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)
903 if (!mm->numa_next_reset || !mm->numa_next_scan) {
904 mm->numa_next_scan = now +
905 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
906 mm->numa_next_reset = now +
907 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
911 * Reset the scan period if enough time has gone by. Objective is that
912 * scanning will be reduced if pages are properly placed. As tasks
913 * can enter different phases this needs to be re-examined. Lacking
914 * proper tracking of reference behaviour, this blunt hammer is used.
916 migrate = mm->numa_next_reset;
917 if (time_after(now, migrate)) {
918 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
919 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
920 xchg(&mm->numa_next_reset, next_scan);
924 * Enforce maximal scan/migration frequency..
926 migrate = mm->numa_next_scan;
927 if (time_before(now, migrate))
930 if (p->numa_scan_period == 0)
931 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
933 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
934 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
938 * Delay this task enough that another task of this mm will likely win
939 * the next time around.
941 p->node_stamp += 2 * TICK_NSEC;
943 start = mm->numa_scan_offset;
944 pages = sysctl_numa_balancing_scan_size;
945 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
949 down_read(&mm->mmap_sem);
950 vma = find_vma(mm, start);
952 reset_ptenuma_scan(p);
956 for (; vma; vma = vma->vm_next) {
957 if (!vma_migratable(vma))
960 /* Skip small VMAs. They are not likely to be of relevance */
961 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
965 start = max(start, vma->vm_start);
966 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
967 end = min(end, vma->vm_end);
968 pages -= change_prot_numa(vma, start, end);
973 } while (end != vma->vm_end);
978 * It is possible to reach the end of the VMA list but the last few
979 * VMAs are not guaranteed to the vma_migratable. If they are not, we
980 * would find the !migratable VMA on the next scan but not reset the
981 * scanner to the start so check it now.
984 mm->numa_scan_offset = start;
986 reset_ptenuma_scan(p);
987 up_read(&mm->mmap_sem);
991 * Drive the periodic memory faults..
993 void task_tick_numa(struct rq *rq, struct task_struct *curr)
995 struct callback_head *work = &curr->numa_work;
999 * We don't care about NUMA placement if we don't have memory.
1001 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1005 * Using runtime rather than walltime has the dual advantage that
1006 * we (mostly) drive the selection from busy threads and that the
1007 * task needs to have done some actual work before we bother with
1010 now = curr->se.sum_exec_runtime;
1011 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1013 if (now - curr->node_stamp > period) {
1014 if (!curr->node_stamp)
1015 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
1016 curr->node_stamp += period;
1018 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1019 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1020 task_work_add(curr, work, true);
1025 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1028 #endif /* CONFIG_NUMA_BALANCING */
1031 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1033 update_load_add(&cfs_rq->load, se->load.weight);
1034 if (!parent_entity(se))
1035 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1037 if (entity_is_task(se))
1038 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1040 cfs_rq->nr_running++;
1044 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1046 update_load_sub(&cfs_rq->load, se->load.weight);
1047 if (!parent_entity(se))
1048 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1049 if (entity_is_task(se))
1050 list_del_init(&se->group_node);
1051 cfs_rq->nr_running--;
1054 #ifdef CONFIG_FAIR_GROUP_SCHED
1056 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1061 * Use this CPU's actual weight instead of the last load_contribution
1062 * to gain a more accurate current total weight. See
1063 * update_cfs_rq_load_contribution().
1065 tg_weight = atomic_long_read(&tg->load_avg);
1066 tg_weight -= cfs_rq->tg_load_contrib;
1067 tg_weight += cfs_rq->load.weight;
1072 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1074 long tg_weight, load, shares;
1076 tg_weight = calc_tg_weight(tg, cfs_rq);
1077 load = cfs_rq->load.weight;
1079 shares = (tg->shares * load);
1081 shares /= tg_weight;
1083 if (shares < MIN_SHARES)
1084 shares = MIN_SHARES;
1085 if (shares > tg->shares)
1086 shares = tg->shares;
1090 # else /* CONFIG_SMP */
1091 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1095 # endif /* CONFIG_SMP */
1096 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1097 unsigned long weight)
1100 /* commit outstanding execution time */
1101 if (cfs_rq->curr == se)
1102 update_curr(cfs_rq);
1103 account_entity_dequeue(cfs_rq, se);
1106 update_load_set(&se->load, weight);
1109 account_entity_enqueue(cfs_rq, se);
1112 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1114 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1116 struct task_group *tg;
1117 struct sched_entity *se;
1121 se = tg->se[cpu_of(rq_of(cfs_rq))];
1122 if (!se || throttled_hierarchy(cfs_rq))
1125 if (likely(se->load.weight == tg->shares))
1128 shares = calc_cfs_shares(cfs_rq, tg);
1130 reweight_entity(cfs_rq_of(se), se, shares);
1132 #else /* CONFIG_FAIR_GROUP_SCHED */
1133 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1136 #endif /* CONFIG_FAIR_GROUP_SCHED */
1140 * We choose a half-life close to 1 scheduling period.
1141 * Note: The tables below are dependent on this value.
1143 #define LOAD_AVG_PERIOD 32
1144 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1145 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1147 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1148 static const u32 runnable_avg_yN_inv[] = {
1149 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1150 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1151 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1152 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1153 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1154 0x85aac367, 0x82cd8698,
1158 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1159 * over-estimates when re-combining.
1161 static const u32 runnable_avg_yN_sum[] = {
1162 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1163 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1164 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1169 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1171 static __always_inline u64 decay_load(u64 val, u64 n)
1173 unsigned int local_n;
1177 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1180 /* after bounds checking we can collapse to 32-bit */
1184 * As y^PERIOD = 1/2, we can combine
1185 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1186 * With a look-up table which covers k^n (n<PERIOD)
1188 * To achieve constant time decay_load.
1190 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1191 val >>= local_n / LOAD_AVG_PERIOD;
1192 local_n %= LOAD_AVG_PERIOD;
1195 val *= runnable_avg_yN_inv[local_n];
1196 /* We don't use SRR here since we always want to round down. */
1201 * For updates fully spanning n periods, the contribution to runnable
1202 * average will be: \Sum 1024*y^n
1204 * We can compute this reasonably efficiently by combining:
1205 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1207 static u32 __compute_runnable_contrib(u64 n)
1211 if (likely(n <= LOAD_AVG_PERIOD))
1212 return runnable_avg_yN_sum[n];
1213 else if (unlikely(n >= LOAD_AVG_MAX_N))
1214 return LOAD_AVG_MAX;
1216 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1218 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1219 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1221 n -= LOAD_AVG_PERIOD;
1222 } while (n > LOAD_AVG_PERIOD);
1224 contrib = decay_load(contrib, n);
1225 return contrib + runnable_avg_yN_sum[n];
1229 * We can represent the historical contribution to runnable average as the
1230 * coefficients of a geometric series. To do this we sub-divide our runnable
1231 * history into segments of approximately 1ms (1024us); label the segment that
1232 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1234 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1236 * (now) (~1ms ago) (~2ms ago)
1238 * Let u_i denote the fraction of p_i that the entity was runnable.
1240 * We then designate the fractions u_i as our co-efficients, yielding the
1241 * following representation of historical load:
1242 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1244 * We choose y based on the with of a reasonably scheduling period, fixing:
1247 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1248 * approximately half as much as the contribution to load within the last ms
1251 * When a period "rolls over" and we have new u_0`, multiplying the previous
1252 * sum again by y is sufficient to update:
1253 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1254 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1256 static __always_inline int __update_entity_runnable_avg(u64 now,
1257 struct sched_avg *sa,
1261 u32 runnable_contrib;
1262 int delta_w, decayed = 0;
1264 delta = now - sa->last_runnable_update;
1266 * This should only happen when time goes backwards, which it
1267 * unfortunately does during sched clock init when we swap over to TSC.
1269 if ((s64)delta < 0) {
1270 sa->last_runnable_update = now;
1275 * Use 1024ns as the unit of measurement since it's a reasonable
1276 * approximation of 1us and fast to compute.
1281 sa->last_runnable_update = now;
1283 /* delta_w is the amount already accumulated against our next period */
1284 delta_w = sa->runnable_avg_period % 1024;
1285 if (delta + delta_w >= 1024) {
1286 /* period roll-over */
1290 * Now that we know we're crossing a period boundary, figure
1291 * out how much from delta we need to complete the current
1292 * period and accrue it.
1294 delta_w = 1024 - delta_w;
1296 sa->runnable_avg_sum += delta_w;
1297 sa->runnable_avg_period += delta_w;
1301 /* Figure out how many additional periods this update spans */
1302 periods = delta / 1024;
1305 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1307 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1310 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1311 runnable_contrib = __compute_runnable_contrib(periods);
1313 sa->runnable_avg_sum += runnable_contrib;
1314 sa->runnable_avg_period += runnable_contrib;
1317 /* Remainder of delta accrued against u_0` */
1319 sa->runnable_avg_sum += delta;
1320 sa->runnable_avg_period += delta;
1325 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1326 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1328 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1329 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1331 decays -= se->avg.decay_count;
1335 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1336 se->avg.decay_count = 0;
1341 #ifdef CONFIG_FAIR_GROUP_SCHED
1342 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1345 struct task_group *tg = cfs_rq->tg;
1348 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1349 tg_contrib -= cfs_rq->tg_load_contrib;
1351 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1352 atomic_long_add(tg_contrib, &tg->load_avg);
1353 cfs_rq->tg_load_contrib += tg_contrib;
1358 * Aggregate cfs_rq runnable averages into an equivalent task_group
1359 * representation for computing load contributions.
1361 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1362 struct cfs_rq *cfs_rq)
1364 struct task_group *tg = cfs_rq->tg;
1367 /* The fraction of a cpu used by this cfs_rq */
1368 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1369 sa->runnable_avg_period + 1);
1370 contrib -= cfs_rq->tg_runnable_contrib;
1372 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1373 atomic_add(contrib, &tg->runnable_avg);
1374 cfs_rq->tg_runnable_contrib += contrib;
1378 static inline void __update_group_entity_contrib(struct sched_entity *se)
1380 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1381 struct task_group *tg = cfs_rq->tg;
1386 contrib = cfs_rq->tg_load_contrib * tg->shares;
1387 se->avg.load_avg_contrib = div_u64(contrib,
1388 atomic_long_read(&tg->load_avg) + 1);
1391 * For group entities we need to compute a correction term in the case
1392 * that they are consuming <1 cpu so that we would contribute the same
1393 * load as a task of equal weight.
1395 * Explicitly co-ordinating this measurement would be expensive, but
1396 * fortunately the sum of each cpus contribution forms a usable
1397 * lower-bound on the true value.
1399 * Consider the aggregate of 2 contributions. Either they are disjoint
1400 * (and the sum represents true value) or they are disjoint and we are
1401 * understating by the aggregate of their overlap.
1403 * Extending this to N cpus, for a given overlap, the maximum amount we
1404 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1405 * cpus that overlap for this interval and w_i is the interval width.
1407 * On a small machine; the first term is well-bounded which bounds the
1408 * total error since w_i is a subset of the period. Whereas on a
1409 * larger machine, while this first term can be larger, if w_i is the
1410 * of consequential size guaranteed to see n_i*w_i quickly converge to
1411 * our upper bound of 1-cpu.
1413 runnable_avg = atomic_read(&tg->runnable_avg);
1414 if (runnable_avg < NICE_0_LOAD) {
1415 se->avg.load_avg_contrib *= runnable_avg;
1416 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1420 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1421 int force_update) {}
1422 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1423 struct cfs_rq *cfs_rq) {}
1424 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1427 static inline void __update_task_entity_contrib(struct sched_entity *se)
1431 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1432 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1433 contrib /= (se->avg.runnable_avg_period + 1);
1434 se->avg.load_avg_contrib = scale_load(contrib);
1437 /* Compute the current contribution to load_avg by se, return any delta */
1438 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1440 long old_contrib = se->avg.load_avg_contrib;
1442 if (entity_is_task(se)) {
1443 __update_task_entity_contrib(se);
1445 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1446 __update_group_entity_contrib(se);
1449 return se->avg.load_avg_contrib - old_contrib;
1452 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1455 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1456 cfs_rq->blocked_load_avg -= load_contrib;
1458 cfs_rq->blocked_load_avg = 0;
1461 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1463 /* Update a sched_entity's runnable average */
1464 static inline void update_entity_load_avg(struct sched_entity *se,
1467 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1472 * For a group entity we need to use their owned cfs_rq_clock_task() in
1473 * case they are the parent of a throttled hierarchy.
1475 if (entity_is_task(se))
1476 now = cfs_rq_clock_task(cfs_rq);
1478 now = cfs_rq_clock_task(group_cfs_rq(se));
1480 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1483 contrib_delta = __update_entity_load_avg_contrib(se);
1489 cfs_rq->runnable_load_avg += contrib_delta;
1491 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1495 * Decay the load contributed by all blocked children and account this so that
1496 * their contribution may appropriately discounted when they wake up.
1498 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1500 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1503 decays = now - cfs_rq->last_decay;
1504 if (!decays && !force_update)
1507 if (atomic_long_read(&cfs_rq->removed_load)) {
1508 unsigned long removed_load;
1509 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1510 subtract_blocked_load_contrib(cfs_rq, removed_load);
1514 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1516 atomic64_add(decays, &cfs_rq->decay_counter);
1517 cfs_rq->last_decay = now;
1520 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1523 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1525 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1526 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1529 /* Add the load generated by se into cfs_rq's child load-average */
1530 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1531 struct sched_entity *se,
1535 * We track migrations using entity decay_count <= 0, on a wake-up
1536 * migration we use a negative decay count to track the remote decays
1537 * accumulated while sleeping.
1539 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1540 * are seen by enqueue_entity_load_avg() as a migration with an already
1541 * constructed load_avg_contrib.
1543 if (unlikely(se->avg.decay_count <= 0)) {
1544 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1545 if (se->avg.decay_count) {
1547 * In a wake-up migration we have to approximate the
1548 * time sleeping. This is because we can't synchronize
1549 * clock_task between the two cpus, and it is not
1550 * guaranteed to be read-safe. Instead, we can
1551 * approximate this using our carried decays, which are
1552 * explicitly atomically readable.
1554 se->avg.last_runnable_update -= (-se->avg.decay_count)
1556 update_entity_load_avg(se, 0);
1557 /* Indicate that we're now synchronized and on-rq */
1558 se->avg.decay_count = 0;
1563 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1564 * would have made count negative); we must be careful to avoid
1565 * double-accounting blocked time after synchronizing decays.
1567 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1571 /* migrated tasks did not contribute to our blocked load */
1573 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1574 update_entity_load_avg(se, 0);
1577 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1578 /* we force update consideration on load-balancer moves */
1579 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1583 * Remove se's load from this cfs_rq child load-average, if the entity is
1584 * transitioning to a blocked state we track its projected decay using
1587 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1588 struct sched_entity *se,
1591 update_entity_load_avg(se, 1);
1592 /* we force update consideration on load-balancer moves */
1593 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1595 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1597 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1598 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1599 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1603 * Update the rq's load with the elapsed running time before entering
1604 * idle. if the last scheduled task is not a CFS task, idle_enter will
1605 * be the only way to update the runnable statistic.
