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);
1107 /* If the preferred nid has capacity, try to use it. */
1108 if (env.dst_stats.has_capacity)
1109 task_numa_find_cpu(&env, imp);
1111 /* No space available on the preferred nid. Look elsewhere. */
1112 if (env.best_cpu == -1) {
1113 for_each_online_node(nid) {
1114 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1117 /* Only consider nodes that recorded more faults */
1118 imp = task_faults(env.p, nid) - faults;
1123 update_numa_stats(&env.dst_stats, env.dst_nid);
1124 task_numa_find_cpu(&env, imp);
1128 /* No better CPU than the current one was found. */
1129 if (env.best_cpu == -1)
1132 if (env.best_task == NULL) {
1133 int ret = migrate_task_to(p, env.best_cpu);
1137 ret = migrate_swap(p, env.best_task);
1138 put_task_struct(env.best_task);
1142 /* Attempt to migrate a task to a CPU on the preferred node. */
1143 static void numa_migrate_preferred(struct task_struct *p)
1145 /* Success if task is already running on preferred CPU */
1146 p->numa_migrate_retry = 0;
1147 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid) {
1149 * If migration is temporarily disabled due to a task migration
1150 * then re-enable it now as the task is running on its
1151 * preferred node and memory should migrate locally
1153 if (!p->numa_migrate_seq)
1154 p->numa_migrate_seq++;
1158 /* This task has no NUMA fault statistics yet */
1159 if (unlikely(p->numa_preferred_nid == -1))
1162 /* Otherwise, try migrate to a CPU on the preferred node */
1163 if (task_numa_migrate(p) != 0)
1164 p->numa_migrate_retry = jiffies + HZ*5;
1167 static void task_numa_placement(struct task_struct *p)
1169 int seq, nid, max_nid = -1;
1170 unsigned long max_faults = 0;
1172 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1173 if (p->numa_scan_seq == seq)
1175 p->numa_scan_seq = seq;
1176 p->numa_migrate_seq++;
1177 p->numa_scan_period_max = task_scan_max(p);
1179 /* Find the node with the highest number of faults */
1180 for_each_online_node(nid) {
1181 unsigned long faults = 0;
1184 for (priv = 0; priv < 2; priv++) {
1185 i = task_faults_idx(nid, priv);
1187 /* Decay existing window, copy faults since last scan */
1188 p->numa_faults[i] >>= 1;
1189 p->numa_faults[i] += p->numa_faults_buffer[i];
1190 p->numa_faults_buffer[i] = 0;
1192 faults += p->numa_faults[i];
1195 if (faults > max_faults) {
1196 max_faults = faults;
1201 /* Preferred node as the node with the most faults */
1202 if (max_faults && max_nid != p->numa_preferred_nid) {
1203 /* Update the preferred nid and migrate task if possible */
1204 p->numa_preferred_nid = max_nid;
1205 p->numa_migrate_seq = 1;
1206 numa_migrate_preferred(p);
1211 * Got a PROT_NONE fault for a page on @node.
1213 void task_numa_fault(int last_cpupid, int node, int pages, bool migrated)
1215 struct task_struct *p = current;
1218 if (!numabalancing_enabled)
1221 /* for example, ksmd faulting in a user's mm */
1226 * First accesses are treated as private, otherwise consider accesses
1227 * to be private if the accessing pid has not changed
1229 if (!cpupid_pid_unset(last_cpupid))
1230 priv = ((p->pid & LAST__PID_MASK) == cpupid_to_pid(last_cpupid));
1234 /* Allocate buffer to track faults on a per-node basis */
1235 if (unlikely(!p->numa_faults)) {
1236 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1238 /* numa_faults and numa_faults_buffer share the allocation */
1239 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1240 if (!p->numa_faults)
1243 BUG_ON(p->numa_faults_buffer);
1244 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1248 * If pages are properly placed (did not migrate) then scan slower.
1249 * This is reset periodically in case of phase changes
1252 /* Initialise if necessary */
1253 if (!p->numa_scan_period_max)
1254 p->numa_scan_period_max = task_scan_max(p);
1256 p->numa_scan_period = min(p->numa_scan_period_max,
1257 p->numa_scan_period + 10);
1260 task_numa_placement(p);
1262 /* Retry task to preferred node migration if it previously failed */
1263 if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry))
1264 numa_migrate_preferred(p);
1266 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1269 static void reset_ptenuma_scan(struct task_struct *p)
1271 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1272 p->mm->numa_scan_offset = 0;
1276 * The expensive part of numa migration is done from task_work context.
1277 * Triggered from task_tick_numa().
1279 void task_numa_work(struct callback_head *work)
1281 unsigned long migrate, next_scan, now = jiffies;
1282 struct task_struct *p = current;
1283 struct mm_struct *mm = p->mm;
1284 struct vm_area_struct *vma;
1285 unsigned long start, end;
1286 unsigned long nr_pte_updates = 0;
1289 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1291 work->next = work; /* protect against double add */
1293 * Who cares about NUMA placement when they're dying.
1295 * NOTE: make sure not to dereference p->mm before this check,
1296 * exit_task_work() happens _after_ exit_mm() so we could be called
1297 * without p->mm even though we still had it when we enqueued this
1300 if (p->flags & PF_EXITING)
1303 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1304 mm->numa_next_scan = now +
1305 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1306 mm->numa_next_reset = now +
1307 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1311 * Reset the scan period if enough time has gone by. Objective is that
1312 * scanning will be reduced if pages are properly placed. As tasks
1313 * can enter different phases this needs to be re-examined. Lacking
1314 * proper tracking of reference behaviour, this blunt hammer is used.
1316 migrate = mm->numa_next_reset;
1317 if (time_after(now, migrate)) {
1318 p->numa_scan_period = task_scan_min(p);
1319 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1320 xchg(&mm->numa_next_reset, next_scan);
1324 * Enforce maximal scan/migration frequency..
1326 migrate = mm->numa_next_scan;
1327 if (time_before(now, migrate))
1330 if (p->numa_scan_period == 0) {
1331 p->numa_scan_period_max = task_scan_max(p);
1332 p->numa_scan_period = task_scan_min(p);
1335 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1336 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1340 * Delay this task enough that another task of this mm will likely win
1341 * the next time around.
1343 p->node_stamp += 2 * TICK_NSEC;
1345 start = mm->numa_scan_offset;
1346 pages = sysctl_numa_balancing_scan_size;
1347 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1351 down_read(&mm->mmap_sem);
1352 vma = find_vma(mm, start);
1354 reset_ptenuma_scan(p);
1358 for (; vma; vma = vma->vm_next) {
1359 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1363 * Shared library pages mapped by multiple processes are not
1364 * migrated as it is expected they are cache replicated. Avoid
1365 * hinting faults in read-only file-backed mappings or the vdso
1366 * as migrating the pages will be of marginal benefit.
1369 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1373 start = max(start, vma->vm_start);
1374 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1375 end = min(end, vma->vm_end);
1376 nr_pte_updates += change_prot_numa(vma, start, end);
1379 * Scan sysctl_numa_balancing_scan_size but ensure that
1380 * at least one PTE is updated so that unused virtual
1381 * address space is quickly skipped.
1384 pages -= (end - start) >> PAGE_SHIFT;
1389 } while (end != vma->vm_end);
1394 * If the whole process was scanned without updates then no NUMA
1395 * hinting faults are being recorded and scan rate should be lower.
1397 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1398 p->numa_scan_period = min(p->numa_scan_period_max,
1399 p->numa_scan_period << 1);
1401 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1402 mm->numa_next_scan = next_scan;
1406 * It is possible to reach the end of the VMA list but the last few
1407 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1408 * would find the !migratable VMA on the next scan but not reset the
1409 * scanner to the start so check it now.
1412 mm->numa_scan_offset = start;
1414 reset_ptenuma_scan(p);
1415 up_read(&mm->mmap_sem);
1419 * Drive the periodic memory faults..
1421 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1423 struct callback_head *work = &curr->numa_work;
1427 * We don't care about NUMA placement if we don't have memory.
1429 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1433 * Using runtime rather than walltime has the dual advantage that
1434 * we (mostly) drive the selection from busy threads and that the
1435 * task needs to have done some actual work before we bother with
1438 now = curr->se.sum_exec_runtime;
1439 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1441 if (now - curr->node_stamp > period) {
1442 if (!curr->node_stamp)
1443 curr->numa_scan_period = task_scan_min(curr);
1444 curr->node_stamp += period;
1446 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1447 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1448 task_work_add(curr, work, true);
1453 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1456 #endif /* CONFIG_NUMA_BALANCING */
1459 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1461 update_load_add(&cfs_rq->load, se->load.weight);
1462 if (!parent_entity(se))
1463 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1465 if (entity_is_task(se))
1466 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1468 cfs_rq->nr_running++;
1472 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1474 update_load_sub(&cfs_rq->load, se->load.weight);
1475 if (!parent_entity(se))
1476 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1477 if (entity_is_task(se))
1478 list_del_init(&se->group_node);
1479 cfs_rq->nr_running--;
1482 #ifdef CONFIG_FAIR_GROUP_SCHED
1484 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1489 * Use this CPU's actual weight instead of the last load_contribution
1490 * to gain a more accurate current total weight. See
1491 * update_cfs_rq_load_contribution().
1493 tg_weight = atomic_long_read(&tg->load_avg);
1494 tg_weight -= cfs_rq->tg_load_contrib;
1495 tg_weight += cfs_rq->load.weight;
1500 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1502 long tg_weight, load, shares;
1504 tg_weight = calc_tg_weight(tg, cfs_rq);
1505 load = cfs_rq->load.weight;
1507 shares = (tg->shares * load);
1509 shares /= tg_weight;
1511 if (shares < MIN_SHARES)
1512 shares = MIN_SHARES;
1513 if (shares > tg->shares)
1514 shares = tg->shares;
1518 # else /* CONFIG_SMP */
1519 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1523 # endif /* CONFIG_SMP */
1524 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1525 unsigned long weight)
1528 /* commit outstanding execution time */
1529 if (cfs_rq->curr == se)
1530 update_curr(cfs_rq);
1531 account_entity_dequeue(cfs_rq, se);
1534 update_load_set(&se->load, weight);
1537 account_entity_enqueue(cfs_rq, se);
1540 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1542 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1544 struct task_group *tg;
1545 struct sched_entity *se;
1549 se = tg->se[cpu_of(rq_of(cfs_rq))];
1550 if (!se || throttled_hierarchy(cfs_rq))
1553 if (likely(se->load.weight == tg->shares))
1556 shares = calc_cfs_shares(cfs_rq, tg);
1558 reweight_entity(cfs_rq_of(se), se, shares);
1560 #else /* CONFIG_FAIR_GROUP_SCHED */
1561 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1564 #endif /* CONFIG_FAIR_GROUP_SCHED */
1568 * We choose a half-life close to 1 scheduling period.
1569 * Note: The tables below are dependent on this value.
1571 #define LOAD_AVG_PERIOD 32
1572 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1573 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1575 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1576 static const u32 runnable_avg_yN_inv[] = {
1577 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1578 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1579 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1580 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1581 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1582 0x85aac367, 0x82cd8698,
1586 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1587 * over-estimates when re-combining.
1589 static const u32 runnable_avg_yN_sum[] = {
1590 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1591 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1592 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1597 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1599 static __always_inline u64 decay_load(u64 val, u64 n)
1601 unsigned int local_n;
1605 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1608 /* after bounds checking we can collapse to 32-bit */
1612 * As y^PERIOD = 1/2, we can combine
1613 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1614 * With a look-up table which covers k^n (n<PERIOD)
1616 * To achieve constant time decay_load.
1618 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1619 val >>= local_n / LOAD_AVG_PERIOD;
1620 local_n %= LOAD_AVG_PERIOD;
1623 val *= runnable_avg_yN_inv[local_n];
1624 /* We don't use SRR here since we always want to round down. */
1629 * For updates fully spanning n periods, the contribution to runnable
1630 * average will be: \Sum 1024*y^n
1632 * We can compute this reasonably efficiently by combining:
1633 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1635 static u32 __compute_runnable_contrib(u64 n)
1639 if (likely(n <= LOAD_AVG_PERIOD))
1640 return runnable_avg_yN_sum[n];
1641 else if (unlikely(n >= LOAD_AVG_MAX_N))
1642 return LOAD_AVG_MAX;
1644 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1646 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1647 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1649 n -= LOAD_AVG_PERIOD;
1650 } while (n > LOAD_AVG_PERIOD);
1652 contrib = decay_load(contrib, n);
1653 return contrib + runnable_avg_yN_sum[n];
1657 * We can represent the historical contribution to runnable average as the
1658 * coefficients of a geometric series. To do this we sub-divide our runnable
1659 * history into segments of approximately 1ms (1024us); label the segment that
1660 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1662 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1664 * (now) (~1ms ago) (~2ms ago)
1666 * Let u_i denote the fraction of p_i that the entity was runnable.
1668 * We then designate the fractions u_i as our co-efficients, yielding the
1669 * following representation of historical load:
1670 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1672 * We choose y based on the with of a reasonably scheduling period, fixing:
1675 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1676 * approximately half as much as the contribution to load within the last ms
1679 * When a period "rolls over" and we have new u_0`, multiplying the previous
1680 * sum again by y is sufficient to update:
1681 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1682 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1684 static __always_inline int __update_entity_runnable_avg(u64 now,
1685 struct sched_avg *sa,
1689 u32 runnable_contrib;
1690 int delta_w, decayed = 0;
1692 delta = now - sa->last_runnable_update;
1694 * This should only happen when time goes backwards, which it
1695 * unfortunately does during sched clock init when we swap over to TSC.
1697 if ((s64)delta < 0) {
1698 sa->last_runnable_update = now;
1703 * Use 1024ns as the unit of measurement since it's a reasonable
1704 * approximation of 1us and fast to compute.
1709 sa->last_runnable_update = now;
1711 /* delta_w is the amount already accumulated against our next period */
1712 delta_w = sa->runnable_avg_period % 1024;
1713 if (delta + delta_w >= 1024) {
1714 /* period roll-over */
1718 * Now that we know we're crossing a period boundary, figure
1719 * out how much from delta we need to complete the current
1720 * period and accrue it.
1722 delta_w = 1024 - delta_w;
1724 sa->runnable_avg_sum += delta_w;
1725 sa->runnable_avg_period += delta_w;
1729 /* Figure out how many additional periods this update spans */
1730 periods = delta / 1024;
1733 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1735 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1738 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1739 runnable_contrib = __compute_runnable_contrib(periods);
1741 sa->runnable_avg_sum += runnable_contrib;
1742 sa->runnable_avg_period += runnable_contrib;
1745 /* Remainder of delta accrued against u_0` */
1747 sa->runnable_avg_sum += delta;
1748 sa->runnable_avg_period += delta;
1753 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1754 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1756 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1757 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1759 decays -= se->avg.decay_count;
1763 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1764 se->avg.decay_count = 0;
1769 #ifdef CONFIG_FAIR_GROUP_SCHED
1770 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1773 struct task_group *tg = cfs_rq->tg;
1776 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1777 tg_contrib -= cfs_rq->tg_load_contrib;
1779 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1780 atomic_long_add(tg_contrib, &tg->load_avg);
1781 cfs_rq->tg_load_contrib += tg_contrib;
1786 * Aggregate cfs_rq runnable averages into an equivalent task_group
1787 * representation for computing load contributions.
1789 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1790 struct cfs_rq *cfs_rq)
1792 struct task_group *tg = cfs_rq->tg;
1795 /* The fraction of a cpu used by this cfs_rq */
1796 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1797 sa->runnable_avg_period + 1);
1798 contrib -= cfs_rq->tg_runnable_contrib;
1800 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1801 atomic_add(contrib, &tg->runnable_avg);
1802 cfs_rq->tg_runnable_contrib += contrib;
1806 static inline void __update_group_entity_contrib(struct sched_entity *se)
1808 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1809 struct task_group *tg = cfs_rq->tg;
1814 contrib = cfs_rq->tg_load_contrib * tg->shares;
1815 se->avg.load_avg_contrib = div_u64(contrib,
1816 atomic_long_read(&tg->load_avg) + 1);
1819 * For group entities we need to compute a correction term in the case
1820 * that they are consuming <1 cpu so that we would contribute the same
1821 * load as a task of equal weight.
