq, prev);
2016 /* in !on_rq case, update occurred at dequeue */
2017 update_entity_load_avg(prev, 1);
2019 cfs_rq->curr = NULL;
2023 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2026 * Update run-time statistics of the 'current'.
2028 update_curr(cfs_rq);
2031 * Ensure that runnable average is periodically updated.
2033 update_entity_load_avg(curr, 1);
2034 update_cfs_rq_blocked_load(cfs_rq, 1);
2035 update_cfs_shares(cfs_rq);
2037 #ifdef CONFIG_SCHED_HRTICK
2039 * queued ticks are scheduled to match the slice, so don't bother
2040 * validating it and just reschedule.
2043 resched_task(rq_of(cfs_rq)->curr);
2047 * don't let the period tick interfere with the hrtick preemption
2049 if (!sched_feat(DOUBLE_TICK) &&
2050 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2054 if (cfs_rq->nr_running > 1)
2055 check_preempt_tick(cfs_rq, curr);
2059 /**************************************************
2060 * CFS bandwidth control machinery
2063 #ifdef CONFIG_CFS_BANDWIDTH
2065 #ifdef HAVE_JUMP_LABEL
2066 static struct static_key __cfs_bandwidth_used;
2068 static inline bool cfs_bandwidth_used(void)
2070 return static_key_false(&__cfs_bandwidth_used);
2073 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2075 /* only need to count groups transitioning between enabled/!enabled */
2076 if (enabled && !was_enabled)
2077 static_key_slow_inc(&__cfs_bandwidth_used);
2078 else if (!enabled && was_enabled)
2079 static_key_slow_dec(&__cfs_bandwidth_used);
2081 #else /* HAVE_JUMP_LABEL */
2082 static bool cfs_bandwidth_used(void)
2087 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2088 #endif /* HAVE_JUMP_LABEL */
2091 * default period for cfs group bandwidth.
2092 * default: 0.1s, units: nanoseconds
2094 static inline u64 default_cfs_period(void)
2096 return 100000000ULL;
2099 static inline u64 sched_cfs_bandwidth_slice(void)
2101 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2105 * Replenish runtime according to assigned quota and update expiration time.
2106 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2107 * additional synchronization around rq->lock.
2109 * requires cfs_b->lock
2111 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2115 if (cfs_b->quota == RUNTIME_INF)
2118 now = sched_clock_cpu(smp_processor_id());
2119 cfs_b->runtime = cfs_b->quota;
2120 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2123 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2125 return &tg->cfs_bandwidth;
2128 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2129 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2131 if (unlikely(cfs_rq->throttle_count))
2132 return cfs_rq->throttled_clock_task;
2134 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2137 /* returns 0 on failure to allocate runtime */
2138 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2140 struct task_group *tg = cfs_rq->tg;
2141 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2142 u64 amount = 0, min_amount, expires;
2144 /* note: this is a positive sum as runtime_remaining <= 0 */
2145 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2147 raw_spin_lock(&cfs_b->lock);
2148 if (cfs_b->quota == RUNTIME_INF)
2149 amount = min_amount;
2152 * If the bandwidth pool has become inactive, then at least one
2153 * period must have elapsed since the last consumption.
2154 * Refresh the global state and ensure bandwidth timer becomes
2157 if (!cfs_b->timer_active) {
2158 __refill_cfs_bandwidth_runtime(cfs_b);
2159 __start_cfs_bandwidth(cfs_b);
2162 if (cfs_b->runtime > 0) {
2163 amount = min(cfs_b->runtime, min_amount);
2164 cfs_b->runtime -= amount;
2168 expires = cfs_b->runtime_expires;
2169 raw_spin_unlock(&cfs_b->lock);
2171 cfs_rq->runtime_remaining += amount;
2173 * we may have advanced our local expiration to account for allowed
2174 * spread between our sched_clock and the one on which runtime was
2177 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2178 cfs_rq->runtime_expires = expires;
2180 return cfs_rq->runtime_remaining > 0;
2184 * Note: This depends on the synchronization provided by sched_clock and the
2185 * fact that rq->clock snapshots this value.
2187 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2189 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2191 /* if the deadline is ahead of our clock, nothing to do */
2192 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2195 if (cfs_rq->runtime_remaining < 0)
2199 * If the local deadline has passed we have to consider the
2200 * possibility that our sched_clock is 'fast' and the global deadline
2201 * has not truly expired.
2203 * Fortunately we can check determine whether this the case by checking
2204 * whether the global deadline has advanced.
2207 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2208 /* extend local deadline, drift is bounded above by 2 ticks */
2209 cfs_rq->runtime_expires += TICK_NSEC;
2211 /* global deadline is ahead, expiration has passed */
2212 cfs_rq->runtime_remaining = 0;
2216 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2217 unsigned long delta_exec)
2219 /* dock delta_exec before expiring quota (as it could span periods) */
2220 cfs_rq->runtime_remaining -= delta_exec;
2221 expire_cfs_rq_runtime(cfs_rq);
2223 if (likely(cfs_rq->runtime_remaining > 0))
2227 * if we're unable to extend our runtime we resched so that the active
2228 * hierarchy can be throttled
2230 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2231 resched_task(rq_of(cfs_rq)->curr);
2234 static __always_inline
2235 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2237 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2240 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2243 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2245 return cfs_bandwidth_used() && cfs_rq->throttled;
2248 /* check whether cfs_rq, or any parent, is throttled */
2249 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2251 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2255 * Ensure that neither of the group entities corresponding to src_cpu or
2256 * dest_cpu are members of a throttled hierarchy when performing group
2257 * load-balance operations.
2259 static inline int throttled_lb_pair(struct task_group *tg,
2260 int src_cpu, int dest_cpu)
2262 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2264 src_cfs_rq = tg->cfs_rq[src_cpu];
2265 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2267 return throttled_hierarchy(src_cfs_rq) ||
2268 throttled_hierarchy(dest_cfs_rq);
2271 /* updated child weight may affect parent so we have to do this bottom up */
2272 static int tg_unthrottle_up(struct task_group *tg, void *data)
2274 struct rq *rq = data;
2275 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2277 cfs_rq->throttle_count--;
2279 if (!cfs_rq->throttle_count) {
2280 /* adjust cfs_rq_clock_task() */
2281 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2282 cfs_rq->throttled_clock_task;
2289 static int tg_throttle_down(struct task_group *tg, void *data)
2291 struct rq *rq = data;
2292 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2294 /* group is entering throttled state, stop time */
2295 if (!cfs_rq->throttle_count)
2296 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2297 cfs_rq->throttle_count++;
2302 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2304 struct rq *rq = rq_of(cfs_rq);
2305 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2306 struct sched_entity *se;
2307 long task_delta, dequeue = 1;
2309 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2311 /* freeze hierarchy runnable averages while throttled */
2313 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2316 task_delta = cfs_rq->h_nr_running;
2317 for_each_sched_entity(se) {
2318 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2319 /* throttled entity or throttle-on-deactivate */
2324 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2325 qcfs_rq->h_nr_running -= task_delta;
2327 if (qcfs_rq->load.weight)
2332 rq->nr_running -= task_delta;
2334 cfs_rq->throttled = 1;
2335 cfs_rq->throttled_clock = rq_clock(rq);
2336 raw_spin_lock(&cfs_b->lock);
2337 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2338 raw_spin_unlock(&cfs_b->lock);
2341 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2343 struct rq *rq = rq_of(cfs_rq);
2344 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2345 struct sched_entity *se;
2349 se = cfs_rq->tg->se[cpu_of(rq)];
2351 cfs_rq->throttled = 0;
2353 update_rq_clock(rq);
2355 raw_spin_lock(&cfs_b->lock);
2356 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2357 list_del_rcu(&cfs_rq->throttled_list);
2358 raw_spin_unlock(&cfs_b->lock);
2360 /* update hierarchical throttle state */
2361 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2363 if (!cfs_rq->load.weight)
2366 task_delta = cfs_rq->h_nr_running;
2367 for_each_sched_entity(se) {
2371 cfs_rq = cfs_rq_of(se);
2373 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2374 cfs_rq->h_nr_running += task_delta;
2376 if (cfs_rq_throttled(cfs_rq))
2381 rq->nr_running += task_delta;
2383 /* determine whether we need to wake up potentially idle cpu */
2384 if (rq->curr == rq->idle && rq->cfs.nr_running)
2385 resched_task(rq->curr);
2388 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2389 u64 remaining, u64 expires)
2391 struct cfs_rq *cfs_rq;
2392 u64 runtime = remaining;
2395 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2397 struct rq *rq = rq_of(cfs_rq);
2399 raw_spin_lock(&rq->lock);
2400 if (!cfs_rq_throttled(cfs_rq))
2403 runtime = -cfs_rq->runtime_remaining + 1;
2404 if (runtime > remaining)
2405 runtime = remaining;
2406 remaining -= runtime;
2408 cfs_rq->runtime_remaining += runtime;
2409 cfs_rq->runtime_expires = expires;
2411 /* we check whether we're throttled above */
2412 if (cfs_rq->runtime_remaining > 0)
2413 unthrottle_cfs_rq(cfs_rq);
2416 raw_spin_unlock(&rq->lock);
2427 * Responsible for refilling a task_group's bandwidth and unthrottling its
2428 * cfs_rqs as appropriate. If there has been no activity within the last
2429 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2430 * used to track this state.
2432 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2434 u64 runtime, runtime_expires;
2435 int idle = 1, throttled;
2437 raw_spin_lock(&cfs_b->lock);
2438 /* no need to continue the timer with no bandwidth constraint */
2439 if (cfs_b->quota == RUNTIME_INF)
2442 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2443 /* idle depends on !throttled (for the case of a large deficit) */
2444 idle = cfs_b->idle && !throttled;
2445 cfs_b->nr_periods += overrun;
2447 /* if we're going inactive then everything else can be deferred */
2451 __refill_cfs_bandwidth_runtime(cfs_b);
2454 /* mark as potentially idle for the upcoming period */
2459 /* account preceding periods in which throttling occurred */
2460 cfs_b->nr_throttled += overrun;
2463 * There are throttled entities so we must first use the new bandwidth
2464 * to unthrottle them before making it generally available. This
2465 * ensures that all existing debts will be paid before a new cfs_rq is
2468 runtime = cfs_b->runtime;
2469 runtime_expires = cfs_b->runtime_expires;
2473 * This check is repeated as we are holding onto the new bandwidth
2474 * while we unthrottle. This can potentially race with an unthrottled
2475 * group trying to acquire new bandwidth from the global pool.
2477 while (throttled && runtime > 0) {
2478 raw_spin_unlock(&cfs_b->lock);
2479 /* we can't nest cfs_b->lock while distributing bandwidth */
2480 runtime = distribute_cfs_runtime(cfs_b, runtime,
2482 raw_spin_lock(&cfs_b->lock);
2484 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2487 /* return (any) remaining runtime */
2488 cfs_b->runtime = runtime;
2490 * While we are ensured activity in the period following an
2491 * unthrottle, this also covers the case in which the new bandwidth is
2492 * insufficient to cover the existing bandwidth deficit. (Forcing the
2493 * timer to remain active while there are any throttled entities.)