1607 void idle_enter_fair(struct rq *this_rq)
1609 update_rq_runnable_avg(this_rq, 1);
1613 * Update the rq's load with the elapsed idle time before a task is
1614 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1615 * be the only way to update the runnable statistic.
1617 void idle_exit_fair(struct rq *this_rq)
1619 update_rq_runnable_avg(this_rq, 0);
1623 static inline void update_entity_load_avg(struct sched_entity *se,
1624 int update_cfs_rq) {}
1625 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1626 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1627 struct sched_entity *se,
1629 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1630 struct sched_entity *se,
1632 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1633 int force_update) {}
1636 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1638 #ifdef CONFIG_SCHEDSTATS
1639 struct task_struct *tsk = NULL;
1641 if (entity_is_task(se))
1644 if (se->statistics.sleep_start) {
1645 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1650 if (unlikely(delta > se->statistics.sleep_max))
1651 se->statistics.sleep_max = delta;
1653 se->statistics.sleep_start = 0;
1654 se->statistics.sum_sleep_runtime += delta;
1657 account_scheduler_latency(tsk, delta >> 10, 1);
1658 trace_sched_stat_sleep(tsk, delta);
1661 if (se->statistics.block_start) {
1662 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1667 if (unlikely(delta > se->statistics.block_max))
1668 se->statistics.block_max = delta;
1670 se->statistics.block_start = 0;
1671 se->statistics.sum_sleep_runtime += delta;
1674 if (tsk->in_iowait) {
1675 se->statistics.iowait_sum += delta;
1676 se->statistics.iowait_count++;
1677 trace_sched_stat_iowait(tsk, delta);
1680 trace_sched_stat_blocked(tsk, delta);
1683 * Blocking time is in units of nanosecs, so shift by
1684 * 20 to get a milliseconds-range estimation of the
1685 * amount of time that the task spent sleeping:
1687 if (unlikely(prof_on == SLEEP_PROFILING)) {
1688 profile_hits(SLEEP_PROFILING,
1689 (void *)get_wchan(tsk),
1692 account_scheduler_latency(tsk, delta >> 10, 0);
1698 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1700 #ifdef CONFIG_SCHED_DEBUG
1701 s64 d = se->vruntime - cfs_rq->min_vruntime;
1706 if (d > 3*sysctl_sched_latency)
1707 schedstat_inc(cfs_rq, nr_spread_over);
1712 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1714 u64 vruntime = cfs_rq->min_vruntime;
1717 * The 'current' period is already promised to the current tasks,
1718 * however the extra weight of the new task will slow them down a
1719 * little, place the new task so that it fits in the slot that
1720 * stays open at the end.
1722 if (initial && sched_feat(START_DEBIT))
1723 vruntime += sched_vslice(cfs_rq, se);
1725 /* sleeps up to a single latency don't count. */
1727 unsigned long thresh = sysctl_sched_latency;
1730 * Halve their sleep time's effect, to allow
1731 * for a gentler effect of sleepers:
1733 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1739 /* ensure we never gain time by being placed backwards. */
1740 se->vruntime = max_vruntime(se->vruntime, vruntime);
1743 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1746 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1749 * Update the normalized vruntime before updating min_vruntime
1750 * through calling update_curr().
1752 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1753 se->vruntime += cfs_rq->min_vruntime;
1756 * Update run-time statistics of the 'current'.
1758 update_curr(cfs_rq);
1759 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1760 account_entity_enqueue(cfs_rq, se);
1761 update_cfs_shares(cfs_rq);
1763 if (flags & ENQUEUE_WAKEUP) {
1764 place_entity(cfs_rq, se, 0);
1765 enqueue_sleeper(cfs_rq, se);
1768 update_stats_enqueue(cfs_rq, se);
1769 check_spread(cfs_rq, se);
1770 if (se != cfs_rq->curr)
1771 __enqueue_entity(cfs_rq, se);
1774 if (cfs_rq->nr_running == 1) {
1775 list_add_leaf_cfs_rq(cfs_rq);
1776 check_enqueue_throttle(cfs_rq);
1780 static void __clear_buddies_last(struct sched_entity *se)
1782 for_each_sched_entity(se) {
1783 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1784 if (cfs_rq->last == se)
1785 cfs_rq->last = NULL;
1791 static void __clear_buddies_next(struct sched_entity *se)
1793 for_each_sched_entity(se) {
1794 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1795 if (cfs_rq->next == se)
1796 cfs_rq->next = NULL;
1802 static void __clear_buddies_skip(struct sched_entity *se)
1804 for_each_sched_entity(se) {
1805 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1806 if (cfs_rq->skip == se)
1807 cfs_rq->skip = NULL;
1813 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1815 if (cfs_rq->last == se)
1816 __clear_buddies_last(se);
1818 if (cfs_rq->next == se)
1819 __clear_buddies_next(se);
1821 if (cfs_rq->skip == se)
1822 __clear_buddies_skip(se);
1825 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1828 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1831 * Update run-time statistics of the 'current'.
1833 update_curr(cfs_rq);
1834 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1836 update_stats_dequeue(cfs_rq, se);
1837 if (flags & DEQUEUE_SLEEP) {
1838 #ifdef CONFIG_SCHEDSTATS
1839 if (entity_is_task(se)) {
1840 struct task_struct *tsk = task_of(se);
1842 if (tsk->state & TASK_INTERRUPTIBLE)
1843 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
1844 if (tsk->state & TASK_UNINTERRUPTIBLE)
1845 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
1850 clear_buddies(cfs_rq, se);
1852 if (se != cfs_rq->curr)
1853 __dequeue_entity(cfs_rq, se);
1855 account_entity_dequeue(cfs_rq, se);
1858 * Normalize the entity after updating the min_vruntime because the
1859 * update can refer to the ->curr item and we need to reflect this
1860 * movement in our normalized position.
1862 if (!(flags & DEQUEUE_SLEEP))
1863 se->vruntime -= cfs_rq->min_vruntime;
1865 /* return excess runtime on last dequeue */
1866 return_cfs_rq_runtime(cfs_rq);
1868 update_min_vruntime(cfs_rq);
1869 update_cfs_shares(cfs_rq);
1873 * Preempt the current task with a newly woken task if needed:
1876 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1878 unsigned long ideal_runtime, delta_exec;
1879 struct sched_entity *se;
1882 ideal_runtime = sched_slice(cfs_rq, curr);
1883 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1884 if (delta_exec > ideal_runtime) {
1885 resched_task(rq_of(cfs_rq)->curr);
1887 * The current task ran long enough, ensure it doesn't get
1888 * re-elected due to buddy favours.
1890 clear_buddies(cfs_rq, curr);
1895 * Ensure that a task that missed wakeup preemption by a
1896 * narrow margin doesn't have to wait for a full slice.
1897 * This also mitigates buddy induced latencies under load.
1899 if (delta_exec < sysctl_sched_min_granularity)
1902 se = __pick_first_entity(cfs_rq);
1903 delta = curr->vruntime - se->vruntime;
1908 if (delta > ideal_runtime)
1909 resched_task(rq_of(cfs_rq)->curr);
1913 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1915 /* 'current' is not kept within the tree. */
1918 * Any task has to be enqueued before it get to execute on
1919 * a CPU. So account for the time it spent waiting on the
1922 update_stats_wait_end(cfs_rq, se);
1923 __dequeue_entity(cfs_rq, se);
1926 update_stats_curr_start(cfs_rq, se);
1928 #ifdef CONFIG_SCHEDSTATS
1930 * Track our maximum slice length, if the CPU's load is at
1931 * least twice that of our own weight (i.e. dont track it
1932 * when there are only lesser-weight tasks around):
1934 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1935 se->statistics.slice_max = max(se->statistics.slice_max,
1936 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1939 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1943 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1946 * Pick the next process, keeping these things in mind, in this order:
1947 * 1) keep things fair between processes/task groups
1948 * 2) pick the "next" process, since someone really wants that to run
1949 * 3) pick the "last" process, for cache locality
1950 * 4) do not run the "skip" process, if something else is available
1952 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1954 struct sched_entity *se = __pick_first_entity(cfs_rq);
1955 struct sched_entity *left = se;
1958 * Avoid running the skip buddy, if running something else can
1959 * be done without getting too unfair.
1961 if (cfs_rq->skip == se) {
1962 struct sched_entity *second = __pick_next_entity(se);
1963 if (second && wakeup_preempt_entity(second, left) < 1)
1968 * Prefer last buddy, try to return the CPU to a preempted task.
1970 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1974 * Someone really wants this to run. If it's not unfair, run it.
1976 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1979 clear_buddies(cfs_rq, se);
1984 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1986 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1989 * If still on the runqueue then deactivate_task()
1990 * was not called and update_curr() has to be done:
1993 update_curr(cfs_rq);
1995 /* throttle cfs_rqs exceeding runtime */
1996 check_cfs_rq_runtime(cfs_rq);
1998 check_spread(cfs_rq, prev);
2000 update_stats_wait_start(cfs_rq, prev);
2001 /* Put 'current' back into the tree. */
2002 __enqueue_entity(cfs_rq, prev);
2003 /* in !on_rq case, update occurred at dequeue */
2004 update_entity_load_avg(prev, 1);
2006 cfs_rq->curr = NULL;
2010 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2013 * Update run-time statistics of the 'current'.
2015 update_curr(cfs_rq);
2018 * Ensure that runnable average is periodically updated.
2020 update_entity_load_avg(curr, 1);
2021 update_cfs_rq_blocked_load(cfs_rq, 1);
2022 update_cfs_shares(cfs_rq);
2024 #ifdef CONFIG_SCHED_HRTICK
2026 * queued ticks are scheduled to match the slice, so don't bother
2027 * validating it and just reschedule.
2030 resched_task(rq_of(cfs_rq)->curr);
2034 * don't let the period tick interfere with the hrtick preemption
2036 if (!sched_feat(DOUBLE_TICK) &&
2037 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2041 if (cfs_rq->nr_running > 1)
2042 check_preempt_tick(cfs_rq, curr);
2046 /**************************************************
2047 * CFS bandwidth control machinery
2050 #ifdef CONFIG_CFS_BANDWIDTH
2052 #ifdef HAVE_JUMP_LABEL
2053 static struct static_key __cfs_bandwidth_used;
2055 static inline bool cfs_bandwidth_used(void)
2057 return static_key_false(&__cfs_bandwidth_used);
2060 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2062 /* only need to count groups transitioning between enabled/!enabled */
2063 if (enabled && !was_enabled)
2064 static_key_slow_inc(&__cfs_bandwidth_used);
2065 else if (!enabled && was_enabled)
2066 static_key_slow_dec(&__cfs_bandwidth_used);
2068 #else /* HAVE_JUMP_LABEL */
2069 static bool cfs_bandwidth_used(void)
2074 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2075 #endif /* HAVE_JUMP_LABEL */
2078 * default period for cfs group bandwidth.
2079 * default: 0.1s, units: nanoseconds
2081 static inline u64 default_cfs_period(void)
2083 return 100000000ULL;
2086 static inline u64 sched_cfs_bandwidth_slice(void)
2088 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2092 * Replenish runtime according to assigned quota and update expiration time.
2093 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2094 * additional synchronization around rq->lock.
2096 * requires cfs_b->lock
2098 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2102 if (cfs_b->quota == RUNTIME_INF)
2105 now = sched_clock_cpu(smp_processor_id());
2106 cfs_b->runtime = cfs_b->quota;
2107 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2110 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2112 return &tg->cfs_bandwidth;
2115 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2116 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2118 if (unlikely(cfs_rq->throttle_count))
2119 return cfs_rq->throttled_clock_task;
2121 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2124 /* returns 0 on failure to allocate runtime */
2125 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2127 struct task_group *tg = cfs_rq->tg;
2128 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2129 u64 amount = 0, min_amount, expires;
2131 /* note: this is a positive sum as runtime_remaining <= 0 */
2132 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2134 raw_spin_lock(&cfs_b->lock);
2135 if (cfs_b->quota == RUNTIME_INF)
2136 amount = min_amount;
2139 * If the bandwidth pool has become inactive, then at least one
2140 * period must have elapsed since the last consumption.
2141 * Refresh the global state and ensure bandwidth timer becomes
2144 if (!cfs_b->timer_active) {
2145 __refill_cfs_bandwidth_runtime(cfs_b);
2146 __start_cfs_bandwidth(cfs_b);
2149 if (cfs_b->runtime > 0) {
2150 amount = min(cfs_b->runtime, min_amount);
2151 cfs_b->runtime -= amount;
2155 expires = cfs_b->runtime_expires;
2156 raw_spin_unlock(&cfs_b->lock);
2158 cfs_rq->runtime_remaining += amount;
2160 * we may have advanced our local expiration to account for allowed
2161 * spread between our sched_clock and the one on which runtime was
2164 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2165 cfs_rq->runtime_expires = expires;
2167 return cfs_rq->runtime_remaining > 0;
2171 * Note: This depends on the synchronization provided by sched_clock and the
2172 * fact that rq->clock snapshots this value.
2174 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2176 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2178 /* if the deadline is ahead of our clock, nothing to do */
2179 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2182 if (cfs_rq->runtime_remaining < 0)
2186 * If the local deadline has passed we have to consider the
2187 * possibility that our sched_clock is 'fast' and the global deadline
2188 * has not truly expired.
2190 * Fortunately we can check determine whether this the case by checking
2191 * whether the global deadline has advanced.
2194 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2195 /* extend local deadline, drift is bounded above by 2 ticks */
2196 cfs_rq->runtime_expires += TICK_NSEC;
2198 /* global deadline is ahead, expiration has passed */
2199 cfs_rq->runtime_remaining = 0;
2203 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2204 unsigned long delta_exec)
2206 /* dock delta_exec before expiring quota (as it could span periods) */
2207 cfs_rq->runtime_remaining -= delta_exec;
2208 expire_cfs_rq_runtime(cfs_rq);
2210 if (likely(cfs_rq->runtime_remaining > 0))
2214 * if we're unable to extend our runtime we resched so that the active
2215 * hierarchy can be throttled
2217 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2218 resched_task(rq_of(cfs_rq)->curr);
2221 static __always_inline
2222 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2224 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2227 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2230 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2232 return cfs_bandwidth_used() && cfs_rq->throttled;
2235 /* check whether cfs_rq, or any parent, is throttled */
2236 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2238 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2242 * Ensure that neither of the group entities corresponding to src_cpu or
2243 * dest_cpu are members of a throttled hierarchy when performing group
2244 * load-balance operations.