1823 * Explicitly co-ordinating this measurement would be expensive, but
1824 * fortunately the sum of each cpus contribution forms a usable
1825 * lower-bound on the true value.
1827 * Consider the aggregate of 2 contributions. Either they are disjoint
1828 * (and the sum represents true value) or they are disjoint and we are
1829 * understating by the aggregate of their overlap.
1831 * Extending this to N cpus, for a given overlap, the maximum amount we
1832 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1833 * cpus that overlap for this interval and w_i is the interval width.
1835 * On a small machine; the first term is well-bounded which bounds the
1836 * total error since w_i is a subset of the period. Whereas on a
1837 * larger machine, while this first term can be larger, if w_i is the
1838 * of consequential size guaranteed to see n_i*w_i quickly converge to
1839 * our upper bound of 1-cpu.
1841 runnable_avg = atomic_read(&tg->runnable_avg);
1842 if (runnable_avg < NICE_0_LOAD) {
1843 se->avg.load_avg_contrib *= runnable_avg;
1844 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1848 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1849 int force_update) {}
1850 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1851 struct cfs_rq *cfs_rq) {}
1852 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1855 static inline void __update_task_entity_contrib(struct sched_entity *se)
1859 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1860 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1861 contrib /= (se->avg.runnable_avg_period + 1);
1862 se->avg.load_avg_contrib = scale_load(contrib);
1865 /* Compute the current contribution to load_avg by se, return any delta */
1866 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1868 long old_contrib = se->avg.load_avg_contrib;
1870 if (entity_is_task(se)) {
1871 __update_task_entity_contrib(se);
1873 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1874 __update_group_entity_contrib(se);
1877 return se->avg.load_avg_contrib - old_contrib;
1880 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1883 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1884 cfs_rq->blocked_load_avg -= load_contrib;
1886 cfs_rq->blocked_load_avg = 0;
1889 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1891 /* Update a sched_entity's runnable average */
1892 static inline void update_entity_load_avg(struct sched_entity *se,
1895 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1900 * For a group entity we need to use their owned cfs_rq_clock_task() in
1901 * case they are the parent of a throttled hierarchy.
1903 if (entity_is_task(se))
1904 now = cfs_rq_clock_task(cfs_rq);
1906 now = cfs_rq_clock_task(group_cfs_rq(se));
1908 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1911 contrib_delta = __update_entity_load_avg_contrib(se);
1917 cfs_rq->runnable_load_avg += contrib_delta;
1919 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1923 * Decay the load contributed by all blocked children and account this so that
1924 * their contribution may appropriately discounted when they wake up.
1926 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1928 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1931 decays = now - cfs_rq->last_decay;
1932 if (!decays && !force_update)
1935 if (atomic_long_read(&cfs_rq->removed_load)) {
1936 unsigned long removed_load;
1937 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1938 subtract_blocked_load_contrib(cfs_rq, removed_load);
1942 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1944 atomic64_add(decays, &cfs_rq->decay_counter);
1945 cfs_rq->last_decay = now;
1948 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1951 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1953 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1954 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1957 /* Add the load generated by se into cfs_rq's child load-average */
1958 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1959 struct sched_entity *se,
1963 * We track migrations using entity decay_count <= 0, on a wake-up
1964 * migration we use a negative decay count to track the remote decays
1965 * accumulated while sleeping.
1967 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1968 * are seen by enqueue_entity_load_avg() as a migration with an already
1969 * constructed load_avg_contrib.
1971 if (unlikely(se->avg.decay_count <= 0)) {
1972 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1973 if (se->avg.decay_count) {
1975 * In a wake-up migration we have to approximate the
1976 * time sleeping. This is because we can't synchronize
1977 * clock_task between the two cpus, and it is not
1978 * guaranteed to be read-safe. Instead, we can
1979 * approximate this using our carried decays, which are
1980 * explicitly atomically readable.
1982 se->avg.last_runnable_update -= (-se->avg.decay_count)
1984 update_entity_load_avg(se, 0);
1985 /* Indicate that we're now synchronized and on-rq */
1986 se->avg.decay_count = 0;
1991 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1992 * would have made count negative); we must be careful to avoid
1993 * double-accounting blocked time after synchronizing decays.
1995 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1999 /* migrated tasks did not contribute to our blocked load */
2001 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2002 update_entity_load_avg(se, 0);
2005 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2006 /* we force update consideration on load-balancer moves */
2007 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2011 * Remove se's load from this cfs_rq child load-average, if the entity is
2012 * transitioning to a blocked state we track its projected decay using
2015 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2016 struct sched_entity *se,
2019 update_entity_load_avg(se, 1);
2020 /* we force update consideration on load-balancer moves */
2021 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2023 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2025 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2026 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2027 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2031 * Update the rq's load with the elapsed running time before entering
2032 * idle. if the last scheduled task is not a CFS task, idle_enter will
2033 * be the only way to update the runnable statistic.
2035 void idle_enter_fair(struct rq *this_rq)
2037 update_rq_runnable_avg(this_rq, 1);
2041 * Update the rq's load with the elapsed idle time before a task is
2042 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2043 * be the only way to update the runnable statistic.
2045 void idle_exit_fair(struct rq *this_rq)
2047 update_rq_runnable_avg(this_rq, 0);
2051 static inline void update_entity_load_avg(struct sched_entity *se,
2052 int update_cfs_rq) {}
2053 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2054 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2055 struct sched_entity *se,
2057 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2058 struct sched_entity *se,
2060 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2061 int force_update) {}
2064 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2066 #ifdef CONFIG_SCHEDSTATS
2067 struct task_struct *tsk = NULL;
2069 if (entity_is_task(se))
2072 if (se->statistics.sleep_start) {
2073 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2078 if (unlikely(delta > se->statistics.sleep_max))
2079 se->statistics.sleep_max = delta;
2081 se->statistics.sleep_start = 0;
2082 se->statistics.sum_sleep_runtime += delta;
2085 account_scheduler_latency(tsk, delta >> 10, 1);
2086 trace_sched_stat_sleep(tsk, delta);
2089 if (se->statistics.block_start) {
2090 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2095 if (unlikely(delta > se->statistics.block_max))
2096 se->statistics.block_max = delta;
2098 se->statistics.block_start = 0;
2099 se->statistics.sum_sleep_runtime += delta;
2102 if (tsk->in_iowait) {
2103 se->statistics.iowait_sum += delta;
2104 se->statistics.iowait_count++;
2105 trace_sched_stat_iowait(tsk, delta);
2108 trace_sched_stat_blocked(tsk, delta);
2111 * Blocking time is in units of nanosecs, so shift by
2112 * 20 to get a milliseconds-range estimation of the
2113 * amount of time that the task spent sleeping:
2115 if (unlikely(prof_on == SLEEP_PROFILING)) {
2116 profile_hits(SLEEP_PROFILING,
2117 (void *)get_wchan(tsk),
2120 account_scheduler_latency(tsk, delta >> 10, 0);
2126 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2128 #ifdef CONFIG_SCHED_DEBUG
2129 s64 d = se->vruntime - cfs_rq->min_vruntime;
2134 if (d > 3*sysctl_sched_latency)
2135 schedstat_inc(cfs_rq, nr_spread_over);
2140 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2142 u64 vruntime = cfs_rq->min_vruntime;
2145 * The 'current' period is already promised to the current tasks,
2146 * however the extra weight of the new task will slow them down a
2147 * little, place the new task so that it fits in the slot that
2148 * stays open at the end.
2150 if (initial && sched_feat(START_DEBIT))
2151 vruntime += sched_vslice(cfs_rq, se);
2153 /* sleeps up to a single latency don't count. */
2155 unsigned long thresh = sysctl_sched_latency;
2158 * Halve their sleep time's effect, to allow
2159 * for a gentler effect of sleepers:
2161 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2167 /* ensure we never gain time by being placed backwards. */
2168 se->vruntime = max_vruntime(se->vruntime, vruntime);
2171 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2174 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2177 * Update the normalized vruntime before updating min_vruntime
2178 * through calling update_curr().
2180 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2181 se->vruntime += cfs_rq->min_vruntime;
2184 * Update run-time statistics of the 'current'.
2186 update_curr(cfs_rq);
2187 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2188 account_entity_enqueue(cfs_rq, se);
2189 update_cfs_shares(cfs_rq);
2191 if (flags & ENQUEUE_WAKEUP) {
2192 place_entity(cfs_rq, se, 0);
2193 enqueue_sleeper(cfs_rq, se);
2196 update_stats_enqueue(cfs_rq, se);
2197 check_spread(cfs_rq, se);
2198 if (se != cfs_rq->curr)
2199 __enqueue_entity(cfs_rq, se);
2202 if (cfs_rq->nr_running == 1) {
2203 list_add_leaf_cfs_rq(cfs_rq);
2204 check_enqueue_throttle(cfs_rq);
2208 static void __clear_buddies_last(struct sched_entity *se)
2210 for_each_sched_entity(se) {
2211 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2212 if (cfs_rq->last == se)
2213 cfs_rq->last = NULL;
2219 static void __clear_buddies_next(struct sched_entity *se)
2221 for_each_sched_entity(se) {
2222 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2223 if (cfs_rq->next == se)
2224 cfs_rq->next = NULL;
2230 static void __clear_buddies_skip(struct sched_entity *se)
2232 for_each_sched_entity(se) {
2233 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2234 if (cfs_rq->skip == se)
2235 cfs_rq->skip = NULL;
2241 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2243 if (cfs_rq->last == se)
2244 __clear_buddies_last(se);
2246 if (cfs_rq->next == se)
2247 __clear_buddies_next(se);
2249 if (cfs_rq->skip == se)
2250 __clear_buddies_skip(se);
2253 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2256 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2259 * Update run-time statistics of the 'current'.
2261 update_curr(cfs_rq);
2262 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2264 update_stats_dequeue(cfs_rq, se);
2265 if (flags & DEQUEUE_SLEEP) {
2266 #ifdef CONFIG_SCHEDSTATS
2267 if (entity_is_task(se)) {
2268 struct task_struct *tsk = task_of(se);
2270 if (tsk->state & TASK_INTERRUPTIBLE)
2271 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2272 if (tsk->state & TASK_UNINTERRUPTIBLE)
2273 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2278 clear_buddies(cfs_rq, se);
2280 if (se != cfs_rq->curr)
2281 __dequeue_entity(cfs_rq, se);
2283 account_entity_dequeue(cfs_rq, se);
2286 * Normalize the entity after updating the min_vruntime because the
2287 * update can refer to the ->curr item and we need to reflect this
2288 * movement in our normalized position.
2290 if (!(flags & DEQUEUE_SLEEP))
2291 se->vruntime -= cfs_rq->min_vruntime;
2293 /* return excess runtime on last dequeue */
2294 return_cfs_rq_runtime(cfs_rq);
2296 update_min_vruntime(cfs_rq);
2297 update_cfs_shares(cfs_rq);
2301 * Preempt the current task with a newly woken task if needed:
2304 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2306 unsigned long ideal_runtime, delta_exec;
2307 struct sched_entity *se;
2310 ideal_runtime = sched_slice(cfs_rq, curr);
2311 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2312 if (delta_exec > ideal_runtime) {
2313 resched_task(rq_of(cfs_rq)->curr);
2315 * The current task ran long enough, ensure it doesn't get
2316 * re-elected due to buddy favours.
2318 clear_buddies(cfs_rq, curr);
2323 * Ensure that a task that missed wakeup preemption by a
2324 * narrow margin doesn't have to wait for a full slice.
2325 * This also mitigates buddy induced latencies under load.
2327 if (delta_exec < sysctl_sched_min_granularity)
2330 se = __pick_first_entity(cfs_rq);
2331 delta = curr->vruntime - se->vruntime;
2336 if (delta > ideal_runtime)
2337 resched_task(rq_of(cfs_rq)->curr);
2341 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2343 /* 'current' is not kept within the tree. */
2346 * Any task has to be enqueued before it get to execute on
2347 * a CPU. So account for the time it spent waiting on the
2350 update_stats_wait_end(cfs_rq, se);
2351 __dequeue_entity(cfs_rq, se);
2354 update_stats_curr_start(cfs_rq, se);
2356 #ifdef CONFIG_SCHEDSTATS
2358 * Track our maximum slice length, if the CPU's load is at
2359 * least twice that of our own weight (i.e. dont track it
2360 * when there are only lesser-weight tasks around):
2362 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2363 se->statistics.slice_max = max(se->statistics.slice_max,
2364 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2367 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2371 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2374 * Pick the next process, keeping these things in mind, in this order:
2375 * 1) keep things fair between processes/task groups
2376 * 2) pick the "next" process, since someone really wants that to run
2377 * 3) pick the "last" process, for cache locality
2378 * 4) do not run the "skip" process, if something else is available
2380 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2382 struct sched_entity *se = __pick_first_entity(cfs_rq);
2383 struct sched_entity *left = se;
2386 * Avoid running the skip buddy, if running something else can
2387 * be done without getting too unfair.
2389 if (cfs_rq->skip == se) {
2390 struct sched_entity *second = __pick_next_entity(se);
2391 if (second && wakeup_preempt_entity(second, left) < 1)
2396 * Prefer last buddy, try to return the CPU to a preempted task.
2398 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2402 * Someone really wants this to run. If it's not unfair, run it.
2404 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2407 clear_buddies(cfs_rq, se);
2412 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2414 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2417 * If still on the runqueue then deactivate_task()
2418 * was not called and update_curr() has to be done:
2421 update_curr(cfs_rq);
2423 /* throttle cfs_rqs exceeding runtime */
2424 check_cfs_rq_runtime(cfs_rq);
2426 check_spread(cfs_rq, prev);
2428 update_stats_wait_start(cfs_rq, prev);
2429 /* Put 'current' back into the tree. */
2430 __enqueue_entity(cfs_rq, prev);
2431 /* in !on_rq case, update occurred at dequeue */
2432 update_entity_load_avg(prev, 1);
2434 cfs_rq->curr = NULL;
2438 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2441 * Update run-time statistics of the 'current'.
2443 update_curr(cfs_rq);
2446 * Ensure that runnable average is periodically updated.
2448 update_entity_load_avg(curr, 1);
2449 update_cfs_rq_blocked_load(cfs_rq, 1);
2450 update_cfs_shares(cfs_rq);
2452 #ifdef CONFIG_SCHED_HRTICK
2454 * queued ticks are scheduled to match the slice, so don't bother
2455 * validating it and just reschedule.
2458 resched_task(rq_of(cfs_rq)->curr);
2462 * don't let the period tick interfere with the hrtick preemption
2464 if (!sched_feat(DOUBLE_TICK) &&
2465 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2469 if (cfs_rq->nr_running > 1)
2470 check_preempt_tick(cfs_rq, curr);
2474 /**************************************************
2475 * CFS bandwidth control machinery
2478 #ifdef CONFIG_CFS_BANDWIDTH
2480 #ifdef HAVE_JUMP_LABEL
2481 static struct static_key __cfs_bandwidth_used;
2483 static inline bool cfs_bandwidth_used(void)
2485 return static_key_false(&__cfs_bandwidth_used);
2488 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2490 /* only need to count groups transitioning between enabled/!enabled */
2491 if (enabled && !was_enabled)
2492 static_key_slow_inc(&__cfs_bandwidth_used);
2493 else if (!enabled && was_enabled)
2494 static_key_slow_dec(&__cfs_bandwidth_used);
2496 #else /* HAVE_JUMP_LABEL */
2497 static bool cfs_bandwidth_used(void)
2502 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2503 #endif /* HAVE_JUMP_LABEL */
2506 * default period for cfs group bandwidth.
2507 * default: 0.1s, units: nanoseconds
2509 static inline u64 default_cfs_period(void)
2511 return 100000000ULL;
2514 static inline u64 sched_cfs_bandwidth_slice(void)
2516 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2520 * Replenish runtime according to assigned quota and update expiration time.
2521 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2522 * additional synchronization around rq->lock.