2498 cfs_b->timer_active = 0;
2499 raw_spin_unlock(&cfs_b->lock);
2504 /* a cfs_rq won't donate quota below this amount */
2505 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2506 /* minimum remaining period time to redistribute slack quota */
2507 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2508 /* how long we wait to gather additional slack before distributing */
2509 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2511 /* are we near the end of the current quota period? */
2512 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2514 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2517 /* if the call-back is running a quota refresh is already occurring */
2518 if (hrtimer_callback_running(refresh_timer))
2521 /* is a quota refresh about to occur? */
2522 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2523 if (remaining < min_expire)
2529 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2531 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2533 /* if there's a quota refresh soon don't bother with slack */
2534 if (runtime_refresh_within(cfs_b, min_left))
2537 start_bandwidth_timer(&cfs_b->slack_timer,
2538 ns_to_ktime(cfs_bandwidth_slack_period));
2541 /* we know any runtime found here is valid as update_curr() precedes return */
2542 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2544 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2545 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2547 if (slack_runtime <= 0)
2550 raw_spin_lock(&cfs_b->lock);
2551 if (cfs_b->quota != RUNTIME_INF &&
2552 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2553 cfs_b->runtime += slack_runtime;
2555 /* we are under rq->lock, defer unthrottling using a timer */
2556 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2557 !list_empty(&cfs_b->throttled_cfs_rq))
2558 start_cfs_slack_bandwidth(cfs_b);
2560 raw_spin_unlock(&cfs_b->lock);
2562 /* even if it's not valid for return we don't want to try again */
2563 cfs_rq->runtime_remaining -= slack_runtime;
2566 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2568 if (!cfs_bandwidth_used())
2571 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2574 __return_cfs_rq_runtime(cfs_rq);
2578 * This is done with a timer (instead of inline with bandwidth return) since
2579 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2581 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2583 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2586 /* confirm we're still not at a refresh boundary */
2587 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2590 raw_spin_lock(&cfs_b->lock);
2591 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2592 runtime = cfs_b->runtime;
2595 expires = cfs_b->runtime_expires;
2596 raw_spin_unlock(&cfs_b->lock);
2601 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2603 raw_spin_lock(&cfs_b->lock);
2604 if (expires == cfs_b->runtime_expires)
2605 cfs_b->runtime = runtime;
2606 raw_spin_unlock(&cfs_b->lock);
2610 * When a group wakes up we want to make sure that its quota is not already
2611 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2612 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2614 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2616 if (!cfs_bandwidth_used())
2619 /* an active group must be handled by the update_curr()->put() path */
2620 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2623 /* ensure the group is not already throttled */
2624 if (cfs_rq_throttled(cfs_rq))
2627 /* update runtime allocation */
2628 account_cfs_rq_runtime(cfs_rq, 0);
2629 if (cfs_rq->runtime_remaining <= 0)
2630 throttle_cfs_rq(cfs_rq);
2633 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2634 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2636 if (!cfs_bandwidth_used())
2639 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2643 * it's possible for a throttled entity to be forced into a running
2644 * state (e.g. set_curr_task), in this case we're finished.
2646 if (cfs_rq_throttled(cfs_rq))
2649 throttle_cfs_rq(cfs_rq);
2652 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2654 struct cfs_bandwidth *cfs_b =
2655 container_of(timer, struct cfs_bandwidth, slack_timer);
2656 do_sched_cfs_slack_timer(cfs_b);
2658 return HRTIMER_NORESTART;
2661 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2663 struct cfs_bandwidth *cfs_b =
2664 container_of(timer, struct cfs_bandwidth, period_timer);
2670 now = hrtimer_cb_get_time(timer);
2671 overrun = hrtimer_forward(timer, now, cfs_b->period);
2676 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2679 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2682 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2684 raw_spin_lock_init(&cfs_b->lock);
2686 cfs_b->quota = RUNTIME_INF;
2687 cfs_b->period = ns_to_ktime(default_cfs_period());
2689 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2690 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2691 cfs_b->period_timer.function = sched_cfs_period_timer;
2692 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2693 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2696 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2698 cfs_rq->runtime_enabled = 0;
2699 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2702 /* requires cfs_b->lock, may release to reprogram timer */
2703 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2706 * The timer may be active because we're trying to set a new bandwidth
2707 * period or because we're racing with the tear-down path
2708 * (timer_active==0 becomes visible before the hrtimer call-back
2709 * terminates). In either case we ensure that it's re-programmed
2711 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2712 raw_spin_unlock(&cfs_b->lock);
2713 /* ensure cfs_b->lock is available while we wait */
2714 hrtimer_cancel(&cfs_b->period_timer);
2716 raw_spin_lock(&cfs_b->lock);
2717 /* if someone else restarted the timer then we're done */
2718 if (cfs_b->timer_active)
2722 cfs_b->timer_active = 1;
2723 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2726 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2728 hrtimer_cancel(&cfs_b->period_timer);
2729 hrtimer_cancel(&cfs_b->slack_timer);
2732 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2734 struct cfs_rq *cfs_rq;
2736 for_each_leaf_cfs_rq(rq, cfs_rq) {
2737 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2739 if (!cfs_rq->runtime_enabled)
2743 * clock_task is not advancing so we just need to make sure
2744 * there's some valid quota amount
2746 cfs_rq->runtime_remaining = cfs_b->quota;
2747 if (cfs_rq_throttled(cfs_rq))
2748 unthrottle_cfs_rq(cfs_rq);
2752 #else /* CONFIG_CFS_BANDWIDTH */
2753 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2755 return rq_clock_task(rq_of(cfs_rq));
2758 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2759 unsigned long delta_exec) {}
2760 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2761 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2762 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2764 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2769 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2774 static inline int throttled_lb_pair(struct task_group *tg,
2775 int src_cpu, int dest_cpu)
2780 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2782 #ifdef CONFIG_FAIR_GROUP_SCHED
2783 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2786 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2790 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2791 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2793 #endif /* CONFIG_CFS_BANDWIDTH */
2795 /**************************************************
2796 * CFS operations on tasks:
2799 #ifdef CONFIG_SCHED_HRTICK
2800 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2802 struct sched_entity *se = &p->se;
2803 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2805 WARN_ON(task_rq(p) != rq);
2807 if (cfs_rq->nr_running > 1) {
2808 u64 slice = sched_slice(cfs_rq, se);
2809 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2810 s64 delta = slice - ran;
2819 * Don't schedule slices shorter than 10000ns, that just
2820 * doesn't make sense. Rely on vruntime for fairness.
2823 delta = max_t(s64, 10000LL, delta);
2825 hrtick_start(rq, delta);
2830 * called from enqueue/dequeue and updates the hrtick when the
2831 * current task is from our class and nr_running is low enough
2834 static void hrtick_update(struct rq *rq)
2836 struct task_struct *curr = rq->curr;
2838 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2841 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2842 hrtick_start_fair(rq, curr);
2844 #else /* !CONFIG_SCHED_HRTICK */
2846 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2850 static inline void hrtick_update(struct rq *rq)
2856 * The enqueue_task method is called before nr_running is
2857 * increased. Here we update the fair scheduling stats and
2858 * then put the task into the rbtree:
2861 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2863 struct cfs_rq *cfs_rq;
2864 struct sched_entity *se = &p->se;
2866 for_each_sched_entity(se) {
2869 cfs_rq = cfs_rq_of(se);
2870 enqueue_entity(cfs_rq, se, flags);
2873 * end evaluation on encountering a throttled cfs_rq
2875 * note: in the case of encountering a throttled cfs_rq we will
2876 * post the final h_nr_running increment below.
2878 if (cfs_rq_throttled(cfs_rq))
2880 cfs_rq->h_nr_running++;
2882 flags = ENQUEUE_WAKEUP;
2885 for_each_sched_entity(se) {
2886 cfs_rq = cfs_rq_of(se);
2887 cfs_rq->h_nr_running++;
2889 if (cfs_rq_throttled(cfs_rq))
2892 update_cfs_shares(cfs_rq);
2893 update_entity_load_avg(se, 1);
2897 update_rq_runnable_avg(rq, rq->nr_running);
2903 static void set_next_buddy(struct sched_entity *se);
2906 * The dequeue_task method is called before nr_running is
2907 * decreased. We remove the task from the rbtree and
2908 * update the fair scheduling stats:
2910 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2912 struct cfs_rq *cfs_rq;
2913 struct sched_entity *se = &p->se;
2914 int task_sleep = flags & DEQUEUE_SLEEP;
2916 for_each_sched_entity(se) {
2917 cfs_rq = cfs_rq_of(se);
2918 dequeue_entity(cfs_rq, se, flags);
2921 * end evaluation on encountering a throttled cfs_rq
2923 * note: in the case of encountering a throttled cfs_rq we will
2924 * post the final h_nr_running decrement below.
2926 if (cfs_rq_throttled(cfs_rq))
2928 cfs_rq->h_nr_running--;
2930 /* Don't dequeue parent if it has other entities besides us */
2931 if (cfs_rq->load.weight) {
2933 * Bias pick_next to pick a task from this cfs_rq, as
2934 * p is sleeping when it is within its sched_slice.
2936 if (task_sleep && parent_entity(se))
2937 set_next_buddy(parent_entity(se));
2939 /* avoid re-evaluating load for this entity */
2940 se = parent_entity(se);
2943 flags |= DEQUEUE_SLEEP;
2946 for_each_sched_entity(se) {
2947 cfs_rq = cfs_rq_of(se);
2948 cfs_rq->h_nr_running--;
2950 if (cfs_rq_throttled(cfs_rq))
2953 update_cfs_shares(cfs_rq);
2954 update_entity_load_avg(se, 1);
2959 update_rq_runnable_avg(rq, 1);
2965 /* Used instead of source_load when we know the type == 0 */
2966 static unsigned long weighted_cpuload(const int cpu)
2968 return cpu_rq(cpu)->cfs.runnable_load_avg;
2972 * Return a low guess at the load of a migration-source cpu weighted
2973 * according to the scheduling class and "nice" value.
2975 * We want to under-estimate the load of migration sources, to
2976 * balance conservatively.
2978 static unsigned long source_load(int cpu, int type)
2980 struct rq *rq = cpu_rq(cpu);
2981 unsigned long total = weighted_cpuload(cpu);
2983 if (type == 0 || !sched_feat(LB_BIAS))
2986 return min(rq->cpu_load[type-1], total);
2990 * Return a high guess at the load of a migration-target cpu weighted
2991 * according to the scheduling class and "nice" value.
2993 static unsigned long target_load(int cpu, int type)
2995 struct rq *rq = cpu_rq(cpu);
2996 unsigned long total = weighted_cpuload(cpu);
2998 if (type == 0 || !sched_feat(LB_BIAS))
3001 return max(rq->cpu_load[type-1], total);
3004 static unsigned long power_of(int cpu)
3006 return cpu_rq(cpu)->cpu_power;
3009 static unsigned long cpu_avg_load_per_task(int cpu)
3011 struct rq *rq = cpu_rq(cpu);
3012 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3013 unsigned long load_avg = rq->cfs.runnable_load_avg;
3016 return load_avg / nr_running;
3022 static void task_waking_fair(struct task_struct *p)
3024 struct sched_entity *se = &p->se;
3025 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3028 #ifndef CONFIG_64BIT
3029 u64 min_vruntime_copy;
3032 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3034 min_vruntime = cfs_rq->min_vruntime;
3035 } while (min_vruntime != min_vruntime_copy);
3037 min_vruntime = cfs_rq->min_vruntime;
3040 se->vruntime -= min_vruntime;
3043 #ifdef CONFIG_FAIR_GROUP_SCHED
3045 * effective_load() calculates the load change as seen from the root_task_group
3047 * Adding load to a group doesn't make a group heavier, but can cause movement
3048 * of group shares between cpus. Assuming the shares were perfectly aligned one
3049 * can calculate the shift in shares.
3051 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3052 * on this @cpu and results in a total addition (subtraction) of @wg to the
3053 * total group weight.
3055 * Given a runqueue weight distribution (rw_i) we can compute a shares
3056 * distribution (s_i) using:
3058 * s_i = rw_i / \Sum rw_j (1)
3060 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3061 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3062 * shares distribution (s_i):
3064 * rw_i = { 2, 4, 1, 0 }
3065 * s_i = { 2/7, 4/7, 1/7, 0 }
3067 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3068 * task used to run on and the CPU the waker is running on), we need to
3069 * compute the effect of waking a task on either CPU and, in case of a sync
3070 * wakeup, compute the effect of the current task going to sleep.
