2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
199 struct load_weight *lw)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
209 tmp = (u64)delta_exec * scale_load_down(weight);
211 tmp = (u64)delta_exec;
213 if (!lw->inv_weight) {
214 unsigned long w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp > WMULT_CONST))
228 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
231 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
233 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
237 const struct sched_class fair_sched_class;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct *task_of(struct sched_entity *se)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se));
259 return container_of(se, struct task_struct, se);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
283 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
312 if (cfs_rq->on_list) {
313 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity *se, struct sched_entity *pse)
326 if (se->cfs_rq == pse->cfs_rq)
332 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity *se)
342 for_each_sched_entity(se)
349 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
351 int se_depth, pse_depth;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth = depth_se(*se);
362 pse_depth = depth_se(*pse);
364 while (se_depth > pse_depth) {
366 *se = parent_entity(*se);
369 while (pse_depth > se_depth) {
371 *pse = parent_entity(*pse);
374 while (!is_same_group(*se, *pse)) {
375 *se = parent_entity(*se);
376 *pse = parent_entity(*pse);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct *task_of(struct sched_entity *se)
384 return container_of(se, struct task_struct, se);
387 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
389 return container_of(cfs_rq, struct rq, cfs);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
399 return &task_rq(p)->cfs;
402 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
404 struct task_struct *p = task_of(se);
405 struct rq *rq = task_rq(p);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity *se, struct sched_entity *pse)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta, struct sched_entity *se)
618 if (unlikely(se->load.weight != NICE_0_LOAD))
619 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64 __sched_period(unsigned long nr_running)
634 u64 period = sysctl_sched_latency;
635 unsigned long nr_latency = sched_nr_latency;
637 if (unlikely(nr_running > nr_latency)) {
638 period = sysctl_sched_min_granularity;
639 period *= nr_running;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
655 for_each_sched_entity(se) {
656 struct load_weight *load;
657 struct load_weight lw;
659 cfs_rq = cfs_rq_of(se);
660 load = &cfs_rq->load;
662 if (unlikely(!se->on_rq)) {
665 update_load_add(&lw, se->load.weight);
668 slice = calc_delta_mine(slice, se->load.weight, load);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
680 return calc_delta_fair(sched_slice(cfs_rq, se), se);
684 static inline void __update_task_entity_contrib(struct sched_entity *se);
686 /* Give new task start runnable values to heavy its load in infant time */
687 void init_task_runnable_average(struct task_struct *p)
691 p->se.avg.decay_count = 0;
692 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
693 p->se.avg.runnable_avg_sum = slice;
694 p->se.avg.runnable_avg_period = slice;
695 __update_task_entity_contrib(&p->se);
698 void init_task_runnable_average(struct task_struct *p)
704 * Update the current task's runtime statistics. Skip current tasks that
705 * are not in our scheduling class.
708 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
709 unsigned long delta_exec)
711 unsigned long delta_exec_weighted;
713 schedstat_set(curr->statistics.exec_max,
714 max((u64)delta_exec, curr->statistics.exec_max));
716 curr->sum_exec_runtime += delta_exec;
717 schedstat_add(cfs_rq, exec_clock, delta_exec);
718 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
720 curr->vruntime += delta_exec_weighted;
721 update_min_vruntime(cfs_rq);
724 static void update_curr(struct cfs_rq *cfs_rq)
726 struct sched_entity *curr = cfs_rq->curr;
727 u64 now = rq_clock_task(rq_of(cfs_rq));
728 unsigned long delta_exec;
734 * Get the amount of time the current task was running
735 * since the last time we changed load (this cannot
736 * overflow on 32 bits):
738 delta_exec = (unsigned long)(now - curr->exec_start);
742 __update_curr(cfs_rq, curr, delta_exec);
743 curr->exec_start = now;
745 if (entity_is_task(curr)) {
746 struct task_struct *curtask = task_of(curr);
748 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
749 cpuacct_charge(curtask, delta_exec);
750 account_group_exec_runtime(curtask, delta_exec);
753 account_cfs_rq_runtime(cfs_rq, delta_exec);
757 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
759 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
763 * Task is being enqueued - update stats:
765 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
768 * Are we enqueueing a waiting task? (for current tasks
769 * a dequeue/enqueue event is a NOP)
771 if (se != cfs_rq->curr)
772 update_stats_wait_start(cfs_rq, se);
776 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
778 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
779 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
780 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
781 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
782 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
783 #ifdef CONFIG_SCHEDSTATS
784 if (entity_is_task(se)) {
785 trace_sched_stat_wait(task_of(se),
786 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
789 schedstat_set(se->statistics.wait_start, 0);
793 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
796 * Mark the end of the wait period if dequeueing a
799 if (se != cfs_rq->curr)
800 update_stats_wait_end(cfs_rq, se);
804 * We are picking a new current task - update its stats:
807 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
810 * We are starting a new run period:
812 se->exec_start = rq_clock_task(rq_of(cfs_rq));
815 /**************************************************
816 * Scheduling class queueing methods:
819 #ifdef CONFIG_NUMA_BALANCING
821 * numa task sample period in ms
823 unsigned int sysctl_numa_balancing_scan_period_min = 100;
824 unsigned int sysctl_numa_balancing_scan_period_max = 100*50;
825 unsigned int sysctl_numa_balancing_scan_period_reset = 100*600;
827 /* Portion of address space to scan in MB */
828 unsigned int sysctl_numa_balancing_scan_size = 256;
830 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
831 unsigned int sysctl_numa_balancing_scan_delay = 1000;
833 static void task_numa_placement(struct task_struct *p)
837 if (!p->mm) /* for example, ksmd faulting in a user's mm */
839 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
840 if (p->numa_scan_seq == seq)
842 p->numa_scan_seq = seq;
844 /* FIXME: Scheduling placement policy hints go here */
848 * Got a PROT_NONE fault for a page on @node.
850 void task_numa_fault(int node, int pages, bool migrated)
852 struct task_struct *p = current;
854 if (!numabalancing_enabled)
857 /* FIXME: Allocate task-specific structure for placement policy here */
860 * If pages are properly placed (did not migrate) then scan slower.
861 * This is reset periodically in case of phase changes
864 p->numa_scan_period = min(sysctl_numa_balancing_scan_period_max,
865 p->numa_scan_period + jiffies_to_msecs(10));
867 task_numa_placement(p);
870 static void reset_ptenuma_scan(struct task_struct *p)
872 ACCESS_ONCE(p->mm->numa_scan_seq)++;
873 p->mm->numa_scan_offset = 0;
877 * The expensive part of numa migration is done from task_work context.
878 * Triggered from task_tick_numa().
880 void task_numa_work(struct callback_head *work)
882 unsigned long migrate, next_scan, now = jiffies;
883 struct task_struct *p = current;
884 struct mm_struct *mm = p->mm;
885 struct vm_area_struct *vma;
886 unsigned long start, end;
889 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
891 work->next = work; /* protect against double add */
893 * Who cares about NUMA placement when they're dying.
895 * NOTE: make sure not to dereference p->mm before this check,
896 * exit_task_work() happens _after_ exit_mm() so we could be called
897 * without p->mm even though we still had it when we enqueued this
900 if (p->flags & PF_EXITING)
904 * We do not care about task placement until a task runs on a node
905 * other than the first one used by the address space. This is
906 * largely because migrations are driven by what CPU the task
907 * is running on. If it's never scheduled on another node, it'll
908 * not migrate so why bother trapping the fault.
910 if (mm->first_nid == NUMA_PTE_SCAN_INIT)
911 mm->first_nid = numa_node_id();
912 if (mm->first_nid != NUMA_PTE_SCAN_ACTIVE) {
913 /* Are we running on a new node yet? */
914 if (numa_node_id() == mm->first_nid &&
915 !sched_feat_numa(NUMA_FORCE))
918 mm->first_nid = NUMA_PTE_SCAN_ACTIVE;
922 * Reset the scan period if enough time has gone by. Objective is that
923 * scanning will be reduced if pages are properly placed. As tasks
924 * can enter different phases this needs to be re-examined. Lacking
925 * proper tracking of reference behaviour, this blunt hammer is used.
927 migrate = mm->numa_next_reset;
928 if (time_after(now, migrate)) {
929 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
930 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
931 xchg(&mm->numa_next_reset, next_scan);
935 * Enforce maximal scan/migration frequency..
937 migrate = mm->numa_next_scan;
938 if (time_before(now, migrate))
941 if (p->numa_scan_period == 0)
942 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
944 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
945 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
949 * Do not set pte_numa if the current running node is rate-limited.
950 * This loses statistics on the fault but if we are unwilling to
951 * migrate to this node, it is less likely we can do useful work
953 if (migrate_ratelimited(numa_node_id()))
956 start = mm->numa_scan_offset;
957 pages = sysctl_numa_balancing_scan_size;
958 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
962 down_read(&mm->mmap_sem);
963 vma = find_vma(mm, start);
965 reset_ptenuma_scan(p);
969 for (; vma; vma = vma->vm_next) {
970 if (!vma_migratable(vma))
973 /* Skip small VMAs. They are not likely to be of relevance */
974 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
978 start = max(start, vma->vm_start);
979 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
980 end = min(end, vma->vm_end);
981 pages -= change_prot_numa(vma, start, end);
986 } while (end != vma->vm_end);
991 * It is possible to reach the end of the VMA list but the last few
992 * VMAs are not guaranteed to the vma_migratable. If they are not, we
993 * would find the !migratable VMA on the next scan but not reset the
994 * scanner to the start so check it now.
997 mm->numa_scan_offset = start;
999 reset_ptenuma_scan(p);
1000 up_read(&mm->mmap_sem);
1004 * Drive the periodic memory faults..
1006 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1008 struct callback_head *work = &curr->numa_work;
1012 * We don't care about NUMA placement if we don't have memory.
1014 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1018 * Using runtime rather than walltime has the dual advantage that
1019 * we (mostly) drive the selection from busy threads and that the
1020 * task needs to have done some actual work before we bother with
1023 now = curr->se.sum_exec_runtime;
1024 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1026 if (now - curr->node_stamp > period) {
1027 if (!curr->node_stamp)
1028 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
1029 curr->node_stamp = now;
1031 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1032 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1033 task_work_add(curr, work, true);
1038 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1041 #endif /* CONFIG_NUMA_BALANCING */
1044 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1046 update_load_add(&cfs_rq->load, se->load.weight);
1047 if (!parent_entity(se))
1048 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1050 if (entity_is_task(se))
1051 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1053 cfs_rq->nr_running++;
1057 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1059 update_load_sub(&cfs_rq->load, se->load.weight);
1060 if (!parent_entity(se))
1061 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1062 if (entity_is_task(se))
1063 list_del_init(&se->group_node);
1064 cfs_rq->nr_running--;
1067 #ifdef CONFIG_FAIR_GROUP_SCHED
1069 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1074 * Use this CPU's actual weight instead of the last load_contribution
1075 * to gain a more accurate current total weight. See
1076 * update_cfs_rq_load_contribution().
1078 tg_weight = atomic_long_read(&tg->load_avg);
1079 tg_weight -= cfs_rq->tg_load_contrib;
1080 tg_weight += cfs_rq->load.weight;
1085 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1087 long tg_weight, load, shares;
1089 tg_weight = calc_tg_weight(tg, cfs_rq);
1090 load = cfs_rq->load.weight;
1092 shares = (tg->shares * load);
1094 shares /= tg_weight;
1096 if (shares < MIN_SHARES)
1097 shares = MIN_SHARES;
1098 if (shares > tg->shares)
1099 shares = tg->shares;
1103 # else /* CONFIG_SMP */
1104 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1108 # endif /* CONFIG_SMP */
1109 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1110 unsigned long weight)
1113 /* commit outstanding execution time */
1114 if (cfs_rq->curr == se)
1115 update_curr(cfs_rq);
1116 account_entity_dequeue(cfs_rq, se);
1119 update_load_set(&se->load, weight);
1122 account_entity_enqueue(cfs_rq, se);
1125 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1127 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1129 struct task_group *tg;
1130 struct sched_entity *se;
1134 se = tg->se[cpu_of(rq_of(cfs_rq))];
1135 if (!se || throttled_hierarchy(cfs_rq))
1138 if (likely(se->load.weight == tg->shares))
1141 shares = calc_cfs_shares(cfs_rq, tg);
1143 reweight_entity(cfs_rq_of(se), se, shares);
1145 #else /* CONFIG_FAIR_GROUP_SCHED */
1146 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1149 #endif /* CONFIG_FAIR_GROUP_SCHED */
1153 * We choose a half-life close to 1 scheduling period.
1154 * Note: The tables below are dependent on this value.
1156 #define LOAD_AVG_PERIOD 32
1157 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1158 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1160 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1161 static const u32 runnable_avg_yN_inv[] = {
1162 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1163 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1164 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1165 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1166 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1167 0x85aac367, 0x82cd8698,
1171 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1172 * over-estimates when re-combining.
1174 static const u32 runnable_avg_yN_sum[] = {
1175 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1176 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1177 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1182 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1184 static __always_inline u64 decay_load(u64 val, u64 n)
1186 unsigned int local_n;
1190 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1193 /* after bounds checking we can collapse to 32-bit */
1197 * As y^PERIOD = 1/2, we can combine
1198 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1199 * With a look-up table which covers k^n (n<PERIOD)
1201 * To achieve constant time decay_load.
1203 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1204 val >>= local_n / LOAD_AVG_PERIOD;
1205 local_n %= LOAD_AVG_PERIOD;
1208 val *= runnable_avg_yN_inv[local_n];
1209 /* We don't use SRR here since we always want to round down. */
1214 * For updates fully spanning n periods, the contribution to runnable
1215 * average will be: \Sum 1024*y^n
1217 * We can compute this reasonably efficiently by combining:
1218 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1220 static u32 __compute_runnable_contrib(u64 n)
1224 if (likely(n <= LOAD_AVG_PERIOD))
1225 return runnable_avg_yN_sum[n];
1226 else if (unlikely(n >= LOAD_AVG_MAX_N))
1227 return LOAD_AVG_MAX;
1229 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1231 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1232 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1234 n -= LOAD_AVG_PERIOD;
1235 } while (n > LOAD_AVG_PERIOD);
1237 contrib = decay_load(contrib, n);
1238 return contrib + runnable_avg_yN_sum[n];
1242 * We can represent the historical contribution to runnable average as the
1243 * coefficients of a geometric series. To do this we sub-divide our runnable
1244 * history into segments of approximately 1ms (1024us); label the segment that
1245 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1247 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1249 * (now) (~1ms ago) (~2ms ago)
1251 * Let u_i denote the fraction of p_i that the entity was runnable.
1253 * We then designate the fractions u_i as our co-efficients, yielding the
1254 * following representation of historical load:
1255 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1257 * We choose y based on the with of a reasonably scheduling period, fixing:
1260 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1261 * approximately half as much as the contribution to load within the last ms
1264 * When a period "rolls over" and we have new u_0`, multiplying the previous
1265 * sum again by y is sufficient to update:
1266 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1267 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1269 static __always_inline int __update_entity_runnable_avg(u64 now,
1270 struct sched_avg *sa,
1274 u32 runnable_contrib;
1275 int delta_w, decayed = 0;
1277 delta = now - sa->last_runnable_update;
1279 * This should only happen when time goes backwards, which it
1280 * unfortunately does during sched clock init when we swap over to TSC.
1282 if ((s64)delta < 0) {
1283 sa->last_runnable_update = now;
1288 * Use 1024ns as the unit of measurement since it's a reasonable
1289 * approximation of 1us and fast to compute.
1294 sa->last_runnable_update = now;
1296 /* delta_w is the amount already accumulated against our next period */
1297 delta_w = sa->runnable_avg_period % 1024;
1298 if (delta + delta_w >= 1024) {
1299 /* period roll-over */
1303 * Now that we know we're crossing a period boundary, figure
1304 * out how much from delta we need to complete the current
1305 * period and accrue it.