2524 * requires cfs_b->lock
2526 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2530 if (cfs_b->quota == RUNTIME_INF)
2533 now = sched_clock_cpu(smp_processor_id());
2534 cfs_b->runtime = cfs_b->quota;
2535 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2538 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2540 return &tg->cfs_bandwidth;
2543 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2544 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2546 if (unlikely(cfs_rq->throttle_count))
2547 return cfs_rq->throttled_clock_task;
2549 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2552 /* returns 0 on failure to allocate runtime */
2553 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2555 struct task_group *tg = cfs_rq->tg;
2556 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2557 u64 amount = 0, min_amount, expires;
2559 /* note: this is a positive sum as runtime_remaining <= 0 */
2560 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2562 raw_spin_lock(&cfs_b->lock);
2563 if (cfs_b->quota == RUNTIME_INF)
2564 amount = min_amount;
2567 * If the bandwidth pool has become inactive, then at least one
2568 * period must have elapsed since the last consumption.
2569 * Refresh the global state and ensure bandwidth timer becomes
2572 if (!cfs_b->timer_active) {
2573 __refill_cfs_bandwidth_runtime(cfs_b);
2574 __start_cfs_bandwidth(cfs_b);
2577 if (cfs_b->runtime > 0) {
2578 amount = min(cfs_b->runtime, min_amount);
2579 cfs_b->runtime -= amount;
2583 expires = cfs_b->runtime_expires;
2584 raw_spin_unlock(&cfs_b->lock);
2586 cfs_rq->runtime_remaining += amount;
2588 * we may have advanced our local expiration to account for allowed
2589 * spread between our sched_clock and the one on which runtime was
2592 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2593 cfs_rq->runtime_expires = expires;
2595 return cfs_rq->runtime_remaining > 0;
2599 * Note: This depends on the synchronization provided by sched_clock and the
2600 * fact that rq->clock snapshots this value.
2602 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2604 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2606 /* if the deadline is ahead of our clock, nothing to do */
2607 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2610 if (cfs_rq->runtime_remaining < 0)
2614 * If the local deadline has passed we have to consider the
2615 * possibility that our sched_clock is 'fast' and the global deadline
2616 * has not truly expired.
2618 * Fortunately we can check determine whether this the case by checking
2619 * whether the global deadline has advanced.
2622 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2623 /* extend local deadline, drift is bounded above by 2 ticks */
2624 cfs_rq->runtime_expires += TICK_NSEC;
2626 /* global deadline is ahead, expiration has passed */
2627 cfs_rq->runtime_remaining = 0;
2631 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2632 unsigned long delta_exec)
2634 /* dock delta_exec before expiring quota (as it could span periods) */
2635 cfs_rq->runtime_remaining -= delta_exec;
2636 expire_cfs_rq_runtime(cfs_rq);
2638 if (likely(cfs_rq->runtime_remaining > 0))
2642 * if we're unable to extend our runtime we resched so that the active
2643 * hierarchy can be throttled
2645 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2646 resched_task(rq_of(cfs_rq)->curr);
2649 static __always_inline
2650 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2652 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2655 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2658 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2660 return cfs_bandwidth_used() && cfs_rq->throttled;
2663 /* check whether cfs_rq, or any parent, is throttled */
2664 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2666 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2670 * Ensure that neither of the group entities corresponding to src_cpu or
2671 * dest_cpu are members of a throttled hierarchy when performing group
2672 * load-balance operations.
2674 static inline int throttled_lb_pair(struct task_group *tg,
2675 int src_cpu, int dest_cpu)
2677 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2679 src_cfs_rq = tg->cfs_rq[src_cpu];
2680 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2682 return throttled_hierarchy(src_cfs_rq) ||
2683 throttled_hierarchy(dest_cfs_rq);
2686 /* updated child weight may affect parent so we have to do this bottom up */
2687 static int tg_unthrottle_up(struct task_group *tg, void *data)
2689 struct rq *rq = data;
2690 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2692 cfs_rq->throttle_count--;
2694 if (!cfs_rq->throttle_count) {
2695 /* adjust cfs_rq_clock_task() */
2696 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2697 cfs_rq->throttled_clock_task;
2704 static int tg_throttle_down(struct task_group *tg, void *data)
2706 struct rq *rq = data;
2707 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2709 /* group is entering throttled state, stop time */
2710 if (!cfs_rq->throttle_count)
2711 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2712 cfs_rq->throttle_count++;
2717 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2719 struct rq *rq = rq_of(cfs_rq);
2720 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2721 struct sched_entity *se;
2722 long task_delta, dequeue = 1;
2724 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2726 /* freeze hierarchy runnable averages while throttled */
2728 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2731 task_delta = cfs_rq->h_nr_running;
2732 for_each_sched_entity(se) {
2733 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2734 /* throttled entity or throttle-on-deactivate */
2739 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2740 qcfs_rq->h_nr_running -= task_delta;
2742 if (qcfs_rq->load.weight)
2747 rq->nr_running -= task_delta;
2749 cfs_rq->throttled = 1;
2750 cfs_rq->throttled_clock = rq_clock(rq);
2751 raw_spin_lock(&cfs_b->lock);
2752 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2753 raw_spin_unlock(&cfs_b->lock);
2756 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2758 struct rq *rq = rq_of(cfs_rq);
2759 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2760 struct sched_entity *se;
2764 se = cfs_rq->tg->se[cpu_of(rq)];
2766 cfs_rq->throttled = 0;
2768 update_rq_clock(rq);
2770 raw_spin_lock(&cfs_b->lock);
2771 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2772 list_del_rcu(&cfs_rq->throttled_list);
2773 raw_spin_unlock(&cfs_b->lock);
2775 /* update hierarchical throttle state */
2776 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2778 if (!cfs_rq->load.weight)
2781 task_delta = cfs_rq->h_nr_running;
2782 for_each_sched_entity(se) {
2786 cfs_rq = cfs_rq_of(se);
2788 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2789 cfs_rq->h_nr_running += task_delta;
2791 if (cfs_rq_throttled(cfs_rq))
2796 rq->nr_running += task_delta;
2798 /* determine whether we need to wake up potentially idle cpu */
2799 if (rq->curr == rq->idle && rq->cfs.nr_running)
2800 resched_task(rq->curr);
2803 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2804 u64 remaining, u64 expires)
2806 struct cfs_rq *cfs_rq;
2807 u64 runtime = remaining;
2810 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2812 struct rq *rq = rq_of(cfs_rq);
2814 raw_spin_lock(&rq->lock);
2815 if (!cfs_rq_throttled(cfs_rq))
2818 runtime = -cfs_rq->runtime_remaining + 1;
2819 if (runtime > remaining)
2820 runtime = remaining;
2821 remaining -= runtime;
2823 cfs_rq->runtime_remaining += runtime;
2824 cfs_rq->runtime_expires = expires;
2826 /* we check whether we're throttled above */
2827 if (cfs_rq->runtime_remaining > 0)
2828 unthrottle_cfs_rq(cfs_rq);
2831 raw_spin_unlock(&rq->lock);
2842 * Responsible for refilling a task_group's bandwidth and unthrottling its
2843 * cfs_rqs as appropriate. If there has been no activity within the last
2844 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2845 * used to track this state.
2847 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2849 u64 runtime, runtime_expires;
2850 int idle = 1, throttled;
2852 raw_spin_lock(&cfs_b->lock);
2853 /* no need to continue the timer with no bandwidth constraint */
2854 if (cfs_b->quota == RUNTIME_INF)
2857 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2858 /* idle depends on !throttled (for the case of a large deficit) */
2859 idle = cfs_b->idle && !throttled;
2860 cfs_b->nr_periods += overrun;
2862 /* if we're going inactive then everything else can be deferred */
2866 __refill_cfs_bandwidth_runtime(cfs_b);
2869 /* mark as potentially idle for the upcoming period */
2874 /* account preceding periods in which throttling occurred */
2875 cfs_b->nr_throttled += overrun;
2878 * There are throttled entities so we must first use the new bandwidth
2879 * to unthrottle them before making it generally available. This
2880 * ensures that all existing debts will be paid before a new cfs_rq is
2883 runtime = cfs_b->runtime;
2884 runtime_expires = cfs_b->runtime_expires;
2888 * This check is repeated as we are holding onto the new bandwidth
2889 * while we unthrottle. This can potentially race with an unthrottled
2890 * group trying to acquire new bandwidth from the global pool.
2892 while (throttled && runtime > 0) {
2893 raw_spin_unlock(&cfs_b->lock);
2894 /* we can't nest cfs_b->lock while distributing bandwidth */
2895 runtime = distribute_cfs_runtime(cfs_b, runtime,
2897 raw_spin_lock(&cfs_b->lock);
2899 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2902 /* return (any) remaining runtime */
2903 cfs_b->runtime = runtime;
2905 * While we are ensured activity in the period following an
2906 * unthrottle, this also covers the case in which the new bandwidth is
2907 * insufficient to cover the existing bandwidth deficit. (Forcing the
2908 * timer to remain active while there are any throttled entities.)
2913 cfs_b->timer_active = 0;
2914 raw_spin_unlock(&cfs_b->lock);
2919 /* a cfs_rq won't donate quota below this amount */
2920 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2921 /* minimum remaining period time to redistribute slack quota */
2922 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2923 /* how long we wait to gather additional slack before distributing */
2924 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2926 /* are we near the end of the current quota period? */
2927 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2929 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2932 /* if the call-back is running a quota refresh is already occurring */
2933 if (hrtimer_callback_running(refresh_timer))
2936 /* is a quota refresh about to occur? */
2937 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2938 if (remaining < min_expire)
2944 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2946 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2948 /* if there's a quota refresh soon don't bother with slack */
2949 if (runtime_refresh_within(cfs_b, min_left))
2952 start_bandwidth_timer(&cfs_b->slack_timer,
2953 ns_to_ktime(cfs_bandwidth_slack_period));
2956 /* we know any runtime found here is valid as update_curr() precedes return */
2957 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2959 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2960 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2962 if (slack_runtime <= 0)
2965 raw_spin_lock(&cfs_b->lock);
2966 if (cfs_b->quota != RUNTIME_INF &&
2967 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2968 cfs_b->runtime += slack_runtime;
2970 /* we are under rq->lock, defer unthrottling using a timer */
2971 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2972 !list_empty(&cfs_b->throttled_cfs_rq))
2973 start_cfs_slack_bandwidth(cfs_b);
2975 raw_spin_unlock(&cfs_b->lock);
2977 /* even if it's not valid for return we don't want to try again */
2978 cfs_rq->runtime_remaining -= slack_runtime;
2981 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2983 if (!cfs_bandwidth_used())
2986 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2989 __return_cfs_rq_runtime(cfs_rq);
2993 * This is done with a timer (instead of inline with bandwidth return) since
2994 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2996 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2998 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3001 /* confirm we're still not at a refresh boundary */
3002 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
3005 raw_spin_lock(&cfs_b->lock);
3006 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3007 runtime = cfs_b->runtime;
3010 expires = cfs_b->runtime_expires;
3011 raw_spin_unlock(&cfs_b->lock);
3016 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3018 raw_spin_lock(&cfs_b->lock);
3019 if (expires == cfs_b->runtime_expires)
3020 cfs_b->runtime = runtime;
3021 raw_spin_unlock(&cfs_b->lock);
3025 * When a group wakes up we want to make sure that its quota is not already
3026 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3027 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3029 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3031 if (!cfs_bandwidth_used())
3034 /* an active group must be handled by the update_curr()->put() path */
3035 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3038 /* ensure the group is not already throttled */
3039 if (cfs_rq_throttled(cfs_rq))
3042 /* update runtime allocation */
3043 account_cfs_rq_runtime(cfs_rq, 0);
3044 if (cfs_rq->runtime_remaining <= 0)
3045 throttle_cfs_rq(cfs_rq);
3048 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3049 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3051 if (!cfs_bandwidth_used())
3054 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3058 * it's possible for a throttled entity to be forced into a running
3059 * state (e.g. set_curr_task), in this case we're finished.
3061 if (cfs_rq_throttled(cfs_rq))
3064 throttle_cfs_rq(cfs_rq);
3067 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3069 struct cfs_bandwidth *cfs_b =
3070 container_of(timer, struct cfs_bandwidth, slack_timer);
3071 do_sched_cfs_slack_timer(cfs_b);
3073 return HRTIMER_NORESTART;
3076 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3078 struct cfs_bandwidth *cfs_b =
3079 container_of(timer, struct cfs_bandwidth, period_timer);
3085 now = hrtimer_cb_get_time(timer);
3086 overrun = hrtimer_forward(timer, now, cfs_b->period);
3091 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3094 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3097 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3099 raw_spin_lock_init(&cfs_b->lock);
3101 cfs_b->quota = RUNTIME_INF;
3102 cfs_b->period = ns_to_ktime(default_cfs_period());
3104 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3105 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3106 cfs_b->period_timer.function = sched_cfs_period_timer;
3107 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3108 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3111 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3113 cfs_rq->runtime_enabled = 0;
3114 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3117 /* requires cfs_b->lock, may release to reprogram timer */
3118 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3121 * The timer may be active because we're trying to set a new bandwidth
3122 * period or because we're racing with the tear-down path
3123 * (timer_active==0 becomes visible before the hrtimer call-back
3124 * terminates). In either case we ensure that it's re-programmed
3126 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
3127 raw_spin_unlock(&cfs_b->lock);
3128 /* ensure cfs_b->lock is available while we wait */
3129 hrtimer_cancel(&cfs_b->period_timer);
3131 raw_spin_lock(&cfs_b->lock);
3132 /* if someone else restarted the timer then we're done */
3133 if (cfs_b->timer_active)
3137 cfs_b->timer_active = 1;
3138 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3141 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3143 hrtimer_cancel(&cfs_b->period_timer);
3144 hrtimer_cancel(&cfs_b->slack_timer);
3147 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3149 struct cfs_rq *cfs_rq;
3151 for_each_leaf_cfs_rq(rq, cfs_rq) {
3152 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3154 if (!cfs_rq->runtime_enabled)
3158 * clock_task is not advancing so we just need to make sure
3159 * there's some valid quota amount
3161 cfs_rq->runtime_remaining = cfs_b->quota;
3162 if (cfs_rq_throttled(cfs_rq))
3163 unthrottle_cfs_rq(cfs_rq);
3167 #else /* CONFIG_CFS_BANDWIDTH */
3168 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3170 return rq_clock_task(rq_of(cfs_rq));
3173 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3174 unsigned long delta_exec) {}
3175 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3176 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3177 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3179 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3184 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3189 static inline int throttled_lb_pair(struct task_group *tg,
3190 int src_cpu, int dest_cpu)
3195 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3197 #ifdef CONFIG_FAIR_GROUP_SCHED
3198 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3201 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3205 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3206 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3208 #endif /* CONFIG_CFS_BANDWIDTH */
3210 /**************************************************
3211 * CFS operations on tasks:
3214 #ifdef CONFIG_SCHED_HRTICK
3215 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3217 struct sched_entity *se = &p->se;
3218 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3220 WARN_ON(task_rq(p) != rq);
3222 if (cfs_rq->nr_running > 1) {
3223 u64 slice = sched_slice(cfs_rq, se);
3224 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3225 s64 delta = slice - ran;
3234 * Don't schedule slices shorter than 10000ns, that just
3235 * doesn't make sense. Rely on vruntime for fairness.
3238 delta = max_t(s64, 10000LL, delta);
3240 hrtick_start(rq, delta);
3245 * called from enqueue/dequeue and updates the hrtick when the
3246 * current task is from our class and nr_running is low enough
3249 static void hrtick_update(struct rq *rq)
3251 struct task_struct *curr = rq->curr;
3253 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3256 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3257 hrtick_start_fair(rq, curr);
3259 #else /* !CONFIG_SCHED_HRTICK */
3261 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3265 static inline void hrtick_update(struct rq *rq)
3271 * The enqueue_task method is called before nr_running is
3272 * increased. Here we update the fair scheduling stats and
3273 * then put the task into the rbtree:
3276 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3278 struct cfs_rq *cfs_rq;
3279 struct sched_entity *se = &p->se;
3281 for_each_sched_entity(se) {
3284 cfs_rq = cfs_rq_of(se);
3285 enqueue_entity(cfs_rq, se, flags);
3288 * end evaluation on encountering a throttled cfs_rq
3290 * note: in the case of encountering a throttled cfs_rq we will
3291 * post the final h_nr_running increment below.