3072 * So for a change of @wl to the local @cpu with an overall group weight change
3073 * of @wl we can compute the new shares distribution (s'_i) using:
3075 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3077 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3078 * differences in waking a task to CPU 0. The additional task changes the
3079 * weight and shares distributions like:
3081 * rw'_i = { 3, 4, 1, 0 }
3082 * s'_i = { 3/8, 4/8, 1/8, 0 }
3084 * We can then compute the difference in effective weight by using:
3086 * dw_i = S * (s'_i - s_i) (3)
3088 * Where 'S' is the group weight as seen by its parent.
3090 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3091 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3092 * 4/7) times the weight of the group.
3094 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3096 struct sched_entity *se = tg->se[cpu];
3098 if (!tg->parent) /* the trivial, non-cgroup case */
3101 for_each_sched_entity(se) {
3107 * W = @wg + \Sum rw_j
3109 W = wg + calc_tg_weight(tg, se->my_q);
3114 w = se->my_q->load.weight + wl;
3117 * wl = S * s'_i; see (2)
3120 wl = (w * tg->shares) / W;
3125 * Per the above, wl is the new se->load.weight value; since
3126 * those are clipped to [MIN_SHARES, ...) do so now. See
3127 * calc_cfs_shares().
3129 if (wl < MIN_SHARES)
3133 * wl = dw_i = S * (s'_i - s_i); see (3)
3135 wl -= se->load.weight;
3138 * Recursively apply this logic to all parent groups to compute
3139 * the final effective load change on the root group. Since
3140 * only the @tg group gets extra weight, all parent groups can
3141 * only redistribute existing shares. @wl is the shift in shares
3142 * resulting from this level per the above.
3151 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3152 unsigned long wl, unsigned long wg)
3159 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3161 s64 this_load, load;
3162 int idx, this_cpu, prev_cpu;
3163 unsigned long tl_per_task;
3164 struct task_group *tg;
3165 unsigned long weight;
3169 this_cpu = smp_processor_id();
3170 prev_cpu = task_cpu(p);
3171 load = source_load(prev_cpu, idx);
3172 this_load = target_load(this_cpu, idx);
3175 * If sync wakeup then subtract the (maximum possible)
3176 * effect of the currently running task from the load
3177 * of the current CPU:
3180 tg = task_group(current);
3181 weight = current->se.load.weight;
3183 this_load += effective_load(tg, this_cpu, -weight, -weight);
3184 load += effective_load(tg, prev_cpu, 0, -weight);
3188 weight = p->se.load.weight;
3191 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3192 * due to the sync cause above having dropped this_load to 0, we'll
3193 * always have an imbalance, but there's really nothing you can do
3194 * about that, so that's good too.
3196 * Otherwise check if either cpus are near enough in load to allow this
3197 * task to be woken on this_cpu.
3199 if (this_load > 0) {
3200 s64 this_eff_load, prev_eff_load;
3202 this_eff_load = 100;
3203 this_eff_load *= power_of(prev_cpu);
3204 this_eff_load *= this_load +
3205 effective_load(tg, this_cpu, weight, weight);
3207 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3208 prev_eff_load *= power_of(this_cpu);
3209 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3211 balanced = this_eff_load <= prev_eff_load;
3216 * If the currently running task will sleep within
3217 * a reasonable amount of time then attract this newly
3220 if (sync && balanced)
3223 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3224 tl_per_task = cpu_avg_load_per_task(this_cpu);
3227 (this_load <= load &&
3228 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3230 * This domain has SD_WAKE_AFFINE and
3231 * p is cache cold in this domain, and
3232 * there is no bad imbalance.
3234 schedstat_inc(sd, ttwu_move_affine);
3235 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3243 * find_idlest_group finds and returns the least busy CPU group within the
3246 static struct sched_group *
3247 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3248 int this_cpu, int load_idx)
3250 struct sched_group *idlest = NULL, *group = sd->groups;
3251 unsigned long min_load = ULONG_MAX, this_load = 0;
3252 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3255 unsigned long load, avg_load;
3259 /* Skip over this group if it has no CPUs allowed */
3260 if (!cpumask_intersects(sched_group_cpus(group),
3261 tsk_cpus_allowed(p)))
3264 local_group = cpumask_test_cpu(this_cpu,
3265 sched_group_cpus(group));
3267 /* Tally up the load of all CPUs in the group */
3270 for_each_cpu(i, sched_group_cpus(group)) {
3271 /* Bias balancing toward cpus of our domain */
3273 load = source_load(i, load_idx);
3275 load = target_load(i, load_idx);
3280 /* Adjust by relative CPU power of the group */
3281 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3284 this_load = avg_load;
3285 } else if (avg_load < min_load) {
3286 min_load = avg_load;
3289 } while (group = group->next, group != sd->groups);
3291 if (!idlest || 100*this_load < imbalance*min_load)
3297 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3300 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3302 unsigned long load, min_load = ULONG_MAX;
3306 /* Traverse only the allowed CPUs */
3307 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3308 load = weighted_cpuload(i);
3310 if (load < min_load || (load == min_load && i == this_cpu)) {
3320 * Try and locate an idle CPU in the sched_domain.
3322 static int select_idle_sibling(struct task_struct *p, int target)
3324 struct sched_domain *sd;
3325 struct sched_group *sg;
3326 int i = task_cpu(p);
3328 if (idle_cpu(target))
3332 * If the prevous cpu is cache affine and idle, don't be stupid.
3334 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3338 * Otherwise, iterate the domains and find an elegible idle cpu.
3340 sd = rcu_dereference(per_cpu(sd_llc, target));
3341 for_each_lower_domain(sd) {
3344 if (!cpumask_intersects(sched_group_cpus(sg),
3345 tsk_cpus_allowed(p)))
3348 for_each_cpu(i, sched_group_cpus(sg)) {
3349 if (i == target || !idle_cpu(i))
3353 target = cpumask_first_and(sched_group_cpus(sg),
3354 tsk_cpus_allowed(p));
3358 } while (sg != sd->groups);
3365 * sched_balance_self: balance the current task (running on cpu) in domains
3366 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3369 * Balance, ie. select the least loaded group.
3371 * Returns the target CPU number, or the same CPU if no balancing is needed.
3373 * preempt must be disabled.
3376 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3378 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3379 int cpu = smp_processor_id();
3380 int prev_cpu = task_cpu(p);
3382 int want_affine = 0;
3383 int sync = wake_flags & WF_SYNC;
3385 if (p->nr_cpus_allowed == 1)
3388 if (sd_flag & SD_BALANCE_WAKE) {
3389 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3395 for_each_domain(cpu, tmp) {
3396 if (!(tmp->flags & SD_LOAD_BALANCE))
3400 * If both cpu and prev_cpu are part of this domain,
3401 * cpu is a valid SD_WAKE_AFFINE target.
3403 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3404 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3409 if (tmp->flags & sd_flag)
3414 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3417 new_cpu = select_idle_sibling(p, prev_cpu);
3422 int load_idx = sd->forkexec_idx;
3423 struct sched_group *group;
3426 if (!(sd->flags & sd_flag)) {
3431 if (sd_flag & SD_BALANCE_WAKE)
3432 load_idx = sd->wake_idx;
3434 group = find_idlest_group(sd, p, cpu, load_idx);
3440 new_cpu = find_idlest_cpu(group, p, cpu);
3441 if (new_cpu == -1 || new_cpu == cpu) {
3442 /* Now try balancing at a lower domain level of cpu */
3447 /* Now try balancing at a lower domain level of new_cpu */
3449 weight = sd->span_weight;
3451 for_each_domain(cpu, tmp) {
3452 if (weight <= tmp->span_weight)
3454 if (tmp->flags & sd_flag)
3457 /* while loop will break here if sd == NULL */
3466 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3467 * cfs_rq_of(p) references at time of call are still valid and identify the
3468 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3469 * other assumptions, including the state of rq->lock, should be made.
3472 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3474 struct sched_entity *se = &p->se;
3475 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3478 * Load tracking: accumulate removed load so that it can be processed
3479 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3480 * to blocked load iff they have a positive decay-count. It can never
3481 * be negative here since on-rq tasks have decay-count == 0.
3483 if (se->avg.decay_count) {
3484 se->avg.decay_count = -__synchronize_entity_decay(se);
3485 atomic_long_add(se->avg.load_avg_contrib,
3486 &cfs_rq->removed_load);
3489 #endif /* CONFIG_SMP */
3491 static unsigned long
3492 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3494 unsigned long gran = sysctl_sched_wakeup_granularity;
3497 * Since its curr running now, convert the gran from real-time
3498 * to virtual-time in his units.
3500 * By using 'se' instead of 'curr' we penalize light tasks, so
3501 * they get preempted easier. That is, if 'se' < 'curr' then
3502 * the resulting gran will be larger, therefore penalizing the
3503 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3504 * be smaller, again penalizing the lighter task.
3506 * This is especially important for buddies when the leftmost
3507 * task is higher priority than the buddy.
3509 return calc_delta_fair(gran, se);
3513 * Should 'se' preempt 'curr'.
3527 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3529 s64 gran, vdiff = curr->vruntime - se->vruntime;
3534 gran = wakeup_gran(curr, se);
3541 static void set_last_buddy(struct sched_entity *se)
3543 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3546 for_each_sched_entity(se)
3547 cfs_rq_of(se)->last = se;
3550 static void set_next_buddy(struct sched_entity *se)
3552 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3555 for_each_sched_entity(se)
3556 cfs_rq_of(se)->next = se;
3559 static void set_skip_buddy(struct sched_entity *se)
3561 for_each_sched_entity(se)
3562 cfs_rq_of(se)->skip = se;
3566 * Preempt the current task with a newly woken task if needed:
3568 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3570 struct task_struct *curr = rq->curr;
3571 struct sched_entity *se = &curr->se, *pse = &p->se;
3572 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3573 int scale = cfs_rq->nr_running >= sched_nr_latency;
3574 int next_buddy_marked = 0;
3576 if (unlikely(se == pse))
3580 * This is possible from callers such as move_task(), in which we
3581 * unconditionally check_prempt_curr() after an enqueue (which may have
3582 * lead to a throttle). This both saves work and prevents false
3583 * next-buddy nomination below.
3585 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3588 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3589 set_next_buddy(pse);
3590 next_buddy_marked = 1;
3594 * We can come here with TIF_NEED_RESCHED already set from new task
3597 * Note: this also catches the edge-case of curr being in a throttled
3598 * group (e.g. via set_curr_task), since update_curr() (in the
3599 * enqueue of curr) will have resulted in resched being set. This
3600 * prevents us from potentially nominating it as a false LAST_BUDDY
3603 if (test_tsk_need_resched(curr))
3606 /* Idle tasks are by definition preempted by non-idle tasks. */
3607 if (unlikely(curr->policy == SCHED_IDLE) &&
3608 likely(p->policy != SCHED_IDLE))
3612 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3613 * is driven by the tick):
3615 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3618 find_matching_se(&se, &pse);
3619 update_curr(cfs_rq_of(se));
3621 if (wakeup_preempt_entity(se, pse) == 1) {
3623 * Bias pick_next to pick the sched entity that is
3624 * triggering this preemption.
3626 if (!next_buddy_marked)
3627 set_next_buddy(pse);
3636 * Only set the backward buddy when the current task is still
3637 * on the rq. This can happen when a wakeup gets interleaved
3638 * with schedule on the ->pre_schedule() or idle_balance()
3639 * point, either of which can * drop the rq lock.
3641 * Also, during early boot the idle thread is in the fair class,
3642 * for obvious reasons its a bad idea to schedule back to it.