1307 delta_w = 1024 - delta_w;
1309 sa->runnable_avg_sum += delta_w;
1310 sa->runnable_avg_period += delta_w;
1314 /* Figure out how many additional periods this update spans */
1315 periods = delta / 1024;
1318 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1320 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1323 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1324 runnable_contrib = __compute_runnable_contrib(periods);
1326 sa->runnable_avg_sum += runnable_contrib;
1327 sa->runnable_avg_period += runnable_contrib;
1330 /* Remainder of delta accrued against u_0` */
1332 sa->runnable_avg_sum += delta;
1333 sa->runnable_avg_period += delta;
1338 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1339 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1341 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1342 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1344 decays -= se->avg.decay_count;
1348 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1349 se->avg.decay_count = 0;
1354 #ifdef CONFIG_FAIR_GROUP_SCHED
1355 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1358 struct task_group *tg = cfs_rq->tg;
1361 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1362 tg_contrib -= cfs_rq->tg_load_contrib;
1364 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1365 atomic_long_add(tg_contrib, &tg->load_avg);
1366 cfs_rq->tg_load_contrib += tg_contrib;
1371 * Aggregate cfs_rq runnable averages into an equivalent task_group
1372 * representation for computing load contributions.
1374 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1375 struct cfs_rq *cfs_rq)
1377 struct task_group *tg = cfs_rq->tg;
1380 /* The fraction of a cpu used by this cfs_rq */
1381 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1382 sa->runnable_avg_period + 1);
1383 contrib -= cfs_rq->tg_runnable_contrib;
1385 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1386 atomic_add(contrib, &tg->runnable_avg);
1387 cfs_rq->tg_runnable_contrib += contrib;
1391 static inline void __update_group_entity_contrib(struct sched_entity *se)
1393 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1394 struct task_group *tg = cfs_rq->tg;
1399 contrib = cfs_rq->tg_load_contrib * tg->shares;
1400 se->avg.load_avg_contrib = div_u64(contrib,
1401 atomic_long_read(&tg->load_avg) + 1);
1404 * For group entities we need to compute a correction term in the case
1405 * that they are consuming <1 cpu so that we would contribute the same
1406 * load as a task of equal weight.
1408 * Explicitly co-ordinating this measurement would be expensive, but
1409 * fortunately the sum of each cpus contribution forms a usable
1410 * lower-bound on the true value.
1412 * Consider the aggregate of 2 contributions. Either they are disjoint
1413 * (and the sum represents true value) or they are disjoint and we are
1414 * understating by the aggregate of their overlap.
1416 * Extending this to N cpus, for a given overlap, the maximum amount we
1417 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1418 * cpus that overlap for this interval and w_i is the interval width.
1420 * On a small machine; the first term is well-bounded which bounds the
1421 * total error since w_i is a subset of the period. Whereas on a
1422 * larger machine, while this first term can be larger, if w_i is the
1423 * of consequential size guaranteed to see n_i*w_i quickly converge to
1424 * our upper bound of 1-cpu.
1426 runnable_avg = atomic_read(&tg->runnable_avg);
1427 if (runnable_avg < NICE_0_LOAD) {
1428 se->avg.load_avg_contrib *= runnable_avg;
1429 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1433 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1434 int force_update) {}
1435 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1436 struct cfs_rq *cfs_rq) {}
1437 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1440 static inline void __update_task_entity_contrib(struct sched_entity *se)
1444 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1445 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1446 contrib /= (se->avg.runnable_avg_period + 1);
1447 se->avg.load_avg_contrib = scale_load(contrib);
1450 /* Compute the current contribution to load_avg by se, return any delta */
1451 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1453 long old_contrib = se->avg.load_avg_contrib;
1455 if (entity_is_task(se)) {
1456 __update_task_entity_contrib(se);
1458 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1459 __update_group_entity_contrib(se);
1462 return se->avg.load_avg_contrib - old_contrib;
1465 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1468 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1469 cfs_rq->blocked_load_avg -= load_contrib;
1471 cfs_rq->blocked_load_avg = 0;
1474 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1476 /* Update a sched_entity's runnable average */
1477 static inline void update_entity_load_avg(struct sched_entity *se,
1480 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1485 * For a group entity we need to use their owned cfs_rq_clock_task() in
1486 * case they are the parent of a throttled hierarchy.
1488 if (entity_is_task(se))
1489 now = cfs_rq_clock_task(cfs_rq);
1491 now = cfs_rq_clock_task(group_cfs_rq(se));
1493 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1496 contrib_delta = __update_entity_load_avg_contrib(se);
1502 cfs_rq->runnable_load_avg += contrib_delta;
1504 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1508 * Decay the load contributed by all blocked children and account this so that
1509 * their contribution may appropriately discounted when they wake up.
1511 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1513 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1516 decays = now - cfs_rq->last_decay;
1517 if (!decays && !force_update)
1520 if (atomic_long_read(&cfs_rq->removed_load)) {
1521 unsigned long removed_load;
1522 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1523 subtract_blocked_load_contrib(cfs_rq, removed_load);
1527 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1529 atomic64_add(decays, &cfs_rq->decay_counter);
1530 cfs_rq->last_decay = now;
1533 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1536 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1538 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1539 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1542 /* Add the load generated by se into cfs_rq's child load-average */
1543 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1544 struct sched_entity *se,
1548 * We track migrations using entity decay_count <= 0, on a wake-up
1549 * migration we use a negative decay count to track the remote decays
1550 * accumulated while sleeping.
1552 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1553 * are seen by enqueue_entity_load_avg() as a migration with an already
1554 * constructed load_avg_contrib.
1556 if (unlikely(se->avg.decay_count <= 0)) {
1557 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1558 if (se->avg.decay_count) {
1560 * In a wake-up migration we have to approximate the
1561 * time sleeping. This is because we can't synchronize
1562 * clock_task between the two cpus, and it is not
1563 * guaranteed to be read-safe. Instead, we can
1564 * approximate this using our carried decays, which are
1565 * explicitly atomically readable.
1567 se->avg.last_runnable_update -= (-se->avg.decay_count)
1569 update_entity_load_avg(se, 0);
1570 /* Indicate that we're now synchronized and on-rq */
1571 se->avg.decay_count = 0;
1576 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1577 * would have made count negative); we must be careful to avoid
1578 * double-accounting blocked time after synchronizing decays.
1580 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1584 /* migrated tasks did not contribute to our blocked load */
1586 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1587 update_entity_load_avg(se, 0);
1590 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1591 /* we force update consideration on load-balancer moves */
1592 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1596 * Remove se's load from this cfs_rq child load-average, if the entity is
1597 * transitioning to a blocked state we track its projected decay using
1600 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1601 struct sched_entity *se,
1604 update_entity_load_avg(se, 1);
1605 /* we force update consideration on load-balancer moves */
1606 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1608 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1610 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1611 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1612 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1616 * Update the rq's load with the elapsed running time before entering
1617 * idle. if the last scheduled task is not a CFS task, idle_enter will
1618 * be the only way to update the runnable statistic.
1620 void idle_enter_fair(struct rq *this_rq)
1622 update_rq_runnable_avg(this_rq, 1);
1626 * Update the rq's load with the elapsed idle time before a task is
1627 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1628 * be the only way to update the runnable statistic.
1630 void idle_exit_fair(struct rq *this_rq)
1632 update_rq_runnable_avg(this_rq, 0);
1636 static inline void update_entity_load_avg(struct sched_entity *se,
1637 int update_cfs_rq) {}
1638 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1639 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1640 struct sched_entity *se,
1642 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1643 struct sched_entity *se,
1645 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1646 int force_update) {}
1649 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1651 #ifdef CONFIG_SCHEDSTATS
1652 struct task_struct *tsk = NULL;
1654 if (entity_is_task(se))
1657 if (se->statistics.sleep_start) {
1658 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1663 if (unlikely(delta > se->statistics.sleep_max))
1664 se->statistics.sleep_max = delta;
1666 se->statistics.sleep_start = 0;
1667 se->statistics.sum_sleep_runtime += delta;
1670 account_scheduler_latency(tsk, delta >> 10, 1);
1671 trace_sched_stat_sleep(tsk, delta);
1674 if (se->statistics.block_start) {
1675 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1680 if (unlikely(delta > se->statistics.block_max))
1681 se->statistics.block_max = delta;
1683 se->statistics.block_start = 0;
1684 se->statistics.sum_sleep_runtime += delta;
1687 if (tsk->in_iowait) {
1688 se->statistics.iowait_sum += delta;
1689 se->statistics.iowait_count++;
1690 trace_sched_stat_iowait(tsk, delta);
1693 trace_sched_stat_blocked(tsk, delta);
1696 * Blocking time is in units of nanosecs, so shift by
1697 * 20 to get a milliseconds-range estimation of the
1698 * amount of time that the task spent sleeping:
1700 if (unlikely(prof_on == SLEEP_PROFILING)) {
1701 profile_hits(SLEEP_PROFILING,
1702 (void *)get_wchan(tsk),
1705 account_scheduler_latency(tsk, delta >> 10, 0);
1711 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1713 #ifdef CONFIG_SCHED_DEBUG
1714 s64 d = se->vruntime - cfs_rq->min_vruntime;
1719 if (d > 3*sysctl_sched_latency)
1720 schedstat_inc(cfs_rq, nr_spread_over);
1725 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1727 u64 vruntime = cfs_rq->min_vruntime;
1730 * The 'current' period is already promised to the current tasks,
1731 * however the extra weight of the new task will slow them down a
1732 * little, place the new task so that it fits in the slot that
1733 * stays open at the end.
1735 if (initial && sched_feat(START_DEBIT))
1736 vruntime += sched_vslice(cfs_rq, se);
1738 /* sleeps up to a single latency don't count. */
1740 unsigned long thresh = sysctl_sched_latency;
1743 * Halve their sleep time's effect, to allow
1744 * for a gentler effect of sleepers:
1746 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1752 /* ensure we never gain time by being placed backwards. */
1753 se->vruntime = max_vruntime(se->vruntime, vruntime);
1756 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1759 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1762 * Update the normalized vruntime before updating min_vruntime
1763 * through calling update_curr().
1765 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1766 se->vruntime += cfs_rq->min_vruntime;
1769 * Update run-time statistics of the 'current'.
1771 update_curr(cfs_rq);
1772 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1773 account_entity_enqueue(cfs_rq, se);
1774 update_cfs_shares(cfs_rq);
1776 if (flags & ENQUEUE_WAKEUP) {
1777 place_entity(cfs_rq, se, 0);
1778 enqueue_sleeper(cfs_rq, se);
1781 update_stats_enqueue(cfs_rq, se);
1782 check_spread(cfs_rq, se);
1783 if (se != cfs_rq->curr)
1784 __enqueue_entity(cfs_rq, se);
1787 if (cfs_rq->nr_running == 1) {
1788 list_add_leaf_cfs_rq(cfs_rq);
1789 check_enqueue_throttle(cfs_rq);
1793 static void __clear_buddies_last(struct sched_entity *se)
1795 for_each_sched_entity(se) {
1796 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1797 if (cfs_rq->last == se)
1798 cfs_rq->last = NULL;
1804 static void __clear_buddies_next(struct sched_entity *se)
1806 for_each_sched_entity(se) {
1807 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1808 if (cfs_rq->next == se)
1809 cfs_rq->next = NULL;
1815 static void __clear_buddies_skip(struct sched_entity *se)
1817 for_each_sched_entity(se) {
1818 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1819 if (cfs_rq->skip == se)
1820 cfs_rq->skip = NULL;
1826 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1828 if (cfs_rq->last == se)
1829 __clear_buddies_last(se);
1831 if (cfs_rq->next == se)
1832 __clear_buddies_next(se);
1834 if (cfs_rq->skip == se)
1835 __clear_buddies_skip(se);
1838 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1841 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1844 * Update run-time statistics of the 'current'.
1846 update_curr(cfs_rq);
1847 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1849 update_stats_dequeue(cfs_rq, se);
1850 if (flags & DEQUEUE_SLEEP) {
1851 #ifdef CONFIG_SCHEDSTATS
1852 if (entity_is_task(se)) {
1853 struct task_struct *tsk = task_of(se);
1855 if (tsk->state & TASK_INTERRUPTIBLE)
1856 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
1857 if (tsk->state & TASK_UNINTERRUPTIBLE)
1858 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
1863 clear_buddies(cfs_rq, se);
1865 if (se != cfs_rq->curr)
1866 __dequeue_entity(cfs_rq, se);
1868 account_entity_dequeue(cfs_rq, se);
1871 * Normalize the entity after updating the min_vruntime because the
1872 * update can refer to the ->curr item and we need to reflect this
1873 * movement in our normalized position.
1875 if (!(flags & DEQUEUE_SLEEP))
1876 se->vruntime -= cfs_rq->min_vruntime;
1878 /* return excess runtime on last dequeue */
1879 return_cfs_rq_runtime(cfs_rq);
1881 update_min_vruntime(cfs_rq);
1882 update_cfs_shares(cfs_rq);
1886 * Preempt the current task with a newly woken task if needed:
1889 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1891 unsigned long ideal_runtime, delta_exec;
1892 struct sched_entity *se;
1895 ideal_runtime = sched_slice(cfs_rq, curr);
1896 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1897 if (delta_exec > ideal_runtime) {
1898 resched_task(rq_of(cfs_rq)->curr);
1900 * The current task ran long enough, ensure it doesn't get
1901 * re-elected due to buddy favours.
1903 clear_buddies(cfs_rq, curr);
1908 * Ensure that a task that missed wakeup preemption by a
1909 * narrow margin doesn't have to wait for a full slice.
1910 * This also mitigates buddy induced latencies under load.
1912 if (delta_exec < sysctl_sched_min_granularity)
1915 se = __pick_first_entity(cfs_rq);
1916 delta = curr->vruntime - se->vruntime;
1921 if (delta > ideal_runtime)
1922 resched_task(rq_of(cfs_rq)->curr);
1926 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1928 /* 'current' is not kept within the tree. */
1931 * Any task has to be enqueued before it get to execute on
1932 * a CPU. So account for the time it spent waiting on the
1935 update_stats_wait_end(cfs_rq, se);
1936 __dequeue_entity(cfs_rq, se);
1939 update_stats_curr_start(cfs_rq, se);
1941 #ifdef CONFIG_SCHEDSTATS
1943 * Track our maximum slice length, if the CPU's load is at
1944 * least twice that of our own weight (i.e. dont track it
1945 * when there are only lesser-weight tasks around):
1947 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1948 se->statistics.slice_max = max(se->statistics.slice_max,
1949 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1952 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1956 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1959 * Pick the next process, keeping these things in mind, in this order:
1960 * 1) keep things fair between processes/task groups
1961 * 2) pick the "next" process, since someone really wants that to run
1962 * 3) pick the "last" process, for cache locality
1963 * 4) do not run the "skip" process, if something else is available
1965 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1967 struct sched_entity *se = __pick_first_entity(cfs_rq);
1968 struct sched_entity *left = se;
1971 * Avoid running the skip buddy, if running something else can
1972 * be done without getting too unfair.
1974 if (cfs_rq->skip == se) {
1975 struct sched_entity *second = __pick_next_entity(se);
1976 if (second && wakeup_preempt_entity(second, left) < 1)
1981 * Prefer last buddy, try to return the CPU to a preempted task.
1983 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1987 * Someone really wants this to run. If it's not unfair, run it.
1989 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1992 clear_buddies(cfs_rq, se);
1997 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1999 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2002 * If still on the runqueue then deactivate_task()
2003 * was not called and update_curr() has to be done:
2006 update_curr(cfs_rq);
2008 /* throttle cfs_rqs exceeding runtime */
2009 check_cfs_rq_runtime(cfs_rq);
2011 check_spread(cfs_rq, prev);
2013 update_stats_wait_start(cfs_rq, prev);
2014 /* Put 'current' back into the tree. */
2015 __enqueue_entity(cfs_rq, prev);
2016 /* in !on_rq case, update occurred at dequeue */
2017 update_entity_load_avg(prev, 1);
2019 cfs_rq->curr = NULL;
2023 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2026 * Update run-time statistics of the 'current'.
2028 update_curr(cfs_rq);
2031 * Ensure that runnable average is periodically updated.
2033 update_entity_load_avg(curr, 1);
2034 update_cfs_rq_blocked_load(cfs_rq, 1);
2035 update_cfs_shares(cfs_rq);
2037 #ifdef CONFIG_SCHED_HRTICK
2039 * queued ticks are scheduled to match the slice, so don't bother
2040 * validating it and just reschedule.