2246 static inline int throttled_lb_pair(struct task_group *tg,
2247 int src_cpu, int dest_cpu)
2249 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2251 src_cfs_rq = tg->cfs_rq[src_cpu];
2252 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2254 return throttled_hierarchy(src_cfs_rq) ||
2255 throttled_hierarchy(dest_cfs_rq);
2258 /* updated child weight may affect parent so we have to do this bottom up */
2259 static int tg_unthrottle_up(struct task_group *tg, void *data)
2261 struct rq *rq = data;
2262 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2264 cfs_rq->throttle_count--;
2266 if (!cfs_rq->throttle_count) {
2267 /* adjust cfs_rq_clock_task() */
2268 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2269 cfs_rq->throttled_clock_task;
2276 static int tg_throttle_down(struct task_group *tg, void *data)
2278 struct rq *rq = data;
2279 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2281 /* group is entering throttled state, stop time */
2282 if (!cfs_rq->throttle_count)
2283 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2284 cfs_rq->throttle_count++;
2289 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2291 struct rq *rq = rq_of(cfs_rq);
2292 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2293 struct sched_entity *se;
2294 long task_delta, dequeue = 1;
2296 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2298 /* freeze hierarchy runnable averages while throttled */
2300 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2303 task_delta = cfs_rq->h_nr_running;
2304 for_each_sched_entity(se) {
2305 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2306 /* throttled entity or throttle-on-deactivate */
2311 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2312 qcfs_rq->h_nr_running -= task_delta;
2314 if (qcfs_rq->load.weight)
2319 rq->nr_running -= task_delta;
2321 cfs_rq->throttled = 1;
2322 cfs_rq->throttled_clock = rq_clock(rq);
2323 raw_spin_lock(&cfs_b->lock);
2324 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2325 raw_spin_unlock(&cfs_b->lock);
2328 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2330 struct rq *rq = rq_of(cfs_rq);
2331 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2332 struct sched_entity *se;
2336 se = cfs_rq->tg->se[cpu_of(rq)];
2338 cfs_rq->throttled = 0;
2340 update_rq_clock(rq);
2342 raw_spin_lock(&cfs_b->lock);
2343 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2344 list_del_rcu(&cfs_rq->throttled_list);
2345 raw_spin_unlock(&cfs_b->lock);
2347 /* update hierarchical throttle state */
2348 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2350 if (!cfs_rq->load.weight)
2353 task_delta = cfs_rq->h_nr_running;
2354 for_each_sched_entity(se) {
2358 cfs_rq = cfs_rq_of(se);
2360 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2361 cfs_rq->h_nr_running += task_delta;
2363 if (cfs_rq_throttled(cfs_rq))
2368 rq->nr_running += task_delta;
2370 /* determine whether we need to wake up potentially idle cpu */
2371 if (rq->curr == rq->idle && rq->cfs.nr_running)
2372 resched_task(rq->curr);
2375 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2376 u64 remaining, u64 expires)
2378 struct cfs_rq *cfs_rq;
2379 u64 runtime = remaining;
2382 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2384 struct rq *rq = rq_of(cfs_rq);
2386 raw_spin_lock(&rq->lock);
2387 if (!cfs_rq_throttled(cfs_rq))
2390 runtime = -cfs_rq->runtime_remaining + 1;
2391 if (runtime > remaining)
2392 runtime = remaining;
2393 remaining -= runtime;
2395 cfs_rq->runtime_remaining += runtime;
2396 cfs_rq->runtime_expires = expires;
2398 /* we check whether we're throttled above */
2399 if (cfs_rq->runtime_remaining > 0)
2400 unthrottle_cfs_rq(cfs_rq);
2403 raw_spin_unlock(&rq->lock);
2414 * Responsible for refilling a task_group's bandwidth and unthrottling its
2415 * cfs_rqs as appropriate. If there has been no activity within the last
2416 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2417 * used to track this state.
2419 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2421 u64 runtime, runtime_expires;
2422 int idle = 1, throttled;
2424 raw_spin_lock(&cfs_b->lock);
2425 /* no need to continue the timer with no bandwidth constraint */
2426 if (cfs_b->quota == RUNTIME_INF)
2429 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2430 /* idle depends on !throttled (for the case of a large deficit) */
2431 idle = cfs_b->idle && !throttled;
2432 cfs_b->nr_periods += overrun;
2434 /* if we're going inactive then everything else can be deferred */
2438 __refill_cfs_bandwidth_runtime(cfs_b);
2441 /* mark as potentially idle for the upcoming period */
2446 /* account preceding periods in which throttling occurred */
2447 cfs_b->nr_throttled += overrun;
2450 * There are throttled entities so we must first use the new bandwidth
2451 * to unthrottle them before making it generally available. This
2452 * ensures that all existing debts will be paid before a new cfs_rq is
2455 runtime = cfs_b->runtime;
2456 runtime_expires = cfs_b->runtime_expires;
2460 * This check is repeated as we are holding onto the new bandwidth
2461 * while we unthrottle. This can potentially race with an unthrottled
2462 * group trying to acquire new bandwidth from the global pool.
2464 while (throttled && runtime > 0) {
2465 raw_spin_unlock(&cfs_b->lock);
2466 /* we can't nest cfs_b->lock while distributing bandwidth */
2467 runtime = distribute_cfs_runtime(cfs_b, runtime,
2469 raw_spin_lock(&cfs_b->lock);
2471 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2474 /* return (any) remaining runtime */
2475 cfs_b->runtime = runtime;
2477 * While we are ensured activity in the period following an
2478 * unthrottle, this also covers the case in which the new bandwidth is
2479 * insufficient to cover the existing bandwidth deficit. (Forcing the
2480 * timer to remain active while there are any throttled entities.)
2485 cfs_b->timer_active = 0;
2486 raw_spin_unlock(&cfs_b->lock);
2491 /* a cfs_rq won't donate quota below this amount */
2492 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2493 /* minimum remaining period time to redistribute slack quota */
2494 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2495 /* how long we wait to gather additional slack before distributing */
2496 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2498 /* are we near the end of the current quota period? */
2499 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2501 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2504 /* if the call-back is running a quota refresh is already occurring */
2505 if (hrtimer_callback_running(refresh_timer))
2508 /* is a quota refresh about to occur? */
2509 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2510 if (remaining < min_expire)
2516 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2518 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2520 /* if there's a quota refresh soon don't bother with slack */
2521 if (runtime_refresh_within(cfs_b, min_left))
2524 start_bandwidth_timer(&cfs_b->slack_timer,
2525 ns_to_ktime(cfs_bandwidth_slack_period));
2528 /* we know any runtime found here is valid as update_curr() precedes return */
2529 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2531 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2532 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2534 if (slack_runtime <= 0)
2537 raw_spin_lock(&cfs_b->lock);
2538 if (cfs_b->quota != RUNTIME_INF &&
2539 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2540 cfs_b->runtime += slack_runtime;
2542 /* we are under rq->lock, defer unthrottling using a timer */
2543 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2544 !list_empty(&cfs_b->throttled_cfs_rq))
2545 start_cfs_slack_bandwidth(cfs_b);
2547 raw_spin_unlock(&cfs_b->lock);
2549 /* even if it's not valid for return we don't want to try again */
2550 cfs_rq->runtime_remaining -= slack_runtime;
2553 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2555 if (!cfs_bandwidth_used())
2558 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2561 __return_cfs_rq_runtime(cfs_rq);
2565 * This is done with a timer (instead of inline with bandwidth return) since
2566 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2568 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2570 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2573 /* confirm we're still not at a refresh boundary */
2574 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2577 raw_spin_lock(&cfs_b->lock);
2578 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2579 runtime = cfs_b->runtime;
2582 expires = cfs_b->runtime_expires;
2583 raw_spin_unlock(&cfs_b->lock);
2588 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2590 raw_spin_lock(&cfs_b->lock);
2591 if (expires == cfs_b->runtime_expires)
2592 cfs_b->runtime = runtime;
2593 raw_spin_unlock(&cfs_b->lock);
2597 * When a group wakes up we want to make sure that its quota is not already
2598 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2599 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2601 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2603 if (!cfs_bandwidth_used())
2606 /* an active group must be handled by the update_curr()->put() path */
2607 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2610 /* ensure the group is not already throttled */
2611 if (cfs_rq_throttled(cfs_rq))
2614 /* update runtime allocation */
2615 account_cfs_rq_runtime(cfs_rq, 0);
2616 if (cfs_rq->runtime_remaining <= 0)
2617 throttle_cfs_rq(cfs_rq);
2620 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2621 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2623 if (!cfs_bandwidth_used())
2626 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2630 * it's possible for a throttled entity to be forced into a running
2631 * state (e.g. set_curr_task), in this case we're finished.
2633 if (cfs_rq_throttled(cfs_rq))
2636 throttle_cfs_rq(cfs_rq);
2639 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2641 struct cfs_bandwidth *cfs_b =
2642 container_of(timer, struct cfs_bandwidth, slack_timer);
2643 do_sched_cfs_slack_timer(cfs_b);
2645 return HRTIMER_NORESTART;
2648 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2650 struct cfs_bandwidth *cfs_b =
2651 container_of(timer, struct cfs_bandwidth, period_timer);
2657 now = hrtimer_cb_get_time(timer);
2658 overrun = hrtimer_forward(timer, now, cfs_b->period);
2663 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2666 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2669 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2671 raw_spin_lock_init(&cfs_b->lock);
2673 cfs_b->quota = RUNTIME_INF;
2674 cfs_b->period = ns_to_ktime(default_cfs_period());
2676 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2677 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2678 cfs_b->period_timer.function = sched_cfs_period_timer;
2679 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2680 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2683 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2685 cfs_rq->runtime_enabled = 0;
2686 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2689 /* requires cfs_b->lock, may release to reprogram timer */
2690 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2693 * The timer may be active because we're trying to set a new bandwidth
2694 * period or because we're racing with the tear-down path
2695 * (timer_active==0 becomes visible before the hrtimer call-back
2696 * terminates). In either case we ensure that it's re-programmed
2698 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2699 raw_spin_unlock(&cfs_b->lock);
2700 /* ensure cfs_b->lock is available while we wait */
2701 hrtimer_cancel(&cfs_b->period_timer);
2703 raw_spin_lock(&cfs_b->lock);
2704 /* if someone else restarted the timer then we're done */
2705 if (cfs_b->timer_active)
2709 cfs_b->timer_active = 1;
2710 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2713 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2715 hrtimer_cancel(&cfs_b->period_timer);
2716 hrtimer_cancel(&cfs_b->slack_timer);
2719 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2721 struct cfs_rq *cfs_rq;
2723 for_each_leaf_cfs_rq(rq, cfs_rq) {
2724 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2726 if (!cfs_rq->runtime_enabled)
2730 * clock_task is not advancing so we just need to make sure
2731 * there's some valid quota amount
2733 cfs_rq->runtime_remaining = cfs_b->quota;
2734 if (cfs_rq_throttled(cfs_rq))
2735 unthrottle_cfs_rq(cfs_rq);
2739 #else /* CONFIG_CFS_BANDWIDTH */
2740 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2742 return rq_clock_task(rq_of(cfs_rq));
2745 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2746 unsigned long delta_exec) {}
2747 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2748 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2749 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2751 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2756 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2761 static inline int throttled_lb_pair(struct task_group *tg,
2762 int src_cpu, int dest_cpu)
2767 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2769 #ifdef CONFIG_FAIR_GROUP_SCHED
2770 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2773 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2777 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2778 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2780 #endif /* CONFIG_CFS_BANDWIDTH */
2782 /**************************************************
2783 * CFS operations on tasks:
2786 #ifdef CONFIG_SCHED_HRTICK
2787 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2789 struct sched_entity *se = &p->se;
2790 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2792 WARN_ON(task_rq(p) != rq);
2794 if (cfs_rq->nr_running > 1) {
2795 u64 slice = sched_slice(cfs_rq, se);
2796 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2797 s64 delta = slice - ran;
2806 * Don't schedule slices shorter than 10000ns, that just
2807 * doesn't make sense. Rely on vruntime for fairness.
2810 delta = max_t(s64, 10000LL, delta);
2812 hrtick_start(rq, delta);
2817 * called from enqueue/dequeue and updates the hrtick when the
2818 * current task is from our class and nr_running is low enough
2821 static void hrtick_update(struct rq *rq)
2823 struct task_struct *curr = rq->curr;
2825 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2828 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2829 hrtick_start_fair(rq, curr);
2831 #else /* !CONFIG_SCHED_HRTICK */
2833 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2837 static inline void hrtick_update(struct rq *rq)
2843 * The enqueue_task method is called before nr_running is
2844 * increased. Here we update the fair scheduling stats and
2845 * then put the task into the rbtree:
2848 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2850 struct cfs_rq *cfs_rq;
2851 struct sched_entity *se = &p->se;
2853 for_each_sched_entity(se) {
2856 cfs_rq = cfs_rq_of(se);
2857 enqueue_entity(cfs_rq, se, flags);
2860 * end evaluation on encountering a throttled cfs_rq
2862 * note: in the case of encountering a throttled cfs_rq we will
2863 * post the final h_nr_running increment below.
2865 if (cfs_rq_throttled(cfs_rq))
2867 cfs_rq->h_nr_running++;
2869 flags = ENQUEUE_WAKEUP;
2872 for_each_sched_entity(se) {
2873 cfs_rq = cfs_rq_of(se);
2874 cfs_rq->h_nr_running++;
2876 if (cfs_rq_throttled(cfs_rq))
2879 update_cfs_shares(cfs_rq);
2880 update_entity_load_avg(se, 1);
2884 update_rq_runnable_avg(rq, rq->nr_running);
2890 static void set_next_buddy(struct sched_entity *se);
2893 * The dequeue_task method is called before nr_running is
2894 * decreased. We remove the task from the rbtree and
2895 * update the fair scheduling stats:
2897 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2899 struct cfs_rq *cfs_rq;
2900 struct sched_entity *se = &p->se;
2901 int task_sleep = flags & DEQUEUE_SLEEP;
2903 for_each_sched_entity(se) {
2904 cfs_rq = cfs_rq_of(se);
2905 dequeue_entity(cfs_rq, se, flags);
2908 * end evaluation on encountering a throttled cfs_rq
2910 * note: in the case of encountering a throttled cfs_rq we will
2911 * post the final h_nr_running decrement below.
2913 if (cfs_rq_throttled(cfs_rq))
2915 cfs_rq->h_nr_running--;
2917 /* Don't dequeue parent if it has other entities besides us */
2918 if (cfs_rq->load.weight) {
2920 * Bias pick_next to pick a task from this cfs_rq, as
2921 * p is sleeping when it is within its sched_slice.
2923 if (task_sleep && parent_entity(se))
2924 set_next_buddy(parent_entity(se));
2926 /* avoid re-evaluating load for this entity */
2927 se = parent_entity(se);
2930 flags |= DEQUEUE_SLEEP;
2933 for_each_sched_entity(se) {
2934 cfs_rq = cfs_rq_of(se);
2935 cfs_rq->h_nr_running--;
2937 if (cfs_rq_throttled(cfs_rq))
2940 update_cfs_shares(cfs_rq);
2941 update_entity_load_avg(se, 1);
2946 update_rq_runnable_avg(rq, 1);
2952 /* Used instead of source_load when we know the type == 0 */
2953 static unsigned long weighted_cpuload(const int cpu)
2955 return cpu_rq(cpu)->cfs.runnable_load_avg;
2959 * Return a low guess at the load of a migration-source cpu weighted
2960 * according to the scheduling class and "nice" value.
2962 * We want to under-estimate the load of migration sources, to
2963 * balance conservatively.
2965 static unsigned long source_load(int cpu, int type)
2967 struct rq *rq = cpu_rq(cpu);
2968 unsigned long total = weighted_cpuload(cpu);
2970 if (type == 0 || !sched_feat(LB_BIAS))
2973 return min(rq->cpu_load[type-1], total);
2977 * Return a high guess at the load of a migration-target cpu weighted
2978 * according to the scheduling class and "nice" value.
