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 unsigned long task_h_load(struct task_struct *p);
686 static inline void __update_task_entity_contrib(struct sched_entity *se);
688 /* Give new task start runnable values to heavy its load in infant time */
689 void init_task_runnable_average(struct task_struct *p)
693 p->se.avg.decay_count = 0;
694 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
695 p->se.avg.runnable_avg_sum = slice;
696 p->se.avg.runnable_avg_period = slice;
697 __update_task_entity_contrib(&p->se);
700 void init_task_runnable_average(struct task_struct *p)
706 * Update the current task's runtime statistics. Skip current tasks that
707 * are not in our scheduling class.
710 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
711 unsigned long delta_exec)
713 unsigned long delta_exec_weighted;
715 schedstat_set(curr->statistics.exec_max,
716 max((u64)delta_exec, curr->statistics.exec_max));
718 curr->sum_exec_runtime += delta_exec;
719 schedstat_add(cfs_rq, exec_clock, delta_exec);
720 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
722 curr->vruntime += delta_exec_weighted;
723 update_min_vruntime(cfs_rq);
726 static void update_curr(struct cfs_rq *cfs_rq)
728 struct sched_entity *curr = cfs_rq->curr;
729 u64 now = rq_clock_task(rq_of(cfs_rq));
730 unsigned long delta_exec;
736 * Get the amount of time the current task was running
737 * since the last time we changed load (this cannot
738 * overflow on 32 bits):
740 delta_exec = (unsigned long)(now - curr->exec_start);
744 __update_curr(cfs_rq, curr, delta_exec);
745 curr->exec_start = now;
747 if (entity_is_task(curr)) {
748 struct task_struct *curtask = task_of(curr);
750 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
751 cpuacct_charge(curtask, delta_exec);
752 account_group_exec_runtime(curtask, delta_exec);
755 account_cfs_rq_runtime(cfs_rq, delta_exec);
759 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
761 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
765 * Task is being enqueued - update stats:
767 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
770 * Are we enqueueing a waiting task? (for current tasks
771 * a dequeue/enqueue event is a NOP)
773 if (se != cfs_rq->curr)
774 update_stats_wait_start(cfs_rq, se);
778 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
780 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
781 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
782 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
783 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
784 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
785 #ifdef CONFIG_SCHEDSTATS
786 if (entity_is_task(se)) {
787 trace_sched_stat_wait(task_of(se),
788 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
791 schedstat_set(se->statistics.wait_start, 0);
795 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
798 * Mark the end of the wait period if dequeueing a
801 if (se != cfs_rq->curr)
802 update_stats_wait_end(cfs_rq, se);
806 * We are picking a new current task - update its stats:
809 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
812 * We are starting a new run period:
814 se->exec_start = rq_clock_task(rq_of(cfs_rq));
817 /**************************************************
818 * Scheduling class queueing methods:
821 #ifdef CONFIG_NUMA_BALANCING
823 * Approximate time to scan a full NUMA task in ms. The task scan period is
824 * calculated based on the tasks virtual memory size and
825 * numa_balancing_scan_size.
827 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
828 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
829 unsigned int sysctl_numa_balancing_scan_period_reset = 60000;
831 /* Portion of address space to scan in MB */
832 unsigned int sysctl_numa_balancing_scan_size = 256;
834 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
835 unsigned int sysctl_numa_balancing_scan_delay = 1000;
837 static unsigned int task_nr_scan_windows(struct task_struct *p)
839 unsigned long rss = 0;
840 unsigned long nr_scan_pages;
843 * Calculations based on RSS as non-present and empty pages are skipped
844 * by the PTE scanner and NUMA hinting faults should be trapped based
847 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
848 rss = get_mm_rss(p->mm);
852 rss = round_up(rss, nr_scan_pages);
853 return rss / nr_scan_pages;
856 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
857 #define MAX_SCAN_WINDOW 2560
859 static unsigned int task_scan_min(struct task_struct *p)
861 unsigned int scan, floor;
862 unsigned int windows = 1;
864 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
865 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
866 floor = 1000 / windows;
868 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
869 return max_t(unsigned int, floor, scan);
872 static unsigned int task_scan_max(struct task_struct *p)
874 unsigned int smin = task_scan_min(p);
877 /* Watch for min being lower than max due to floor calculations */
878 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
879 return max(smin, smax);
883 * Once a preferred node is selected the scheduler balancer will prefer moving
884 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
885 * scans. This will give the process the chance to accumulate more faults on
886 * the preferred node but still allow the scheduler to move the task again if
887 * the nodes CPUs are overloaded.
889 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 4;
891 static inline int task_faults_idx(int nid, int priv)
893 return 2 * nid + priv;
896 static inline unsigned long task_faults(struct task_struct *p, int nid)
901 return p->numa_faults[task_faults_idx(nid, 0)] +
902 p->numa_faults[task_faults_idx(nid, 1)];
905 static unsigned long weighted_cpuload(const int cpu);
906 static unsigned long source_load(int cpu, int type);
907 static unsigned long target_load(int cpu, int type);
908 static unsigned long power_of(int cpu);
909 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
911 /* Cached statistics for all CPUs within a node */
913 unsigned long nr_running;
916 /* Total compute capacity of CPUs on a node */
919 /* Approximate capacity in terms of runnable tasks on a node */
920 unsigned long capacity;
925 * XXX borrowed from update_sg_lb_stats
927 static void update_numa_stats(struct numa_stats *ns, int nid)
931 memset(ns, 0, sizeof(*ns));
932 for_each_cpu(cpu, cpumask_of_node(nid)) {
933 struct rq *rq = cpu_rq(cpu);
935 ns->nr_running += rq->nr_running;
936 ns->load += weighted_cpuload(cpu);
937 ns->power += power_of(cpu);
940 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
941 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
942 ns->has_capacity = (ns->nr_running < ns->capacity);
945 struct task_numa_env {
946 struct task_struct *p;
948 int src_cpu, src_nid;
949 int dst_cpu, dst_nid;
951 struct numa_stats src_stats, dst_stats;
953 int imbalance_pct, idx;
955 struct task_struct *best_task;
960 static void task_numa_assign(struct task_numa_env *env,
961 struct task_struct *p, long imp)
964 put_task_struct(env->best_task);
970 env->best_cpu = env->dst_cpu;
974 * This checks if the overall compute and NUMA accesses of the system would
975 * be improved if the source tasks was migrated to the target dst_cpu taking
976 * into account that it might be best if task running on the dst_cpu should
977 * be exchanged with the source task
979 static void task_numa_compare(struct task_numa_env *env, long imp)
981 struct rq *src_rq = cpu_rq(env->src_cpu);
982 struct rq *dst_rq = cpu_rq(env->dst_cpu);
983 struct task_struct *cur;
984 long dst_load, src_load;
988 cur = ACCESS_ONCE(dst_rq->curr);
989 if (cur->pid == 0) /* idle */
993 * "imp" is the fault differential for the source task between the
994 * source and destination node. Calculate the total differential for
995 * the source task and potential destination task. The more negative
996 * the value is, the more rmeote accesses that would be expected to
997 * be incurred if the tasks were swapped.
1000 /* Skip this swap candidate if cannot move to the source cpu */
1001 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1004 imp += task_faults(cur, env->src_nid) -
1005 task_faults(cur, env->dst_nid);
1008 if (imp < env->best_imp)
1012 /* Is there capacity at our destination? */
1013 if (env->src_stats.has_capacity &&
1014 !env->dst_stats.has_capacity)
1020 /* Balance doesn't matter much if we're running a task per cpu */
1021 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1025 * In the overloaded case, try and keep the load balanced.
1028 dst_load = env->dst_stats.load;
1029 src_load = env->src_stats.load;
1031 /* XXX missing power terms */
1032 load = task_h_load(env->p);
1037 load = task_h_load(cur);
1042 /* make src_load the smaller */
1043 if (dst_load < src_load)
1044 swap(dst_load, src_load);
1046 if (src_load * env->imbalance_pct < dst_load * 100)
1050 task_numa_assign(env, cur, imp);
1055 static void task_numa_find_cpu(struct task_numa_env *env, long imp)
1059 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1060 /* Skip this CPU if the source task cannot migrate */
1061 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1065 task_numa_compare(env, imp);
1069 static int task_numa_migrate(struct task_struct *p)
1071 struct task_numa_env env = {
1074 .src_cpu = task_cpu(p),
1075 .src_nid = cpu_to_node(task_cpu(p)),
1077 .imbalance_pct = 112,
1083 struct sched_domain *sd;
1084 unsigned long faults;
1089 * Pick the lowest SD_NUMA domain, as that would have the smallest
1090 * imbalance and would be the first to start moving tasks about.
1092 * And we want to avoid any moving of tasks about, as that would create
1093 * random movement of tasks -- counter the numa conditions we're trying
1097 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1098 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1101 faults = task_faults(p, env.src_nid);
1102 update_numa_stats(&env.src_stats, env.src_nid);
1103 env.dst_nid = p->numa_preferred_nid;
1104 imp = task_faults(env.p, env.dst_nid) - faults;
1105 update_numa_stats(&env.dst_stats, env.dst_nid);
1108 * If the preferred nid has capacity then use it. Otherwise find an
1109 * alternative node with relatively better statistics.
1111 if (env.dst_stats.has_capacity) {
1112 task_numa_find_cpu(&env, imp);
1114 for_each_online_node(nid) {
1115 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1118 /* Only consider nodes that recorded more faults */
1119 imp = task_faults(env.p, nid) - faults;
1124 update_numa_stats(&env.dst_stats, env.dst_nid);
1125 task_numa_find_cpu(&env, imp);
1129 /* No better CPU than the current one was found. */
1130 if (env.best_cpu == -1)
1133 if (env.best_task == NULL) {
1134 int ret = migrate_task_to(p, env.best_cpu);
1138 ret = migrate_swap(p, env.best_task);
1139 put_task_struct(env.best_task);
1143 /* Attempt to migrate a task to a CPU on the preferred node. */
1144 static void numa_migrate_preferred(struct task_struct *p)
1146 /* Success if task is already running on preferred CPU */
1147 p->numa_migrate_retry = 0;
1148 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid) {
1150 * If migration is temporarily disabled due to a task migration
1151 * then re-enable it now as the task is running on its
1152 * preferred node and memory should migrate locally
1154 if (!p->numa_migrate_seq)
1155 p->numa_migrate_seq++;
1159 /* This task has no NUMA fault statistics yet */
1160 if (unlikely(p->numa_preferred_nid == -1))
1163 /* Otherwise, try migrate to a CPU on the preferred node */
1164 if (task_numa_migrate(p) != 0)
1165 p->numa_migrate_retry = jiffies + HZ*5;
1168 static void task_numa_placement(struct task_struct *p)
1170 int seq, nid, max_nid = -1;
1171 unsigned long max_faults = 0;
1173 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1174 if (p->numa_scan_seq == seq)
1176 p->numa_scan_seq = seq;
1177 p->numa_migrate_seq++;
1178 p->numa_scan_period_max = task_scan_max(p);
1180 /* Find the node with the highest number of faults */
1181 for_each_online_node(nid) {
1182 unsigned long faults = 0;
1185 for (priv = 0; priv < 2; priv++) {
1186 i = task_faults_idx(nid, priv);
1188 /* Decay existing window, copy faults since last scan */
1189 p->numa_faults[i] >>= 1;
1190 p->numa_faults[i] += p->numa_faults_buffer[i];
1191 p->numa_faults_buffer[i] = 0;
1193 faults += p->numa_faults[i];
1196 if (faults > max_faults) {
1197 max_faults = faults;
1202 /* Preferred node as the node with the most faults */
1203 if (max_faults && max_nid != p->numa_preferred_nid) {
1204 /* Update the preferred nid and migrate task if possible */
1205 p->numa_preferred_nid = max_nid;
1206 p->numa_migrate_seq = 1;
1207 numa_migrate_preferred(p);
1212 * Got a PROT_NONE fault for a page on @node.
1214 void task_numa_fault(int last_nidpid, int node, int pages, bool migrated)
1216 struct task_struct *p = current;
1219 if (!numabalancing_enabled)
1222 /* for example, ksmd faulting in a user's mm */
1227 * First accesses are treated as private, otherwise consider accesses
1228 * to be private if the accessing pid has not changed
1230 if (!nidpid_pid_unset(last_nidpid))
1231 priv = ((p->pid & LAST__PID_MASK) == nidpid_to_pid(last_nidpid));
1235 /* Allocate buffer to track faults on a per-node basis */
1236 if (unlikely(!p->numa_faults)) {
1237 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1239 /* numa_faults and numa_faults_buffer share the allocation */
1240 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1241 if (!p->numa_faults)
1244 BUG_ON(p->numa_faults_buffer);
1245 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1249 * If pages are properly placed (did not migrate) then scan slower.
1250 * This is reset periodically in case of phase changes
1253 /* Initialise if necessary */
1254 if (!p->numa_scan_period_max)
1255 p->numa_scan_period_max = task_scan_max(p);
1257 p->numa_scan_period = min(p->numa_scan_period_max,
1258 p->numa_scan_period + 10);
1261 task_numa_placement(p);
1263 /* Retry task to preferred node migration if it previously failed */
1264 if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry))
1265 numa_migrate_preferred(p);
1267 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1270 static void reset_ptenuma_scan(struct task_struct *p)
1272 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1273 p->mm->numa_scan_offset = 0;
1277 * The expensive part of numa migration is done from task_work context.
1278 * Triggered from task_tick_numa().
1280 void task_numa_work(struct callback_head *work)
1282 unsigned long migrate, next_scan, now = jiffies;
1283 struct task_struct *p = current;
1284 struct mm_struct *mm = p->mm;
1285 struct vm_area_struct *vma;
1286 unsigned long start, end;
1287 unsigned long nr_pte_updates = 0;
1290 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1292 work->next = work; /* protect against double add */
1294 * Who cares about NUMA placement when they're dying.
1296 * NOTE: make sure not to dereference p->mm before this check,
1297 * exit_task_work() happens _after_ exit_mm() so we could be called
1298 * without p->mm even though we still had it when we enqueued this
1301 if (p->flags & PF_EXITING)
1304 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1305 mm->numa_next_scan = now +
1306 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1307 mm->numa_next_reset = now +
1308 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1312 * Reset the scan period if enough time has gone by. Objective is that
1313 * scanning will be reduced if pages are properly placed. As tasks
1314 * can enter different phases this needs to be re-examined. Lacking
1315 * proper tracking of reference behaviour, this blunt hammer is used.
1317 migrate = mm->numa_next_reset;
1318 if (time_after(now, migrate)) {
1319 p->numa_scan_period = task_scan_min(p);
1320 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1321 xchg(&mm->numa_next_reset, next_scan);
1325 * Enforce maximal scan/migration frequency..
1327 migrate = mm->numa_next_scan;
1328 if (time_before(now, migrate))
1331 if (p->numa_scan_period == 0) {
1332 p->numa_scan_period_max = task_scan_max(p);
1333 p->numa_scan_period = task_scan_min(p);
1336 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1337 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1341 * Delay this task enough that another task of this mm will likely win
1342 * the next time around.
1344 p->node_stamp += 2 * TICK_NSEC;
1346 start = mm->numa_scan_offset;
1347 pages = sysctl_numa_balancing_scan_size;
1348 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1352 down_read(&mm->mmap_sem);
1353 vma = find_vma(mm, start);
1355 reset_ptenuma_scan(p);
1359 for (; vma; vma = vma->vm_next) {
1360 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1364 * Shared library pages mapped by multiple processes are not
1365 * migrated as it is expected they are cache replicated. Avoid
1366 * hinting faults in read-only file-backed mappings or the vdso
1367 * as migrating the pages will be of marginal benefit.
1370 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1374 start = max(start, vma->vm_start);
1375 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1376 end = min(end, vma->vm_end);
1377 nr_pte_updates += change_prot_numa(vma, start, end);
1380 * Scan sysctl_numa_balancing_scan_size but ensure that
1381 * at least one PTE is updated so that unused virtual
1382 * address space is quickly skipped.
1385 pages -= (end - start) >> PAGE_SHIFT;
1390 } while (end != vma->vm_end);
1395 * If the whole process was scanned without updates then no NUMA
1396 * hinting faults are being recorded and scan rate should be lower.
1398 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1399 p->numa_scan_period = min(p->numa_scan_period_max,
1400 p->numa_scan_period << 1);
1402 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1403 mm->numa_next_scan = next_scan;
1407 * It is possible to reach the end of the VMA list but the last few
1408 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1409 * would find the !migratable VMA on the next scan but not reset the
1410 * scanner to the start so check it now.
1413 mm->numa_scan_offset = start;
1415 reset_ptenuma_scan(p);
1416 up_read(&mm->mmap_sem);
1420 * Drive the periodic memory faults..
1422 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1424 struct callback_head *work = &curr->numa_work;
1428 * We don't care about NUMA placement if we don't have memory.