2043 resched_task(rq_of(cfs_rq)->curr);
2047 * don't let the period tick interfere with the hrtick preemption
2049 if (!sched_feat(DOUBLE_TICK) &&
2050 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2054 if (cfs_rq->nr_running > 1)
2055 check_preempt_tick(cfs_rq, curr);
2059 /**************************************************
2060 * CFS bandwidth control machinery
2063 #ifdef CONFIG_CFS_BANDWIDTH
2065 #ifdef HAVE_JUMP_LABEL
2066 static struct static_key __cfs_bandwidth_used;
2068 static inline bool cfs_bandwidth_used(void)
2070 return static_key_false(&__cfs_bandwidth_used);
2073 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2075 /* only need to count groups transitioning between enabled/!enabled */
2076 if (enabled && !was_enabled)
2077 static_key_slow_inc(&__cfs_bandwidth_used);
2078 else if (!enabled && was_enabled)
2079 static_key_slow_dec(&__cfs_bandwidth_used);
2081 #else /* HAVE_JUMP_LABEL */
2082 static bool cfs_bandwidth_used(void)
2087 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2088 #endif /* HAVE_JUMP_LABEL */
2091 * default period for cfs group bandwidth.
2092 * default: 0.1s, units: nanoseconds
2094 static inline u64 default_cfs_period(void)
2096 return 100000000ULL;
2099 static inline u64 sched_cfs_bandwidth_slice(void)
2101 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2105 * Replenish runtime according to assigned quota and update expiration time.
2106 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2107 * additional synchronization around rq->lock.
2109 * requires cfs_b->lock
2111 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2115 if (cfs_b->quota == RUNTIME_INF)
2118 now = sched_clock_cpu(smp_processor_id());
2119 cfs_b->runtime = cfs_b->quota;
2120 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2123 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2125 return &tg->cfs_bandwidth;
2128 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2129 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2131 if (unlikely(cfs_rq->throttle_count))
2132 return cfs_rq->throttled_clock_task;
2134 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2137 /* returns 0 on failure to allocate runtime */
2138 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2140 struct task_group *tg = cfs_rq->tg;
2141 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2142 u64 amount = 0, min_amount, expires;
2144 /* note: this is a positive sum as runtime_remaining <= 0 */
2145 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2147 raw_spin_lock(&cfs_b->lock);
2148 if (cfs_b->quota == RUNTIME_INF)
2149 amount = min_amount;
2152 * If the bandwidth pool has become inactive, then at least one
2153 * period must have elapsed since the last consumption.
2154 * Refresh the global state and ensure bandwidth timer becomes
2157 if (!cfs_b->timer_active) {
2158 __refill_cfs_bandwidth_runtime(cfs_b);
2159 __start_cfs_bandwidth(cfs_b);
2162 if (cfs_b->runtime > 0) {
2163 amount = min(cfs_b->runtime, min_amount);
2164 cfs_b->runtime -= amount;
2168 expires = cfs_b->runtime_expires;
2169 raw_spin_unlock(&cfs_b->lock);
2171 cfs_rq->runtime_remaining += amount;
2173 * we may have advanced our local expiration to account for allowed
2174 * spread between our sched_clock and the one on which runtime was
2177 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2178 cfs_rq->runtime_expires = expires;
2180 return cfs_rq->runtime_remaining > 0;
2184 * Note: This depends on the synchronization provided by sched_clock and the
2185 * fact that rq->clock snapshots this value.
2187 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2189 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2191 /* if the deadline is ahead of our clock, nothing to do */
2192 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2195 if (cfs_rq->runtime_remaining < 0)
2199 * If the local deadline has passed we have to consider the
2200 * possibility that our sched_clock is 'fast' and the global deadline
2201 * has not truly expired.
2203 * Fortunately we can check determine whether this the case by checking
2204 * whether the global deadline has advanced.
2207 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2208 /* extend local deadline, drift is bounded above by 2 ticks */
2209 cfs_rq->runtime_expires += TICK_NSEC;
2211 /* global deadline is ahead, expiration has passed */
2212 cfs_rq->runtime_remaining = 0;
2216 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2217 unsigned long delta_exec)
2219 /* dock delta_exec before expiring quota (as it could span periods) */
2220 cfs_rq->runtime_remaining -= delta_exec;
2221 expire_cfs_rq_runtime(cfs_rq);
2223 if (likely(cfs_rq->runtime_remaining > 0))
2227 * if we're unable to extend our runtime we resched so that the active
2228 * hierarchy can be throttled
2230 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2231 resched_task(rq_of(cfs_rq)->curr);
2234 static __always_inline
2235 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2237 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2240 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2243 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2245 return cfs_bandwidth_used() && cfs_rq->throttled;
2248 /* check whether cfs_rq, or any parent, is throttled */
2249 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2251 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2255 * Ensure that neither of the group entities corresponding to src_cpu or
2256 * dest_cpu are members of a throttled hierarchy when performing group
2257 * load-balance operations.
2259 static inline int throttled_lb_pair(struct task_group *tg,
2260 int src_cpu, int dest_cpu)
2262 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2264 src_cfs_rq = tg->cfs_rq[src_cpu];
2265 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2267 return throttled_hierarchy(src_cfs_rq) ||
2268 throttled_hierarchy(dest_cfs_rq);
2271 /* updated child weight may affect parent so we have to do this bottom up */
2272 static int tg_unthrottle_up(struct task_group *tg, void *data)
2274 struct rq *rq = data;
2275 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2277 cfs_rq->throttle_count--;
2279 if (!cfs_rq->throttle_count) {
2280 /* adjust cfs_rq_clock_task() */
2281 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2282 cfs_rq->throttled_clock_task;
2289 static int tg_throttle_down(struct task_group *tg, void *data)
2291 struct rq *rq = data;
2292 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2294 /* group is entering throttled state, stop time */
2295 if (!cfs_rq->throttle_count)
2296 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2297 cfs_rq->throttle_count++;
2302 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2304 struct rq *rq = rq_of(cfs_rq);
2305 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2306 struct sched_entity *se;
2307 long task_delta, dequeue = 1;
2309 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2311 /* freeze hierarchy runnable averages while throttled */
2313 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2316 task_delta = cfs_rq->h_nr_running;
2317 for_each_sched_entity(se) {
2318 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2319 /* throttled entity or throttle-on-deactivate */
2324 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2325 qcfs_rq->h_nr_running -= task_delta;
2327 if (qcfs_rq->load.weight)
2332 rq->nr_running -= task_delta;
2334 cfs_rq->throttled = 1;
2335 cfs_rq->throttled_clock = rq_clock(rq);
2336 raw_spin_lock(&cfs_b->lock);
2337 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2338 raw_spin_unlock(&cfs_b->lock);
2341 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2343 struct rq *rq = rq_of(cfs_rq);
2344 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2345 struct sched_entity *se;
2349 se = cfs_rq->tg->se[cpu_of(rq)];
2351 cfs_rq->throttled = 0;
2353 update_rq_clock(rq);
2355 raw_spin_lock(&cfs_b->lock);
2356 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2357 list_del_rcu(&cfs_rq->throttled_list);
2358 raw_spin_unlock(&cfs_b->lock);
2360 /* update hierarchical throttle state */
2361 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2363 if (!cfs_rq->load.weight)
2366 task_delta = cfs_rq->h_nr_running;
2367 for_each_sched_entity(se) {
2371 cfs_rq = cfs_rq_of(se);
2373 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2374 cfs_rq->h_nr_running += task_delta;
2376 if (cfs_rq_throttled(cfs_rq))
2381 rq->nr_running += task_delta;
2383 /* determine whether we need to wake up potentially idle cpu */
2384 if (rq->curr == rq->idle && rq->cfs.nr_running)
2385 resched_task(rq->curr);
2388 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2389 u64 remaining, u64 expires)
2391 struct cfs_rq *cfs_rq;
2392 u64 runtime = remaining;
2395 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2397 struct rq *rq = rq_of(cfs_rq);
2399 raw_spin_lock(&rq->lock);
2400 if (!cfs_rq_throttled(cfs_rq))
2403 runtime = -cfs_rq->runtime_remaining + 1;
2404 if (runtime > remaining)
2405 runtime = remaining;
2406 remaining -= runtime;
2408 cfs_rq->runtime_remaining += runtime;
2409 cfs_rq->runtime_expires = expires;
2411 /* we check whether we're throttled above */
2412 if (cfs_rq->runtime_remaining > 0)
2413 unthrottle_cfs_rq(cfs_rq);
2416 raw_spin_unlock(&rq->lock);
2427 * Responsible for refilling a task_group's bandwidth and unthrottling its
2428 * cfs_rqs as appropriate. If there has been no activity within the last
2429 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2430 * used to track this state.
2432 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2434 u64 runtime, runtime_expires;
2435 int idle = 1, throttled;
2437 raw_spin_lock(&cfs_b->lock);
2438 /* no need to continue the timer with no bandwidth constraint */
2439 if (cfs_b->quota == RUNTIME_INF)
2442 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2443 /* idle depends on !throttled (for the case of a large deficit) */
2444 idle = cfs_b->idle && !throttled;
2445 cfs_b->nr_periods += overrun;
2447 /* if we're going inactive then everything else can be deferred */
2451 __refill_cfs_bandwidth_runtime(cfs_b);
2454 /* mark as potentially idle for the upcoming period */
2459 /* account preceding periods in which throttling occurred */
2460 cfs_b->nr_throttled += overrun;
2463 * There are throttled entities so we must first use the new bandwidth
2464 * to unthrottle them before making it generally available. This
2465 * ensures that all existing debts will be paid before a new cfs_rq is
2468 runtime = cfs_b->runtime;
2469 runtime_expires = cfs_b->runtime_expires;
2473 * This check is repeated as we are holding onto the new bandwidth
2474 * while we unthrottle. This can potentially race with an unthrottled
2475 * group trying to acquire new bandwidth from the global pool.
2477 while (throttled && runtime > 0) {
2478 raw_spin_unlock(&cfs_b->lock);
2479 /* we can't nest cfs_b->lock while distributing bandwidth */
2480 runtime = distribute_cfs_runtime(cfs_b, runtime,
2482 raw_spin_lock(&cfs_b->lock);
2484 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2487 /* return (any) remaining runtime */
2488 cfs_b->runtime = runtime;
2490 * While we are ensured activity in the period following an
2491 * unthrottle, this also covers the case in which the new bandwidth is
2492 * insufficient to cover the existing bandwidth deficit. (Forcing the
2493 * timer to remain active while there are any throttled entities.)
2498 cfs_b->timer_active = 0;
2499 raw_spin_unlock(&cfs_b->lock);
2504 /* a cfs_rq won't donate quota below this amount */
2505 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2506 /* minimum remaining period time to redistribute slack quota */
2507 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2508 /* how long we wait to gather additional slack before distributing */
2509 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2511 /* are we near the end of the current quota period? */
2512 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2514 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2517 /* if the call-back is running a quota refresh is already occurring */
2518 if (hrtimer_callback_running(refresh_timer))
2521 /* is a quota refresh about to occur? */
2522 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2523 if (remaining < min_expire)
2529 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2531 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2533 /* if there's a quota refresh soon don't bother with slack */
2534 if (runtime_refresh_within(cfs_b, min_left))
2537 start_bandwidth_timer(&cfs_b->slack_timer,
2538 ns_to_ktime(cfs_bandwidth_slack_period));
2541 /* we know any runtime found here is valid as update_curr() precedes return */
2542 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2544 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2545 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2547 if (slack_runtime <= 0)
2550 raw_spin_lock(&cfs_b->lock);
2551 if (cfs_b->quota != RUNTIME_INF &&
2552 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2553 cfs_b->runtime += slack_runtime;
2555 /* we are under rq->lock, defer unthrottling using a timer */
2556 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2557 !list_empty(&cfs_b->throttled_cfs_rq))
2558 start_cfs_slack_bandwidth(cfs_b);
2560 raw_spin_unlock(&cfs_b->lock);
2562 /* even if it's not valid for return we don't want to try again */
2563 cfs_rq->runtime_remaining -= slack_runtime;
2566 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2568 if (!cfs_bandwidth_used())
2571 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2574 __return_cfs_rq_runtime(cfs_rq);
2578 * This is done with a timer (instead of inline with bandwidth return) since
2579 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2581 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2583 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2586 /* confirm we're still not at a refresh boundary */
2587 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2590 raw_spin_lock(&cfs_b->lock);
2591 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2592 runtime = cfs_b->runtime;
2595 expires = cfs_b->runtime_expires;
2596 raw_spin_unlock(&cfs_b->lock);
2601 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2603 raw_spin_lock(&cfs_b->lock);
2604 if (expires == cfs_b->runtime_expires)
2605 cfs_b->runtime = runtime;
2606 raw_spin_unlock(&cfs_b->lock);
2610 * When a group wakes up we want to make sure that its quota is not already
2611 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2612 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2614 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2616 if (!cfs_bandwidth_used())
2619 /* an active group must be handled by the update_curr()->put() path */
2620 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2623 /* ensure the group is not already throttled */
2624 if (cfs_rq_throttled(cfs_rq))
2627 /* update runtime allocation */
2628 account_cfs_rq_runtime(cfs_rq, 0);
2629 if (cfs_rq->runtime_remaining <= 0)
2630 throttle_cfs_rq(cfs_rq);
2633 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2634 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2636 if (!cfs_bandwidth_used())
2639 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2643 * it's possible for a throttled entity to be forced into a running
2644 * state (e.g. set_curr_task), in this case we're finished.
2646 if (cfs_rq_throttled(cfs_rq))
2649 throttle_cfs_rq(cfs_rq);
2652 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2654 struct cfs_bandwidth *cfs_b =
2655 container_of(timer, struct cfs_bandwidth, slack_timer);
2656 do_sched_cfs_slack_timer(cfs_b);
2658 return HRTIMER_NORESTART;
2661 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2663 struct cfs_bandwidth *cfs_b =
2664 container_of(timer, struct cfs_bandwidth, period_timer);
2670 now = hrtimer_cb_get_time(timer);
2671 overrun = hrtimer_forward(timer, now, cfs_b->period);
2676 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2679 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2682 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2684 raw_spin_lock_init(&cfs_b->lock);
2686 cfs_b->quota = RUNTIME_INF;
2687 cfs_b->period = ns_to_ktime(default_cfs_period());
2689 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2690 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2691 cfs_b->period_timer.function = sched_cfs_period_timer;
2692 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2693 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2696 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2698 cfs_rq->runtime_enabled = 0;
2699 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2702 /* requires cfs_b->lock, may release to reprogram timer */
2703 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2706 * The timer may be active because we're trying to set a new bandwidth
2707 * period or because we're racing with the tear-down path
2708 * (timer_active==0 becomes visible before the hrtimer call-back
2709 * terminates). In either case we ensure that it's re-programmed
2711 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2712 raw_spin_unlock(&cfs_b->lock);
2713 /* ensure cfs_b->lock is available while we wait */
2714 hrtimer_cancel(&cfs_b->period_timer);
2716 raw_spin_lock(&cfs_b->lock);
2717 /* if someone else restarted the timer then we're done */
2718 if (cfs_b->timer_active)
2722 cfs_b->timer_active = 1;
2723 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2726 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2728 hrtimer_cancel(&cfs_b->period_timer);
2729 hrtimer_cancel(&cfs_b->slack_timer);
2732 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2734 struct cfs_rq *cfs_rq;
2736 for_each_leaf_cfs_rq(rq, cfs_rq) {
2737 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2739 if (!cfs_rq->runtime_enabled)
2743 * clock_task is not advancing so we just need to make sure
2744 * there's some valid quota amount
2746 cfs_rq->runtime_remaining = cfs_b->quota;
2747 if (cfs_rq_throttled(cfs_rq))
2748 unthrottle_cfs_rq(cfs_rq);
2752 #else /* CONFIG_CFS_BANDWIDTH */
2753 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2755 return rq_clock_task(rq_of(cfs_rq));
2758 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2759 unsigned long delta_exec) {}
2760 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2761 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2762 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2764 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2769 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2774 static inline int throttled_lb_pair(struct task_group *tg,
2775 int src_cpu, int dest_cpu)
2780 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2782 #ifdef CONFIG_FAIR_GROUP_SCHED
2783 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2786 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2790 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2791 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2793 #endif /* CONFIG_CFS_BANDWIDTH */
2795 /**************************************************
2796 * CFS operations on tasks:
2799 #ifdef CONFIG_SCHED_HRTICK
2800 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2802 struct sched_entity *se = &p->se;
2803 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2805 WARN_ON(task_rq(p) != rq);
2807 if (cfs_rq->nr_running > 1) {
2808 u64 slice = sched_slice(cfs_rq, se);
2809 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2810 s64 delta = slice - ran;
2819 * Don't schedule slices shorter than 10000ns, that just
2820 * doesn't make sense. Rely on vruntime for fairness.