3644 if (unlikely(!se->on_rq || curr == rq->idle))
3647 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3651 static struct task_struct *pick_next_task_fair(struct rq *rq)
3653 struct task_struct *p;
3654 struct cfs_rq *cfs_rq = &rq->cfs;
3655 struct sched_entity *se;
3657 if (!cfs_rq->nr_running)
3661 se = pick_next_entity(cfs_rq);
3662 set_next_entity(cfs_rq, se);
3663 cfs_rq = group_cfs_rq(se);
3667 if (hrtick_enabled(rq))
3668 hrtick_start_fair(rq, p);
3674 * Account for a descheduled task:
3676 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3678 struct sched_entity *se = &prev->se;
3679 struct cfs_rq *cfs_rq;
3681 for_each_sched_entity(se) {
3682 cfs_rq = cfs_rq_of(se);
3683 put_prev_entity(cfs_rq, se);
3688 * sched_yield() is very simple
3690 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3692 static void yield_task_fair(struct rq *rq)
3694 struct task_struct *curr = rq->curr;
3695 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3696 struct sched_entity *se = &curr->se;
3699 * Are we the only task in the tree?
3701 if (unlikely(rq->nr_running == 1))
3704 clear_buddies(cfs_rq, se);
3706 if (curr->policy != SCHED_BATCH) {
3707 update_rq_clock(rq);
3709 * Update run-time statistics of the 'current'.
3711 update_curr(cfs_rq);
3713 * Tell update_rq_clock() that we've just updated,
3714 * so we don't do microscopic update in schedule()
3715 * and double the fastpath cost.
3717 rq->skip_clock_update = 1;
3723 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3725 struct sched_entity *se = &p->se;
3727 /* throttled hierarchies are not runnable */
3728 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3731 /* Tell the scheduler that we'd really like pse to run next. */
3734 yield_task_fair(rq);
3740 /**************************************************
3741 * Fair scheduling class load-balancing methods.
3745 * The purpose of load-balancing is to achieve the same basic fairness the
3746 * per-cpu scheduler provides, namely provide a proportional amount of compute
3747 * time to each task. This is expressed in the following equation:
3749 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3751 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3752 * W_i,0 is defined as:
3754 * W_i,0 = \Sum_j w_i,j (2)
3756 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3757 * is derived from the nice value as per prio_to_weight[].
3759 * The weight average is an exponential decay average of the instantaneous
3762 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3764 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3765 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3766 * can also include other factors [XXX].
3768 * To achieve this balance we define a measure of imbalance which follows
3769 * directly from (1):
3771 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3773 * We them move tasks around to minimize the imbalance. In the continuous
3774 * function space it is obvious this converges, in the discrete case we get
3775 * a few fun cases generally called infeasible weight scenarios.
3778 * - infeasible weights;
3779 * - local vs global optima in the discrete case. ]
3784 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3785 * for all i,j solution, we create a tree of cpus that follows the hardware
3786 * topology where each level pairs two lower groups (or better). This results
3787 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3788 * tree to only the first of the previous level and we decrease the frequency
3789 * of load-balance at each level inv. proportional to the number of cpus in
3795 * \Sum { --- * --- * 2^i } = O(n) (5)
3797 * `- size of each group
3798 * | | `- number of cpus doing load-balance
3800 * `- sum over all levels
3802 * Coupled with a limit on how many tasks we can migrate every balance pass,
3803 * this makes (5) the runtime complexity of the balancer.
3805 * An important property here is that each CPU is still (indirectly) connected
3806 * to every other cpu in at most O(log n) steps:
3808 * The adjacency matrix of the resulting graph is given by:
3811 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3814 * And you'll find that:
3816 * A^(log_2 n)_i,j != 0 for all i,j (7)
3818 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3819 * The task movement gives a factor of O(m), giving a convergence complexity
3822 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3827 * In order to avoid CPUs going idle while there's still work to do, new idle
3828 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3829 * tree itself instead of relying on other CPUs to bring it work.
3831 * This adds some complexity to both (5) and (8) but it reduces the total idle
3839 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3842 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3847 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3849 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3851 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3854 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3855 * rewrite all of this once again.]
3858 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3860 #define LBF_ALL_PINNED 0x01
3861 #define LBF_NEED_BREAK 0x02
3862 #define LBF_SOME_PINNED 0x04
3865 struct sched_domain *sd;
3873 struct cpumask *dst_grpmask;
3875 enum cpu_idle_type idle;
3877 /* The set of CPUs under consideration for load-balancing */
3878 struct cpumask *cpus;
3883 unsigned int loop_break;
3884 unsigned int loop_max;
3888 * move_task - move a task from one runqueue to another runqueue.
3889 * Both runqueues must be locked.
3891 static void move_task(struct task_struct *p, struct lb_env *env)
3893 deactivate_task(env->src_rq, p, 0);
3894 set_task_cpu(p, env->dst_cpu);
3895 activate_task(env->dst_rq, p, 0);
3896 check_preempt_curr(env->dst_rq, p, 0);
3900 * Is this task likely cache-hot:
3903 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3907 if (p->sched_class != &fair_sched_class)
3910 if (unlikely(p->policy == SCHED_IDLE))
3914 * Buddy candidates are cache hot:
3916 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3917 (&p->se == cfs_rq_of(&p->se)->next ||
3918 &p->se == cfs_rq_of(&p->se)->last))
3921 if (sysctl_sched_migration_cost == -1)
3923 if (sysctl_sched_migration_cost == 0)
3926 delta = now - p->se.exec_start;
3928 return delta < (s64)sysctl_sched_migration_cost;
3932 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3935 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3937 int tsk_cache_hot = 0;
3939 * We do not migrate tasks that are:
3940 * 1) throttled_lb_pair, or
3941 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3942 * 3) running (obviously), or
3943 * 4) are cache-hot on their current CPU.
3945 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3948 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3951 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3954 * Remember if this task can be migrated to any other cpu in
3955 * our sched_group. We may want to revisit it if we couldn't
3956 * meet load balance goals by pulling other tasks on src_cpu.
3958 * Also avoid computing new_dst_cpu if we have already computed
3959 * one in current iteration.
3961 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3964 /* Prevent to re-select dst_cpu via env's cpus */
3965 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
3966 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
3967 env->flags |= LBF_SOME_PINNED;
3968 env->new_dst_cpu = cpu;
3976 /* Record that we found atleast one task that could run on dst_cpu */
3977 env->flags &= ~LBF_ALL_PINNED;
3979 if (task_running(env->src_rq, p)) {
3980 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3985 * Aggressive migration if:
3986 * 1) task is cache cold, or
3987 * 2) too many balance attempts have failed.
3990 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
3991 if (!tsk_cache_hot ||
3992 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3994 if (tsk_cache_hot) {
3995 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3996 schedstat_inc(p, se.statistics.nr_forced_migrations);
4002 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4007 * move_one_task tries to move exactly one task from busiest to this_rq, as
4008 * part of active balancing operations within "domain".
4009 * Returns 1 if successful and 0 otherwise.
4011 * Called with both runqueues locked.
4013 static int move_one_task(struct lb_env *env)
4015 struct task_struct *p, *n;
4017 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4018 if (!can_migrate_task(p, env))
4023 * Right now, this is only the second place move_task()
4024 * is called, so we can safely collect move_task()
4025 * stats here rather than inside move_task().
4027 schedstat_inc(env->sd, lb_gained[env->idle]);
4033 static unsigned long task_h_load(struct task_struct *p);
4035 static const unsigned int sched_nr_migrate_break = 32;
4038 * move_tasks tries to move up to imbalance weighted load from busiest to
4039 * this_rq, as part of a balancing operation within domain "sd".
4040 * Returns 1 if successful and 0 otherwise.
4042 * Called with both runqueues locked.
4044 static int move_tasks(struct lb_env *env)
4046 struct list_head *tasks = &env->src_rq->cfs_tasks;
4047 struct task_struct *p;
4051 if (env->imbalance <= 0)
4054 while (!list_empty(tasks)) {
4055 p = list_first_entry(tasks, struct task_struct, se.group_node);
4058 /* We've more or less seen every task there is, call it quits */
4059 if (env->loop > env->loop_max)
4062 /* take a breather every nr_migrate tasks */
4063 if (env->loop > env->loop_break) {
4064 env->loop_break += sched_nr_migrate_break;
4065 env->flags |= LBF_NEED_BREAK;
4069 if (!can_migrate_task(p, env))
4072 load = task_h_load(p);
4074 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4077 if ((load / 2) > env->imbalance)
4082 env->imbalance -= load;
4084 #ifdef CONFIG_PREEMPT
4086 * NEWIDLE balancing is a source of latency, so preemptible
4087 * kernels will stop after the first task is pulled to minimize
4088 * the critical section.
4090 if (env->idle == CPU_NEWLY_IDLE)
4095 * We only want to steal up to the prescribed amount of
4098 if (env->imbalance <= 0)
4103 list_move_tail(&p->se.group_node, tasks);
4107 * Right now, this is one of only two places move_task() is called,
4108 * so we can safely collect move_task() stats here rather than
4109 * inside move_task().
4111 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4116 #ifdef CONFIG_FAIR_GROUP_SCHED
4118 * update tg->load_weight by folding this cpu's load_avg
4120 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4122 struct sched_entity *se = tg->se[cpu];
4123 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4125 /* throttled entities do not contribute to load */
4126 if (throttled_hierarchy(cfs_rq))
4129 update_cfs_rq_blocked_load(cfs_rq, 1);
4132 update_entity_load_avg(se, 1);
4134 * We pivot on our runnable average having decayed to zero for
4135 * list removal. This generally implies that all our children
4136 * have also been removed (modulo rounding error or bandwidth
4137 * control); however, such cases are rare and we can fix these
4140 * TODO: fix up out-of-order children on enqueue.
4142 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4143 list_del_leaf_cfs_rq(cfs_rq);
4145 struct rq *rq = rq_of(cfs_rq);
4146 update_rq_runnable_avg(rq, rq->nr_running);
4150 static void update_blocked_averages(int cpu)
4152 struct rq *rq = cpu_rq(cpu);
4153 struct cfs_rq *cfs_rq;
4154 unsigned long flags;
4156 raw_spin_lock_irqsave(&rq->lock, flags);
4157 update_rq_clock(rq);
4159 * Iterates the task_group tree in a bottom up fashion, see
4160 * list_add_leaf_cfs_rq() for details.
4162 for_each_leaf_cfs_rq(rq, cfs_rq) {
4164 * Note: We may want to consider periodically releasing
4165 * rq->lock about these updates so that creating many task
4166 * groups does not result in continually extending hold time.
4168 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4171 raw_spin_unlock_irqrestore(&rq->lock, flags);
4175 * Compute the cpu's hierarchical load factor for each task group.
4176 * This needs to be done in a top-down fashion because the load of a child
4177 * group is a fraction of its parents load.
4179 static int tg_load_down(struct task_group *tg, void *data)
4182 long cpu = (long)data;
4185 load = cpu_rq(cpu)->avg.load_avg_contrib;
4187 load = tg->parent->cfs_rq[cpu]->h_load;
4188 load = div64_ul(load * tg->se[cpu]->avg.load_avg_contrib,
4189 tg->parent->cfs_rq[cpu]->runnable_load_avg + 1);
4192 tg->cfs_rq[cpu]->h_load = load;
4197 static void update_h_load(long cpu)
4199 struct rq *rq = cpu_rq(cpu);
4200 unsigned long now = jiffies;
4202 if (rq->h_load_throttle == now)
4205 rq->h_load_throttle = now;
4208 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
4212 static unsigned long task_h_load(struct task_struct *p)
4214 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4216 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4217 cfs_rq->runnable_load_avg + 1);
4220 static inline void update_blocked_averages(int cpu)
4224 static inline void update_h_load(long cpu)
4228 static unsigned long task_h_load(struct task_struct *p)
4230 return p->se.avg.load_avg_contrib;
4234 /********** Helpers for find_busiest_group ************************/
4236 * sd_lb_stats - Structure to store the statistics of a sched_domain
4237 * during load balancing.