1430 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1434 * Using runtime rather than walltime has the dual advantage that
1435 * we (mostly) drive the selection from busy threads and that the
1436 * task needs to have done some actual work before we bother with
1439 now = curr->se.sum_exec_runtime;
1440 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1442 if (now - curr->node_stamp > period) {
1443 if (!curr->node_stamp)
1444 curr->numa_scan_period = task_scan_min(curr);
1445 curr->node_stamp += period;
1447 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1448 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1449 task_work_add(curr, work, true);
1454 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1457 #endif /* CONFIG_NUMA_BALANCING */
1460 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1462 update_load_add(&cfs_rq->load, se->load.weight);
1463 if (!parent_entity(se))
1464 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1466 if (entity_is_task(se))
1467 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1469 cfs_rq->nr_running++;
1473 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1475 update_load_sub(&cfs_rq->load, se->load.weight);
1476 if (!parent_entity(se))
1477 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1478 if (entity_is_task(se))
1479 list_del_init(&se->group_node);
1480 cfs_rq->nr_running--;
1483 #ifdef CONFIG_FAIR_GROUP_SCHED
1485 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1490 * Use this CPU's actual weight instead of the last load_contribution
1491 * to gain a more accurate current total weight. See
1492 * update_cfs_rq_load_contribution().
1494 tg_weight = atomic_long_read(&tg->load_avg);
1495 tg_weight -= cfs_rq->tg_load_contrib;
1496 tg_weight += cfs_rq->load.weight;
1501 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1503 long tg_weight, load, shares;
1505 tg_weight = calc_tg_weight(tg, cfs_rq);
1506 load = cfs_rq->load.weight;
1508 shares = (tg->shares * load);
1510 shares /= tg_weight;
1512 if (shares < MIN_SHARES)
1513 shares = MIN_SHARES;
1514 if (shares > tg->shares)
1515 shares = tg->shares;
1519 # else /* CONFIG_SMP */
1520 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1524 # endif /* CONFIG_SMP */
1525 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1526 unsigned long weight)
1529 /* commit outstanding execution time */
1530 if (cfs_rq->curr == se)
1531 update_curr(cfs_rq);
1532 account_entity_dequeue(cfs_rq, se);
1535 update_load_set(&se->load, weight);
1538 account_entity_enqueue(cfs_rq, se);
1541 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1543 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1545 struct task_group *tg;
1546 struct sched_entity *se;
1550 se = tg->se[cpu_of(rq_of(cfs_rq))];
1551 if (!se || throttled_hierarchy(cfs_rq))
1554 if (likely(se->load.weight == tg->shares))
1557 shares = calc_cfs_shares(cfs_rq, tg);
1559 reweight_entity(cfs_rq_of(se), se, shares);
1561 #else /* CONFIG_FAIR_GROUP_SCHED */
1562 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1565 #endif /* CONFIG_FAIR_GROUP_SCHED */
1569 * We choose a half-life close to 1 scheduling period.
1570 * Note: The tables below are dependent on this value.
1572 #define LOAD_AVG_PERIOD 32
1573 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1574 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1576 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1577 static const u32 runnable_avg_yN_inv[] = {
1578 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1579 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1580 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1581 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1582 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1583 0x85aac367, 0x82cd8698,
1587 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1588 * over-estimates when re-combining.
1590 static const u32 runnable_avg_yN_sum[] = {
1591 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1592 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1593 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1598 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1600 static __always_inline u64 decay_load(u64 val, u64 n)
1602 unsigned int local_n;
1606 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1609 /* after bounds checking we can collapse to 32-bit */
1613 * As y^PERIOD = 1/2, we can combine
1614 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1615 * With a look-up table which covers k^n (n<PERIOD)
1617 * To achieve constant time decay_load.
1619 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1620 val >>= local_n / LOAD_AVG_PERIOD;
1621 local_n %= LOAD_AVG_PERIOD;
1624 val *= runnable_avg_yN_inv[local_n];
1625 /* We don't use SRR here since we always want to round down. */
1630 * For updates fully spanning n periods, the contribution to runnable
1631 * average will be: \Sum 1024*y^n
1633 * We can compute this reasonably efficiently by combining:
1634 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1636 static u32 __compute_runnable_contrib(u64 n)
1640 if (likely(n <= LOAD_AVG_PERIOD))
1641 return runnable_avg_yN_sum[n];
1642 else if (unlikely(n >= LOAD_AVG_MAX_N))
1643 return LOAD_AVG_MAX;
1645 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1647 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1648 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1650 n -= LOAD_AVG_PERIOD;
1651 } while (n > LOAD_AVG_PERIOD);
1653 contrib = decay_load(contrib, n);
1654 return contrib + runnable_avg_yN_sum[n];
1658 * We can represent the historical contribution to runnable average as the
1659 * coefficients of a geometric series. To do this we sub-divide our runnable
1660 * history into segments of approximately 1ms (1024us); label the segment that
1661 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1663 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1665 * (now) (~1ms ago) (~2ms ago)
1667 * Let u_i denote the fraction of p_i that the entity was runnable.
1669 * We then designate the fractions u_i as our co-efficients, yielding the
1670 * following representation of historical load:
1671 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1673 * We choose y based on the with of a reasonably scheduling period, fixing:
1676 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1677 * approximately half as much as the contribution to load within the last ms
1680 * When a period "rolls over" and we have new u_0`, multiplying the previous
1681 * sum again by y is sufficient to update:
1682 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1683 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1685 static __always_inline int __update_entity_runnable_avg(u64 now,
1686 struct sched_avg *sa,
1690 u32 runnable_contrib;
1691 int delta_w, decayed = 0;
1693 delta = now - sa->last_runnable_update;
1695 * This should only happen when time goes backwards, which it
1696 * unfortunately does during sched clock init when we swap over to TSC.
1698 if ((s64)delta < 0) {
1699 sa->last_runnable_update = now;
1704 * Use 1024ns as the unit of measurement since it's a reasonable
1705 * approximation of 1us and fast to compute.
1710 sa->last_runnable_update = now;
1712 /* delta_w is the amount already accumulated against our next period */
1713 delta_w = sa->runnable_avg_period % 1024;
1714 if (delta + delta_w >= 1024) {
1715 /* period roll-over */
1719 * Now that we know we're crossing a period boundary, figure
1720 * out how much from delta we need to complete the current
1721 * period and accrue it.
1723 delta_w = 1024 - delta_w;
1725 sa->runnable_avg_sum += delta_w;
1726 sa->runnable_avg_period += delta_w;
1730 /* Figure out how many additional periods this update spans */
1731 periods = delta / 1024;
1734 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1736 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1739 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1740 runnable_contrib = __compute_runnable_contrib(periods);
1742 sa->runnable_avg_sum += runnable_contrib;
1743 sa->runnable_avg_period += runnable_contrib;
1746 /* Remainder of delta accrued against u_0` */
1748 sa->runnable_avg_sum += delta;
1749 sa->runnable_avg_period += delta;
1754 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1755 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1757 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1758 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1760 decays -= se->avg.decay_count;
1764 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1765 se->avg.decay_count = 0;
1770 #ifdef CONFIG_FAIR_GROUP_SCHED
1771 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1774 struct task_group *tg = cfs_rq->tg;
1777 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1778 tg_contrib -= cfs_rq->tg_load_contrib;
1780 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1781 atomic_long_add(tg_contrib, &tg->load_avg);
1782 cfs_rq->tg_load_contrib += tg_contrib;
1787 * Aggregate cfs_rq runnable averages into an equivalent task_group
1788 * representation for computing load contributions.
1790 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1791 struct cfs_rq *cfs_rq)
1793 struct task_group *tg = cfs_rq->tg;
1796 /* The fraction of a cpu used by this cfs_rq */
1797 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1798 sa->runnable_avg_period + 1);
1799 contrib -= cfs_rq->tg_runnable_contrib;
1801 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1802 atomic_add(contrib, &tg->runnable_avg);
1803 cfs_rq->tg_runnable_contrib += contrib;
1807 static inline void __update_group_entity_contrib(struct sched_entity *se)
1809 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1810 struct task_group *tg = cfs_rq->tg;
1815 contrib = cfs_rq->tg_load_contrib * tg->shares;
1816 se->avg.load_avg_contrib = div_u64(contrib,
1817 atomic_long_read(&tg->load_avg) + 1);
1820 * For group entities we need to compute a correction term in the case
1821 * that they are consuming <1 cpu so that we would contribute the same
1822 * load as a task of equal weight.
1824 * Explicitly co-ordinating this measurement would be expensive, but
1825 * fortunately the sum of each cpus contribution forms a usable
1826 * lower-bound on the true value.
1828 * Consider the aggregate of 2 contributions. Either they are disjoint
1829 * (and the sum represents true value) or they are disjoint and we are
1830 * understating by the aggregate of their overlap.
1832 * Extending this to N cpus, for a given overlap, the maximum amount we
1833 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1834 * cpus that overlap for this interval and w_i is the interval width.
1836 * On a small machine; the first term is well-bounded which bounds the
1837 * total error since w_i is a subset of the period. Whereas on a
1838 * larger machine, while this first term can be larger, if w_i is the
1839 * of consequential size guaranteed to see n_i*w_i quickly converge to
1840 * our upper bound of 1-cpu.
1842 runnable_avg = atomic_read(&tg->runnable_avg);
1843 if (runnable_avg < NICE_0_LOAD) {
1844 se->avg.load_avg_contrib *= runnable_avg;
1845 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1849 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1850 int force_update) {}
1851 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1852 struct cfs_rq *cfs_rq) {}
1853 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1856 static inline void __update_task_entity_contrib(struct sched_entity *se)
1860 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1861 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1862 contrib /= (se->avg.runnable_avg_period + 1);
1863 se->avg.load_avg_contrib = scale_load(contrib);
1866 /* Compute the current contribution to load_avg by se, return any delta */
1867 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1869 long old_contrib = se->avg.load_avg_contrib;
1871 if (entity_is_task(se)) {
1872 __update_task_entity_contrib(se);
1874 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1875 __update_group_entity_contrib(se);
1878 return se->avg.load_avg_contrib - old_contrib;
1881 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1884 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1885 cfs_rq->blocked_load_avg -= load_contrib;
1887 cfs_rq->blocked_load_avg = 0;
1890 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1892 /* Update a sched_entity's runnable average */
1893 static inline void update_entity_load_avg(struct sched_entity *se,
1896 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1901 * For a group entity we need to use their owned cfs_rq_clock_task() in
1902 * case they are the parent of a throttled hierarchy.
1904 if (entity_is_task(se))
1905 now = cfs_rq_clock_task(cfs_rq);
1907 now = cfs_rq_clock_task(group_cfs_rq(se));
1909 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1912 contrib_delta = __update_entity_load_avg_contrib(se);
1918 cfs_rq->runnable_load_avg += contrib_delta;
1920 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1924 * Decay the load contributed by all blocked children and account this so that
1925 * their contribution may appropriately discounted when they wake up.
1927 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1929 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1932 decays = now - cfs_rq->last_decay;
1933 if (!decays && !force_update)
1936 if (atomic_long_read(&cfs_rq->removed_load)) {
1937 unsigned long removed_load;
1938 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1939 subtract_blocked_load_contrib(cfs_rq, removed_load);
1943 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1945 atomic64_add(decays, &cfs_rq->decay_counter);
1946 cfs_rq->last_decay = now;
1949 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1952 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1954 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1955 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1958 /* Add the load generated by se into cfs_rq's child load-average */
1959 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1960 struct sched_entity *se,
1964 * We track migrations using entity decay_count <= 0, on a wake-up
1965 * migration we use a negative decay count to track the remote decays
1966 * accumulated while sleeping.
1968 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1969 * are seen by enqueue_entity_load_avg() as a migration with an already
1970 * constructed load_avg_contrib.
1972 if (unlikely(se->avg.decay_count <= 0)) {
1973 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1974 if (se->avg.decay_count) {
1976 * In a wake-up migration we have to approximate the
1977 * time sleeping. This is because we can't synchronize
1978 * clock_task between the two cpus, and it is not
1979 * guaranteed to be read-safe. Instead, we can
1980 * approximate this using our carried decays, which are
1981 * explicitly atomically readable.
1983 se->avg.last_runnable_update -= (-se->avg.decay_count)
1985 update_entity_load_avg(se, 0);
1986 /* Indicate that we're now synchronized and on-rq */
1987 se->avg.decay_count = 0;
1992 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1993 * would have made count negative); we must be careful to avoid
1994 * double-accounting blocked time after synchronizing decays.
1996 se->avg.last_runnable_update += __synchronize_entity_decay(se)
2000 /* migrated tasks did not contribute to our blocked load */
2002 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2003 update_entity_load_avg(se, 0);
2006 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2007 /* we force update consideration on load-balancer moves */
2008 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2012 * Remove se's load from this cfs_rq child load-average, if the entity is
2013 * transitioning to a blocked state we track its projected decay using
2016 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2017 struct sched_entity *se,
2020 update_entity_load_avg(se, 1);
2021 /* we force update consideration on load-balancer moves */
2022 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2024 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2026 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2027 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2028 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2032 * Update the rq's load with the elapsed running time before entering
2033 * idle. if the last scheduled task is not a CFS task, idle_enter will
2034 * be the only way to update the runnable statistic.
2036 void idle_enter_fair(struct rq *this_rq)
2038 update_rq_runnable_avg(this_rq, 1);
2042 * Update the rq's load with the elapsed idle time before a task is
2043 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2044 * be the only way to update the runnable statistic.
2046 void idle_exit_fair(struct rq *this_rq)
2048 update_rq_runnable_avg(this_rq, 0);
2052 static inline void update_entity_load_avg(struct sched_entity *se,
2053 int update_cfs_rq) {}
2054 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2055 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2056 struct sched_entity *se,
2058 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2059 struct sched_entity *se,
2061 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2062 int force_update) {}
2065 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2067 #ifdef CONFIG_SCHEDSTATS
2068 struct task_struct *tsk = NULL;
2070 if (entity_is_task(se))
2073 if (se->statistics.sleep_start) {
2074 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2079 if (unlikely(delta > se->statistics.sleep_max))
2080 se->statistics.sleep_max = delta;
2082 se->statistics.sleep_start = 0;
2083 se->statistics.sum_sleep_runtime += delta;
2086 account_scheduler_latency(tsk, delta >> 10, 1);
2087 trace_sched_stat_sleep(tsk, delta);
2090 if (se->statistics.block_start) {
2091 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2096 if (unlikely(delta > se->statistics.block_max))
2097 se->statistics.block_max = delta;
2099 se->statistics.block_start = 0;
2100 se->statistics.sum_sleep_runtime += delta;
2103 if (tsk->in_iowait) {
2104 se->statistics.iowait_sum += delta;
2105 se->statistics.iowait_count++;
2106 trace_sched_stat_iowait(tsk, delta);
2109 trace_sched_stat_blocked(tsk, delta);
2112 * Blocking time is in units of nanosecs, so shift by
2113 * 20 to get a milliseconds-range estimation of the
2114 * amount of time that the task spent sleeping:
2116 if (unlikely(prof_on == SLEEP_PROFILING)) {
2117 profile_hits(SLEEP_PROFILING,
2118 (void *)get_wchan(tsk),
2121 account_scheduler_latency(tsk, delta >> 10, 0);
2127 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2129 #ifdef CONFIG_SCHED_DEBUG
2130 s64 d = se->vruntime - cfs_rq->min_vruntime;
2135 if (d > 3*sysctl_sched_latency)
2136 schedstat_inc(cfs_rq, nr_spread_over);
2141 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2143 u64 vruntime = cfs_rq->min_vruntime;
2146 * The 'current' period is already promised to the current tasks,
2147 * however the extra weight of the new task will slow them down a
2148 * little, place the new task so that it fits in the slot that
2149 * stays open at the end.
2151 if (initial && sched_feat(START_DEBIT))
2152 vruntime += sched_vslice(cfs_rq, se);
2154 /* sleeps up to a single latency don't count. */
2156 unsigned long thresh = sysctl_sched_latency;
2159 * Halve their sleep time's effect, to allow
2160 * for a gentler effect of sleepers:
2162 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2168 /* ensure we never gain time by being placed backwards. */
2169 se->vruntime = max_vruntime(se->vruntime, vruntime);
2172 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2175 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2178 * Update the normalized vruntime before updating min_vruntime
2179 * through calling update_curr().
2181 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2182 se->vruntime += cfs_rq->min_vruntime;
2185 * Update run-time statistics of the 'current'.
2187 update_curr(cfs_rq);
2188 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2189 account_entity_enqueue(cfs_rq, se);
2190 update_cfs_shares(cfs_rq);
2192 if (flags & ENQUEUE_WAKEUP) {
2193 place_entity(cfs_rq, se, 0);
2194 enqueue_sleeper(cfs_rq, se);
2197 update_stats_enqueue(cfs_rq, se);
2198 check_spread(cfs_rq, se);
2199 if (se != cfs_rq->curr)
2200 __enqueue_entity(cfs_rq, se);
2203 if (cfs_rq->nr_running == 1) {
2204 list_add_leaf_cfs_rq(cfs_rq);
2205 check_enqueue_throttle(cfs_rq);
2209 static void __clear_buddies_last(struct sched_entity *se)
2211 for_each_sched_entity(se) {
2212 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2213 if (cfs_rq->last == se)
2214 cfs_rq->last = NULL;
2220 static void __clear_buddies_next(struct sched_entity *se)
2222 for_each_sched_entity(se) {
2223 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2224 if (cfs_rq->next == se)
2225 cfs_rq->next = NULL;
2231 static void __clear_buddies_skip(struct sched_entity *se)
2233 for_each_sched_entity(se) {
2234 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2235 if (cfs_rq->skip == se)
2236 cfs_rq->skip = NULL;
2242 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2244 if (cfs_rq->last == se)
2245 __clear_buddies_last(se);
2247 if (cfs_rq->next == se)
2248 __clear_buddies_next(se);
2250 if (cfs_rq->skip == se)
2251 __clear_buddies_skip(se);
2254 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2257 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2260 * Update run-time statistics of the 'current'.