2980 static unsigned long target_load(int cpu, int type)
2982 struct rq *rq = cpu_rq(cpu);
2983 unsigned long total = weighted_cpuload(cpu);
2985 if (type == 0 || !sched_feat(LB_BIAS))
2988 return max(rq->cpu_load[type-1], total);
2991 static unsigned long power_of(int cpu)
2993 return cpu_rq(cpu)->cpu_power;
2996 static unsigned long cpu_avg_load_per_task(int cpu)
2998 struct rq *rq = cpu_rq(cpu);
2999 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3000 unsigned long load_avg = rq->cfs.runnable_load_avg;
3003 return load_avg / nr_running;
3008 static void record_wakee(struct task_struct *p)
3011 * Rough decay (wiping) for cost saving, don't worry
3012 * about the boundary, really active task won't care
3015 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3016 current->wakee_flips = 0;
3017 current->wakee_flip_decay_ts = jiffies;
3020 if (current->last_wakee != p) {
3021 current->last_wakee = p;
3022 current->wakee_flips++;
3026 static void task_waking_fair(struct task_struct *p)
3028 struct sched_entity *se = &p->se;
3029 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3032 #ifndef CONFIG_64BIT
3033 u64 min_vruntime_copy;
3036 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3038 min_vruntime = cfs_rq->min_vruntime;
3039 } while (min_vruntime != min_vruntime_copy);
3041 min_vruntime = cfs_rq->min_vruntime;
3044 se->vruntime -= min_vruntime;
3048 #ifdef CONFIG_FAIR_GROUP_SCHED
3050 * effective_load() calculates the load change as seen from the root_task_group
3052 * Adding load to a group doesn't make a group heavier, but can cause movement
3053 * of group shares between cpus. Assuming the shares were perfectly aligned one
3054 * can calculate the shift in shares.
3056 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3057 * on this @cpu and results in a total addition (subtraction) of @wg to the
3058 * total group weight.
3060 * Given a runqueue weight distribution (rw_i) we can compute a shares
3061 * distribution (s_i) using:
3063 * s_i = rw_i / \Sum rw_j (1)
3065 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3066 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3067 * shares distribution (s_i):
3069 * rw_i = { 2, 4, 1, 0 }
3070 * s_i = { 2/7, 4/7, 1/7, 0 }
3072 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3073 * task used to run on and the CPU the waker is running on), we need to
3074 * compute the effect of waking a task on either CPU and, in case of a sync
3075 * wakeup, compute the effect of the current task going to sleep.
3077 * So for a change of @wl to the local @cpu with an overall group weight change
3078 * of @wl we can compute the new shares distribution (s'_i) using:
3080 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3082 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3083 * differences in waking a task to CPU 0. The additional task changes the
3084 * weight and shares distributions like:
3086 * rw'_i = { 3, 4, 1, 0 }
3087 * s'_i = { 3/8, 4/8, 1/8, 0 }
3089 * We can then compute the difference in effective weight by using:
3091 * dw_i = S * (s'_i - s_i) (3)
3093 * Where 'S' is the group weight as seen by its parent.
3095 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3096 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3097 * 4/7) times the weight of the group.
3099 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3101 struct sched_entity *se = tg->se[cpu];
3103 if (!tg->parent) /* the trivial, non-cgroup case */
3106 for_each_sched_entity(se) {
3112 * W = @wg + \Sum rw_j
3114 W = wg + calc_tg_weight(tg, se->my_q);
3119 w = se->my_q->load.weight + wl;
3122 * wl = S * s'_i; see (2)
3125 wl = (w * tg->shares) / W;
3130 * Per the above, wl is the new se->load.weight value; since
3131 * those are clipped to [MIN_SHARES, ...) do so now. See
3132 * calc_cfs_shares().
3134 if (wl < MIN_SHARES)
3138 * wl = dw_i = S * (s'_i - s_i); see (3)
3140 wl -= se->load.weight;
3143 * Recursively apply this logic to all parent groups to compute
3144 * the final effective load change on the root group. Since
3145 * only the @tg group gets extra weight, all parent groups can
3146 * only redistribute existing shares. @wl is the shift in shares
3147 * resulting from this level per the above.
3156 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3157 unsigned long wl, unsigned long wg)
3164 static int wake_wide(struct task_struct *p)
3166 int factor = this_cpu_read(sd_llc_size);
3169 * Yeah, it's the switching-frequency, could means many wakee or
3170 * rapidly switch, use factor here will just help to automatically
3171 * adjust the loose-degree, so bigger node will lead to more pull.
3173 if (p->wakee_flips > factor) {
3175 * wakee is somewhat hot, it needs certain amount of cpu
3176 * resource, so if waker is far more hot, prefer to leave
3179 if (current->wakee_flips > (factor * p->wakee_flips))
3186 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3188 s64 this_load, load;
3189 int idx, this_cpu, prev_cpu;
3190 unsigned long tl_per_task;
3191 struct task_group *tg;
3192 unsigned long weight;
3196 * If we wake multiple tasks be careful to not bounce
3197 * ourselves around too much.
3203 this_cpu = smp_processor_id();
3204 prev_cpu = task_cpu(p);
3205 load = source_load(prev_cpu, idx);
3206 this_load = target_load(this_cpu, idx);
3209 * If sync wakeup then subtract the (maximum possible)
3210 * effect of the currently running task from the load
3211 * of the current CPU:
3214 tg = task_group(current);
3215 weight = current->se.load.weight;
3217 this_load += effective_load(tg, this_cpu, -weight, -weight);
3218 load += effective_load(tg, prev_cpu, 0, -weight);
3222 weight = p->se.load.weight;
3225 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3226 * due to the sync cause above having dropped this_load to 0, we'll
3227 * always have an imbalance, but there's really nothing you can do
3228 * about that, so that's good too.
3230 * Otherwise check if either cpus are near enough in load to allow this
3231 * task to be woken on this_cpu.
3233 if (this_load > 0) {
3234 s64 this_eff_load, prev_eff_load;
3236 this_eff_load = 100;
3237 this_eff_load *= power_of(prev_cpu);
3238 this_eff_load *= this_load +
3239 effective_load(tg, this_cpu, weight, weight);
3241 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3242 prev_eff_load *= power_of(this_cpu);
3243 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3245 balanced = this_eff_load <= prev_eff_load;
3250 * If the currently running task will sleep within
3251 * a reasonable amount of time then attract this newly
3254 if (sync && balanced)
3257 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3258 tl_per_task = cpu_avg_load_per_task(this_cpu);
3261 (this_load <= load &&
3262 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3264 * This domain has SD_WAKE_AFFINE and
3265 * p is cache cold in this domain, and
3266 * there is no bad imbalance.
3268 schedstat_inc(sd, ttwu_move_affine);
3269 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3277 * find_idlest_group finds and returns the least busy CPU group within the
3280 static struct sched_group *
3281 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3282 int this_cpu, int load_idx)
3284 struct sched_group *idlest = NULL, *group = sd->groups;
3285 unsigned long min_load = ULONG_MAX, this_load = 0;
3286 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3289 unsigned long load, avg_load;
3293 /* Skip over this group if it has no CPUs allowed */
3294 if (!cpumask_intersects(sched_group_cpus(group),
3295 tsk_cpus_allowed(p)))
3298 local_group = cpumask_test_cpu(this_cpu,
3299 sched_group_cpus(group));
3301 /* Tally up the load of all CPUs in the group */
3304 for_each_cpu(i, sched_group_cpus(group)) {
3305 /* Bias balancing toward cpus of our domain */
3307 load = source_load(i, load_idx);
3309 load = target_load(i, load_idx);
3314 /* Adjust by relative CPU power of the group */
3315 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3318 this_load = avg_load;
3319 } else if (avg_load < min_load) {
3320 min_load = avg_load;
3323 } while (group = group->next, group != sd->groups);
3325 if (!idlest || 100*this_load < imbalance*min_load)
3331 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3334 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3336 unsigned long load, min_load = ULONG_MAX;
3340 /* Traverse only the allowed CPUs */
3341 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3342 load = weighted_cpuload(i);
3344 if (load < min_load || (load == min_load && i == this_cpu)) {
3354 * Try and locate an idle CPU in the sched_domain.
3356 static int select_idle_sibling(struct task_struct *p, int target)
3358 struct sched_domain *sd;
3359 struct sched_group *sg;
3360 int i = task_cpu(p);
3362 if (idle_cpu(target))
3366 * If the prevous cpu is cache affine and idle, don't be stupid.
3368 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3372 * Otherwise, iterate the domains and find an elegible idle cpu.
3374 sd = rcu_dereference(per_cpu(sd_llc, target));
3375 for_each_lower_domain(sd) {
3378 if (!cpumask_intersects(sched_group_cpus(sg),
3379 tsk_cpus_allowed(p)))
3382 for_each_cpu(i, sched_group_cpus(sg)) {
3383 if (i == target || !idle_cpu(i))
3387 target = cpumask_first_and(sched_group_cpus(sg),
3388 tsk_cpus_allowed(p));
3392 } while (sg != sd->groups);
3399 * sched_balance_self: balance the current task (running on cpu) in domains
3400 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3403 * Balance, ie. select the least loaded group.
3405 * Returns the target CPU number, or the same CPU if no balancing is needed.
3407 * preempt must be disabled.
3410 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3412 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3413 int cpu = smp_processor_id();
3414 int prev_cpu = task_cpu(p);
3416 int want_affine = 0;
3417 int sync = wake_flags & WF_SYNC;
3419 if (p->nr_cpus_allowed == 1)
3422 if (sd_flag & SD_BALANCE_WAKE) {
3423 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3429 for_each_domain(cpu, tmp) {
3430 if (!(tmp->flags & SD_LOAD_BALANCE))
3434 * If both cpu and prev_cpu are part of this domain,
3435 * cpu is a valid SD_WAKE_AFFINE target.
3437 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3438 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3443 if (tmp->flags & sd_flag)
3448 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3451 new_cpu = select_idle_sibling(p, prev_cpu);
3456 int load_idx = sd->forkexec_idx;
3457 struct sched_group *group;
3460 if (!(sd->flags & sd_flag)) {
3465 if (sd_flag & SD_BALANCE_WAKE)
3466 load_idx = sd->wake_idx;
3468 group = find_idlest_group(sd, p, cpu, load_idx);
3474 new_cpu = find_idlest_cpu(group, p, cpu);
3475 if (new_cpu == -1 || new_cpu == cpu) {
3476 /* Now try balancing at a lower domain level of cpu */
3481 /* Now try balancing at a lower domain level of new_cpu */
3483 weight = sd->span_weight;
3485 for_each_domain(cpu, tmp) {
3486 if (weight <= tmp->span_weight)
3488 if (tmp->flags & sd_flag)
3491 /* while loop will break here if sd == NULL */
3500 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3501 * cfs_rq_of(p) references at time of call are still valid and identify the
3502 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3503 * other assumptions, including the state of rq->lock, should be made.
3506 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3508 struct sched_entity *se = &p->se;
3509 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3512 * Load tracking: accumulate removed load so that it can be processed
3513 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3514 * to blocked load iff they have a positive decay-count. It can never
3515 * be negative here since on-rq tasks have decay-count == 0.
3517 if (se->avg.decay_count) {
3518 se->avg.decay_count = -__synchronize_entity_decay(se);
3519 atomic_long_add(se->avg.load_avg_contrib,
3520 &cfs_rq->removed_load);
3523 #endif /* CONFIG_SMP */
3525 static unsigned long
3526 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3528 unsigned long gran = sysctl_sched_wakeup_granularity;
3531 * Since its curr running now, convert the gran from real-time
3532 * to virtual-time in his units.
3534 * By using 'se' instead of 'curr' we penalize light tasks, so
3535 * they get preempted easier. That is, if 'se' < 'curr' then
3536 * the resulting gran will be larger, therefore penalizing the
3537 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3538 * be smaller, again penalizing the lighter task.
3540 * This is especially important for buddies when the leftmost
3541 * task is higher priority than the buddy.
3543 return calc_delta_fair(gran, se);
3547 * Should 'se' preempt 'curr'.
3561 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3563 s64 gran, vdiff = curr->vruntime - se->vruntime;
3568 gran = wakeup_gran(curr, se);
3575 static void set_last_buddy(struct sched_entity *se)
3577 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3580 for_each_sched_entity(se)
3581 cfs_rq_of(se)->last = se;
3584 static void set_next_buddy(struct sched_entity *se)
3586 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3589 for_each_sched_entity(se)
3590 cfs_rq_of(se)->next = se;
3593 static void set_skip_buddy(struct sched_entity *se)
3595 for_each_sched_entity(se)
3596 cfs_rq_of(se)->skip = se;
3600 * Preempt the current task with a newly woken task if needed:
3602 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3604 struct task_struct *curr = rq->curr;
3605 struct sched_entity *se = &curr->se, *pse = &p->se;
3606 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3607 int scale = cfs_rq->nr_running >= sched_nr_latency;
3608 int next_buddy_marked = 0;
3610 if (unlikely(se == pse))
3614 * This is possible from callers such as move_task(), in which we
3615 * unconditionally check_prempt_curr() after an enqueue (which may have
3616 * lead to a throttle). This both saves work and prevents false
3617 * next-buddy nomination below.
3619 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3622 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3623 set_next_buddy(pse);
3624 next_buddy_marked = 1;
3628 * We can come here with TIF_NEED_RESCHED already set from new task
3631 * Note: this also catches the edge-case of curr being in a throttled
3632 * group (e.g. via set_curr_task), since update_curr() (in the
3633 * enqueue of curr) will have resulted in resched being set. This
3634 * prevents us from potentially nominating it as a false LAST_BUDDY
3637 if (test_tsk_need_resched(curr))
3640 /* Idle tasks are by definition preempted by non-idle tasks. */
3641 if (unlikely(curr->policy == SCHED_IDLE) &&
3642 likely(p->policy != SCHED_IDLE))
3646 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3647 * is driven by the tick):
3649 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3652 find_matching_se(&se, &pse);
3653 update_curr(cfs_rq_of(se));
3655 if (wakeup_preempt_entity(se, pse) == 1) {
3657 * Bias pick_next to pick the sched entity that is
3658 * triggering this preemption.
3660 if (!next_buddy_marked)
3661 set_next_buddy(pse);
3670 * Only set the backward buddy when the current task is still
3671 * on the rq. This can happen when a wakeup gets interleaved
3672 * with schedule on the ->pre_schedule() or idle_balance()
3673 * point, either of which can * drop the rq lock.
3675 * Also, during early boot the idle thread is in the fair class,
3676 * for obvious reasons its a bad idea to schedule back to it.
3678 if (unlikely(!se->on_rq || curr == rq->idle))
3681 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3685 static struct task_struct *pick_next_task_fair(struct rq *rq)
3687 struct task_struct *p;
3688 struct cfs_rq *cfs_rq = &rq->cfs;
3689 struct sched_entity *se;
3691 if (!cfs_rq->nr_running)
3695 se = pick_next_entity(cfs_rq);
3696 set_next_entity(cfs_rq, se);
3697 cfs_rq = group_cfs_rq(se);
3701 if (hrtick_enabled(rq))
3702 hrtick_start_fair(rq, p);
3708 * Account for a descheduled task:
3710 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3712 struct sched_entity *se = &prev->se;
3713 struct cfs_rq *cfs_rq;
3715 for_each_sched_entity(se) {
3716 cfs_rq = cfs_rq_of(se);
3717 put_prev_entity(cfs_rq, se);
3722 * sched_yield() is very simple
3724 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3726 static void yield_task_fair(struct rq *rq)
3728 struct task_struct *curr = rq->curr;
3729 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3730 struct sched_entity *se = &curr->se;
3733 * Are we the only task in the tree?