3293 if (cfs_rq_throttled(cfs_rq))
3295 cfs_rq->h_nr_running++;
3297 flags = ENQUEUE_WAKEUP;
3300 for_each_sched_entity(se) {
3301 cfs_rq = cfs_rq_of(se);
3302 cfs_rq->h_nr_running++;
3304 if (cfs_rq_throttled(cfs_rq))
3307 update_cfs_shares(cfs_rq);
3308 update_entity_load_avg(se, 1);
3312 update_rq_runnable_avg(rq, rq->nr_running);
3318 static void set_next_buddy(struct sched_entity *se);
3321 * The dequeue_task method is called before nr_running is
3322 * decreased. We remove the task from the rbtree and
3323 * update the fair scheduling stats:
3325 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3327 struct cfs_rq *cfs_rq;
3328 struct sched_entity *se = &p->se;
3329 int task_sleep = flags & DEQUEUE_SLEEP;
3331 for_each_sched_entity(se) {
3332 cfs_rq = cfs_rq_of(se);
3333 dequeue_entity(cfs_rq, se, flags);
3336 * end evaluation on encountering a throttled cfs_rq
3338 * note: in the case of encountering a throttled cfs_rq we will
3339 * post the final h_nr_running decrement below.
3341 if (cfs_rq_throttled(cfs_rq))
3343 cfs_rq->h_nr_running--;
3345 /* Don't dequeue parent if it has other entities besides us */
3346 if (cfs_rq->load.weight) {
3348 * Bias pick_next to pick a task from this cfs_rq, as
3349 * p is sleeping when it is within its sched_slice.
3351 if (task_sleep && parent_entity(se))
3352 set_next_buddy(parent_entity(se));
3354 /* avoid re-evaluating load for this entity */
3355 se = parent_entity(se);
3358 flags |= DEQUEUE_SLEEP;
3361 for_each_sched_entity(se) {
3362 cfs_rq = cfs_rq_of(se);
3363 cfs_rq->h_nr_running--;
3365 if (cfs_rq_throttled(cfs_rq))
3368 update_cfs_shares(cfs_rq);
3369 update_entity_load_avg(se, 1);
3374 update_rq_runnable_avg(rq, 1);
3380 /* Used instead of source_load when we know the type == 0 */
3381 static unsigned long weighted_cpuload(const int cpu)
3383 return cpu_rq(cpu)->cfs.runnable_load_avg;
3387 * Return a low guess at the load of a migration-source cpu weighted
3388 * according to the scheduling class and "nice" value.
3390 * We want to under-estimate the load of migration sources, to
3391 * balance conservatively.
3393 static unsigned long source_load(int cpu, int type)
3395 struct rq *rq = cpu_rq(cpu);
3396 unsigned long total = weighted_cpuload(cpu);
3398 if (type == 0 || !sched_feat(LB_BIAS))
3401 return min(rq->cpu_load[type-1], total);
3405 * Return a high guess at the load of a migration-target cpu weighted
3406 * according to the scheduling class and "nice" value.
3408 static unsigned long target_load(int cpu, int type)
3410 struct rq *rq = cpu_rq(cpu);
3411 unsigned long total = weighted_cpuload(cpu);
3413 if (type == 0 || !sched_feat(LB_BIAS))
3416 return max(rq->cpu_load[type-1], total);
3419 static unsigned long power_of(int cpu)
3421 return cpu_rq(cpu)->cpu_power;
3424 static unsigned long cpu_avg_load_per_task(int cpu)
3426 struct rq *rq = cpu_rq(cpu);
3427 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3428 unsigned long load_avg = rq->cfs.runnable_load_avg;
3431 return load_avg / nr_running;
3436 static void record_wakee(struct task_struct *p)
3439 * Rough decay (wiping) for cost saving, don't worry
3440 * about the boundary, really active task won't care
3443 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3444 current->wakee_flips = 0;
3445 current->wakee_flip_decay_ts = jiffies;
3448 if (current->last_wakee != p) {
3449 current->last_wakee = p;
3450 current->wakee_flips++;
3454 static void task_waking_fair(struct task_struct *p)
3456 struct sched_entity *se = &p->se;
3457 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3460 #ifndef CONFIG_64BIT
3461 u64 min_vruntime_copy;
3464 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3466 min_vruntime = cfs_rq->min_vruntime;
3467 } while (min_vruntime != min_vruntime_copy);
3469 min_vruntime = cfs_rq->min_vruntime;
3472 se->vruntime -= min_vruntime;
3476 #ifdef CONFIG_FAIR_GROUP_SCHED
3478 * effective_load() calculates the load change as seen from the root_task_group
3480 * Adding load to a group doesn't make a group heavier, but can cause movement
3481 * of group shares between cpus. Assuming the shares were perfectly aligned one
3482 * can calculate the shift in shares.
3484 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3485 * on this @cpu and results in a total addition (subtraction) of @wg to the
3486 * total group weight.
3488 * Given a runqueue weight distribution (rw_i) we can compute a shares
3489 * distribution (s_i) using:
3491 * s_i = rw_i / \Sum rw_j (1)
3493 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3494 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3495 * shares distribution (s_i):
3497 * rw_i = { 2, 4, 1, 0 }
3498 * s_i = { 2/7, 4/7, 1/7, 0 }
3500 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3501 * task used to run on and the CPU the waker is running on), we need to
3502 * compute the effect of waking a task on either CPU and, in case of a sync
3503 * wakeup, compute the effect of the current task going to sleep.
3505 * So for a change of @wl to the local @cpu with an overall group weight change
3506 * of @wl we can compute the new shares distribution (s'_i) using:
3508 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3510 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3511 * differences in waking a task to CPU 0. The additional task changes the
3512 * weight and shares distributions like:
3514 * rw'_i = { 3, 4, 1, 0 }
3515 * s'_i = { 3/8, 4/8, 1/8, 0 }
3517 * We can then compute the difference in effective weight by using:
3519 * dw_i = S * (s'_i - s_i) (3)
3521 * Where 'S' is the group weight as seen by its parent.
3523 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3524 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3525 * 4/7) times the weight of the group.
3527 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3529 struct sched_entity *se = tg->se[cpu];
3531 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3534 for_each_sched_entity(se) {
3540 * W = @wg + \Sum rw_j
3542 W = wg + calc_tg_weight(tg, se->my_q);
3547 w = se->my_q->load.weight + wl;
3550 * wl = S * s'_i; see (2)
3553 wl = (w * tg->shares) / W;
3558 * Per the above, wl is the new se->load.weight value; since
3559 * those are clipped to [MIN_SHARES, ...) do so now. See
3560 * calc_cfs_shares().
3562 if (wl < MIN_SHARES)
3566 * wl = dw_i = S * (s'_i - s_i); see (3)
3568 wl -= se->load.weight;
3571 * Recursively apply this logic to all parent groups to compute
3572 * the final effective load change on the root group. Since
3573 * only the @tg group gets extra weight, all parent groups can
3574 * only redistribute existing shares. @wl is the shift in shares
3575 * resulting from this level per the above.
3584 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3591 static int wake_wide(struct task_struct *p)
3593 int factor = this_cpu_read(sd_llc_size);
3596 * Yeah, it's the switching-frequency, could means many wakee or
3597 * rapidly switch, use factor here will just help to automatically
3598 * adjust the loose-degree, so bigger node will lead to more pull.
3600 if (p->wakee_flips > factor) {
3602 * wakee is somewhat hot, it needs certain amount of cpu
3603 * resource, so if waker is far more hot, prefer to leave
3606 if (current->wakee_flips > (factor * p->wakee_flips))
3613 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3615 s64 this_load, load;
3616 int idx, this_cpu, prev_cpu;
3617 unsigned long tl_per_task;
3618 struct task_group *tg;
3619 unsigned long weight;
3623 * If we wake multiple tasks be careful to not bounce
3624 * ourselves around too much.
3630 this_cpu = smp_processor_id();
3631 prev_cpu = task_cpu(p);
3632 load = source_load(prev_cpu, idx);
3633 this_load = target_load(this_cpu, idx);
3636 * If sync wakeup then subtract the (maximum possible)
3637 * effect of the currently running task from the load
3638 * of the current CPU:
3641 tg = task_group(current);
3642 weight = current->se.load.weight;
3644 this_load += effective_load(tg, this_cpu, -weight, -weight);
3645 load += effective_load(tg, prev_cpu, 0, -weight);
3649 weight = p->se.load.weight;
3652 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3653 * due to the sync cause above having dropped this_load to 0, we'll
3654 * always have an imbalance, but there's really nothing you can do
3655 * about that, so that's good too.
3657 * Otherwise check if either cpus are near enough in load to allow this
3658 * task to be woken on this_cpu.
3660 if (this_load > 0) {
3661 s64 this_eff_load, prev_eff_load;
3663 this_eff_load = 100;
3664 this_eff_load *= power_of(prev_cpu);
3665 this_eff_load *= this_load +
3666 effective_load(tg, this_cpu, weight, weight);
3668 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3669 prev_eff_load *= power_of(this_cpu);
3670 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3672 balanced = this_eff_load <= prev_eff_load;
3677 * If the currently running task will sleep within
3678 * a reasonable amount of time then attract this newly
3681 if (sync && balanced)
3684 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3685 tl_per_task = cpu_avg_load_per_task(this_cpu);
3688 (this_load <= load &&
3689 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3691 * This domain has SD_WAKE_AFFINE and
3692 * p is cache cold in this domain, and
3693 * there is no bad imbalance.
3695 schedstat_inc(sd, ttwu_move_affine);
3696 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3704 * find_idlest_group finds and returns the least busy CPU group within the
3707 static struct sched_group *
3708 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3709 int this_cpu, int load_idx)
3711 struct sched_group *idlest = NULL, *group = sd->groups;
3712 unsigned long min_load = ULONG_MAX, this_load = 0;
3713 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3716 unsigned long load, avg_load;
3720 /* Skip over this group if it has no CPUs allowed */
3721 if (!cpumask_intersects(sched_group_cpus(group),
3722 tsk_cpus_allowed(p)))
3725 local_group = cpumask_test_cpu(this_cpu,
3726 sched_group_cpus(group));
3728 /* Tally up the load of all CPUs in the group */
3731 for_each_cpu(i, sched_group_cpus(group)) {
3732 /* Bias balancing toward cpus of our domain */
3734 load = source_load(i, load_idx);
3736 load = target_load(i, load_idx);
3741 /* Adjust by relative CPU power of the group */
3742 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3745 this_load = avg_load;
3746 } else if (avg_load < min_load) {
3747 min_load = avg_load;
3750 } while (group = group->next, group != sd->groups);
3752 if (!idlest || 100*this_load < imbalance*min_load)
3758 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3761 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3763 unsigned long load, min_load = ULONG_MAX;
3767 /* Traverse only the allowed CPUs */
3768 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3769 load = weighted_cpuload(i);
3771 if (load < min_load || (load == min_load && i == this_cpu)) {
3781 * Try and locate an idle CPU in the sched_domain.
3783 static int select_idle_sibling(struct task_struct *p, int target)
3785 struct sched_domain *sd;
3786 struct sched_group *sg;
3787 int i = task_cpu(p);
3789 if (idle_cpu(target))
3793 * If the prevous cpu is cache affine and idle, don't be stupid.
3795 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3799 * Otherwise, iterate the domains and find an elegible idle cpu.
3801 sd = rcu_dereference(per_cpu(sd_llc, target));
3802 for_each_lower_domain(sd) {
3805 if (!cpumask_intersects(sched_group_cpus(sg),
3806 tsk_cpus_allowed(p)))
3809 for_each_cpu(i, sched_group_cpus(sg)) {
3810 if (i == target || !idle_cpu(i))
3814 target = cpumask_first_and(sched_group_cpus(sg),
3815 tsk_cpus_allowed(p));
3819 } while (sg != sd->groups);
3826 * sched_balance_self: balance the current task (running on cpu) in domains
3827 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3830 * Balance, ie. select the least loaded group.
3832 * Returns the target CPU number, or the same CPU if no balancing is needed.
3834 * preempt must be disabled.
3837 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
3839 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3840 int cpu = smp_processor_id();
3842 int want_affine = 0;
3843 int sync = wake_flags & WF_SYNC;
3845 if (p->nr_cpus_allowed == 1)
3848 if (sd_flag & SD_BALANCE_WAKE) {
3849 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3855 for_each_domain(cpu, tmp) {
3856 if (!(tmp->flags & SD_LOAD_BALANCE))
3860 * If both cpu and prev_cpu are part of this domain,
3861 * cpu is a valid SD_WAKE_AFFINE target.
3863 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3864 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3869 if (tmp->flags & sd_flag)
3874 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3877 new_cpu = select_idle_sibling(p, prev_cpu);
3882 int load_idx = sd->forkexec_idx;
3883 struct sched_group *group;
3886 if (!(sd->flags & sd_flag)) {
3891 if (sd_flag & SD_BALANCE_WAKE)
3892 load_idx = sd->wake_idx;
3894 group = find_idlest_group(sd, p, cpu, load_idx);
3900 new_cpu = find_idlest_cpu(group, p, cpu);
3901 if (new_cpu == -1 || new_cpu == cpu) {
3902 /* Now try balancing at a lower domain level of cpu */
3907 /* Now try balancing at a lower domain level of new_cpu */
3909 weight = sd->span_weight;
3911 for_each_domain(cpu, tmp) {
3912 if (weight <= tmp->span_weight)
3914 if (tmp->flags & sd_flag)
3917 /* while loop will break here if sd == NULL */
3926 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3927 * cfs_rq_of(p) references at time of call are still valid and identify the
3928 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3929 * other assumptions, including the state of rq->lock, should be made.
3932 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3934 struct sched_entity *se = &p->se;
3935 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3938 * Load tracking: accumulate removed load so that it can be processed
3939 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3940 * to blocked load iff they have a positive decay-count. It can never
3941 * be negative here since on-rq tasks have decay-count == 0.
3943 if (se->avg.decay_count) {
3944 se->avg.decay_count = -__synchronize_entity_decay(se);
3945 atomic_long_add(se->avg.load_avg_contrib,
3946 &cfs_rq->removed_load);
3949 #endif /* CONFIG_SMP */
3951 static unsigned long
3952 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3954 unsigned long gran = sysctl_sched_wakeup_granularity;
3957 * Since its curr running now, convert the gran from real-time
3958 * to virtual-time in his units.
3960 * By using 'se' instead of 'curr' we penalize light tasks, so
3961 * they get preempted easier. That is, if 'se' < 'curr' then
3962 * the resulting gran will be larger, therefore penalizing the
3963 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3964 * be smaller, again penalizing the lighter task.
3966 * This is especially important for buddies when the leftmost
3967 * task is higher priority than the buddy.
3969 return calc_delta_fair(gran, se);
3973 * Should 'se' preempt 'curr'.
3987 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3989 s64 gran, vdiff = curr->vruntime - se->vruntime;
3994 gran = wakeup_gran(curr, se);
4001 static void set_last_buddy(struct sched_entity *se)
4003 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4006 for_each_sched_entity(se)
4007 cfs_rq_of(se)->last = se;
4010 static void set_next_buddy(struct sched_entity *se)
4012 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4015 for_each_sched_entity(se)
4016 cfs_rq_of(se)->next = se;
4019 static void set_skip_buddy(struct sched_entity *se)
4021 for_each_sched_entity(se)
4022 cfs_rq_of(se)->skip = se;
4026 * Preempt the current task with a newly woken task if needed:
4028 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4030 struct task_struct *curr = rq->curr;
4031 struct sched_entity *se = &curr->se, *pse = &p->se;
4032 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4033 int scale = cfs_rq->nr_running >= sched_nr_latency;
4034 int next_buddy_marked = 0;
4036 if (unlikely(se == pse))
4040 * This is possible from callers such as move_task(), in which we
4041 * unconditionally check_prempt_curr() after an enqueue (which may have
4042 * lead to a throttle). This both saves work and prevents false
4043 * next-buddy nomination below.