4239 struct sd_lb_stats {
4240 struct sched_group *busiest; /* Busiest group in this sd */
4241 struct sched_group *this; /* Local group in this sd */
4242 unsigned long total_load; /* Total load of all groups in sd */
4243 unsigned long total_pwr; /* Total power of all groups in sd */
4244 unsigned long avg_load; /* Average load across all groups in sd */
4246 /** Statistics of this group */
4247 unsigned long this_load;
4248 unsigned long this_load_per_task;
4249 unsigned long this_nr_running;
4250 unsigned long this_has_capacity;
4251 unsigned int this_idle_cpus;
4253 /* Statistics of the busiest group */
4254 unsigned int busiest_idle_cpus;
4255 unsigned long max_load;
4256 unsigned long busiest_load_per_task;
4257 unsigned long busiest_nr_running;
4258 unsigned long busiest_group_capacity;
4259 unsigned long busiest_has_capacity;
4260 unsigned int busiest_group_weight;
4262 int group_imb; /* Is there imbalance in this sd */
4266 * sg_lb_stats - stats of a sched_group required for load_balancing
4268 struct sg_lb_stats {
4269 unsigned long avg_load; /*Avg load across the CPUs of the group */
4270 unsigned long group_load; /* Total load over the CPUs of the group */
4271 unsigned long sum_nr_running; /* Nr tasks running in the group */
4272 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4273 unsigned long group_capacity;
4274 unsigned long idle_cpus;
4275 unsigned long group_weight;
4276 int group_imb; /* Is there an imbalance in the group ? */
4277 int group_has_capacity; /* Is there extra capacity in the group? */
4281 * get_sd_load_idx - Obtain the load index for a given sched domain.
4282 * @sd: The sched_domain whose load_idx is to be obtained.
4283 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4285 * Return: The load index.
4287 static inline int get_sd_load_idx(struct sched_domain *sd,
4288 enum cpu_idle_type idle)
4294 load_idx = sd->busy_idx;
4297 case CPU_NEWLY_IDLE:
4298 load_idx = sd->newidle_idx;
4301 load_idx = sd->idle_idx;
4308 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4310 return SCHED_POWER_SCALE;
4313 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4315 return default_scale_freq_power(sd, cpu);
4318 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4320 unsigned long weight = sd->span_weight;
4321 unsigned long smt_gain = sd->smt_gain;
4328 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4330 return default_scale_smt_power(sd, cpu);
4333 static unsigned long scale_rt_power(int cpu)
4335 struct rq *rq = cpu_rq(cpu);
4336 u64 total, available, age_stamp, avg;
4339 * Since we're reading these variables without serialization make sure
4340 * we read them once before doing sanity checks on them.
4342 age_stamp = ACCESS_ONCE(rq->age_stamp);
4343 avg = ACCESS_ONCE(rq->rt_avg);
4345 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4347 if (unlikely(total < avg)) {
4348 /* Ensures that power won't end up being negative */
4351 available = total - avg;
4354 if (unlikely((s64)total < SCHED_POWER_SCALE))
4355 total = SCHED_POWER_SCALE;
4357 total >>= SCHED_POWER_SHIFT;
4359 return div_u64(available, total);
4362 static void update_cpu_power(struct sched_domain *sd, int cpu)
4364 unsigned long weight = sd->span_weight;
4365 unsigned long power = SCHED_POWER_SCALE;
4366 struct sched_group *sdg = sd->groups;
4368 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4369 if (sched_feat(ARCH_POWER))
4370 power *= arch_scale_smt_power(sd, cpu);
4372 power *= default_scale_smt_power(sd, cpu);
4374 power >>= SCHED_POWER_SHIFT;
4377 sdg->sgp->power_orig = power;
4379 if (sched_feat(ARCH_POWER))
4380 power *= arch_scale_freq_power(sd, cpu);
4382 power *= default_scale_freq_power(sd, cpu);
4384 power >>= SCHED_POWER_SHIFT;
4386 power *= scale_rt_power(cpu);
4387 power >>= SCHED_POWER_SHIFT;
4392 cpu_rq(cpu)->cpu_power = power;
4393 sdg->sgp->power = power;
4396 void update_group_power(struct sched_domain *sd, int cpu)
4398 struct sched_domain *child = sd->child;
4399 struct sched_group *group, *sdg = sd->groups;
4400 unsigned long power;
4401 unsigned long interval;
4403 interval = msecs_to_jiffies(sd->balance_interval);
4404 interval = clamp(interval, 1UL, max_load_balance_interval);
4405 sdg->sgp->next_update = jiffies + interval;
4408 update_cpu_power(sd, cpu);
4414 if (child->flags & SD_OVERLAP) {
4416 * SD_OVERLAP domains cannot assume that child groups
4417 * span the current group.
4420 for_each_cpu(cpu, sched_group_cpus(sdg))
4421 power += power_of(cpu);
4424 * !SD_OVERLAP domains can assume that child groups
4425 * span the current group.
4428 group = child->groups;
4430 power += group->sgp->power;
4431 group = group->next;
4432 } while (group != child->groups);
4435 sdg->sgp->power_orig = sdg->sgp->power = power;
4439 * Try and fix up capacity for tiny siblings, this is needed when
4440 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4441 * which on its own isn't powerful enough.
4443 * See update_sd_pick_busiest() and check_asym_packing().
4446 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4449 * Only siblings can have significantly less than SCHED_POWER_SCALE
4451 if (!(sd->flags & SD_SHARE_CPUPOWER))
4455 * If ~90% of the cpu_power is still there, we're good.
4457 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4464 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4465 * @env: The load balancing environment.
4466 * @group: sched_group whose statistics are to be updated.
4467 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4468 * @local_group: Does group contain this_cpu.
4469 * @balance: Should we balance.
4470 * @sgs: variable to hold the statistics for this group.
4472 static inline void update_sg_lb_stats(struct lb_env *env,
4473 struct sched_group *group, int load_idx,
4474 int local_group, int *balance, struct sg_lb_stats *sgs)
4476 unsigned long nr_running, max_nr_running, min_nr_running;
4477 unsigned long load, max_cpu_load, min_cpu_load;
4478 unsigned int balance_cpu = -1, first_idle_cpu = 0;
4479 unsigned long avg_load_per_task = 0;
4483 balance_cpu = group_balance_cpu(group);
4485 /* Tally up the load of all CPUs in the group */
4487 min_cpu_load = ~0UL;
4489 min_nr_running = ~0UL;
4491 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4492 struct rq *rq = cpu_rq(i);
4494 nr_running = rq->nr_running;
4496 /* Bias balancing toward cpus of our domain */
4498 if (idle_cpu(i) && !first_idle_cpu &&
4499 cpumask_test_cpu(i, sched_group_mask(group))) {
4504 load = target_load(i, load_idx);
4506 load = source_load(i, load_idx);
4507 if (load > max_cpu_load)
4508 max_cpu_load = load;
4509 if (min_cpu_load > load)
4510 min_cpu_load = load;
4512 if (nr_running > max_nr_running)
4513 max_nr_running = nr_running;
4514 if (min_nr_running > nr_running)
4515 min_nr_running = nr_running;
4518 sgs->group_load += load;
4519 sgs->sum_nr_running += nr_running;
4520 sgs->sum_weighted_load += weighted_cpuload(i);
4526 * First idle cpu or the first cpu(busiest) in this sched group
4527 * is eligible for doing load balancing at this and above
4528 * domains. In the newly idle case, we will allow all the cpu's
4529 * to do the newly idle load balance.
4532 if (env->idle != CPU_NEWLY_IDLE) {
4533 if (balance_cpu != env->dst_cpu) {
4537 update_group_power(env->sd, env->dst_cpu);
4538 } else if (time_after_eq(jiffies, group->sgp->next_update))
4539 update_group_power(env->sd, env->dst_cpu);
4542 /* Adjust by relative CPU power of the group */
4543 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
4546 * Consider the group unbalanced when the imbalance is larger
4547 * than the average weight of a task.
4549 * APZ: with cgroup the avg task weight can vary wildly and
4550 * might not be a suitable number - should we keep a
4551 * normalized nr_running number somewhere that negates
4554 if (sgs->sum_nr_running)
4555 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4557 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
4558 (max_nr_running - min_nr_running) > 1)
4561 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
4563 if (!sgs->group_capacity)
4564 sgs->group_capacity = fix_small_capacity(env->sd, group);
4565 sgs->group_weight = group->group_weight;
4567 if (sgs->group_capacity > sgs->sum_nr_running)
4568 sgs->group_has_capacity = 1;
4572 * update_sd_pick_busiest - return 1 on busiest group
4573 * @env: The load balancing environment.
4574 * @sds: sched_domain statistics
4575 * @sg: sched_group candidate to be checked for being the busiest
4576 * @sgs: sched_group statistics
4578 * Determine if @sg is a busier group than the previously selected
4581 * Return: %true if @sg is a busier group than the previously selected
4582 * busiest group. %false otherwise.
4584 static bool update_sd_pick_busiest(struct lb_env *env,
4585 struct sd_lb_stats *sds,
4586 struct sched_group *sg,
4587 struct sg_lb_stats *sgs)
4589 if (sgs->avg_load <= sds->max_load)
4592 if (sgs->sum_nr_running > sgs->group_capacity)
4599 * ASYM_PACKING needs to move all the work to the lowest
4600 * numbered CPUs in the group, therefore mark all groups
4601 * higher than ourself as busy.
4603 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4604 env->dst_cpu < group_first_cpu(sg)) {
4608 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4616 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4617 * @env: The load balancing environment.
4618 * @balance: Should we balance.
4619 * @sds: variable to hold the statistics for this sched_domain.
4621 static inline void update_sd_lb_stats(struct lb_env *env,
4622 int *balance, struct sd_lb_stats *sds)
4624 struct sched_domain *child = env->sd->child;
4625 struct sched_group *sg = env->sd->groups;
4626 struct sg_lb_stats sgs;
4627 int load_idx, prefer_sibling = 0;
4629 if (child && child->flags & SD_PREFER_SIBLING)
4632 load_idx = get_sd_load_idx(env->sd, env->idle);
4637 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4638 memset(&sgs, 0, sizeof(sgs));
4639 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4641 if (local_group && !(*balance))
4644 sds->total_load += sgs.group_load;
4645 sds->total_pwr += sg->sgp->power;
4648 * In case the child domain prefers tasks go to siblings
4649 * first, lower the sg capacity to one so that we'll try
4650 * and move all the excess tasks away. We lower the capacity
4651 * of a group only if the local group has the capacity to fit
4652 * these excess tasks, i.e. nr_running < group_capacity. The
4653 * extra check prevents the case where you always pull from the
4654 * heaviest group when it is already under-utilized (possible
4655 * with a large weight task outweighs the tasks on the system).
4657 if (prefer_sibling && !local_group && sds->this_has_capacity)
4658 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4661 sds->this_load = sgs.avg_load;
4663 sds->this_nr_running = sgs.sum_nr_running;
4664 sds->this_load_per_task = sgs.sum_weighted_load;
4665 sds->this_has_capacity = sgs.group_has_capacity;
4666 sds->this_idle_cpus = sgs.idle_cpus;
4667 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4668 sds->max_load = sgs.avg_load;
4670 sds->busiest_nr_running = sgs.sum_nr_running;
4671 sds->busiest_idle_cpus = sgs.idle_cpus;
4672 sds->busiest_group_capacity = sgs.group_capacity;
4673 sds->busiest_load_per_task = sgs.sum_weighted_load;
4674 sds->busiest_has_capacity = sgs.group_has_capacity;
4675 sds->busiest_group_weight = sgs.group_weight;
4676 sds->group_imb = sgs.group_imb;
4680 } while (sg != env->sd->groups);
4684 * check_asym_packing - Check to see if the group is packed into the
4687 * This is primarily intended to used at the sibling level. Some
4688 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4689 * case of POWER7, it can move to lower SMT modes only when higher
4690 * threads are idle. When in lower SMT modes, the threads will
4691 * perform better since they share less core resources. Hence when we
4692 * have idle threads, we want them to be the higher ones.
4694 * This packing function is run on idle threads. It checks to see if
4695 * the busiest CPU in this domain (core in the P7 case) has a higher
4696 * CPU number than the packing function is being run on. Here we are
4697 * assuming lower CPU number will be equivalent to lower a SMT thread
4700 * Return: 1 when packing is required and a task should be moved to
4701 * this CPU. The amount of the imbalance is returned in *imbalance.