2262 update_curr(cfs_rq);
2263 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2265 update_stats_dequeue(cfs_rq, se);
2266 if (flags & DEQUEUE_SLEEP) {
2267 #ifdef CONFIG_SCHEDSTATS
2268 if (entity_is_task(se)) {
2269 struct task_struct *tsk = task_of(se);
2271 if (tsk->state & TASK_INTERRUPTIBLE)
2272 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2273 if (tsk->state & TASK_UNINTERRUPTIBLE)
2274 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2279 clear_buddies(cfs_rq, se);
2281 if (se != cfs_rq->curr)
2282 __dequeue_entity(cfs_rq, se);
2284 account_entity_dequeue(cfs_rq, se);
2287 * Normalize the entity after updating the min_vruntime because the
2288 * update can refer to the ->curr item and we need to reflect this
2289 * movement in our normalized position.
2291 if (!(flags & DEQUEUE_SLEEP))
2292 se->vruntime -= cfs_rq->min_vruntime;
2294 /* return excess runtime on last dequeue */
2295 return_cfs_rq_runtime(cfs_rq);
2297 update_min_vruntime(cfs_rq);
2298 update_cfs_shares(cfs_rq);
2302 * Preempt the current task with a newly woken task if needed:
2305 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2307 unsigned long ideal_runtime, delta_exec;
2308 struct sched_entity *se;
2311 ideal_runtime = sched_slice(cfs_rq, curr);
2312 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2313 if (delta_exec > ideal_runtime) {
2314 resched_task(rq_of(cfs_rq)->curr);
2316 * The current task ran long enough, ensure it doesn't get
2317 * re-elected due to buddy favours.
2319 clear_buddies(cfs_rq, curr);
2324 * Ensure that a task that missed wakeup preemption by a
2325 * narrow margin doesn't have to wait for a full slice.
2326 * This also mitigates buddy induced latencies under load.
2328 if (delta_exec < sysctl_sched_min_granularity)
2331 se = __pick_first_entity(cfs_rq);
2332 delta = curr->vruntime - se->vruntime;
2337 if (delta > ideal_runtime)
2338 resched_task(rq_of(cfs_rq)->curr);
2342 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2344 /* 'current' is not kept within the tree. */
2347 * Any task has to be enqueued before it get to execute on
2348 * a CPU. So account for the time it spent waiting on the
2351 update_stats_wait_end(cfs_rq, se);
2352 __dequeue_entity(cfs_rq, se);
2355 update_stats_curr_start(cfs_rq, se);
2357 #ifdef CONFIG_SCHEDSTATS
2359 * Track our maximum slice length, if the CPU's load is at
2360 * least twice that of our own weight (i.e. dont track it
2361 * when there are only lesser-weight tasks around):
2363 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2364 se->statistics.slice_max = max(se->statistics.slice_max,
2365 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2368 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2372 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2375 * Pick the next process, keeping these things in mind, in this order:
2376 * 1) keep things fair between processes/task groups
2377 * 2) pick the "next" process, since someone really wants that to run
2378 * 3) pick the "last" process, for cache locality
2379 * 4) do not run the "skip" process, if something else is available
2381 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2383 struct sched_entity *se = __pick_first_entity(cfs_rq);
2384 struct sched_entity *left = se;
2387 * Avoid running the skip buddy, if running something else can
2388 * be done without getting too unfair.
2390 if (cfs_rq->skip == se) {
2391 struct sched_entity *second = __pick_next_entity(se);
2392 if (second && wakeup_preempt_entity(second, left) < 1)
2397 * Prefer last buddy, try to return the CPU to a preempted task.
2399 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2403 * Someone really wants this to run. If it's not unfair, run it.
2405 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2408 clear_buddies(cfs_rq, se);
2413 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2415 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2418 * If still on the runqueue then deactivate_task()
2419 * was not called and update_curr() has to be done:
2422 update_curr(cfs_rq);
2424 /* throttle cfs_rqs exceeding runtime */
2425 check_cfs_rq_runtime(cfs_rq);
2427 check_spread(cfs_rq, prev);
2429 update_stats_wait_start(cfs_rq, prev);
2430 /* Put 'current' back into the tree. */
2431 __enqueue_entity(cfs_rq, prev);
2432 /* in !on_rq case, update occurred at dequeue */
2433 update_entity_load_avg(prev, 1);
2435 cfs_rq->curr = NULL;
2439 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2442 * Update run-time statistics of the 'current'.
2444 update_curr(cfs_rq);
2447 * Ensure that runnable average is periodically updated.
2449 update_entity_load_avg(curr, 1);
2450 update_cfs_rq_blocked_load(cfs_rq, 1);
2451 update_cfs_shares(cfs_rq);
2453 #ifdef CONFIG_SCHED_HRTICK
2455 * queued ticks are scheduled to match the slice, so don't bother
2456 * validating it and just reschedule.
2459 resched_task(rq_of(cfs_rq)->curr);
2463 * don't let the period tick interfere with the hrtick preemption
2465 if (!sched_feat(DOUBLE_TICK) &&
2466 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2470 if (cfs_rq->nr_running > 1)
2471 check_preempt_tick(cfs_rq, curr);
2475 /**************************************************
2476 * CFS bandwidth control machinery
2479 #ifdef CONFIG_CFS_BANDWIDTH
2481 #ifdef HAVE_JUMP_LABEL
2482 static struct static_key __cfs_bandwidth_used;
2484 static inline bool cfs_bandwidth_used(void)
2486 return static_key_false(&__cfs_bandwidth_used);
2489 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2491 /* only need to count groups transitioning between enabled/!enabled */
2492 if (enabled && !was_enabled)
2493 static_key_slow_inc(&__cfs_bandwidth_used);
2494 else if (!enabled && was_enabled)
2495 static_key_slow_dec(&__cfs_bandwidth_used);
2497 #else /* HAVE_JUMP_LABEL */
2498 static bool cfs_bandwidth_used(void)
2503 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2504 #endif /* HAVE_JUMP_LABEL */
2507 * default period for cfs group bandwidth.
2508 * default: 0.1s, units: nanoseconds
2510 static inline u64 default_cfs_period(void)
2512 return 100000000ULL;
2515 static inline u64 sched_cfs_bandwidth_slice(void)
2517 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2521 * Replenish runtime according to assigned quota and update expiration time.
2522 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2523 * additional synchronization around rq->lock.
2525 * requires cfs_b->lock
2527 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2531 if (cfs_b->quota == RUNTIME_INF)
2534 now = sched_clock_cpu(smp_processor_id());
2535 cfs_b->runtime = cfs_b->quota;
2536 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2539 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2541 return &tg->cfs_bandwidth;
2544 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2545 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2547 if (unlikely(cfs_rq->throttle_count))
2548 return cfs_rq->throttled_clock_task;
2550 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2553 /* returns 0 on failure to allocate runtime */
2554 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2556 struct task_group *tg = cfs_rq->tg;
2557 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2558 u64 amount = 0, min_amount, expires;
2560 /* note: this is a positive sum as runtime_remaining <= 0 */
2561 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2563 raw_spin_lock(&cfs_b->lock);
2564 if (cfs_b->quota == RUNTIME_INF)
2565 amount = min_amount;
2568 * If the bandwidth pool has become inactive, then at least one
2569 * period must have elapsed since the last consumption.
2570 * Refresh the global state and ensure bandwidth timer becomes
2573 if (!cfs_b->timer_active) {
2574 __refill_cfs_bandwidth_runtime(cfs_b);
2575 __start_cfs_bandwidth(cfs_b);
2578 if (cfs_b->runtime > 0) {
2579 amount = min(cfs_b->runtime, min_amount);
2580 cfs_b->runtime -= amount;
2584 expires = cfs_b->runtime_expires;
2585 raw_spin_unlock(&cfs_b->lock);
2587 cfs_rq->runtime_remaining += amount;
2589 * we may have advanced our local expiration to account for allowed
2590 * spread between our sched_clock and the one on which runtime was
2593 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2594 cfs_rq->runtime_expires = expires;
2596 return cfs_rq->runtime_remaining > 0;
2600 * Note: This depends on the synchronization provided by sched_clock and the
2601 * fact that rq->clock snapshots this value.
2603 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2605 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2607 /* if the deadline is ahead of our clock, nothing to do */
2608 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2611 if (cfs_rq->runtime_remaining < 0)
2615 * If the local deadline has passed we have to consider the
2616 * possibility that our sched_clock is 'fast' and the global deadline
2617 * has not truly expired.
2619 * Fortunately we can check determine whether this the case by checking
2620 * whether the global deadline has advanced.
2623 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2624 /* extend local deadline, drift is bounded above by 2 ticks */
2625 cfs_rq->runtime_expires += TICK_NSEC;
2627 /* global deadline is ahead, expiration has passed */
2628 cfs_rq->runtime_remaining = 0;
2632 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2633 unsigned long delta_exec)
2635 /* dock delta_exec before expiring quota (as it could span periods) */
2636 cfs_rq->runtime_remaining -= delta_exec;
2637 expire_cfs_rq_runtime(cfs_rq);
2639 if (likely(cfs_rq->runtime_remaining > 0))
2643 * if we're unable to extend our runtime we resched so that the active
2644 * hierarchy can be throttled
2646 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2647 resched_task(rq_of(cfs_rq)->curr);
2650 static __always_inline
2651 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2653 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2656 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2659 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2661 return cfs_bandwidth_used() && cfs_rq->throttled;
2664 /* check whether cfs_rq, or any parent, is throttled */
2665 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2667 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2671 * Ensure that neither of the group entities corresponding to src_cpu or
2672 * dest_cpu are members of a throttled hierarchy when performing group
2673 * load-balance operations.
2675 static inline int throttled_lb_pair(struct task_group *tg,
2676 int src_cpu, int dest_cpu)
2678 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2680 src_cfs_rq = tg->cfs_rq[src_cpu];
2681 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2683 return throttled_hierarchy(src_cfs_rq) ||
2684 throttled_hierarchy(dest_cfs_rq);
2687 /* updated child weight may affect parent so we have to do this bottom up */
2688 static int tg_unthrottle_up(struct task_group *tg, void *data)
2690 struct rq *rq = data;
2691 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2693 cfs_rq->throttle_count--;
2695 if (!cfs_rq->throttle_count) {
2696 /* adjust cfs_rq_clock_task() */
2697 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2698 cfs_rq->throttled_clock_task;
2705 static int tg_throttle_down(struct task_group *tg, void *data)
2707 struct rq *rq = data;
2708 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2710 /* group is entering throttled state, stop time */
2711 if (!cfs_rq->throttle_count)
2712 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2713 cfs_rq->throttle_count++;
2718 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2720 struct rq *rq = rq_of(cfs_rq);
2721 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2722 struct sched_entity *se;
2723 long task_delta, dequeue = 1;
2725 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2727 /* freeze hierarchy runnable averages while throttled */
2729 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2732 task_delta = cfs_rq->h_nr_running;
2733 for_each_sched_entity(se) {
2734 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2735 /* throttled entity or throttle-on-deactivate */
2740 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2741 qcfs_rq->h_nr_running -= task_delta;
2743 if (qcfs_rq->load.weight)
2748 rq->nr_running -= task_delta;
2750 cfs_rq->throttled = 1;
2751 cfs_rq->throttled_clock = rq_clock(rq);
2752 raw_spin_lock(&cfs_b->lock);
2753 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2754 raw_spin_unlock(&cfs_b->lock);
2757 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2759 struct rq *rq = rq_of(cfs_rq);
2760 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2761 struct sched_entity *se;
2765 se = cfs_rq->tg->se[cpu_of(rq)];
2767 cfs_rq->throttled = 0;
2769 update_rq_clock(rq);
2771 raw_spin_lock(&cfs_b->lock);
2772 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2773 list_del_rcu(&cfs_rq->throttled_list);
2774 raw_spin_unlock(&cfs_b->lock);
2776 /* update hierarchical throttle state */
2777 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2779 if (!cfs_rq->load.weight)
2782 task_delta = cfs_rq->h_nr_running;
2783 for_each_sched_entity(se) {
2787 cfs_rq = cfs_rq_of(se);
2789 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2790 cfs_rq->h_nr_running += task_delta;
2792 if (cfs_rq_throttled(cfs_rq))
2797 rq->nr_running += task_delta;
2799 /* determine whether we need to wake up potentially idle cpu */
2800 if (rq->curr == rq->idle && rq->cfs.nr_running)
2801 resched_task(rq->curr);
2804 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2805 u64 remaining, u64 expires)
2807 struct cfs_rq *cfs_rq;
2808 u64 runtime = remaining;
2811 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2813 struct rq *rq = rq_of(cfs_rq);
2815 raw_spin_lock(&rq->lock);
2816 if (!cfs_rq_throttled(cfs_rq))
2819 runtime = -cfs_rq->runtime_remaining + 1;
2820 if (runtime > remaining)
2821 runtime = remaining;
2822 remaining -= runtime;
2824 cfs_rq->runtime_remaining += runtime;
2825 cfs_rq->runtime_expires = expires;
2827 /* we check whether we're throttled above */
2828 if (cfs_rq->runtime_remaining > 0)
2829 unthrottle_cfs_rq(cfs_rq);
2832 raw_spin_unlock(&rq->lock);
2843 * Responsible for refilling a task_group's bandwidth and unthrottling its
2844 * cfs_rqs as appropriate. If there has been no activity within the last
2845 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2846 * used to track this state.
2848 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2850 u64 runtime, runtime_expires;
2851 int idle = 1, throttled;
2853 raw_spin_lock(&cfs_b->lock);
2854 /* no need to continue the timer with no bandwidth constraint */
2855 if (cfs_b->quota == RUNTIME_INF)
2858 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2859 /* idle depends on !throttled (for the case of a large deficit) */
2860 idle = cfs_b->idle && !throttled;
2861 cfs_b->nr_periods += overrun;
2863 /* if we're going inactive then everything else can be deferred */
2867 __refill_cfs_bandwidth_runtime(cfs_b);
2870 /* mark as potentially idle for the upcoming period */
2875 /* account preceding periods in which throttling occurred */
2876 cfs_b->nr_throttled += overrun;
2879 * There are throttled entities so we must first use the new bandwidth
2880 * to unthrottle them before making it generally available. This
2881 * ensures that all existing debts will be paid before a new cfs_rq is
2884 runtime = cfs_b->runtime;
2885 runtime_expires = cfs_b->runtime_expires;
2889 * This check is repeated as we are holding onto the new bandwidth
2890 * while we unthrottle. This can potentially race with an unthrottled
2891 * group trying to acquire new bandwidth from the global pool.
2893 while (throttled && runtime > 0) {
2894 raw_spin_unlock(&cfs_b->lock);
2895 /* we can't nest cfs_b->lock while distributing bandwidth */
2896 runtime = distribute_cfs_runtime(cfs_b, runtime,
2898 raw_spin_lock(&cfs_b->lock);
2900 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2903 /* return (any) remaining runtime */
2904 cfs_b->runtime = runtime;
2906 * While we are ensured activity in the period following an
2907 * unthrottle, this also covers the case in which the new bandwidth is
2908 * insufficient to cover the existing bandwidth deficit. (Forcing the
2909 * timer to remain active while there are any throttled entities.)
2914 cfs_b->timer_active = 0;
2915 raw_spin_unlock(&cfs_b->lock);
2920 /* a cfs_rq won't donate quota below this amount */
2921 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2922 /* minimum remaining period time to redistribute slack quota */
2923 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2924 /* how long we wait to gather additional slack before distributing */
2925 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2927 /* are we near the end of the current quota period? */
2928 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2930 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2933 /* if the call-back is running a quota refresh is already occurring */
2934 if (hrtimer_callback_running(refresh_timer))
2937 /* is a quota refresh about to occur? */
2938 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2939 if (remaining < min_expire)
2945 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2947 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2949 /* if there's a quota refresh soon don't bother with slack */
2950 if (runtime_refresh_within(cfs_b, min_left))
2953 start_bandwidth_timer(&cfs_b->slack_timer,
2954 ns_to_ktime(cfs_bandwidth_slack_period));
2957 /* we know any runtime found here is valid as update_curr() precedes return */
2958 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2960 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2961 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2963 if (slack_runtime <= 0)
2966 raw_spin_lock(&cfs_b->lock);
2967 if (cfs_b->quota != RUNTIME_INF &&
2968 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2969 cfs_b->runtime += slack_runtime;
2971 /* we are under rq->lock, defer unthrottling using a timer */
2972 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2973 !list_empty(&cfs_b->throttled_cfs_rq))
2974 start_cfs_slack_bandwidth(cfs_b);
2976 raw_spin_unlock(&cfs_b->lock);
2978 /* even if it's not valid for return we don't want to try again */
2979 cfs_rq->runtime_remaining -= slack_runtime;
2982 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2984 if (!cfs_bandwidth_used())
2987 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2990 __return_cfs_rq_runtime(cfs_rq);
2994 * This is done with a timer (instead of inline with bandwidth return) since
2995 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2997 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2999 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3002 /* confirm we're still not at a refresh boundary */
3003 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
3006 raw_spin_lock(&cfs_b->lock);
3007 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3008 runtime = cfs_b->runtime;
3011 expires = cfs_b->runtime_expires;
3012 raw_spin_unlock(&cfs_b->lock);
3017 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3019 raw_spin_lock(&cfs_b->lock);
3020 if (expires == cfs_b->runtime_expires)
3021 cfs_b->runtime = runtime;
3022 raw_spin_unlock(&cfs_b->lock);
3026 * When a group wakes up we want to make sure that its quota is not already
3027 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3028 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3030 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3032 if (!cfs_bandwidth_used())
3035 /* an active group must be handled by the update_curr()->put() path */
3036 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3039 /* ensure the group is not already throttled */
3040 if (cfs_rq_throttled(cfs_rq))
3043 /* update runtime allocation */
3044 account_cfs_rq_runtime(cfs_rq, 0);
3045 if (cfs_rq->runtime_remaining <= 0)
3046 throttle_cfs_rq(cfs_rq);
3049 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3050 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3052 if (!cfs_bandwidth_used())
3055 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3059 * it's possible for a throttled entity to be forced into a running
3060 * state (e.g. set_curr_task), in this case we're finished.