4045 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4048 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4049 set_next_buddy(pse);
4050 next_buddy_marked = 1;
4054 * We can come here with TIF_NEED_RESCHED already set from new task
4057 * Note: this also catches the edge-case of curr being in a throttled
4058 * group (e.g. via set_curr_task), since update_curr() (in the
4059 * enqueue of curr) will have resulted in resched being set. This
4060 * prevents us from potentially nominating it as a false LAST_BUDDY
4063 if (test_tsk_need_resched(curr))
4066 /* Idle tasks are by definition preempted by non-idle tasks. */
4067 if (unlikely(curr->policy == SCHED_IDLE) &&
4068 likely(p->policy != SCHED_IDLE))
4072 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4073 * is driven by the tick):
4075 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4078 find_matching_se(&se, &pse);
4079 update_curr(cfs_rq_of(se));
4081 if (wakeup_preempt_entity(se, pse) == 1) {
4083 * Bias pick_next to pick the sched entity that is
4084 * triggering this preemption.
4086 if (!next_buddy_marked)
4087 set_next_buddy(pse);
4096 * Only set the backward buddy when the current task is still
4097 * on the rq. This can happen when a wakeup gets interleaved
4098 * with schedule on the ->pre_schedule() or idle_balance()
4099 * point, either of which can * drop the rq lock.
4101 * Also, during early boot the idle thread is in the fair class,
4102 * for obvious reasons its a bad idea to schedule back to it.
4104 if (unlikely(!se->on_rq || curr == rq->idle))
4107 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4111 static struct task_struct *pick_next_task_fair(struct rq *rq)
4113 struct task_struct *p;
4114 struct cfs_rq *cfs_rq = &rq->cfs;
4115 struct sched_entity *se;
4117 if (!cfs_rq->nr_running)
4121 se = pick_next_entity(cfs_rq);
4122 set_next_entity(cfs_rq, se);
4123 cfs_rq = group_cfs_rq(se);
4127 if (hrtick_enabled(rq))
4128 hrtick_start_fair(rq, p);
4134 * Account for a descheduled task:
4136 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4138 struct sched_entity *se = &prev->se;
4139 struct cfs_rq *cfs_rq;
4141 for_each_sched_entity(se) {
4142 cfs_rq = cfs_rq_of(se);
4143 put_prev_entity(cfs_rq, se);
4148 * sched_yield() is very simple
4150 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4152 static void yield_task_fair(struct rq *rq)
4154 struct task_struct *curr = rq->curr;
4155 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4156 struct sched_entity *se = &curr->se;
4159 * Are we the only task in the tree?
4161 if (unlikely(rq->nr_running == 1))
4164 clear_buddies(cfs_rq, se);
4166 if (curr->policy != SCHED_BATCH) {
4167 update_rq_clock(rq);
4169 * Update run-time statistics of the 'current'.
4171 update_curr(cfs_rq);
4173 * Tell update_rq_clock() that we've just updated,
4174 * so we don't do microscopic update in schedule()
4175 * and double the fastpath cost.
4177 rq->skip_clock_update = 1;
4183 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4185 struct sched_entity *se = &p->se;
4187 /* throttled hierarchies are not runnable */
4188 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4191 /* Tell the scheduler that we'd really like pse to run next. */
4194 yield_task_fair(rq);
4200 /**************************************************
4201 * Fair scheduling class load-balancing methods.
4205 * The purpose of load-balancing is to achieve the same basic fairness the
4206 * per-cpu scheduler provides, namely provide a proportional amount of compute
4207 * time to each task. This is expressed in the following equation:
4209 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4211 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4212 * W_i,0 is defined as:
4214 * W_i,0 = \Sum_j w_i,j (2)
4216 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4217 * is derived from the nice value as per prio_to_weight[].
4219 * The weight average is an exponential decay average of the instantaneous
4222 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4224 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4225 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4226 * can also include other factors [XXX].
4228 * To achieve this balance we define a measure of imbalance which follows
4229 * directly from (1):
4231 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4233 * We them move tasks around to minimize the imbalance. In the continuous
4234 * function space it is obvious this converges, in the discrete case we get
4235 * a few fun cases generally called infeasible weight scenarios.
4238 * - infeasible weights;
4239 * - local vs global optima in the discrete case. ]
4244 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4245 * for all i,j solution, we create a tree of cpus that follows the hardware
4246 * topology where each level pairs two lower groups (or better). This results
4247 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4248 * tree to only the first of the previous level and we decrease the frequency
4249 * of load-balance at each level inv. proportional to the number of cpus in
4255 * \Sum { --- * --- * 2^i } = O(n) (5)
4257 * `- size of each group
4258 * | | `- number of cpus doing load-balance
4260 * `- sum over all levels
4262 * Coupled with a limit on how many tasks we can migrate every balance pass,
4263 * this makes (5) the runtime complexity of the balancer.
4265 * An important property here is that each CPU is still (indirectly) connected
4266 * to every other cpu in at most O(log n) steps:
4268 * The adjacency matrix of the resulting graph is given by:
4271 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4274 * And you'll find that:
4276 * A^(log_2 n)_i,j != 0 for all i,j (7)
4278 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4279 * The task movement gives a factor of O(m), giving a convergence complexity
4282 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4287 * In order to avoid CPUs going idle while there's still work to do, new idle
4288 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4289 * tree itself instead of relying on other CPUs to bring it work.
4291 * This adds some complexity to both (5) and (8) but it reduces the total idle
4299 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4302 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4307 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4309 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4311 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4314 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4315 * rewrite all of this once again.]
4318 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4320 #define LBF_ALL_PINNED 0x01
4321 #define LBF_NEED_BREAK 0x02
4322 #define LBF_DST_PINNED 0x04
4323 #define LBF_SOME_PINNED 0x08
4326 struct sched_domain *sd;
4334 struct cpumask *dst_grpmask;
4336 enum cpu_idle_type idle;
4338 /* The set of CPUs under consideration for load-balancing */
4339 struct cpumask *cpus;
4344 unsigned int loop_break;
4345 unsigned int loop_max;
4349 * move_task - move a task from one runqueue to another runqueue.
4350 * Both runqueues must be locked.
4352 static void move_task(struct task_struct *p, struct lb_env *env)
4354 deactivate_task(env->src_rq, p, 0);
4355 set_task_cpu(p, env->dst_cpu);
4356 activate_task(env->dst_rq, p, 0);
4357 check_preempt_curr(env->dst_rq, p, 0);
4358 #ifdef CONFIG_NUMA_BALANCING
4359 if (p->numa_preferred_nid != -1) {
4360 int src_nid = cpu_to_node(env->src_cpu);
4361 int dst_nid = cpu_to_node(env->dst_cpu);
4364 * If the load balancer has moved the task then limit
4365 * migrations from taking place in the short term in
4366 * case this is a short-lived migration.
4368 if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid)
4369 p->numa_migrate_seq = 0;
4375 * Is this task likely cache-hot:
4378 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4382 if (p->sched_class != &fair_sched_class)
4385 if (unlikely(p->policy == SCHED_IDLE))
4389 * Buddy candidates are cache hot:
4391 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4392 (&p->se == cfs_rq_of(&p->se)->next ||
4393 &p->se == cfs_rq_of(&p->se)->last))
4396 if (sysctl_sched_migration_cost == -1)
4398 if (sysctl_sched_migration_cost == 0)
4401 delta = now - p->se.exec_start;
4403 return delta < (s64)sysctl_sched_migration_cost;
4406 #ifdef CONFIG_NUMA_BALANCING
4407 /* Returns true if the destination node has incurred more faults */
4408 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4410 int src_nid, dst_nid;
4412 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4413 !(env->sd->flags & SD_NUMA)) {
4417 src_nid = cpu_to_node(env->src_cpu);
4418 dst_nid = cpu_to_node(env->dst_cpu);
4420 if (src_nid == dst_nid ||
4421 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4424 if (dst_nid == p->numa_preferred_nid ||
4425 task_faults(p, dst_nid) > task_faults(p, src_nid))
4432 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4434 int src_nid, dst_nid;
4436 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4439 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4442 src_nid = cpu_to_node(env->src_cpu);
4443 dst_nid = cpu_to_node(env->dst_cpu);
4445 if (src_nid == dst_nid ||
4446 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4449 if (task_faults(p, dst_nid) < task_faults(p, src_nid))
4456 static inline bool migrate_improves_locality(struct task_struct *p,
4462 static inline bool migrate_degrades_locality(struct task_struct *p,
4470 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4473 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4475 int tsk_cache_hot = 0;
4477 * We do not migrate tasks that are:
4478 * 1) throttled_lb_pair, or
4479 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4480 * 3) running (obviously), or
4481 * 4) are cache-hot on their current CPU.
4483 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4486 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4489 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4491 env->flags |= LBF_SOME_PINNED;
4494 * Remember if this task can be migrated to any other cpu in
4495 * our sched_group. We may want to revisit it if we couldn't
4496 * meet load balance goals by pulling other tasks on src_cpu.
4498 * Also avoid computing new_dst_cpu if we have already computed
4499 * one in current iteration.
4501 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4504 /* Prevent to re-select dst_cpu via env's cpus */
4505 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4506 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4507 env->flags |= LBF_DST_PINNED;
4508 env->new_dst_cpu = cpu;
4516 /* Record that we found atleast one task that could run on dst_cpu */
4517 env->flags &= ~LBF_ALL_PINNED;
4519 if (task_running(env->src_rq, p)) {
4520 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4525 * Aggressive migration if:
4526 * 1) destination numa is preferred
4527 * 2) task is cache cold, or
4528 * 3) too many balance attempts have failed.
4530 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4532 tsk_cache_hot = migrate_degrades_locality(p, env);
4534 if (migrate_improves_locality(p, env)) {
4535 #ifdef CONFIG_SCHEDSTATS
4536 if (tsk_cache_hot) {
4537 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4538 schedstat_inc(p, se.statistics.nr_forced_migrations);
4544 if (!tsk_cache_hot ||
4545 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4547 if (tsk_cache_hot) {
4548 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4549 schedstat_inc(p, se.statistics.nr_forced_migrations);
4555 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4560 * move_one_task tries to move exactly one task from busiest to this_rq, as
4561 * part of active balancing operations within "domain".
4562 * Returns 1 if successful and 0 otherwise.
4564 * Called with both runqueues locked.
4566 static int move_one_task(struct lb_env *env)
4568 struct task_struct *p, *n;
4570 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4571 if (!can_migrate_task(p, env))
4576 * Right now, this is only the second place move_task()
4577 * is called, so we can safely collect move_task()
4578 * stats here rather than inside move_task().
4580 schedstat_inc(env->sd, lb_gained[env->idle]);
4586 static const unsigned int sched_nr_migrate_break = 32;
4589 * move_tasks tries to move up to imbalance weighted load from busiest to
4590 * this_rq, as part of a balancing operation within domain "sd".
4591 * Returns 1 if successful and 0 otherwise.
4593 * Called with both runqueues locked.
4595 static int move_tasks(struct lb_env *env)
4597 struct list_head *tasks = &env->src_rq->cfs_tasks;
4598 struct task_struct *p;
4602 if (env->imbalance <= 0)
4605 while (!list_empty(tasks)) {
4606 p = list_first_entry(tasks, struct task_struct, se.group_node);
4609 /* We've more or less seen every task there is, call it quits */
4610 if (env->loop > env->loop_max)
4613 /* take a breather every nr_migrate tasks */
4614 if (env->loop > env->loop_break) {
4615 env->loop_break += sched_nr_migrate_break;
4616 env->flags |= LBF_NEED_BREAK;
4620 if (!can_migrate_task(p, env))
4623 load = task_h_load(p);
4625 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4628 if ((load / 2) > env->imbalance)
4633 env->imbalance -= load;
4635 #ifdef CONFIG_PREEMPT
4637 * NEWIDLE balancing is a source of latency, so preemptible
4638 * kernels will stop after the first task is pulled to minimize
4639 * the critical section.
4641 if (env->idle == CPU_NEWLY_IDLE)
4646 * We only want to steal up to the prescribed amount of
4649 if (env->imbalance <= 0)
4654 list_move_tail(&p->se.group_node, tasks);
4658 * Right now, this is one of only two places move_task() is called,
4659 * so we can safely collect move_task() stats here rather than
4660 * inside move_task().
4662 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4667 #ifdef CONFIG_FAIR_GROUP_SCHED
4669 * update tg->load_weight by folding this cpu's load_avg
4671 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4673 struct sched_entity *se = tg->se[cpu];
4674 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4676 /* throttled entities do not contribute to load */
4677 if (throttled_hierarchy(cfs_rq))
4680 update_cfs_rq_blocked_load(cfs_rq, 1);
4683 update_entity_load_avg(se, 1);
4685 * We pivot on our runnable average having decayed to zero for
4686 * list removal. This generally implies that all our children
4687 * have also been removed (modulo rounding error or bandwidth
4688 * control); however, such cases are rare and we can fix these
4691 * TODO: fix up out-of-order children on enqueue.
4693 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4694 list_del_leaf_cfs_rq(cfs_rq);
4696 struct rq *rq = rq_of(cfs_rq);
4697 update_rq_runnable_avg(rq, rq->nr_running);
4701 static void update_blocked_averages(int cpu)
4703 struct rq *rq = cpu_rq(cpu);
4704 struct cfs_rq *cfs_rq;
4705 unsigned long flags;
4707 raw_spin_lock_irqsave(&rq->lock, flags);
4708 update_rq_clock(rq);
4710 * Iterates the task_group tree in a bottom up fashion, see
4711 * list_add_leaf_cfs_rq() for details.
4713 for_each_leaf_cfs_rq(rq, cfs_rq) {
4715 * Note: We may want to consider periodically releasing
4716 * rq->lock about these updates so that creating many task
4717 * groups does not result in continually extending hold time.
4719 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4722 raw_spin_unlock_irqrestore(&rq->lock, flags);
4726 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4727 * This needs to be done in a top-down fashion because the load of a child
4728 * group is a fraction of its parents load.
4730 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4732 struct rq *rq = rq_of(cfs_rq);
4733 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4734 unsigned long now = jiffies;
4737 if (cfs_rq->last_h_load_update == now)
4740 cfs_rq->h_load_next = NULL;
4741 for_each_sched_entity(se) {
4742 cfs_rq = cfs_rq_of(se);
4743 cfs_rq->h_load_next = se;
4744 if (cfs_rq->last_h_load_update == now)
4749 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4750 cfs_rq->last_h_load_update = now;
4753 while ((se = cfs_rq->h_load_next) != NULL) {
4754 load = cfs_rq->h_load;
4755 load = div64_ul(load * se->avg.load_avg_contrib,
4756 cfs_rq->runnable_load_avg + 1);
4757 cfs_rq = group_cfs_rq(se);
4758 cfs_rq->h_load = load;
4759 cfs_rq->last_h_load_update = now;
4763 static unsigned long task_h_load(struct task_struct *p)
4765 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4767 update_cfs_rq_h_load(cfs_rq);
4768 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4769 cfs_rq->runnable_load_avg + 1);
4772 static inline void update_blocked_averages(int cpu)
4776 static unsigned long task_h_load(struct task_struct *p)
4778 return p->se.avg.load_avg_contrib;
4782 /********** Helpers for find_busiest_group ************************/
4784 * sg_lb_stats - stats of a sched_group required for load_balancing
4786 struct sg_lb_stats {
4787 unsigned long avg_load; /*Avg load across the CPUs of the group */
4788 unsigned long group_load; /* Total load over the CPUs of the group */
4789 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4790 unsigned long load_per_task;
4791 unsigned long group_power;
4792 unsigned int sum_nr_running; /* Nr tasks running in the group */
4793 unsigned int group_capacity;
4794 unsigned int idle_cpus;
4795 unsigned int group_weight;
4796 int group_imb; /* Is there an imbalance in the group ? */
4797 int group_has_capacity; /* Is there extra capacity in the group? */
4801 * sd_lb_stats - Structure to store the statistics of a sched_domain
4802 * during load balancing.
4804 struct sd_lb_stats {
4805 struct sched_group *busiest; /* Busiest group in this sd */
4806 struct sched_group *local; /* Local group in this sd */
4807 unsigned long total_load; /* Total load of all groups in sd */
4808 unsigned long total_pwr; /* Total power of all groups in sd */
4809 unsigned long avg_load; /* Average load across all groups in sd */
4811 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4812 struct sg_lb_stats local_stat; /* Statistics of the local group */
4815 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4818 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4819 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4820 * We must however clear busiest_stat::avg_load because
4821 * update_sd_pick_busiest() reads this before assignment.