3735 if (unlikely(rq->nr_running == 1))
3738 clear_buddies(cfs_rq, se);
3740 if (curr->policy != SCHED_BATCH) {
3741 update_rq_clock(rq);
3743 * Update run-time statistics of the 'current'.
3745 update_curr(cfs_rq);
3747 * Tell update_rq_clock() that we've just updated,
3748 * so we don't do microscopic update in schedule()
3749 * and double the fastpath cost.
3751 rq->skip_clock_update = 1;
3757 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3759 struct sched_entity *se = &p->se;
3761 /* throttled hierarchies are not runnable */
3762 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3765 /* Tell the scheduler that we'd really like pse to run next. */
3768 yield_task_fair(rq);
3774 /**************************************************
3775 * Fair scheduling class load-balancing methods.
3779 * The purpose of load-balancing is to achieve the same basic fairness the
3780 * per-cpu scheduler provides, namely provide a proportional amount of compute
3781 * time to each task. This is expressed in the following equation:
3783 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3785 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3786 * W_i,0 is defined as:
3788 * W_i,0 = \Sum_j w_i,j (2)
3790 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3791 * is derived from the nice value as per prio_to_weight[].
3793 * The weight average is an exponential decay average of the instantaneous
3796 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3798 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3799 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3800 * can also include other factors [XXX].
3802 * To achieve this balance we define a measure of imbalance which follows
3803 * directly from (1):
3805 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3807 * We them move tasks around to minimize the imbalance. In the continuous
3808 * function space it is obvious this converges, in the discrete case we get
3809 * a few fun cases generally called infeasible weight scenarios.
3812 * - infeasible weights;
3813 * - local vs global optima in the discrete case. ]
3818 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3819 * for all i,j solution, we create a tree of cpus that follows the hardware
3820 * topology where each level pairs two lower groups (or better). This results
3821 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3822 * tree to only the first of the previous level and we decrease the frequency
3823 * of load-balance at each level inv. proportional to the number of cpus in
3829 * \Sum { --- * --- * 2^i } = O(n) (5)
3831 * `- size of each group
3832 * | | `- number of cpus doing load-balance
3834 * `- sum over all levels
3836 * Coupled with a limit on how many tasks we can migrate every balance pass,
3837 * this makes (5) the runtime complexity of the balancer.
3839 * An important property here is that each CPU is still (indirectly) connected
3840 * to every other cpu in at most O(log n) steps:
3842 * The adjacency matrix of the resulting graph is given by:
3845 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3848 * And you'll find that:
3850 * A^(log_2 n)_i,j != 0 for all i,j (7)
3852 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3853 * The task movement gives a factor of O(m), giving a convergence complexity
3856 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3861 * In order to avoid CPUs going idle while there's still work to do, new idle
3862 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3863 * tree itself instead of relying on other CPUs to bring it work.
3865 * This adds some complexity to both (5) and (8) but it reduces the total idle
3873 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3876 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3881 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3883 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3885 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3888 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3889 * rewrite all of this once again.]
3892 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3894 #define LBF_ALL_PINNED 0x01
3895 #define LBF_NEED_BREAK 0x02
3896 #define LBF_DST_PINNED 0x04
3897 #define LBF_SOME_PINNED 0x08
3900 struct sched_domain *sd;
3908 struct cpumask *dst_grpmask;
3910 enum cpu_idle_type idle;
3912 /* The set of CPUs under consideration for load-balancing */
3913 struct cpumask *cpus;
3918 unsigned int loop_break;
3919 unsigned int loop_max;
3923 * move_task - move a task from one runqueue to another runqueue.
3924 * Both runqueues must be locked.
3926 static void move_task(struct task_struct *p, struct lb_env *env)
3928 deactivate_task(env->src_rq, p, 0);
3929 set_task_cpu(p, env->dst_cpu);
3930 activate_task(env->dst_rq, p, 0);
3931 check_preempt_curr(env->dst_rq, p, 0);
3935 * Is this task likely cache-hot:
3938 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3942 if (p->sched_class != &fair_sched_class)
3945 if (unlikely(p->policy == SCHED_IDLE))
3949 * Buddy candidates are cache hot:
3951 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3952 (&p->se == cfs_rq_of(&p->se)->next ||
3953 &p->se == cfs_rq_of(&p->se)->last))
3956 if (sysctl_sched_migration_cost == -1)
3958 if (sysctl_sched_migration_cost == 0)
3961 delta = now - p->se.exec_start;
3963 return delta < (s64)sysctl_sched_migration_cost;
3967 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3970 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3972 int tsk_cache_hot = 0;
3974 * We do not migrate tasks that are:
3975 * 1) throttled_lb_pair, or
3976 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3977 * 3) running (obviously), or
3978 * 4) are cache-hot on their current CPU.
3980 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3983 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3986 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3988 env->flags |= LBF_SOME_PINNED;
3991 * Remember if this task can be migrated to any other cpu in
3992 * our sched_group. We may want to revisit it if we couldn't
3993 * meet load balance goals by pulling other tasks on src_cpu.
3995 * Also avoid computing new_dst_cpu if we have already computed
3996 * one in current iteration.
3998 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4001 /* Prevent to re-select dst_cpu via env's cpus */
4002 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4003 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4004 env->flags |= LBF_DST_PINNED;
4005 env->new_dst_cpu = cpu;
4013 /* Record that we found atleast one task that could run on dst_cpu */
4014 env->flags &= ~LBF_ALL_PINNED;
4016 if (task_running(env->src_rq, p)) {
4017 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4022 * Aggressive migration if:
4023 * 1) task is cache cold, or
4024 * 2) too many balance attempts have failed.
4027 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4028 if (!tsk_cache_hot ||
4029 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4031 if (tsk_cache_hot) {
4032 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4033 schedstat_inc(p, se.statistics.nr_forced_migrations);
4039 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4044 * move_one_task tries to move exactly one task from busiest to this_rq, as
4045 * part of active balancing operations within "domain".
4046 * Returns 1 if successful and 0 otherwise.
4048 * Called with both runqueues locked.
4050 static int move_one_task(struct lb_env *env)
4052 struct task_struct *p, *n;
4054 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4055 if (!can_migrate_task(p, env))
4060 * Right now, this is only the second place move_task()
4061 * is called, so we can safely collect move_task()
4062 * stats here rather than inside move_task().
4064 schedstat_inc(env->sd, lb_gained[env->idle]);
4070 static unsigned long task_h_load(struct task_struct *p);
4072 static const unsigned int sched_nr_migrate_break = 32;
4075 * move_tasks tries to move up to imbalance weighted load from busiest to
4076 * this_rq, as part of a balancing operation within domain "sd".
4077 * Returns 1 if successful and 0 otherwise.
4079 * Called with both runqueues locked.
4081 static int move_tasks(struct lb_env *env)
4083 struct list_head *tasks = &env->src_rq->cfs_tasks;
4084 struct task_struct *p;
4088 if (env->imbalance <= 0)
4091 while (!list_empty(tasks)) {
4092 p = list_first_entry(tasks, struct task_struct, se.group_node);
4095 /* We've more or less seen every task there is, call it quits */
4096 if (env->loop > env->loop_max)
4099 /* take a breather every nr_migrate tasks */
4100 if (env->loop > env->loop_break) {
4101 env->loop_break += sched_nr_migrate_break;
4102 env->flags |= LBF_NEED_BREAK;
4106 if (!can_migrate_task(p, env))
4109 load = task_h_load(p);
4111 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4114 if ((load / 2) > env->imbalance)
4119 env->imbalance -= load;
4121 #ifdef CONFIG_PREEMPT
4123 * NEWIDLE balancing is a source of latency, so preemptible
4124 * kernels will stop after the first task is pulled to minimize
4125 * the critical section.
4127 if (env->idle == CPU_NEWLY_IDLE)
4132 * We only want to steal up to the prescribed amount of
4135 if (env->imbalance <= 0)
4140 list_move_tail(&p->se.group_node, tasks);
4144 * Right now, this is one of only two places move_task() is called,
4145 * so we can safely collect move_task() stats here rather than
4146 * inside move_task().
4148 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4153 #ifdef CONFIG_FAIR_GROUP_SCHED
4155 * update tg->load_weight by folding this cpu's load_avg
4157 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4159 struct sched_entity *se = tg->se[cpu];
4160 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4162 /* throttled entities do not contribute to load */
4163 if (throttled_hierarchy(cfs_rq))
4166 update_cfs_rq_blocked_load(cfs_rq, 1);
4169 update_entity_load_avg(se, 1);
4171 * We pivot on our runnable average having decayed to zero for
4172 * list removal. This generally implies that all our children
4173 * have also been removed (modulo rounding error or bandwidth
4174 * control); however, such cases are rare and we can fix these
4177 * TODO: fix up out-of-order children on enqueue.
4179 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4180 list_del_leaf_cfs_rq(cfs_rq);
4182 struct rq *rq = rq_of(cfs_rq);
4183 update_rq_runnable_avg(rq, rq->nr_running);
4187 static void update_blocked_averages(int cpu)
4189 struct rq *rq = cpu_rq(cpu);
4190 struct cfs_rq *cfs_rq;
4191 unsigned long flags;
4193 raw_spin_lock_irqsave(&rq->lock, flags);
4194 update_rq_clock(rq);
4196 * Iterates the task_group tree in a bottom up fashion, see
4197 * list_add_leaf_cfs_rq() for details.
4199 for_each_leaf_cfs_rq(rq, cfs_rq) {
4201 * Note: We may want to consider periodically releasing
4202 * rq->lock about these updates so that creating many task
4203 * groups does not result in continually extending hold time.
4205 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4208 raw_spin_unlock_irqrestore(&rq->lock, flags);
4212 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4213 * This needs to be done in a top-down fashion because the load of a child
4214 * group is a fraction of its parents load.
4216 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4218 struct rq *rq = rq_of(cfs_rq);
4219 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4220 unsigned long now = jiffies;
4223 if (cfs_rq->last_h_load_update == now)
4226 cfs_rq->h_load_next = NULL;
4227 for_each_sched_entity(se) {
4228 cfs_rq = cfs_rq_of(se);
4229 cfs_rq->h_load_next = se;
4230 if (cfs_rq->last_h_load_update == now)
4235 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4236 cfs_rq->last_h_load_update = now;
4239 while ((se = cfs_rq->h_load_next) != NULL) {
4240 load = cfs_rq->h_load;
4241 load = div64_ul(load * se->avg.load_avg_contrib,
4242 cfs_rq->runnable_load_avg + 1);
4243 cfs_rq = group_cfs_rq(se);
4244 cfs_rq->h_load = load;
4245 cfs_rq->last_h_load_update = now;
4249 static unsigned long task_h_load(struct task_struct *p)
4251 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4253 update_cfs_rq_h_load(cfs_rq);
4254 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4255 cfs_rq->runnable_load_avg + 1);
4258 static inline void update_blocked_averages(int cpu)
4262 static unsigned long task_h_load(struct task_struct *p)
4264 return p->se.avg.load_avg_contrib;
4268 /********** Helpers for find_busiest_group ************************/
4270 * sg_lb_stats - stats of a sched_group required for load_balancing
4272 struct sg_lb_stats {
4273 unsigned long avg_load; /*Avg load across the CPUs of the group */
4274 unsigned long group_load; /* Total load over the CPUs of the group */
4275 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4276 unsigned long load_per_task;
4277 unsigned long group_power;
4278 unsigned int sum_nr_running; /* Nr tasks running in the group */
4279 unsigned int group_capacity;
4280 unsigned int idle_cpus;
4281 unsigned int group_weight;
4282 int group_imb; /* Is there an imbalance in the group ? */
4283 int group_has_capacity; /* Is there extra capacity in the group? */
4287 * sd_lb_stats - Structure to store the statistics of a sched_domain
4288 * during load balancing.
4290 struct sd_lb_stats {
4291 struct sched_group *busiest; /* Busiest group in this sd */
4292 struct sched_group *local; /* Local group in this sd */
4293 unsigned long total_load; /* Total load of all groups in sd */
4294 unsigned long total_pwr; /* Total power of all groups in sd */
4295 unsigned long avg_load; /* Average load across all groups in sd */
4297 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4298 struct sg_lb_stats local_stat; /* Statistics of the local group */
4301 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4304 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4305 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4306 * We must however clear busiest_stat::avg_load because
4307 * update_sd_pick_busiest() reads this before assignment.
4309 *sds = (struct sd_lb_stats){
4321 * get_sd_load_idx - Obtain the load index for a given sched domain.
4322 * @sd: The sched_domain whose load_idx is to be obtained.
4323 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4325 * Return: The load index.
4327 static inline int get_sd_load_idx(struct sched_domain *sd,
4328 enum cpu_idle_type idle)
4334 load_idx = sd->busy_idx;
4337 case CPU_NEWLY_IDLE:
4338 load_idx = sd->newidle_idx;
4341 load_idx = sd->idle_idx;
4348 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4350 return SCHED_POWER_SCALE;
4353 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4355 return default_scale_freq_power(sd, cpu);
4358 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4360 unsigned long weight = sd->span_weight;
4361 unsigned long smt_gain = sd->smt_gain;
4368 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4370 return default_scale_smt_power(sd, cpu);
4373 static unsigned long scale_rt_power(int cpu)
4375 struct rq *rq = cpu_rq(cpu);
4376 u64 total, available, age_stamp, avg;
4379 * Since we're reading these variables without serialization make sure
4380 * we read them once before doing sanity checks on them.
4382 age_stamp = ACCESS_ONCE(rq->age_stamp);
4383 avg = ACCESS_ONCE(rq->rt_avg);
4385 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4387 if (unlikely(total < avg)) {
4388 /* Ensures that power won't end up being negative */
4391 available = total - avg;
4394 if (unlikely((s64)total < SCHED_POWER_SCALE))
4395 total = SCHED_POWER_SCALE;
4397 total >>= SCHED_POWER_SHIFT;
4399 return div_u64(available, total);
4402 static void update_cpu_power(struct sched_domain *sd, int cpu)
4404 unsigned long weight = sd->span_weight;
4405 unsigned long power = SCHED_POWER_SCALE;
4406 struct sched_group *sdg = sd->groups;
4408 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4409 if (sched_feat(ARCH_POWER))
4410 power *= arch_scale_smt_power(sd, cpu);
4412 power *= default_scale_smt_power(sd, cpu);
4414 power >>= SCHED_POWER_SHIFT;
4417 sdg->sgp->power_orig = power;
4419 if (sched_feat(ARCH_POWER))
4420 power *= arch_scale_freq_power(sd, cpu);
4422 power *= default_scale_freq_power(sd, cpu);
4424 power >>= SCHED_POWER_SHIFT;
4426 power *= scale_rt_power(cpu);
4427 power >>= SCHED_POWER_SHIFT;
4432 cpu_rq(cpu)->cpu_power = power;
4433 sdg->sgp->power = power;
4436 void update_group_power(struct sched_domain *sd, int cpu)
4438 struct sched_domain *child = sd->child;
4439 struct sched_group *group, *sdg = sd->groups;
4440 unsigned long power, power_orig;
4441 unsigned long interval;
4443 interval = msecs_to_jiffies(sd->balance_interval);
4444 interval = clamp(interval, 1UL, max_load_balance_interval);
4445 sdg->sgp->next_update = jiffies + interval;
4448 update_cpu_power(sd, cpu);
4452 power_orig = power = 0;
4454 if (child->flags & SD_OVERLAP) {
4456 * SD_OVERLAP domains cannot assume that child groups
4457 * span the current group.
