2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
199 struct load_weight *lw)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
209 tmp = (u64)delta_exec * scale_load_down(weight);
211 tmp = (u64)delta_exec;
213 if (!lw->inv_weight) {
214 unsigned long w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp > WMULT_CONST))
228 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
231 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
233 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
237 const struct sched_class fair_sched_class;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct *task_of(struct sched_entity *se)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se));
259 return container_of(se, struct task_struct, se);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
283 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
312 if (cfs_rq->on_list) {
313 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity *se, struct sched_entity *pse)
326 if (se->cfs_rq == pse->cfs_rq)
332 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity *se)
342 for_each_sched_entity(se)
349 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
351 int se_depth, pse_depth;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth = depth_se(*se);
362 pse_depth = depth_se(*pse);
364 while (se_depth > pse_depth) {
366 *se = parent_entity(*se);
369 while (pse_depth > se_depth) {
371 *pse = parent_entity(*pse);
374 while (!is_same_group(*se, *pse)) {
375 *se = parent_entity(*se);
376 *pse = parent_entity(*pse);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct *task_of(struct sched_entity *se)
384 return container_of(se, struct task_struct, se);
387 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
389 return container_of(cfs_rq, struct rq, cfs);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
399 return &task_rq(p)->cfs;
402 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
404 struct task_struct *p = task_of(se);
405 struct rq *rq = task_rq(p);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity *se, struct sched_entity *pse)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta, struct sched_entity *se)
618 if (unlikely(se->load.weight != NICE_0_LOAD))
619 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64 __sched_period(unsigned long nr_running)
634 u64 period = sysctl_sched_latency;
635 unsigned long nr_latency = sched_nr_latency;
637 if (unlikely(nr_running > nr_latency)) {
638 period = sysctl_sched_min_granularity;
639 period *= nr_running;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
655 for_each_sched_entity(se) {
656 struct load_weight *load;
657 struct load_weight lw;
659 cfs_rq = cfs_rq_of(se);
660 load = &cfs_rq->load;
662 if (unlikely(!se->on_rq)) {
665 update_load_add(&lw, se->load.weight);
668 slice = calc_delta_mine(slice, se->load.weight, load);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
680 return calc_delta_fair(sched_slice(cfs_rq, se), se);
684 static inline void __update_task_entity_contrib(struct sched_entity *se);
686 /* Give new task start runnable values to heavy its load in infant time */
687 void init_task_runnable_average(struct task_struct *p)
691 p->se.avg.decay_count = 0;
692 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
693 p->se.avg.runnable_avg_sum = slice;
694 p->se.avg.runnable_avg_period = slice;
695 __update_task_entity_contrib(&p->se);
698 void init_task_runnable_average(struct task_struct *p)
704 * Update the current task's runtime statistics. Skip current tasks that
705 * are not in our scheduling class.
708 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
709 unsigned long delta_exec)
711 unsigned long delta_exec_weighted;
713 schedstat_set(curr->statistics.exec_max,
714 max((u64)delta_exec, curr->statistics.exec_max));
716 curr->sum_exec_runtime += delta_exec;
717 schedstat_add(cfs_rq, exec_clock, delta_exec);
718 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
720 curr->vruntime += delta_exec_weighted;
721 update_min_vruntime(cfs_rq);
724 static void update_curr(struct cfs_rq *cfs_rq)
726 struct sched_entity *curr = cfs_rq->curr;
727 u64 now = rq_clock_task(rq_of(cfs_rq));
728 unsigned long delta_exec;
734 * Get the amount of time the current task was running
735 * since the last time we changed load (this cannot
736 * overflow on 32 bits):
738 delta_exec = (unsigned long)(now - curr->exec_start);
742 __update_curr(cfs_rq, curr, delta_exec);
743 curr->exec_start = now;
745 if (entity_is_task(curr)) {
746 struct task_struct *curtask = task_of(curr);
748 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
749 cpuacct_charge(curtask, delta_exec);
750 account_group_exec_runtime(curtask, delta_exec);
753 account_cfs_rq_runtime(cfs_rq, delta_exec);
757 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
759 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
763 * Task is being enqueued - update stats:
765 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
768 * Are we enqueueing a waiting task? (for current tasks
769 * a dequeue/enqueue event is a NOP)
771 if (se != cfs_rq->curr)
772 update_stats_wait_start(cfs_rq, se);
776 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
778 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
779 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
780 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
781 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
782 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
783 #ifdef CONFIG_SCHEDSTATS
784 if (entity_is_task(se)) {
785 trace_sched_stat_wait(task_of(se),
786 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
789 schedstat_set(se->statistics.wait_start, 0);
793 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
796 * Mark the end of the wait period if dequeueing a
799 if (se != cfs_rq->curr)
800 update_stats_wait_end(cfs_rq, se);
804 * We are picking a new current task - update its stats:
807 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
810 * We are starting a new run period:
812 se->exec_start = rq_clock_task(rq_of(cfs_rq));
815 /**************************************************
816 * Scheduling class queueing methods:
819 #ifdef CONFIG_NUMA_BALANCING
821 * numa task sample period in ms
823 unsigned int sysctl_numa_balancing_scan_period_min = 100;
824 unsigned int sysctl_numa_balancing_scan_period_max = 100*50;
825 unsigned int sysctl_numa_balancing_scan_period_reset = 100*600;
827 /* Portion of address space to scan in MB */
828 unsigned int sysctl_numa_balancing_scan_size = 256;
830 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
831 unsigned int sysctl_numa_balancing_scan_delay = 1000;
833 static void task_numa_placement(struct task_struct *p)
837 if (!p->mm) /* for example, ksmd faulting in a user's mm */
839 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
840 if (p->numa_scan_seq == seq)
842 p->numa_scan_seq = seq;
844 /* FIXME: Scheduling placement policy hints go here */
848 * Got a PROT_NONE fault for a page on @node.
850 void task_numa_fault(int node, int pages, bool migrated)
852 struct task_struct *p = current;
854 if (!numabalancing_enabled)
857 /* FIXME: Allocate task-specific structure for placement policy here */
860 * If pages are properly placed (did not migrate) then scan slower.
861 * This is reset periodically in case of phase changes
864 p->numa_scan_period = min(sysctl_numa_balancing_scan_period_max,
865 p->numa_scan_period + jiffies_to_msecs(10));
867 task_numa_placement(p);
870 static void reset_ptenuma_scan(struct task_struct *p)
872 ACCESS_ONCE(p->mm->numa_scan_seq)++;
873 p->mm->numa_scan_offset = 0;
877 * The expensive part of numa migration is done from task_work context.
878 * Triggered from task_tick_numa().
880 void task_numa_work(struct callback_head *work)
882 unsigned long migrate, next_scan, now = jiffies;
883 struct task_struct *p = current;
884 struct mm_struct *mm = p->mm;
885 struct vm_area_struct *vma;
886 unsigned long start, end;
889 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
891 work->next = work; /* protect against double add */
893 * Who cares about NUMA placement when they're dying.
895 * NOTE: make sure not to dereference p->mm before this check,
896 * exit_task_work() happens _after_ exit_mm() so we could be called
897 * without p->mm even though we still had it when we enqueued this
900 if (p->flags & PF_EXITING)
904 * We do not care about task placement until a task runs on a node
905 * other than the first one used by the address space. This is
906 * largely because migrations are driven by what CPU the task
907 * is running on. If it's never scheduled on another node, it'll
908 * not migrate so why bother trapping the fault.
910 if (mm->first_nid == NUMA_PTE_SCAN_INIT)
911 mm->first_nid = numa_node_id();
912 if (mm->first_nid != NUMA_PTE_SCAN_ACTIVE) {
913 /* Are we running on a new node yet? */
914 if (numa_node_id() == mm->first_nid &&
915 !sched_feat_numa(NUMA_FORCE))
918 mm->first_nid = NUMA_PTE_SCAN_ACTIVE;
922 * Reset the scan period if enough time has gone by. Objective is that
923 * scanning will be reduced if pages are properly placed. As tasks
924 * can enter different phases this needs to be re-examined. Lacking
925 * proper tracking of reference behaviour, this blunt hammer is used.
927 migrate = mm->numa_next_reset;
928 if (time_after(now, migrate)) {
929 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
930 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
931 xchg(&mm->numa_next_reset, next_scan);
935 * Enforce maximal scan/migration frequency..
937 migrate = mm->numa_next_scan;
938 if (time_before(now, migrate))
941 if (p->numa_scan_period == 0)
942 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
944 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
945 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
949 * Delay this task enough that another task of this mm will likely win
950 * the next time around.
952 p->node_stamp += 2 * TICK_NSEC;
954 start = mm->numa_scan_offset;
955 pages = sysctl_numa_balancing_scan_size;
956 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
960 down_read(&mm->mmap_sem);
961 vma = find_vma(mm, start);
963 reset_ptenuma_scan(p);
967 for (; vma; vma = vma->vm_next) {
968 if (!vma_migratable(vma))
971 /* Skip small VMAs. They are not likely to be of relevance */
972 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
976 start = max(start, vma->vm_start);
977 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
978 end = min(end, vma->vm_end);
979 pages -= change_prot_numa(vma, start, end);
984 } while (end != vma->vm_end);
989 * It is possible to reach the end of the VMA list but the last few
990 * VMAs are not guaranteed to the vma_migratable. If they are not, we
991 * would find the !migratable VMA on the next scan but not reset the
992 * scanner to the start so check it now.
995 mm->numa_scan_offset = start;
997 reset_ptenuma_scan(p);
998 up_read(&mm->mmap_sem);
1002 * Drive the periodic memory faults..
1004 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1006 struct callback_head *work = &curr->numa_work;
1010 * We don't care about NUMA placement if we don't have memory.
1012 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1016 * Using runtime rather than walltime has the dual advantage that
1017 * we (mostly) drive the selection from busy threads and that the
1018 * task needs to have done some actual work before we bother with
1021 now = curr->se.sum_exec_runtime;
1022 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1024 if (now - curr->node_stamp > period) {
1025 if (!curr->node_stamp)
1026 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
1027 curr->node_stamp += period;
1029 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1030 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1031 task_work_add(curr, work, true);
1036 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1039 #endif /* CONFIG_NUMA_BALANCING */
1042 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1044 update_load_add(&cfs_rq->load, se->load.weight);
1045 if (!parent_entity(se))
1046 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1048 if (entity_is_task(se))
1049 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1051 cfs_rq->nr_running++;
1055 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1057 update_load_sub(&cfs_rq->load, se->load.weight);
1058 if (!parent_entity(se))
1059 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1060 if (entity_is_task(se))
1061 list_del_init(&se->group_node);
1062 cfs_rq->nr_running--;
1065 #ifdef CONFIG_FAIR_GROUP_SCHED
1067 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1072 * Use this CPU's actual weight instead of the last load_contribution
1073 * to gain a more accurate current total weight. See
1074 * update_cfs_rq_load_contribution().
1076 tg_weight = atomic_long_read(&tg->load_avg);
1077 tg_weight -= cfs_rq->tg_load_contrib;
1078 tg_weight += cfs_rq->load.weight;
1083 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1085 long tg_weight, load, shares;
1087 tg_weight = calc_tg_weight(tg, cfs_rq);
1088 load = cfs_rq->load.weight;
1090 shares = (tg->shares * load);
1092 shares /= tg_weight;
1094 if (shares < MIN_SHARES)
1095 shares = MIN_SHARES;
1096 if (shares > tg->shares)
1097 shares = tg->shares;
1101 # else /* CONFIG_SMP */
1102 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1106 # endif /* CONFIG_SMP */
1107 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1108 unsigned long weight)
1111 /* commit outstanding execution time */
1112 if (cfs_rq->curr == se)
1113 update_curr(cfs_rq);
1114 account_entity_dequeue(cfs_rq, se);
1117 update_load_set(&se->load, weight);
1120 account_entity_enqueue(cfs_rq, se);
1123 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1125 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1127 struct task_group *tg;
1128 struct sched_entity *se;
1132 se = tg->se[cpu_of(rq_of(cfs_rq))];
1133 if (!se || throttled_hierarchy(cfs_rq))
1136 if (likely(se->load.weight == tg->shares))
1139 shares = calc_cfs_shares(cfs_rq, tg);
1141 reweight_entity(cfs_rq_of(se), se, shares);
1143 #else /* CONFIG_FAIR_GROUP_SCHED */
1144 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1147 #endif /* CONFIG_FAIR_GROUP_SCHED */
1151 * We choose a half-life close to 1 scheduling period.
1152 * Note: The tables below are dependent on this value.
1154 #define LOAD_AVG_PERIOD 32
1155 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1156 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1158 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1159 static const u32 runnable_avg_yN_inv[] = {
1160 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1161 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1162 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1163 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1164 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1165 0x85aac367, 0x82cd8698,
1169 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1170 * over-estimates when re-combining.
1172 static const u32 runnable_avg_yN_sum[] = {
1173 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1174 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1175 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1180 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1182 static __always_inline u64 decay_load(u64 val, u64 n)
1184 unsigned int local_n;
1188 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1191 /* after bounds checking we can collapse to 32-bit */
1195 * As y^PERIOD = 1/2, we can combine
1196 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1197 * With a look-up table which covers k^n (n<PERIOD)
1199 * To achieve constant time decay_load.
1201 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1202 val >>= local_n / LOAD_AVG_PERIOD;
1203 local_n %= LOAD_AVG_PERIOD;
1206 val *= runnable_avg_yN_inv[local_n];
1207 /* We don't use SRR here since we always want to round down. */
1212 * For updates fully spanning n periods, the contribution to runnable
1213 * average will be: \Sum 1024*y^n
1215 * We can compute this reasonably efficiently by combining:
1216 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1218 static u32 __compute_runnable_contrib(u64 n)
1222 if (likely(n <= LOAD_AVG_PERIOD))
1223 return runnable_avg_yN_sum[n];
1224 else if (unlikely(n >= LOAD_AVG_MAX_N))
1225 return LOAD_AVG_MAX;
1227 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1229 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1230 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1232 n -= LOAD_AVG_PERIOD;
1233 } while (n > LOAD_AVG_PERIOD);
1235 contrib = decay_load(contrib, n);
1236 return contrib + runnable_avg_yN_sum[n];
1240 * We can represent the historical contribution to runnable average as the
1241 * coefficients of a geometric series. To do this we sub-divide our runnable
1242 * history into segments of approximately 1ms (1024us); label the segment that
1243 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1245 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1247 * (now) (~1ms ago) (~2ms ago)
1249 * Let u_i denote the fraction of p_i that the entity was runnable.
1251 * We then designate the fractions u_i as our co-efficients, yielding the
1252 * following representation of historical load:
1253 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1255 * We choose y based on the with of a reasonably scheduling period, fixing:
1258 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1259 * approximately half as much as the contribution to load within the last ms
1262 * When a period "rolls over" and we have new u_0`, multiplying the previous
1263 * sum again by y is sufficient to update:
1264 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1265 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1267 static __always_inline int __update_entity_runnable_avg(u64 now,
1268 struct sched_avg *sa,
1272 u32 runnable_contrib;
1273 int delta_w, decayed = 0;
1275 delta = now - sa->last_runnable_update;
1277 * This should only happen when time goes backwards, which it
1278 * unfortunately does during sched clock init when we swap over to TSC.
1280 if ((s64)delta < 0) {
1281 sa->last_runnable_update = now;
1286 * Use 1024ns as the unit of measurement since it's a reasonable
1287 * approximation of 1us and fast to compute.
1292 sa->last_runnable_update = now;
1294 /* delta_w is the amount already accumulated against our next period */
1295 delta_w = sa->runnable_avg_period % 1024;
1296 if (delta + delta_w >= 1024) {
1297 /* period roll-over */
1301 * Now that we know we're crossing a period boundary, figure
1302 * out how much from delta we need to complete the current
1303 * period and accrue it.