2823 delta = max_t(s64, 10000LL, delta);
2825 hrtick_start(rq, delta);
2830 * called from enqueue/dequeue and updates the hrtick when the
2831 * current task is from our class and nr_running is low enough
2834 static void hrtick_update(struct rq *rq)
2836 struct task_struct *curr = rq->curr;
2838 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2841 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2842 hrtick_start_fair(rq, curr);
2844 #else /* !CONFIG_SCHED_HRTICK */
2846 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2850 static inline void hrtick_update(struct rq *rq)
2856 * The enqueue_task method is called before nr_running is
2857 * increased. Here we update the fair scheduling stats and
2858 * then put the task into the rbtree:
2861 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2863 struct cfs_rq *cfs_rq;
2864 struct sched_entity *se = &p->se;
2866 for_each_sched_entity(se) {
2869 cfs_rq = cfs_rq_of(se);
2870 enqueue_entity(cfs_rq, se, flags);
2873 * end evaluation on encountering a throttled cfs_rq
2875 * note: in the case of encountering a throttled cfs_rq we will
2876 * post the final h_nr_running increment below.
2878 if (cfs_rq_throttled(cfs_rq))
2880 cfs_rq->h_nr_running++;
2882 flags = ENQUEUE_WAKEUP;
2885 for_each_sched_entity(se) {
2886 cfs_rq = cfs_rq_of(se);
2887 cfs_rq->h_nr_running++;
2889 if (cfs_rq_throttled(cfs_rq))
2892 update_cfs_shares(cfs_rq);
2893 update_entity_load_avg(se, 1);
2897 update_rq_runnable_avg(rq, rq->nr_running);
2903 static void set_next_buddy(struct sched_entity *se);
2906 * The dequeue_task method is called before nr_running is
2907 * decreased. We remove the task from the rbtree and
2908 * update the fair scheduling stats:
2910 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2912 struct cfs_rq *cfs_rq;
2913 struct sched_entity *se = &p->se;
2914 int task_sleep = flags & DEQUEUE_SLEEP;
2916 for_each_sched_entity(se) {
2917 cfs_rq = cfs_rq_of(se);
2918 dequeue_entity(cfs_rq, se, flags);
2921 * end evaluation on encountering a throttled cfs_rq
2923 * note: in the case of encountering a throttled cfs_rq we will
2924 * post the final h_nr_running decrement below.
2926 if (cfs_rq_throttled(cfs_rq))
2928 cfs_rq->h_nr_running--;
2930 /* Don't dequeue parent if it has other entities besides us */
2931 if (cfs_rq->load.weight) {
2933 * Bias pick_next to pick a task from this cfs_rq, as
2934 * p is sleeping when it is within its sched_slice.
2936 if (task_sleep && parent_entity(se))
2937 set_next_buddy(parent_entity(se));
2939 /* avoid re-evaluating load for this entity */
2940 se = parent_entity(se);
2943 flags |= DEQUEUE_SLEEP;
2946 for_each_sched_entity(se) {
2947 cfs_rq = cfs_rq_of(se);
2948 cfs_rq->h_nr_running--;
2950 if (cfs_rq_throttled(cfs_rq))
2953 update_cfs_shares(cfs_rq);
2954 update_entity_load_avg(se, 1);
2959 update_rq_runnable_avg(rq, 1);
2965 /* Used instead of source_load when we know the type == 0 */
2966 static unsigned long weighted_cpuload(const int cpu)
2968 return cpu_rq(cpu)->cfs.runnable_load_avg;
2972 * Return a low guess at the load of a migration-source cpu weighted
2973 * according to the scheduling class and "nice" value.
2975 * We want to under-estimate the load of migration sources, to
2976 * balance conservatively.
2978 static unsigned long source_load(int cpu, int type)
2980 struct rq *rq = cpu_rq(cpu);
2981 unsigned long total = weighted_cpuload(cpu);
2983 if (type == 0 || !sched_feat(LB_BIAS))
2986 return min(rq->cpu_load[type-1], total);
2990 * Return a high guess at the load of a migration-target cpu weighted
2991 * according to the scheduling class and "nice" value.
2993 static unsigned long target_load(int cpu, int type)
2995 struct rq *rq = cpu_rq(cpu);
2996 unsigned long total = weighted_cpuload(cpu);
2998 if (type == 0 || !sched_feat(LB_BIAS))
3001 return max(rq->cpu_load[type-1], total);
3004 static unsigned long power_of(int cpu)
3006 return cpu_rq(cpu)->cpu_power;
3009 static unsigned long cpu_avg_load_per_task(int cpu)
3011 struct rq *rq = cpu_rq(cpu);
3012 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3013 unsigned long load_avg = rq->cfs.runnable_load_avg;
3016 return load_avg / nr_running;
3021 static void record_wakee(struct task_struct *p)
3024 * Rough decay (wiping) for cost saving, don't worry
3025 * about the boundary, really active task won't care
3028 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3029 current->wakee_flips = 0;
3030 current->wakee_flip_decay_ts = jiffies;
3033 if (current->last_wakee != p) {
3034 current->last_wakee = p;
3035 current->wakee_flips++;
3039 static void task_waking_fair(struct task_struct *p)
3041 struct sched_entity *se = &p->se;
3042 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3045 #ifndef CONFIG_64BIT
3046 u64 min_vruntime_copy;
3049 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3051 min_vruntime = cfs_rq->min_vruntime;
3052 } while (min_vruntime != min_vruntime_copy);
3054 min_vruntime = cfs_rq->min_vruntime;
3057 se->vruntime -= min_vruntime;
3061 #ifdef CONFIG_FAIR_GROUP_SCHED
3063 * effective_load() calculates the load change as seen from the root_task_group
3065 * Adding load to a group doesn't make a group heavier, but can cause movement
3066 * of group shares between cpus. Assuming the shares were perfectly aligned one
3067 * can calculate the shift in shares.
3069 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3070 * on this @cpu and results in a total addition (subtraction) of @wg to the
3071 * total group weight.
3073 * Given a runqueue weight distribution (rw_i) we can compute a shares
3074 * distribution (s_i) using:
3076 * s_i = rw_i / \Sum rw_j (1)
3078 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3079 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3080 * shares distribution (s_i):
3082 * rw_i = { 2, 4, 1, 0 }
3083 * s_i = { 2/7, 4/7, 1/7, 0 }
3085 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3086 * task used to run on and the CPU the waker is running on), we need to
3087 * compute the effect of waking a task on either CPU and, in case of a sync
3088 * wakeup, compute the effect of the current task going to sleep.
3090 * So for a change of @wl to the local @cpu with an overall group weight change
3091 * of @wl we can compute the new shares distribution (s'_i) using:
3093 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3095 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3096 * differences in waking a task to CPU 0. The additional task changes the
3097 * weight and shares distributions like:
3099 * rw'_i = { 3, 4, 1, 0 }
3100 * s'_i = { 3/8, 4/8, 1/8, 0 }
3102 * We can then compute the difference in effective weight by using:
3104 * dw_i = S * (s'_i - s_i) (3)
3106 * Where 'S' is the group weight as seen by its parent.
3108 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3109 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3110 * 4/7) times the weight of the group.
3112 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3114 struct sched_entity *se = tg->se[cpu];
3116 if (!tg->parent) /* the trivial, non-cgroup case */
3119 for_each_sched_entity(se) {
3125 * W = @wg + \Sum rw_j
3127 W = wg + calc_tg_weight(tg, se->my_q);
3132 w = se->my_q->load.weight + wl;
3135 * wl = S * s'_i; see (2)
3138 wl = (w * tg->shares) / W;
3143 * Per the above, wl is the new se->load.weight value; since
3144 * those are clipped to [MIN_SHARES, ...) do so now. See
3145 * calc_cfs_shares().
3147 if (wl < MIN_SHARES)
3151 * wl = dw_i = S * (s'_i - s_i); see (3)
3153 wl -= se->load.weight;
3156 * Recursively apply this logic to all parent groups to compute
3157 * the final effective load change on the root group. Since
3158 * only the @tg group gets extra weight, all parent groups can
3159 * only redistribute existing shares. @wl is the shift in shares
3160 * resulting from this level per the above.
3169 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3170 unsigned long wl, unsigned long wg)
3177 static int wake_wide(struct task_struct *p)
3179 int factor = this_cpu_read(sd_llc_size);
3182 * Yeah, it's the switching-frequency, could means many wakee or
3183 * rapidly switch, use factor here will just help to automatically
3184 * adjust the loose-degree, so bigger node will lead to more pull.
3186 if (p->wakee_flips > factor) {
3188 * wakee is somewhat hot, it needs certain amount of cpu
3189 * resource, so if waker is far more hot, prefer to leave
3192 if (current->wakee_flips > (factor * p->wakee_flips))
3199 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3201 s64 this_load, load;
3202 int idx, this_cpu, prev_cpu;
3203 unsigned long tl_per_task;
3204 struct task_group *tg;
3205 unsigned long weight;
3209 * If we wake multiple tasks be careful to not bounce
3210 * ourselves around too much.
3216 this_cpu = smp_processor_id();
3217 prev_cpu = task_cpu(p);
3218 load = source_load(prev_cpu, idx);
3219 this_load = target_load(this_cpu, idx);
3222 * If sync wakeup then subtract the (maximum possible)
3223 * effect of the currently running task from the load
3224 * of the current CPU:
3227 tg = task_group(current);
3228 weight = current->se.load.weight;
3230 this_load += effective_load(tg, this_cpu, -weight, -weight);
3231 load += effective_load(tg, prev_cpu, 0, -weight);
3235 weight = p->se.load.weight;
3238 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3239 * due to the sync cause above having dropped this_load to 0, we'll
3240 * always have an imbalance, but there's really nothing you can do
3241 * about that, so that's good too.
3243 * Otherwise check if either cpus are near enough in load to allow this
3244 * task to be woken on this_cpu.
3246 if (this_load > 0) {
3247 s64 this_eff_load, prev_eff_load;
3249 this_eff_load = 100;
3250 this_eff_load *= power_of(prev_cpu);
3251 this_eff_load *= this_load +
3252 effective_load(tg, this_cpu, weight, weight);
3254 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3255 prev_eff_load *= power_of(this_cpu);
3256 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3258 balanced = this_eff_load <= prev_eff_load;
3263 * If the currently running task will sleep within
3264 * a reasonable amount of time then attract this newly
3267 if (sync && balanced)
3270 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3271 tl_per_task = cpu_avg_load_per_task(this_cpu);
3274 (this_load <= load &&
3275 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3277 * This domain has SD_WAKE_AFFINE and
3278 * p is cache cold in this domain, and
3279 * there is no bad imbalance.
3281 schedstat_inc(sd, ttwu_move_affine);
3282 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3290 * find_idlest_group finds and returns the least busy CPU group within the
3293 static struct sched_group *
3294 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3295 int this_cpu, int load_idx)
3297 struct sched_group *idlest = NULL, *group = sd->groups;
3298 unsigned long min_load = ULONG_MAX, this_load = 0;
3299 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3302 unsigned long load, avg_load;
3306 /* Skip over this group if it has no CPUs allowed */
3307 if (!cpumask_intersects(sched_group_cpus(group),
3308 tsk_cpus_allowed(p)))
3311 local_group = cpumask_test_cpu(this_cpu,
3312 sched_group_cpus(group));
3314 /* Tally up the load of all CPUs in the group */
3317 for_each_cpu(i, sched_group_cpus(group)) {
3318 /* Bias balancing toward cpus of our domain */
3320 load = source_load(i, load_idx);
3322 load = target_load(i, load_idx);
3327 /* Adjust by relative CPU power of the group */
3328 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3331 this_load = avg_load;
3332 } else if (avg_load < min_load) {
3333 min_load = avg_load;
3336 } while (group = group->next, group != sd->groups);
3338 if (!idlest || 100*this_load < imbalance*min_load)
3344 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3347 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3349 unsigned long load, min_load = ULONG_MAX;
3353 /* Traverse only the allowed CPUs */
3354 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3355 load = weighted_cpuload(i);
3357 if (load < min_load || (load == min_load && i == this_cpu)) {
3367 * Try and locate an idle CPU in the sched_domain.
3369 static int select_idle_sibling(struct task_struct *p, int target)
3371 struct sched_domain *sd;
3372 struct sched_group *sg;
3373 int i = task_cpu(p);
3375 if (idle_cpu(target))
3379 * If the prevous cpu is cache affine and idle, don't be stupid.
3381 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3385 * Otherwise, iterate the domains and find an elegible idle cpu.
3387 sd = rcu_dereference(per_cpu(sd_llc, target));
3388 for_each_lower_domain(sd) {
3391 if (!cpumask_intersects(sched_group_cpus(sg),
3392 tsk_cpus_allowed(p)))
3395 for_each_cpu(i, sched_group_cpus(sg)) {
3396 if (i == target || !idle_cpu(i))
3400 target = cpumask_first_and(sched_group_cpus(sg),
3401 tsk_cpus_allowed(p));
3405 } while (sg != sd->groups);
3412 * sched_balance_self: balance the current task (running on cpu) in domains
3413 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3416 * Balance, ie. select the least loaded group.
3418 * Returns the target CPU number, or the same CPU if no balancing is needed.
3420 * preempt must be disabled.
3423 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3425 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3426 int cpu = smp_processor_id();
3427 int prev_cpu = task_cpu(p);
3429 int want_affine = 0;
3430 int sync = wake_flags & WF_SYNC;
3432 if (p->nr_cpus_allowed == 1)
3435 if (sd_flag & SD_BALANCE_WAKE) {
3436 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3442 for_each_domain(cpu, tmp) {
3443 if (!(tmp->flags & SD_LOAD_BALANCE))
3447 * If both cpu and prev_cpu are part of this domain,
3448 * cpu is a valid SD_WAKE_AFFINE target.
3450 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3451 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3456 if (tmp->flags & sd_flag)
3461 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3464 new_cpu = select_idle_sibling(p, prev_cpu);
3469 int load_idx = sd->forkexec_idx;
3470 struct sched_group *group;
3473 if (!(sd->flags & sd_flag)) {
3478 if (sd_flag & SD_BALANCE_WAKE)
3479 load_idx = sd->wake_idx;
3481 group = find_idlest_group(sd, p, cpu, load_idx);
3487 new_cpu = find_idlest_cpu(group, p, cpu);
3488 if (new_cpu == -1 || new_cpu == cpu) {
3489 /* Now try balancing at a lower domain level of cpu */
3494 /* Now try balancing at a lower domain level of new_cpu */
3496 weight = sd->span_weight;
3498 for_each_domain(cpu, tmp) {
3499 if (weight <= tmp->span_weight)
3501 if (tmp->flags & sd_flag)
3504 /* while loop will break here if sd == NULL */
3513 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3514 * cfs_rq_of(p) references at time of call are still valid and identify the
3515 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3516 * other assumptions, including the state of rq->lock, should be made.
3519 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3521 struct sched_entity *se = &p->se;
3522 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3525 * Load tracking: accumulate removed load so that it can be processed
3526 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3527 * to blocked load iff they have a positive decay-count. It can never
3528 * be negative here since on-rq tasks have decay-count == 0.
3530 if (se->avg.decay_count) {
3531 se->avg.decay_count = -__synchronize_entity_decay(se);
3532 atomic_long_add(se->avg.load_avg_contrib,
3533 &cfs_rq->removed_load);
3536 #endif /* CONFIG_SMP */
3538 static unsigned long
3539 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3541 unsigned long gran = sysctl_sched_wakeup_granularity;
3544 * Since its curr running now, convert the gran from real-time
3545 * to virtual-time in his units.