3062 if (cfs_rq_throttled(cfs_rq))
3065 throttle_cfs_rq(cfs_rq);
3068 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3070 struct cfs_bandwidth *cfs_b =
3071 container_of(timer, struct cfs_bandwidth, slack_timer);
3072 do_sched_cfs_slack_timer(cfs_b);
3074 return HRTIMER_NORESTART;
3077 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3079 struct cfs_bandwidth *cfs_b =
3080 container_of(timer, struct cfs_bandwidth, period_timer);
3086 now = hrtimer_cb_get_time(timer);
3087 overrun = hrtimer_forward(timer, now, cfs_b->period);
3092 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3095 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3098 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3100 raw_spin_lock_init(&cfs_b->lock);
3102 cfs_b->quota = RUNTIME_INF;
3103 cfs_b->period = ns_to_ktime(default_cfs_period());
3105 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3106 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3107 cfs_b->period_timer.function = sched_cfs_period_timer;
3108 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3109 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3112 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3114 cfs_rq->runtime_enabled = 0;
3115 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3118 /* requires cfs_b->lock, may release to reprogram timer */
3119 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3122 * The timer may be active because we're trying to set a new bandwidth
3123 * period or because we're racing with the tear-down path
3124 * (timer_active==0 becomes visible before the hrtimer call-back
3125 * terminates). In either case we ensure that it's re-programmed
3127 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
3128 raw_spin_unlock(&cfs_b->lock);
3129 /* ensure cfs_b->lock is available while we wait */
3130 hrtimer_cancel(&cfs_b->period_timer);
3132 raw_spin_lock(&cfs_b->lock);
3133 /* if someone else restarted the timer then we're done */
3134 if (cfs_b->timer_active)
3138 cfs_b->timer_active = 1;
3139 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3142 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3144 hrtimer_cancel(&cfs_b->period_timer);
3145 hrtimer_cancel(&cfs_b->slack_timer);
3148 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3150 struct cfs_rq *cfs_rq;
3152 for_each_leaf_cfs_rq(rq, cfs_rq) {
3153 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3155 if (!cfs_rq->runtime_enabled)
3159 * clock_task is not advancing so we just need to make sure
3160 * there's some valid quota amount
3162 cfs_rq->runtime_remaining = cfs_b->quota;
3163 if (cfs_rq_throttled(cfs_rq))
3164 unthrottle_cfs_rq(cfs_rq);
3168 #else /* CONFIG_CFS_BANDWIDTH */
3169 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3171 return rq_clock_task(rq_of(cfs_rq));
3174 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3175 unsigned long delta_exec) {}
3176 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3177 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3178 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3180 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3185 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3190 static inline int throttled_lb_pair(struct task_group *tg,
3191 int src_cpu, int dest_cpu)
3196 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3198 #ifdef CONFIG_FAIR_GROUP_SCHED
3199 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3202 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3206 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3207 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3209 #endif /* CONFIG_CFS_BANDWIDTH */
3211 /**************************************************
3212 * CFS operations on tasks:
3215 #ifdef CONFIG_SCHED_HRTICK
3216 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3218 struct sched_entity *se = &p->se;
3219 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3221 WARN_ON(task_rq(p) != rq);
3223 if (cfs_rq->nr_running > 1) {
3224 u64 slice = sched_slice(cfs_rq, se);
3225 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3226 s64 delta = slice - ran;
3235 * Don't schedule slices shorter than 10000ns, that just
3236 * doesn't make sense. Rely on vruntime for fairness.
3239 delta = max_t(s64, 10000LL, delta);
3241 hrtick_start(rq, delta);
3246 * called from enqueue/dequeue and updates the hrtick when the
3247 * current task is from our class and nr_running is low enough
3250 static void hrtick_update(struct rq *rq)
3252 struct task_struct *curr = rq->curr;
3254 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3257 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3258 hrtick_start_fair(rq, curr);
3260 #else /* !CONFIG_SCHED_HRTICK */
3262 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3266 static inline void hrtick_update(struct rq *rq)
3272 * The enqueue_task method is called before nr_running is
3273 * increased. Here we update the fair scheduling stats and
3274 * then put the task into the rbtree:
3277 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3279 struct cfs_rq *cfs_rq;
3280 struct sched_entity *se = &p->se;
3282 for_each_sched_entity(se) {
3285 cfs_rq = cfs_rq_of(se);
3286 enqueue_entity(cfs_rq, se, flags);
3289 * end evaluation on encountering a throttled cfs_rq
3291 * note: in the case of encountering a throttled cfs_rq we will
3292 * post the final h_nr_running increment below.
3294 if (cfs_rq_throttled(cfs_rq))
3296 cfs_rq->h_nr_running++;
3298 flags = ENQUEUE_WAKEUP;
3301 for_each_sched_entity(se) {
3302 cfs_rq = cfs_rq_of(se);
3303 cfs_rq->h_nr_running++;
3305 if (cfs_rq_throttled(cfs_rq))
3308 update_cfs_shares(cfs_rq);
3309 update_entity_load_avg(se, 1);
3313 update_rq_runnable_avg(rq, rq->nr_running);
3319 static void set_next_buddy(struct sched_entity *se);
3322 * The dequeue_task method is called before nr_running is
3323 * decreased. We remove the task from the rbtree and
3324 * update the fair scheduling stats:
3326 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3328 struct cfs_rq *cfs_rq;
3329 struct sched_entity *se = &p->se;
3330 int task_sleep = flags & DEQUEUE_SLEEP;
3332 for_each_sched_entity(se) {
3333 cfs_rq = cfs_rq_of(se);
3334 dequeue_entity(cfs_rq, se, flags);
3337 * end evaluation on encountering a throttled cfs_rq
3339 * note: in the case of encountering a throttled cfs_rq we will
3340 * post the final h_nr_running decrement below.
3342 if (cfs_rq_throttled(cfs_rq))
3344 cfs_rq->h_nr_running--;
3346 /* Don't dequeue parent if it has other entities besides us */
3347 if (cfs_rq->load.weight) {
3349 * Bias pick_next to pick a task from this cfs_rq, as
3350 * p is sleeping when it is within its sched_slice.
3352 if (task_sleep && parent_entity(se))
3353 set_next_buddy(parent_entity(se));
3355 /* avoid re-evaluating load for this entity */
3356 se = parent_entity(se);
3359 flags |= DEQUEUE_SLEEP;
3362 for_each_sched_entity(se) {
3363 cfs_rq = cfs_rq_of(se);
3364 cfs_rq->h_nr_running--;
3366 if (cfs_rq_throttled(cfs_rq))
3369 update_cfs_shares(cfs_rq);
3370 update_entity_load_avg(se, 1);
3375 update_rq_runnable_avg(rq, 1);
3381 /* Used instead of source_load when we know the type == 0 */
3382 static unsigned long weighted_cpuload(const int cpu)
3384 return cpu_rq(cpu)->cfs.runnable_load_avg;
3388 * Return a low guess at the load of a migration-source cpu weighted
3389 * according to the scheduling class and "nice" value.
3391 * We want to under-estimate the load of migration sources, to
3392 * balance conservatively.
3394 static unsigned long source_load(int cpu, int type)
3396 struct rq *rq = cpu_rq(cpu);
3397 unsigned long total = weighted_cpuload(cpu);
3399 if (type == 0 || !sched_feat(LB_BIAS))
3402 return min(rq->cpu_load[type-1], total);
3406 * Return a high guess at the load of a migration-target cpu weighted
3407 * according to the scheduling class and "nice" value.
3409 static unsigned long target_load(int cpu, int type)
3411 struct rq *rq = cpu_rq(cpu);
3412 unsigned long total = weighted_cpuload(cpu);
3414 if (type == 0 || !sched_feat(LB_BIAS))
3417 return max(rq->cpu_load[type-1], total);
3420 static unsigned long power_of(int cpu)
3422 return cpu_rq(cpu)->cpu_power;
3425 static unsigned long cpu_avg_load_per_task(int cpu)
3427 struct rq *rq = cpu_rq(cpu);
3428 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3429 unsigned long load_avg = rq->cfs.runnable_load_avg;
3432 return load_avg / nr_running;
3437 static void record_wakee(struct task_struct *p)
3440 * Rough decay (wiping) for cost saving, don't worry
3441 * about the boundary, really active task won't care
3444 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3445 current->wakee_flips = 0;
3446 current->wakee_flip_decay_ts = jiffies;
3449 if (current->last_wakee != p) {
3450 current->last_wakee = p;
3451 current->wakee_flips++;
3455 static void task_waking_fair(struct task_struct *p)
3457 struct sched_entity *se = &p->se;
3458 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3461 #ifndef CONFIG_64BIT
3462 u64 min_vruntime_copy;
3465 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3467 min_vruntime = cfs_rq->min_vruntime;
3468 } while (min_vruntime != min_vruntime_copy);
3470 min_vruntime = cfs_rq->min_vruntime;
3473 se->vruntime -= min_vruntime;
3477 #ifdef CONFIG_FAIR_GROUP_SCHED
3479 * effective_load() calculates the load change as seen from the root_task_group
3481 * Adding load to a group doesn't make a group heavier, but can cause movement
3482 * of group shares between cpus. Assuming the shares were perfectly aligned one
3483 * can calculate the shift in shares.
3485 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3486 * on this @cpu and results in a total addition (subtraction) of @wg to the
3487 * total group weight.
3489 * Given a runqueue weight distribution (rw_i) we can compute a shares
3490 * distribution (s_i) using:
3492 * s_i = rw_i / \Sum rw_j (1)
3494 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3495 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3496 * shares distribution (s_i):
3498 * rw_i = { 2, 4, 1, 0 }
3499 * s_i = { 2/7, 4/7, 1/7, 0 }
3501 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3502 * task used to run on and the CPU the waker is running on), we need to
3503 * compute the effect of waking a task on either CPU and, in case of a sync
3504 * wakeup, compute the effect of the current task going to sleep.
3506 * So for a change of @wl to the local @cpu with an overall group weight change
3507 * of @wl we can compute the new shares distribution (s'_i) using:
3509 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3511 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3512 * differences in waking a task to CPU 0. The additional task changes the
3513 * weight and shares distributions like:
3515 * rw'_i = { 3, 4, 1, 0 }
3516 * s'_i = { 3/8, 4/8, 1/8, 0 }
3518 * We can then compute the difference in effective weight by using:
3520 * dw_i = S * (s'_i - s_i) (3)
3522 * Where 'S' is the group weight as seen by its parent.
3524 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3525 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3526 * 4/7) times the weight of the group.
3528 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3530 struct sched_entity *se = tg->se[cpu];
3532 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3535 for_each_sched_entity(se) {
3541 * W = @wg + \Sum rw_j
3543 W = wg + calc_tg_weight(tg, se->my_q);
3548 w = se->my_q->load.weight + wl;
3551 * wl = S * s'_i; see (2)
3554 wl = (w * tg->shares) / W;
3559 * Per the above, wl is the new se->load.weight value; since
3560 * those are clipped to [MIN_SHARES, ...) do so now. See
3561 * calc_cfs_shares().
3563 if (wl < MIN_SHARES)
3567 * wl = dw_i = S * (s'_i - s_i); see (3)
3569 wl -= se->load.weight;
3572 * Recursively apply this logic to all parent groups to compute
3573 * the final effective load change on the root group. Since
3574 * only the @tg group gets extra weight, all parent groups can
3575 * only redistribute existing shares. @wl is the shift in shares
3576 * resulting from this level per the above.
3585 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3592 static int wake_wide(struct task_struct *p)
3594 int factor = this_cpu_read(sd_llc_size);
3597 * Yeah, it's the switching-frequency, could means many wakee or
3598 * rapidly switch, use factor here will just help to automatically
3599 * adjust the loose-degree, so bigger node will lead to more pull.
3601 if (p->wakee_flips > factor) {
3603 * wakee is somewhat hot, it needs certain amount of cpu
3604 * resource, so if waker is far more hot, prefer to leave
3607 if (current->wakee_flips > (factor * p->wakee_flips))
3614 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3616 s64 this_load, load;
3617 int idx, this_cpu, prev_cpu;
3618 unsigned long tl_per_task;
3619 struct task_group *tg;
3620 unsigned long weight;
3624 * If we wake multiple tasks be careful to not bounce
3625 * ourselves around too much.
3631 this_cpu = smp_processor_id();
3632 prev_cpu = task_cpu(p);
3633 load = source_load(prev_cpu, idx);
3634 this_load = target_load(this_cpu, idx);
3637 * If sync wakeup then subtract the (maximum possible)
3638 * effect of the currently running task from the load
3639 * of the current CPU:
3642 tg = task_group(current);
3643 weight = current->se.load.weight;
3645 this_load += effective_load(tg, this_cpu, -weight, -weight);
3646 load += effective_load(tg, prev_cpu, 0, -weight);
3650 weight = p->se.load.weight;
3653 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3654 * due to the sync cause above having dropped this_load to 0, we'll
3655 * always have an imbalance, but there's really nothing you can do
3656 * about that, so that's good too.
3658 * Otherwise check if either cpus are near enough in load to allow this
3659 * task to be woken on this_cpu.
3661 if (this_load > 0) {
3662 s64 this_eff_load, prev_eff_load;
3664 this_eff_load = 100;
3665 this_eff_load *= power_of(prev_cpu);
3666 this_eff_load *= this_load +
3667 effective_load(tg, this_cpu, weight, weight);
3669 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3670 prev_eff_load *= power_of(this_cpu);
3671 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3673 balanced = this_eff_load <= prev_eff_load;
3678 * If the currently running task will sleep within
3679 * a reasonable amount of time then attract this newly
3682 if (sync && balanced)
3685 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3686 tl_per_task = cpu_avg_load_per_task(this_cpu);
3689 (this_load <= load &&
3690 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3692 * This domain has SD_WAKE_AFFINE and
3693 * p is cache cold in this domain, and
3694 * there is no bad imbalance.
3696 schedstat_inc(sd, ttwu_move_affine);
3697 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3705 * find_idlest_group finds and returns the least busy CPU group within the
3708 static struct sched_group *
3709 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3710 int this_cpu, int load_idx)
3712 struct sched_group *idlest = NULL, *group = sd->groups;
3713 unsigned long min_load = ULONG_MAX, this_load = 0;
3714 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3717 unsigned long load, avg_load;
3721 /* Skip over this group if it has no CPUs allowed */
3722 if (!cpumask_intersects(sched_group_cpus(group),
3723 tsk_cpus_allowed(p)))
3726 local_group = cpumask_test_cpu(this_cpu,
3727 sched_group_cpus(group));
3729 /* Tally up the load of all CPUs in the group */
3732 for_each_cpu(i, sched_group_cpus(group)) {
3733 /* Bias balancing toward cpus of our domain */
3735 load = source_load(i, load_idx);
3737 load = target_load(i, load_idx);
3742 /* Adjust by relative CPU power of the group */
3743 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3746 this_load = avg_load;
3747 } else if (avg_load < min_load) {
3748 min_load = avg_load;
3751 } while (group = group->next, group != sd->groups);
3753 if (!idlest || 100*this_load < imbalance*min_load)
3759 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3762 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3764 unsigned long load, min_load = ULONG_MAX;
3768 /* Traverse only the allowed CPUs */
3769 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3770 load = weighted_cpuload(i);
3772 if (load < min_load || (load == min_load && i == this_cpu)) {
3782 * Try and locate an idle CPU in the sched_domain.
3784 static int select_idle_sibling(struct task_struct *p, int target)
3786 struct sched_domain *sd;
3787 struct sched_group *sg;
3788 int i = task_cpu(p);
3790 if (idle_cpu(target))
3794 * If the prevous cpu is cache affine and idle, don't be stupid.
3796 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3800 * Otherwise, iterate the domains and find an elegible idle cpu.
3802 sd = rcu_dereference(per_cpu(sd_llc, target));
3803 for_each_lower_domain(sd) {
3806 if (!cpumask_intersects(sched_group_cpus(sg),
3807 tsk_cpus_allowed(p)))
3810 for_each_cpu(i, sched_group_cpus(sg)) {
3811 if (i == target || !idle_cpu(i))
3815 target = cpumask_first_and(sched_group_cpus(sg),
3816 tsk_cpus_allowed(p));
3820 } while (sg != sd->groups);
3827 * sched_balance_self: balance the current task (running on cpu) in domains
3828 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3831 * Balance, ie. select the least loaded group.