4823 *sds = (struct sd_lb_stats){
4835 * get_sd_load_idx - Obtain the load index for a given sched domain.
4836 * @sd: The sched_domain whose load_idx is to be obtained.
4837 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4839 * Return: The load index.
4841 static inline int get_sd_load_idx(struct sched_domain *sd,
4842 enum cpu_idle_type idle)
4848 load_idx = sd->busy_idx;
4851 case CPU_NEWLY_IDLE:
4852 load_idx = sd->newidle_idx;
4855 load_idx = sd->idle_idx;
4862 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4864 return SCHED_POWER_SCALE;
4867 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4869 return default_scale_freq_power(sd, cpu);
4872 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4874 unsigned long weight = sd->span_weight;
4875 unsigned long smt_gain = sd->smt_gain;
4882 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4884 return default_scale_smt_power(sd, cpu);
4887 static unsigned long scale_rt_power(int cpu)
4889 struct rq *rq = cpu_rq(cpu);
4890 u64 total, available, age_stamp, avg;
4893 * Since we're reading these variables without serialization make sure
4894 * we read them once before doing sanity checks on them.
4896 age_stamp = ACCESS_ONCE(rq->age_stamp);
4897 avg = ACCESS_ONCE(rq->rt_avg);
4899 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4901 if (unlikely(total < avg)) {
4902 /* Ensures that power won't end up being negative */
4905 available = total - avg;
4908 if (unlikely((s64)total < SCHED_POWER_SCALE))
4909 total = SCHED_POWER_SCALE;
4911 total >>= SCHED_POWER_SHIFT;
4913 return div_u64(available, total);
4916 static void update_cpu_power(struct sched_domain *sd, int cpu)
4918 unsigned long weight = sd->span_weight;
4919 unsigned long power = SCHED_POWER_SCALE;
4920 struct sched_group *sdg = sd->groups;
4922 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4923 if (sched_feat(ARCH_POWER))
4924 power *= arch_scale_smt_power(sd, cpu);
4926 power *= default_scale_smt_power(sd, cpu);
4928 power >>= SCHED_POWER_SHIFT;
4931 sdg->sgp->power_orig = power;
4933 if (sched_feat(ARCH_POWER))
4934 power *= arch_scale_freq_power(sd, cpu);
4936 power *= default_scale_freq_power(sd, cpu);
4938 power >>= SCHED_POWER_SHIFT;
4940 power *= scale_rt_power(cpu);
4941 power >>= SCHED_POWER_SHIFT;
4946 cpu_rq(cpu)->cpu_power = power;
4947 sdg->sgp->power = power;
4950 void update_group_power(struct sched_domain *sd, int cpu)
4952 struct sched_domain *child = sd->child;
4953 struct sched_group *group, *sdg = sd->groups;
4954 unsigned long power, power_orig;
4955 unsigned long interval;
4957 interval = msecs_to_jiffies(sd->balance_interval);
4958 interval = clamp(interval, 1UL, max_load_balance_interval);
4959 sdg->sgp->next_update = jiffies + interval;
4962 update_cpu_power(sd, cpu);
4966 power_orig = power = 0;
4968 if (child->flags & SD_OVERLAP) {
4970 * SD_OVERLAP domains cannot assume that child groups
4971 * span the current group.
4974 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4975 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4977 power_orig += sg->sgp->power_orig;
4978 power += sg->sgp->power;
4982 * !SD_OVERLAP domains can assume that child groups
4983 * span the current group.
4986 group = child->groups;
4988 power_orig += group->sgp->power_orig;
4989 power += group->sgp->power;
4990 group = group->next;
4991 } while (group != child->groups);
4994 sdg->sgp->power_orig = power_orig;
4995 sdg->sgp->power = power;
4999 * Try and fix up capacity for tiny siblings, this is needed when
5000 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5001 * which on its own isn't powerful enough.
5003 * See update_sd_pick_busiest() and check_asym_packing().
5006 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5009 * Only siblings can have significantly less than SCHED_POWER_SCALE
5011 if (!(sd->flags & SD_SHARE_CPUPOWER))
5015 * If ~90% of the cpu_power is still there, we're good.
5017 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5024 * Group imbalance indicates (and tries to solve) the problem where balancing
5025 * groups is inadequate due to tsk_cpus_allowed() constraints.
5027 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5028 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5031 * { 0 1 2 3 } { 4 5 6 7 }
5034 * If we were to balance group-wise we'd place two tasks in the first group and
5035 * two tasks in the second group. Clearly this is undesired as it will overload
5036 * cpu 3 and leave one of the cpus in the second group unused.
5038 * The current solution to this issue is detecting the skew in the first group
5039 * by noticing the lower domain failed to reach balance and had difficulty
5040 * moving tasks due to affinity constraints.
5042 * When this is so detected; this group becomes a candidate for busiest; see
5043 * update_sd_pick_busiest(). And calculcate_imbalance() and
5044 * find_busiest_group() avoid some of the usual balance conditions to allow it
5045 * to create an effective group imbalance.
5047 * This is a somewhat tricky proposition since the next run might not find the
5048 * group imbalance and decide the groups need to be balanced again. A most
5049 * subtle and fragile situation.
5052 static inline int sg_imbalanced(struct sched_group *group)
5054 return group->sgp->imbalance;
5058 * Compute the group capacity.
5060 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5061 * first dividing out the smt factor and computing the actual number of cores
5062 * and limit power unit capacity with that.
5064 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5066 unsigned int capacity, smt, cpus;
5067 unsigned int power, power_orig;
5069 power = group->sgp->power;
5070 power_orig = group->sgp->power_orig;
5071 cpus = group->group_weight;
5073 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5074 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5075 capacity = cpus / smt; /* cores */
5077 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5079 capacity = fix_small_capacity(env->sd, group);
5085 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5086 * @env: The load balancing environment.
5087 * @group: sched_group whose statistics are to be updated.
5088 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5089 * @local_group: Does group contain this_cpu.
5090 * @sgs: variable to hold the statistics for this group.
5092 static inline void update_sg_lb_stats(struct lb_env *env,
5093 struct sched_group *group, int load_idx,
5094 int local_group, struct sg_lb_stats *sgs)
5096 unsigned long nr_running;
5100 memset(sgs, 0, sizeof(*sgs));
5102 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5103 struct rq *rq = cpu_rq(i);
5105 nr_running = rq->nr_running;
5107 /* Bias balancing toward cpus of our domain */
5109 load = target_load(i, load_idx);
5111 load = source_load(i, load_idx);
5113 sgs->group_load += load;
5114 sgs->sum_nr_running += nr_running;
5115 sgs->sum_weighted_load += weighted_cpuload(i);
5120 /* Adjust by relative CPU power of the group */
5121 sgs->group_power = group->sgp->power;
5122 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5124 if (sgs->sum_nr_running)
5125 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5127 sgs->group_weight = group->group_weight;
5129 sgs->group_imb = sg_imbalanced(group);
5130 sgs->group_capacity = sg_capacity(env, group);
5132 if (sgs->group_capacity > sgs->sum_nr_running)
5133 sgs->group_has_capacity = 1;
5137 * update_sd_pick_busiest - return 1 on busiest group
5138 * @env: The load balancing environment.
5139 * @sds: sched_domain statistics
5140 * @sg: sched_group candidate to be checked for being the busiest
5141 * @sgs: sched_group statistics
5143 * Determine if @sg is a busier group than the previously selected
5146 * Return: %true if @sg is a busier group than the previously selected
5147 * busiest group. %false otherwise.
5149 static bool update_sd_pick_busiest(struct lb_env *env,
5150 struct sd_lb_stats *sds,
5151 struct sched_group *sg,
5152 struct sg_lb_stats *sgs)
5154 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5157 if (sgs->sum_nr_running > sgs->group_capacity)
5164 * ASYM_PACKING needs to move all the work to the lowest
5165 * numbered CPUs in the group, therefore mark all groups
5166 * higher than ourself as busy.
5168 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5169 env->dst_cpu < group_first_cpu(sg)) {
5173 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5181 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5182 * @env: The load balancing environment.
5183 * @balance: Should we balance.
5184 * @sds: variable to hold the statistics for this sched_domain.
5186 static inline void update_sd_lb_stats(struct lb_env *env,
5187 struct sd_lb_stats *sds)
5189 struct sched_domain *child = env->sd->child;
5190 struct sched_group *sg = env->sd->groups;
5191 struct sg_lb_stats tmp_sgs;
5192 int load_idx, prefer_sibling = 0;
5194 if (child && child->flags & SD_PREFER_SIBLING)
5197 load_idx = get_sd_load_idx(env->sd, env->idle);
5200 struct sg_lb_stats *sgs = &tmp_sgs;
5203 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5206 sgs = &sds->local_stat;
5208 if (env->idle != CPU_NEWLY_IDLE ||
5209 time_after_eq(jiffies, sg->sgp->next_update))
5210 update_group_power(env->sd, env->dst_cpu);
5213 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5219 * In case the child domain prefers tasks go to siblings
5220 * first, lower the sg capacity to one so that we'll try
5221 * and move all the excess tasks away. We lower the capacity
5222 * of a group only if the local group has the capacity to fit
5223 * these excess tasks, i.e. nr_running < group_capacity. The
5224 * extra check prevents the case where you always pull from the
5225 * heaviest group when it is already under-utilized (possible
5226 * with a large weight task outweighs the tasks on the system).
5228 if (prefer_sibling && sds->local &&
5229 sds->local_stat.group_has_capacity)
5230 sgs->group_capacity = min(sgs->group_capacity, 1U);
5232 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5234 sds->busiest_stat = *sgs;
5238 /* Now, start updating sd_lb_stats */
5239 sds->total_load += sgs->group_load;
5240 sds->total_pwr += sgs->group_power;
5243 } while (sg != env->sd->groups);
5247 * check_asym_packing - Check to see if the group is packed into the
5250 * This is primarily intended to used at the sibling level. Some
5251 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5252 * case of POWER7, it can move to lower SMT modes only when higher
5253 * threads are idle. When in lower SMT modes, the threads will
5254 * perform better since they share less core resources. Hence when we
5255 * have idle threads, we want them to be the higher ones.
5257 * This packing function is run on idle threads. It checks to see if
5258 * the busiest CPU in this domain (core in the P7 case) has a higher
5259 * CPU number than the packing function is being run on. Here we are
5260 * assuming lower CPU number will be equivalent to lower a SMT thread
5263 * Return: 1 when packing is required and a task should be moved to
5264 * this CPU. The amount of the imbalance is returned in *imbalance.
5266 * @env: The load balancing environment.
5267 * @sds: Statistics of the sched_domain which is to be packed
5269 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5273 if (!(env->sd->flags & SD_ASYM_PACKING))
5279 busiest_cpu = group_first_cpu(sds->busiest);
5280 if (env->dst_cpu > busiest_cpu)
5283 env->imbalance = DIV_ROUND_CLOSEST(
5284 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5291 * fix_small_imbalance - Calculate the minor imbalance that exists
5292 * amongst the groups of a sched_domain, during
5294 * @env: The load balancing environment.
5295 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5298 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5300 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5301 unsigned int imbn = 2;
5302 unsigned long scaled_busy_load_per_task;
5303 struct sg_lb_stats *local, *busiest;
5305 local = &sds->local_stat;
5306 busiest = &sds->busiest_stat;
5308 if (!local->sum_nr_running)
5309 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5310 else if (busiest->load_per_task > local->load_per_task)
5313 scaled_busy_load_per_task =
5314 (busiest->load_per_task * SCHED_POWER_SCALE) /
5315 busiest->group_power;
5317 if (busiest->avg_load + scaled_busy_load_per_task >=
5318 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5319 env->imbalance = busiest->load_per_task;
5324 * OK, we don't have enough imbalance to justify moving tasks,
5325 * however we may be able to increase total CPU power used by
5329 pwr_now += busiest->group_power *
5330 min(busiest->load_per_task, busiest->avg_load);
5331 pwr_now += local->group_power *
5332 min(local->load_per_task, local->avg_load);
5333 pwr_now /= SCHED_POWER_SCALE;
5335 /* Amount of load we'd subtract */
5336 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5337 busiest->group_power;
5338 if (busiest->avg_load > tmp) {
5339 pwr_move += busiest->group_power *
5340 min(busiest->load_per_task,
5341 busiest->avg_load - tmp);
5344 /* Amount of load we'd add */
5345 if (busiest->avg_load * busiest->group_power <
5346 busiest->load_per_task * SCHED_POWER_SCALE) {
5347 tmp = (busiest->avg_load * busiest->group_power) /
5350 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5353 pwr_move += local->group_power *
5354 min(local->load_per_task, local->avg_load + tmp);
5355 pwr_move /= SCHED_POWER_SCALE;
5357 /* Move if we gain throughput */
5358 if (pwr_move > pwr_now)
5359 env->imbalance = busiest->load_per_task;
5363 * calculate_imbalance - Calculate the amount of imbalance present within the
5364 * groups of a given sched_domain during load balance.
5365 * @env: load balance environment
5366 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5368 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5370 unsigned long max_pull, load_above_capacity = ~0UL;
5371 struct sg_lb_stats *local, *busiest;
5373 local = &sds->local_stat;
5374 busiest = &sds->busiest_stat;
5376 if (busiest->group_imb) {
5378 * In the group_imb case we cannot rely on group-wide averages
5379 * to ensure cpu-load equilibrium, look at wider averages. XXX
5381 busiest->load_per_task =
5382 min(busiest->load_per_task, sds->avg_load);
5386 * In the presence of smp nice balancing, certain scenarios can have
5387 * max load less than avg load(as we skip the groups at or below
5388 * its cpu_power, while calculating max_load..)
5390 if (busiest->avg_load <= sds->avg_load ||
5391 local->avg_load >= sds->avg_load) {
5393 return fix_small_imbalance(env, sds);
5396 if (!busiest->group_imb) {
5398 * Don't want to pull so many tasks that a group would go idle.
5399 * Except of course for the group_imb case, since then we might
5400 * have to drop below capacity to reach cpu-load equilibrium.
5402 load_above_capacity =
5403 (busiest->sum_nr_running - busiest->group_capacity);
5405 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5406 load_above_capacity /= busiest->group_power;
5410 * We're trying to get all the cpus to the average_load, so we don't
5411 * want to push ourselves above the average load, nor do we wish to
5412 * reduce the max loaded cpu below the average load. At the same time,
5413 * we also don't want to reduce the group load below the group capacity
5414 * (so that we can implement power-savings policies etc). Thus we look
5415 * for the minimum possible imbalance.
5417 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5419 /* How much load to actually move to equalise the imbalance */
5420 env->imbalance = min(
5421 max_pull * busiest->group_power,
5422 (sds->avg_load - local->avg_load) * local->group_power
5423 ) / SCHED_POWER_SCALE;
5426 * if *imbalance is less than the average load per runnable task
5427 * there is no guarantee that any tasks will be moved so we'll have
5428 * a think about bumping its value to force at least one task to be
5431 if (env->imbalance < busiest->load_per_task)
5432 return fix_small_imbalance(env, sds);
5435 /******* find_busiest_group() helpers end here *********************/
5438 * find_busiest_group - Returns the busiest group within the sched_domain
5439 * if there is an imbalance. If there isn't an imbalance, and
5440 * the user has opted for power-savings, it returns a group whose
5441 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5442 * such a group exists.
5444 * Also calculates the amount of weighted load which should be moved
5445 * to restore balance.
5447 * @env: The load balancing environment.
5449 * Return: - The busiest group if imbalance exists.
5450 * - If no imbalance and user has opted for power-savings balance,
5451 * return the least loaded group whose CPUs can be
5452 * put to idle by rebalancing its tasks onto our group.
5454 static struct sched_group *find_busiest_group(struct lb_env *env)
5456 struct sg_lb_stats *local, *busiest;
5457 struct sd_lb_stats sds;
5459 init_sd_lb_stats(&sds);
5462 * Compute the various statistics relavent for load balancing at
5465 update_sd_lb_stats(env, &sds);
5466 local = &sds.local_stat;
5467 busiest = &sds.busiest_stat;
5469 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5470 check_asym_packing(env, &sds))
5473 /* There is no busy sibling group to pull tasks from */
5474 if (!sds.busiest || busiest->sum_nr_running == 0)
5477 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5480 * If the busiest group is imbalanced the below checks don't
5481 * work because they assume all things are equal, which typically
5482 * isn't true due to cpus_allowed constraints and the like.