4460 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4461 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4463 power_orig += sg->sgp->power_orig;
4464 power += sg->sgp->power;
4468 * !SD_OVERLAP domains can assume that child groups
4469 * span the current group.
4472 group = child->groups;
4474 power_orig += group->sgp->power_orig;
4475 power += group->sgp->power;
4476 group = group->next;
4477 } while (group != child->groups);
4480 sdg->sgp->power_orig = power_orig;
4481 sdg->sgp->power = power;
4485 * Try and fix up capacity for tiny siblings, this is needed when
4486 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4487 * which on its own isn't powerful enough.
4489 * See update_sd_pick_busiest() and check_asym_packing().
4492 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4495 * Only siblings can have significantly less than SCHED_POWER_SCALE
4497 if (!(sd->flags & SD_SHARE_CPUPOWER))
4501 * If ~90% of the cpu_power is still there, we're good.
4503 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4510 * Group imbalance indicates (and tries to solve) the problem where balancing
4511 * groups is inadequate due to tsk_cpus_allowed() constraints.
4513 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4514 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4517 * { 0 1 2 3 } { 4 5 6 7 }
4520 * If we were to balance group-wise we'd place two tasks in the first group and
4521 * two tasks in the second group. Clearly this is undesired as it will overload
4522 * cpu 3 and leave one of the cpus in the second group unused.
4524 * The current solution to this issue is detecting the skew in the first group
4525 * by noticing the lower domain failed to reach balance and had difficulty
4526 * moving tasks due to affinity constraints.
4528 * When this is so detected; this group becomes a candidate for busiest; see
4529 * update_sd_pick_busiest(). And calculcate_imbalance() and
4530 * find_busiest_group() avoid some of the usual balance conditions to allow it
4531 * to create an effective group imbalance.
4533 * This is a somewhat tricky proposition since the next run might not find the
4534 * group imbalance and decide the groups need to be balanced again. A most
4535 * subtle and fragile situation.
4538 static inline int sg_imbalanced(struct sched_group *group)
4540 return group->sgp->imbalance;
4544 * Compute the group capacity.
4546 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4547 * first dividing out the smt factor and computing the actual number of cores
4548 * and limit power unit capacity with that.
4550 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4552 unsigned int capacity, smt, cpus;
4553 unsigned int power, power_orig;
4555 power = group->sgp->power;
4556 power_orig = group->sgp->power_orig;
4557 cpus = group->group_weight;
4559 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4560 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4561 capacity = cpus / smt; /* cores */
4563 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4565 capacity = fix_small_capacity(env->sd, group);
4571 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4572 * @env: The load balancing environment.
4573 * @group: sched_group whose statistics are to be updated.
4574 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4575 * @local_group: Does group contain this_cpu.
4576 * @sgs: variable to hold the statistics for this group.
4578 static inline void update_sg_lb_stats(struct lb_env *env,
4579 struct sched_group *group, int load_idx,
4580 int local_group, struct sg_lb_stats *sgs)
4582 unsigned long nr_running;
4586 memset(sgs, 0, sizeof(*sgs));
4588 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4589 struct rq *rq = cpu_rq(i);
4591 nr_running = rq->nr_running;
4593 /* Bias balancing toward cpus of our domain */
4595 load = target_load(i, load_idx);
4597 load = source_load(i, load_idx);
4599 sgs->group_load += load;
4600 sgs->sum_nr_running += nr_running;
4601 sgs->sum_weighted_load += weighted_cpuload(i);
4606 /* Adjust by relative CPU power of the group */
4607 sgs->group_power = group->sgp->power;
4608 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4610 if (sgs->sum_nr_running)
4611 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4613 sgs->group_weight = group->group_weight;
4615 sgs->group_imb = sg_imbalanced(group);
4616 sgs->group_capacity = sg_capacity(env, group);
4618 if (sgs->group_capacity > sgs->sum_nr_running)
4619 sgs->group_has_capacity = 1;
4623 * update_sd_pick_busiest - return 1 on busiest group
4624 * @env: The load balancing environment.
4625 * @sds: sched_domain statistics
4626 * @sg: sched_group candidate to be checked for being the busiest
4627 * @sgs: sched_group statistics
4629 * Determine if @sg is a busier group than the previously selected
4632 * Return: %true if @sg is a busier group than the previously selected
4633 * busiest group. %false otherwise.
4635 static bool update_sd_pick_busiest(struct lb_env *env,
4636 struct sd_lb_stats *sds,
4637 struct sched_group *sg,
4638 struct sg_lb_stats *sgs)
4640 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4643 if (sgs->sum_nr_running > sgs->group_capacity)
4650 * ASYM_PACKING needs to move all the work to the lowest
4651 * numbered CPUs in the group, therefore mark all groups
4652 * higher than ourself as busy.
4654 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4655 env->dst_cpu < group_first_cpu(sg)) {
4659 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4667 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4668 * @env: The load balancing environment.
4669 * @balance: Should we balance.
4670 * @sds: variable to hold the statistics for this sched_domain.
4672 static inline void update_sd_lb_stats(struct lb_env *env,
4673 struct sd_lb_stats *sds)
4675 struct sched_domain *child = env->sd->child;
4676 struct sched_group *sg = env->sd->groups;
4677 struct sg_lb_stats tmp_sgs;
4678 int load_idx, prefer_sibling = 0;
4680 if (child && child->flags & SD_PREFER_SIBLING)
4683 load_idx = get_sd_load_idx(env->sd, env->idle);
4686 struct sg_lb_stats *sgs = &tmp_sgs;
4689 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4692 sgs = &sds->local_stat;
4694 if (env->idle != CPU_NEWLY_IDLE ||
4695 time_after_eq(jiffies, sg->sgp->next_update))
4696 update_group_power(env->sd, env->dst_cpu);
4699 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4705 * In case the child domain prefers tasks go to siblings
4706 * first, lower the sg capacity to one so that we'll try
4707 * and move all the excess tasks away. We lower the capacity
4708 * of a group only if the local group has the capacity to fit
4709 * these excess tasks, i.e. nr_running < group_capacity. The
4710 * extra check prevents the case where you always pull from the
4711 * heaviest group when it is already under-utilized (possible
4712 * with a large weight task outweighs the tasks on the system).
4714 if (prefer_sibling && sds->local &&
4715 sds->local_stat.group_has_capacity)
4716 sgs->group_capacity = min(sgs->group_capacity, 1U);
4718 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4720 sds->busiest_stat = *sgs;
4724 /* Now, start updating sd_lb_stats */
4725 sds->total_load += sgs->group_load;
4726 sds->total_pwr += sgs->group_power;
4729 } while (sg != env->sd->groups);
4733 * check_asym_packing - Check to see if the group is packed into the
4736 * This is primarily intended to used at the sibling level. Some
4737 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4738 * case of POWER7, it can move to lower SMT modes only when higher
4739 * threads are idle. When in lower SMT modes, the threads will
4740 * perform better since they share less core resources. Hence when we
4741 * have idle threads, we want them to be the higher ones.
4743 * This packing function is run on idle threads. It checks to see if
4744 * the busiest CPU in this domain (core in the P7 case) has a higher
4745 * CPU number than the packing function is being run on. Here we are
4746 * assuming lower CPU number will be equivalent to lower a SMT thread
4749 * Return: 1 when packing is required and a task should be moved to
4750 * this CPU. The amount of the imbalance is returned in *imbalance.
4752 * @env: The load balancing environment.
4753 * @sds: Statistics of the sched_domain which is to be packed
4755 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4759 if (!(env->sd->flags & SD_ASYM_PACKING))
4765 busiest_cpu = group_first_cpu(sds->busiest);
4766 if (env->dst_cpu > busiest_cpu)
4769 env->imbalance = DIV_ROUND_CLOSEST(
4770 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
4777 * fix_small_imbalance - Calculate the minor imbalance that exists
4778 * amongst the groups of a sched_domain, during
4780 * @env: The load balancing environment.
4781 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4784 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4786 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4787 unsigned int imbn = 2;
4788 unsigned long scaled_busy_load_per_task;
4789 struct sg_lb_stats *local, *busiest;
4791 local = &sds->local_stat;
4792 busiest = &sds->busiest_stat;
4794 if (!local->sum_nr_running)
4795 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
4796 else if (busiest->load_per_task > local->load_per_task)
4799 scaled_busy_load_per_task =
4800 (busiest->load_per_task * SCHED_POWER_SCALE) /
4801 busiest->group_power;
4803 if (busiest->avg_load + scaled_busy_load_per_task >=
4804 local->avg_load + (scaled_busy_load_per_task * imbn)) {
4805 env->imbalance = busiest->load_per_task;
4810 * OK, we don't have enough imbalance to justify moving tasks,
4811 * however we may be able to increase total CPU power used by
4815 pwr_now += busiest->group_power *
4816 min(busiest->load_per_task, busiest->avg_load);
4817 pwr_now += local->group_power *
4818 min(local->load_per_task, local->avg_load);
4819 pwr_now /= SCHED_POWER_SCALE;
4821 /* Amount of load we'd subtract */
4822 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4823 busiest->group_power;
4824 if (busiest->avg_load > tmp) {
4825 pwr_move += busiest->group_power *
4826 min(busiest->load_per_task,
4827 busiest->avg_load - tmp);
4830 /* Amount of load we'd add */
4831 if (busiest->avg_load * busiest->group_power <
4832 busiest->load_per_task * SCHED_POWER_SCALE) {
4833 tmp = (busiest->avg_load * busiest->group_power) /
4836 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4839 pwr_move += local->group_power *
4840 min(local->load_per_task, local->avg_load + tmp);
4841 pwr_move /= SCHED_POWER_SCALE;
4843 /* Move if we gain throughput */
4844 if (pwr_move > pwr_now)
4845 env->imbalance = busiest->load_per_task;
4849 * calculate_imbalance - Calculate the amount of imbalance present within the
4850 * groups of a given sched_domain during load balance.
4851 * @env: load balance environment
4852 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4854 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4856 unsigned long max_pull, load_above_capacity = ~0UL;
4857 struct sg_lb_stats *local, *busiest;
4859 local = &sds->local_stat;
4860 busiest = &sds->busiest_stat;
4862 if (busiest->group_imb) {
4864 * In the group_imb case we cannot rely on group-wide averages
4865 * to ensure cpu-load equilibrium, look at wider averages. XXX
4867 busiest->load_per_task =
4868 min(busiest->load_per_task, sds->avg_load);
4872 * In the presence of smp nice balancing, certain scenarios can have
4873 * max load less than avg load(as we skip the groups at or below
4874 * its cpu_power, while calculating max_load..)
4876 if (busiest->avg_load <= sds->avg_load ||
4877 local->avg_load >= sds->avg_load) {
4879 return fix_small_imbalance(env, sds);
4882 if (!busiest->group_imb) {
4884 * Don't want to pull so many tasks that a group would go idle.
4885 * Except of course for the group_imb case, since then we might
4886 * have to drop below capacity to reach cpu-load equilibrium.
4888 load_above_capacity =
4889 (busiest->sum_nr_running - busiest->group_capacity);
4891 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4892 load_above_capacity /= busiest->group_power;
4896 * We're trying to get all the cpus to the average_load, so we don't
4897 * want to push ourselves above the average load, nor do we wish to
4898 * reduce the max loaded cpu below the average load. At the same time,
4899 * we also don't want to reduce the group load below the group capacity
4900 * (so that we can implement power-savings policies etc). Thus we look
4901 * for the minimum possible imbalance.
4903 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
4905 /* How much load to actually move to equalise the imbalance */
4906 env->imbalance = min(
4907 max_pull * busiest->group_power,
4908 (sds->avg_load - local->avg_load) * local->group_power
4909 ) / SCHED_POWER_SCALE;
4912 * if *imbalance is less than the average load per runnable task
4913 * there is no guarantee that any tasks will be moved so we'll have
4914 * a think about bumping its value to force at least one task to be
4917 if (env->imbalance < busiest->load_per_task)
4918 return fix_small_imbalance(env, sds);
4921 /******* find_busiest_group() helpers end here *********************/
4924 * find_busiest_group - Returns the busiest group within the sched_domain
4925 * if there is an imbalance. If there isn't an imbalance, and
4926 * the user has opted for power-savings, it returns a group whose
4927 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4928 * such a group exists.
4930 * Also calculates the amount of weighted load which should be moved
4931 * to restore balance.
4933 * @env: The load balancing environment.
4935 * Return: - The busiest group if imbalance exists.
4936 * - If no imbalance and user has opted for power-savings balance,
4937 * return the least loaded group whose CPUs can be
4938 * put to idle by rebalancing its tasks onto our group.
4940 static struct sched_group *find_busiest_group(struct lb_env *env)
4942 struct sg_lb_stats *local, *busiest;
4943 struct sd_lb_stats sds;
4945 init_sd_lb_stats(&sds);
4948 * Compute the various statistics relavent for load balancing at
4951 update_sd_lb_stats(env, &sds);
4952 local = &sds.local_stat;
4953 busiest = &sds.busiest_stat;
4955 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4956 check_asym_packing(env, &sds))
4959 /* There is no busy sibling group to pull tasks from */
4960 if (!sds.busiest || busiest->sum_nr_running == 0)
4963 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4966 * If the busiest group is imbalanced the below checks don't
4967 * work because they assume all things are equal, which typically
4968 * isn't true due to cpus_allowed constraints and the like.
4970 if (busiest->group_imb)
4973 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4974 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
4975 !busiest->group_has_capacity)
4979 * If the local group is more busy than the selected busiest group
4980 * don't try and pull any tasks.
4982 if (local->avg_load >= busiest->avg_load)
4986 * Don't pull any tasks if this group is already above the domain
4989 if (local->avg_load >= sds.avg_load)
4992 if (env->idle == CPU_IDLE) {
4994 * This cpu is idle. If the busiest group load doesn't
4995 * have more tasks than the number of available cpu's and
4996 * there is no imbalance between this and busiest group
4997 * wrt to idle cpu's, it is balanced.
4999 if ((local->idle_cpus < busiest->idle_cpus) &&
5000 busiest->sum_nr_running <= busiest->group_weight)
5004 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5005 * imbalance_pct to be conservative.
5007 if (100 * busiest->avg_load <=
5008 env->sd->imbalance_pct * local->avg_load)
5013 /* Looks like there is an imbalance. Compute it */
5014 calculate_imbalance(env, &sds);
5023 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5025 static struct rq *find_busiest_queue(struct lb_env *env,
5026 struct sched_group *group)
5028 struct rq *busiest = NULL, *rq;
5029 unsigned long busiest_load = 0, busiest_power = 1;
5032 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5033 unsigned long power = power_of(i);
5034 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5039 capacity = fix_small_capacity(env->sd, group);
5042 wl = weighted_cpuload(i);
5045 * When comparing with imbalance, use weighted_cpuload()
5046 * which is not scaled with the cpu power.