3833 * Returns the target CPU number, or the same CPU if no balancing is needed.
3835 * preempt must be disabled.
3838 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
3840 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3841 int cpu = smp_processor_id();
3843 int want_affine = 0;
3844 int sync = wake_flags & WF_SYNC;
3846 if (p->nr_cpus_allowed == 1)
3849 if (sd_flag & SD_BALANCE_WAKE) {
3850 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3856 for_each_domain(cpu, tmp) {
3857 if (!(tmp->flags & SD_LOAD_BALANCE))
3861 * If both cpu and prev_cpu are part of this domain,
3862 * cpu is a valid SD_WAKE_AFFINE target.
3864 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3865 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3870 if (tmp->flags & sd_flag)
3875 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3878 new_cpu = select_idle_sibling(p, prev_cpu);
3883 int load_idx = sd->forkexec_idx;
3884 struct sched_group *group;
3887 if (!(sd->flags & sd_flag)) {
3892 if (sd_flag & SD_BALANCE_WAKE)
3893 load_idx = sd->wake_idx;
3895 group = find_idlest_group(sd, p, cpu, load_idx);
3901 new_cpu = find_idlest_cpu(group, p, cpu);
3902 if (new_cpu == -1 || new_cpu == cpu) {
3903 /* Now try balancing at a lower domain level of cpu */
3908 /* Now try balancing at a lower domain level of new_cpu */
3910 weight = sd->span_weight;
3912 for_each_domain(cpu, tmp) {
3913 if (weight <= tmp->span_weight)
3915 if (tmp->flags & sd_flag)
3918 /* while loop will break here if sd == NULL */
3927 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3928 * cfs_rq_of(p) references at time of call are still valid and identify the
3929 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3930 * other assumptions, including the state of rq->lock, should be made.
3933 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3935 struct sched_entity *se = &p->se;
3936 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3939 * Load tracking: accumulate removed load so that it can be processed
3940 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3941 * to blocked load iff they have a positive decay-count. It can never
3942 * be negative here since on-rq tasks have decay-count == 0.
3944 if (se->avg.decay_count) {
3945 se->avg.decay_count = -__synchronize_entity_decay(se);
3946 atomic_long_add(se->avg.load_avg_contrib,
3947 &cfs_rq->removed_load);
3950 #endif /* CONFIG_SMP */
3952 static unsigned long
3953 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3955 unsigned long gran = sysctl_sched_wakeup_granularity;
3958 * Since its curr running now, convert the gran from real-time
3959 * to virtual-time in his units.
3961 * By using 'se' instead of 'curr' we penalize light tasks, so
3962 * they get preempted easier. That is, if 'se' < 'curr' then
3963 * the resulting gran will be larger, therefore penalizing the
3964 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3965 * be smaller, again penalizing the lighter task.
3967 * This is especially important for buddies when the leftmost
3968 * task is higher priority than the buddy.
3970 return calc_delta_fair(gran, se);
3974 * Should 'se' preempt 'curr'.
3988 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3990 s64 gran, vdiff = curr->vruntime - se->vruntime;
3995 gran = wakeup_gran(curr, se);
4002 static void set_last_buddy(struct sched_entity *se)
4004 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4007 for_each_sched_entity(se)
4008 cfs_rq_of(se)->last = se;
4011 static void set_next_buddy(struct sched_entity *se)
4013 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4016 for_each_sched_entity(se)
4017 cfs_rq_of(se)->next = se;
4020 static void set_skip_buddy(struct sched_entity *se)
4022 for_each_sched_entity(se)
4023 cfs_rq_of(se)->skip = se;
4027 * Preempt the current task with a newly woken task if needed:
4029 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4031 struct task_struct *curr = rq->curr;
4032 struct sched_entity *se = &curr->se, *pse = &p->se;
4033 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4034 int scale = cfs_rq->nr_running >= sched_nr_latency;
4035 int next_buddy_marked = 0;
4037 if (unlikely(se == pse))
4041 * This is possible from callers such as move_task(), in which we
4042 * unconditionally check_prempt_curr() after an enqueue (which may have
4043 * lead to a throttle). This both saves work and prevents false
4044 * next-buddy nomination below.
4046 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4049 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4050 set_next_buddy(pse);
4051 next_buddy_marked = 1;
4055 * We can come here with TIF_NEED_RESCHED already set from new task
4058 * Note: this also catches the edge-case of curr being in a throttled
4059 * group (e.g. via set_curr_task), since update_curr() (in the
4060 * enqueue of curr) will have resulted in resched being set. This
4061 * prevents us from potentially nominating it as a false LAST_BUDDY
4064 if (test_tsk_need_resched(curr))
4067 /* Idle tasks are by definition preempted by non-idle tasks. */
4068 if (unlikely(curr->policy == SCHED_IDLE) &&
4069 likely(p->policy != SCHED_IDLE))
4073 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4074 * is driven by the tick):
4076 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4079 find_matching_se(&se, &pse);
4080 update_curr(cfs_rq_of(se));
4082 if (wakeup_preempt_entity(se, pse) == 1) {
4084 * Bias pick_next to pick the sched entity that is
4085 * triggering this preemption.
4087 if (!next_buddy_marked)
4088 set_next_buddy(pse);
4097 * Only set the backward buddy when the current task is still
4098 * on the rq. This can happen when a wakeup gets interleaved
4099 * with schedule on the ->pre_schedule() or idle_balance()
4100 * point, either of which can * drop the rq lock.
4102 * Also, during early boot the idle thread is in the fair class,
4103 * for obvious reasons its a bad idea to schedule back to it.
4105 if (unlikely(!se->on_rq || curr == rq->idle))
4108 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4112 static struct task_struct *pick_next_task_fair(struct rq *rq)
4114 struct task_struct *p;
4115 struct cfs_rq *cfs_rq = &rq->cfs;
4116 struct sched_entity *se;
4118 if (!cfs_rq->nr_running)
4122 se = pick_next_entity(cfs_rq);
4123 set_next_entity(cfs_rq, se);
4124 cfs_rq = group_cfs_rq(se);
4128 if (hrtick_enabled(rq))
4129 hrtick_start_fair(rq, p);
4135 * Account for a descheduled task:
4137 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4139 struct sched_entity *se = &prev->se;
4140 struct cfs_rq *cfs_rq;
4142 for_each_sched_entity(se) {
4143 cfs_rq = cfs_rq_of(se);
4144 put_prev_entity(cfs_rq, se);
4149 * sched_yield() is very simple
4151 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4153 static void yield_task_fair(struct rq *rq)
4155 struct task_struct *curr = rq->curr;
4156 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4157 struct sched_entity *se = &curr->se;
4160 * Are we the only task in the tree?
4162 if (unlikely(rq->nr_running == 1))
4165 clear_buddies(cfs_rq, se);
4167 if (curr->policy != SCHED_BATCH) {
4168 update_rq_clock(rq);
4170 * Update run-time statistics of the 'current'.
4172 update_curr(cfs_rq);
4174 * Tell update_rq_clock() that we've just updated,
4175 * so we don't do microscopic update in schedule()
4176 * and double the fastpath cost.
4178 rq->skip_clock_update = 1;
4184 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4186 struct sched_entity *se = &p->se;
4188 /* throttled hierarchies are not runnable */
4189 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4192 /* Tell the scheduler that we'd really like pse to run next. */
4195 yield_task_fair(rq);
4201 /**************************************************
4202 * Fair scheduling class load-balancing methods.
4206 * The purpose of load-balancing is to achieve the same basic fairness the
4207 * per-cpu scheduler provides, namely provide a proportional amount of compute
4208 * time to each task. This is expressed in the following equation:
4210 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4212 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4213 * W_i,0 is defined as:
4215 * W_i,0 = \Sum_j w_i,j (2)
4217 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4218 * is derived from the nice value as per prio_to_weight[].
4220 * The weight average is an exponential decay average of the instantaneous
4223 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4225 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4226 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4227 * can also include other factors [XXX].
4229 * To achieve this balance we define a measure of imbalance which follows
4230 * directly from (1):
4232 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4234 * We them move tasks around to minimize the imbalance. In the continuous
4235 * function space it is obvious this converges, in the discrete case we get
4236 * a few fun cases generally called infeasible weight scenarios.
4239 * - infeasible weights;
4240 * - local vs global optima in the discrete case. ]
4245 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4246 * for all i,j solution, we create a tree of cpus that follows the hardware
4247 * topology where each level pairs two lower groups (or better). This results
4248 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4249 * tree to only the first of the previous level and we decrease the frequency
4250 * of load-balance at each level inv. proportional to the number of cpus in
4256 * \Sum { --- * --- * 2^i } = O(n) (5)
4258 * `- size of each group
4259 * | | `- number of cpus doing load-balance
4261 * `- sum over all levels
4263 * Coupled with a limit on how many tasks we can migrate every balance pass,
4264 * this makes (5) the runtime complexity of the balancer.
4266 * An important property here is that each CPU is still (indirectly) connected
4267 * to every other cpu in at most O(log n) steps:
4269 * The adjacency matrix of the resulting graph is given by:
4272 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4275 * And you'll find that:
4277 * A^(log_2 n)_i,j != 0 for all i,j (7)
4279 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4280 * The task movement gives a factor of O(m), giving a convergence complexity
4283 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4288 * In order to avoid CPUs going idle while there's still work to do, new idle
4289 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4290 * tree itself instead of relying on other CPUs to bring it work.
4292 * This adds some complexity to both (5) and (8) but it reduces the total idle
4300 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4303 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4308 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4310 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4312 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4315 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4316 * rewrite all of this once again.]
4319 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4321 #define LBF_ALL_PINNED 0x01
4322 #define LBF_NEED_BREAK 0x02
4323 #define LBF_DST_PINNED 0x04
4324 #define LBF_SOME_PINNED 0x08
4327 struct sched_domain *sd;
4335 struct cpumask *dst_grpmask;
4337 enum cpu_idle_type idle;
4339 /* The set of CPUs under consideration for load-balancing */
4340 struct cpumask *cpus;
4345 unsigned int loop_break;
4346 unsigned int loop_max;
4350 * move_task - move a task from one runqueue to another runqueue.
4351 * Both runqueues must be locked.
4353 static void move_task(struct task_struct *p, struct lb_env *env)
4355 deactivate_task(env->src_rq, p, 0);
4356 set_task_cpu(p, env->dst_cpu);
4357 activate_task(env->dst_rq, p, 0);
4358 check_preempt_curr(env->dst_rq, p, 0);
4359 #ifdef CONFIG_NUMA_BALANCING
4360 if (p->numa_preferred_nid != -1) {
4361 int src_nid = cpu_to_node(env->src_cpu);
4362 int dst_nid = cpu_to_node(env->dst_cpu);
4365 * If the load balancer has moved the task then limit
4366 * migrations from taking place in the short term in
4367 * case this is a short-lived migration.
4369 if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid)
4370 p->numa_migrate_seq = 0;
4376 * Is this task likely cache-hot:
4379 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4383 if (p->sched_class != &fair_sched_class)
4386 if (unlikely(p->policy == SCHED_IDLE))
4390 * Buddy candidates are cache hot:
4392 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4393 (&p->se == cfs_rq_of(&p->se)->next ||
4394 &p->se == cfs_rq_of(&p->se)->last))
4397 if (sysctl_sched_migration_cost == -1)
4399 if (sysctl_sched_migration_cost == 0)
4402 delta = now - p->se.exec_start;
4404 return delta < (s64)sysctl_sched_migration_cost;
4407 #ifdef CONFIG_NUMA_BALANCING
4408 /* Returns true if the destination node has incurred more faults */
4409 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4411 int src_nid, dst_nid;
4413 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4414 !(env->sd->flags & SD_NUMA)) {
4418 src_nid = cpu_to_node(env->src_cpu);
4419 dst_nid = cpu_to_node(env->dst_cpu);
4421 if (src_nid == dst_nid ||
4422 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4425 if (dst_nid == p->numa_preferred_nid ||
4426 task_faults(p, dst_nid) > task_faults(p, src_nid))
4433 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4435 int src_nid, dst_nid;
4437 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4440 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4443 src_nid = cpu_to_node(env->src_cpu);
4444 dst_nid = cpu_to_node(env->dst_cpu);
4446 if (src_nid == dst_nid ||
4447 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4450 if (task_faults(p, dst_nid) < task_faults(p, src_nid))
4457 static inline bool migrate_improves_locality(struct task_struct *p,
4463 static inline bool migrate_degrades_locality(struct task_struct *p,
4471 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4474 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4476 int tsk_cache_hot = 0;
4478 * We do not migrate tasks that are:
4479 * 1) throttled_lb_pair, or
4480 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4481 * 3) running (obviously), or
4482 * 4) are cache-hot on their current CPU.
4484 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4487 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4490 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4492 env->flags |= LBF_SOME_PINNED;
4495 * Remember if this task can be migrated to any other cpu in
4496 * our sched_group. We may want to revisit it if we couldn't
4497 * meet load balance goals by pulling other tasks on src_cpu.
4499 * Also avoid computing new_dst_cpu if we have already computed
4500 * one in current iteration.
4502 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4505 /* Prevent to re-select dst_cpu via env's cpus */
4506 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4507 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4508 env->flags |= LBF_DST_PINNED;
4509 env->new_dst_cpu = cpu;
4517 /* Record that we found atleast one task that could run on dst_cpu */
4518 env->flags &= ~LBF_ALL_PINNED;
4520 if (task_running(env->src_rq, p)) {
4521 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4526 * Aggressive migration if:
4527 * 1) destination numa is preferred
4528 * 2) task is cache cold, or
4529 * 3) too many balance attempts have failed.
4531 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4533 tsk_cache_hot = migrate_degrades_locality(p, env);
4535 if (migrate_improves_locality(p, env)) {
4536 #ifdef CONFIG_SCHEDSTATS
4537 if (tsk_cache_hot) {
4538 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4539 schedstat_inc(p, se.statistics.nr_forced_migrations);
4545 if (!tsk_cache_hot ||
4546 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4548 if (tsk_cache_hot) {
4549 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4550 schedstat_inc(p, se.statistics.nr_forced_migrations);
4556 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4561 * move_one_task tries to move exactly one task from busiest to this_rq, as
4562 * part of active balancing operations within "domain".
4563 * Returns 1 if successful and 0 otherwise.
4565 * Called with both runqueues locked.
4567 static int move_one_task(struct lb_env *env)
4569 struct task_struct *p, *n;
4571 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4572 if (!can_migrate_task(p, env))
4577 * Right now, this is only the second place move_task()
4578 * is called, so we can safely collect move_task()
4579 * stats here rather than inside move_task().
4581 schedstat_inc(env->sd, lb_gained[env->idle]);
4587 static const unsigned int sched_nr_migrate_break = 32;
4590 * move_tasks tries to move up to imbalance weighted load from busiest to
4591 * this_rq, as part of a balancing operation within domain "sd".
4592 * Returns 1 if successful and 0 otherwise.
4594 * Called with both runqueues locked.
4596 static int move_tasks(struct lb_env *env)
4598 struct list_head *tasks = &env->src_rq->cfs_tasks;
4599 struct task_struct *p;
4603 if (env->imbalance <= 0)
4606 while (!list_empty(tasks)) {
4607 p = list_first_entry(tasks, struct task_struct, se.group_node);
4610 /* We've more or less seen every task there is, call it quits */
4611 if (env->loop > env->loop_max)
4614 /* take a breather every nr_migrate tasks */
4615 if (env->loop > env->loop_break) {
4616 env->loop_break += sched_nr_migrate_break;
4617 env->flags |= LBF_NEED_BREAK;
4621 if (!can_migrate_task(p, env))
4624 load = task_h_load(p);
4626 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4629 if ((load / 2) > env->imbalance)
4634 env->imbalance -= load;
4636 #ifdef CONFIG_PREEMPT
4638 * NEWIDLE balancing is a source of latency, so preemptible
4639 * kernels will stop after the first task is pulled to minimize
4640 * the critical section.
4642 if (env->idle == CPU_NEWLY_IDLE)
4647 * We only want to steal up to the prescribed amount of
4650 if (env->imbalance <= 0)
4655 list_move_tail(&p->se.group_node, tasks);
4659 * Right now, this is one of only two places move_task() is called,
4660 * so we can safely collect move_task() stats here rather than
4661 * inside move_task().
4663 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4668 #ifdef CONFIG_FAIR_GROUP_SCHED
4670 * update tg->load_weight by folding this cpu's load_avg
4672 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4674 struct sched_entity *se = tg->se[cpu];
4675 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4677 /* throttled entities do not contribute to load */
4678 if (throttled_hierarchy(cfs_rq))
4681 update_cfs_rq_blocked_load(cfs_rq, 1);
4684 update_entity_load_avg(se, 1);
4686 * We pivot on our runnable average having decayed to zero for
4687 * list removal. This generally implies that all our children
4688 * have also been removed (modulo rounding error or bandwidth
4689 * control); however, such cases are rare and we can fix these
4692 * TODO: fix up out-of-order children on enqueue.