1305 delta_w = 1024 - delta_w;
1307 sa->runnable_avg_sum += delta_w;
1308 sa->runnable_avg_period += delta_w;
1312 /* Figure out how many additional periods this update spans */
1313 periods = delta / 1024;
1316 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1318 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1321 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1322 runnable_contrib = __compute_runnable_contrib(periods);
1324 sa->runnable_avg_sum += runnable_contrib;
1325 sa->runnable_avg_period += runnable_contrib;
1328 /* Remainder of delta accrued against u_0` */
1330 sa->runnable_avg_sum += delta;
1331 sa->runnable_avg_period += delta;
1336 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1337 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1339 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1340 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1342 decays -= se->avg.decay_count;
1346 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1347 se->avg.decay_count = 0;
1352 #ifdef CONFIG_FAIR_GROUP_SCHED
1353 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1356 struct task_group *tg = cfs_rq->tg;
1359 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1360 tg_contrib -= cfs_rq->tg_load_contrib;
1362 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1363 atomic_long_add(tg_contrib, &tg->load_avg);
1364 cfs_rq->tg_load_contrib += tg_contrib;
1369 * Aggregate cfs_rq runnable averages into an equivalent task_group
1370 * representation for computing load contributions.
1372 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1373 struct cfs_rq *cfs_rq)
1375 struct task_group *tg = cfs_rq->tg;
1378 /* The fraction of a cpu used by this cfs_rq */
1379 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1380 sa->runnable_avg_period + 1);
1381 contrib -= cfs_rq->tg_runnable_contrib;
1383 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1384 atomic_add(contrib, &tg->runnable_avg);
1385 cfs_rq->tg_runnable_contrib += contrib;
1389 static inline void __update_group_entity_contrib(struct sched_entity *se)
1391 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1392 struct task_group *tg = cfs_rq->tg;
1397 contrib = cfs_rq->tg_load_contrib * tg->shares;
1398 se->avg.load_avg_contrib = div_u64(contrib,
1399 atomic_long_read(&tg->load_avg) + 1);
1402 * For group entities we need to compute a correction term in the case
1403 * that they are consuming <1 cpu so that we would contribute the same
1404 * load as a task of equal weight.
1406 * Explicitly co-ordinating this measurement would be expensive, but
1407 * fortunately the sum of each cpus contribution forms a usable
1408 * lower-bound on the true value.
1410 * Consider the aggregate of 2 contributions. Either they are disjoint
1411 * (and the sum represents true value) or they are disjoint and we are
1412 * understating by the aggregate of their overlap.
1414 * Extending this to N cpus, for a given overlap, the maximum amount we
1415 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1416 * cpus that overlap for this interval and w_i is the interval width.
1418 * On a small machine; the first term is well-bounded which bounds the
1419 * total error since w_i is a subset of the period. Whereas on a
1420 * larger machine, while this first term can be larger, if w_i is the
1421 * of consequential size guaranteed to see n_i*w_i quickly converge to
1422 * our upper bound of 1-cpu.
1424 runnable_avg = atomic_read(&tg->runnable_avg);
1425 if (runnable_avg < NICE_0_LOAD) {
1426 se->avg.load_avg_contrib *= runnable_avg;
1427 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1431 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1432 int force_update) {}
1433 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1434 struct cfs_rq *cfs_rq) {}
1435 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1438 static inline void __update_task_entity_contrib(struct sched_entity *se)
1442 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1443 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1444 contrib /= (se->avg.runnable_avg_period + 1);
1445 se->avg.load_avg_contrib = scale_load(contrib);
1448 /* Compute the current contribution to load_avg by se, return any delta */
1449 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1451 long old_contrib = se->avg.load_avg_contrib;
1453 if (entity_is_task(se)) {
1454 __update_task_entity_contrib(se);
1456 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1457 __update_group_entity_contrib(se);
1460 return se->avg.load_avg_contrib - old_contrib;
1463 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1466 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1467 cfs_rq->blocked_load_avg -= load_contrib;
1469 cfs_rq->blocked_load_avg = 0;
1472 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1474 /* Update a sched_entity's runnable average */
1475 static inline void update_entity_load_avg(struct sched_entity *se,
1478 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1483 * For a group entity we need to use their owned cfs_rq_clock_task() in
1484 * case they are the parent of a throttled hierarchy.
1486 if (entity_is_task(se))
1487 now = cfs_rq_clock_task(cfs_rq);
1489 now = cfs_rq_clock_task(group_cfs_rq(se));
1491 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1494 contrib_delta = __update_entity_load_avg_contrib(se);
1500 cfs_rq->runnable_load_avg += contrib_delta;
1502 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1506 * Decay the load contributed by all blocked children and account this so that
1507 * their contribution may appropriately discounted when they wake up.
1509 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1511 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1514 decays = now - cfs_rq->last_decay;
1515 if (!decays && !force_update)
1518 if (atomic_long_read(&cfs_rq->removed_load)) {
1519 unsigned long removed_load;
1520 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1521 subtract_blocked_load_contrib(cfs_rq, removed_load);
1525 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1527 atomic64_add(decays, &cfs_rq->decay_counter);
1528 cfs_rq->last_decay = now;
1531 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1534 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1536 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1537 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1540 /* Add the load generated by se into cfs_rq's child load-average */
1541 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1542 struct sched_entity *se,
1546 * We track migrations using entity decay_count <= 0, on a wake-up
1547 * migration we use a negative decay count to track the remote decays
1548 * accumulated while sleeping.
1550 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1551 * are seen by enqueue_entity_load_avg() as a migration with an already
1552 * constructed load_avg_contrib.
1554 if (unlikely(se->avg.decay_count <= 0)) {
1555 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1556 if (se->avg.decay_count) {
1558 * In a wake-up migration we have to approximate the
1559 * time sleeping. This is because we can't synchronize
1560 * clock_task between the two cpus, and it is not
1561 * guaranteed to be read-safe. Instead, we can
1562 * approximate this using our carried decays, which are
1563 * explicitly atomically readable.
1565 se->avg.last_runnable_update -= (-se->avg.decay_count)
1567 update_entity_load_avg(se, 0);
1568 /* Indicate that we're now synchronized and on-rq */
1569 se->avg.decay_count = 0;
1574 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1575 * would have made count negative); we must be careful to avoid
1576 * double-accounting blocked time after synchronizing decays.
1578 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1582 /* migrated tasks did not contribute to our blocked load */
1584 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1585 update_entity_load_avg(se, 0);
1588 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1589 /* we force update consideration on load-balancer moves */
1590 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1594 * Remove se's load from this cfs_rq child load-average, if the entity is
1595 * transitioning to a blocked state we track its projected decay using
1598 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1599 struct sched_entity *se,
1602 update_entity_load_avg(se, 1);
1603 /* we force update consideration on load-balancer moves */
1604 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1606 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1608 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1609 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1610 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1614 * Update the rq's load with the elapsed running time before entering
1615 * idle. if the last scheduled task is not a CFS task, idle_enter will
1616 * be the only way to update the runnable statistic.
1618 void idle_enter_fair(struct rq *this_rq)
1620 update_rq_runnable_avg(this_rq, 1);
1624 * Update the rq's load with the elapsed idle time before a task is
1625 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1626 * be the only way to update the runnable statistic.
1628 void idle_exit_fair(struct rq *this_rq)
1630 update_rq_runnable_avg(this_rq, 0);
1634 static inline void update_entity_load_avg(struct sched_entity *se,
1635 int update_cfs_rq) {}
1636 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1637 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1638 struct sched_entity *se,
1640 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1641 struct sched_entity *se,
1643 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1644 int force_update) {}
1647 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1649 #ifdef CONFIG_SCHEDSTATS
1650 struct task_struct *tsk = NULL;
1652 if (entity_is_task(se))
1655 if (se->statistics.sleep_start) {
1656 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1661 if (unlikely(delta > se->statistics.sleep_max))
1662 se->statistics.sleep_max = delta;
1664 se->statistics.sleep_start = 0;
1665 se->statistics.sum_sleep_runtime += delta;
1668 account_scheduler_latency(tsk, delta >> 10, 1);
1669 trace_sched_stat_sleep(tsk, delta);
1672 if (se->statistics.block_start) {
1673 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1678 if (unlikely(delta > se->statistics.block_max))
1679 se->statistics.block_max = delta;
1681 se->statistics.block_start = 0;
1682 se->statistics.sum_sleep_runtime += delta;
1685 if (tsk->in_iowait) {
1686 se->statistics.iowait_sum += delta;
1687 se->statistics.iowait_count++;
1688 trace_sched_stat_iowait(tsk, delta);
1691 trace_sched_stat_blocked(tsk, delta);
1694 * Blocking time is in units of nanosecs, so shift by
1695 * 20 to get a milliseconds-range estimation of the
1696 * amount of time that the task spent sleeping:
1698 if (unlikely(prof_on == SLEEP_PROFILING)) {
1699 profile_hits(SLEEP_PROFILING,
1700 (void *)get_wchan(tsk),
1703 account_scheduler_latency(tsk, delta >> 10, 0);
1709 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1711 #ifdef CONFIG_SCHED_DEBUG
1712 s64 d = se->vruntime - cfs_rq->min_vruntime;
1717 if (d > 3*sysctl_sched_latency)
1718 schedstat_inc(cfs_rq, nr_spread_over);
1723 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1725 u64 vruntime = cfs_rq->min_vruntime;
1728 * The 'current' period is already promised to the current tasks,
1729 * however the extra weight of the new task will slow them down a
1730 * little, place the new task so that it fits in the slot that
1731 * stays open at the end.
1733 if (initial && sched_feat(START_DEBIT))
1734 vruntime += sched_vslice(cfs_rq, se);
1736 /* sleeps up to a single latency don't count. */
1738 unsigned long thresh = sysctl_sched_latency;
1741 * Halve their sleep time's effect, to allow
1742 * for a gentler effect of sleepers:
1744 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1750 /* ensure we never gain time by being placed backwards. */
1751 se->vruntime = max_vruntime(se->vruntime, vruntime);
1754 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1757 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1760 * Update the normalized vruntime before updating min_vruntime
1761 * through calling update_curr().
1763 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1764 se->vruntime += cfs_rq->min_vruntime;
1767 * Update run-time statistics of the 'current'.
1769 update_curr(cfs_rq);
1770 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1771 account_entity_enqueue(cfs_rq, se);
1772 update_cfs_shares(cfs_rq);
1774 if (flags & ENQUEUE_WAKEUP) {
1775 place_entity(cfs_rq, se, 0);
1776 enqueue_sleeper(cfs_rq, se);
1779 update_stats_enqueue(cfs_rq, se);
1780 check_spread(cfs_rq, se);
1781 if (se != cfs_rq->curr)
1782 __enqueue_entity(cfs_rq, se);
1785 if (cfs_rq->nr_running == 1) {
1786 list_add_leaf_cfs_rq(cfs_rq);
1787 check_enqueue_throttle(cfs_rq);
1791 static void __clear_buddies_last(struct sched_entity *se)
1793 for_each_sched_entity(se) {
1794 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1795 if (cfs_rq->last == se)
1796 cfs_rq->last = NULL;
1802 static void __clear_buddies_next(struct sched_entity *se)
1804 for_each_sched_entity(se) {
1805 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1806 if (cfs_rq->next == se)
1807 cfs_rq->next = NULL;
1813 static void __clear_buddies_skip(struct sched_entity *se)
1815 for_each_sched_entity(se) {
1816 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1817 if (cfs_rq->skip == se)
1818 cfs_rq->skip = NULL;
1824 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1826 if (cfs_rq->last == se)
1827 __clear_buddies_last(se);
1829 if (cfs_rq->next == se)
1830 __clear_buddies_next(se);
1832 if (cfs_rq->skip == se)
1833 __clear_buddies_skip(se);
1836 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1839 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1842 * Update run-time statistics of the 'current'.
1844 update_curr(cfs_rq);
1845 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1847 update_stats_dequeue(cfs_rq, se);
1848 if (flags & DEQUEUE_SLEEP) {
1849 #ifdef CONFIG_SCHEDSTATS
1850 if (entity_is_task(se)) {
1851 struct task_struct *tsk = task_of(se);
1853 if (tsk->state & TASK_INTERRUPTIBLE)
1854 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
1855 if (tsk->state & TASK_UNINTERRUPTIBLE)
1856 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
1861 clear_buddies(cfs_rq, se);
1863 if (se != cfs_rq->curr)
1864 __dequeue_entity(cfs_rq, se);
1866 account_entity_dequeue(cfs_rq, se);
1869 * Normalize the entity after updating the min_vruntime because the
1870 * update can refer to the ->curr item and we need to reflect this
1871 * movement in our normalized position.
1873 if (!(flags & DEQUEUE_SLEEP))
1874 se->vruntime -= cfs_rq->min_vruntime;
1876 /* return excess runtime on last dequeue */
1877 return_cfs_rq_runtime(cfs_rq);
1879 update_min_vruntime(cfs_rq);
1880 update_cfs_shares(cfs_rq);
1884 * Preempt the current task with a newly woken task if needed:
1887 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1889 unsigned long ideal_runtime, delta_exec;
1890 struct sched_entity *se;
1893 ideal_runtime = sched_slice(cfs_rq, curr);
1894 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1895 if (delta_exec > ideal_runtime) {
1896 resched_task(rq_of(cfs_rq)->curr);
1898 * The current task ran long enough, ensure it doesn't get
1899 * re-elected due to buddy favours.
1901 clear_buddies(cfs_rq, curr);
1906 * Ensure that a task that missed wakeup preemption by a
1907 * narrow margin doesn't have to wait for a full slice.
1908 * This also mitigates buddy induced latencies under load.
1910 if (delta_exec < sysctl_sched_min_granularity)
1913 se = __pick_first_entity(cfs_rq);
1914 delta = curr->vruntime - se->vruntime;
1919 if (delta > ideal_runtime)
1920 resched_task(rq_of(cfs_rq)->curr);
1924 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1926 /* 'current' is not kept within the tree. */
1929 * Any task has to be enqueued before it get to execute on
1930 * a CPU. So account for the time it spent waiting on the
1933 update_stats_wait_end(cfs_rq, se);
1934 __dequeue_entity(cfs_rq, se);
1937 update_stats_curr_start(cfs_rq, se);
1939 #ifdef CONFIG_SCHEDSTATS
1941 * Track our maximum slice length, if the CPU's load is at
1942 * least twice that of our own weight (i.e. dont track it
1943 * when there are only lesser-weight tasks around):
1945 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1946 se->statistics.slice_max = max(se->statistics.slice_max,
1947 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1950 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1954 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1957 * Pick the next process, keeping these things in mind, in this order:
1958 * 1) keep things fair between processes/task groups
1959 * 2) pick the "next" process, since someone really wants that to run
1960 * 3) pick the "last" process, for cache locality
1961 * 4) do not run the "skip" process, if something else is available
1963 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1965 struct sched_entity *se = __pick_first_entity(cfs_rq);
1966 struct sched_entity *left = se;
1969 * Avoid running the skip buddy, if running something else can
1970 * be done without getting too unfair.
1972 if (cfs_rq->skip == se) {
1973 struct sched_entity *second = __pick_next_entity(se);
1974 if (second && wakeup_preempt_entity(second, left) < 1)
1979 * Prefer last buddy, try to return the CPU to a preempted task.
1981 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1985 * Someone really wants this to run. If it's not unfair, run it.
1987 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1990 clear_buddies(cfs_rq, se);
1995 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1997 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2000 * If still on the runqueue then deactivate_task()
2001 * was not called and update_curr() has to be done:
2004 update_curr(cfs_rq);
2006 /* throttle cfs_rqs exceeding runtime */
2007 check_cfs_rq_runtime(cfs_rq);
2009 check_spread(cfs_rq, prev);
2011 update_stats_wait_start(cfs_rq, prev);
2012 /* Put 'current' back into the tree. */
2013 __enqueue_entity(cfs_rq, prev);
2014 /* in !on_rq case, update occurred at dequeue */
2015 update_entity_load_avg(prev, 1);
2017 cfs_rq->curr = NULL;
2021 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2024 * Update run-time statistics of the 'current'.
2026 update_curr(cfs_rq);
2029 * Ensure that runnable average is periodically updated.
2031 update_entity_load_avg(curr, 1);
2032 update_cfs_rq_blocked_load(cfs_rq, 1);
2033 update_cfs_shares(cfs_rq);
2035 #ifdef CONFIG_SCHED_HRTICK
2037 * queued ticks are scheduled to match the slice, so don't bother
2038 * validating it and just reschedule.