3547 * By using 'se' instead of 'curr' we penalize light tasks, so
3548 * they get preempted easier. That is, if 'se' < 'curr' then
3549 * the resulting gran will be larger, therefore penalizing the
3550 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3551 * be smaller, again penalizing the lighter task.
3553 * This is especially important for buddies when the leftmost
3554 * task is higher priority than the buddy.
3556 return calc_delta_fair(gran, se);
3560 * Should 'se' preempt 'curr'.
3574 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3576 s64 gran, vdiff = curr->vruntime - se->vruntime;
3581 gran = wakeup_gran(curr, se);
3588 static void set_last_buddy(struct sched_entity *se)
3590 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3593 for_each_sched_entity(se)
3594 cfs_rq_of(se)->last = se;
3597 static void set_next_buddy(struct sched_entity *se)
3599 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3602 for_each_sched_entity(se)
3603 cfs_rq_of(se)->next = se;
3606 static void set_skip_buddy(struct sched_entity *se)
3608 for_each_sched_entity(se)
3609 cfs_rq_of(se)->skip = se;
3613 * Preempt the current task with a newly woken task if needed:
3615 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3617 struct task_struct *curr = rq->curr;
3618 struct sched_entity *se = &curr->se, *pse = &p->se;
3619 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3620 int scale = cfs_rq->nr_running >= sched_nr_latency;
3621 int next_buddy_marked = 0;
3623 if (unlikely(se == pse))
3627 * This is possible from callers such as move_task(), in which we
3628 * unconditionally check_prempt_curr() after an enqueue (which may have
3629 * lead to a throttle). This both saves work and prevents false
3630 * next-buddy nomination below.
3632 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3635 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3636 set_next_buddy(pse);
3637 next_buddy_marked = 1;
3641 * We can come here with TIF_NEED_RESCHED already set from new task
3644 * Note: this also catches the edge-case of curr being in a throttled
3645 * group (e.g. via set_curr_task), since update_curr() (in the
3646 * enqueue of curr) will have resulted in resched being set. This
3647 * prevents us from potentially nominating it as a false LAST_BUDDY
3650 if (test_tsk_need_resched(curr))
3653 /* Idle tasks are by definition preempted by non-idle tasks. */
3654 if (unlikely(curr->policy == SCHED_IDLE) &&
3655 likely(p->policy != SCHED_IDLE))
3659 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3660 * is driven by the tick):
3662 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3665 find_matching_se(&se, &pse);
3666 update_curr(cfs_rq_of(se));
3668 if (wakeup_preempt_entity(se, pse) == 1) {
3670 * Bias pick_next to pick the sched entity that is
3671 * triggering this preemption.
3673 if (!next_buddy_marked)
3674 set_next_buddy(pse);
3683 * Only set the backward buddy when the current task is still
3684 * on the rq. This can happen when a wakeup gets interleaved
3685 * with schedule on the ->pre_schedule() or idle_balance()
3686 * point, either of which can * drop the rq lock.
3688 * Also, during early boot the idle thread is in the fair class,
3689 * for obvious reasons its a bad idea to schedule back to it.
3691 if (unlikely(!se->on_rq || curr == rq->idle))
3694 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3698 static struct task_struct *pick_next_task_fair(struct rq *rq)
3700 struct task_struct *p;
3701 struct cfs_rq *cfs_rq = &rq->cfs;
3702 struct sched_entity *se;
3704 if (!cfs_rq->nr_running)
3708 se = pick_next_entity(cfs_rq);
3709 set_next_entity(cfs_rq, se);
3710 cfs_rq = group_cfs_rq(se);
3714 if (hrtick_enabled(rq))
3715 hrtick_start_fair(rq, p);
3721 * Account for a descheduled task:
3723 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3725 struct sched_entity *se = &prev->se;
3726 struct cfs_rq *cfs_rq;
3728 for_each_sched_entity(se) {
3729 cfs_rq = cfs_rq_of(se);
3730 put_prev_entity(cfs_rq, se);
3735 * sched_yield() is very simple
3737 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3739 static void yield_task_fair(struct rq *rq)
3741 struct task_struct *curr = rq->curr;
3742 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3743 struct sched_entity *se = &curr->se;
3746 * Are we the only task in the tree?
3748 if (unlikely(rq->nr_running == 1))
3751 clear_buddies(cfs_rq, se);
3753 if (curr->policy != SCHED_BATCH) {
3754 update_rq_clock(rq);
3756 * Update run-time statistics of the 'current'.
3758 update_curr(cfs_rq);
3760 * Tell update_rq_clock() that we've just updated,
3761 * so we don't do microscopic update in schedule()
3762 * and double the fastpath cost.
3764 rq->skip_clock_update = 1;
3770 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3772 struct sched_entity *se = &p->se;
3774 /* throttled hierarchies are not runnable */
3775 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3778 /* Tell the scheduler that we'd really like pse to run next. */
3781 yield_task_fair(rq);
3787 /**************************************************
3788 * Fair scheduling class load-balancing methods.
3792 * The purpose of load-balancing is to achieve the same basic fairness the
3793 * per-cpu scheduler provides, namely provide a proportional amount of compute
3794 * time to each task. This is expressed in the following equation:
3796 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3798 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3799 * W_i,0 is defined as:
3801 * W_i,0 = \Sum_j w_i,j (2)
3803 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3804 * is derived from the nice value as per prio_to_weight[].
3806 * The weight average is an exponential decay average of the instantaneous
3809 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3811 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3812 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3813 * can also include other factors [XXX].
3815 * To achieve this balance we define a measure of imbalance which follows
3816 * directly from (1):
3818 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3820 * We them move tasks around to minimize the imbalance. In the continuous
3821 * function space it is obvious this converges, in the discrete case we get
3822 * a few fun cases generally called infeasible weight scenarios.
3825 * - infeasible weights;
3826 * - local vs global optima in the discrete case. ]
3831 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3832 * for all i,j solution, we create a tree of cpus that follows the hardware
3833 * topology where each level pairs two lower groups (or better). This results
3834 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3835 * tree to only the first of the previous level and we decrease the frequency
3836 * of load-balance at each level inv. proportional to the number of cpus in
3842 * \Sum { --- * --- * 2^i } = O(n) (5)
3844 * `- size of each group
3845 * | | `- number of cpus doing load-balance
3847 * `- sum over all levels
3849 * Coupled with a limit on how many tasks we can migrate every balance pass,
3850 * this makes (5) the runtime complexity of the balancer.
3852 * An important property here is that each CPU is still (indirectly) connected
3853 * to every other cpu in at most O(log n) steps:
3855 * The adjacency matrix of the resulting graph is given by:
3858 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3861 * And you'll find that:
3863 * A^(log_2 n)_i,j != 0 for all i,j (7)
3865 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3866 * The task movement gives a factor of O(m), giving a convergence complexity
3869 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3874 * In order to avoid CPUs going idle while there's still work to do, new idle
3875 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3876 * tree itself instead of relying on other CPUs to bring it work.
3878 * This adds some complexity to both (5) and (8) but it reduces the total idle
3886 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3889 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3894 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3896 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3898 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3901 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3902 * rewrite all of this once again.]
3905 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3907 #define LBF_ALL_PINNED 0x01
3908 #define LBF_NEED_BREAK 0x02
3909 #define LBF_DST_PINNED 0x04
3910 #define LBF_SOME_PINNED 0x08
3913 struct sched_domain *sd;
3921 struct cpumask *dst_grpmask;
3923 enum cpu_idle_type idle;
3925 /* The set of CPUs under consideration for load-balancing */
3926 struct cpumask *cpus;
3931 unsigned int loop_break;
3932 unsigned int loop_max;
3936 * move_task - move a task from one runqueue to another runqueue.
3937 * Both runqueues must be locked.
3939 static void move_task(struct task_struct *p, struct lb_env *env)
3941 deactivate_task(env->src_rq, p, 0);
3942 set_task_cpu(p, env->dst_cpu);
3943 activate_task(env->dst_rq, p, 0);
3944 check_preempt_curr(env->dst_rq, p, 0);
3948 * Is this task likely cache-hot:
3951 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3955 if (p->sched_class != &fair_sched_class)
3958 if (unlikely(p->policy == SCHED_IDLE))
3962 * Buddy candidates are cache hot:
3964 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3965 (&p->se == cfs_rq_of(&p->se)->next ||
3966 &p->se == cfs_rq_of(&p->se)->last))
3969 if (sysctl_sched_migration_cost == -1)
3971 if (sysctl_sched_migration_cost == 0)
3974 delta = now - p->se.exec_start;
3976 return delta < (s64)sysctl_sched_migration_cost;
3980 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3983 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3985 int tsk_cache_hot = 0;
3987 * We do not migrate tasks that are:
3988 * 1) throttled_lb_pair, or
3989 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3990 * 3) running (obviously), or
3991 * 4) are cache-hot on their current CPU.
3993 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3996 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3999 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4001 env->flags |= LBF_SOME_PINNED;
4004 * Remember if this task can be migrated to any other cpu in
4005 * our sched_group. We may want to revisit it if we couldn't
4006 * meet load balance goals by pulling other tasks on src_cpu.
4008 * Also avoid computing new_dst_cpu if we have already computed
4009 * one in current iteration.
4011 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4014 /* Prevent to re-select dst_cpu via env's cpus */
4015 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4016 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4017 env->flags |= LBF_DST_PINNED;
4018 env->new_dst_cpu = cpu;
4026 /* Record that we found atleast one task that could run on dst_cpu */
4027 env->flags &= ~LBF_ALL_PINNED;
4029 if (task_running(env->src_rq, p)) {
4030 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4035 * Aggressive migration if:
4036 * 1) task is cache cold, or
4037 * 2) too many balance attempts have failed.
4040 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4041 if (!tsk_cache_hot ||
4042 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4044 if (tsk_cache_hot) {
4045 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4046 schedstat_inc(p, se.statistics.nr_forced_migrations);
4052 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4057 * move_one_task tries to move exactly one task from busiest to this_rq, as
4058 * part of active balancing operations within "domain".
4059 * Returns 1 if successful and 0 otherwise.
4061 * Called with both runqueues locked.
4063 static int move_one_task(struct lb_env *env)
4065 struct task_struct *p, *n;
4067 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4068 if (!can_migrate_task(p, env))
4073 * Right now, this is only the second place move_task()
4074 * is called, so we can safely collect move_task()
4075 * stats here rather than inside move_task().
4077 schedstat_inc(env->sd, lb_gained[env->idle]);
4083 static unsigned long task_h_load(struct task_struct *p);
4085 static const unsigned int sched_nr_migrate_break = 32;
4088 * move_tasks tries to move up to imbalance weighted load from busiest to
4089 * this_rq, as part of a balancing operation within domain "sd".
4090 * Returns 1 if successful and 0 otherwise.
4092 * Called with both runqueues locked.
4094 static int move_tasks(struct lb_env *env)
4096 struct list_head *tasks = &env->src_rq->cfs_tasks;
4097 struct task_struct *p;
4101 if (env->imbalance <= 0)
4104 while (!list_empty(tasks)) {
4105 p = list_first_entry(tasks, struct task_struct, se.group_node);
4108 /* We've more or less seen every task there is, call it quits */
4109 if (env->loop > env->loop_max)
4112 /* take a breather every nr_migrate tasks */
4113 if (env->loop > env->loop_break) {
4114 env->loop_break += sched_nr_migrate_break;
4115 env->flags |= LBF_NEED_BREAK;
4119 if (!can_migrate_task(p, env))
4122 load = task_h_load(p);
4124 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4127 if ((load / 2) > env->imbalance)
4132 env->imbalance -= load;
4134 #ifdef CONFIG_PREEMPT
4136 * NEWIDLE balancing is a source of latency, so preemptible
4137 * kernels will stop after the first task is pulled to minimize
4138 * the critical section.
4140 if (env->idle == CPU_NEWLY_IDLE)
4145 * We only want to steal up to the prescribed amount of
4148 if (env->imbalance <= 0)
4153 list_move_tail(&p->se.group_node, tasks);
4157 * Right now, this is one of only two places move_task() is called,
4158 * so we can safely collect move_task() stats here rather than
4159 * inside move_task().
4161 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4166 #ifdef CONFIG_FAIR_GROUP_SCHED
4168 * update tg->load_weight by folding this cpu's load_avg
4170 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4172 struct sched_entity *se = tg->se[cpu];
4173 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4175 /* throttled entities do not contribute to load */
4176 if (throttled_hierarchy(cfs_rq))
4179 update_cfs_rq_blocked_load(cfs_rq, 1);
4182 update_entity_load_avg(se, 1);
4184 * We pivot on our runnable average having decayed to zero for
4185 * list removal. This generally implies that all our children
4186 * have also been removed (modulo rounding error or bandwidth
4187 * control); however, such cases are rare and we can fix these
4190 * TODO: fix up out-of-order children on enqueue.
4192 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4193 list_del_leaf_cfs_rq(cfs_rq);
4195 struct rq *rq = rq_of(cfs_rq);
4196 update_rq_runnable_avg(rq, rq->nr_running);
4200 static void update_blocked_averages(int cpu)
4202 struct rq *rq = cpu_rq(cpu);
4203 struct cfs_rq *cfs_rq;
4204 unsigned long flags;
4206 raw_spin_lock_irqsave(&rq->lock, flags);
4207 update_rq_clock(rq);
4209 * Iterates the task_group tree in a bottom up fashion, see
4210 * list_add_leaf_cfs_rq() for details.
4212 for_each_leaf_cfs_rq(rq, cfs_rq) {
4214 * Note: We may want to consider periodically releasing
4215 * rq->lock about these updates so that creating many task
4216 * groups does not result in continually extending hold time.
4218 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4221 raw_spin_unlock_irqrestore(&rq->lock, flags);
4225 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4226 * This needs to be done in a top-down fashion because the load of a child
4227 * group is a fraction of its parents load.
4229 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4231 struct rq *rq = rq_of(cfs_rq);
4232 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4233 unsigned long now = jiffies;
4236 if (cfs_rq->last_h_load_update == now)
4239 cfs_rq->h_load_next = NULL;
4240 for_each_sched_entity(se) {
4241 cfs_rq = cfs_rq_of(se);
4242 cfs_rq->h_load_next = se;
4243 if (cfs_rq->last_h_load_update == now)
4248 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4249 cfs_rq->last_h_load_update = now;
4252 while ((se = cfs_rq->h_load_next) != NULL) {
4253 load = cfs_rq->h_load;
4254 load = div64_ul(load * se->avg.load_avg_contrib,
4255 cfs_rq->runnable_load_avg + 1);
4256 cfs_rq = group_cfs_rq(se);
4257 cfs_rq->h_load = load;
4258 cfs_rq->last_h_load_update = now;
4262 static unsigned long task_h_load(struct task_struct *p)
4264 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4266 update_cfs_rq_h_load(cfs_rq);
4267 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4268 cfs_rq->runnable_load_avg + 1);
4271 static inline void update_blocked_averages(int cpu)
4275 static unsigned long task_h_load(struct task_struct *p)
4277 return p->se.avg.load_avg_contrib;
4281 /********** Helpers for find_busiest_group ************************/
4283 * sg_lb_stats - stats of a sched_group required for load_balancing
4285 struct sg_lb_stats {
4286 unsigned long avg_load; /*Avg load across the CPUs of the group */
4287 unsigned long group_load; /* Total load over the CPUs of the group */
4288 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4289 unsigned long load_per_task;
4290 unsigned long group_power;
4291 unsigned int sum_nr_running; /* Nr tasks running in the group */
4292 unsigned int group_capacity;
4293 unsigned int idle_cpus;
4294 unsigned int group_weight;
4295 int group_imb; /* Is there an imbalance in the group ? */
4296 int group_has_capacity; /* Is there extra capacity in the group? */
4300 * sd_lb_stats - Structure to store the statistics of a sched_domain
4301 * during load balancing.
4303 struct sd_lb_stats {
4304 struct sched_group *busiest; /* Busiest group in this sd */
4305 struct sched_group *local; /* Local group in this sd */
4306 unsigned long total_load; /* Total load of all groups in sd */
4307 unsigned long total_pwr; /* Total power of all groups in sd */
4308 unsigned long avg_load; /* Average load across all groups in sd */
4310 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4311 struct sg_lb_stats local_stat; /* Statistics of the local group */
4314 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4317 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4318 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4319 * We must however clear busiest_stat::avg_load because
4320 * update_sd_pick_busiest() reads this before assignment.