4694 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4695 list_del_leaf_cfs_rq(cfs_rq);
4697 struct rq *rq = rq_of(cfs_rq);
4698 update_rq_runnable_avg(rq, rq->nr_running);
4702 static void update_blocked_averages(int cpu)
4704 struct rq *rq = cpu_rq(cpu);
4705 struct cfs_rq *cfs_rq;
4706 unsigned long flags;
4708 raw_spin_lock_irqsave(&rq->lock, flags);
4709 update_rq_clock(rq);
4711 * Iterates the task_group tree in a bottom up fashion, see
4712 * list_add_leaf_cfs_rq() for details.
4714 for_each_leaf_cfs_rq(rq, cfs_rq) {
4716 * Note: We may want to consider periodically releasing
4717 * rq->lock about these updates so that creating many task
4718 * groups does not result in continually extending hold time.
4720 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4723 raw_spin_unlock_irqrestore(&rq->lock, flags);
4727 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4728 * This needs to be done in a top-down fashion because the load of a child
4729 * group is a fraction of its parents load.
4731 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4733 struct rq *rq = rq_of(cfs_rq);
4734 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4735 unsigned long now = jiffies;
4738 if (cfs_rq->last_h_load_update == now)
4741 cfs_rq->h_load_next = NULL;
4742 for_each_sched_entity(se) {
4743 cfs_rq = cfs_rq_of(se);
4744 cfs_rq->h_load_next = se;
4745 if (cfs_rq->last_h_load_update == now)
4750 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4751 cfs_rq->last_h_load_update = now;
4754 while ((se = cfs_rq->h_load_next) != NULL) {
4755 load = cfs_rq->h_load;
4756 load = div64_ul(load * se->avg.load_avg_contrib,
4757 cfs_rq->runnable_load_avg + 1);
4758 cfs_rq = group_cfs_rq(se);
4759 cfs_rq->h_load = load;
4760 cfs_rq->last_h_load_update = now;
4764 static unsigned long task_h_load(struct task_struct *p)
4766 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4768 update_cfs_rq_h_load(cfs_rq);
4769 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4770 cfs_rq->runnable_load_avg + 1);
4773 static inline void update_blocked_averages(int cpu)
4777 static unsigned long task_h_load(struct task_struct *p)
4779 return p->se.avg.load_avg_contrib;
4783 /********** Helpers for find_busiest_group ************************/
4785 * sg_lb_stats - stats of a sched_group required for load_balancing
4787 struct sg_lb_stats {
4788 unsigned long avg_load; /*Avg load across the CPUs of the group */
4789 unsigned long group_load; /* Total load over the CPUs of the group */
4790 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4791 unsigned long load_per_task;
4792 unsigned long group_power;
4793 unsigned int sum_nr_running; /* Nr tasks running in the group */
4794 unsigned int group_capacity;
4795 unsigned int idle_cpus;
4796 unsigned int group_weight;
4797 int group_imb; /* Is there an imbalance in the group ? */
4798 int group_has_capacity; /* Is there extra capacity in the group? */
4802 * sd_lb_stats - Structure to store the statistics of a sched_domain
4803 * during load balancing.
4805 struct sd_lb_stats {
4806 struct sched_group *busiest; /* Busiest group in this sd */
4807 struct sched_group *local; /* Local group in this sd */
4808 unsigned long total_load; /* Total load of all groups in sd */
4809 unsigned long total_pwr; /* Total power of all groups in sd */
4810 unsigned long avg_load; /* Average load across all groups in sd */
4812 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4813 struct sg_lb_stats local_stat; /* Statistics of the local group */
4816 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4819 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4820 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4821 * We must however clear busiest_stat::avg_load because
4822 * update_sd_pick_busiest() reads this before assignment.
4824 *sds = (struct sd_lb_stats){
4836 * get_sd_load_idx - Obtain the load index for a given sched domain.
4837 * @sd: The sched_domain whose load_idx is to be obtained.
4838 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4840 * Return: The load index.
4842 static inline int get_sd_load_idx(struct sched_domain *sd,
4843 enum cpu_idle_type idle)
4849 load_idx = sd->busy_idx;
4852 case CPU_NEWLY_IDLE:
4853 load_idx = sd->newidle_idx;
4856 load_idx = sd->idle_idx;
4863 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4865 return SCHED_POWER_SCALE;
4868 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4870 return default_scale_freq_power(sd, cpu);
4873 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4875 unsigned long weight = sd->span_weight;
4876 unsigned long smt_gain = sd->smt_gain;
4883 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4885 return default_scale_smt_power(sd, cpu);
4888 static unsigned long scale_rt_power(int cpu)
4890 struct rq *rq = cpu_rq(cpu);
4891 u64 total, available, age_stamp, avg;
4894 * Since we're reading these variables without serialization make sure
4895 * we read them once before doing sanity checks on them.
4897 age_stamp = ACCESS_ONCE(rq->age_stamp);
4898 avg = ACCESS_ONCE(rq->rt_avg);
4900 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4902 if (unlikely(total < avg)) {
4903 /* Ensures that power won't end up being negative */
4906 available = total - avg;
4909 if (unlikely((s64)total < SCHED_POWER_SCALE))
4910 total = SCHED_POWER_SCALE;
4912 total >>= SCHED_POWER_SHIFT;
4914 return div_u64(available, total);
4917 static void update_cpu_power(struct sched_domain *sd, int cpu)
4919 unsigned long weight = sd->span_weight;
4920 unsigned long power = SCHED_POWER_SCALE;
4921 struct sched_group *sdg = sd->groups;
4923 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4924 if (sched_feat(ARCH_POWER))
4925 power *= arch_scale_smt_power(sd, cpu);
4927 power *= default_scale_smt_power(sd, cpu);
4929 power >>= SCHED_POWER_SHIFT;
4932 sdg->sgp->power_orig = power;
4934 if (sched_feat(ARCH_POWER))
4935 power *= arch_scale_freq_power(sd, cpu);
4937 power *= default_scale_freq_power(sd, cpu);
4939 power >>= SCHED_POWER_SHIFT;
4941 power *= scale_rt_power(cpu);
4942 power >>= SCHED_POWER_SHIFT;
4947 cpu_rq(cpu)->cpu_power = power;
4948 sdg->sgp->power = power;
4951 void update_group_power(struct sched_domain *sd, int cpu)
4953 struct sched_domain *child = sd->child;
4954 struct sched_group *group, *sdg = sd->groups;
4955 unsigned long power, power_orig;
4956 unsigned long interval;
4958 interval = msecs_to_jiffies(sd->balance_interval);
4959 interval = clamp(interval, 1UL, max_load_balance_interval);
4960 sdg->sgp->next_update = jiffies + interval;
4963 update_cpu_power(sd, cpu);
4967 power_orig = power = 0;
4969 if (child->flags & SD_OVERLAP) {
4971 * SD_OVERLAP domains cannot assume that child groups
4972 * span the current group.
4975 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4976 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4978 power_orig += sg->sgp->power_orig;
4979 power += sg->sgp->power;
4983 * !SD_OVERLAP domains can assume that child groups
4984 * span the current group.
4987 group = child->groups;
4989 power_orig += group->sgp->power_orig;
4990 power += group->sgp->power;
4991 group = group->next;
4992 } while (group != child->groups);
4995 sdg->sgp->power_orig = power_orig;
4996 sdg->sgp->power = power;
5000 * Try and fix up capacity for tiny siblings, this is needed when
5001 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5002 * which on its own isn't powerful enough.
5004 * See update_sd_pick_busiest() and check_asym_packing().
5007 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5010 * Only siblings can have significantly less than SCHED_POWER_SCALE
5012 if (!(sd->flags & SD_SHARE_CPUPOWER))
5016 * If ~90% of the cpu_power is still there, we're good.
5018 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5025 * Group imbalance indicates (and tries to solve) the problem where balancing
5026 * groups is inadequate due to tsk_cpus_allowed() constraints.
5028 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5029 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5032 * { 0 1 2 3 } { 4 5 6 7 }
5035 * If we were to balance group-wise we'd place two tasks in the first group and
5036 * two tasks in the second group. Clearly this is undesired as it will overload
5037 * cpu 3 and leave one of the cpus in the second group unused.
5039 * The current solution to this issue is detecting the skew in the first group
5040 * by noticing the lower domain failed to reach balance and had difficulty
5041 * moving tasks due to affinity constraints.
5043 * When this is so detected; this group becomes a candidate for busiest; see
5044 * update_sd_pick_busiest(). And calculcate_imbalance() and
5045 * find_busiest_group() avoid some of the usual balance conditions to allow it
5046 * to create an effective group imbalance.
5048 * This is a somewhat tricky proposition since the next run might not find the
5049 * group imbalance and decide the groups need to be balanced again. A most
5050 * subtle and fragile situation.
5053 static inline int sg_imbalanced(struct sched_group *group)
5055 return group->sgp->imbalance;
5059 * Compute the group capacity.
5061 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5062 * first dividing out the smt factor and computing the actual number of cores
5063 * and limit power unit capacity with that.
5065 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5067 unsigned int capacity, smt, cpus;
5068 unsigned int power, power_orig;
5070 power = group->sgp->power;
5071 power_orig = group->sgp->power_orig;
5072 cpus = group->group_weight;
5074 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5075 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5076 capacity = cpus / smt; /* cores */
5078 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5080 capacity = fix_small_capacity(env->sd, group);
5086 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5087 * @env: The load balancing environment.
5088 * @group: sched_group whose statistics are to be updated.
5089 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5090 * @local_group: Does group contain this_cpu.
5091 * @sgs: variable to hold the statistics for this group.
5093 static inline void update_sg_lb_stats(struct lb_env *env,
5094 struct sched_group *group, int load_idx,
5095 int local_group, struct sg_lb_stats *sgs)
5097 unsigned long nr_running;
5101 memset(sgs, 0, sizeof(*sgs));
5103 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5104 struct rq *rq = cpu_rq(i);
5106 nr_running = rq->nr_running;
5108 /* Bias balancing toward cpus of our domain */
5110 load = target_load(i, load_idx);
5112 load = source_load(i, load_idx);
5114 sgs->group_load += load;
5115 sgs->sum_nr_running += nr_running;
5116 sgs->sum_weighted_load += weighted_cpuload(i);
5121 /* Adjust by relative CPU power of the group */
5122 sgs->group_power = group->sgp->power;
5123 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5125 if (sgs->sum_nr_running)
5126 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5128 sgs->group_weight = group->group_weight;
5130 sgs->group_imb = sg_imbalanced(group);
5131 sgs->group_capacity = sg_capacity(env, group);
5133 if (sgs->group_capacity > sgs->sum_nr_running)
5134 sgs->group_has_capacity = 1;
5138 * update_sd_pick_busiest - return 1 on busiest group
5139 * @env: The load balancing environment.
5140 * @sds: sched_domain statistics
5141 * @sg: sched_group candidate to be checked for being the busiest
5142 * @sgs: sched_group statistics
5144 * Determine if @sg is a busier group than the previously selected
5147 * Return: %true if @sg is a busier group than the previously selected
5148 * busiest group. %false otherwise.
5150 static bool update_sd_pick_busiest(struct lb_env *env,
5151 struct sd_lb_stats *sds,
5152 struct sched_group *sg,
5153 struct sg_lb_stats *sgs)
5155 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5158 if (sgs->sum_nr_running > sgs->group_capacity)
5165 * ASYM_PACKING needs to move all the work to the lowest
5166 * numbered CPUs in the group, therefore mark all groups
5167 * higher than ourself as busy.
5169 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5170 env->dst_cpu < group_first_cpu(sg)) {
5174 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5182 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5183 * @env: The load balancing environment.
5184 * @balance: Should we balance.
5185 * @sds: variable to hold the statistics for this sched_domain.
5187 static inline void update_sd_lb_stats(struct lb_env *env,
5188 struct sd_lb_stats *sds)
5190 struct sched_domain *child = env->sd->child;
5191 struct sched_group *sg = env->sd->groups;
5192 struct sg_lb_stats tmp_sgs;
5193 int load_idx, prefer_sibling = 0;
5195 if (child && child->flags & SD_PREFER_SIBLING)
5198 load_idx = get_sd_load_idx(env->sd, env->idle);
5201 struct sg_lb_stats *sgs = &tmp_sgs;
5204 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5207 sgs = &sds->local_stat;
5209 if (env->idle != CPU_NEWLY_IDLE ||
5210 time_after_eq(jiffies, sg->sgp->next_update))
5211 update_group_power(env->sd, env->dst_cpu);
5214 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5220 * In case the child domain prefers tasks go to siblings
5221 * first, lower the sg capacity to one so that we'll try
5222 * and move all the excess tasks away. We lower the capacity
5223 * of a group only if the local group has the capacity to fit
5224 * these excess tasks, i.e. nr_running < group_capacity. The
5225 * extra check prevents the case where you always pull from the
5226 * heaviest group when it is already under-utilized (possible
5227 * with a large weight task outweighs the tasks on the system).
5229 if (prefer_sibling && sds->local &&
5230 sds->local_stat.group_has_capacity)
5231 sgs->group_capacity = min(sgs->group_capacity, 1U);
5233 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5235 sds->busiest_stat = *sgs;
5239 /* Now, start updating sd_lb_stats */
5240 sds->total_load += sgs->group_load;
5241 sds->total_pwr += sgs->group_power;
5244 } while (sg != env->sd->groups);
5248 * check_asym_packing - Check to see if the group is packed into the
5251 * This is primarily intended to used at the sibling level. Some
5252 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5253 * case of POWER7, it can move to lower SMT modes only when higher
5254 * threads are idle. When in lower SMT modes, the threads will
5255 * perform better since they share less core resources. Hence when we
5256 * have idle threads, we want them to be the higher ones.
5258 * This packing function is run on idle threads. It checks to see if
5259 * the busiest CPU in this domain (core in the P7 case) has a higher
5260 * CPU number than the packing function is being run on. Here we are
5261 * assuming lower CPU number will be equivalent to lower a SMT thread
5264 * Return: 1 when packing is required and a task should be moved to
5265 * this CPU. The amount of the imbalance is returned in *imbalance.
5267 * @env: The load balancing environment.
5268 * @sds: Statistics of the sched_domain which is to be packed
5270 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5274 if (!(env->sd->flags & SD_ASYM_PACKING))
5280 busiest_cpu = group_first_cpu(sds->busiest);
5281 if (env->dst_cpu > busiest_cpu)
5284 env->imbalance = DIV_ROUND_CLOSEST(
5285 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5292 * fix_small_imbalance - Calculate the minor imbalance that exists
5293 * amongst the groups of a sched_domain, during
5295 * @env: The load balancing environment.
5296 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5299 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5301 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5302 unsigned int imbn = 2;
5303 unsigned long scaled_busy_load_per_task;
5304 struct sg_lb_stats *local, *busiest;
5306 local = &sds->local_stat;
5307 busiest = &sds->busiest_stat;
5309 if (!local->sum_nr_running)
5310 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5311 else if (busiest->load_per_task > local->load_per_task)
5314 scaled_busy_load_per_task =
5315 (busiest->load_per_task * SCHED_POWER_SCALE) /
5316 busiest->group_power;
5318 if (busiest->avg_load + scaled_busy_load_per_task >=
5319 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5320 env->imbalance = busiest->load_per_task;
5325 * OK, we don't have enough imbalance to justify moving tasks,
5326 * however we may be able to increase total CPU power used by
5330 pwr_now += busiest->group_power *
5331 min(busiest->load_per_task, busiest->avg_load);
5332 pwr_now += local->group_power *
5333 min(local->load_per_task, local->avg_load);
5334 pwr_now /= SCHED_POWER_SCALE;
5336 /* Amount of load we'd subtract */
5337 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5338 busiest->group_power;
5339 if (busiest->avg_load > tmp) {
5340 pwr_move += busiest->group_power *
5341 min(busiest->load_per_task,
5342 busiest->avg_load - tmp);
5345 /* Amount of load we'd add */
5346 if (busiest->avg_load * busiest->group_power <
5347 busiest->load_per_task * SCHED_POWER_SCALE) {
5348 tmp = (busiest->avg_load * busiest->group_power) /
5351 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5354 pwr_move += local->group_power *
5355 min(local->load_per_task, local->avg_load + tmp);
5356 pwr_move /= SCHED_POWER_SCALE;
5358 /* Move if we gain throughput */
5359 if (pwr_move > pwr_now)
5360 env->imbalance = busiest->load_per_task;
5364 * calculate_imbalance - Calculate the amount of imbalance present within the
5365 * groups of a given sched_domain during load balance.
5366 * @env: load balance environment
5367 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5369 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5371 unsigned long max_pull, load_above_capacity = ~0UL;
5372 struct sg_lb_stats *local, *busiest;
5374 local = &sds->local_stat;
5375 busiest = &sds->busiest_stat;
5377 if (busiest->group_imb) {
5379 * In the group_imb case we cannot rely on group-wide averages
5380 * to ensure cpu-load equilibrium, look at wider averages. XXX
5382 busiest->load_per_task =
5383 min(busiest->load_per_task, sds->avg_load);
5387 * In the presence of smp nice balancing, certain scenarios can have
5388 * max load less than avg load(as we skip the groups at or below
5389 * its cpu_power, while calculating max_load..)
5391 if (busiest->avg_load <= sds->avg_load ||
5392 local->avg_load >= sds->avg_load) {
5394 return fix_small_imbalance(env, sds);
5397 if (!busiest->group_imb) {
5399 * Don't want to pull so many tasks that a group would go idle.