4703 * @env: The load balancing environment.
4704 * @sds: Statistics of the sched_domain which is to be packed
4706 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4710 if (!(env->sd->flags & SD_ASYM_PACKING))
4716 busiest_cpu = group_first_cpu(sds->busiest);
4717 if (env->dst_cpu > busiest_cpu)
4720 env->imbalance = DIV_ROUND_CLOSEST(
4721 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4727 * fix_small_imbalance - Calculate the minor imbalance that exists
4728 * amongst the groups of a sched_domain, during
4730 * @env: The load balancing environment.
4731 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4734 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4736 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4737 unsigned int imbn = 2;
4738 unsigned long scaled_busy_load_per_task;
4740 if (sds->this_nr_running) {
4741 sds->this_load_per_task /= sds->this_nr_running;
4742 if (sds->busiest_load_per_task >
4743 sds->this_load_per_task)
4746 sds->this_load_per_task =
4747 cpu_avg_load_per_task(env->dst_cpu);
4750 scaled_busy_load_per_task = sds->busiest_load_per_task
4751 * SCHED_POWER_SCALE;
4752 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4754 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4755 (scaled_busy_load_per_task * imbn)) {
4756 env->imbalance = sds->busiest_load_per_task;
4761 * OK, we don't have enough imbalance to justify moving tasks,
4762 * however we may be able to increase total CPU power used by
4766 pwr_now += sds->busiest->sgp->power *
4767 min(sds->busiest_load_per_task, sds->max_load);
4768 pwr_now += sds->this->sgp->power *
4769 min(sds->this_load_per_task, sds->this_load);
4770 pwr_now /= SCHED_POWER_SCALE;
4772 /* Amount of load we'd subtract */
4773 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4774 sds->busiest->sgp->power;
4775 if (sds->max_load > tmp)
4776 pwr_move += sds->busiest->sgp->power *
4777 min(sds->busiest_load_per_task, sds->max_load - tmp);
4779 /* Amount of load we'd add */
4780 if (sds->max_load * sds->busiest->sgp->power <
4781 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4782 tmp = (sds->max_load * sds->busiest->sgp->power) /
4783 sds->this->sgp->power;
4785 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4786 sds->this->sgp->power;
4787 pwr_move += sds->this->sgp->power *
4788 min(sds->this_load_per_task, sds->this_load + tmp);
4789 pwr_move /= SCHED_POWER_SCALE;
4791 /* Move if we gain throughput */
4792 if (pwr_move > pwr_now)
4793 env->imbalance = sds->busiest_load_per_task;
4797 * calculate_imbalance - Calculate the amount of imbalance present within the
4798 * groups of a given sched_domain during load balance.
4799 * @env: load balance environment
4800 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4802 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4804 unsigned long max_pull, load_above_capacity = ~0UL;
4806 sds->busiest_load_per_task /= sds->busiest_nr_running;
4807 if (sds->group_imb) {
4808 sds->busiest_load_per_task =
4809 min(sds->busiest_load_per_task, sds->avg_load);
4813 * In the presence of smp nice balancing, certain scenarios can have
4814 * max load less than avg load(as we skip the groups at or below
4815 * its cpu_power, while calculating max_load..)
4817 if (sds->max_load < sds->avg_load) {
4819 return fix_small_imbalance(env, sds);
4822 if (!sds->group_imb) {
4824 * Don't want to pull so many tasks that a group would go idle.
4826 load_above_capacity = (sds->busiest_nr_running -
4827 sds->busiest_group_capacity);
4829 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4831 load_above_capacity /= sds->busiest->sgp->power;
4835 * We're trying to get all the cpus to the average_load, so we don't
4836 * want to push ourselves above the average load, nor do we wish to
4837 * reduce the max loaded cpu below the average load. At the same time,
4838 * we also don't want to reduce the group load below the group capacity
4839 * (so that we can implement power-savings policies etc). Thus we look
4840 * for the minimum possible imbalance.
4841 * Be careful of negative numbers as they'll appear as very large values
4842 * with unsigned longs.
4844 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4846 /* How much load to actually move to equalise the imbalance */
4847 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4848 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4849 / SCHED_POWER_SCALE;
4852 * if *imbalance is less than the average load per runnable task
4853 * there is no guarantee that any tasks will be moved so we'll have
4854 * a think about bumping its value to force at least one task to be
4857 if (env->imbalance < sds->busiest_load_per_task)
4858 return fix_small_imbalance(env, sds);
4862 /******* find_busiest_group() helpers end here *********************/
4865 * find_busiest_group - Returns the busiest group within the sched_domain
4866 * if there is an imbalance. If there isn't an imbalance, and
4867 * the user has opted for power-savings, it returns a group whose
4868 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4869 * such a group exists.
4871 * Also calculates the amount of weighted load which should be moved
4872 * to restore balance.
4874 * @env: The load balancing environment.
4875 * @balance: Pointer to a variable indicating if this_cpu
4876 * is the appropriate cpu to perform load balancing at this_level.
4878 * Return: - The busiest group if imbalance exists.
4879 * - If no imbalance and user has opted for power-savings balance,
4880 * return the least loaded group whose CPUs can be
4881 * put to idle by rebalancing its tasks onto our group.
4883 static struct sched_group *
4884 find_busiest_group(struct lb_env *env, int *balance)
4886 struct sd_lb_stats sds;
4888 memset(&sds, 0, sizeof(sds));
4891 * Compute the various statistics relavent for load balancing at
4894 update_sd_lb_stats(env, balance, &sds);
4897 * this_cpu is not the appropriate cpu to perform load balancing at
4903 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4904 check_asym_packing(env, &sds))
4907 /* There is no busy sibling group to pull tasks from */
4908 if (!sds.busiest || sds.busiest_nr_running == 0)
4911 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4914 * If the busiest group is imbalanced the below checks don't
4915 * work because they assumes all things are equal, which typically
4916 * isn't true due to cpus_allowed constraints and the like.
4921 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4922 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4923 !sds.busiest_has_capacity)
4927 * If the local group is more busy than the selected busiest group
4928 * don't try and pull any tasks.
4930 if (sds.this_load >= sds.max_load)
4934 * Don't pull any tasks if this group is already above the domain
4937 if (sds.this_load >= sds.avg_load)
4940 if (env->idle == CPU_IDLE) {
4942 * This cpu is idle. If the busiest group load doesn't
4943 * have more tasks than the number of available cpu's and
4944 * there is no imbalance between this and busiest group
4945 * wrt to idle cpu's, it is balanced.
4947 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4948 sds.busiest_nr_running <= sds.busiest_group_weight)
4952 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4953 * imbalance_pct to be conservative.
4955 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4960 /* Looks like there is an imbalance. Compute it */
4961 calculate_imbalance(env, &sds);
4971 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4973 static struct rq *find_busiest_queue(struct lb_env *env,
4974 struct sched_group *group)
4976 struct rq *busiest = NULL, *rq;
4977 unsigned long max_load = 0;
4980 for_each_cpu(i, sched_group_cpus(group)) {
4981 unsigned long power = power_of(i);
4982 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4987 capacity = fix_small_capacity(env->sd, group);
4989 if (!cpumask_test_cpu(i, env->cpus))
4993 wl = weighted_cpuload(i);
4996 * When comparing with imbalance, use weighted_cpuload()
4997 * which is not scaled with the cpu power.
4999 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5003 * For the load comparisons with the other cpu's, consider
5004 * the weighted_cpuload() scaled with the cpu power, so that
5005 * the load can be moved away from the cpu that is potentially
5006 * running at a lower capacity.
5008 wl = (wl * SCHED_POWER_SCALE) / power;
5010 if (wl > max_load) {
5020 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5021 * so long as it is large enough.
5023 #define MAX_PINNED_INTERVAL 512
5025 /* Working cpumask for load_balance and load_balance_newidle. */
5026 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5028 static int need_active_balance(struct lb_env *env)
5030 struct sched_domain *sd = env->sd;
5032 if (env->idle == CPU_NEWLY_IDLE) {
5035 * ASYM_PACKING needs to force migrate tasks from busy but
5036 * higher numbered CPUs in order to pack all tasks in the
5037 * lowest numbered CPUs.
5039 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5043 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5046 static int active_load_balance_cpu_stop(void *data);
5049 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5050 * tasks if there is an imbalance.
5052 static int load_balance(int this_cpu, struct rq *this_rq,
5053 struct sched_domain *sd, enum cpu_idle_type idle,
5056 int ld_moved, cur_ld_moved, active_balance = 0;
5057 struct sched_group *group;
5059 unsigned long flags;
5060 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5062 struct lb_env env = {
5064 .dst_cpu = this_cpu,
5066 .dst_grpmask = sched_group_cpus(sd->groups),
5068 .loop_break = sched_nr_migrate_break,
5073 * For NEWLY_IDLE load_balancing, we don't need to consider
5074 * other cpus in our group
5076 if (idle == CPU_NEWLY_IDLE)
5077 env.dst_grpmask = NULL;
5079 cpumask_copy(cpus, cpu_active_mask);
5081 schedstat_inc(sd, lb_count[idle]);
5084 group = find_busiest_group(&env, balance);
5090 schedstat_inc(sd, lb_nobusyg[idle]);
5094 busiest = find_busiest_queue(&env, group);
5096 schedstat_inc(sd, lb_nobusyq[idle]);
5100 BUG_ON(busiest == env.dst_rq);
5102 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5105 if (busiest->nr_running > 1) {
5107 * Attempt to move tasks. If find_busiest_group has found
5108 * an imbalance but busiest->nr_running <= 1, the group is
5109 * still unbalanced. ld_moved simply stays zero, so it is
5110 * correctly treated as an imbalance.
5112 env.flags |= LBF_ALL_PINNED;
5113 env.src_cpu = busiest->cpu;
5114 env.src_rq = busiest;
5115 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5117 update_h_load(env.src_cpu);
5119 local_irq_save(flags);
5120 double_rq_lock(env.dst_rq, busiest);
5123 * cur_ld_moved - load moved in current iteration
5124 * ld_moved - cumulative load moved across iterations
5126 cur_ld_moved = move_tasks(&env);
5127 ld_moved += cur_ld_moved;
5128 double_rq_unlock(env.dst_rq, busiest);
5129 local_irq_restore(flags);
5132 * some other cpu did the load balance for us.
5134 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5135 resched_cpu(env.dst_cpu);
5137 if (env.flags & LBF_NEED_BREAK) {
5138 env.flags &= ~LBF_NEED_BREAK;
5143 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5144 * us and move them to an alternate dst_cpu in our sched_group
5145 * where they can run. The upper limit on how many times we
5146 * iterate on same src_cpu is dependent on number of cpus in our
5149 * This changes load balance semantics a bit on who can move
5150 * load to a given_cpu. In addition to the given_cpu itself
5151 * (or a ilb_cpu acting on its behalf where given_cpu is
5152 * nohz-idle), we now have balance_cpu in a position to move
5153 * load to given_cpu. In rare situations, this may cause
5154 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5155 * _independently_ and at _same_ time to move some load to
5156 * given_cpu) causing exceess load to be moved to given_cpu.
5157 * This however should not happen so much in practice and
5158 * moreover subsequent load balance cycles should correct the
5159 * excess load moved.
5161 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5163 env.dst_rq = cpu_rq(env.new_dst_cpu);
5164 env.dst_cpu = env.new_dst_cpu;
5165 env.flags &= ~LBF_SOME_PINNED;
5167 env.loop_break = sched_nr_migrate_break;
5169 /* Prevent to re-select dst_cpu via env's cpus */
5170 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5173 * Go back to "more_balance" rather than "redo" since we
5174 * need to continue with same src_cpu.
5179 /* All tasks on this runqueue were pinned by CPU affinity */
5180 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5181 cpumask_clear_cpu(cpu_of(busiest), cpus);
5182 if (!cpumask_empty(cpus)) {
5184 env.loop_break = sched_nr_migrate_break;
5192 schedstat_inc(sd, lb_failed[idle]);
5194 * Increment the failure counter only on periodic balance.