2041 resched_task(rq_of(cfs_rq)->curr);
2045 * don't let the period tick interfere with the hrtick preemption
2047 if (!sched_feat(DOUBLE_TICK) &&
2048 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2052 if (cfs_rq->nr_running > 1)
2053 check_preempt_tick(cfs_rq, curr);
2057 /**************************************************
2058 * CFS bandwidth control machinery
2061 #ifdef CONFIG_CFS_BANDWIDTH
2063 #ifdef HAVE_JUMP_LABEL
2064 static struct static_key __cfs_bandwidth_used;
2066 static inline bool cfs_bandwidth_used(void)
2068 return static_key_false(&__cfs_bandwidth_used);
2071 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2073 /* only need to count groups transitioning between enabled/!enabled */
2074 if (enabled && !was_enabled)
2075 static_key_slow_inc(&__cfs_bandwidth_used);
2076 else if (!enabled && was_enabled)
2077 static_key_slow_dec(&__cfs_bandwidth_used);
2079 #else /* HAVE_JUMP_LABEL */
2080 static bool cfs_bandwidth_used(void)
2085 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2086 #endif /* HAVE_JUMP_LABEL */
2089 * default period for cfs group bandwidth.
2090 * default: 0.1s, units: nanoseconds
2092 static inline u64 default_cfs_period(void)
2094 return 100000000ULL;
2097 static inline u64 sched_cfs_bandwidth_slice(void)
2099 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2103 * Replenish runtime according to assigned quota and update expiration time.
2104 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2105 * additional synchronization around rq->lock.
2107 * requires cfs_b->lock
2109 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2113 if (cfs_b->quota == RUNTIME_INF)
2116 now = sched_clock_cpu(smp_processor_id());
2117 cfs_b->runtime = cfs_b->quota;
2118 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2121 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2123 return &tg->cfs_bandwidth;
2126 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2127 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2129 if (unlikely(cfs_rq->throttle_count))
2130 return cfs_rq->throttled_clock_task;
2132 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2135 /* returns 0 on failure to allocate runtime */
2136 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2138 struct task_group *tg = cfs_rq->tg;
2139 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2140 u64 amount = 0, min_amount, expires;
2142 /* note: this is a positive sum as runtime_remaining <= 0 */
2143 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2145 raw_spin_lock(&cfs_b->lock);
2146 if (cfs_b->quota == RUNTIME_INF)
2147 amount = min_amount;
2150 * If the bandwidth pool has become inactive, then at least one
2151 * period must have elapsed since the last consumption.
2152 * Refresh the global state and ensure bandwidth timer becomes
2155 if (!cfs_b->timer_active) {
2156 __refill_cfs_bandwidth_runtime(cfs_b);
2157 __start_cfs_bandwidth(cfs_b);
2160 if (cfs_b->runtime > 0) {
2161 amount = min(cfs_b->runtime, min_amount);
2162 cfs_b->runtime -= amount;
2166 expires = cfs_b->runtime_expires;
2167 raw_spin_unlock(&cfs_b->lock);
2169 cfs_rq->runtime_remaining += amount;
2171 * we may have advanced our local expiration to account for allowed
2172 * spread between our sched_clock and the one on which runtime was
2175 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2176 cfs_rq->runtime_expires = expires;
2178 return cfs_rq->runtime_remaining > 0;
2182 * Note: This depends on the synchronization provided by sched_clock and the
2183 * fact that rq->clock snapshots this value.
2185 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2187 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2189 /* if the deadline is ahead of our clock, nothing to do */
2190 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2193 if (cfs_rq->runtime_remaining < 0)
2197 * If the local deadline has passed we have to consider the
2198 * possibility that our sched_clock is 'fast' and the global deadline
2199 * has not truly expired.
2201 * Fortunately we can check determine whether this the case by checking
2202 * whether the global deadline has advanced.
2205 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2206 /* extend local deadline, drift is bounded above by 2 ticks */
2207 cfs_rq->runtime_expires += TICK_NSEC;
2209 /* global deadline is ahead, expiration has passed */
2210 cfs_rq->runtime_remaining = 0;
2214 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2215 unsigned long delta_exec)
2217 /* dock delta_exec before expiring quota (as it could span periods) */
2218 cfs_rq->runtime_remaining -= delta_exec;
2219 expire_cfs_rq_runtime(cfs_rq);
2221 if (likely(cfs_rq->runtime_remaining > 0))
2225 * if we're unable to extend our runtime we resched so that the active
2226 * hierarchy can be throttled
2228 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2229 resched_task(rq_of(cfs_rq)->curr);
2232 static __always_inline
2233 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2235 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2238 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2241 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2243 return cfs_bandwidth_used() && cfs_rq->throttled;
2246 /* check whether cfs_rq, or any parent, is throttled */
2247 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2249 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2253 * Ensure that neither of the group entities corresponding to src_cpu or
2254 * dest_cpu are members of a throttled hierarchy when performing group
2255 * load-balance operations.
2257 static inline int throttled_lb_pair(struct task_group *tg,
2258 int src_cpu, int dest_cpu)
2260 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2262 src_cfs_rq = tg->cfs_rq[src_cpu];
2263 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2265 return throttled_hierarchy(src_cfs_rq) ||
2266 throttled_hierarchy(dest_cfs_rq);
2269 /* updated child weight may affect parent so we have to do this bottom up */
2270 static int tg_unthrottle_up(struct task_group *tg, void *data)
2272 struct rq *rq = data;
2273 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2275 cfs_rq->throttle_count--;
2277 if (!cfs_rq->throttle_count) {
2278 /* adjust cfs_rq_clock_task() */
2279 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2280 cfs_rq->throttled_clock_task;
2287 static int tg_throttle_down(struct task_group *tg, void *data)
2289 struct rq *rq = data;
2290 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2292 /* group is entering throttled state, stop time */
2293 if (!cfs_rq->throttle_count)
2294 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2295 cfs_rq->throttle_count++;
2300 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2302 struct rq *rq = rq_of(cfs_rq);
2303 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2304 struct sched_entity *se;
2305 long task_delta, dequeue = 1;
2307 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2309 /* freeze hierarchy runnable averages while throttled */
2311 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2314 task_delta = cfs_rq->h_nr_running;
2315 for_each_sched_entity(se) {
2316 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2317 /* throttled entity or throttle-on-deactivate */
2322 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2323 qcfs_rq->h_nr_running -= task_delta;
2325 if (qcfs_rq->load.weight)
2330 rq->nr_running -= task_delta;
2332 cfs_rq->throttled = 1;
2333 cfs_rq->throttled_clock = rq_clock(rq);
2334 raw_spin_lock(&cfs_b->lock);
2335 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2336 raw_spin_unlock(&cfs_b->lock);
2339 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2341 struct rq *rq = rq_of(cfs_rq);
2342 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2343 struct sched_entity *se;
2347 se = cfs_rq->tg->se[cpu_of(rq)];
2349 cfs_rq->throttled = 0;
2351 update_rq_clock(rq);
2353 raw_spin_lock(&cfs_b->lock);
2354 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2355 list_del_rcu(&cfs_rq->throttled_list);
2356 raw_spin_unlock(&cfs_b->lock);
2358 /* update hierarchical throttle state */
2359 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2361 if (!cfs_rq->load.weight)
2364 task_delta = cfs_rq->h_nr_running;
2365 for_each_sched_entity(se) {
2369 cfs_rq = cfs_rq_of(se);
2371 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2372 cfs_rq->h_nr_running += task_delta;
2374 if (cfs_rq_throttled(cfs_rq))
2379 rq->nr_running += task_delta;
2381 /* determine whether we need to wake up potentially idle cpu */
2382 if (rq->curr == rq->idle && rq->cfs.nr_running)
2383 resched_task(rq->curr);
2386 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2387 u64 remaining, u64 expires)
2389 struct cfs_rq *cfs_rq;
2390 u64 runtime = remaining;
2393 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2395 struct rq *rq = rq_of(cfs_rq);
2397 raw_spin_lock(&rq->lock);
2398 if (!cfs_rq_throttled(cfs_rq))
2401 runtime = -cfs_rq->runtime_remaining + 1;
2402 if (runtime > remaining)
2403 runtime = remaining;
2404 remaining -= runtime;
2406 cfs_rq->runtime_remaining += runtime;
2407 cfs_rq->runtime_expires = expires;
2409 /* we check whether we're throttled above */
2410 if (cfs_rq->runtime_remaining > 0)
2411 unthrottle_cfs_rq(cfs_rq);
2414 raw_spin_unlock(&rq->lock);
2425 * Responsible for refilling a task_group's bandwidth and unthrottling its
2426 * cfs_rqs as appropriate. If there has been no activity within the last
2427 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2428 * used to track this state.
2430 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2432 u64 runtime, runtime_expires;
2433 int idle = 1, throttled;
2435 raw_spin_lock(&cfs_b->lock);
2436 /* no need to continue the timer with no bandwidth constraint */
2437 if (cfs_b->quota == RUNTIME_INF)
2440 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2441 /* idle depends on !throttled (for the case of a large deficit) */
2442 idle = cfs_b->idle && !throttled;
2443 cfs_b->nr_periods += overrun;
2445 /* if we're going inactive then everything else can be deferred */
2449 __refill_cfs_bandwidth_runtime(cfs_b);
2452 /* mark as potentially idle for the upcoming period */
2457 /* account preceding periods in which throttling occurred */
2458 cfs_b->nr_throttled += overrun;
2461 * There are throttled entities so we must first use the new bandwidth
2462 * to unthrottle them before making it generally available. This
2463 * ensures that all existing debts will be paid before a new cfs_rq is
2466 runtime = cfs_b->runtime;
2467 runtime_expires = cfs_b->runtime_expires;
2471 * This check is repeated as we are holding onto the new bandwidth
2472 * while we unthrottle. This can potentially race with an unthrottled
2473 * group trying to acquire new bandwidth from the global pool.
2475 while (throttled && runtime > 0) {
2476 raw_spin_unlock(&cfs_b->lock);
2477 /* we can't nest cfs_b->lock while distributing bandwidth */
2478 runtime = distribute_cfs_runtime(cfs_b, runtime,
2480 raw_spin_lock(&cfs_b->lock);
2482 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2485 /* return (any) remaining runtime */
2486 cfs_b->runtime = runtime;
2488 * While we are ensured activity in the period following an
2489 * unthrottle, this also covers the case in which the new bandwidth is
2490 * insufficient to cover the existing bandwidth deficit. (Forcing the
2491 * timer to remain active while there are any throttled entities.)
2496 cfs_b->timer_active = 0;
2497 raw_spin_unlock(&cfs_b->lock);
2502 /* a cfs_rq won't donate quota below this amount */
2503 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2504 /* minimum remaining period time to redistribute slack quota */
2505 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2506 /* how long we wait to gather additional slack before distributing */
2507 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2509 /* are we near the end of the current quota period? */
2510 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2512 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2515 /* if the call-back is running a quota refresh is already occurring */
2516 if (hrtimer_callback_running(refresh_timer))
2519 /* is a quota refresh about to occur? */
2520 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2521 if (remaining < min_expire)
2527 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2529 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2531 /* if there's a quota refresh soon don't bother with slack */
2532 if (runtime_refresh_within(cfs_b, min_left))
2535 start_bandwidth_timer(&cfs_b->slack_timer,
2536 ns_to_ktime(cfs_bandwidth_slack_period));
2539 /* we know any runtime found here is valid as update_curr() precedes return */
2540 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2542 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2543 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2545 if (slack_runtime <= 0)
2548 raw_spin_lock(&cfs_b->lock);
2549 if (cfs_b->quota != RUNTIME_INF &&
2550 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2551 cfs_b->runtime += slack_runtime;
2553 /* we are under rq->lock, defer unthrottling using a timer */
2554 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2555 !list_empty(&cfs_b->throttled_cfs_rq))
2556 start_cfs_slack_bandwidth(cfs_b);
2558 raw_spin_unlock(&cfs_b->lock);
2560 /* even if it's not valid for return we don't want to try again */
2561 cfs_rq->runtime_remaining -= slack_runtime;
2564 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2566 if (!cfs_bandwidth_used())
2569 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2572 __return_cfs_rq_runtime(cfs_rq);
2576 * This is done with a timer (instead of inline with bandwidth return) since
2577 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2579 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2581 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2584 /* confirm we're still not at a refresh boundary */
2585 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2588 raw_spin_lock(&cfs_b->lock);
2589 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2590 runtime = cfs_b->runtime;
2593 expires = cfs_b->runtime_expires;
2594 raw_spin_unlock(&cfs_b->lock);
2599 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2601 raw_spin_lock(&cfs_b->lock);
2602 if (expires == cfs_b->runtime_expires)
2603 cfs_b->runtime = runtime;
2604 raw_spin_unlock(&cfs_b->lock);
2608 * When a group wakes up we want to make sure that its quota is not already
2609 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2610 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2612 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2614 if (!cfs_bandwidth_used())
2617 /* an active group must be handled by the update_curr()->put() path */
2618 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2621 /* ensure the group is not already throttled */
2622 if (cfs_rq_throttled(cfs_rq))
2625 /* update runtime allocation */
2626 account_cfs_rq_runtime(cfs_rq, 0);
2627 if (cfs_rq->runtime_remaining <= 0)
2628 throttle_cfs_rq(cfs_rq);
2631 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2632 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2634 if (!cfs_bandwidth_used())
2637 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2641 * it's possible for a throttled entity to be forced into a running
2642 * state (e.g. set_curr_task), in this case we're finished.
2644 if (cfs_rq_throttled(cfs_rq))
2647 throttle_cfs_rq(cfs_rq);
2650 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2652 struct cfs_bandwidth *cfs_b =
2653 container_of(timer, struct cfs_bandwidth, slack_timer);
2654 do_sched_cfs_slack_timer(cfs_b);
2656 return HRTIMER_NORESTART;
2659 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2661 struct cfs_bandwidth *cfs_b =
2662 container_of(timer, struct cfs_bandwidth, period_timer);
2668 now = hrtimer_cb_get_time(timer);
2669 overrun = hrtimer_forward(timer, now, cfs_b->period);
2674 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2677 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2680 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2682 raw_spin_lock_init(&cfs_b->lock);
2684 cfs_b->quota = RUNTIME_INF;
2685 cfs_b->period = ns_to_ktime(default_cfs_period());
2687 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2688 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2689 cfs_b->period_timer.function = sched_cfs_period_timer;
2690 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2691 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2694 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2696 cfs_rq->runtime_enabled = 0;
2697 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2700 /* requires cfs_b->lock, may release to reprogram timer */
2701 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2704 * The timer may be active because we're trying to set a new bandwidth
2705 * period or because we're racing with the tear-down path
2706 * (timer_active==0 becomes visible before the hrtimer call-back
2707 * terminates). In either case we ensure that it's re-programmed
2709 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2710 raw_spin_unlock(&cfs_b->lock);
2711 /* ensure cfs_b->lock is available while we wait */
2712 hrtimer_cancel(&cfs_b->period_timer);
2714 raw_spin_lock(&cfs_b->lock);
2715 /* if someone else restarted the timer then we're done */
2716 if (cfs_b->timer_active)
2720 cfs_b->timer_active = 1;
2721 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2724 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2726 hrtimer_cancel(&cfs_b->period_timer);
2727 hrtimer_cancel(&cfs_b->slack_timer);
2730 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2732 struct cfs_rq *cfs_rq;
2734 for_each_leaf_cfs_rq(rq, cfs_rq) {
2735 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2737 if (!cfs_rq->runtime_enabled)
2741 * clock_task is not advancing so we just need to make sure
2742 * there's some valid quota amount
2744 cfs_rq->runtime_remaining = cfs_b->quota;
2745 if (cfs_rq_throttled(cfs_rq))
2746 unthrottle_cfs_rq(cfs_rq);
2750 #else /* CONFIG_CFS_BANDWIDTH */
2751 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2753 return rq_clock_task(rq_of(cfs_rq));
2756 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2757 unsigned long delta_exec) {}
2758 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2759 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2760 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2762 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2767 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2772 static inline int throttled_lb_pair(struct task_group *tg,
2773 int src_cpu, int dest_cpu)
2778 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2780 #ifdef CONFIG_FAIR_GROUP_SCHED
2781 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2784 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2788 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2789 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2791 #endif /* CONFIG_CFS_BANDWIDTH */
2793 /**************************************************
2794 * CFS operations on tasks:
2797 #ifdef CONFIG_SCHED_HRTICK
2798 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2800 struct sched_entity *se = &p->se;
2801 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2803 WARN_ON(task_rq(p) != rq);
2805 if (cfs_rq->nr_running > 1) {
2806 u64 slice = sched_slice(cfs_rq, se);
2807 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2808 s64 delta = slice - ran;
2817 * Don't schedule slices shorter than 10000ns, that just
2818 * doesn't make sense. Rely on vruntime for fairness.