4322 *sds = (struct sd_lb_stats){
4334 * get_sd_load_idx - Obtain the load index for a given sched domain.
4335 * @sd: The sched_domain whose load_idx is to be obtained.
4336 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4338 * Return: The load index.
4340 static inline int get_sd_load_idx(struct sched_domain *sd,
4341 enum cpu_idle_type idle)
4347 load_idx = sd->busy_idx;
4350 case CPU_NEWLY_IDLE:
4351 load_idx = sd->newidle_idx;
4354 load_idx = sd->idle_idx;
4361 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4363 return SCHED_POWER_SCALE;
4366 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4368 return default_scale_freq_power(sd, cpu);
4371 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4373 unsigned long weight = sd->span_weight;
4374 unsigned long smt_gain = sd->smt_gain;
4381 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4383 return default_scale_smt_power(sd, cpu);
4386 static unsigned long scale_rt_power(int cpu)
4388 struct rq *rq = cpu_rq(cpu);
4389 u64 total, available, age_stamp, avg;
4392 * Since we're reading these variables without serialization make sure
4393 * we read them once before doing sanity checks on them.
4395 age_stamp = ACCESS_ONCE(rq->age_stamp);
4396 avg = ACCESS_ONCE(rq->rt_avg);
4398 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4400 if (unlikely(total < avg)) {
4401 /* Ensures that power won't end up being negative */
4404 available = total - avg;
4407 if (unlikely((s64)total < SCHED_POWER_SCALE))
4408 total = SCHED_POWER_SCALE;
4410 total >>= SCHED_POWER_SHIFT;
4412 return div_u64(available, total);
4415 static void update_cpu_power(struct sched_domain *sd, int cpu)
4417 unsigned long weight = sd->span_weight;
4418 unsigned long power = SCHED_POWER_SCALE;
4419 struct sched_group *sdg = sd->groups;
4421 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4422 if (sched_feat(ARCH_POWER))
4423 power *= arch_scale_smt_power(sd, cpu);
4425 power *= default_scale_smt_power(sd, cpu);
4427 power >>= SCHED_POWER_SHIFT;
4430 sdg->sgp->power_orig = power;
4432 if (sched_feat(ARCH_POWER))
4433 power *= arch_scale_freq_power(sd, cpu);
4435 power *= default_scale_freq_power(sd, cpu);
4437 power >>= SCHED_POWER_SHIFT;
4439 power *= scale_rt_power(cpu);
4440 power >>= SCHED_POWER_SHIFT;
4445 cpu_rq(cpu)->cpu_power = power;
4446 sdg->sgp->power = power;
4449 void update_group_power(struct sched_domain *sd, int cpu)
4451 struct sched_domain *child = sd->child;
4452 struct sched_group *group, *sdg = sd->groups;
4453 unsigned long power, power_orig;
4454 unsigned long interval;
4456 interval = msecs_to_jiffies(sd->balance_interval);
4457 interval = clamp(interval, 1UL, max_load_balance_interval);
4458 sdg->sgp->next_update = jiffies + interval;
4461 update_cpu_power(sd, cpu);
4465 power_orig = power = 0;
4467 if (child->flags & SD_OVERLAP) {
4469 * SD_OVERLAP domains cannot assume that child groups
4470 * span the current group.
4473 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4474 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4476 power_orig += sg->sgp->power_orig;
4477 power += sg->sgp->power;
4481 * !SD_OVERLAP domains can assume that child groups
4482 * span the current group.
4485 group = child->groups;
4487 power_orig += group->sgp->power_orig;
4488 power += group->sgp->power;
4489 group = group->next;
4490 } while (group != child->groups);
4493 sdg->sgp->power_orig = power_orig;
4494 sdg->sgp->power = power;
4498 * Try and fix up capacity for tiny siblings, this is needed when
4499 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4500 * which on its own isn't powerful enough.
4502 * See update_sd_pick_busiest() and check_asym_packing().
4505 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4508 * Only siblings can have significantly less than SCHED_POWER_SCALE
4510 if (!(sd->flags & SD_SHARE_CPUPOWER))
4514 * If ~90% of the cpu_power is still there, we're good.
4516 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4523 * Group imbalance indicates (and tries to solve) the problem where balancing
4524 * groups is inadequate due to tsk_cpus_allowed() constraints.
4526 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4527 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4530 * { 0 1 2 3 } { 4 5 6 7 }
4533 * If we were to balance group-wise we'd place two tasks in the first group and
4534 * two tasks in the second group. Clearly this is undesired as it will overload
4535 * cpu 3 and leave one of the cpus in the second group unused.
4537 * The current solution to this issue is detecting the skew in the first group
4538 * by noticing the lower domain failed to reach balance and had difficulty
4539 * moving tasks due to affinity constraints.
4541 * When this is so detected; this group becomes a candidate for busiest; see
4542 * update_sd_pick_busiest(). And calculcate_imbalance() and
4543 * find_busiest_group() avoid some of the usual balance conditions to allow it
4544 * to create an effective group imbalance.
4546 * This is a somewhat tricky proposition since the next run might not find the
4547 * group imbalance and decide the groups need to be balanced again. A most
4548 * subtle and fragile situation.
4551 static inline int sg_imbalanced(struct sched_group *group)
4553 return group->sgp->imbalance;
4557 * Compute the group capacity.
4559 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4560 * first dividing out the smt factor and computing the actual number of cores
4561 * and limit power unit capacity with that.
4563 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4565 unsigned int capacity, smt, cpus;
4566 unsigned int power, power_orig;
4568 power = group->sgp->power;
4569 power_orig = group->sgp->power_orig;
4570 cpus = group->group_weight;
4572 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4573 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4574 capacity = cpus / smt; /* cores */
4576 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4578 capacity = fix_small_capacity(env->sd, group);
4584 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4585 * @env: The load balancing environment.
4586 * @group: sched_group whose statistics are to be updated.
4587 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4588 * @local_group: Does group contain this_cpu.
4589 * @sgs: variable to hold the statistics for this group.
4591 static inline void update_sg_lb_stats(struct lb_env *env,
4592 struct sched_group *group, int load_idx,
4593 int local_group, struct sg_lb_stats *sgs)
4595 unsigned long nr_running;
4599 memset(sgs, 0, sizeof(*sgs));
4601 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4602 struct rq *rq = cpu_rq(i);
4604 nr_running = rq->nr_running;
4606 /* Bias balancing toward cpus of our domain */
4608 load = target_load(i, load_idx);
4610 load = source_load(i, load_idx);
4612 sgs->group_load += load;
4613 sgs->sum_nr_running += nr_running;
4614 sgs->sum_weighted_load += weighted_cpuload(i);
4619 /* Adjust by relative CPU power of the group */
4620 sgs->group_power = group->sgp->power;
4621 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4623 if (sgs->sum_nr_running)
4624 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4626 sgs->group_weight = group->group_weight;
4628 sgs->group_imb = sg_imbalanced(group);
4629 sgs->group_capacity = sg_capacity(env, group);
4631 if (sgs->group_capacity > sgs->sum_nr_running)
4632 sgs->group_has_capacity = 1;
4636 * update_sd_pick_busiest - return 1 on busiest group
4637 * @env: The load balancing environment.
4638 * @sds: sched_domain statistics
4639 * @sg: sched_group candidate to be checked for being the busiest
4640 * @sgs: sched_group statistics
4642 * Determine if @sg is a busier group than the previously selected
4645 * Return: %true if @sg is a busier group than the previously selected
4646 * busiest group. %false otherwise.
4648 static bool update_sd_pick_busiest(struct lb_env *env,
4649 struct sd_lb_stats *sds,
4650 struct sched_group *sg,
4651 struct sg_lb_stats *sgs)
4653 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4656 if (sgs->sum_nr_running > sgs->group_capacity)
4663 * ASYM_PACKING needs to move all the work to the lowest
4664 * numbered CPUs in the group, therefore mark all groups
4665 * higher than ourself as busy.
4667 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4668 env->dst_cpu < group_first_cpu(sg)) {
4672 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4680 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4681 * @env: The load balancing environment.
4682 * @balance: Should we balance.
4683 * @sds: variable to hold the statistics for this sched_domain.
4685 static inline void update_sd_lb_stats(struct lb_env *env,
4686 struct sd_lb_stats *sds)
4688 struct sched_domain *child = env->sd->child;
4689 struct sched_group *sg = env->sd->groups;
4690 struct sg_lb_stats tmp_sgs;
4691 int load_idx, prefer_sibling = 0;
4693 if (child && child->flags & SD_PREFER_SIBLING)
4696 load_idx = get_sd_load_idx(env->sd, env->idle);
4699 struct sg_lb_stats *sgs = &tmp_sgs;
4702 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4705 sgs = &sds->local_stat;
4707 if (env->idle != CPU_NEWLY_IDLE ||
4708 time_after_eq(jiffies, sg->sgp->next_update))
4709 update_group_power(env->sd, env->dst_cpu);
4712 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4718 * In case the child domain prefers tasks go to siblings
4719 * first, lower the sg capacity to one so that we'll try
4720 * and move all the excess tasks away. We lower the capacity
4721 * of a group only if the local group has the capacity to fit
4722 * these excess tasks, i.e. nr_running < group_capacity. The
4723 * extra check prevents the case where you always pull from the
4724 * heaviest group when it is already under-utilized (possible
4725 * with a large weight task outweighs the tasks on the system).
4727 if (prefer_sibling && sds->local &&
4728 sds->local_stat.group_has_capacity)
4729 sgs->group_capacity = min(sgs->group_capacity, 1U);
4731 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4733 sds->busiest_stat = *sgs;
4737 /* Now, start updating sd_lb_stats */
4738 sds->total_load += sgs->group_load;
4739 sds->total_pwr += sgs->group_power;
4742 } while (sg != env->sd->groups);
4746 * check_asym_packing - Check to see if the group is packed into the
4749 * This is primarily intended to used at the sibling level. Some
4750 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4751 * case of POWER7, it can move to lower SMT modes only when higher
4752 * threads are idle. When in lower SMT modes, the threads will
4753 * perform better since they share less core resources. Hence when we
4754 * have idle threads, we want them to be the higher ones.
4756 * This packing function is run on idle threads. It checks to see if
4757 * the busiest CPU in this domain (core in the P7 case) has a higher
4758 * CPU number than the packing function is being run on. Here we are
4759 * assuming lower CPU number will be equivalent to lower a SMT thread
4762 * Return: 1 when packing is required and a task should be moved to
4763 * this CPU. The amount of the imbalance is returned in *imbalance.
4765 * @env: The load balancing environment.
4766 * @sds: Statistics of the sched_domain which is to be packed
4768 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4772 if (!(env->sd->flags & SD_ASYM_PACKING))
4778 busiest_cpu = group_first_cpu(sds->busiest);
4779 if (env->dst_cpu > busiest_cpu)
4782 env->imbalance = DIV_ROUND_CLOSEST(
4783 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
4790 * fix_small_imbalance - Calculate the minor imbalance that exists
4791 * amongst the groups of a sched_domain, during
4793 * @env: The load balancing environment.
4794 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4797 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4799 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4800 unsigned int imbn = 2;
4801 unsigned long scaled_busy_load_per_task;
4802 struct sg_lb_stats *local, *busiest;
4804 local = &sds->local_stat;
4805 busiest = &sds->busiest_stat;
4807 if (!local->sum_nr_running)
4808 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
4809 else if (busiest->load_per_task > local->load_per_task)
4812 scaled_busy_load_per_task =
4813 (busiest->load_per_task * SCHED_POWER_SCALE) /
4814 busiest->group_power;
4816 if (busiest->avg_load + scaled_busy_load_per_task >=
4817 local->avg_load + (scaled_busy_load_per_task * imbn)) {
4818 env->imbalance = busiest->load_per_task;
4823 * OK, we don't have enough imbalance to justify moving tasks,
4824 * however we may be able to increase total CPU power used by
4828 pwr_now += busiest->group_power *
4829 min(busiest->load_per_task, busiest->avg_load);
4830 pwr_now += local->group_power *
4831 min(local->load_per_task, local->avg_load);
4832 pwr_now /= SCHED_POWER_SCALE;
4834 /* Amount of load we'd subtract */
4835 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4836 busiest->group_power;
4837 if (busiest->avg_load > tmp) {
4838 pwr_move += busiest->group_power *
4839 min(busiest->load_per_task,
4840 busiest->avg_load - tmp);
4843 /* Amount of load we'd add */
4844 if (busiest->avg_load * busiest->group_power <
4845 busiest->load_per_task * SCHED_POWER_SCALE) {
4846 tmp = (busiest->avg_load * busiest->group_power) /
4849 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4852 pwr_move += local->group_power *
4853 min(local->load_per_task, local->avg_load + tmp);
4854 pwr_move /= SCHED_POWER_SCALE;
4856 /* Move if we gain throughput */
4857 if (pwr_move > pwr_now)
4858 env->imbalance = busiest->load_per_task;
4862 * calculate_imbalance - Calculate the amount of imbalance present within the
4863 * groups of a given sched_domain during load balance.
4864 * @env: load balance environment
4865 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4867 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4869 unsigned long max_pull, load_above_capacity = ~0UL;
4870 struct sg_lb_stats *local, *busiest;
4872 local = &sds->local_stat;
4873 busiest = &sds->busiest_stat;
4875 if (busiest->group_imb) {
4877 * In the group_imb case we cannot rely on group-wide averages
4878 * to ensure cpu-load equilibrium, look at wider averages. XXX
4880 busiest->load_per_task =
4881 min(busiest->load_per_task, sds->avg_load);
4885 * In the presence of smp nice balancing, certain scenarios can have
4886 * max load less than avg load(as we skip the groups at or below
4887 * its cpu_power, while calculating max_load..)
4889 if (busiest->avg_load <= sds->avg_load ||
4890 local->avg_load >= sds->avg_load) {
4892 return fix_small_imbalance(env, sds);
4895 if (!busiest->group_imb) {
4897 * Don't want to pull so many tasks that a group would go idle.
4898 * Except of course for the group_imb case, since then we might
4899 * have to drop below capacity to reach cpu-load equilibrium.
4901 load_above_capacity =
4902 (busiest->sum_nr_running - busiest->group_capacity);
4904 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4905 load_above_capacity /= busiest->group_power;
4909 * We're trying to get all the cpus to the average_load, so we don't
4910 * want to push ourselves above the average load, nor do we wish to
4911 * reduce the max loaded cpu below the average load. At the same time,
4912 * we also don't want to reduce the group load below the group capacity
4913 * (so that we can implement power-savings policies etc). Thus we look
4914 * for the minimum possible imbalance.
4916 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
4918 /* How much load to actually move to equalise the imbalance */
4919 env->imbalance = min(
4920 max_pull * busiest->group_power,
4921 (sds->avg_load - local->avg_load) * local->group_power
4922 ) / SCHED_POWER_SCALE;
4925 * if *imbalance is less than the average load per runnable task
4926 * there is no guarantee that any tasks will be moved so we'll have
4927 * a think about bumping its value to force at least one task to be
4930 if (env->imbalance < busiest->load_per_task)
4931 return fix_small_imbalance(env, sds);
4934 /******* find_busiest_group() helpers end here *********************/
4937 * find_busiest_group - Returns the busiest group within the sched_domain
4938 * if there is an imbalance. If there isn't an imbalance, and
4939 * the user has opted for power-savings, it returns a group whose
4940 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4941 * such a group exists.
4943 * Also calculates the amount of weighted load which should be moved
4944 * to restore balance.
4946 * @env: The load balancing environment.
4948 * Return: - The busiest group if imbalance exists.
4949 * - If no imbalance and user has opted for power-savings balance,
4950 * return the least loaded group whose CPUs can be
4951 * put to idle by rebalancing its tasks onto our group.
4953 static struct sched_group *find_busiest_group(struct lb_env *env)
4955 struct sg_lb_stats *local, *busiest;
4956 struct sd_lb_stats sds;
4958 init_sd_lb_stats(&sds);
4961 * Compute the various statistics relavent for load balancing at
4964 update_sd_lb_stats(env, &sds);
4965 local = &sds.local_stat;
4966 busiest = &sds.busiest_stat;
4968 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4969 check_asym_packing(env, &sds))
4972 /* There is no busy sibling group to pull tasks from */
4973 if (!sds.busiest || busiest->sum_nr_running == 0)
4976 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4979 * If the busiest group is imbalanced the below checks don't
4980 * work because they assume all things are equal, which typically
4981 * isn't true due to cpus_allowed constraints and the like.