5195 * We do not want newidle balance, which can be very
5196 * frequent, pollute the failure counter causing
5197 * excessive cache_hot migrations and active balances.
5199 if (idle != CPU_NEWLY_IDLE)
5200 sd->nr_balance_failed++;
5202 if (need_active_balance(&env)) {
5203 raw_spin_lock_irqsave(&busiest->lock, flags);
5205 /* don't kick the active_load_balance_cpu_stop,
5206 * if the curr task on busiest cpu can't be
5209 if (!cpumask_test_cpu(this_cpu,
5210 tsk_cpus_allowed(busiest->curr))) {
5211 raw_spin_unlock_irqrestore(&busiest->lock,
5213 env.flags |= LBF_ALL_PINNED;
5214 goto out_one_pinned;
5218 * ->active_balance synchronizes accesses to
5219 * ->active_balance_work. Once set, it's cleared
5220 * only after active load balance is finished.
5222 if (!busiest->active_balance) {
5223 busiest->active_balance = 1;
5224 busiest->push_cpu = this_cpu;
5227 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5229 if (active_balance) {
5230 stop_one_cpu_nowait(cpu_of(busiest),
5231 active_load_balance_cpu_stop, busiest,
5232 &busiest->active_balance_work);
5236 * We've kicked active balancing, reset the failure
5239 sd->nr_balance_failed = sd->cache_nice_tries+1;
5242 sd->nr_balance_failed = 0;
5244 if (likely(!active_balance)) {
5245 /* We were unbalanced, so reset the balancing interval */
5246 sd->balance_interval = sd->min_interval;
5249 * If we've begun active balancing, start to back off. This
5250 * case may not be covered by the all_pinned logic if there
5251 * is only 1 task on the busy runqueue (because we don't call
5254 if (sd->balance_interval < sd->max_interval)
5255 sd->balance_interval *= 2;
5261 schedstat_inc(sd, lb_balanced[idle]);
5263 sd->nr_balance_failed = 0;
5266 /* tune up the balancing interval */
5267 if (((env.flags & LBF_ALL_PINNED) &&
5268 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5269 (sd->balance_interval < sd->max_interval))
5270 sd->balance_interval *= 2;
5278 * idle_balance is called by schedule() if this_cpu is about to become
5279 * idle. Attempts to pull tasks from other CPUs.
5281 void idle_balance(int this_cpu, struct rq *this_rq)
5283 struct sched_domain *sd;
5284 int pulled_task = 0;
5285 unsigned long next_balance = jiffies + HZ;
5287 this_rq->idle_stamp = rq_clock(this_rq);
5289 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5293 * Drop the rq->lock, but keep IRQ/preempt disabled.
5295 raw_spin_unlock(&this_rq->lock);
5297 update_blocked_averages(this_cpu);
5299 for_each_domain(this_cpu, sd) {
5300 unsigned long interval;
5303 if (!(sd->flags & SD_LOAD_BALANCE))
5306 if (sd->flags & SD_BALANCE_NEWIDLE) {
5307 /* If we've pulled tasks over stop searching: */
5308 pulled_task = load_balance(this_cpu, this_rq,
5309 sd, CPU_NEWLY_IDLE, &balance);
5312 interval = msecs_to_jiffies(sd->balance_interval);
5313 if (time_after(next_balance, sd->last_balance + interval))
5314 next_balance = sd->last_balance + interval;
5316 this_rq->idle_stamp = 0;
5322 raw_spin_lock(&this_rq->lock);
5324 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5326 * We are going idle. next_balance may be set based on
5327 * a busy processor. So reset next_balance.
5329 this_rq->next_balance = next_balance;
5334 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5335 * running tasks off the busiest CPU onto idle CPUs. It requires at
5336 * least 1 task to be running on each physical CPU where possible, and
5337 * avoids physical / logical imbalances.
5339 static int active_load_balance_cpu_stop(void *data)
5341 struct rq *busiest_rq = data;
5342 int busiest_cpu = cpu_of(busiest_rq);
5343 int target_cpu = busiest_rq->push_cpu;
5344 struct rq *target_rq = cpu_rq(target_cpu);
5345 struct sched_domain *sd;
5347 raw_spin_lock_irq(&busiest_rq->lock);
5349 /* make sure the requested cpu hasn't gone down in the meantime */
5350 if (unlikely(busiest_cpu != smp_processor_id() ||
5351 !busiest_rq->active_balance))
5354 /* Is there any task to move? */
5355 if (busiest_rq->nr_running <= 1)
5359 * This condition is "impossible", if it occurs
5360 * we need to fix it. Originally reported by
5361 * Bjorn Helgaas on a 128-cpu setup.
5363 BUG_ON(busiest_rq == target_rq);
5365 /* move a task from busiest_rq to target_rq */
5366 double_lock_balance(busiest_rq, target_rq);
5368 /* Search for an sd spanning us and the target CPU. */
5370 for_each_domain(target_cpu, sd) {
5371 if ((sd->flags & SD_LOAD_BALANCE) &&
5372 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5377 struct lb_env env = {
5379 .dst_cpu = target_cpu,
5380 .dst_rq = target_rq,
5381 .src_cpu = busiest_rq->cpu,
5382 .src_rq = busiest_rq,
5386 schedstat_inc(sd, alb_count);
5388 if (move_one_task(&env))
5389 schedstat_inc(sd, alb_pushed);
5391 schedstat_inc(sd, alb_failed);
5394 double_unlock_balance(busiest_rq, target_rq);
5396 busiest_rq->active_balance = 0;
5397 raw_spin_unlock_irq(&busiest_rq->lock);
5401 #ifdef CONFIG_NO_HZ_COMMON
5403 * idle load balancing details
5404 * - When one of the busy CPUs notice that there may be an idle rebalancing
5405 * needed, they will kick the idle load balancer, which then does idle
5406 * load balancing for all the idle CPUs.
5409 cpumask_var_t idle_cpus_mask;
5411 unsigned long next_balance; /* in jiffy units */
5412 } nohz ____cacheline_aligned;
5414 static inline int find_new_ilb(int call_cpu)
5416 int ilb = cpumask_first(nohz.idle_cpus_mask);
5418 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5425 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5426 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5427 * CPU (if there is one).
5429 static void nohz_balancer_kick(int cpu)
5433 nohz.next_balance++;
5435 ilb_cpu = find_new_ilb(cpu);
5437 if (ilb_cpu >= nr_cpu_ids)
5440 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5443 * Use smp_send_reschedule() instead of resched_cpu().
5444 * This way we generate a sched IPI on the target cpu which
5445 * is idle. And the softirq performing nohz idle load balance
5446 * will be run before returning from the IPI.
5448 smp_send_reschedule(ilb_cpu);
5452 static inline void nohz_balance_exit_idle(int cpu)
5454 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5455 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5456 atomic_dec(&nohz.nr_cpus);
5457 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5461 static inline void set_cpu_sd_state_busy(void)
5463 struct sched_domain *sd;
5466 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5468 if (!sd || !sd->nohz_idle)
5472 for (; sd; sd = sd->parent)
5473 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5478 void set_cpu_sd_state_idle(void)
5480 struct sched_domain *sd;
5483 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5485 if (!sd || sd->nohz_idle)
5489 for (; sd; sd = sd->parent)
5490 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5496 * This routine will record that the cpu is going idle with tick stopped.
5497 * This info will be used in performing idle load balancing in the future.
5499 void nohz_balance_enter_idle(int cpu)
5502 * If this cpu is going down, then nothing needs to be done.
5504 if (!cpu_active(cpu))
5507 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5510 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5511 atomic_inc(&nohz.nr_cpus);
5512 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5515 static int sched_ilb_notifier(struct notifier_block *nfb,
5516 unsigned long action, void *hcpu)
5518 switch (action & ~CPU_TASKS_FROZEN) {
5520 nohz_balance_exit_idle(smp_processor_id());
5528 static DEFINE_SPINLOCK(balancing);
5531 * Scale the max load_balance interval with the number of CPUs in the system.
5532 * This trades load-balance latency on larger machines for less cross talk.
5534 void update_max_interval(void)
5536 max_load_balance_interval = HZ*num_online_cpus()/10;
5540 * It checks each scheduling domain to see if it is due to be balanced,
5541 * and initiates a balancing operation if so.
5543 * Balancing parameters are set up in init_sched_domains.
5545 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5548 struct rq *rq = cpu_rq(cpu);
5549 unsigned long interval;
5550 struct sched_domain *sd;
5551 /* Earliest time when we have to do rebalance again */
5552 unsigned long next_balance = jiffies + 60*HZ;
5553 int update_next_balance = 0;
5556 update_blocked_averages(cpu);
5559 for_each_domain(cpu, sd) {
5560 if (!(sd->flags & SD_LOAD_BALANCE))
5563 interval = sd->balance_interval;
5564 if (idle != CPU_IDLE)
5565 interval *= sd->busy_factor;
5567 /* scale ms to jiffies */
5568 interval = msecs_to_jiffies(interval);
5569 interval = clamp(interval, 1UL, max_load_balance_interval);
5571 need_serialize = sd->flags & SD_SERIALIZE;
5573 if (need_serialize) {
5574 if (!spin_trylock(&balancing))
5578 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5579 if (load_balance(cpu, rq, sd, idle, &balance)) {
5581 * The LBF_SOME_PINNED logic could have changed
5582 * env->dst_cpu, so we can't know our idle
5583 * state even if we migrated tasks. Update it.
5585 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5587 sd->last_balance = jiffies;
5590 spin_unlock(&balancing);
5592 if (time_after(next_balance, sd->last_balance + interval)) {
5593 next_balance = sd->last_balance + interval;
5594 update_next_balance = 1;
5598 * Stop the load balance at this level. There is another
5599 * CPU in our sched group which is doing load balancing more
5608 * next_balance will be updated only when there is a need.
5609 * When the cpu is attached to null domain for ex, it will not be
5612 if (likely(update_next_balance))
5613 rq->next_balance = next_balance;
5616 #ifdef CONFIG_NO_HZ_COMMON
5618 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5619 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5621 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5623 struct rq *this_rq = cpu_rq(this_cpu);
5627 if (idle != CPU_IDLE ||
5628 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5631 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5632 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5636 * If this cpu gets work to do, stop the load balancing
5637 * work being done for other cpus. Next load
5638 * balancing owner will pick it up.
5643 rq = cpu_rq(balance_cpu);
5645 raw_spin_lock_irq(&rq->lock);
5646 update_rq_clock(rq);
5647 update_idle_cpu_load(rq);
5648 raw_spin_unlock_irq(&rq->lock);
5650 rebalance_domains(balance_cpu, CPU_IDLE);
5652 if (time_after(this_rq->next_balance, rq->next_balance))
5653 this_rq->next_balance = rq->next_balance;
5655 nohz.next_balance = this_rq->next_balance;
5657 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5661 * Current heuristic for kicking the idle load balancer in the presence
5662 * of an idle cpu is the system.
5663 * - This rq has more than one task.
5664 * - At any scheduler domain level, this cpu's scheduler group has multiple
5665 * busy cpu's exceeding the group's power.
5666 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5667 * domain span are idle.
5669 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5671 unsigned long now = jiffies;
5672 struct sched_domain *sd;
5674 if (unlikely(idle_cpu(cpu)))
5678 * We may be recently in ticked or tickless idle mode. At the first
5679 * busy tick after returning from idle, we will update the busy stats.