2821 delta = max_t(s64, 10000LL, delta);
2823 hrtick_start(rq, delta);
2828 * called from enqueue/dequeue and updates the hrtick when the
2829 * current task is from our class and nr_running is low enough
2832 static void hrtick_update(struct rq *rq)
2834 struct task_struct *curr = rq->curr;
2836 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2839 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2840 hrtick_start_fair(rq, curr);
2842 #else /* !CONFIG_SCHED_HRTICK */
2844 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2848 static inline void hrtick_update(struct rq *rq)
2854 * The enqueue_task method is called before nr_running is
2855 * increased. Here we update the fair scheduling stats and
2856 * then put the task into the rbtree:
2859 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2861 struct cfs_rq *cfs_rq;
2862 struct sched_entity *se = &p->se;
2864 for_each_sched_entity(se) {
2867 cfs_rq = cfs_rq_of(se);
2868 enqueue_entity(cfs_rq, se, flags);
2871 * end evaluation on encountering a throttled cfs_rq
2873 * note: in the case of encountering a throttled cfs_rq we will
2874 * post the final h_nr_running increment below.
2876 if (cfs_rq_throttled(cfs_rq))
2878 cfs_rq->h_nr_running++;
2880 flags = ENQUEUE_WAKEUP;
2883 for_each_sched_entity(se) {
2884 cfs_rq = cfs_rq_of(se);
2885 cfs_rq->h_nr_running++;
2887 if (cfs_rq_throttled(cfs_rq))
2890 update_cfs_shares(cfs_rq);
2891 update_entity_load_avg(se, 1);
2895 update_rq_runnable_avg(rq, rq->nr_running);
2901 static void set_next_buddy(struct sched_entity *se);
2904 * The dequeue_task method is called before nr_running is
2905 * decreased. We remove the task from the rbtree and
2906 * update the fair scheduling stats:
2908 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2910 struct cfs_rq *cfs_rq;
2911 struct sched_entity *se = &p->se;
2912 int task_sleep = flags & DEQUEUE_SLEEP;
2914 for_each_sched_entity(se) {
2915 cfs_rq = cfs_rq_of(se);
2916 dequeue_entity(cfs_rq, se, flags);
2919 * end evaluation on encountering a throttled cfs_rq
2921 * note: in the case of encountering a throttled cfs_rq we will
2922 * post the final h_nr_running decrement below.
2924 if (cfs_rq_throttled(cfs_rq))
2926 cfs_rq->h_nr_running--;
2928 /* Don't dequeue parent if it has other entities besides us */
2929 if (cfs_rq->load.weight) {
2931 * Bias pick_next to pick a task from this cfs_rq, as
2932 * p is sleeping when it is within its sched_slice.
2934 if (task_sleep && parent_entity(se))
2935 set_next_buddy(parent_entity(se));
2937 /* avoid re-evaluating load for this entity */
2938 se = parent_entity(se);
2941 flags |= DEQUEUE_SLEEP;
2944 for_each_sched_entity(se) {
2945 cfs_rq = cfs_rq_of(se);
2946 cfs_rq->h_nr_running--;
2948 if (cfs_rq_throttled(cfs_rq))
2951 update_cfs_shares(cfs_rq);
2952 update_entity_load_avg(se, 1);
2957 update_rq_runnable_avg(rq, 1);
2963 /* Used instead of source_load when we know the type == 0 */
2964 static unsigned long weighted_cpuload(const int cpu)
2966 return cpu_rq(cpu)->cfs.runnable_load_avg;
2970 * Return a low guess at the load of a migration-source cpu weighted
2971 * according to the scheduling class and "nice" value.
2973 * We want to under-estimate the load of migration sources, to
2974 * balance conservatively.
2976 static unsigned long source_load(int cpu, int type)
2978 struct rq *rq = cpu_rq(cpu);
2979 unsigned long total = weighted_cpuload(cpu);
2981 if (type == 0 || !sched_feat(LB_BIAS))
2984 return min(rq->cpu_load[type-1], total);
2988 * Return a high guess at the load of a migration-target cpu weighted
2989 * according to the scheduling class and "nice" value.
2991 static unsigned long target_load(int cpu, int type)
2993 struct rq *rq = cpu_rq(cpu);
2994 unsigned long total = weighted_cpuload(cpu);
2996 if (type == 0 || !sched_feat(LB_BIAS))
2999 return max(rq->cpu_load[type-1], total);
3002 static unsigned long power_of(int cpu)
3004 return cpu_rq(cpu)->cpu_power;
3007 static unsigned long cpu_avg_load_per_task(int cpu)
3009 struct rq *rq = cpu_rq(cpu);
3010 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3011 unsigned long load_avg = rq->cfs.runnable_load_avg;
3014 return load_avg / nr_running;
3019 static void record_wakee(struct task_struct *p)
3022 * Rough decay (wiping) for cost saving, don't worry
3023 * about the boundary, really active task won't care
3026 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3027 current->wakee_flips = 0;
3028 current->wakee_flip_decay_ts = jiffies;
3031 if (current->last_wakee != p) {
3032 current->last_wakee = p;
3033 current->wakee_flips++;
3037 static void task_waking_fair(struct task_struct *p)
3039 struct sched_entity *se = &p->se;
3040 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3043 #ifndef CONFIG_64BIT
3044 u64 min_vruntime_copy;
3047 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3049 min_vruntime = cfs_rq->min_vruntime;
3050 } while (min_vruntime != min_vruntime_copy);
3052 min_vruntime = cfs_rq->min_vruntime;
3055 se->vruntime -= min_vruntime;
3059 #ifdef CONFIG_FAIR_GROUP_SCHED
3061 * effective_load() calculates the load change as seen from the root_task_group
3063 * Adding load to a group doesn't make a group heavier, but can cause movement
3064 * of group shares between cpus. Assuming the shares were perfectly aligned one
3065 * can calculate the shift in shares.
3067 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3068 * on this @cpu and results in a total addition (subtraction) of @wg to the
3069 * total group weight.
3071 * Given a runqueue weight distribution (rw_i) we can compute a shares
3072 * distribution (s_i) using:
3074 * s_i = rw_i / \Sum rw_j (1)
3076 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3077 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3078 * shares distribution (s_i):
3080 * rw_i = { 2, 4, 1, 0 }
3081 * s_i = { 2/7, 4/7, 1/7, 0 }
3083 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3084 * task used to run on and the CPU the waker is running on), we need to
3085 * compute the effect of waking a task on either CPU and, in case of a sync
3086 * wakeup, compute the effect of the current task going to sleep.
3088 * So for a change of @wl to the local @cpu with an overall group weight change
3089 * of @wl we can compute the new shares distribution (s'_i) using:
3091 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3093 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3094 * differences in waking a task to CPU 0. The additional task changes the
3095 * weight and shares distributions like:
3097 * rw'_i = { 3, 4, 1, 0 }
3098 * s'_i = { 3/8, 4/8, 1/8, 0 }
3100 * We can then compute the difference in effective weight by using:
3102 * dw_i = S * (s'_i - s_i) (3)
3104 * Where 'S' is the group weight as seen by its parent.
3106 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3107 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3108 * 4/7) times the weight of the group.
3110 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3112 struct sched_entity *se = tg->se[cpu];
3114 if (!tg->parent) /* the trivial, non-cgroup case */
3117 for_each_sched_entity(se) {
3123 * W = @wg + \Sum rw_j
3125 W = wg + calc_tg_weight(tg, se->my_q);
3130 w = se->my_q->load.weight + wl;
3133 * wl = S * s'_i; see (2)
3136 wl = (w * tg->shares) / W;
3141 * Per the above, wl is the new se->load.weight value; since
3142 * those are clipped to [MIN_SHARES, ...) do so now. See
3143 * calc_cfs_shares().
3145 if (wl < MIN_SHARES)
3149 * wl = dw_i = S * (s'_i - s_i); see (3)
3151 wl -= se->load.weight;
3154 * Recursively apply this logic to all parent groups to compute
3155 * the final effective load change on the root group. Since
3156 * only the @tg group gets extra weight, all parent groups can
3157 * only redistribute existing shares. @wl is the shift in shares
3158 * resulting from this level per the above.
3167 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3168 unsigned long wl, unsigned long wg)
3175 static int wake_wide(struct task_struct *p)
3177 int factor = this_cpu_read(sd_llc_size);
3180 * Yeah, it's the switching-frequency, could means many wakee or
3181 * rapidly switch, use factor here will just help to automatically
3182 * adjust the loose-degree, so bigger node will lead to more pull.
3184 if (p->wakee_flips > factor) {
3186 * wakee is somewhat hot, it needs certain amount of cpu
3187 * resource, so if waker is far more hot, prefer to leave
3190 if (current->wakee_flips > (factor * p->wakee_flips))
3197 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3199 s64 this_load, load;
3200 int idx, this_cpu, prev_cpu;
3201 unsigned long tl_per_task;
3202 struct task_group *tg;
3203 unsigned long weight;
3207 * If we wake multiple tasks be careful to not bounce
3208 * ourselves around too much.
3214 this_cpu = smp_processor_id();
3215 prev_cpu = task_cpu(p);
3216 load = source_load(prev_cpu, idx);
3217 this_load = target_load(this_cpu, idx);
3220 * If sync wakeup then subtract the (maximum possible)
3221 * effect of the currently running task from the load
3222 * of the current CPU:
3225 tg = task_group(current);
3226 weight = current->se.load.weight;
3228 this_load += effective_load(tg, this_cpu, -weight, -weight);
3229 load += effective_load(tg, prev_cpu, 0, -weight);
3233 weight = p->se.load.weight;
3236 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3237 * due to the sync cause above having dropped this_load to 0, we'll
3238 * always have an imbalance, but there's really nothing you can do
3239 * about that, so that's good too.
3241 * Otherwise check if either cpus are near enough in load to allow this
3242 * task to be woken on this_cpu.
3244 if (this_load > 0) {
3245 s64 this_eff_load, prev_eff_load;
3247 this_eff_load = 100;
3248 this_eff_load *= power_of(prev_cpu);
3249 this_eff_load *= this_load +
3250 effective_load(tg, this_cpu, weight, weight);
3252 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3253 prev_eff_load *= power_of(this_cpu);
3254 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3256 balanced = this_eff_load <= prev_eff_load;
3261 * If the currently running task will sleep within
3262 * a reasonable amount of time then attract this newly
3265 if (sync && balanced)
3268 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3269 tl_per_task = cpu_avg_load_per_task(this_cpu);
3272 (this_load <= load &&
3273 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3275 * This domain has SD_WAKE_AFFINE and
3276 * p is cache cold in this domain, and
3277 * there is no bad imbalance.
3279 schedstat_inc(sd, ttwu_move_affine);
3280 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3288 * find_idlest_group finds and returns the least busy CPU group within the
3291 static struct sched_group *
3292 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3293 int this_cpu, int load_idx)
3295 struct sched_group *idlest = NULL, *group = sd->groups;
3296 unsigned long min_load = ULONG_MAX, this_load = 0;
3297 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3300 unsigned long load, avg_load;
3304 /* Skip over this group if it has no CPUs allowed */
3305 if (!cpumask_intersects(sched_group_cpus(group),
3306 tsk_cpus_allowed(p)))
3309 local_group = cpumask_test_cpu(this_cpu,
3310 sched_group_cpus(group));
3312 /* Tally up the load of all CPUs in the group */
3315 for_each_cpu(i, sched_group_cpus(group)) {
3316 /* Bias balancing toward cpus of our domain */
3318 load = source_load(i, load_idx);
3320 load = target_load(i, load_idx);
3325 /* Adjust by relative CPU power of the group */
3326 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3329 this_load = avg_load;
3330 } else if (avg_load < min_load) {
3331 min_load = avg_load;
3334 } while (group = group->next, group != sd->groups);
3336 if (!idlest || 100*this_load < imbalance*min_load)
3342 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3345 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3347 unsigned long load, min_load = ULONG_MAX;
3351 /* Traverse only the allowed CPUs */
3352 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3353 load = weighted_cpuload(i);
3355 if (load < min_load || (load == min_load && i == this_cpu)) {
3365 * Try and locate an idle CPU in the sched_domain.
3367 static int select_idle_sibling(struct task_struct *p, int target)
3369 struct sched_domain *sd;
3370 struct sched_group *sg;
3371 int i = task_cpu(p);
3373 if (idle_cpu(target))
3377 * If the prevous cpu is cache affine and idle, don't be stupid.
3379 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3383 * Otherwise, iterate the domains and find an elegible idle cpu.
3385 sd = rcu_dereference(per_cpu(sd_llc, target));
3386 for_each_lower_domain(sd) {
3389 if (!cpumask_intersects(sched_group_cpus(sg),
3390 tsk_cpus_allowed(p)))
3393 for_each_cpu(i, sched_group_cpus(sg)) {
3394 if (i == target || !idle_cpu(i))
3398 target = cpumask_first_and(sched_group_cpus(sg),
3399 tsk_cpus_allowed(p));
3403 } while (sg != sd->groups);
3410 * sched_balance_self: balance the current task (running on cpu) in domains
3411 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3414 * Balance, ie. select the least loaded group.
3416 * Returns the target CPU number, or the same CPU if no balancing is needed.
3418 * preempt must be disabled.
3421 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3423 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3424 int cpu = smp_processor_id();
3425 int prev_cpu = task_cpu(p);
3427 int want_affine = 0;
3428 int sync = wake_flags & WF_SYNC;
3430 if (p->nr_cpus_allowed == 1)
3433 if (sd_flag & SD_BALANCE_WAKE) {
3434 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3440 for_each_domain(cpu, tmp) {
3441 if (!(tmp->flags & SD_LOAD_BALANCE))
3445 * If both cpu and prev_cpu are part of this domain,
3446 * cpu is a valid SD_WAKE_AFFINE target.
3448 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3449 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3454 if (tmp->flags & sd_flag)
3459 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3462 new_cpu = select_idle_sibling(p, prev_cpu);
3467 int load_idx = sd->forkexec_idx;
3468 struct sched_group *group;
3471 if (!(sd->flags & sd_flag)) {
3476 if (sd_flag & SD_BALANCE_WAKE)
3477 load_idx = sd->wake_idx;
3479 group = find_idlest_group(sd, p, cpu, load_idx);
3485 new_cpu = find_idlest_cpu(group, p, cpu);
3486 if (new_cpu == -1 || new_cpu == cpu) {
3487 /* Now try balancing at a lower domain level of cpu */
3492 /* Now try balancing at a lower domain level of new_cpu */
3494 weight = sd->span_weight;
3496 for_each_domain(cpu, tmp) {
3497 if (weight <= tmp->span_weight)
3499 if (tmp->flags & sd_flag)
3502 /* while loop will break here if sd == NULL */
3511 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3512 * cfs_rq_of(p) references at time of call are still valid and identify the
3513 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3514 * other assumptions, including the state of rq->lock, should be made.
3517 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3519 struct sched_entity *se = &p->se;
3520 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3523 * Load tracking: accumulate removed load so that it can be processed
3524 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3525 * to blocked load iff they have a positive decay-count. It can never
3526 * be negative here since on-rq tasks have decay-count == 0.
3528 if (se->avg.decay_count) {
3529 se->avg.decay_count = -__synchronize_entity_decay(se);
3530 atomic_long_add(se->avg.load_avg_contrib,
3531 &cfs_rq->removed_load);
3534 #endif /* CONFIG_SMP */
3536 static unsigned long
3537 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3539 unsigned long gran = sysctl_sched_wakeup_granularity;
3542 * Since its curr running now, convert the gran from real-time
3543 * to virtual-time in his units.
3545 * By using 'se' instead of 'curr' we penalize light tasks, so
3546 * they get preempted easier. That is, if 'se' < 'curr' then
3547 * the resulting gran will be larger, therefore penalizing the
3548 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3549 * be smaller, again penalizing the lighter task.