5048 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5052 * For the load comparisons with the other cpu's, consider
5053 * the weighted_cpuload() scaled with the cpu power, so that
5054 * the load can be moved away from the cpu that is potentially
5055 * running at a lower capacity.
5057 * Thus we're looking for max(wl_i / power_i), crosswise
5058 * multiplication to rid ourselves of the division works out
5059 * to: wl_i * power_j > wl_j * power_i; where j is our
5062 if (wl * busiest_power > busiest_load * power) {
5064 busiest_power = power;
5073 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5074 * so long as it is large enough.
5076 #define MAX_PINNED_INTERVAL 512
5078 /* Working cpumask for load_balance and load_balance_newidle. */
5079 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5081 static int need_active_balance(struct lb_env *env)
5083 struct sched_domain *sd = env->sd;
5085 if (env->idle == CPU_NEWLY_IDLE) {
5088 * ASYM_PACKING needs to force migrate tasks from busy but
5089 * higher numbered CPUs in order to pack all tasks in the
5090 * lowest numbered CPUs.
5092 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5096 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5099 static int active_load_balance_cpu_stop(void *data);
5101 static int should_we_balance(struct lb_env *env)
5103 struct sched_group *sg = env->sd->groups;
5104 struct cpumask *sg_cpus, *sg_mask;
5105 int cpu, balance_cpu = -1;
5108 * In the newly idle case, we will allow all the cpu's
5109 * to do the newly idle load balance.
5111 if (env->idle == CPU_NEWLY_IDLE)
5114 sg_cpus = sched_group_cpus(sg);
5115 sg_mask = sched_group_mask(sg);
5116 /* Try to find first idle cpu */
5117 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5118 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5125 if (balance_cpu == -1)
5126 balance_cpu = group_balance_cpu(sg);
5129 * First idle cpu or the first cpu(busiest) in this sched group
5130 * is eligible for doing load balancing at this and above domains.
5132 return balance_cpu == env->dst_cpu;
5136 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5137 * tasks if there is an imbalance.
5139 static int load_balance(int this_cpu, struct rq *this_rq,
5140 struct sched_domain *sd, enum cpu_idle_type idle,
5141 int *continue_balancing)
5143 int ld_moved, cur_ld_moved, active_balance = 0;
5144 struct sched_domain *sd_parent = sd->parent;
5145 struct sched_group *group;
5147 unsigned long flags;
5148 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5150 struct lb_env env = {
5152 .dst_cpu = this_cpu,
5154 .dst_grpmask = sched_group_cpus(sd->groups),
5156 .loop_break = sched_nr_migrate_break,
5161 * For NEWLY_IDLE load_balancing, we don't need to consider
5162 * other cpus in our group
5164 if (idle == CPU_NEWLY_IDLE)
5165 env.dst_grpmask = NULL;
5167 cpumask_copy(cpus, cpu_active_mask);
5169 schedstat_inc(sd, lb_count[idle]);
5172 if (!should_we_balance(&env)) {
5173 *continue_balancing = 0;
5177 group = find_busiest_group(&env);
5179 schedstat_inc(sd, lb_nobusyg[idle]);
5183 busiest = find_busiest_queue(&env, group);
5185 schedstat_inc(sd, lb_nobusyq[idle]);
5189 BUG_ON(busiest == env.dst_rq);
5191 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5194 if (busiest->nr_running > 1) {
5196 * Attempt to move tasks. If find_busiest_group has found
5197 * an imbalance but busiest->nr_running <= 1, the group is
5198 * still unbalanced. ld_moved simply stays zero, so it is
5199 * correctly treated as an imbalance.
5201 env.flags |= LBF_ALL_PINNED;
5202 env.src_cpu = busiest->cpu;
5203 env.src_rq = busiest;
5204 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5207 local_irq_save(flags);
5208 double_rq_lock(env.dst_rq, busiest);
5211 * cur_ld_moved - load moved in current iteration
5212 * ld_moved - cumulative load moved across iterations
5214 cur_ld_moved = move_tasks(&env);
5215 ld_moved += cur_ld_moved;
5216 double_rq_unlock(env.dst_rq, busiest);
5217 local_irq_restore(flags);
5220 * some other cpu did the load balance for us.
5222 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5223 resched_cpu(env.dst_cpu);
5225 if (env.flags & LBF_NEED_BREAK) {
5226 env.flags &= ~LBF_NEED_BREAK;
5231 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5232 * us and move them to an alternate dst_cpu in our sched_group
5233 * where they can run. The upper limit on how many times we
5234 * iterate on same src_cpu is dependent on number of cpus in our
5237 * This changes load balance semantics a bit on who can move
5238 * load to a given_cpu. In addition to the given_cpu itself
5239 * (or a ilb_cpu acting on its behalf where given_cpu is
5240 * nohz-idle), we now have balance_cpu in a position to move
5241 * load to given_cpu. In rare situations, this may cause
5242 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5243 * _independently_ and at _same_ time to move some load to
5244 * given_cpu) causing exceess load to be moved to given_cpu.
5245 * This however should not happen so much in practice and
5246 * moreover subsequent load balance cycles should correct the
5247 * excess load moved.
5249 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5251 /* Prevent to re-select dst_cpu via env's cpus */
5252 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5254 env.dst_rq = cpu_rq(env.new_dst_cpu);
5255 env.dst_cpu = env.new_dst_cpu;
5256 env.flags &= ~LBF_DST_PINNED;
5258 env.loop_break = sched_nr_migrate_break;
5261 * Go back to "more_balance" rather than "redo" since we
5262 * need to continue with same src_cpu.
5268 * We failed to reach balance because of affinity.
5271 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5273 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5274 *group_imbalance = 1;
5275 } else if (*group_imbalance)
5276 *group_imbalance = 0;
5279 /* All tasks on this runqueue were pinned by CPU affinity */
5280 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5281 cpumask_clear_cpu(cpu_of(busiest), cpus);
5282 if (!cpumask_empty(cpus)) {
5284 env.loop_break = sched_nr_migrate_break;
5292 schedstat_inc(sd, lb_failed[idle]);
5294 * Increment the failure counter only on periodic balance.
5295 * We do not want newidle balance, which can be very
5296 * frequent, pollute the failure counter causing
5297 * excessive cache_hot migrations and active balances.
5299 if (idle != CPU_NEWLY_IDLE)
5300 sd->nr_balance_failed++;
5302 if (need_active_balance(&env)) {
5303 raw_spin_lock_irqsave(&busiest->lock, flags);
5305 /* don't kick the active_load_balance_cpu_stop,
5306 * if the curr task on busiest cpu can't be
5309 if (!cpumask_test_cpu(this_cpu,
5310 tsk_cpus_allowed(busiest->curr))) {
5311 raw_spin_unlock_irqrestore(&busiest->lock,
5313 env.flags |= LBF_ALL_PINNED;
5314 goto out_one_pinned;
5318 * ->active_balance synchronizes accesses to
5319 * ->active_balance_work. Once set, it's cleared
5320 * only after active load balance is finished.
5322 if (!busiest->active_balance) {
5323 busiest->active_balance = 1;
5324 busiest->push_cpu = this_cpu;
5327 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5329 if (active_balance) {
5330 stop_one_cpu_nowait(cpu_of(busiest),
5331 active_load_balance_cpu_stop, busiest,
5332 &busiest->active_balance_work);
5336 * We've kicked active balancing, reset the failure
5339 sd->nr_balance_failed = sd->cache_nice_tries+1;
5342 sd->nr_balance_failed = 0;
5344 if (likely(!active_balance)) {
5345 /* We were unbalanced, so reset the balancing interval */
5346 sd->balance_interval = sd->min_interval;
5349 * If we've begun active balancing, start to back off. This
5350 * case may not be covered by the all_pinned logic if there
5351 * is only 1 task on the busy runqueue (because we don't call
5354 if (sd->balance_interval < sd->max_interval)
5355 sd->balance_interval *= 2;
5361 schedstat_inc(sd, lb_balanced[idle]);
5363 sd->nr_balance_failed = 0;
5366 /* tune up the balancing interval */
5367 if (((env.flags & LBF_ALL_PINNED) &&
5368 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5369 (sd->balance_interval < sd->max_interval))
5370 sd->balance_interval *= 2;
5378 * idle_balance is called by schedule() if this_cpu is about to become
5379 * idle. Attempts to pull tasks from other CPUs.
5381 void idle_balance(int this_cpu, struct rq *this_rq)
5383 struct sched_domain *sd;
5384 int pulled_task = 0;
5385 unsigned long next_balance = jiffies + HZ;
5388 this_rq->idle_stamp = rq_clock(this_rq);
5390 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5394 * Drop the rq->lock, but keep IRQ/preempt disabled.
5396 raw_spin_unlock(&this_rq->lock);
5398 update_blocked_averages(this_cpu);
5400 for_each_domain(this_cpu, sd) {
5401 unsigned long interval;
5402 int continue_balancing = 1;
5403 u64 t0, domain_cost;
5405 if (!(sd->flags & SD_LOAD_BALANCE))
5408 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5411 if (sd->flags & SD_BALANCE_NEWIDLE) {
5412 t0 = sched_clock_cpu(this_cpu);
5414 /* If we've pulled tasks over stop searching: */
5415 pulled_task = load_balance(this_cpu, this_rq,
5417 &continue_balancing);
5419 domain_cost = sched_clock_cpu(this_cpu) - t0;
5420 if (domain_cost > sd->max_newidle_lb_cost)
5421 sd->max_newidle_lb_cost = domain_cost;
5423 curr_cost += domain_cost;
5426 interval = msecs_to_jiffies(sd->balance_interval);
5427 if (time_after(next_balance, sd->last_balance + interval))
5428 next_balance = sd->last_balance + interval;
5430 this_rq->idle_stamp = 0;
5436 raw_spin_lock(&this_rq->lock);
5438 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5440 * We are going idle. next_balance may be set based on
5441 * a busy processor. So reset next_balance.
5443 this_rq->next_balance = next_balance;
5446 if (curr_cost > this_rq->max_idle_balance_cost)
5447 this_rq->max_idle_balance_cost = curr_cost;
5451 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5452 * running tasks off the busiest CPU onto idle CPUs. It requires at
5453 * least 1 task to be running on each physical CPU where possible, and
5454 * avoids physical / logical imbalances.
5456 static int active_load_balance_cpu_stop(void *data)
5458 struct rq *busiest_rq = data;
5459 int busiest_cpu = cpu_of(busiest_rq);
5460 int target_cpu = busiest_rq->push_cpu;
5461 struct rq *target_rq = cpu_rq(target_cpu);
5462 struct sched_domain *sd;
5464 raw_spin_lock_irq(&busiest_rq->lock);
5466 /* make sure the requested cpu hasn't gone down in the meantime */
5467 if (unlikely(busiest_cpu != smp_processor_id() ||
5468 !busiest_rq->active_balance))
5471 /* Is there any task to move? */
5472 if (busiest_rq->nr_running <= 1)
5476 * This condition is "impossible", if it occurs
5477 * we need to fix it. Originally reported by
5478 * Bjorn Helgaas on a 128-cpu setup.
5480 BUG_ON(busiest_rq == target_rq);
5482 /* move a task from busiest_rq to target_rq */
5483 double_lock_balance(busiest_rq, target_rq);
5485 /* Search for an sd spanning us and the target CPU. */
5487 for_each_domain(target_cpu, sd) {
5488 if ((sd->flags & SD_LOAD_BALANCE) &&
5489 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5494 struct lb_env env = {
5496 .dst_cpu = target_cpu,
5497 .dst_rq = target_rq,
5498 .src_cpu = busiest_rq->cpu,
5499 .src_rq = busiest_rq,
5503 schedstat_inc(sd, alb_count);
5505 if (move_one_task(&env))
5506 schedstat_inc(sd, alb_pushed);
5508 schedstat_inc(sd, alb_failed);
5511 double_unlock_balance(busiest_rq, target_rq);
5513 busiest_rq->active_balance = 0;
5514 raw_spin_unlock_irq(&busiest_rq->lock);
5518 #ifdef CONFIG_NO_HZ_COMMON
5520 * idle load balancing details
5521 * - When one of the busy CPUs notice that there may be an idle rebalancing
5522 * needed, they will kick the idle load balancer, which then does idle
5523 * load balancing for all the idle CPUs.
5526 cpumask_var_t idle_cpus_mask;
5528 unsigned long next_balance; /* in jiffy units */
5529 } nohz ____cacheline_aligned;
5531 static inline int find_new_ilb(int call_cpu)
5533 int ilb = cpumask_first(nohz.idle_cpus_mask);
5535 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5542 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5543 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5544 * CPU (if there is one).
5546 static void nohz_balancer_kick(int cpu)
5550 nohz.next_balance++;
5552 ilb_cpu = find_new_ilb(cpu);
5554 if (ilb_cpu >= nr_cpu_ids)
5557 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5560 * Use smp_send_reschedule() instead of resched_cpu().
5561 * This way we generate a sched IPI on the target cpu which
5562 * is idle. And the softirq performing nohz idle load balance
5563 * will be run before returning from the IPI.
5565 smp_send_reschedule(ilb_cpu);
5569 static inline void nohz_balance_exit_idle(int cpu)
5571 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5572 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5573 atomic_dec(&nohz.nr_cpus);
5574 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5578 static inline void set_cpu_sd_state_busy(void)
5580 struct sched_domain *sd;
5583 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5585 if (!sd || !sd->nohz_idle)
5589 for (; sd; sd = sd->parent)
5590 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5595 void set_cpu_sd_state_idle(void)
5597 struct sched_domain *sd;
5600 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5602 if (!sd || sd->nohz_idle)
5606 for (; sd; sd = sd->parent)
5607 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5613 * This routine will record that the cpu is going idle with tick stopped.
5614 * This info will be used in performing idle load balancing in the future.
5616 void nohz_balance_enter_idle(int cpu)
5619 * If this cpu is going down, then nothing needs to be done.
5621 if (!cpu_active(cpu))
5624 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5627 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5628 atomic_inc(&nohz.nr_cpus);
5629 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5632 static int sched_ilb_notifier(struct notifier_block *nfb,
5633 unsigned long action, void *hcpu)
5635 switch (action & ~CPU_TASKS_FROZEN) {
5637 nohz_balance_exit_idle(smp_processor_id());
5645 static DEFINE_SPINLOCK(balancing);
5648 * Scale the max load_balance interval with the number of CPUs in the system.
5649 * This trades load-balance latency on larger machines for less cross talk.
5651 void update_max_interval(void)
5653 max_load_balance_interval = HZ*num_online_cpus()/10;
5657 * It checks each scheduling domain to see if it is due to be balanced,
5658 * and initiates a balancing operation if so.
5660 * Balancing parameters are set up in init_sched_domains.
5662 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5664 int continue_balancing = 1;
5665 struct rq *rq = cpu_rq(cpu);
5666 unsigned long interval;
5667 struct sched_domain *sd;
5668 /* Earliest time when we have to do rebalance again */
5669 unsigned long next_balance = jiffies + 60*HZ;
5670 int update_next_balance = 0;
5671 int need_serialize, need_decay = 0;
5674 update_blocked_averages(cpu);
5677 for_each_domain(cpu, sd) {
5679 * Decay the newidle max times here because this is a regular
5680 * visit to all the domains. Decay ~1% per second.