4983 if (busiest->group_imb)
4986 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4987 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
4988 !busiest->group_has_capacity)
4992 * If the local group is more busy than the selected busiest group
4993 * don't try and pull any tasks.
4995 if (local->avg_load >= busiest->avg_load)
4999 * Don't pull any tasks if this group is already above the domain
5002 if (local->avg_load >= sds.avg_load)
5005 if (env->idle == CPU_IDLE) {
5007 * This cpu is idle. If the busiest group load doesn't
5008 * have more tasks than the number of available cpu's and
5009 * there is no imbalance between this and busiest group
5010 * wrt to idle cpu's, it is balanced.
5012 if ((local->idle_cpus < busiest->idle_cpus) &&
5013 busiest->sum_nr_running <= busiest->group_weight)
5017 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5018 * imbalance_pct to be conservative.
5020 if (100 * busiest->avg_load <=
5021 env->sd->imbalance_pct * local->avg_load)
5026 /* Looks like there is an imbalance. Compute it */
5027 calculate_imbalance(env, &sds);
5036 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5038 static struct rq *find_busiest_queue(struct lb_env *env,
5039 struct sched_group *group)
5041 struct rq *busiest = NULL, *rq;
5042 unsigned long busiest_load = 0, busiest_power = 1;
5045 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5046 unsigned long power = power_of(i);
5047 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5052 capacity = fix_small_capacity(env->sd, group);
5055 wl = weighted_cpuload(i);
5058 * When comparing with imbalance, use weighted_cpuload()
5059 * which is not scaled with the cpu power.
5061 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5065 * For the load comparisons with the other cpu's, consider
5066 * the weighted_cpuload() scaled with the cpu power, so that
5067 * the load can be moved away from the cpu that is potentially
5068 * running at a lower capacity.
5070 * Thus we're looking for max(wl_i / power_i), crosswise
5071 * multiplication to rid ourselves of the division works out
5072 * to: wl_i * power_j > wl_j * power_i; where j is our
5075 if (wl * busiest_power > busiest_load * power) {
5077 busiest_power = power;
5086 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5087 * so long as it is large enough.
5089 #define MAX_PINNED_INTERVAL 512
5091 /* Working cpumask for load_balance and load_balance_newidle. */
5092 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5094 static int need_active_balance(struct lb_env *env)
5096 struct sched_domain *sd = env->sd;
5098 if (env->idle == CPU_NEWLY_IDLE) {
5101 * ASYM_PACKING needs to force migrate tasks from busy but
5102 * higher numbered CPUs in order to pack all tasks in the
5103 * lowest numbered CPUs.
5105 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5109 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5112 static int active_load_balance_cpu_stop(void *data);
5114 static int should_we_balance(struct lb_env *env)
5116 struct sched_group *sg = env->sd->groups;
5117 struct cpumask *sg_cpus, *sg_mask;
5118 int cpu, balance_cpu = -1;
5121 * In the newly idle case, we will allow all the cpu's
5122 * to do the newly idle load balance.
5124 if (env->idle == CPU_NEWLY_IDLE)
5127 sg_cpus = sched_group_cpus(sg);
5128 sg_mask = sched_group_mask(sg);
5129 /* Try to find first idle cpu */
5130 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5131 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5138 if (balance_cpu == -1)
5139 balance_cpu = group_balance_cpu(sg);
5142 * First idle cpu or the first cpu(busiest) in this sched group
5143 * is eligible for doing load balancing at this and above domains.
5145 return balance_cpu == env->dst_cpu;
5149 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5150 * tasks if there is an imbalance.
5152 static int load_balance(int this_cpu, struct rq *this_rq,
5153 struct sched_domain *sd, enum cpu_idle_type idle,
5154 int *continue_balancing)
5156 int ld_moved, cur_ld_moved, active_balance = 0;
5157 struct sched_domain *sd_parent = sd->parent;
5158 struct sched_group *group;
5160 unsigned long flags;
5161 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5163 struct lb_env env = {
5165 .dst_cpu = this_cpu,
5167 .dst_grpmask = sched_group_cpus(sd->groups),
5169 .loop_break = sched_nr_migrate_break,
5174 * For NEWLY_IDLE load_balancing, we don't need to consider
5175 * other cpus in our group
5177 if (idle == CPU_NEWLY_IDLE)
5178 env.dst_grpmask = NULL;
5180 cpumask_copy(cpus, cpu_active_mask);
5182 schedstat_inc(sd, lb_count[idle]);
5185 if (!should_we_balance(&env)) {
5186 *continue_balancing = 0;
5190 group = find_busiest_group(&env);
5192 schedstat_inc(sd, lb_nobusyg[idle]);
5196 busiest = find_busiest_queue(&env, group);
5198 schedstat_inc(sd, lb_nobusyq[idle]);
5202 BUG_ON(busiest == env.dst_rq);
5204 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5207 if (busiest->nr_running > 1) {
5209 * Attempt to move tasks. If find_busiest_group has found
5210 * an imbalance but busiest->nr_running <= 1, the group is
5211 * still unbalanced. ld_moved simply stays zero, so it is
5212 * correctly treated as an imbalance.
5214 env.flags |= LBF_ALL_PINNED;
5215 env.src_cpu = busiest->cpu;
5216 env.src_rq = busiest;
5217 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5220 local_irq_save(flags);
5221 double_rq_lock(env.dst_rq, busiest);
5224 * cur_ld_moved - load moved in current iteration
5225 * ld_moved - cumulative load moved across iterations
5227 cur_ld_moved = move_tasks(&env);
5228 ld_moved += cur_ld_moved;
5229 double_rq_unlock(env.dst_rq, busiest);
5230 local_irq_restore(flags);
5233 * some other cpu did the load balance for us.
5235 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5236 resched_cpu(env.dst_cpu);
5238 if (env.flags & LBF_NEED_BREAK) {
5239 env.flags &= ~LBF_NEED_BREAK;
5244 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5245 * us and move them to an alternate dst_cpu in our sched_group
5246 * where they can run. The upper limit on how many times we
5247 * iterate on same src_cpu is dependent on number of cpus in our
5250 * This changes load balance semantics a bit on who can move
5251 * load to a given_cpu. In addition to the given_cpu itself
5252 * (or a ilb_cpu acting on its behalf where given_cpu is
5253 * nohz-idle), we now have balance_cpu in a position to move
5254 * load to given_cpu. In rare situations, this may cause
5255 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5256 * _independently_ and at _same_ time to move some load to
5257 * given_cpu) causing exceess load to be moved to given_cpu.
5258 * This however should not happen so much in practice and
5259 * moreover subsequent load balance cycles should correct the
5260 * excess load moved.
5262 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5264 /* Prevent to re-select dst_cpu via env's cpus */
5265 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5267 env.dst_rq = cpu_rq(env.new_dst_cpu);
5268 env.dst_cpu = env.new_dst_cpu;
5269 env.flags &= ~LBF_DST_PINNED;
5271 env.loop_break = sched_nr_migrate_break;
5274 * Go back to "more_balance" rather than "redo" since we
5275 * need to continue with same src_cpu.
5281 * We failed to reach balance because of affinity.
5284 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5286 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5287 *group_imbalance = 1;
5288 } else if (*group_imbalance)
5289 *group_imbalance = 0;
5292 /* All tasks on this runqueue were pinned by CPU affinity */
5293 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5294 cpumask_clear_cpu(cpu_of(busiest), cpus);
5295 if (!cpumask_empty(cpus)) {
5297 env.loop_break = sched_nr_migrate_break;
5305 schedstat_inc(sd, lb_failed[idle]);
5307 * Increment the failure counter only on periodic balance.
5308 * We do not want newidle balance, which can be very
5309 * frequent, pollute the failure counter causing
5310 * excessive cache_hot migrations and active balances.
5312 if (idle != CPU_NEWLY_IDLE)
5313 sd->nr_balance_failed++;
5315 if (need_active_balance(&env)) {
5316 raw_spin_lock_irqsave(&busiest->lock, flags);
5318 /* don't kick the active_load_balance_cpu_stop,
5319 * if the curr task on busiest cpu can't be
5322 if (!cpumask_test_cpu(this_cpu,
5323 tsk_cpus_allowed(busiest->curr))) {
5324 raw_spin_unlock_irqrestore(&busiest->lock,
5326 env.flags |= LBF_ALL_PINNED;
5327 goto out_one_pinned;
5331 * ->active_balance synchronizes accesses to
5332 * ->active_balance_work. Once set, it's cleared
5333 * only after active load balance is finished.
5335 if (!busiest->active_balance) {
5336 busiest->active_balance = 1;
5337 busiest->push_cpu = this_cpu;
5340 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5342 if (active_balance) {
5343 stop_one_cpu_nowait(cpu_of(busiest),
5344 active_load_balance_cpu_stop, busiest,
5345 &busiest->active_balance_work);
5349 * We've kicked active balancing, reset the failure
5352 sd->nr_balance_failed = sd->cache_nice_tries+1;
5355 sd->nr_balance_failed = 0;
5357 if (likely(!active_balance)) {
5358 /* We were unbalanced, so reset the balancing interval */
5359 sd->balance_interval = sd->min_interval;
5362 * If we've begun active balancing, start to back off. This
5363 * case may not be covered by the all_pinned logic if there
5364 * is only 1 task on the busy runqueue (because we don't call
5367 if (sd->balance_interval < sd->max_interval)
5368 sd->balance_interval *= 2;
5374 schedstat_inc(sd, lb_balanced[idle]);
5376 sd->nr_balance_failed = 0;
5379 /* tune up the balancing interval */
5380 if (((env.flags & LBF_ALL_PINNED) &&
5381 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5382 (sd->balance_interval < sd->max_interval))
5383 sd->balance_interval *= 2;
5391 * idle_balance is called by schedule() if this_cpu is about to become
5392 * idle. Attempts to pull tasks from other CPUs.
5394 void idle_balance(int this_cpu, struct rq *this_rq)
5396 struct sched_domain *sd;
5397 int pulled_task = 0;
5398 unsigned long next_balance = jiffies + HZ;
5401 this_rq->idle_stamp = rq_clock(this_rq);
5403 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5407 * Drop the rq->lock, but keep IRQ/preempt disabled.
5409 raw_spin_unlock(&this_rq->lock);
5411 update_blocked_averages(this_cpu);
5413 for_each_domain(this_cpu, sd) {
5414 unsigned long interval;
5415 int continue_balancing = 1;
5416 u64 t0, domain_cost;
5418 if (!(sd->flags & SD_LOAD_BALANCE))
5421 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5424 if (sd->flags & SD_BALANCE_NEWIDLE) {
5425 t0 = sched_clock_cpu(this_cpu);
5427 /* If we've pulled tasks over stop searching: */
5428 pulled_task = load_balance(this_cpu, this_rq,
5430 &continue_balancing);
5432 domain_cost = sched_clock_cpu(this_cpu) - t0;
5433 if (domain_cost > sd->max_newidle_lb_cost)
5434 sd->max_newidle_lb_cost = domain_cost;
5436 curr_cost += domain_cost;
5439 interval = msecs_to_jiffies(sd->balance_interval);
5440 if (time_after(next_balance, sd->last_balance + interval))
5441 next_balance = sd->last_balance + interval;
5443 this_rq->idle_stamp = 0;
5449 raw_spin_lock(&this_rq->lock);
5451 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5453 * We are going idle. next_balance may be set based on
5454 * a busy processor. So reset next_balance.
5456 this_rq->next_balance = next_balance;
5459 if (curr_cost > this_rq->max_idle_balance_cost)
5460 this_rq->max_idle_balance_cost = curr_cost;
5464 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5465 * running tasks off the busiest CPU onto idle CPUs. It requires at
5466 * least 1 task to be running on each physical CPU where possible, and
5467 * avoids physical / logical imbalances.
5469 static int active_load_balance_cpu_stop(void *data)
5471 struct rq *busiest_rq = data;
5472 int busiest_cpu = cpu_of(busiest_rq);
5473 int target_cpu = busiest_rq->push_cpu;
5474 struct rq *target_rq = cpu_rq(target_cpu);
5475 struct sched_domain *sd;
5477 raw_spin_lock_irq(&busiest_rq->lock);
5479 /* make sure the requested cpu hasn't gone down in the meantime */
5480 if (unlikely(busiest_cpu != smp_processor_id() ||
5481 !busiest_rq->active_balance))
5484 /* Is there any task to move? */
5485 if (busiest_rq->nr_running <= 1)
5489 * This condition is "impossible", if it occurs
5490 * we need to fix it. Originally reported by
5491 * Bjorn Helgaas on a 128-cpu setup.
5493 BUG_ON(busiest_rq == target_rq);
5495 /* move a task from busiest_rq to target_rq */
5496 double_lock_balance(busiest_rq, target_rq);
5498 /* Search for an sd spanning us and the target CPU. */
5500 for_each_domain(target_cpu, sd) {
5501 if ((sd->flags & SD_LOAD_BALANCE) &&
5502 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5507 struct lb_env env = {
5509 .dst_cpu = target_cpu,
5510 .dst_rq = target_rq,
5511 .src_cpu = busiest_rq->cpu,
5512 .src_rq = busiest_rq,
5516 schedstat_inc(sd, alb_count);
5518 if (move_one_task(&env))
5519 schedstat_inc(sd, alb_pushed);
5521 schedstat_inc(sd, alb_failed);
5524 double_unlock_balance(busiest_rq, target_rq);
5526 busiest_rq->active_balance = 0;
5527 raw_spin_unlock_irq(&busiest_rq->lock);
5531 #ifdef CONFIG_NO_HZ_COMMON
5533 * idle load balancing details
5534 * - When one of the busy CPUs notice that there may be an idle rebalancing
5535 * needed, they will kick the idle load balancer, which then does idle
5536 * load balancing for all the idle CPUs.
5539 cpumask_var_t idle_cpus_mask;
5541 unsigned long next_balance; /* in jiffy units */
5542 } nohz ____cacheline_aligned;
5544 static inline int find_new_ilb(int call_cpu)
5546 int ilb = cpumask_first(nohz.idle_cpus_mask);
5548 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5555 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5556 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5557 * CPU (if there is one).
5559 static void nohz_balancer_kick(int cpu)
5563 nohz.next_balance++;
5565 ilb_cpu = find_new_ilb(cpu);
5567 if (ilb_cpu >= nr_cpu_ids)
5570 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5573 * Use smp_send_reschedule() instead of resched_cpu().
5574 * This way we generate a sched IPI on the target cpu which
5575 * is idle. And the softirq performing nohz idle load balance
5576 * will be run before returning from the IPI.
5578 smp_send_reschedule(ilb_cpu);
5582 static inline void nohz_balance_exit_idle(int cpu)
5584 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5585 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5586 atomic_dec(&nohz.nr_cpus);
5587 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5591 static inline void set_cpu_sd_state_busy(void)
5593 struct sched_domain *sd;
5596 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5598 if (!sd || !sd->nohz_idle)
5602 for (; sd; sd = sd->parent)
5603 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5608 void set_cpu_sd_state_idle(void)
5610 struct sched_domain *sd;
5613 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5615 if (!sd || sd->nohz_idle)
5619 for (; sd; sd = sd->parent)
5620 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5626 * This routine will record that the cpu is going idle with tick stopped.
5627 * This info will be used in performing idle load balancing in the future.
5629 void nohz_balance_enter_idle(int cpu)
5632 * If this cpu is going down, then nothing needs to be done.
5634 if (!cpu_active(cpu))
5637 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5640 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5641 atomic_inc(&nohz.nr_cpus);
5642 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5645 static int sched_ilb_notifier(struct notifier_block *nfb,
5646 unsigned long action, void *hcpu)
5648 switch (action & ~CPU_TASKS_FROZEN) {
5650 nohz_balance_exit_idle(smp_processor_id());
5658 static DEFINE_SPINLOCK(balancing);
5661 * Scale the max load_balance interval with the number of CPUs in the system.
5662 * This trades load-balance latency on larger machines for less cross talk.