3551 * This is especially important for buddies when the leftmost
3552 * task is higher priority than the buddy.
3554 return calc_delta_fair(gran, se);
3558 * Should 'se' preempt 'curr'.
3572 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3574 s64 gran, vdiff = curr->vruntime - se->vruntime;
3579 gran = wakeup_gran(curr, se);
3586 static void set_last_buddy(struct sched_entity *se)
3588 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3591 for_each_sched_entity(se)
3592 cfs_rq_of(se)->last = se;
3595 static void set_next_buddy(struct sched_entity *se)
3597 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3600 for_each_sched_entity(se)
3601 cfs_rq_of(se)->next = se;
3604 static void set_skip_buddy(struct sched_entity *se)
3606 for_each_sched_entity(se)
3607 cfs_rq_of(se)->skip = se;
3611 * Preempt the current task with a newly woken task if needed:
3613 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3615 struct task_struct *curr = rq->curr;
3616 struct sched_entity *se = &curr->se, *pse = &p->se;
3617 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3618 int scale = cfs_rq->nr_running >= sched_nr_latency;
3619 int next_buddy_marked = 0;
3621 if (unlikely(se == pse))
3625 * This is possible from callers such as move_task(), in which we
3626 * unconditionally check_prempt_curr() after an enqueue (which may have
3627 * lead to a throttle). This both saves work and prevents false
3628 * next-buddy nomination below.
3630 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3633 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3634 set_next_buddy(pse);
3635 next_buddy_marked = 1;
3639 * We can come here with TIF_NEED_RESCHED already set from new task
3642 * Note: this also catches the edge-case of curr being in a throttled
3643 * group (e.g. via set_curr_task), since update_curr() (in the
3644 * enqueue of curr) will have resulted in resched being set. This
3645 * prevents us from potentially nominating it as a false LAST_BUDDY
3648 if (test_tsk_need_resched(curr))
3651 /* Idle tasks are by definition preempted by non-idle tasks. */
3652 if (unlikely(curr->policy == SCHED_IDLE) &&
3653 likely(p->policy != SCHED_IDLE))
3657 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3658 * is driven by the tick):
3660 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3663 find_matching_se(&se, &pse);
3664 update_curr(cfs_rq_of(se));
3666 if (wakeup_preempt_entity(se, pse) == 1) {
3668 * Bias pick_next to pick the sched entity that is
3669 * triggering this preemption.
3671 if (!next_buddy_marked)
3672 set_next_buddy(pse);
3681 * Only set the backward buddy when the current task is still
3682 * on the rq. This can happen when a wakeup gets interleaved
3683 * with schedule on the ->pre_schedule() or idle_balance()
3684 * point, either of which can * drop the rq lock.
3686 * Also, during early boot the idle thread is in the fair class,
3687 * for obvious reasons its a bad idea to schedule back to it.
3689 if (unlikely(!se->on_rq || curr == rq->idle))
3692 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3696 static struct task_struct *pick_next_task_fair(struct rq *rq)
3698 struct task_struct *p;
3699 struct cfs_rq *cfs_rq = &rq->cfs;
3700 struct sched_entity *se;
3702 if (!cfs_rq->nr_running)
3706 se = pick_next_entity(cfs_rq);
3707 set_next_entity(cfs_rq, se);
3708 cfs_rq = group_cfs_rq(se);
3712 if (hrtick_enabled(rq))
3713 hrtick_start_fair(rq, p);
3719 * Account for a descheduled task:
3721 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3723 struct sched_entity *se = &prev->se;
3724 struct cfs_rq *cfs_rq;
3726 for_each_sched_entity(se) {
3727 cfs_rq = cfs_rq_of(se);
3728 put_prev_entity(cfs_rq, se);
3733 * sched_yield() is very simple
3735 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3737 static void yield_task_fair(struct rq *rq)
3739 struct task_struct *curr = rq->curr;
3740 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3741 struct sched_entity *se = &curr->se;
3744 * Are we the only task in the tree?
3746 if (unlikely(rq->nr_running == 1))
3749 clear_buddies(cfs_rq, se);
3751 if (curr->policy != SCHED_BATCH) {
3752 update_rq_clock(rq);
3754 * Update run-time statistics of the 'current'.
3756 update_curr(cfs_rq);
3758 * Tell update_rq_clock() that we've just updated,
3759 * so we don't do microscopic update in schedule()
3760 * and double the fastpath cost.
3762 rq->skip_clock_update = 1;
3768 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3770 struct sched_entity *se = &p->se;
3772 /* throttled hierarchies are not runnable */
3773 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3776 /* Tell the scheduler that we'd really like pse to run next. */
3779 yield_task_fair(rq);
3785 /**************************************************
3786 * Fair scheduling class load-balancing methods.
3790 * The purpose of load-balancing is to achieve the same basic fairness the
3791 * per-cpu scheduler provides, namely provide a proportional amount of compute
3792 * time to each task. This is expressed in the following equation:
3794 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3796 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3797 * W_i,0 is defined as:
3799 * W_i,0 = \Sum_j w_i,j (2)
3801 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3802 * is derived from the nice value as per prio_to_weight[].
3804 * The weight average is an exponential decay average of the instantaneous
3807 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3809 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3810 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3811 * can also include other factors [XXX].
3813 * To achieve this balance we define a measure of imbalance which follows
3814 * directly from (1):
3816 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3818 * We them move tasks around to minimize the imbalance. In the continuous
3819 * function space it is obvious this converges, in the discrete case we get
3820 * a few fun cases generally called infeasible weight scenarios.
3823 * - infeasible weights;
3824 * - local vs global optima in the discrete case. ]
3829 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3830 * for all i,j solution, we create a tree of cpus that follows the hardware
3831 * topology where each level pairs two lower groups (or better). This results
3832 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3833 * tree to only the first of the previous level and we decrease the frequency
3834 * of load-balance at each level inv. proportional to the number of cpus in
3840 * \Sum { --- * --- * 2^i } = O(n) (5)
3842 * `- size of each group
3843 * | | `- number of cpus doing load-balance
3845 * `- sum over all levels
3847 * Coupled with a limit on how many tasks we can migrate every balance pass,
3848 * this makes (5) the runtime complexity of the balancer.
3850 * An important property here is that each CPU is still (indirectly) connected
3851 * to every other cpu in at most O(log n) steps:
3853 * The adjacency matrix of the resulting graph is given by:
3856 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3859 * And you'll find that:
3861 * A^(log_2 n)_i,j != 0 for all i,j (7)
3863 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3864 * The task movement gives a factor of O(m), giving a convergence complexity
3867 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3872 * In order to avoid CPUs going idle while there's still work to do, new idle
3873 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3874 * tree itself instead of relying on other CPUs to bring it work.
3876 * This adds some complexity to both (5) and (8) but it reduces the total idle
3884 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3887 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3892 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3894 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3896 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3899 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3900 * rewrite all of this once again.]
3903 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3905 #define LBF_ALL_PINNED 0x01
3906 #define LBF_NEED_BREAK 0x02
3907 #define LBF_DST_PINNED 0x04
3908 #define LBF_SOME_PINNED 0x08
3911 struct sched_domain *sd;
3919 struct cpumask *dst_grpmask;
3921 enum cpu_idle_type idle;
3923 /* The set of CPUs under consideration for load-balancing */
3924 struct cpumask *cpus;
3929 unsigned int loop_break;
3930 unsigned int loop_max;
3934 * move_task - move a task from one runqueue to another runqueue.
3935 * Both runqueues must be locked.
3937 static void move_task(struct task_struct *p, struct lb_env *env)
3939 deactivate_task(env->src_rq, p, 0);
3940 set_task_cpu(p, env->dst_cpu);
3941 activate_task(env->dst_rq, p, 0);
3942 check_preempt_curr(env->dst_rq, p, 0);
3946 * Is this task likely cache-hot:
3949 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3953 if (p->sched_class != &fair_sched_class)
3956 if (unlikely(p->policy == SCHED_IDLE))
3960 * Buddy candidates are cache hot:
3962 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3963 (&p->se == cfs_rq_of(&p->se)->next ||
3964 &p->se == cfs_rq_of(&p->se)->last))
3967 if (sysctl_sched_migration_cost == -1)
3969 if (sysctl_sched_migration_cost == 0)
3972 delta = now - p->se.exec_start;
3974 return delta < (s64)sysctl_sched_migration_cost;
3978 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3981 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3983 int tsk_cache_hot = 0;
3985 * We do not migrate tasks that are:
3986 * 1) throttled_lb_pair, or
3987 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3988 * 3) running (obviously), or
3989 * 4) are cache-hot on their current CPU.
3991 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3994 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3997 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3999 env->flags |= LBF_SOME_PINNED;
4002 * Remember if this task can be migrated to any other cpu in
4003 * our sched_group. We may want to revisit it if we couldn't
4004 * meet load balance goals by pulling other tasks on src_cpu.
4006 * Also avoid computing new_dst_cpu if we have already computed
4007 * one in current iteration.
4009 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4012 /* Prevent to re-select dst_cpu via env's cpus */
4013 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4014 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4015 env->flags |= LBF_DST_PINNED;
4016 env->new_dst_cpu = cpu;
4024 /* Record that we found atleast one task that could run on dst_cpu */
4025 env->flags &= ~LBF_ALL_PINNED;
4027 if (task_running(env->src_rq, p)) {
4028 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4033 * Aggressive migration if:
4034 * 1) task is cache cold, or
4035 * 2) too many balance attempts have failed.
4038 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4039 if (!tsk_cache_hot ||
4040 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4042 if (tsk_cache_hot) {
4043 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4044 schedstat_inc(p, se.statistics.nr_forced_migrations);
4050 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4055 * move_one_task tries to move exactly one task from busiest to this_rq, as
4056 * part of active balancing operations within "domain".
4057 * Returns 1 if successful and 0 otherwise.
4059 * Called with both runqueues locked.
4061 static int move_one_task(struct lb_env *env)
4063 struct task_struct *p, *n;
4065 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4066 if (!can_migrate_task(p, env))
4071 * Right now, this is only the second place move_task()
4072 * is called, so we can safely collect move_task()
4073 * stats here rather than inside move_task().
4075 schedstat_inc(env->sd, lb_gained[env->idle]);
4081 static unsigned long task_h_load(struct task_struct *p);
4083 static const unsigned int sched_nr_migrate_break = 32;
4086 * move_tasks tries to move up to imbalance weighted load from busiest to
4087 * this_rq, as part of a balancing operation within domain "sd".
4088 * Returns 1 if successful and 0 otherwise.
4090 * Called with both runqueues locked.
4092 static int move_tasks(struct lb_env *env)
4094 struct list_head *tasks = &env->src_rq->cfs_tasks;
4095 struct task_struct *p;
4099 if (env->imbalance <= 0)
4102 while (!list_empty(tasks)) {
4103 p = list_first_entry(tasks, struct task_struct, se.group_node);
4106 /* We've more or less seen every task there is, call it quits */
4107 if (env->loop > env->loop_max)
4110 /* take a breather every nr_migrate tasks */
4111 if (env->loop > env->loop_break) {
4112 env->loop_break += sched_nr_migrate_break;
4113 env->flags |= LBF_NEED_BREAK;
4117 if (!can_migrate_task(p, env))
4120 load = task_h_load(p);
4122 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4125 if ((load / 2) > env->imbalance)
4130 env->imbalance -= load;
4132 #ifdef CONFIG_PREEMPT
4134 * NEWIDLE balancing is a source of latency, so preemptible
4135 * kernels will stop after the first task is pulled to minimize
4136 * the critical section.
4138 if (env->idle == CPU_NEWLY_IDLE)
4143 * We only want to steal up to the prescribed amount of
4146 if (env->imbalance <= 0)
4151 list_move_tail(&p->se.group_node, tasks);
4155 * Right now, this is one of only two places move_task() is called,
4156 * so we can safely collect move_task() stats here rather than
4157 * inside move_task().
4159 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4164 #ifdef CONFIG_FAIR_GROUP_SCHED
4166 * update tg->load_weight by folding this cpu's load_avg
4168 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4170 struct sched_entity *se = tg->se[cpu];
4171 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4173 /* throttled entities do not contribute to load */
4174 if (throttled_hierarchy(cfs_rq))
4177 update_cfs_rq_blocked_load(cfs_rq, 1);
4180 update_entity_load_avg(se, 1);
4182 * We pivot on our runnable average having decayed to zero for
4183 * list removal. This generally implies that all our children
4184 * have also been removed (modulo rounding error or bandwidth
4185 * control); however, such cases are rare and we can fix these
4188 * TODO: fix up out-of-order children on enqueue.
4190 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4191 list_del_leaf_cfs_rq(cfs_rq);
4193 struct rq *rq = rq_of(cfs_rq);
4194 update_rq_runnable_avg(rq, rq->nr_running);
4198 static void update_blocked_averages(int cpu)
4200 struct rq *rq = cpu_rq(cpu);
4201 struct cfs_rq *cfs_rq;
4202 unsigned long flags;
4204 raw_spin_lock_irqsave(&rq->lock, flags);
4205 update_rq_clock(rq);
4207 * Iterates the task_group tree in a bottom up fashion, see
4208 * list_add_leaf_cfs_rq() for details.
4210 for_each_leaf_cfs_rq(rq, cfs_rq) {
4212 * Note: We may want to consider periodically releasing
4213 * rq->lock about these updates so that creating many task
4214 * groups does not result in continually extending hold time.
4216 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4219 raw_spin_unlock_irqrestore(&rq->lock, flags);
4223 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4224 * This needs to be done in a top-down fashion because the load of a child
4225 * group is a fraction of its parents load.
4227 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4229 struct rq *rq = rq_of(cfs_rq);
4230 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4231 unsigned long now = jiffies;
4234 if (cfs_rq->last_h_load_update == now)
4237 cfs_rq->h_load_next = NULL;
4238 for_each_sched_entity(se) {
4239 cfs_rq = cfs_rq_of(se);
4240 cfs_rq->h_load_next = se;
4241 if (cfs_rq->last_h_load_update == now)
4246 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4247 cfs_rq->last_h_load_update = now;
4250 while ((se = cfs_rq->h_load_next) != NULL) {
4251 load = cfs_rq->h_load;
4252 load = div64_ul(load * se->avg.load_avg_contrib,
4253 cfs_rq->runnable_load_avg + 1);
4254 cfs_rq = group_cfs_rq(se);
4255 cfs_rq->h_load = load;
4256 cfs_rq->last_h_load_update = now;
4260 static unsigned long task_h_load(struct task_struct *p)
4262 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4264 update_cfs_rq_h_load(cfs_rq);
4265 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4266 cfs_rq->runnable_load_avg + 1);
4269 static inline void update_blocked_averages(int cpu)
4273 static unsigned long task_h_load(struct task_struct *p)
4275 return p->se.avg.load_avg_contrib;
4279 /********** Helpers for find_busiest_group ************************/
4281 * sg_lb_stats - stats of a sched_group required for load_balancing
4283 struct sg_lb_stats {
4284 unsigned long avg_load; /*Avg load across the CPUs of the group */
4285 unsigned long group_load; /* Total load over the CPUs of the group */
4286 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4287 unsigned long load_per_task;
4288 unsigned long group_power;
4289 unsigned int sum_nr_running; /* Nr tasks running in the group */
4290 unsigned int group_capacity;
4291 unsigned int idle_cpus;
4292 unsigned int group_weight;
4293 int group_imb; /* Is there an imbalance in the group ? */
4294 int group_has_capacity; /* Is there extra capacity in the group? */
4298 * sd_lb_stats - Structure to store the statistics of a sched_domain
4299 * during load balancing.
4301 struct sd_lb_stats {
4302 struct sched_group *busiest; /* Busiest group in this sd */
4303 struct sched_group *local; /* Local group in this sd */
4304 unsigned long total_load; /* Total load of all groups in sd */
4305 unsigned long total_pwr; /* Total power of all groups in sd */
4306 unsigned long avg_load; /* Average load across all groups in sd */
4308 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4309 struct sg_lb_stats local_stat; /* Statistics of the local group */
4312 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4315 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4316 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4317 * We must however clear busiest_stat::avg_load because
4318 * update_sd_pick_busiest() reads this before assignment.