5484 if (busiest->group_imb)
5487 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5488 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5489 !busiest->group_has_capacity)
5493 * If the local group is more busy than the selected busiest group
5494 * don't try and pull any tasks.
5496 if (local->avg_load >= busiest->avg_load)
5500 * Don't pull any tasks if this group is already above the domain
5503 if (local->avg_load >= sds.avg_load)
5506 if (env->idle == CPU_IDLE) {
5508 * This cpu is idle. If the busiest group load doesn't
5509 * have more tasks than the number of available cpu's and
5510 * there is no imbalance between this and busiest group
5511 * wrt to idle cpu's, it is balanced.
5513 if ((local->idle_cpus < busiest->idle_cpus) &&
5514 busiest->sum_nr_running <= busiest->group_weight)
5518 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5519 * imbalance_pct to be conservative.
5521 if (100 * busiest->avg_load <=
5522 env->sd->imbalance_pct * local->avg_load)
5527 /* Looks like there is an imbalance. Compute it */
5528 calculate_imbalance(env, &sds);
5537 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5539 static struct rq *find_busiest_queue(struct lb_env *env,
5540 struct sched_group *group)
5542 struct rq *busiest = NULL, *rq;
5543 unsigned long busiest_load = 0, busiest_power = 1;
5546 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5547 unsigned long power = power_of(i);
5548 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5553 capacity = fix_small_capacity(env->sd, group);
5556 wl = weighted_cpuload(i);
5559 * When comparing with imbalance, use weighted_cpuload()
5560 * which is not scaled with the cpu power.
5562 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5566 * For the load comparisons with the other cpu's, consider
5567 * the weighted_cpuload() scaled with the cpu power, so that
5568 * the load can be moved away from the cpu that is potentially
5569 * running at a lower capacity.
5571 * Thus we're looking for max(wl_i / power_i), crosswise
5572 * multiplication to rid ourselves of the division works out
5573 * to: wl_i * power_j > wl_j * power_i; where j is our
5576 if (wl * busiest_power > busiest_load * power) {
5578 busiest_power = power;
5587 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5588 * so long as it is large enough.
5590 #define MAX_PINNED_INTERVAL 512
5592 /* Working cpumask for load_balance and load_balance_newidle. */
5593 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5595 static int need_active_balance(struct lb_env *env)
5597 struct sched_domain *sd = env->sd;
5599 if (env->idle == CPU_NEWLY_IDLE) {
5602 * ASYM_PACKING needs to force migrate tasks from busy but
5603 * higher numbered CPUs in order to pack all tasks in the
5604 * lowest numbered CPUs.
5606 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5610 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5613 static int active_load_balance_cpu_stop(void *data);
5615 static int should_we_balance(struct lb_env *env)
5617 struct sched_group *sg = env->sd->groups;
5618 struct cpumask *sg_cpus, *sg_mask;
5619 int cpu, balance_cpu = -1;
5622 * In the newly idle case, we will allow all the cpu's
5623 * to do the newly idle load balance.
5625 if (env->idle == CPU_NEWLY_IDLE)
5628 sg_cpus = sched_group_cpus(sg);
5629 sg_mask = sched_group_mask(sg);
5630 /* Try to find first idle cpu */
5631 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5632 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5639 if (balance_cpu == -1)
5640 balance_cpu = group_balance_cpu(sg);
5643 * First idle cpu or the first cpu(busiest) in this sched group
5644 * is eligible for doing load balancing at this and above domains.
5646 return balance_cpu == env->dst_cpu;
5650 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5651 * tasks if there is an imbalance.
5653 static int load_balance(int this_cpu, struct rq *this_rq,
5654 struct sched_domain *sd, enum cpu_idle_type idle,
5655 int *continue_balancing)
5657 int ld_moved, cur_ld_moved, active_balance = 0;
5658 struct sched_domain *sd_parent = sd->parent;
5659 struct sched_group *group;
5661 unsigned long flags;
5662 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5664 struct lb_env env = {
5666 .dst_cpu = this_cpu,
5668 .dst_grpmask = sched_group_cpus(sd->groups),
5670 .loop_break = sched_nr_migrate_break,
5675 * For NEWLY_IDLE load_balancing, we don't need to consider
5676 * other cpus in our group
5678 if (idle == CPU_NEWLY_IDLE)
5679 env.dst_grpmask = NULL;
5681 cpumask_copy(cpus, cpu_active_mask);
5683 schedstat_inc(sd, lb_count[idle]);
5686 if (!should_we_balance(&env)) {
5687 *continue_balancing = 0;
5691 group = find_busiest_group(&env);
5693 schedstat_inc(sd, lb_nobusyg[idle]);
5697 busiest = find_busiest_queue(&env, group);
5699 schedstat_inc(sd, lb_nobusyq[idle]);
5703 BUG_ON(busiest == env.dst_rq);
5705 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5708 if (busiest->nr_running > 1) {
5710 * Attempt to move tasks. If find_busiest_group has found
5711 * an imbalance but busiest->nr_running <= 1, the group is
5712 * still unbalanced. ld_moved simply stays zero, so it is
5713 * correctly treated as an imbalance.
5715 env.flags |= LBF_ALL_PINNED;
5716 env.src_cpu = busiest->cpu;
5717 env.src_rq = busiest;
5718 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5721 local_irq_save(flags);
5722 double_rq_lock(env.dst_rq, busiest);
5725 * cur_ld_moved - load moved in current iteration
5726 * ld_moved - cumulative load moved across iterations
5728 cur_ld_moved = move_tasks(&env);
5729 ld_moved += cur_ld_moved;
5730 double_rq_unlock(env.dst_rq, busiest);
5731 local_irq_restore(flags);
5734 * some other cpu did the load balance for us.
5736 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5737 resched_cpu(env.dst_cpu);
5739 if (env.flags & LBF_NEED_BREAK) {
5740 env.flags &= ~LBF_NEED_BREAK;
5745 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5746 * us and move them to an alternate dst_cpu in our sched_group
5747 * where they can run. The upper limit on how many times we
5748 * iterate on same src_cpu is dependent on number of cpus in our
5751 * This changes load balance semantics a bit on who can move
5752 * load to a given_cpu. In addition to the given_cpu itself
5753 * (or a ilb_cpu acting on its behalf where given_cpu is
5754 * nohz-idle), we now have balance_cpu in a position to move
5755 * load to given_cpu. In rare situations, this may cause
5756 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5757 * _independently_ and at _same_ time to move some load to
5758 * given_cpu) causing exceess load to be moved to given_cpu.
5759 * This however should not happen so much in practice and
5760 * moreover subsequent load balance cycles should correct the
5761 * excess load moved.
5763 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5765 /* Prevent to re-select dst_cpu via env's cpus */
5766 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5768 env.dst_rq = cpu_rq(env.new_dst_cpu);
5769 env.dst_cpu = env.new_dst_cpu;
5770 env.flags &= ~LBF_DST_PINNED;
5772 env.loop_break = sched_nr_migrate_break;
5775 * Go back to "more_balance" rather than "redo" since we
5776 * need to continue with same src_cpu.
5782 * We failed to reach balance because of affinity.
5785 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5787 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5788 *group_imbalance = 1;
5789 } else if (*group_imbalance)
5790 *group_imbalance = 0;
5793 /* All tasks on this runqueue were pinned by CPU affinity */
5794 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5795 cpumask_clear_cpu(cpu_of(busiest), cpus);
5796 if (!cpumask_empty(cpus)) {
5798 env.loop_break = sched_nr_migrate_break;
5806 schedstat_inc(sd, lb_failed[idle]);
5808 * Increment the failure counter only on periodic balance.
5809 * We do not want newidle balance, which can be very
5810 * frequent, pollute the failure counter causing
5811 * excessive cache_hot migrations and active balances.
5813 if (idle != CPU_NEWLY_IDLE)
5814 sd->nr_balance_failed++;
5816 if (need_active_balance(&env)) {
5817 raw_spin_lock_irqsave(&busiest->lock, flags);
5819 /* don't kick the active_load_balance_cpu_stop,
5820 * if the curr task on busiest cpu can't be
5823 if (!cpumask_test_cpu(this_cpu,
5824 tsk_cpus_allowed(busiest->curr))) {
5825 raw_spin_unlock_irqrestore(&busiest->lock,
5827 env.flags |= LBF_ALL_PINNED;
5828 goto out_one_pinned;
5832 * ->active_balance synchronizes accesses to
5833 * ->active_balance_work. Once set, it's cleared
5834 * only after active load balance is finished.
5836 if (!busiest->active_balance) {
5837 busiest->active_balance = 1;
5838 busiest->push_cpu = this_cpu;
5841 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5843 if (active_balance) {
5844 stop_one_cpu_nowait(cpu_of(busiest),
5845 active_load_balance_cpu_stop, busiest,
5846 &busiest->active_balance_work);
5850 * We've kicked active balancing, reset the failure
5853 sd->nr_balance_failed = sd->cache_nice_tries+1;
5856 sd->nr_balance_failed = 0;
5858 if (likely(!active_balance)) {
5859 /* We were unbalanced, so reset the balancing interval */
5860 sd->balance_interval = sd->min_interval;
5863 * If we've begun active balancing, start to back off. This
5864 * case may not be covered by the all_pinned logic if there
5865 * is only 1 task on the busy runqueue (because we don't call
5868 if (sd->balance_interval < sd->max_interval)
5869 sd->balance_interval *= 2;
5875 schedstat_inc(sd, lb_balanced[idle]);
5877 sd->nr_balance_failed = 0;
5880 /* tune up the balancing interval */
5881 if (((env.flags & LBF_ALL_PINNED) &&
5882 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5883 (sd->balance_interval < sd->max_interval))
5884 sd->balance_interval *= 2;
5892 * idle_balance is called by schedule() if this_cpu is about to become
5893 * idle. Attempts to pull tasks from other CPUs.
5895 void idle_balance(int this_cpu, struct rq *this_rq)
5897 struct sched_domain *sd;
5898 int pulled_task = 0;
5899 unsigned long next_balance = jiffies + HZ;
5902 this_rq->idle_stamp = rq_clock(this_rq);
5904 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5908 * Drop the rq->lock, but keep IRQ/preempt disabled.
5910 raw_spin_unlock(&this_rq->lock);
5912 update_blocked_averages(this_cpu);
5914 for_each_domain(this_cpu, sd) {
5915 unsigned long interval;
5916 int continue_balancing = 1;
5917 u64 t0, domain_cost;
5919 if (!(sd->flags & SD_LOAD_BALANCE))
5922 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5925 if (sd->flags & SD_BALANCE_NEWIDLE) {
5926 t0 = sched_clock_cpu(this_cpu);
5928 /* If we've pulled tasks over stop searching: */
5929 pulled_task = load_balance(this_cpu, this_rq,
5931 &continue_balancing);
5933 domain_cost = sched_clock_cpu(this_cpu) - t0;
5934 if (domain_cost > sd->max_newidle_lb_cost)
5935 sd->max_newidle_lb_cost = domain_cost;
5937 curr_cost += domain_cost;
5940 interval = msecs_to_jiffies(sd->balance_interval);
5941 if (time_after(next_balance, sd->last_balance + interval))
5942 next_balance = sd->last_balance + interval;
5944 this_rq->idle_stamp = 0;
5950 raw_spin_lock(&this_rq->lock);
5952 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5954 * We are going idle. next_balance may be set based on
5955 * a busy processor. So reset next_balance.
5957 this_rq->next_balance = next_balance;
5960 if (curr_cost > this_rq->max_idle_balance_cost)
5961 this_rq->max_idle_balance_cost = curr_cost;
5965 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5966 * running tasks off the busiest CPU onto idle CPUs. It requires at
5967 * least 1 task to be running on each physical CPU where possible, and
5968 * avoids physical / logical imbalances.
5970 static int active_load_balance_cpu_stop(void *data)
5972 struct rq *busiest_rq = data;
5973 int busiest_cpu = cpu_of(busiest_rq);
5974 int target_cpu = busiest_rq->push_cpu;
5975 struct rq *target_rq = cpu_rq(target_cpu);
5976 struct sched_domain *sd;
5978 raw_spin_lock_irq(&busiest_rq->lock);
5980 /* make sure the requested cpu hasn't gone down in the meantime */
5981 if (unlikely(busiest_cpu != smp_processor_id() ||
5982 !busiest_rq->active_balance))
5985 /* Is there any task to move? */
5986 if (busiest_rq->nr_running <= 1)
5990 * This condition is "impossible", if it occurs
5991 * we need to fix it. Originally reported by
5992 * Bjorn Helgaas on a 128-cpu setup.
5994 BUG_ON(busiest_rq == target_rq);
5996 /* move a task from busiest_rq to target_rq */
5997 double_lock_balance(busiest_rq, target_rq);
5999 /* Search for an sd spanning us and the target CPU. */
6001 for_each_domain(target_cpu, sd) {
6002 if ((sd->flags & SD_LOAD_BALANCE) &&
6003 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6008 struct lb_env env = {
6010 .dst_cpu = target_cpu,
6011 .dst_rq = target_rq,
6012 .src_cpu = busiest_rq->cpu,
6013 .src_rq = busiest_rq,
6017 schedstat_inc(sd, alb_count);
6019 if (move_one_task(&env))
6020 schedstat_inc(sd, alb_pushed);
6022 schedstat_inc(sd, alb_failed);
6025 double_unlock_balance(busiest_rq, target_rq);
6027 busiest_rq->active_balance = 0;
6028 raw_spin_unlock_irq(&busiest_rq->lock);
6032 #ifdef CONFIG_NO_HZ_COMMON
6034 * idle load balancing details
6035 * - When one of the busy CPUs notice that there may be an idle rebalancing
6036 * needed, they will kick the idle load balancer, which then does idle
6037 * load balancing for all the idle CPUs.
6040 cpumask_var_t idle_cpus_mask;
6042 unsigned long next_balance; /* in jiffy units */
6043 } nohz ____cacheline_aligned;
6045 static inline int find_new_ilb(int call_cpu)
6047 int ilb = cpumask_first(nohz.idle_cpus_mask);
6049 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6056 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6057 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6058 * CPU (if there is one).
6060 static void nohz_balancer_kick(int cpu)
6064 nohz.next_balance++;
6066 ilb_cpu = find_new_ilb(cpu);
6068 if (ilb_cpu >= nr_cpu_ids)
6071 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6074 * Use smp_send_reschedule() instead of resched_cpu().
6075 * This way we generate a sched IPI on the target cpu which
6076 * is idle. And the softirq performing nohz idle load balance
6077 * will be run before returning from the IPI.
6079 smp_send_reschedule(ilb_cpu);
6083 static inline void nohz_balance_exit_idle(int cpu)
6085 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6086 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6087 atomic_dec(&nohz.nr_cpus);
6088 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6092 static inline void set_cpu_sd_state_busy(void)
6094 struct sched_domain *sd;
6097 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6099 if (!sd || !sd->nohz_idle)
6103 for (; sd; sd = sd->parent)
6104 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6109 void set_cpu_sd_state_idle(void)
6111 struct sched_domain *sd;
6114 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6116 if (!sd || sd->nohz_idle)
6120 for (; sd; sd = sd->parent)
6121 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6127 * This routine will record that the cpu is going idle with tick stopped.
6128 * This info will be used in performing idle load balancing in the future.
6130 void nohz_balance_enter_idle(int cpu)
6133 * If this cpu is going down, then nothing needs to be done.
6135 if (!cpu_active(cpu))
6138 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6141 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6142 atomic_inc(&nohz.nr_cpus);
6143 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6146 static int sched_ilb_notifier(struct notifier_block *nfb,
6147 unsigned long action, void *hcpu)
6149 switch (action & ~CPU_TASKS_FROZEN) {
6151 nohz_balance_exit_idle(smp_processor_id());
6159 static DEFINE_SPINLOCK(balancing);
6162 * Scale the max load_balance interval with the number of CPUs in the system.
6163 * This trades load-balance latency on larger machines for less cross talk.