5664 void update_max_interval(void)
5666 max_load_balance_interval = HZ*num_online_cpus()/10;
5670 * It checks each scheduling domain to see if it is due to be balanced,
5671 * and initiates a balancing operation if so.
5673 * Balancing parameters are set up in init_sched_domains.
5675 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5677 int continue_balancing = 1;
5678 struct rq *rq = cpu_rq(cpu);
5679 unsigned long interval;
5680 struct sched_domain *sd;
5681 /* Earliest time when we have to do rebalance again */
5682 unsigned long next_balance = jiffies + 60*HZ;
5683 int update_next_balance = 0;
5684 int need_serialize, need_decay = 0;
5687 update_blocked_averages(cpu);
5690 for_each_domain(cpu, sd) {
5692 * Decay the newidle max times here because this is a regular
5693 * visit to all the domains. Decay ~1% per second.
5695 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5696 sd->max_newidle_lb_cost =
5697 (sd->max_newidle_lb_cost * 253) / 256;
5698 sd->next_decay_max_lb_cost = jiffies + HZ;
5701 max_cost += sd->max_newidle_lb_cost;
5703 if (!(sd->flags & SD_LOAD_BALANCE))
5707 * Stop the load balance at this level. There is another
5708 * CPU in our sched group which is doing load balancing more
5711 if (!continue_balancing) {
5717 interval = sd->balance_interval;
5718 if (idle != CPU_IDLE)
5719 interval *= sd->busy_factor;
5721 /* scale ms to jiffies */
5722 interval = msecs_to_jiffies(interval);
5723 interval = clamp(interval, 1UL, max_load_balance_interval);
5725 need_serialize = sd->flags & SD_SERIALIZE;
5727 if (need_serialize) {
5728 if (!spin_trylock(&balancing))
5732 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5733 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5735 * The LBF_DST_PINNED logic could have changed
5736 * env->dst_cpu, so we can't know our idle
5737 * state even if we migrated tasks. Update it.
5739 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5741 sd->last_balance = jiffies;
5744 spin_unlock(&balancing);
5746 if (time_after(next_balance, sd->last_balance + interval)) {
5747 next_balance = sd->last_balance + interval;
5748 update_next_balance = 1;
5753 * Ensure the rq-wide value also decays but keep it at a
5754 * reasonable floor to avoid funnies with rq->avg_idle.
5756 rq->max_idle_balance_cost =
5757 max((u64)sysctl_sched_migration_cost, max_cost);
5762 * next_balance will be updated only when there is a need.
5763 * When the cpu is attached to null domain for ex, it will not be
5766 if (likely(update_next_balance))
5767 rq->next_balance = next_balance;
5770 #ifdef CONFIG_NO_HZ_COMMON
5772 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5773 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5775 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5777 struct rq *this_rq = cpu_rq(this_cpu);
5781 if (idle != CPU_IDLE ||
5782 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5785 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5786 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5790 * If this cpu gets work to do, stop the load balancing
5791 * work being done for other cpus. Next load
5792 * balancing owner will pick it up.
5797 rq = cpu_rq(balance_cpu);
5799 raw_spin_lock_irq(&rq->lock);
5800 update_rq_clock(rq);
5801 update_idle_cpu_load(rq);
5802 raw_spin_unlock_irq(&rq->lock);
5804 rebalance_domains(balance_cpu, CPU_IDLE);
5806 if (time_after(this_rq->next_balance, rq->next_balance))
5807 this_rq->next_balance = rq->next_balance;
5809 nohz.next_balance = this_rq->next_balance;
5811 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5815 * Current heuristic for kicking the idle load balancer in the presence
5816 * of an idle cpu is the system.
5817 * - This rq has more than one task.
5818 * - At any scheduler domain level, this cpu's scheduler group has multiple
5819 * busy cpu's exceeding the group's power.
5820 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5821 * domain span are idle.
5823 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5825 unsigned long now = jiffies;
5826 struct sched_domain *sd;
5828 if (unlikely(idle_cpu(cpu)))
5832 * We may be recently in ticked or tickless idle mode. At the first
5833 * busy tick after returning from idle, we will update the busy stats.
5835 set_cpu_sd_state_busy();
5836 nohz_balance_exit_idle(cpu);
5839 * None are in tickless mode and hence no need for NOHZ idle load
5842 if (likely(!atomic_read(&nohz.nr_cpus)))
5845 if (time_before(now, nohz.next_balance))
5848 if (rq->nr_running >= 2)
5852 for_each_domain(cpu, sd) {
5853 struct sched_group *sg = sd->groups;
5854 struct sched_group_power *sgp = sg->sgp;
5855 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5857 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5858 goto need_kick_unlock;
5860 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5861 && (cpumask_first_and(nohz.idle_cpus_mask,
5862 sched_domain_span(sd)) < cpu))
5863 goto need_kick_unlock;
5865 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5877 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5881 * run_rebalance_domains is triggered when needed from the scheduler tick.
5882 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5884 static void run_rebalance_domains(struct softirq_action *h)
5886 int this_cpu = smp_processor_id();
5887 struct rq *this_rq = cpu_rq(this_cpu);
5888 enum cpu_idle_type idle = this_rq->idle_balance ?
5889 CPU_IDLE : CPU_NOT_IDLE;
5891 rebalance_domains(this_cpu, idle);
5894 * If this cpu has a pending nohz_balance_kick, then do the
5895 * balancing on behalf of the other idle cpus whose ticks are
5898 nohz_idle_balance(this_cpu, idle);
5901 static inline int on_null_domain(int cpu)
5903 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5907 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5909 void trigger_load_balance(struct rq *rq, int cpu)
5911 /* Don't need to rebalance while attached to NULL domain */
5912 if (time_after_eq(jiffies, rq->next_balance) &&
5913 likely(!on_null_domain(cpu)))
5914 raise_softirq(SCHED_SOFTIRQ);
5915 #ifdef CONFIG_NO_HZ_COMMON
5916 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5917 nohz_balancer_kick(cpu);
5921 static void rq_online_fair(struct rq *rq)
5926 static void rq_offline_fair(struct rq *rq)
5930 /* Ensure any throttled groups are reachable by pick_next_task */
5931 unthrottle_offline_cfs_rqs(rq);
5934 #endif /* CONFIG_SMP */
5937 * scheduler tick hitting a task of our scheduling class:
5939 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5941 struct cfs_rq *cfs_rq;
5942 struct sched_entity *se = &curr->se;
5944 for_each_sched_entity(se) {
5945 cfs_rq = cfs_rq_of(se);
5946 entity_tick(cfs_rq, se, queued);
5949 if (numabalancing_enabled)
5950 task_tick_numa(rq, curr);
5952 update_rq_runnable_avg(rq, 1);
5956 * called on fork with the child task as argument from the parent's context
5957 * - child not yet on the tasklist
5958 * - preemption disabled
5960 static void task_fork_fair(struct task_struct *p)
5962 struct cfs_rq *cfs_rq;
5963 struct sched_entity *se = &p->se, *curr;
5964 int this_cpu = smp_processor_id();
5965 struct rq *rq = this_rq();
5966 unsigned long flags;
5968 raw_spin_lock_irqsave(&rq->lock, flags);
5970 update_rq_clock(rq);
5972 cfs_rq = task_cfs_rq(current);
5973 curr = cfs_rq->curr;
5976 * Not only the cpu but also the task_group of the parent might have
5977 * been changed after parent->se.parent,cfs_rq were copied to
5978 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
5979 * of child point to valid ones.
5982 __set_task_cpu(p, this_cpu);
5985 update_curr(cfs_rq);
5988 se->vruntime = curr->vruntime;
5989 place_entity(cfs_rq, se, 1);
5991 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5993 * Upon rescheduling, sched_class::put_prev_task() will place
5994 * 'current' within the tree based on its new key value.
5996 swap(curr->vruntime, se->vruntime);
5997 resched_task(rq->curr);
6000 se->vruntime -= cfs_rq->min_vruntime;
6002 raw_spin_unlock_irqrestore(&rq->lock, flags);
6006 * Priority of the task has changed. Check to see if we preempt
6010 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6016 * Reschedule if we are currently running on this runqueue and
6017 * our priority decreased, or if we are not currently running on
6018 * this runqueue and our priority is higher than the current's
6020 if (rq->curr == p) {
6021 if (p->prio > oldprio)
6022 resched_task(rq->curr);
6024 check_preempt_curr(rq, p, 0);
6027 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6029 struct sched_entity *se = &p->se;
6030 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6033 * Ensure the task's vruntime is normalized, so that when its
6034 * switched back to the fair class the enqueue_entity(.flags=0) will
6035 * do the right thing.
6037 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6038 * have normalized the vruntime, if it was !on_rq, then only when
6039 * the task is sleeping will it still have non-normalized vruntime.
6041 if (!se->on_rq && p->state != TASK_RUNNING) {
6043 * Fix up our vruntime so that the current sleep doesn't
6044 * cause 'unlimited' sleep bonus.
6046 place_entity(cfs_rq, se, 0);
6047 se->vruntime -= cfs_rq->min_vruntime;
6052 * Remove our load from contribution when we leave sched_fair
6053 * and ensure we don't carry in an old decay_count if we
6056 if (se->avg.decay_count) {
6057 __synchronize_entity_decay(se);
6058 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6064 * We switched to the sched_fair class.
6066 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6072 * We were most likely switched from sched_rt, so
6073 * kick off the schedule if running, otherwise just see
6074 * if we can still preempt the current task.
6077 resched_task(rq->curr);
6079 check_preempt_curr(rq, p, 0);
6082 /* Account for a task changing its policy or group.
6084 * This routine is mostly called to set cfs_rq->curr field when a task
6085 * migrates between groups/classes.
6087 static void set_curr_task_fair(struct rq *rq)
6089 struct sched_entity *se = &rq->curr->se;
6091 for_each_sched_entity(se) {
6092 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6094 set_next_entity(cfs_rq, se);
6095 /* ensure bandwidth has been allocated on our new cfs_rq */
6096 account_cfs_rq_runtime(cfs_rq, 0);
6100 void init_cfs_rq(struct cfs_rq *cfs_rq)
6102 cfs_rq->tasks_timeline = RB_ROOT;
6103 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6104 #ifndef CONFIG_64BIT
6105 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6108 atomic64_set(&cfs_rq->decay_counter, 1);
6109 atomic_long_set(&cfs_rq->removed_load, 0);
6113 #ifdef CONFIG_FAIR_GROUP_SCHED
6114 static void task_move_group_fair(struct task_struct *p, int on_rq)
6116 struct cfs_rq *cfs_rq;
6118 * If the task was not on the rq at the time of this cgroup movement
6119 * it must have been asleep, sleeping tasks keep their ->vruntime
6120 * absolute on their old rq until wakeup (needed for the fair sleeper
6121 * bonus in place_entity()).
6123 * If it was on the rq, we've just 'preempted' it, which does convert
6124 * ->vruntime to a relative base.
6126 * Make sure both cases convert their relative position when migrating
6127 * to another cgroup's rq. This does somewhat interfere with the
6128 * fair sleeper stuff for the first placement, but who cares.
6131 * When !on_rq, vruntime of the task has usually NOT been normalized.
6132 * But there are some cases where it has already been normalized:
6134 * - Moving a forked child which is waiting for being woken up by
6135 * wake_up_new_task().
6136 * - Moving a task which has been woken up by try_to_wake_up() and
6137 * waiting for actually being woken up by sched_ttwu_pending().
6139 * To prevent boost or penalty in the new cfs_rq caused by delta
6140 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6142 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6146 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6147 set_task_rq(p, task_cpu(p));
6149 cfs_rq = cfs_rq_of(&p->se);
6150 p->se.vruntime += cfs_rq->min_vruntime;
6153 * migrate_task_rq_fair() will have removed our previous
6154 * contribution, but we must synchronize for ongoing future
6157 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6158 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6163 void free_fair_sched_group(struct task_group *tg)
6167 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6169 for_each_possible_cpu(i) {
6171 kfree(tg->cfs_rq[i]);
6180 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6182 struct cfs_rq *cfs_rq;
6183 struct sched_entity *se;
6186 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6189 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6193 tg->shares = NICE_0_LOAD;
6195 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6197 for_each_possible_cpu(i) {
6198 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6199 GFP_KERNEL, cpu_to_node(i));
6203 se = kzalloc_node(sizeof(struct sched_entity),
6204 GFP_KERNEL, cpu_to_node(i));
6208 init_cfs_rq(cfs_rq);
6209 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6220 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6222 struct rq *rq = cpu_rq(cpu);
6223 unsigned long flags;
6226 * Only empty task groups can be destroyed; so we can speculatively
6227 * check on_list without danger of it being re-added.
6229 if (!tg->cfs_rq[cpu]->on_list)
6232 raw_spin_lock_irqsave(&rq->lock, flags);
6233 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6234 raw_spin_unlock_irqrestore(&rq->lock, flags);
6237 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6238 struct sched_entity *se, int cpu,
6239 struct sched_entity *parent)
6241 struct rq *rq = cpu_rq(cpu);
6245 init_cfs_rq_runtime(cfs_rq);
6247 tg->cfs_rq[cpu] = cfs_rq;
6250 /* se could be NULL for root_task_group */
6255 se->cfs_rq = &rq->cfs;
6257 se->cfs_rq = parent->my_q;
6260 update_load_set(&se->load, 0);
6261 se->parent = parent;
6264 static DEFINE_MUTEX(shares_mutex);
6266 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6269 unsigned long flags;
6272 * We can't change the weight of the root cgroup.
6277 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6279 mutex_lock(&shares_mutex);
6280 if (tg->shares == shares)
6283 tg->shares = shares;
6284 for_each_possible_cpu(i) {
6285 struct rq *rq = cpu_rq(i);
6286 struct sched_entity *se;
6289 /* Propagate contribution to hierarchy */
6290 raw_spin_lock_irqsave(&rq->lock, flags);
6292 /* Possible calls to update_curr() need rq clock */
6293 update_rq_clock(rq);
6294 for_each_sched_entity(se)
6295 update_cfs_shares(group_cfs_rq(se));
6296 raw_spin_unlock_irqrestore(&rq->lock, flags);
6300 mutex_unlock(&shares_mutex);
6303 #else /* CONFIG_FAIR_GROUP_SCHED */
6305 void free_fair_sched_group(struct task_group *tg) { }
6307 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6312 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6314 #endif /* CONFIG_FAIR_GROUP_SCHED */
6317 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6319 struct sched_entity *se = &task->se;
6320 unsigned int rr_interval = 0;
6323 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6326 if (rq->cfs.load.weight)
6327 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6333 * All the scheduling class methods:
6335 const struct sched_class fair_sched_class = {
6336 .next = &idle_sched_class,
6337 .enqueue_task = enqueue_task_fair,
6338 .dequeue_task = dequeue_task_fair,
6339 .yield_task = yield_task_fair,
6340 .yield_to_task = yield_to_task_fair,
6342 .check_preempt_curr = check_preempt_wakeup,
6344 .pick_next_task = pick_next_task_fair,
6345 .put_prev_task = put_prev_task_fair,
6348 .select_task_rq = select_task_rq_fair,
6349 .migrate_task_rq = migrate_task_rq_fair,
6351 .rq_online = rq_online_fair,
6352 .rq_offline = rq_offline_fair,
6354 .task_waking = task_waking_fair,
6357 .set_curr_task = set_curr_task_fair,
6358 .task_tick = task_tick_fair,
6359 .task_fork = task_fork_fair,
6361 .prio_changed = prio_changed_fair,
6362 .switched_from = switched_from_fair,
6363 .switched_to = switched_to_fair,
6365 .get_rr_interval = get_rr_interval_fair,
6367 #ifdef CONFIG_FAIR_GROUP_SCHED
6368 .task_move_group = task_move_group_fair,
6372 #ifdef CONFIG_SCHED_DEBUG
6373 void print_cfs_stats(struct seq_file *m, int cpu)
6375 struct cfs_rq *cfs_rq;
6378 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6379 print_cfs_rq(m, cpu, cfs_rq);
6384 __init void init_sched_fair_class(void)
6387 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6389 #ifdef CONFIG_NO_HZ_COMMON
6390 nohz.next_balance = jiffies;
6391 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6392 cpu_notifier(sched_ilb_notifier, 0);