4320 *sds = (struct sd_lb_stats){
4332 * get_sd_load_idx - Obtain the load index for a given sched domain.
4333 * @sd: The sched_domain whose load_idx is to be obtained.
4334 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4336 * Return: The load index.
4338 static inline int get_sd_load_idx(struct sched_domain *sd,
4339 enum cpu_idle_type idle)
4345 load_idx = sd->busy_idx;
4348 case CPU_NEWLY_IDLE:
4349 load_idx = sd->newidle_idx;
4352 load_idx = sd->idle_idx;
4359 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4361 return SCHED_POWER_SCALE;
4364 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4366 return default_scale_freq_power(sd, cpu);
4369 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4371 unsigned long weight = sd->span_weight;
4372 unsigned long smt_gain = sd->smt_gain;
4379 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4381 return default_scale_smt_power(sd, cpu);
4384 static unsigned long scale_rt_power(int cpu)
4386 struct rq *rq = cpu_rq(cpu);
4387 u64 total, available, age_stamp, avg;
4390 * Since we're reading these variables without serialization make sure
4391 * we read them once before doing sanity checks on them.
4393 age_stamp = ACCESS_ONCE(rq->age_stamp);
4394 avg = ACCESS_ONCE(rq->rt_avg);
4396 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4398 if (unlikely(total < avg)) {
4399 /* Ensures that power won't end up being negative */
4402 available = total - avg;
4405 if (unlikely((s64)total < SCHED_POWER_SCALE))
4406 total = SCHED_POWER_SCALE;
4408 total >>= SCHED_POWER_SHIFT;
4410 return div_u64(available, total);
4413 static void update_cpu_power(struct sched_domain *sd, int cpu)
4415 unsigned long weight = sd->span_weight;
4416 unsigned long power = SCHED_POWER_SCALE;
4417 struct sched_group *sdg = sd->groups;
4419 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4420 if (sched_feat(ARCH_POWER))
4421 power *= arch_scale_smt_power(sd, cpu);
4423 power *= default_scale_smt_power(sd, cpu);
4425 power >>= SCHED_POWER_SHIFT;
4428 sdg->sgp->power_orig = power;
4430 if (sched_feat(ARCH_POWER))
4431 power *= arch_scale_freq_power(sd, cpu);
4433 power *= default_scale_freq_power(sd, cpu);
4435 power >>= SCHED_POWER_SHIFT;
4437 power *= scale_rt_power(cpu);
4438 power >>= SCHED_POWER_SHIFT;
4443 cpu_rq(cpu)->cpu_power = power;
4444 sdg->sgp->power = power;
4447 void update_group_power(struct sched_domain *sd, int cpu)
4449 struct sched_domain *child = sd->child;
4450 struct sched_group *group, *sdg = sd->groups;
4451 unsigned long power, power_orig;
4452 unsigned long interval;
4454 interval = msecs_to_jiffies(sd->balance_interval);
4455 interval = clamp(interval, 1UL, max_load_balance_interval);
4456 sdg->sgp->next_update = jiffies + interval;
4459 update_cpu_power(sd, cpu);
4463 power_orig = power = 0;
4465 if (child->flags & SD_OVERLAP) {
4467 * SD_OVERLAP domains cannot assume that child groups
4468 * span the current group.
4471 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4472 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4474 power_orig += sg->sgp->power_orig;
4475 power += sg->sgp->power;
4479 * !SD_OVERLAP domains can assume that child groups
4480 * span the current group.
4483 group = child->groups;
4485 power_orig += group->sgp->power_orig;
4486 power += group->sgp->power;
4487 group = group->next;
4488 } while (group != child->groups);
4491 sdg->sgp->power_orig = power_orig;
4492 sdg->sgp->power = power;
4496 * Try and fix up capacity for tiny siblings, this is needed when
4497 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4498 * which on its own isn't powerful enough.
4500 * See update_sd_pick_busiest() and check_asym_packing().
4503 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4506 * Only siblings can have significantly less than SCHED_POWER_SCALE
4508 if (!(sd->flags & SD_SHARE_CPUPOWER))
4512 * If ~90% of the cpu_power is still there, we're good.
4514 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4521 * Group imbalance indicates (and tries to solve) the problem where balancing
4522 * groups is inadequate due to tsk_cpus_allowed() constraints.
4524 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4525 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4528 * { 0 1 2 3 } { 4 5 6 7 }
4531 * If we were to balance group-wise we'd place two tasks in the first group and
4532 * two tasks in the second group. Clearly this is undesired as it will overload
4533 * cpu 3 and leave one of the cpus in the second group unused.
4535 * The current solution to this issue is detecting the skew in the first group
4536 * by noticing the lower domain failed to reach balance and had difficulty
4537 * moving tasks due to affinity constraints.
4539 * When this is so detected; this group becomes a candidate for busiest; see
4540 * update_sd_pick_busiest(). And calculcate_imbalance() and
4541 * find_busiest_group() avoid some of the usual balance conditions to allow it
4542 * to create an effective group imbalance.
4544 * This is a somewhat tricky proposition since the next run might not find the
4545 * group imbalance and decide the groups need to be balanced again. A most
4546 * subtle and fragile situation.
4549 static inline int sg_imbalanced(struct sched_group *group)
4551 return group->sgp->imbalance;
4555 * Compute the group capacity.
4557 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4558 * first dividing out the smt factor and computing the actual number of cores
4559 * and limit power unit capacity with that.
4561 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4563 unsigned int capacity, smt, cpus;
4564 unsigned int power, power_orig;
4566 power = group->sgp->power;
4567 power_orig = group->sgp->power_orig;
4568 cpus = group->group_weight;
4570 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4571 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4572 capacity = cpus / smt; /* cores */
4574 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4576 capacity = fix_small_capacity(env->sd, group);
4582 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4583 * @env: The load balancing environment.
4584 * @group: sched_group whose statistics are to be updated.
4585 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4586 * @local_group: Does group contain this_cpu.
4587 * @sgs: variable to hold the statistics for this group.
4589 static inline void update_sg_lb_stats(struct lb_env *env,
4590 struct sched_group *group, int load_idx,
4591 int local_group, struct sg_lb_stats *sgs)
4593 unsigned long nr_running;
4597 memset(sgs, 0, sizeof(*sgs));
4599 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4600 struct rq *rq = cpu_rq(i);
4602 nr_running = rq->nr_running;
4604 /* Bias balancing toward cpus of our domain */
4606 load = target_load(i, load_idx);
4608 load = source_load(i, load_idx);
4610 sgs->group_load += load;
4611 sgs->sum_nr_running += nr_running;
4612 sgs->sum_weighted_load += weighted_cpuload(i);
4617 /* Adjust by relative CPU power of the group */
4618 sgs->group_power = group->sgp->power;
4619 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4621 if (sgs->sum_nr_running)
4622 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4624 sgs->group_weight = group->group_weight;
4626 sgs->group_imb = sg_imbalanced(group);
4627 sgs->group_capacity = sg_capacity(env, group);
4629 if (sgs->group_capacity > sgs->sum_nr_running)
4630 sgs->group_has_capacity = 1;
4634 * update_sd_pick_busiest - return 1 on busiest group
4635 * @env: The load balancing environment.
4636 * @sds: sched_domain statistics
4637 * @sg: sched_group candidate to be checked for being the busiest
4638 * @sgs: sched_group statistics
4640 * Determine if @sg is a busier group than the previously selected
4643 * Return: %true if @sg is a busier group than the previously selected
4644 * busiest group. %false otherwise.
4646 static bool update_sd_pick_busiest(struct lb_env *env,
4647 struct sd_lb_stats *sds,
4648 struct sched_group *sg,
4649 struct sg_lb_stats *sgs)
4651 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4654 if (sgs->sum_nr_running > sgs->group_capacity)
4661 * ASYM_PACKING needs to move all the work to the lowest
4662 * numbered CPUs in the group, therefore mark all groups
4663 * higher than ourself as busy.
4665 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4666 env->dst_cpu < group_first_cpu(sg)) {
4670 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4678 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4679 * @env: The load balancing environment.
4680 * @balance: Should we balance.
4681 * @sds: variable to hold the statistics for this sched_domain.
4683 static inline void update_sd_lb_stats(struct lb_env *env,
4684 struct sd_lb_stats *sds)
4686 struct sched_domain *child = env->sd->child;
4687 struct sched_group *sg = env->sd->groups;
4688 struct sg_lb_stats tmp_sgs;
4689 int load_idx, prefer_sibling = 0;
4691 if (child && child->flags & SD_PREFER_SIBLING)
4694 load_idx = get_sd_load_idx(env->sd, env->idle);
4697 struct sg_lb_stats *sgs = &tmp_sgs;
4700 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4703 sgs = &sds->local_stat;
4705 if (env->idle != CPU_NEWLY_IDLE ||
4706 time_after_eq(jiffies, sg->sgp->next_update))
4707 update_group_power(env->sd, env->dst_cpu);
4710 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4716 * In case the child domain prefers tasks go to siblings
4717 * first, lower the sg capacity to one so that we'll try
4718 * and move all the excess tasks away. We lower the capacity
4719 * of a group only if the local group has the capacity to fit
4720 * these excess tasks, i.e. nr_running < group_capacity. The
4721 * extra check prevents the case where you always pull from the
4722 * heaviest group when it is already under-utilized (possible
4723 * with a large weight task outweighs the tasks on the system).
4725 if (prefer_sibling && sds->local &&
4726 sds->local_stat.group_has_capacity)
4727 sgs->group_capacity = min(sgs->group_capacity, 1U);
4729 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4731 sds->busiest_stat = *sgs;
4735 /* Now, start updating sd_lb_stats */
4736 sds->total_load += sgs->group_load;
4737 sds->total_pwr += sgs->group_power;
4740 } while (sg != env->sd->groups);
4744 * check_asym_packing - Check to see if the group is packed into the
4747 * This is primarily intended to used at the sibling level. Some
4748 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4749 * case of POWER7, it can move to lower SMT modes only when higher
4750 * threads are idle. When in lower SMT modes, the threads will
4751 * perform better since they share less core resources. Hence when we
4752 * have idle threads, we want them to be the higher ones.
4754 * This packing function is run on idle threads. It checks to see if
4755 * the busiest CPU in this domain (core in the P7 case) has a higher
4756 * CPU number than the packing function is being run on. Here we are
4757 * assuming lower CPU number will be equivalent to lower a SMT thread
4760 * Return: 1 when packing is required and a task should be moved to
4761 * this CPU. The amount of the imbalance is returned in *imbalance.
4763 * @env: The load balancing environment.
4764 * @sds: Statistics of the sched_domain which is to be packed
4766 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4770 if (!(env->sd->flags & SD_ASYM_PACKING))
4776 busiest_cpu = group_first_cpu(sds->busiest);
4777 if (env->dst_cpu > busiest_cpu)
4780 env->imbalance = DIV_ROUND_CLOSEST(
4781 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
4788 * fix_small_imbalance - Calculate the minor imbalance that exists
4789 * amongst the groups of a sched_domain, during
4791 * @env: The load balancing environment.
4792 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4795 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4797 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4798 unsigned int imbn = 2;
4799 unsigned long scaled_busy_load_per_task;
4800 struct sg_lb_stats *local, *busiest;
4802 local = &sds->local_stat;
4803 busiest = &sds->busiest_stat;
4805 if (!local->sum_nr_running)
4806 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
4807 else if (busiest->load_per_task > local->load_per_task)
4810 scaled_busy_load_per_task =
4811 (busiest->load_per_task * SCHED_POWER_SCALE) /
4812 busiest->group_power;
4814 if (busiest->avg_load + scaled_busy_load_per_task >=
4815 local->avg_load + (scaled_busy_load_per_task * imbn)) {
4816 env->imbalance = busiest->load_per_task;
4821 * OK, we don't have enough imbalance to justify moving tasks,
4822 * however we may be able to increase total CPU power used by
4826 pwr_now += busiest->group_power *
4827 min(busiest->load_per_task, busiest->avg_load);
4828 pwr_now += local->group_power *
4829 min(local->load_per_task, local->avg_load);
4830 pwr_now /= SCHED_POWER_SCALE;
4832 /* Amount of load we'd subtract */
4833 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4834 busiest->group_power;
4835 if (busiest->avg_load > tmp) {
4836 pwr_move += busiest->group_power *
4837 min(busiest->load_per_task,
4838 busiest->avg_load - tmp);
4841 /* Amount of load we'd add */
4842 if (busiest->avg_load * busiest->group_power <
4843 busiest->load_per_task * SCHED_POWER_SCALE) {
4844 tmp = (busiest->avg_load * busiest->group_power) /
4847 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4850 pwr_move += local->group_power *
4851 min(local->load_per_task, local->avg_load + tmp);
4852 pwr_move /= SCHED_POWER_SCALE;
4854 /* Move if we gain throughput */
4855 if (pwr_move > pwr_now)
4856 env->imbalance = busiest->load_per_task;
4860 * calculate_imbalance - Calculate the amount of imbalance present within the
4861 * groups of a given sched_domain during load balance.
4862 * @env: load balance environment
4863 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4865 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4867 unsigned long max_pull, load_above_capacity = ~0UL;
4868 struct sg_lb_stats *local, *busiest;
4870 local = &sds->local_stat;
4871 busiest = &sds->busiest_stat;
4873 if (busiest->group_imb) {
4875 * In the group_imb case we cannot rely on group-wide averages
4876 * to ensure cpu-load equilibrium, look at wider averages. XXX
4878 busiest->load_per_task =
4879 min(busiest->load_per_task, sds->avg_load);
4883 * In the presence of smp nice balancing, certain scenarios can have
4884 * max load less than avg load(as we skip the groups at or below
4885 * its cpu_power, while calculating max_load..)
4887 if (busiest->avg_load <= sds->avg_load ||
4888 local->avg_load >= sds->avg_load) {
4890 return fix_small_imbalance(env, sds);
4893 if (!busiest->group_imb) {
4895 * Don't want to pull so many tasks that a group would go idle.
4896 * Except of course for the group_imb case, since then we might
4897 * have to drop below capacity to reach cpu-load equilibrium.
4899 load_above_capacity =
4900 (busiest->sum_nr_running - busiest->group_capacity);
4902 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4903 load_above_capacity /= busiest->group_power;
4907 * We're trying to get all the cpus to the average_load, so we don't
4908 * want to push ourselves above the average load, nor do we wish to
4909 * reduce the max loaded cpu below the average load. At the same time,
4910 * we also don't want to reduce the group load below the group capacity
4911 * (so that we can implement power-savings policies etc). Thus we look
4912 * for the minimum possible imbalance.
4914 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
4916 /* How much load to actually move to equalise the imbalance */
4917 env->imbalance = min(
4918 max_pull * busiest->group_power,
4919 (sds->avg_load - local->avg_load) * local->group_power
4920 ) / SCHED_POWER_SCALE;
4923 * if *imbalance is less than the average load per runnable task
4924 * there is no guarantee that any tasks will be moved so we'll have
4925 * a think about bumping its value to force at least one task to be
4928 if (env->imbalance < busiest->load_per_task)
4929 return fix_small_imbalance(env, sds);
4932 /******* find_busiest_group() helpers end here *********************/
4935 * find_busiest_group - Returns the busiest group within the sched_domain
4936 * if there is an imbalance. If there isn't an imbalance, and
4937 * the user has opted for power-savings, it returns a group whose
4938 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4939 * such a group exists.
4941 * Also calculates the amount of weighted load which should be moved
4942 * to restore balance.
4944 * @env: The load balancing environment.
4946 * Return: - The busiest group if imbalance exists.
4947 * - If no imbalance and user has opted for power-savings balance,
4948 * return the least loaded group whose CPUs can be
4949 * put to idle by rebalancing its tasks onto our group.
4951 static struct sched_group *find_busiest_group(struct lb_env *env)
4953 struct sg_lb_stats *local, *busiest;
4954 struct sd_lb_stats sds;
4956 init_sd_lb_stats(&sds);
4959 * Compute the various statistics relavent for load balancing at
4962 update_sd_lb_stats(env, &sds);
4963 local = &sds.local_stat;
4964 busiest = &sds.busiest_stat;
4966 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4967 check_asym_packing(env, &sds))
4970 /* There is no busy sibling group to pull tasks from */
4971 if (!sds.busiest || busiest->sum_nr_running == 0)
4974 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4977 * If the busiest group is imbalanced the below checks don't
4978 * work because they assume all things are equal, which typically
4979 * isn't true due to cpus_allowed constraints and the like.