5681 set_cpu_sd_state_busy();
5682 nohz_balance_exit_idle(cpu);
5685 * None are in tickless mode and hence no need for NOHZ idle load
5688 if (likely(!atomic_read(&nohz.nr_cpus)))
5691 if (time_before(now, nohz.next_balance))
5694 if (rq->nr_running >= 2)
5698 for_each_domain(cpu, sd) {
5699 struct sched_group *sg = sd->groups;
5700 struct sched_group_power *sgp = sg->sgp;
5701 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5703 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5704 goto need_kick_unlock;
5706 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5707 && (cpumask_first_and(nohz.idle_cpus_mask,
5708 sched_domain_span(sd)) < cpu))
5709 goto need_kick_unlock;
5711 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5723 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5727 * run_rebalance_domains is triggered when needed from the scheduler tick.
5728 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5730 static void run_rebalance_domains(struct softirq_action *h)
5732 int this_cpu = smp_processor_id();
5733 struct rq *this_rq = cpu_rq(this_cpu);
5734 enum cpu_idle_type idle = this_rq->idle_balance ?
5735 CPU_IDLE : CPU_NOT_IDLE;
5737 rebalance_domains(this_cpu, idle);
5740 * If this cpu has a pending nohz_balance_kick, then do the
5741 * balancing on behalf of the other idle cpus whose ticks are
5744 nohz_idle_balance(this_cpu, idle);
5747 static inline int on_null_domain(int cpu)
5749 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5753 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5755 void trigger_load_balance(struct rq *rq, int cpu)
5757 /* Don't need to rebalance while attached to NULL domain */
5758 if (time_after_eq(jiffies, rq->next_balance) &&
5759 likely(!on_null_domain(cpu)))
5760 raise_softirq(SCHED_SOFTIRQ);
5761 #ifdef CONFIG_NO_HZ_COMMON
5762 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5763 nohz_balancer_kick(cpu);
5767 static void rq_online_fair(struct rq *rq)
5772 static void rq_offline_fair(struct rq *rq)
5776 /* Ensure any throttled groups are reachable by pick_next_task */
5777 unthrottle_offline_cfs_rqs(rq);
5780 #endif /* CONFIG_SMP */
5783 * scheduler tick hitting a task of our scheduling class:
5785 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5787 struct cfs_rq *cfs_rq;
5788 struct sched_entity *se = &curr->se;
5790 for_each_sched_entity(se) {
5791 cfs_rq = cfs_rq_of(se);
5792 entity_tick(cfs_rq, se, queued);
5795 if (numabalancing_enabled)
5796 task_tick_numa(rq, curr);
5798 update_rq_runnable_avg(rq, 1);
5802 * called on fork with the child task as argument from the parent's context
5803 * - child not yet on the tasklist
5804 * - preemption disabled
5806 static void task_fork_fair(struct task_struct *p)
5808 struct cfs_rq *cfs_rq;
5809 struct sched_entity *se = &p->se, *curr;
5810 int this_cpu = smp_processor_id();
5811 struct rq *rq = this_rq();
5812 unsigned long flags;
5814 raw_spin_lock_irqsave(&rq->lock, flags);
5816 update_rq_clock(rq);
5818 cfs_rq = task_cfs_rq(current);
5819 curr = cfs_rq->curr;
5821 if (unlikely(task_cpu(p) != this_cpu)) {
5823 __set_task_cpu(p, this_cpu);
5827 update_curr(cfs_rq);
5830 se->vruntime = curr->vruntime;
5831 place_entity(cfs_rq, se, 1);
5833 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5835 * Upon rescheduling, sched_class::put_prev_task() will place
5836 * 'current' within the tree based on its new key value.
5838 swap(curr->vruntime, se->vruntime);
5839 resched_task(rq->curr);
5842 se->vruntime -= cfs_rq->min_vruntime;
5844 raw_spin_unlock_irqrestore(&rq->lock, flags);
5848 * Priority of the task has changed. Check to see if we preempt
5852 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5858 * Reschedule if we are currently running on this runqueue and
5859 * our priority decreased, or if we are not currently running on
5860 * this runqueue and our priority is higher than the current's
5862 if (rq->curr == p) {
5863 if (p->prio > oldprio)
5864 resched_task(rq->curr);
5866 check_preempt_curr(rq, p, 0);
5869 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5871 struct sched_entity *se = &p->se;
5872 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5875 * Ensure the task's vruntime is normalized, so that when its
5876 * switched back to the fair class the enqueue_entity(.flags=0) will
5877 * do the right thing.
5879 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5880 * have normalized the vruntime, if it was !on_rq, then only when
5881 * the task is sleeping will it still have non-normalized vruntime.
5883 if (!se->on_rq && p->state != TASK_RUNNING) {
5885 * Fix up our vruntime so that the current sleep doesn't
5886 * cause 'unlimited' sleep bonus.
5888 place_entity(cfs_rq, se, 0);
5889 se->vruntime -= cfs_rq->min_vruntime;
5894 * Remove our load from contribution when we leave sched_fair
5895 * and ensure we don't carry in an old decay_count if we
5898 if (p->se.avg.decay_count) {
5899 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
5900 __synchronize_entity_decay(&p->se);
5901 subtract_blocked_load_contrib(cfs_rq,
5902 p->se.avg.load_avg_contrib);
5908 * We switched to the sched_fair class.
5910 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5916 * We were most likely switched from sched_rt, so
5917 * kick off the schedule if running, otherwise just see
5918 * if we can still preempt the current task.
5921 resched_task(rq->curr);
5923 check_preempt_curr(rq, p, 0);
5926 /* Account for a task changing its policy or group.
5928 * This routine is mostly called to set cfs_rq->curr field when a task
5929 * migrates between groups/classes.
5931 static void set_curr_task_fair(struct rq *rq)
5933 struct sched_entity *se = &rq->curr->se;
5935 for_each_sched_entity(se) {
5936 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5938 set_next_entity(cfs_rq, se);
5939 /* ensure bandwidth has been allocated on our new cfs_rq */
5940 account_cfs_rq_runtime(cfs_rq, 0);
5944 void init_cfs_rq(struct cfs_rq *cfs_rq)
5946 cfs_rq->tasks_timeline = RB_ROOT;
5947 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5948 #ifndef CONFIG_64BIT
5949 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5952 atomic64_set(&cfs_rq->decay_counter, 1);
5953 atomic_long_set(&cfs_rq->removed_load, 0);
5957 #ifdef CONFIG_FAIR_GROUP_SCHED
5958 static void task_move_group_fair(struct task_struct *p, int on_rq)
5960 struct cfs_rq *cfs_rq;
5962 * If the task was not on the rq at the time of this cgroup movement
5963 * it must have been asleep, sleeping tasks keep their ->vruntime
5964 * absolute on their old rq until wakeup (needed for the fair sleeper
5965 * bonus in place_entity()).
5967 * If it was on the rq, we've just 'preempted' it, which does convert
5968 * ->vruntime to a relative base.
5970 * Make sure both cases convert their relative position when migrating
5971 * to another cgroup's rq. This does somewhat interfere with the
5972 * fair sleeper stuff for the first placement, but who cares.
5975 * When !on_rq, vruntime of the task has usually NOT been normalized.
5976 * But there are some cases where it has already been normalized:
5978 * - Moving a forked child which is waiting for being woken up by
5979 * wake_up_new_task().
5980 * - Moving a task which has been woken up by try_to_wake_up() and
5981 * waiting for actually being woken up by sched_ttwu_pending().
5983 * To prevent boost or penalty in the new cfs_rq caused by delta
5984 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5986 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5990 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5991 set_task_rq(p, task_cpu(p));
5993 cfs_rq = cfs_rq_of(&p->se);
5994 p->se.vruntime += cfs_rq->min_vruntime;
5997 * migrate_task_rq_fair() will have removed our previous
5998 * contribution, but we must synchronize for ongoing future
6001 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6002 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6007 void free_fair_sched_group(struct task_group *tg)
6011 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6013 for_each_possible_cpu(i) {
6015 kfree(tg->cfs_rq[i]);
6024 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6026 struct cfs_rq *cfs_rq;
6027 struct sched_entity *se;
6030 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6033 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6037 tg->shares = NICE_0_LOAD;
6039 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6041 for_each_possible_cpu(i) {
6042 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6043 GFP_KERNEL, cpu_to_node(i));
6047 se = kzalloc_node(sizeof(struct sched_entity),
6048 GFP_KERNEL, cpu_to_node(i));
6052 init_cfs_rq(cfs_rq);
6053 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6064 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6066 struct rq *rq = cpu_rq(cpu);
6067 unsigned long flags;
6070 * Only empty task groups can be destroyed; so we can speculatively
6071 * check on_list without danger of it being re-added.
6073 if (!tg->cfs_rq[cpu]->on_list)
6076 raw_spin_lock_irqsave(&rq->lock, flags);
6077 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6078 raw_spin_unlock_irqrestore(&rq->lock, flags);
6081 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6082 struct sched_entity *se, int cpu,
6083 struct sched_entity *parent)
6085 struct rq *rq = cpu_rq(cpu);
6089 init_cfs_rq_runtime(cfs_rq);
6091 tg->cfs_rq[cpu] = cfs_rq;
6094 /* se could be NULL for root_task_group */
6099 se->cfs_rq = &rq->cfs;
6101 se->cfs_rq = parent->my_q;
6104 update_load_set(&se->load, 0);
6105 se->parent = parent;
6108 static DEFINE_MUTEX(shares_mutex);
6110 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6113 unsigned long flags;
6116 * We can't change the weight of the root cgroup.
6121 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6123 mutex_lock(&shares_mutex);
6124 if (tg->shares == shares)
6127 tg->shares = shares;
6128 for_each_possible_cpu(i) {
6129 struct rq *rq = cpu_rq(i);
6130 struct sched_entity *se;
6133 /* Propagate contribution to hierarchy */
6134 raw_spin_lock_irqsave(&rq->lock, flags);
6136 /* Possible calls to update_curr() need rq clock */
6137 update_rq_clock(rq);
6138 for_each_sched_entity(se)
6139 update_cfs_shares(group_cfs_rq(se));
6140 raw_spin_unlock_irqrestore(&rq->lock, flags);
6144 mutex_unlock(&shares_mutex);
6147 #else /* CONFIG_FAIR_GROUP_SCHED */
6149 void free_fair_sched_group(struct task_group *tg) { }
6151 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6156 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6158 #endif /* CONFIG_FAIR_GROUP_SCHED */
6161 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6163 struct sched_entity *se = &task->se;
6164 unsigned int rr_interval = 0;
6167 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6170 if (rq->cfs.load.weight)
6171 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6177 * All the scheduling class methods:
6179 const struct sched_class fair_sched_class = {
6180 .next = &idle_sched_class,
6181 .enqueue_task = enqueue_task_fair,
6182 .dequeue_task = dequeue_task_fair,
6183 .yield_task = yield_task_fair,
6184 .yield_to_task = yield_to_task_fair,
6186 .check_preempt_curr = check_preempt_wakeup,
6188 .pick_next_task = pick_next_task_fair,
6189 .put_prev_task = put_prev_task_fair,
6192 .select_task_rq = select_task_rq_fair,
6193 .migrate_task_rq = migrate_task_rq_fair,
6195 .rq_online = rq_online_fair,
6196 .rq_offline = rq_offline_fair,
6198 .task_waking = task_waking_fair,
6201 .set_curr_task = set_curr_task_fair,
6202 .task_tick = task_tick_fair,
6203 .task_fork = task_fork_fair,
6205 .prio_changed = prio_changed_fair,
6206 .switched_from = switched_from_fair,
6207 .switched_to = switched_to_fair,
6209 .get_rr_interval = get_rr_interval_fair,
6211 #ifdef CONFIG_FAIR_GROUP_SCHED
6212 .task_move_group = task_move_group_fair,
6216 #ifdef CONFIG_SCHED_DEBUG
6217 void print_cfs_stats(struct seq_file *m, int cpu)
6219 struct cfs_rq *cfs_rq;
6222 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6223 print_cfs_rq(m, cpu, cfs_rq);
6228 __init void init_sched_fair_class(void)
6231 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6233 #ifdef CONFIG_NO_HZ_COMMON
6234 nohz.next_balance = jiffies;
6235 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6236 cpu_notifier(sched_ilb_notifier, 0);