5682 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5683 sd->max_newidle_lb_cost =
5684 (sd->max_newidle_lb_cost * 253) / 256;
5685 sd->next_decay_max_lb_cost = jiffies + HZ;
5688 max_cost += sd->max_newidle_lb_cost;
5690 if (!(sd->flags & SD_LOAD_BALANCE))
5694 * Stop the load balance at this level. There is another
5695 * CPU in our sched group which is doing load balancing more
5698 if (!continue_balancing) {
5704 interval = sd->balance_interval;
5705 if (idle != CPU_IDLE)
5706 interval *= sd->busy_factor;
5708 /* scale ms to jiffies */
5709 interval = msecs_to_jiffies(interval);
5710 interval = clamp(interval, 1UL, max_load_balance_interval);
5712 need_serialize = sd->flags & SD_SERIALIZE;
5714 if (need_serialize) {
5715 if (!spin_trylock(&balancing))
5719 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5720 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5722 * The LBF_DST_PINNED logic could have changed
5723 * env->dst_cpu, so we can't know our idle
5724 * state even if we migrated tasks. Update it.
5726 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5728 sd->last_balance = jiffies;
5731 spin_unlock(&balancing);
5733 if (time_after(next_balance, sd->last_balance + interval)) {
5734 next_balance = sd->last_balance + interval;
5735 update_next_balance = 1;
5740 * Ensure the rq-wide value also decays but keep it at a
5741 * reasonable floor to avoid funnies with rq->avg_idle.
5743 rq->max_idle_balance_cost =
5744 max((u64)sysctl_sched_migration_cost, max_cost);
5749 * next_balance will be updated only when there is a need.
5750 * When the cpu is attached to null domain for ex, it will not be
5753 if (likely(update_next_balance))
5754 rq->next_balance = next_balance;
5757 #ifdef CONFIG_NO_HZ_COMMON
5759 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5760 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5762 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5764 struct rq *this_rq = cpu_rq(this_cpu);
5768 if (idle != CPU_IDLE ||
5769 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5772 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5773 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5777 * If this cpu gets work to do, stop the load balancing
5778 * work being done for other cpus. Next load
5779 * balancing owner will pick it up.
5784 rq = cpu_rq(balance_cpu);
5786 raw_spin_lock_irq(&rq->lock);
5787 update_rq_clock(rq);
5788 update_idle_cpu_load(rq);
5789 raw_spin_unlock_irq(&rq->lock);
5791 rebalance_domains(balance_cpu, CPU_IDLE);
5793 if (time_after(this_rq->next_balance, rq->next_balance))
5794 this_rq->next_balance = rq->next_balance;
5796 nohz.next_balance = this_rq->next_balance;
5798 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5802 * Current heuristic for kicking the idle load balancer in the presence
5803 * of an idle cpu is the system.
5804 * - This rq has more than one task.
5805 * - At any scheduler domain level, this cpu's scheduler group has multiple
5806 * busy cpu's exceeding the group's power.
5807 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5808 * domain span are idle.
5810 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5812 unsigned long now = jiffies;
5813 struct sched_domain *sd;
5815 if (unlikely(idle_cpu(cpu)))
5819 * We may be recently in ticked or tickless idle mode. At the first
5820 * busy tick after returning from idle, we will update the busy stats.
5822 set_cpu_sd_state_busy();
5823 nohz_balance_exit_idle(cpu);
5826 * None are in tickless mode and hence no need for NOHZ idle load
5829 if (likely(!atomic_read(&nohz.nr_cpus)))
5832 if (time_before(now, nohz.next_balance))
5835 if (rq->nr_running >= 2)
5839 for_each_domain(cpu, sd) {
5840 struct sched_group *sg = sd->groups;
5841 struct sched_group_power *sgp = sg->sgp;
5842 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5844 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5845 goto need_kick_unlock;
5847 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5848 && (cpumask_first_and(nohz.idle_cpus_mask,
5849 sched_domain_span(sd)) < cpu))
5850 goto need_kick_unlock;
5852 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5864 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5868 * run_rebalance_domains is triggered when needed from the scheduler tick.
5869 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5871 static void run_rebalance_domains(struct softirq_action *h)
5873 int this_cpu = smp_processor_id();
5874 struct rq *this_rq = cpu_rq(this_cpu);
5875 enum cpu_idle_type idle = this_rq->idle_balance ?
5876 CPU_IDLE : CPU_NOT_IDLE;
5878 rebalance_domains(this_cpu, idle);
5881 * If this cpu has a pending nohz_balance_kick, then do the
5882 * balancing on behalf of the other idle cpus whose ticks are
5885 nohz_idle_balance(this_cpu, idle);
5888 static inline int on_null_domain(int cpu)
5890 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5894 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5896 void trigger_load_balance(struct rq *rq, int cpu)
5898 /* Don't need to rebalance while attached to NULL domain */
5899 if (time_after_eq(jiffies, rq->next_balance) &&
5900 likely(!on_null_domain(cpu)))
5901 raise_softirq(SCHED_SOFTIRQ);
5902 #ifdef CONFIG_NO_HZ_COMMON
5903 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5904 nohz_balancer_kick(cpu);
5908 static void rq_online_fair(struct rq *rq)
5913 static void rq_offline_fair(struct rq *rq)
5917 /* Ensure any throttled groups are reachable by pick_next_task */
5918 unthrottle_offline_cfs_rqs(rq);
5921 #endif /* CONFIG_SMP */
5924 * scheduler tick hitting a task of our scheduling class:
5926 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5928 struct cfs_rq *cfs_rq;
5929 struct sched_entity *se = &curr->se;
5931 for_each_sched_entity(se) {
5932 cfs_rq = cfs_rq_of(se);
5933 entity_tick(cfs_rq, se, queued);
5936 if (numabalancing_enabled)
5937 task_tick_numa(rq, curr);
5939 update_rq_runnable_avg(rq, 1);
5943 * called on fork with the child task as argument from the parent's context
5944 * - child not yet on the tasklist
5945 * - preemption disabled
5947 static void task_fork_fair(struct task_struct *p)
5949 struct cfs_rq *cfs_rq;
5950 struct sched_entity *se = &p->se, *curr;
5951 int this_cpu = smp_processor_id();
5952 struct rq *rq = this_rq();
5953 unsigned long flags;
5955 raw_spin_lock_irqsave(&rq->lock, flags);
5957 update_rq_clock(rq);
5959 cfs_rq = task_cfs_rq(current);
5960 curr = cfs_rq->curr;
5963 * Not only the cpu but also the task_group of the parent might have
5964 * been changed after parent->se.parent,cfs_rq were copied to
5965 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
5966 * of child point to valid ones.
5969 __set_task_cpu(p, this_cpu);
5972 update_curr(cfs_rq);
5975 se->vruntime = curr->vruntime;
5976 place_entity(cfs_rq, se, 1);
5978 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5980 * Upon rescheduling, sched_class::put_prev_task() will place
5981 * 'current' within the tree based on its new key value.
5983 swap(curr->vruntime, se->vruntime);
5984 resched_task(rq->curr);
5987 se->vruntime -= cfs_rq->min_vruntime;
5989 raw_spin_unlock_irqrestore(&rq->lock, flags);
5993 * Priority of the task has changed. Check to see if we preempt
5997 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6003 * Reschedule if we are currently running on this runqueue and
6004 * our priority decreased, or if we are not currently running on
6005 * this runqueue and our priority is higher than the current's
6007 if (rq->curr == p) {
6008 if (p->prio > oldprio)
6009 resched_task(rq->curr);
6011 check_preempt_curr(rq, p, 0);
6014 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6016 struct sched_entity *se = &p->se;
6017 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6020 * Ensure the task's vruntime is normalized, so that when its
6021 * switched back to the fair class the enqueue_entity(.flags=0) will
6022 * do the right thing.
6024 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6025 * have normalized the vruntime, if it was !on_rq, then only when
6026 * the task is sleeping will it still have non-normalized vruntime.
6028 if (!se->on_rq && p->state != TASK_RUNNING) {
6030 * Fix up our vruntime so that the current sleep doesn't
6031 * cause 'unlimited' sleep bonus.
6033 place_entity(cfs_rq, se, 0);
6034 se->vruntime -= cfs_rq->min_vruntime;
6039 * Remove our load from contribution when we leave sched_fair
6040 * and ensure we don't carry in an old decay_count if we
6043 if (se->avg.decay_count) {
6044 __synchronize_entity_decay(se);
6045 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6051 * We switched to the sched_fair class.
6053 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6059 * We were most likely switched from sched_rt, so
6060 * kick off the schedule if running, otherwise just see
6061 * if we can still preempt the current task.
6064 resched_task(rq->curr);
6066 check_preempt_curr(rq, p, 0);
6069 /* Account for a task changing its policy or group.
6071 * This routine is mostly called to set cfs_rq->curr field when a task
6072 * migrates between groups/classes.
6074 static void set_curr_task_fair(struct rq *rq)
6076 struct sched_entity *se = &rq->curr->se;
6078 for_each_sched_entity(se) {
6079 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6081 set_next_entity(cfs_rq, se);
6082 /* ensure bandwidth has been allocated on our new cfs_rq */
6083 account_cfs_rq_runtime(cfs_rq, 0);
6087 void init_cfs_rq(struct cfs_rq *cfs_rq)
6089 cfs_rq->tasks_timeline = RB_ROOT;
6090 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6091 #ifndef CONFIG_64BIT
6092 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6095 atomic64_set(&cfs_rq->decay_counter, 1);
6096 atomic_long_set(&cfs_rq->removed_load, 0);
6100 #ifdef CONFIG_FAIR_GROUP_SCHED
6101 static void task_move_group_fair(struct task_struct *p, int on_rq)
6103 struct cfs_rq *cfs_rq;
6105 * If the task was not on the rq at the time of this cgroup movement
6106 * it must have been asleep, sleeping tasks keep their ->vruntime
6107 * absolute on their old rq until wakeup (needed for the fair sleeper
6108 * bonus in place_entity()).
6110 * If it was on the rq, we've just 'preempted' it, which does convert
6111 * ->vruntime to a relative base.
6113 * Make sure both cases convert their relative position when migrating
6114 * to another cgroup's rq. This does somewhat interfere with the
6115 * fair sleeper stuff for the first placement, but who cares.
6118 * When !on_rq, vruntime of the task has usually NOT been normalized.
6119 * But there are some cases where it has already been normalized:
6121 * - Moving a forked child which is waiting for being woken up by
6122 * wake_up_new_task().
6123 * - Moving a task which has been woken up by try_to_wake_up() and
6124 * waiting for actually being woken up by sched_ttwu_pending().
6126 * To prevent boost or penalty in the new cfs_rq caused by delta
6127 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6129 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6133 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6134 set_task_rq(p, task_cpu(p));
6136 cfs_rq = cfs_rq_of(&p->se);
6137 p->se.vruntime += cfs_rq->min_vruntime;
6140 * migrate_task_rq_fair() will have removed our previous
6141 * contribution, but we must synchronize for ongoing future
6144 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6145 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6150 void free_fair_sched_group(struct task_group *tg)
6154 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6156 for_each_possible_cpu(i) {
6158 kfree(tg->cfs_rq[i]);
6167 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6169 struct cfs_rq *cfs_rq;
6170 struct sched_entity *se;
6173 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6176 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6180 tg->shares = NICE_0_LOAD;
6182 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6184 for_each_possible_cpu(i) {
6185 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6186 GFP_KERNEL, cpu_to_node(i));
6190 se = kzalloc_node(sizeof(struct sched_entity),
6191 GFP_KERNEL, cpu_to_node(i));
6195 init_cfs_rq(cfs_rq);
6196 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6207 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6209 struct rq *rq = cpu_rq(cpu);
6210 unsigned long flags;
6213 * Only empty task groups can be destroyed; so we can speculatively
6214 * check on_list without danger of it being re-added.
6216 if (!tg->cfs_rq[cpu]->on_list)
6219 raw_spin_lock_irqsave(&rq->lock, flags);
6220 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6221 raw_spin_unlock_irqrestore(&rq->lock, flags);
6224 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6225 struct sched_entity *se, int cpu,
6226 struct sched_entity *parent)
6228 struct rq *rq = cpu_rq(cpu);
6232 init_cfs_rq_runtime(cfs_rq);
6234 tg->cfs_rq[cpu] = cfs_rq;
6237 /* se could be NULL for root_task_group */
6242 se->cfs_rq = &rq->cfs;
6244 se->cfs_rq = parent->my_q;
6247 update_load_set(&se->load, 0);
6248 se->parent = parent;
6251 static DEFINE_MUTEX(shares_mutex);
6253 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6256 unsigned long flags;
6259 * We can't change the weight of the root cgroup.
6264 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6266 mutex_lock(&shares_mutex);
6267 if (tg->shares == shares)
6270 tg->shares = shares;
6271 for_each_possible_cpu(i) {
6272 struct rq *rq = cpu_rq(i);
6273 struct sched_entity *se;
6276 /* Propagate contribution to hierarchy */
6277 raw_spin_lock_irqsave(&rq->lock, flags);
6279 /* Possible calls to update_curr() need rq clock */
6280 update_rq_clock(rq);
6281 for_each_sched_entity(se)
6282 update_cfs_shares(group_cfs_rq(se));
6283 raw_spin_unlock_irqrestore(&rq->lock, flags);
6287 mutex_unlock(&shares_mutex);
6290 #else /* CONFIG_FAIR_GROUP_SCHED */
6292 void free_fair_sched_group(struct task_group *tg) { }
6294 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6299 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6301 #endif /* CONFIG_FAIR_GROUP_SCHED */
6304 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6306 struct sched_entity *se = &task->se;
6307 unsigned int rr_interval = 0;
6310 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6313 if (rq->cfs.load.weight)
6314 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6320 * All the scheduling class methods:
6322 const struct sched_class fair_sched_class = {
6323 .next = &idle_sched_class,
6324 .enqueue_task = enqueue_task_fair,
6325 .dequeue_task = dequeue_task_fair,
6326 .yield_task = yield_task_fair,
6327 .yield_to_task = yield_to_task_fair,
6329 .check_preempt_curr = check_preempt_wakeup,
6331 .pick_next_task = pick_next_task_fair,
6332 .put_prev_task = put_prev_task_fair,
6335 .select_task_rq = select_task_rq_fair,
6336 .migrate_task_rq = migrate_task_rq_fair,
6338 .rq_online = rq_online_fair,
6339 .rq_offline = rq_offline_fair,
6341 .task_waking = task_waking_fair,
6344 .set_curr_task = set_curr_task_fair,
6345 .task_tick = task_tick_fair,
6346 .task_fork = task_fork_fair,
6348 .prio_changed = prio_changed_fair,
6349 .switched_from = switched_from_fair,
6350 .switched_to = switched_to_fair,
6352 .get_rr_interval = get_rr_interval_fair,
6354 #ifdef CONFIG_FAIR_GROUP_SCHED
6355 .task_move_group = task_move_group_fair,
6359 #ifdef CONFIG_SCHED_DEBUG
6360 void print_cfs_stats(struct seq_file *m, int cpu)
6362 struct cfs_rq *cfs_rq;
6365 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6366 print_cfs_rq(m, cpu, cfs_rq);
6371 __init void init_sched_fair_class(void)
6374 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6376 #ifdef CONFIG_NO_HZ_COMMON
6377 nohz.next_balance = jiffies;
6378 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6379 cpu_notifier(sched_ilb_notifier, 0);