4981 if (busiest->group_imb)
4984 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4985 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
4986 !busiest->group_has_capacity)
4990 * If the local group is more busy than the selected busiest group
4991 * don't try and pull any tasks.
4993 if (local->avg_load >= busiest->avg_load)
4997 * Don't pull any tasks if this group is already above the domain
5000 if (local->avg_load >= sds.avg_load)
5003 if (env->idle == CPU_IDLE) {
5005 * This cpu is idle. If the busiest group load doesn't
5006 * have more tasks than the number of available cpu's and
5007 * there is no imbalance between this and busiest group
5008 * wrt to idle cpu's, it is balanced.
5010 if ((local->idle_cpus < busiest->idle_cpus) &&
5011 busiest->sum_nr_running <= busiest->group_weight)
5015 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5016 * imbalance_pct to be conservative.
5018 if (100 * busiest->avg_load <=
5019 env->sd->imbalance_pct * local->avg_load)
5024 /* Looks like there is an imbalance. Compute it */
5025 calculate_imbalance(env, &sds);
5034 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5036 static struct rq *find_busiest_queue(struct lb_env *env,
5037 struct sched_group *group)
5039 struct rq *busiest = NULL, *rq;
5040 unsigned long busiest_load = 0, busiest_power = 1;
5043 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5044 unsigned long power = power_of(i);
5045 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5050 capacity = fix_small_capacity(env->sd, group);
5053 wl = weighted_cpuload(i);
5056 * When comparing with imbalance, use weighted_cpuload()
5057 * which is not scaled with the cpu power.
5059 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5063 * For the load comparisons with the other cpu's, consider
5064 * the weighted_cpuload() scaled with the cpu power, so that
5065 * the load can be moved away from the cpu that is potentially
5066 * running at a lower capacity.
5068 * Thus we're looking for max(wl_i / power_i), crosswise
5069 * multiplication to rid ourselves of the division works out
5070 * to: wl_i * power_j > wl_j * power_i; where j is our
5073 if (wl * busiest_power > busiest_load * power) {
5075 busiest_power = power;
5084 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5085 * so long as it is large enough.
5087 #define MAX_PINNED_INTERVAL 512
5089 /* Working cpumask for load_balance and load_balance_newidle. */
5090 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5092 static int need_active_balance(struct lb_env *env)
5094 struct sched_domain *sd = env->sd;
5096 if (env->idle == CPU_NEWLY_IDLE) {
5099 * ASYM_PACKING needs to force migrate tasks from busy but
5100 * higher numbered CPUs in order to pack all tasks in the
5101 * lowest numbered CPUs.
5103 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5107 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5110 static int active_load_balance_cpu_stop(void *data);
5112 static int should_we_balance(struct lb_env *env)
5114 struct sched_group *sg = env->sd->groups;
5115 struct cpumask *sg_cpus, *sg_mask;
5116 int cpu, balance_cpu = -1;
5119 * In the newly idle case, we will allow all the cpu's
5120 * to do the newly idle load balance.
5122 if (env->idle == CPU_NEWLY_IDLE)
5125 sg_cpus = sched_group_cpus(sg);
5126 sg_mask = sched_group_mask(sg);
5127 /* Try to find first idle cpu */
5128 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5129 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5136 if (balance_cpu == -1)
5137 balance_cpu = group_balance_cpu(sg);
5140 * First idle cpu or the first cpu(busiest) in this sched group
5141 * is eligible for doing load balancing at this and above domains.
5143 return balance_cpu == env->dst_cpu;
5147 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5148 * tasks if there is an imbalance.
5150 static int load_balance(int this_cpu, struct rq *this_rq,
5151 struct sched_domain *sd, enum cpu_idle_type idle,
5152 int *continue_balancing)
5154 int ld_moved, cur_ld_moved, active_balance = 0;
5155 struct sched_domain *sd_parent = sd->parent;
5156 struct sched_group *group;
5158 unsigned long flags;
5159 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5161 struct lb_env env = {
5163 .dst_cpu = this_cpu,
5165 .dst_grpmask = sched_group_cpus(sd->groups),
5167 .loop_break = sched_nr_migrate_break,
5172 * For NEWLY_IDLE load_balancing, we don't need to consider
5173 * other cpus in our group
5175 if (idle == CPU_NEWLY_IDLE)
5176 env.dst_grpmask = NULL;
5178 cpumask_copy(cpus, cpu_active_mask);
5180 schedstat_inc(sd, lb_count[idle]);
5183 if (!should_we_balance(&env)) {
5184 *continue_balancing = 0;
5188 group = find_busiest_group(&env);
5190 schedstat_inc(sd, lb_nobusyg[idle]);
5194 busiest = find_busiest_queue(&env, group);
5196 schedstat_inc(sd, lb_nobusyq[idle]);
5200 BUG_ON(busiest == env.dst_rq);
5202 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5205 if (busiest->nr_running > 1) {
5207 * Attempt to move tasks. If find_busiest_group has found
5208 * an imbalance but busiest->nr_running <= 1, the group is
5209 * still unbalanced. ld_moved simply stays zero, so it is
5210 * correctly treated as an imbalance.
5212 env.flags |= LBF_ALL_PINNED;
5213 env.src_cpu = busiest->cpu;
5214 env.src_rq = busiest;
5215 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5218 local_irq_save(flags);
5219 double_rq_lock(env.dst_rq, busiest);
5222 * cur_ld_moved - load moved in current iteration
5223 * ld_moved - cumulative load moved across iterations
5225 cur_ld_moved = move_tasks(&env);
5226 ld_moved += cur_ld_moved;
5227 double_rq_unlock(env.dst_rq, busiest);
5228 local_irq_restore(flags);
5231 * some other cpu did the load balance for us.
5233 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5234 resched_cpu(env.dst_cpu);
5236 if (env.flags & LBF_NEED_BREAK) {
5237 env.flags &= ~LBF_NEED_BREAK;
5242 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5243 * us and move them to an alternate dst_cpu in our sched_group
5244 * where they can run. The upper limit on how many times we
5245 * iterate on same src_cpu is dependent on number of cpus in our
5248 * This changes load balance semantics a bit on who can move
5249 * load to a given_cpu. In addition to the given_cpu itself
5250 * (or a ilb_cpu acting on its behalf where given_cpu is
5251 * nohz-idle), we now have balance_cpu in a position to move
5252 * load to given_cpu. In rare situations, this may cause
5253 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5254 * _independently_ and at _same_ time to move some load to
5255 * given_cpu) causing exceess load to be moved to given_cpu.
5256 * This however should not happen so much in practice and
5257 * moreover subsequent load balance cycles should correct the
5258 * excess load moved.
5260 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5262 /* Prevent to re-select dst_cpu via env's cpus */
5263 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5265 env.dst_rq = cpu_rq(env.new_dst_cpu);
5266 env.dst_cpu = env.new_dst_cpu;
5267 env.flags &= ~LBF_DST_PINNED;
5269 env.loop_break = sched_nr_migrate_break;
5272 * Go back to "more_balance" rather than "redo" since we
5273 * need to continue with same src_cpu.
5279 * We failed to reach balance because of affinity.
5282 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5284 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5285 *group_imbalance = 1;
5286 } else if (*group_imbalance)
5287 *group_imbalance = 0;
5290 /* All tasks on this runqueue were pinned by CPU affinity */
5291 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5292 cpumask_clear_cpu(cpu_of(busiest), cpus);
5293 if (!cpumask_empty(cpus)) {
5295 env.loop_break = sched_nr_migrate_break;
5303 schedstat_inc(sd, lb_failed[idle]);
5305 * Increment the failure counter only on periodic balance.
5306 * We do not want newidle balance, which can be very
5307 * frequent, pollute the failure counter causing
5308 * excessive cache_hot migrations and active balances.
5310 if (idle != CPU_NEWLY_IDLE)
5311 sd->nr_balance_failed++;
5313 if (need_active_balance(&env)) {
5314 raw_spin_lock_irqsave(&busiest->lock, flags);
5316 /* don't kick the active_load_balance_cpu_stop,
5317 * if the curr task on busiest cpu can't be
5320 if (!cpumask_test_cpu(this_cpu,
5321 tsk_cpus_allowed(busiest->curr))) {
5322 raw_spin_unlock_irqrestore(&busiest->lock,
5324 env.flags |= LBF_ALL_PINNED;
5325 goto out_one_pinned;
5329 * ->active_balance synchronizes accesses to
5330 * ->active_balance_work. Once set, it's cleared
5331 * only after active load balance is finished.
5333 if (!busiest->active_balance) {
5334 busiest->active_balance = 1;
5335 busiest->push_cpu = this_cpu;
5338 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5340 if (active_balance) {
5341 stop_one_cpu_nowait(cpu_of(busiest),
5342 active_load_balance_cpu_stop, busiest,
5343 &busiest->active_balance_work);
5347 * We've kicked active balancing, reset the failure
5350 sd->nr_balance_failed = sd->cache_nice_tries+1;
5353 sd->nr_balance_failed = 0;
5355 if (likely(!active_balance)) {
5356 /* We were unbalanced, so reset the balancing interval */
5357 sd->balance_interval = sd->min_interval;
5360 * If we've begun active balancing, start to back off. This
5361 * case may not be covered by the all_pinned logic if there
5362 * is only 1 task on the busy runqueue (because we don't call
5365 if (sd->balance_interval < sd->max_interval)
5366 sd->balance_interval *= 2;
5372 schedstat_inc(sd, lb_balanced[idle]);
5374 sd->nr_balance_failed = 0;
5377 /* tune up the balancing interval */
5378 if (((env.flags & LBF_ALL_PINNED) &&
5379 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5380 (sd->balance_interval < sd->max_interval))
5381 sd->balance_interval *= 2;
5389 * idle_balance is called by schedule() if this_cpu is about to become
5390 * idle. Attempts to pull tasks from other CPUs.
5392 void idle_balance(int this_cpu, struct rq *this_rq)
5394 struct sched_domain *sd;
5395 int pulled_task = 0;
5396 unsigned long next_balance = jiffies + HZ;
5399 this_rq->idle_stamp = rq_clock(this_rq);
5401 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5405 * Drop the rq->lock, but keep IRQ/preempt disabled.
5407 raw_spin_unlock(&this_rq->lock);
5409 update_blocked_averages(this_cpu);
5411 for_each_domain(this_cpu, sd) {
5412 unsigned long interval;
5413 int continue_balancing = 1;
5414 u64 t0, domain_cost;
5416 if (!(sd->flags & SD_LOAD_BALANCE))
5419 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5422 if (sd->flags & SD_BALANCE_NEWIDLE) {
5423 t0 = sched_clock_cpu(this_cpu);
5425 /* If we've pulled tasks over stop searching: */
5426 pulled_task = load_balance(this_cpu, this_rq,
5428 &continue_balancing);
5430 domain_cost = sched_clock_cpu(this_cpu) - t0;
5431 if (domain_cost > sd->max_newidle_lb_cost)
5432 sd->max_newidle_lb_cost = domain_cost;
5434 curr_cost += domain_cost;
5437 interval = msecs_to_jiffies(sd->balance_interval);
5438 if (time_after(next_balance, sd->last_balance + interval))
5439 next_balance = sd->last_balance + interval;
5441 this_rq->idle_stamp = 0;
5447 raw_spin_lock(&this_rq->lock);
5449 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5451 * We are going idle. next_balance may be set based on
5452 * a busy processor. So reset next_balance.
5454 this_rq->next_balance = next_balance;
5457 if (curr_cost > this_rq->max_idle_balance_cost)
5458 this_rq->max_idle_balance_cost = curr_cost;
5462 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5463 * running tasks off the busiest CPU onto idle CPUs. It requires at
5464 * least 1 task to be running on each physical CPU where possible, and
5465 * avoids physical / logical imbalances.
5467 static int active_load_balance_cpu_stop(void *data)
5469 struct rq *busiest_rq = data;
5470 int busiest_cpu = cpu_of(busiest_rq);
5471 int target_cpu = busiest_rq->push_cpu;
5472 struct rq *target_rq = cpu_rq(target_cpu);
5473 struct sched_domain *sd;
5475 raw_spin_lock_irq(&busiest_rq->lock);
5477 /* make sure the requested cpu hasn't gone down in the meantime */
5478 if (unlikely(busiest_cpu != smp_processor_id() ||
5479 !busiest_rq->active_balance))
5482 /* Is there any task to move? */
5483 if (busiest_rq->nr_running <= 1)
5487 * This condition is "impossible", if it occurs
5488 * we need to fix it. Originally reported by
5489 * Bjorn Helgaas on a 128-cpu setup.
5491 BUG_ON(busiest_rq == target_rq);
5493 /* move a task from busiest_rq to target_rq */
5494 double_lock_balance(busiest_rq, target_rq);
5496 /* Search for an sd spanning us and the target CPU. */
5498 for_each_domain(target_cpu, sd) {
5499 if ((sd->flags & SD_LOAD_BALANCE) &&
5500 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5505 struct lb_env env = {
5507 .dst_cpu = target_cpu,
5508 .dst_rq = target_rq,
5509 .src_cpu = busiest_rq->cpu,
5510 .src_rq = busiest_rq,
5514 schedstat_inc(sd, alb_count);
5516 if (move_one_task(&env))
5517 schedstat_inc(sd, alb_pushed);
5519 schedstat_inc(sd, alb_failed);
5522 double_unlock_balance(busiest_rq, target_rq);
5524 busiest_rq->active_balance = 0;
5525 raw_spin_unlock_irq(&busiest_rq->lock);
5529 #ifdef CONFIG_NO_HZ_COMMON
5531 * idle load balancing details
5532 * - When one of the busy CPUs notice that there may be an idle rebalancing
5533 * needed, they will kick the idle load balancer, which then does idle
5534 * load balancing for all the idle CPUs.
5537 cpumask_var_t idle_cpus_mask;
5539 unsigned long next_balance; /* in jiffy units */
5540 } nohz ____cacheline_aligned;
5542 static inline int find_new_ilb(int call_cpu)
5544 int ilb = cpumask_first(nohz.idle_cpus_mask);
5546 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5553 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5554 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5555 * CPU (if there is one).
5557 static void nohz_balancer_kick(int cpu)
5561 nohz.next_balance++;
5563 ilb_cpu = find_new_ilb(cpu);
5565 if (ilb_cpu >= nr_cpu_ids)
5568 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5571 * Use smp_send_reschedule() instead of resched_cpu().
5572 * This way we generate a sched IPI on the target cpu which
5573 * is idle. And the softirq performing nohz idle load balance
5574 * will be run before returning from the IPI.
5576 smp_send_reschedule(ilb_cpu);
5580 static inline void nohz_balance_exit_idle(int cpu)
5582 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5583 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5584 atomic_dec(&nohz.nr_cpus);
5585 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5589 static inline void set_cpu_sd_state_busy(void)
5591 struct sched_domain *sd;
5594 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5596 if (!sd || !sd->nohz_idle)
5600 for (; sd; sd = sd->parent)
5601 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5606 void set_cpu_sd_state_idle(void)
5608 struct sched_domain *sd;
5611 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5613 if (!sd || sd->nohz_idle)
5617 for (; sd; sd = sd->parent)
5618 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5624 * This routine will record that the cpu is going idle with tick stopped.
5625 * This info will be used in performing idle load balancing in the future.
5627 void nohz_balance_enter_idle(int cpu)
5630 * If this cpu is going down, then nothing needs to be done.
5632 if (!cpu_active(cpu))
5635 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5638 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5639 atomic_inc(&nohz.nr_cpus);
5640 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5643 static int sched_ilb_notifier(struct notifier_block *nfb,
5644 unsigned long action, void *hcpu)
5646 switch (action & ~CPU_TASKS_FROZEN) {
5648 nohz_balance_exit_idle(smp_processor_id());
5656 static DEFINE_SPINLOCK(balancing);
5659 * Scale the max load_balance interval with the number of CPUs in the system.
5660 * This trades load-balance latency on larger machines for less cross talk.