5400 * Except of course for the group_imb case, since then we might
5401 * have to drop below capacity to reach cpu-load equilibrium.
5403 load_above_capacity =
5404 (busiest->sum_nr_running - busiest->group_capacity);
5406 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5407 load_above_capacity /= busiest->group_power;
5411 * We're trying to get all the cpus to the average_load, so we don't
5412 * want to push ourselves above the average load, nor do we wish to
5413 * reduce the max loaded cpu below the average load. At the same time,
5414 * we also don't want to reduce the group load below the group capacity
5415 * (so that we can implement power-savings policies etc). Thus we look
5416 * for the minimum possible imbalance.
5418 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5420 /* How much load to actually move to equalise the imbalance */
5421 env->imbalance = min(
5422 max_pull * busiest->group_power,
5423 (sds->avg_load - local->avg_load) * local->group_power
5424 ) / SCHED_POWER_SCALE;
5427 * if *imbalance is less than the average load per runnable task
5428 * there is no guarantee that any tasks will be moved so we'll have
5429 * a think about bumping its value to force at least one task to be
5432 if (env->imbalance < busiest->load_per_task)
5433 return fix_small_imbalance(env, sds);
5436 /******* find_busiest_group() helpers end here *********************/
5439 * find_busiest_group - Returns the busiest group within the sched_domain
5440 * if there is an imbalance. If there isn't an imbalance, and
5441 * the user has opted for power-savings, it returns a group whose
5442 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5443 * such a group exists.
5445 * Also calculates the amount of weighted load which should be moved
5446 * to restore balance.
5448 * @env: The load balancing environment.
5450 * Return: - The busiest group if imbalance exists.
5451 * - If no imbalance and user has opted for power-savings balance,
5452 * return the least loaded group whose CPUs can be
5453 * put to idle by rebalancing its tasks onto our group.
5455 static struct sched_group *find_busiest_group(struct lb_env *env)
5457 struct sg_lb_stats *local, *busiest;
5458 struct sd_lb_stats sds;
5460 init_sd_lb_stats(&sds);
5463 * Compute the various statistics relavent for load balancing at
5466 update_sd_lb_stats(env, &sds);
5467 local = &sds.local_stat;
5468 busiest = &sds.busiest_stat;
5470 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5471 check_asym_packing(env, &sds))
5474 /* There is no busy sibling group to pull tasks from */
5475 if (!sds.busiest || busiest->sum_nr_running == 0)
5478 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5481 * If the busiest group is imbalanced the below checks don't
5482 * work because they assume all things are equal, which typically
5483 * isn't true due to cpus_allowed constraints and the like.
5485 if (busiest->group_imb)
5488 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5489 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5490 !busiest->group_has_capacity)
5494 * If the local group is more busy than the selected busiest group
5495 * don't try and pull any tasks.
5497 if (local->avg_load >= busiest->avg_load)
5501 * Don't pull any tasks if this group is already above the domain
5504 if (local->avg_load >= sds.avg_load)
5507 if (env->idle == CPU_IDLE) {
5509 * This cpu is idle. If the busiest group load doesn't
5510 * have more tasks than the number of available cpu's and
5511 * there is no imbalance between this and busiest group
5512 * wrt to idle cpu's, it is balanced.
5514 if ((local->idle_cpus < busiest->idle_cpus) &&
5515 busiest->sum_nr_running <= busiest->group_weight)
5519 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5520 * imbalance_pct to be conservative.
5522 if (100 * busiest->avg_load <=
5523 env->sd->imbalance_pct * local->avg_load)
5528 /* Looks like there is an imbalance. Compute it */
5529 calculate_imbalance(env, &sds);
5538 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5540 static struct rq *find_busiest_queue(struct lb_env *env,
5541 struct sched_group *group)
5543 struct rq *busiest = NULL, *rq;
5544 unsigned long busiest_load = 0, busiest_power = 1;
5547 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5548 unsigned long power = power_of(i);
5549 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5554 capacity = fix_small_capacity(env->sd, group);
5557 wl = weighted_cpuload(i);
5560 * When comparing with imbalance, use weighted_cpuload()
5561 * which is not scaled with the cpu power.
5563 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5567 * For the load comparisons with the other cpu's, consider
5568 * the weighted_cpuload() scaled with the cpu power, so that
5569 * the load can be moved away from the cpu that is potentially
5570 * running at a lower capacity.
5572 * Thus we're looking for max(wl_i / power_i), crosswise
5573 * multiplication to rid ourselves of the division works out
5574 * to: wl_i * power_j > wl_j * power_i; where j is our
5577 if (wl * busiest_power > busiest_load * power) {
5579 busiest_power = power;
5588 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5589 * so long as it is large enough.
5591 #define MAX_PINNED_INTERVAL 512
5593 /* Working cpumask for load_balance and load_balance_newidle. */
5594 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5596 static int need_active_balance(struct lb_env *env)
5598 struct sched_domain *sd = env->sd;
5600 if (env->idle == CPU_NEWLY_IDLE) {
5603 * ASYM_PACKING needs to force migrate tasks from busy but
5604 * higher numbered CPUs in order to pack all tasks in the
5605 * lowest numbered CPUs.
5607 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5611 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5614 static int active_load_balance_cpu_stop(void *data);
5616 static int should_we_balance(struct lb_env *env)
5618 struct sched_group *sg = env->sd->groups;
5619 struct cpumask *sg_cpus, *sg_mask;
5620 int cpu, balance_cpu = -1;
5623 * In the newly idle case, we will allow all the cpu's
5624 * to do the newly idle load balance.
5626 if (env->idle == CPU_NEWLY_IDLE)
5629 sg_cpus = sched_group_cpus(sg);
5630 sg_mask = sched_group_mask(sg);
5631 /* Try to find first idle cpu */
5632 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5633 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5640 if (balance_cpu == -1)
5641 balance_cpu = group_balance_cpu(sg);
5644 * First idle cpu or the first cpu(busiest) in this sched group
5645 * is eligible for doing load balancing at this and above domains.
5647 return balance_cpu == env->dst_cpu;
5651 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5652 * tasks if there is an imbalance.
5654 static int load_balance(int this_cpu, struct rq *this_rq,
5655 struct sched_domain *sd, enum cpu_idle_type idle,
5656 int *continue_balancing)
5658 int ld_moved, cur_ld_moved, active_balance = 0;
5659 struct sched_domain *sd_parent = sd->parent;
5660 struct sched_group *group;
5662 unsigned long flags;
5663 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5665 struct lb_env env = {
5667 .dst_cpu = this_cpu,
5669 .dst_grpmask = sched_group_cpus(sd->groups),
5671 .loop_break = sched_nr_migrate_break,
5676 * For NEWLY_IDLE load_balancing, we don't need to consider
5677 * other cpus in our group
5679 if (idle == CPU_NEWLY_IDLE)
5680 env.dst_grpmask = NULL;
5682 cpumask_copy(cpus, cpu_active_mask);
5684 schedstat_inc(sd, lb_count[idle]);
5687 if (!should_we_balance(&env)) {
5688 *continue_balancing = 0;
5692 group = find_busiest_group(&env);
5694 schedstat_inc(sd, lb_nobusyg[idle]);
5698 busiest = find_busiest_queue(&env, group);
5700 schedstat_inc(sd, lb_nobusyq[idle]);
5704 BUG_ON(busiest == env.dst_rq);
5706 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5709 if (busiest->nr_running > 1) {
5711 * Attempt to move tasks. If find_busiest_group has found
5712 * an imbalance but busiest->nr_running <= 1, the group is
5713 * still unbalanced. ld_moved simply stays zero, so it is
5714 * correctly treated as an imbalance.
5716 env.flags |= LBF_ALL_PINNED;
5717 env.src_cpu = busiest->cpu;
5718 env.src_rq = busiest;
5719 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5722 local_irq_save(flags);
5723 double_rq_lock(env.dst_rq, busiest);
5726 * cur_ld_moved - load moved in current iteration
5727 * ld_moved - cumulative load moved across iterations
5729 cur_ld_moved = move_tasks(&env);
5730 ld_moved += cur_ld_moved;
5731 double_rq_unlock(env.dst_rq, busiest);
5732 local_irq_restore(flags);
5735 * some other cpu did the load balance for us.
5737 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5738 resched_cpu(env.dst_cpu);
5740 if (env.flags & LBF_NEED_BREAK) {
5741 env.flags &= ~LBF_NEED_BREAK;
5746 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5747 * us and move them to an alternate dst_cpu in our sched_group
5748 * where they can run. The upper limit on how many times we
5749 * iterate on same src_cpu is dependent on number of cpus in our
5752 * This changes load balance semantics a bit on who can move
5753 * load to a given_cpu. In addition to the given_cpu itself
5754 * (or a ilb_cpu acting on its behalf where given_cpu is
5755 * nohz-idle), we now have balance_cpu in a position to move
5756 * load to given_cpu. In rare situations, this may cause
5757 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5758 * _independently_ and at _same_ time to move some load to
5759 * given_cpu) causing exceess load to be moved to given_cpu.
5760 * This however should not happen so much in practice and
5761 * moreover subsequent load balance cycles should correct the
5762 * excess load moved.
5764 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5766 /* Prevent to re-select dst_cpu via env's cpus */
5767 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5769 env.dst_rq = cpu_rq(env.new_dst_cpu);
5770 env.dst_cpu = env.new_dst_cpu;
5771 env.flags &= ~LBF_DST_PINNED;
5773 env.loop_break = sched_nr_migrate_break;
5776 * Go back to "more_balance" rather than "redo" since we
5777 * need to continue with same src_cpu.
5783 * We failed to reach balance because of affinity.
5786 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5788 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5789 *group_imbalance = 1;
5790 } else if (*group_imbalance)
5791 *group_imbalance = 0;
5794 /* All tasks on this runqueue were pinned by CPU affinity */
5795 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5796 cpumask_clear_cpu(cpu_of(busiest), cpus);
5797 if (!cpumask_empty(cpus)) {
5799 env.loop_break = sched_nr_migrate_break;
5807 schedstat_inc(sd, lb_failed[idle]);
5809 * Increment the failure counter only on periodic balance.
5810 * We do not want newidle balance, which can be very
5811 * frequent, pollute the failure counter causing
5812 * excessive cache_hot migrations and active balances.
5814 if (idle != CPU_NEWLY_IDLE)
5815 sd->nr_balance_failed++;
5817 if (need_active_balance(&env)) {
5818 raw_spin_lock_irqsave(&busiest->lock, flags);
5820 /* don't kick the active_load_balance_cpu_stop,
5821 * if the curr task on busiest cpu can't be
5824 if (!cpumask_test_cpu(this_cpu,
5825 tsk_cpus_allowed(busiest->curr))) {
5826 raw_spin_unlock_irqrestore(&busiest->lock,
5828 env.flags |= LBF_ALL_PINNED;
5829 goto out_one_pinned;
5833 * ->active_balance synchronizes accesses to
5834 * ->active_balance_work. Once set, it's cleared
5835 * only after active load balance is finished.
5837 if (!busiest->active_balance) {
5838 busiest->active_balance = 1;
5839 busiest->push_cpu = this_cpu;
5842 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5844 if (active_balance) {
5845 stop_one_cpu_nowait(cpu_of(busiest),
5846 active_load_balance_cpu_stop, busiest,
5847 &busiest->active_balance_work);
5851 * We've kicked active balancing, reset the failure
5854 sd->nr_balance_failed = sd->cache_nice_tries+1;
5857 sd->nr_balance_failed = 0;
5859 if (likely(!active_balance)) {
5860 /* We were unbalanced, so reset the balancing interval */
5861 sd->balance_interval = sd->min_interval;
5864 * If we've begun active balancing, start to back off. This
5865 * case may not be covered by the all_pinned logic if there
5866 * is only 1 task on the busy runqueue (because we don't call
5869 if (sd->balance_interval < sd->max_interval)
5870 sd->balance_interval *= 2;
5876 schedstat_inc(sd, lb_balanced[idle]);
5878 sd->nr_balance_failed = 0;
5881 /* tune up the balancing interval */
5882 if (((env.flags & LBF_ALL_PINNED) &&
5883 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5884 (sd->balance_interval < sd->max_interval))
5885 sd->balance_interval *= 2;
5893 * idle_balance is called by schedule() if this_cpu is about to become
5894 * idle. Attempts to pull tasks from other CPUs.
5896 void idle_balance(int this_cpu, struct rq *this_rq)
5898 struct sched_domain *sd;
5899 int pulled_task = 0;
5900 unsigned long next_balance = jiffies + HZ;
5903 this_rq->idle_stamp = rq_clock(this_rq);
5905 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5909 * Drop the rq->lock, but keep IRQ/preempt disabled.
5911 raw_spin_unlock(&this_rq->lock);
5913 update_blocked_averages(this_cpu);
5915 for_each_domain(this_cpu, sd) {
5916 unsigned long interval;
5917 int continue_balancing = 1;
5918 u64 t0, domain_cost;
5920 if (!(sd->flags & SD_LOAD_BALANCE))
5923 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5926 if (sd->flags & SD_BALANCE_NEWIDLE) {
5927 t0 = sched_clock_cpu(this_cpu);
5929 /* If we've pulled tasks over stop searching: */
5930 pulled_task = load_balance(this_cpu, this_rq,
5932 &continue_balancing);
5934 domain_cost = sched_clock_cpu(this_cpu) - t0;
5935 if (domain_cost > sd->max_newidle_lb_cost)
5936 sd->max_newidle_lb_cost = domain_cost;
5938 curr_cost += domain_cost;
5941 interval = msecs_to_jiffies(sd->balance_interval);
5942 if (time_after(next_balance, sd->last_balance + interval))
5943 next_balance = sd->last_balance + interval;
5945 this_rq->idle_stamp = 0;
5951 raw_spin_lock(&this_rq->lock);
5953 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5955 * We are going idle. next_balance may be set based on
5956 * a busy processor. So reset next_balance.
5958 this_rq->next_balance = next_balance;
5961 if (curr_cost > this_rq->max_idle_balance_cost)
5962 this_rq->max_idle_balance_cost = curr_cost;
5966 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5967 * running tasks off the busiest CPU onto idle CPUs. It requires at
5968 * least 1 task to be running on each physical CPU where possible, and
5969 * avoids physical / logical imbalances.
5971 static int active_load_balance_cpu_stop(void *data)
5973 struct rq *busiest_rq = data;
5974 int busiest_cpu = cpu_of(busiest_rq);
5975 int target_cpu = busiest_rq->push_cpu;
5976 struct rq *target_rq = cpu_rq(target_cpu);
5977 struct sched_domain *sd;
5979 raw_spin_lock_irq(&busiest_rq->lock);
5981 /* make sure the requested cpu hasn't gone down in the meantime */
5982 if (unlikely(busiest_cpu != smp_processor_id() ||
5983 !busiest_rq->active_balance))
5986 /* Is there any task to move? */
5987 if (busiest_rq->nr_running <= 1)
5991 * This condition is "impossible", if it occurs
5992 * we need to fix it. Originally reported by
5993 * Bjorn Helgaas on a 128-cpu setup.
5995 BUG_ON(busiest_rq == target_rq);
5997 /* move a task from busiest_rq to target_rq */
5998 double_lock_balance(busiest_rq, target_rq);
6000 /* Search for an sd spanning us and the target CPU. */
6002 for_each_domain(target_cpu, sd) {
6003 if ((sd->flags & SD_LOAD_BALANCE) &&
6004 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6009 struct lb_env env = {
6011 .dst_cpu = target_cpu,
6012 .dst_rq = target_rq,
6013 .src_cpu = busiest_rq->cpu,
6014 .src_rq = busiest_rq,
6018 schedstat_inc(sd, alb_count);
6020 if (move_one_task(&env))
6021 schedstat_inc(sd, alb_pushed);
6023 schedstat_inc(sd, alb_failed);
6026 double_unlock_balance(busiest_rq, target_rq);
6028 busiest_rq->active_balance = 0;
6029 raw_spin_unlock_irq(&busiest_rq->lock);
6033 #ifdef CONFIG_NO_HZ_COMMON
6035 * idle load balancing details
6036 * - When one of the busy CPUs notice that there may be an idle rebalancing
6037 * needed, they will kick the idle load balancer, which then does idle
6038 * load balancing for all the idle CPUs.
6041 cpumask_var_t idle_cpus_mask;
6043 unsigned long next_balance; /* in jiffy units */
6044 } nohz ____cacheline_aligned;
6046 static inline int find_new_ilb(int call_cpu)
6048 int ilb = cpumask_first(nohz.idle_cpus_mask);
6050 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6057 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6058 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6059 * CPU (if there is one).
6061 static void nohz_balancer_kick(int cpu)
6065 nohz.next_balance++;
6067 ilb_cpu = find_new_ilb(cpu);
6069 if (ilb_cpu >= nr_cpu_ids)
6072 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6075 * Use smp_send_reschedule() instead of resched_cpu().
6076 * This way we generate a sched IPI on the target cpu which
6077 * is idle. And the softirq performing nohz idle load balance
6078 * will be run before returning from the IPI.