6165 void update_max_interval(void)
6167 max_load_balance_interval = HZ*num_online_cpus()/10;
6171 * It checks each scheduling domain to see if it is due to be balanced,
6172 * and initiates a balancing operation if so.
6174 * Balancing parameters are set up in init_sched_domains.
6176 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6178 int continue_balancing = 1;
6179 struct rq *rq = cpu_rq(cpu);
6180 unsigned long interval;
6181 struct sched_domain *sd;
6182 /* Earliest time when we have to do rebalance again */
6183 unsigned long next_balance = jiffies + 60*HZ;
6184 int update_next_balance = 0;
6185 int need_serialize, need_decay = 0;
6188 update_blocked_averages(cpu);
6191 for_each_domain(cpu, sd) {
6193 * Decay the newidle max times here because this is a regular
6194 * visit to all the domains. Decay ~1% per second.
6196 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6197 sd->max_newidle_lb_cost =
6198 (sd->max_newidle_lb_cost * 253) / 256;
6199 sd->next_decay_max_lb_cost = jiffies + HZ;
6202 max_cost += sd->max_newidle_lb_cost;
6204 if (!(sd->flags & SD_LOAD_BALANCE))
6208 * Stop the load balance at this level. There is another
6209 * CPU in our sched group which is doing load balancing more
6212 if (!continue_balancing) {
6218 interval = sd->balance_interval;
6219 if (idle != CPU_IDLE)
6220 interval *= sd->busy_factor;
6222 /* scale ms to jiffies */
6223 interval = msecs_to_jiffies(interval);
6224 interval = clamp(interval, 1UL, max_load_balance_interval);
6226 need_serialize = sd->flags & SD_SERIALIZE;
6228 if (need_serialize) {
6229 if (!spin_trylock(&balancing))
6233 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6234 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6236 * The LBF_DST_PINNED logic could have changed
6237 * env->dst_cpu, so we can't know our idle
6238 * state even if we migrated tasks. Update it.
6240 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6242 sd->last_balance = jiffies;
6245 spin_unlock(&balancing);
6247 if (time_after(next_balance, sd->last_balance + interval)) {
6248 next_balance = sd->last_balance + interval;
6249 update_next_balance = 1;
6254 * Ensure the rq-wide value also decays but keep it at a
6255 * reasonable floor to avoid funnies with rq->avg_idle.
6257 rq->max_idle_balance_cost =
6258 max((u64)sysctl_sched_migration_cost, max_cost);
6263 * next_balance will be updated only when there is a need.
6264 * When the cpu is attached to null domain for ex, it will not be
6267 if (likely(update_next_balance))
6268 rq->next_balance = next_balance;
6271 #ifdef CONFIG_NO_HZ_COMMON
6273 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6274 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6276 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6278 struct rq *this_rq = cpu_rq(this_cpu);
6282 if (idle != CPU_IDLE ||
6283 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6286 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6287 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6291 * If this cpu gets work to do, stop the load balancing
6292 * work being done for other cpus. Next load
6293 * balancing owner will pick it up.
6298 rq = cpu_rq(balance_cpu);
6300 raw_spin_lock_irq(&rq->lock);
6301 update_rq_clock(rq);
6302 update_idle_cpu_load(rq);
6303 raw_spin_unlock_irq(&rq->lock);
6305 rebalance_domains(balance_cpu, CPU_IDLE);
6307 if (time_after(this_rq->next_balance, rq->next_balance))
6308 this_rq->next_balance = rq->next_balance;
6310 nohz.next_balance = this_rq->next_balance;
6312 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6316 * Current heuristic for kicking the idle load balancer in the presence
6317 * of an idle cpu is the system.
6318 * - This rq has more than one task.
6319 * - At any scheduler domain level, this cpu's scheduler group has multiple
6320 * busy cpu's exceeding the group's power.
6321 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6322 * domain span are idle.
6324 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6326 unsigned long now = jiffies;
6327 struct sched_domain *sd;
6329 if (unlikely(idle_cpu(cpu)))
6333 * We may be recently in ticked or tickless idle mode. At the first
6334 * busy tick after returning from idle, we will update the busy stats.
6336 set_cpu_sd_state_busy();
6337 nohz_balance_exit_idle(cpu);
6340 * None are in tickless mode and hence no need for NOHZ idle load
6343 if (likely(!atomic_read(&nohz.nr_cpus)))
6346 if (time_before(now, nohz.next_balance))
6349 if (rq->nr_running >= 2)
6353 for_each_domain(cpu, sd) {
6354 struct sched_group *sg = sd->groups;
6355 struct sched_group_power *sgp = sg->sgp;
6356 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6358 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6359 goto need_kick_unlock;
6361 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6362 && (cpumask_first_and(nohz.idle_cpus_mask,
6363 sched_domain_span(sd)) < cpu))
6364 goto need_kick_unlock;
6366 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6378 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6382 * run_rebalance_domains is triggered when needed from the scheduler tick.
6383 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6385 static void run_rebalance_domains(struct softirq_action *h)
6387 int this_cpu = smp_processor_id();
6388 struct rq *this_rq = cpu_rq(this_cpu);
6389 enum cpu_idle_type idle = this_rq->idle_balance ?
6390 CPU_IDLE : CPU_NOT_IDLE;
6392 rebalance_domains(this_cpu, idle);
6395 * If this cpu has a pending nohz_balance_kick, then do the
6396 * balancing on behalf of the other idle cpus whose ticks are
6399 nohz_idle_balance(this_cpu, idle);
6402 static inline int on_null_domain(int cpu)
6404 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6408 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6410 void trigger_load_balance(struct rq *rq, int cpu)
6412 /* Don't need to rebalance while attached to NULL domain */
6413 if (time_after_eq(jiffies, rq->next_balance) &&
6414 likely(!on_null_domain(cpu)))
6415 raise_softirq(SCHED_SOFTIRQ);
6416 #ifdef CONFIG_NO_HZ_COMMON
6417 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6418 nohz_balancer_kick(cpu);
6422 static void rq_online_fair(struct rq *rq)
6427 static void rq_offline_fair(struct rq *rq)
6431 /* Ensure any throttled groups are reachable by pick_next_task */
6432 unthrottle_offline_cfs_rqs(rq);
6435 #endif /* CONFIG_SMP */
6438 * scheduler tick hitting a task of our scheduling class:
6440 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6442 struct cfs_rq *cfs_rq;
6443 struct sched_entity *se = &curr->se;
6445 for_each_sched_entity(se) {
6446 cfs_rq = cfs_rq_of(se);
6447 entity_tick(cfs_rq, se, queued);
6450 if (numabalancing_enabled)
6451 task_tick_numa(rq, curr);
6453 update_rq_runnable_avg(rq, 1);
6457 * called on fork with the child task as argument from the parent's context
6458 * - child not yet on the tasklist
6459 * - preemption disabled
6461 static void task_fork_fair(struct task_struct *p)
6463 struct cfs_rq *cfs_rq;
6464 struct sched_entity *se = &p->se, *curr;
6465 int this_cpu = smp_processor_id();
6466 struct rq *rq = this_rq();
6467 unsigned long flags;
6469 raw_spin_lock_irqsave(&rq->lock, flags);
6471 update_rq_clock(rq);
6473 cfs_rq = task_cfs_rq(current);
6474 curr = cfs_rq->curr;
6477 * Not only the cpu but also the task_group of the parent might have
6478 * been changed after parent->se.parent,cfs_rq were copied to
6479 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6480 * of child point to valid ones.
6483 __set_task_cpu(p, this_cpu);
6486 update_curr(cfs_rq);
6489 se->vruntime = curr->vruntime;
6490 place_entity(cfs_rq, se, 1);
6492 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6494 * Upon rescheduling, sched_class::put_prev_task() will place
6495 * 'current' within the tree based on its new key value.
6497 swap(curr->vruntime, se->vruntime);
6498 resched_task(rq->curr);
6501 se->vruntime -= cfs_rq->min_vruntime;
6503 raw_spin_unlock_irqrestore(&rq->lock, flags);
6507 * Priority of the task has changed. Check to see if we preempt
6511 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6517 * Reschedule if we are currently running on this runqueue and
6518 * our priority decreased, or if we are not currently running on
6519 * this runqueue and our priority is higher than the current's
6521 if (rq->curr == p) {
6522 if (p->prio > oldprio)
6523 resched_task(rq->curr);
6525 check_preempt_curr(rq, p, 0);
6528 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6530 struct sched_entity *se = &p->se;
6531 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6534 * Ensure the task's vruntime is normalized, so that when its
6535 * switched back to the fair class the enqueue_entity(.flags=0) will
6536 * do the right thing.
6538 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6539 * have normalized the vruntime, if it was !on_rq, then only when
6540 * the task is sleeping will it still have non-normalized vruntime.
6542 if (!se->on_rq && p->state != TASK_RUNNING) {
6544 * Fix up our vruntime so that the current sleep doesn't
6545 * cause 'unlimited' sleep bonus.
6547 place_entity(cfs_rq, se, 0);
6548 se->vruntime -= cfs_rq->min_vruntime;
6553 * Remove our load from contribution when we leave sched_fair
6554 * and ensure we don't carry in an old decay_count if we
6557 if (se->avg.decay_count) {
6558 __synchronize_entity_decay(se);
6559 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6565 * We switched to the sched_fair class.
6567 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6573 * We were most likely switched from sched_rt, so
6574 * kick off the schedule if running, otherwise just see
6575 * if we can still preempt the current task.
6578 resched_task(rq->curr);
6580 check_preempt_curr(rq, p, 0);
6583 /* Account for a task changing its policy or group.
6585 * This routine is mostly called to set cfs_rq->curr field when a task
6586 * migrates between groups/classes.
6588 static void set_curr_task_fair(struct rq *rq)
6590 struct sched_entity *se = &rq->curr->se;
6592 for_each_sched_entity(se) {
6593 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6595 set_next_entity(cfs_rq, se);
6596 /* ensure bandwidth has been allocated on our new cfs_rq */
6597 account_cfs_rq_runtime(cfs_rq, 0);
6601 void init_cfs_rq(struct cfs_rq *cfs_rq)
6603 cfs_rq->tasks_timeline = RB_ROOT;
6604 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6605 #ifndef CONFIG_64BIT
6606 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6609 atomic64_set(&cfs_rq->decay_counter, 1);
6610 atomic_long_set(&cfs_rq->removed_load, 0);
6614 #ifdef CONFIG_FAIR_GROUP_SCHED
6615 static void task_move_group_fair(struct task_struct *p, int on_rq)
6617 struct cfs_rq *cfs_rq;
6619 * If the task was not on the rq at the time of this cgroup movement
6620 * it must have been asleep, sleeping tasks keep their ->vruntime
6621 * absolute on their old rq until wakeup (needed for the fair sleeper
6622 * bonus in place_entity()).
6624 * If it was on the rq, we've just 'preempted' it, which does convert
6625 * ->vruntime to a relative base.
6627 * Make sure both cases convert their relative position when migrating
6628 * to another cgroup's rq. This does somewhat interfere with the
6629 * fair sleeper stuff for the first placement, but who cares.
6632 * When !on_rq, vruntime of the task has usually NOT been normalized.
6633 * But there are some cases where it has already been normalized:
6635 * - Moving a forked child which is waiting for being woken up by
6636 * wake_up_new_task().
6637 * - Moving a task which has been woken up by try_to_wake_up() and
6638 * waiting for actually being woken up by sched_ttwu_pending().
6640 * To prevent boost or penalty in the new cfs_rq caused by delta
6641 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6643 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6647 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6648 set_task_rq(p, task_cpu(p));
6650 cfs_rq = cfs_rq_of(&p->se);
6651 p->se.vruntime += cfs_rq->min_vruntime;
6654 * migrate_task_rq_fair() will have removed our previous
6655 * contribution, but we must synchronize for ongoing future
6658 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6659 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6664 void free_fair_sched_group(struct task_group *tg)
6668 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6670 for_each_possible_cpu(i) {
6672 kfree(tg->cfs_rq[i]);
6681 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6683 struct cfs_rq *cfs_rq;
6684 struct sched_entity *se;
6687 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6690 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6694 tg->shares = NICE_0_LOAD;
6696 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6698 for_each_possible_cpu(i) {
6699 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6700 GFP_KERNEL, cpu_to_node(i));
6704 se = kzalloc_node(sizeof(struct sched_entity),
6705 GFP_KERNEL, cpu_to_node(i));
6709 init_cfs_rq(cfs_rq);
6710 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6721 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6723 struct rq *rq = cpu_rq(cpu);
6724 unsigned long flags;
6727 * Only empty task groups can be destroyed; so we can speculatively
6728 * check on_list without danger of it being re-added.
6730 if (!tg->cfs_rq[cpu]->on_list)
6733 raw_spin_lock_irqsave(&rq->lock, flags);
6734 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6735 raw_spin_unlock_irqrestore(&rq->lock, flags);
6738 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6739 struct sched_entity *se, int cpu,
6740 struct sched_entity *parent)
6742 struct rq *rq = cpu_rq(cpu);
6746 init_cfs_rq_runtime(cfs_rq);
6748 tg->cfs_rq[cpu] = cfs_rq;
6751 /* se could be NULL for root_task_group */
6756 se->cfs_rq = &rq->cfs;
6758 se->cfs_rq = parent->my_q;
6761 update_load_set(&se->load, 0);
6762 se->parent = parent;
6765 static DEFINE_MUTEX(shares_mutex);
6767 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6770 unsigned long flags;
6773 * We can't change the weight of the root cgroup.
6778 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6780 mutex_lock(&shares_mutex);
6781 if (tg->shares == shares)
6784 tg->shares = shares;
6785 for_each_possible_cpu(i) {
6786 struct rq *rq = cpu_rq(i);
6787 struct sched_entity *se;
6790 /* Propagate contribution to hierarchy */
6791 raw_spin_lock_irqsave(&rq->lock, flags);
6793 /* Possible calls to update_curr() need rq clock */
6794 update_rq_clock(rq);
6795 for_each_sched_entity(se)
6796 update_cfs_shares(group_cfs_rq(se));
6797 raw_spin_unlock_irqrestore(&rq->lock, flags);
6801 mutex_unlock(&shares_mutex);
6804 #else /* CONFIG_FAIR_GROUP_SCHED */
6806 void free_fair_sched_group(struct task_group *tg) { }
6808 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6813 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6815 #endif /* CONFIG_FAIR_GROUP_SCHED */
6818 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6820 struct sched_entity *se = &task->se;
6821 unsigned int rr_interval = 0;
6824 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6827 if (rq->cfs.load.weight)
6828 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6834 * All the scheduling class methods:
6836 const struct sched_class fair_sched_class = {
6837 .next = &idle_sched_class,
6838 .enqueue_task = enqueue_task_fair,
6839 .dequeue_task = dequeue_task_fair,
6840 .yield_task = yield_task_fair,
6841 .yield_to_task = yield_to_task_fair,
6843 .check_preempt_curr = check_preempt_wakeup,
6845 .pick_next_task = pick_next_task_fair,
6846 .put_prev_task = put_prev_task_fair,
6849 .select_task_rq = select_task_rq_fair,
6850 .migrate_task_rq = migrate_task_rq_fair,
6852 .rq_online = rq_online_fair,
6853 .rq_offline = rq_offline_fair,
6855 .task_waking = task_waking_fair,
6858 .set_curr_task = set_curr_task_fair,
6859 .task_tick = task_tick_fair,
6860 .task_fork = task_fork_fair,
6862 .prio_changed = prio_changed_fair,
6863 .switched_from = switched_from_fair,
6864 .switched_to = switched_to_fair,
6866 .get_rr_interval = get_rr_interval_fair,
6868 #ifdef CONFIG_FAIR_GROUP_SCHED
6869 .task_move_group = task_move_group_fair,
6873 #ifdef CONFIG_SCHED_DEBUG
6874 void print_cfs_stats(struct seq_file *m, int cpu)
6876 struct cfs_rq *cfs_rq;
6879 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6880 print_cfs_rq(m, cpu, cfs_rq);
6885 __init void init_sched_fair_class(void)
6888 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6890 #ifdef CONFIG_NO_HZ_COMMON
6891 nohz.next_balance = jiffies;
6892 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6893 cpu_notifier(sched_ilb_notifier, 0);