5662 void update_max_interval(void)
5664 max_load_balance_interval = HZ*num_online_cpus()/10;
5668 * It checks each scheduling domain to see if it is due to be balanced,
5669 * and initiates a balancing operation if so.
5671 * Balancing parameters are set up in init_sched_domains.
5673 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5675 int continue_balancing = 1;
5676 struct rq *rq = cpu_rq(cpu);
5677 unsigned long interval;
5678 struct sched_domain *sd;
5679 /* Earliest time when we have to do rebalance again */
5680 unsigned long next_balance = jiffies + 60*HZ;
5681 int update_next_balance = 0;
5682 int need_serialize, need_decay = 0;
5685 update_blocked_averages(cpu);
5688 for_each_domain(cpu, sd) {
5690 * Decay the newidle max times here because this is a regular
5691 * visit to all the domains. Decay ~1% per second.
5693 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5694 sd->max_newidle_lb_cost =
5695 (sd->max_newidle_lb_cost * 253) / 256;
5696 sd->next_decay_max_lb_cost = jiffies + HZ;
5699 max_cost += sd->max_newidle_lb_cost;
5701 if (!(sd->flags & SD_LOAD_BALANCE))
5705 * Stop the load balance at this level. There is another
5706 * CPU in our sched group which is doing load balancing more
5709 if (!continue_balancing) {
5715 interval = sd->balance_interval;
5716 if (idle != CPU_IDLE)
5717 interval *= sd->busy_factor;
5719 /* scale ms to jiffies */
5720 interval = msecs_to_jiffies(interval);
5721 interval = clamp(interval, 1UL, max_load_balance_interval);
5723 need_serialize = sd->flags & SD_SERIALIZE;
5725 if (need_serialize) {
5726 if (!spin_trylock(&balancing))
5730 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5731 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5733 * The LBF_DST_PINNED logic could have changed
5734 * env->dst_cpu, so we can't know our idle
5735 * state even if we migrated tasks. Update it.
5737 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5739 sd->last_balance = jiffies;
5742 spin_unlock(&balancing);
5744 if (time_after(next_balance, sd->last_balance + interval)) {
5745 next_balance = sd->last_balance + interval;
5746 update_next_balance = 1;
5751 * Ensure the rq-wide value also decays but keep it at a
5752 * reasonable floor to avoid funnies with rq->avg_idle.
5754 rq->max_idle_balance_cost =
5755 max((u64)sysctl_sched_migration_cost, max_cost);
5760 * next_balance will be updated only when there is a need.
5761 * When the cpu is attached to null domain for ex, it will not be
5764 if (likely(update_next_balance))
5765 rq->next_balance = next_balance;
5768 #ifdef CONFIG_NO_HZ_COMMON
5770 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5771 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5773 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5775 struct rq *this_rq = cpu_rq(this_cpu);
5779 if (idle != CPU_IDLE ||
5780 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5783 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5784 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5788 * If this cpu gets work to do, stop the load balancing
5789 * work being done for other cpus. Next load
5790 * balancing owner will pick it up.
5795 rq = cpu_rq(balance_cpu);
5797 raw_spin_lock_irq(&rq->lock);
5798 update_rq_clock(rq);
5799 update_idle_cpu_load(rq);
5800 raw_spin_unlock_irq(&rq->lock);
5802 rebalance_domains(balance_cpu, CPU_IDLE);
5804 if (time_after(this_rq->next_balance, rq->next_balance))
5805 this_rq->next_balance = rq->next_balance;
5807 nohz.next_balance = this_rq->next_balance;
5809 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5813 * Current heuristic for kicking the idle load balancer in the presence
5814 * of an idle cpu is the system.
5815 * - This rq has more than one task.
5816 * - At any scheduler domain level, this cpu's scheduler group has multiple
5817 * busy cpu's exceeding the group's power.
5818 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5819 * domain span are idle.
5821 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5823 unsigned long now = jiffies;
5824 struct sched_domain *sd;
5826 if (unlikely(idle_cpu(cpu)))
5830 * We may be recently in ticked or tickless idle mode. At the first
5831 * busy tick after returning from idle, we will update the busy stats.
5833 set_cpu_sd_state_busy();
5834 nohz_balance_exit_idle(cpu);
5837 * None are in tickless mode and hence no need for NOHZ idle load
5840 if (likely(!atomic_read(&nohz.nr_cpus)))
5843 if (time_before(now, nohz.next_balance))
5846 if (rq->nr_running >= 2)
5850 for_each_domain(cpu, sd) {
5851 struct sched_group *sg = sd->groups;
5852 struct sched_group_power *sgp = sg->sgp;
5853 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5855 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5856 goto need_kick_unlock;
5858 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5859 && (cpumask_first_and(nohz.idle_cpus_mask,
5860 sched_domain_span(sd)) < cpu))
5861 goto need_kick_unlock;
5863 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5875 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5879 * run_rebalance_domains is triggered when needed from the scheduler tick.
5880 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5882 static void run_rebalance_domains(struct softirq_action *h)
5884 int this_cpu = smp_processor_id();
5885 struct rq *this_rq = cpu_rq(this_cpu);
5886 enum cpu_idle_type idle = this_rq->idle_balance ?
5887 CPU_IDLE : CPU_NOT_IDLE;
5889 rebalance_domains(this_cpu, idle);
5892 * If this cpu has a pending nohz_balance_kick, then do the
5893 * balancing on behalf of the other idle cpus whose ticks are
5896 nohz_idle_balance(this_cpu, idle);
5899 static inline int on_null_domain(int cpu)
5901 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5905 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5907 void trigger_load_balance(struct rq *rq, int cpu)
5909 /* Don't need to rebalance while attached to NULL domain */
5910 if (time_after_eq(jiffies, rq->next_balance) &&
5911 likely(!on_null_domain(cpu)))
5912 raise_softirq(SCHED_SOFTIRQ);
5913 #ifdef CONFIG_NO_HZ_COMMON
5914 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5915 nohz_balancer_kick(cpu);
5919 static void rq_online_fair(struct rq *rq)
5924 static void rq_offline_fair(struct rq *rq)
5928 /* Ensure any throttled groups are reachable by pick_next_task */
5929 unthrottle_offline_cfs_rqs(rq);
5932 #endif /* CONFIG_SMP */
5935 * scheduler tick hitting a task of our scheduling class:
5937 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5939 struct cfs_rq *cfs_rq;
5940 struct sched_entity *se = &curr->se;
5942 for_each_sched_entity(se) {
5943 cfs_rq = cfs_rq_of(se);
5944 entity_tick(cfs_rq, se, queued);
5947 if (numabalancing_enabled)
5948 task_tick_numa(rq, curr);
5950 update_rq_runnable_avg(rq, 1);
5954 * called on fork with the child task as argument from the parent's context
5955 * - child not yet on the tasklist
5956 * - preemption disabled
5958 static void task_fork_fair(struct task_struct *p)
5960 struct cfs_rq *cfs_rq;
5961 struct sched_entity *se = &p->se, *curr;
5962 int this_cpu = smp_processor_id();
5963 struct rq *rq = this_rq();
5964 unsigned long flags;
5966 raw_spin_lock_irqsave(&rq->lock, flags);
5968 update_rq_clock(rq);
5970 cfs_rq = task_cfs_rq(current);
5971 curr = cfs_rq->curr;
5974 * Not only the cpu but also the task_group of the parent might have
5975 * been changed after parent->se.parent,cfs_rq were copied to
5976 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
5977 * of child point to valid ones.
5980 __set_task_cpu(p, this_cpu);
5983 update_curr(cfs_rq);
5986 se->vruntime = curr->vruntime;
5987 place_entity(cfs_rq, se, 1);
5989 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5991 * Upon rescheduling, sched_class::put_prev_task() will place
5992 * 'current' within the tree based on its new key value.
5994 swap(curr->vruntime, se->vruntime);
5995 resched_task(rq->curr);
5998 se->vruntime -= cfs_rq->min_vruntime;
6000 raw_spin_unlock_irqrestore(&rq->lock, flags);
6004 * Priority of the task has changed. Check to see if we preempt
6008 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6014 * Reschedule if we are currently running on this runqueue and
6015 * our priority decreased, or if we are not currently running on
6016 * this runqueue and our priority is higher than the current's
6018 if (rq->curr == p) {
6019 if (p->prio > oldprio)
6020 resched_task(rq->curr);
6022 check_preempt_curr(rq, p, 0);
6025 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6027 struct sched_entity *se = &p->se;
6028 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6031 * Ensure the task's vruntime is normalized, so that when its
6032 * switched back to the fair class the enqueue_entity(.flags=0) will
6033 * do the right thing.
6035 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6036 * have normalized the vruntime, if it was !on_rq, then only when
6037 * the task is sleeping will it still have non-normalized vruntime.
6039 if (!se->on_rq && p->state != TASK_RUNNING) {
6041 * Fix up our vruntime so that the current sleep doesn't
6042 * cause 'unlimited' sleep bonus.
6044 place_entity(cfs_rq, se, 0);
6045 se->vruntime -= cfs_rq->min_vruntime;
6050 * Remove our load from contribution when we leave sched_fair
6051 * and ensure we don't carry in an old decay_count if we
6054 if (se->avg.decay_count) {
6055 __synchronize_entity_decay(se);
6056 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6062 * We switched to the sched_fair class.
6064 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6070 * We were most likely switched from sched_rt, so
6071 * kick off the schedule if running, otherwise just see
6072 * if we can still preempt the current task.
6075 resched_task(rq->curr);
6077 check_preempt_curr(rq, p, 0);
6080 /* Account for a task changing its policy or group.
6082 * This routine is mostly called to set cfs_rq->curr field when a task
6083 * migrates between groups/classes.
6085 static void set_curr_task_fair(struct rq *rq)
6087 struct sched_entity *se = &rq->curr->se;
6089 for_each_sched_entity(se) {
6090 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6092 set_next_entity(cfs_rq, se);
6093 /* ensure bandwidth has been allocated on our new cfs_rq */
6094 account_cfs_rq_runtime(cfs_rq, 0);
6098 void init_cfs_rq(struct cfs_rq *cfs_rq)
6100 cfs_rq->tasks_timeline = RB_ROOT;
6101 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6102 #ifndef CONFIG_64BIT
6103 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6106 atomic64_set(&cfs_rq->decay_counter, 1);
6107 atomic_long_set(&cfs_rq->removed_load, 0);
6111 #ifdef CONFIG_FAIR_GROUP_SCHED
6112 static void task_move_group_fair(struct task_struct *p, int on_rq)
6114 struct cfs_rq *cfs_rq;
6116 * If the task was not on the rq at the time of this cgroup movement
6117 * it must have been asleep, sleeping tasks keep their ->vruntime
6118 * absolute on their old rq until wakeup (needed for the fair sleeper
6119 * bonus in place_entity()).
6121 * If it was on the rq, we've just 'preempted' it, which does convert
6122 * ->vruntime to a relative base.
6124 * Make sure both cases convert their relative position when migrating
6125 * to another cgroup's rq. This does somewhat interfere with the
6126 * fair sleeper stuff for the first placement, but who cares.
6129 * When !on_rq, vruntime of the task has usually NOT been normalized.
6130 * But there are some cases where it has already been normalized:
6132 * - Moving a forked child which is waiting for being woken up by
6133 * wake_up_new_task().
6134 * - Moving a task which has been woken up by try_to_wake_up() and
6135 * waiting for actually being woken up by sched_ttwu_pending().
6137 * To prevent boost or penalty in the new cfs_rq caused by delta
6138 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6140 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6144 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6145 set_task_rq(p, task_cpu(p));
6147 cfs_rq = cfs_rq_of(&p->se);
6148 p->se.vruntime += cfs_rq->min_vruntime;
6151 * migrate_task_rq_fair() will have removed our previous
6152 * contribution, but we must synchronize for ongoing future
6155 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6156 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6161 void free_fair_sched_group(struct task_group *tg)
6165 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6167 for_each_possible_cpu(i) {
6169 kfree(tg->cfs_rq[i]);
6178 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6180 struct cfs_rq *cfs_rq;
6181 struct sched_entity *se;
6184 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6187 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6191 tg->shares = NICE_0_LOAD;
6193 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6195 for_each_possible_cpu(i) {
6196 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6197 GFP_KERNEL, cpu_to_node(i));
6201 se = kzalloc_node(sizeof(struct sched_entity),
6202 GFP_KERNEL, cpu_to_node(i));
6206 init_cfs_rq(cfs_rq);
6207 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6218 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6220 struct rq *rq = cpu_rq(cpu);
6221 unsigned long flags;
6224 * Only empty task groups can be destroyed; so we can speculatively
6225 * check on_list without danger of it being re-added.
6227 if (!tg->cfs_rq[cpu]->on_list)
6230 raw_spin_lock_irqsave(&rq->lock, flags);
6231 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6232 raw_spin_unlock_irqrestore(&rq->lock, flags);
6235 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6236 struct sched_entity *se, int cpu,
6237 struct sched_entity *parent)
6239 struct rq *rq = cpu_rq(cpu);
6243 init_cfs_rq_runtime(cfs_rq);
6245 tg->cfs_rq[cpu] = cfs_rq;
6248 /* se could be NULL for root_task_group */
6253 se->cfs_rq = &rq->cfs;
6255 se->cfs_rq = parent->my_q;
6258 update_load_set(&se->load, 0);
6259 se->parent = parent;
6262 static DEFINE_MUTEX(shares_mutex);
6264 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6267 unsigned long flags;
6270 * We can't change the weight of the root cgroup.
6275 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6277 mutex_lock(&shares_mutex);
6278 if (tg->shares == shares)
6281 tg->shares = shares;
6282 for_each_possible_cpu(i) {
6283 struct rq *rq = cpu_rq(i);
6284 struct sched_entity *se;
6287 /* Propagate contribution to hierarchy */
6288 raw_spin_lock_irqsave(&rq->lock, flags);
6290 /* Possible calls to update_curr() need rq clock */
6291 update_rq_clock(rq);
6292 for_each_sched_entity(se)
6293 update_cfs_shares(group_cfs_rq(se));
6294 raw_spin_unlock_irqrestore(&rq->lock, flags);
6298 mutex_unlock(&shares_mutex);
6301 #else /* CONFIG_FAIR_GROUP_SCHED */
6303 void free_fair_sched_group(struct task_group *tg) { }
6305 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6310 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6312 #endif /* CONFIG_FAIR_GROUP_SCHED */
6315 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6317 struct sched_entity *se = &task->se;
6318 unsigned int rr_interval = 0;
6321 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6324 if (rq->cfs.load.weight)
6325 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6331 * All the scheduling class methods:
6333 const struct sched_class fair_sched_class = {
6334 .next = &idle_sched_class,
6335 .enqueue_task = enqueue_task_fair,
6336 .dequeue_task = dequeue_task_fair,
6337 .yield_task = yield_task_fair,
6338 .yield_to_task = yield_to_task_fair,
6340 .check_preempt_curr = check_preempt_wakeup,
6342 .pick_next_task = pick_next_task_fair,
6343 .put_prev_task = put_prev_task_fair,
6346 .select_task_rq = select_task_rq_fair,
6347 .migrate_task_rq = migrate_task_rq_fair,
6349 .rq_online = rq_online_fair,
6350 .rq_offline = rq_offline_fair,
6352 .task_waking = task_waking_fair,
6355 .set_curr_task = set_curr_task_fair,
6356 .task_tick = task_tick_fair,
6357 .task_fork = task_fork_fair,
6359 .prio_changed = prio_changed_fair,
6360 .switched_from = switched_from_fair,
6361 .switched_to = switched_to_fair,
6363 .get_rr_interval = get_rr_interval_fair,
6365 #ifdef CONFIG_FAIR_GROUP_SCHED
6366 .task_move_group = task_move_group_fair,
6370 #ifdef CONFIG_SCHED_DEBUG
6371 void print_cfs_stats(struct seq_file *m, int cpu)
6373 struct cfs_rq *cfs_rq;
6376 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6377 print_cfs_rq(m, cpu, cfs_rq);
6382 __init void init_sched_fair_class(void)
6385 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6387 #ifdef CONFIG_NO_HZ_COMMON
6388 nohz.next_balance = jiffies;
6389 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6390 cpu_notifier(sched_ilb_notifier, 0);