6080 smp_send_reschedule(ilb_cpu);
6084 static inline void nohz_balance_exit_idle(int cpu)
6086 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6087 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6088 atomic_dec(&nohz.nr_cpus);
6089 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6093 static inline void set_cpu_sd_state_busy(void)
6095 struct sched_domain *sd;
6098 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6100 if (!sd || !sd->nohz_idle)
6104 for (; sd; sd = sd->parent)
6105 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6110 void set_cpu_sd_state_idle(void)
6112 struct sched_domain *sd;
6115 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6117 if (!sd || sd->nohz_idle)
6121 for (; sd; sd = sd->parent)
6122 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6128 * This routine will record that the cpu is going idle with tick stopped.
6129 * This info will be used in performing idle load balancing in the future.
6131 void nohz_balance_enter_idle(int cpu)
6134 * If this cpu is going down, then nothing needs to be done.
6136 if (!cpu_active(cpu))
6139 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6142 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6143 atomic_inc(&nohz.nr_cpus);
6144 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6147 static int sched_ilb_notifier(struct notifier_block *nfb,
6148 unsigned long action, void *hcpu)
6150 switch (action & ~CPU_TASKS_FROZEN) {
6152 nohz_balance_exit_idle(smp_processor_id());
6160 static DEFINE_SPINLOCK(balancing);
6163 * Scale the max load_balance interval with the number of CPUs in the system.
6164 * This trades load-balance latency on larger machines for less cross talk.
6166 void update_max_interval(void)
6168 max_load_balance_interval = HZ*num_online_cpus()/10;
6172 * It checks each scheduling domain to see if it is due to be balanced,
6173 * and initiates a balancing operation if so.
6175 * Balancing parameters are set up in init_sched_domains.
6177 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6179 int continue_balancing = 1;
6180 struct rq *rq = cpu_rq(cpu);
6181 unsigned long interval;
6182 struct sched_domain *sd;
6183 /* Earliest time when we have to do rebalance again */
6184 unsigned long next_balance = jiffies + 60*HZ;
6185 int update_next_balance = 0;
6186 int need_serialize, need_decay = 0;
6189 update_blocked_averages(cpu);
6192 for_each_domain(cpu, sd) {
6194 * Decay the newidle max times here because this is a regular
6195 * visit to all the domains. Decay ~1% per second.
6197 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6198 sd->max_newidle_lb_cost =
6199 (sd->max_newidle_lb_cost * 253) / 256;
6200 sd->next_decay_max_lb_cost = jiffies + HZ;
6203 max_cost += sd->max_newidle_lb_cost;
6205 if (!(sd->flags & SD_LOAD_BALANCE))
6209 * Stop the load balance at this level. There is another
6210 * CPU in our sched group which is doing load balancing more
6213 if (!continue_balancing) {
6219 interval = sd->balance_interval;
6220 if (idle != CPU_IDLE)
6221 interval *= sd->busy_factor;
6223 /* scale ms to jiffies */
6224 interval = msecs_to_jiffies(interval);
6225 interval = clamp(interval, 1UL, max_load_balance_interval);
6227 need_serialize = sd->flags & SD_SERIALIZE;
6229 if (need_serialize) {
6230 if (!spin_trylock(&balancing))
6234 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6235 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6237 * The LBF_DST_PINNED logic could have changed
6238 * env->dst_cpu, so we can't know our idle
6239 * state even if we migrated tasks. Update it.
6241 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6243 sd->last_balance = jiffies;
6246 spin_unlock(&balancing);
6248 if (time_after(next_balance, sd->last_balance + interval)) {
6249 next_balance = sd->last_balance + interval;
6250 update_next_balance = 1;
6255 * Ensure the rq-wide value also decays but keep it at a
6256 * reasonable floor to avoid funnies with rq->avg_idle.
6258 rq->max_idle_balance_cost =
6259 max((u64)sysctl_sched_migration_cost, max_cost);
6264 * next_balance will be updated only when there is a need.
6265 * When the cpu is attached to null domain for ex, it will not be
6268 if (likely(update_next_balance))
6269 rq->next_balance = next_balance;
6272 #ifdef CONFIG_NO_HZ_COMMON
6274 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6275 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6277 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6279 struct rq *this_rq = cpu_rq(this_cpu);
6283 if (idle != CPU_IDLE ||
6284 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6287 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6288 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6292 * If this cpu gets work to do, stop the load balancing
6293 * work being done for other cpus. Next load
6294 * balancing owner will pick it up.
6299 rq = cpu_rq(balance_cpu);
6301 raw_spin_lock_irq(&rq->lock);
6302 update_rq_clock(rq);
6303 update_idle_cpu_load(rq);
6304 raw_spin_unlock_irq(&rq->lock);
6306 rebalance_domains(balance_cpu, CPU_IDLE);
6308 if (time_after(this_rq->next_balance, rq->next_balance))
6309 this_rq->next_balance = rq->next_balance;
6311 nohz.next_balance = this_rq->next_balance;
6313 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6317 * Current heuristic for kicking the idle load balancer in the presence
6318 * of an idle cpu is the system.
6319 * - This rq has more than one task.
6320 * - At any scheduler domain level, this cpu's scheduler group has multiple
6321 * busy cpu's exceeding the group's power.
6322 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6323 * domain span are idle.
6325 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6327 unsigned long now = jiffies;
6328 struct sched_domain *sd;
6330 if (unlikely(idle_cpu(cpu)))
6334 * We may be recently in ticked or tickless idle mode. At the first
6335 * busy tick after returning from idle, we will update the busy stats.
6337 set_cpu_sd_state_busy();
6338 nohz_balance_exit_idle(cpu);
6341 * None are in tickless mode and hence no need for NOHZ idle load
6344 if (likely(!atomic_read(&nohz.nr_cpus)))
6347 if (time_before(now, nohz.next_balance))
6350 if (rq->nr_running >= 2)
6354 for_each_domain(cpu, sd) {
6355 struct sched_group *sg = sd->groups;
6356 struct sched_group_power *sgp = sg->sgp;
6357 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6359 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6360 goto need_kick_unlock;
6362 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6363 && (cpumask_first_and(nohz.idle_cpus_mask,
6364 sched_domain_span(sd)) < cpu))
6365 goto need_kick_unlock;
6367 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6379 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6383 * run_rebalance_domains is triggered when needed from the scheduler tick.
6384 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6386 static void run_rebalance_domains(struct softirq_action *h)
6388 int this_cpu = smp_processor_id();
6389 struct rq *this_rq = cpu_rq(this_cpu);
6390 enum cpu_idle_type idle = this_rq->idle_balance ?
6391 CPU_IDLE : CPU_NOT_IDLE;
6393 rebalance_domains(this_cpu, idle);
6396 * If this cpu has a pending nohz_balance_kick, then do the
6397 * balancing on behalf of the other idle cpus whose ticks are
6400 nohz_idle_balance(this_cpu, idle);
6403 static inline int on_null_domain(int cpu)
6405 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6409 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6411 void trigger_load_balance(struct rq *rq, int cpu)
6413 /* Don't need to rebalance while attached to NULL domain */
6414 if (time_after_eq(jiffies, rq->next_balance) &&
6415 likely(!on_null_domain(cpu)))
6416 raise_softirq(SCHED_SOFTIRQ);
6417 #ifdef CONFIG_NO_HZ_COMMON
6418 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6419 nohz_balancer_kick(cpu);
6423 static void rq_online_fair(struct rq *rq)
6428 static void rq_offline_fair(struct rq *rq)
6432 /* Ensure any throttled groups are reachable by pick_next_task */
6433 unthrottle_offline_cfs_rqs(rq);
6436 #endif /* CONFIG_SMP */
6439 * scheduler tick hitting a task of our scheduling class:
6441 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6443 struct cfs_rq *cfs_rq;
6444 struct sched_entity *se = &curr->se;
6446 for_each_sched_entity(se) {
6447 cfs_rq = cfs_rq_of(se);
6448 entity_tick(cfs_rq, se, queued);
6451 if (numabalancing_enabled)
6452 task_tick_numa(rq, curr);
6454 update_rq_runnable_avg(rq, 1);
6458 * called on fork with the child task as argument from the parent's context
6459 * - child not yet on the tasklist
6460 * - preemption disabled
6462 static void task_fork_fair(struct task_struct *p)
6464 struct cfs_rq *cfs_rq;
6465 struct sched_entity *se = &p->se, *curr;
6466 int this_cpu = smp_processor_id();
6467 struct rq *rq = this_rq();
6468 unsigned long flags;
6470 raw_spin_lock_irqsave(&rq->lock, flags);
6472 update_rq_clock(rq);
6474 cfs_rq = task_cfs_rq(current);
6475 curr = cfs_rq->curr;
6478 * Not only the cpu but also the task_group of the parent might have
6479 * been changed after parent->se.parent,cfs_rq were copied to
6480 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6481 * of child point to valid ones.
6484 __set_task_cpu(p, this_cpu);
6487 update_curr(cfs_rq);
6490 se->vruntime = curr->vruntime;
6491 place_entity(cfs_rq, se, 1);
6493 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6495 * Upon rescheduling, sched_class::put_prev_task() will place
6496 * 'current' within the tree based on its new key value.
6498 swap(curr->vruntime, se->vruntime);
6499 resched_task(rq->curr);
6502 se->vruntime -= cfs_rq->min_vruntime;
6504 raw_spin_unlock_irqrestore(&rq->lock, flags);
6508 * Priority of the task has changed. Check to see if we preempt
6512 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6518 * Reschedule if we are currently running on this runqueue and
6519 * our priority decreased, or if we are not currently running on
6520 * this runqueue and our priority is higher than the current's
6522 if (rq->curr == p) {
6523 if (p->prio > oldprio)
6524 resched_task(rq->curr);
6526 check_preempt_curr(rq, p, 0);
6529 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6531 struct sched_entity *se = &p->se;
6532 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6535 * Ensure the task's vruntime is normalized, so that when its
6536 * switched back to the fair class the enqueue_entity(.flags=0) will
6537 * do the right thing.
6539 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6540 * have normalized the vruntime, if it was !on_rq, then only when
6541 * the task is sleeping will it still have non-normalized vruntime.
6543 if (!se->on_rq && p->state != TASK_RUNNING) {
6545 * Fix up our vruntime so that the current sleep doesn't
6546 * cause 'unlimited' sleep bonus.
6548 place_entity(cfs_rq, se, 0);
6549 se->vruntime -= cfs_rq->min_vruntime;
6554 * Remove our load from contribution when we leave sched_fair
6555 * and ensure we don't carry in an old decay_count if we
6558 if (se->avg.decay_count) {
6559 __synchronize_entity_decay(se);
6560 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6566 * We switched to the sched_fair class.
6568 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6574 * We were most likely switched from sched_rt, so
6575 * kick off the schedule if running, otherwise just see
6576 * if we can still preempt the current task.
6579 resched_task(rq->curr);
6581 check_preempt_curr(rq, p, 0);
6584 /* Account for a task changing its policy or group.
6586 * This routine is mostly called to set cfs_rq->curr field when a task
6587 * migrates between groups/classes.
6589 static void set_curr_task_fair(struct rq *rq)
6591 struct sched_entity *se = &rq->curr->se;
6593 for_each_sched_entity(se) {
6594 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6596 set_next_entity(cfs_rq, se);
6597 /* ensure bandwidth has been allocated on our new cfs_rq */
6598 account_cfs_rq_runtime(cfs_rq, 0);
6602 void init_cfs_rq(struct cfs_rq *cfs_rq)
6604 cfs_rq->tasks_timeline = RB_ROOT;
6605 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6606 #ifndef CONFIG_64BIT
6607 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6610 atomic64_set(&cfs_rq->decay_counter, 1);
6611 atomic_long_set(&cfs_rq->removed_load, 0);
6615 #ifdef CONFIG_FAIR_GROUP_SCHED
6616 static void task_move_group_fair(struct task_struct *p, int on_rq)
6618 struct cfs_rq *cfs_rq;
6620 * If the task was not on the rq at the time of this cgroup movement
6621 * it must have been asleep, sleeping tasks keep their ->vruntime
6622 * absolute on their old rq until wakeup (needed for the fair sleeper
6623 * bonus in place_entity()).
6625 * If it was on the rq, we've just 'preempted' it, which does convert
6626 * ->vruntime to a relative base.
6628 * Make sure both cases convert their relative position when migrating
6629 * to another cgroup's rq. This does somewhat interfere with the
6630 * fair sleeper stuff for the first placement, but who cares.
6633 * When !on_rq, vruntime of the task has usually NOT been normalized.
6634 * But there are some cases where it has already been normalized:
6636 * - Moving a forked child which is waiting for being woken up by
6637 * wake_up_new_task().
6638 * - Moving a task which has been woken up by try_to_wake_up() and
6639 * waiting for actually being woken up by sched_ttwu_pending().
6641 * To prevent boost or penalty in the new cfs_rq caused by delta
6642 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6644 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6648 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6649 set_task_rq(p, task_cpu(p));
6651 cfs_rq = cfs_rq_of(&p->se);
6652 p->se.vruntime += cfs_rq->min_vruntime;
6655 * migrate_task_rq_fair() will have removed our previous
6656 * contribution, but we must synchronize for ongoing future
6659 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6660 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6665 void free_fair_sched_group(struct task_group *tg)
6669 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6671 for_each_possible_cpu(i) {
6673 kfree(tg->cfs_rq[i]);
6682 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6684 struct cfs_rq *cfs_rq;
6685 struct sched_entity *se;
6688 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6691 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6695 tg->shares = NICE_0_LOAD;
6697 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6699 for_each_possible_cpu(i) {
6700 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6701 GFP_KERNEL, cpu_to_node(i));
6705 se = kzalloc_node(sizeof(struct sched_entity),
6706 GFP_KERNEL, cpu_to_node(i));
6710 init_cfs_rq(cfs_rq);
6711 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6722 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6724 struct rq *rq = cpu_rq(cpu);
6725 unsigned long flags;
6728 * Only empty task groups can be destroyed; so we can speculatively
6729 * check on_list without danger of it being re-added.
6731 if (!tg->cfs_rq[cpu]->on_list)
6734 raw_spin_lock_irqsave(&rq->lock, flags);
6735 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6736 raw_spin_unlock_irqrestore(&rq->lock, flags);
6739 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6740 struct sched_entity *se, int cpu,
6741 struct sched_entity *parent)
6743 struct rq *rq = cpu_rq(cpu);
6747 init_cfs_rq_runtime(cfs_rq);
6749 tg->cfs_rq[cpu] = cfs_rq;
6752 /* se could be NULL for root_task_group */
6757 se->cfs_rq = &rq->cfs;
6759 se->cfs_rq = parent->my_q;
6762 update_load_set(&se->load, 0);
6763 se->parent = parent;
6766 static DEFINE_MUTEX(shares_mutex);
6768 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6771 unsigned long flags;
6774 * We can't change the weight of the root cgroup.
6779 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6781 mutex_lock(&shares_mutex);
6782 if (tg->shares == shares)
6785 tg->shares = shares;
6786 for_each_possible_cpu(i) {
6787 struct rq *rq = cpu_rq(i);
6788 struct sched_entity *se;
6791 /* Propagate contribution to hierarchy */
6792 raw_spin_lock_irqsave(&rq->lock, flags);
6794 /* Possible calls to update_curr() need rq clock */
6795 update_rq_clock(rq);
6796 for_each_sched_entity(se)
6797 update_cfs_shares(group_cfs_rq(se));
6798 raw_spin_unlock_irqrestore(&rq->lock, flags);
6802 mutex_unlock(&shares_mutex);
6805 #else /* CONFIG_FAIR_GROUP_SCHED */
6807 void free_fair_sched_group(struct task_group *tg) { }
6809 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6814 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6816 #endif /* CONFIG_FAIR_GROUP_SCHED */
6819 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6821 struct sched_entity *se = &task->se;
6822 unsigned int rr_interval = 0;
6825 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6828 if (rq->cfs.load.weight)
6829 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6835 * All the scheduling class methods:
6837 const struct sched_class fair_sched_class = {
6838 .next = &idle_sched_class,
6839 .enqueue_task = enqueue_task_fair,
6840 .dequeue_task = dequeue_task_fair,
6841 .yield_task = yield_task_fair,
6842 .yield_to_task = yield_to_task_fair,
6844 .check_preempt_curr = check_preempt_wakeup,
6846 .pick_next_task = pick_next_task_fair,
6847 .put_prev_task = put_prev_task_fair,
6850 .select_task_rq = select_task_rq_fair,
6851 .migrate_task_rq = migrate_task_rq_fair,
6853 .rq_online = rq_online_fair,
6854 .rq_offline = rq_offline_fair,
6856 .task_waking = task_waking_fair,
6859 .set_curr_task = set_curr_task_fair,
6860 .task_tick = task_tick_fair,
6861 .task_fork = task_fork_fair,
6863 .prio_changed = prio_changed_fair,
6864 .switched_from = switched_from_fair,
6865 .switched_to = switched_to_fair,
6867 .get_rr_interval = get_rr_interval_fair,
6869 #ifdef CONFIG_FAIR_GROUP_SCHED
6870 .task_move_group = task_move_group_fair,
6874 #ifdef CONFIG_SCHED_DEBUG
6875 void print_cfs_stats(struct seq_file *m, int cpu)
6877 struct cfs_rq *cfs_rq;
6880 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6881 print_cfs_rq(m, cpu, cfs_rq);
6886 __init void init_sched_fair_class(void)
6889 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6891 #ifdef CONFIG_NO_HZ_COMMON
6892 nohz.next_balance = jiffies;
6893 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6894 cpu_notifier(sched_ilb_notifier, 0);