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 #define WMULT_CONST (~0U)
182 #define WMULT_SHIFT 32
184 static void __update_inv_weight(struct load_weight *lw)
188 if (likely(lw->inv_weight))
191 w = scale_load_down(lw->weight);
193 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
195 else if (unlikely(!w))
196 lw->inv_weight = WMULT_CONST;
198 lw->inv_weight = WMULT_CONST / w;
202 * delta_exec * weight / lw.weight
204 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
207 * we're guaranteed shift stays positive because inv_weight is guaranteed to
208 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
211 * weight/lw.weight <= 1, and therefore our shift will also be positive.
213 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
215 u64 fact = scale_load_down(weight);
216 int shift = WMULT_SHIFT;
218 __update_inv_weight(lw);
220 if (unlikely(fact >> 32)) {
227 /* hint to use a 32x32->64 mul */
228 fact = (u64)(u32)fact * lw->inv_weight;
235 return mul_u64_u32_shr(delta_exec, fact, shift);
239 const struct sched_class fair_sched_class;
241 /**************************************************************
242 * CFS operations on generic schedulable entities:
245 #ifdef CONFIG_FAIR_GROUP_SCHED
247 /* cpu runqueue to which this cfs_rq is attached */
248 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
253 /* An entity is a task if it doesn't "own" a runqueue */
254 #define entity_is_task(se) (!se->my_q)
256 static inline struct task_struct *task_of(struct sched_entity *se)
258 #ifdef CONFIG_SCHED_DEBUG
259 WARN_ON_ONCE(!entity_is_task(se));
261 return container_of(se, struct task_struct, se);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
285 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 if (!cfs_rq->on_list) {
292 * Ensure we either appear before our parent (if already
293 * enqueued) or force our parent to appear after us when it is
294 * enqueued. The fact that we always enqueue bottom-up
295 * reduces this to two cases.
297 if (cfs_rq->tg->parent &&
298 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
299 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
300 &rq_of(cfs_rq)->leaf_cfs_rq_list);
302 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
303 &rq_of(cfs_rq)->leaf_cfs_rq_list);
307 /* We should have no load, but we need to update last_decay. */
308 update_cfs_rq_blocked_load(cfs_rq, 0);
312 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
314 if (cfs_rq->on_list) {
315 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
320 /* Iterate thr' all leaf cfs_rq's on a runqueue */
321 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
322 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
324 /* Do the two (enqueued) entities belong to the same group ? */
326 is_same_group(struct sched_entity *se, struct sched_entity *pse)
328 if (se->cfs_rq == pse->cfs_rq)
334 static inline struct sched_entity *parent_entity(struct sched_entity *se)
339 /* return depth at which a sched entity is present in the hierarchy */
340 static inline int depth_se(struct sched_entity *se)
344 for_each_sched_entity(se)
351 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
353 int se_depth, pse_depth;
356 * preemption test can be made between sibling entities who are in the
357 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
358 * both tasks until we find their ancestors who are siblings of common
362 /* First walk up until both entities are at same depth */
363 se_depth = depth_se(*se);
364 pse_depth = depth_se(*pse);
366 while (se_depth > pse_depth) {
368 *se = parent_entity(*se);
371 while (pse_depth > se_depth) {
373 *pse = parent_entity(*pse);
376 while (!is_same_group(*se, *pse)) {
377 *se = parent_entity(*se);
378 *pse = parent_entity(*pse);
382 #else /* !CONFIG_FAIR_GROUP_SCHED */
384 static inline struct task_struct *task_of(struct sched_entity *se)
386 return container_of(se, struct task_struct, se);
389 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
391 return container_of(cfs_rq, struct rq, cfs);
394 #define entity_is_task(se) 1
396 #define for_each_sched_entity(se) \
397 for (; se; se = NULL)
399 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
401 return &task_rq(p)->cfs;
404 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
406 struct task_struct *p = task_of(se);
407 struct rq *rq = task_rq(p);
412 /* runqueue "owned" by this group */
413 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
418 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
422 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
426 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
427 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
430 is_same_group(struct sched_entity *se, struct sched_entity *pse)
435 static inline struct sched_entity *parent_entity(struct sched_entity *se)
441 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
445 #endif /* CONFIG_FAIR_GROUP_SCHED */
447 static __always_inline
448 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
450 /**************************************************************
451 * Scheduling class tree data structure manipulation methods:
454 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
456 s64 delta = (s64)(vruntime - max_vruntime);
458 max_vruntime = vruntime;
463 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
465 s64 delta = (s64)(vruntime - min_vruntime);
467 min_vruntime = vruntime;
472 static inline int entity_before(struct sched_entity *a,
473 struct sched_entity *b)
475 return (s64)(a->vruntime - b->vruntime) < 0;
478 static void update_min_vruntime(struct cfs_rq *cfs_rq)
480 u64 vruntime = cfs_rq->min_vruntime;
483 vruntime = cfs_rq->curr->vruntime;
485 if (cfs_rq->rb_leftmost) {
486 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
491 vruntime = se->vruntime;
493 vruntime = min_vruntime(vruntime, se->vruntime);
496 /* ensure we never gain time by being placed backwards. */
497 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
500 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
505 * Enqueue an entity into the rb-tree:
507 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
509 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
510 struct rb_node *parent = NULL;
511 struct sched_entity *entry;
515 * Find the right place in the rbtree:
519 entry = rb_entry(parent, struct sched_entity, run_node);
521 * We dont care about collisions. Nodes with
522 * the same key stay together.
524 if (entity_before(se, entry)) {
525 link = &parent->rb_left;
527 link = &parent->rb_right;
533 * Maintain a cache of leftmost tree entries (it is frequently
537 cfs_rq->rb_leftmost = &se->run_node;
539 rb_link_node(&se->run_node, parent, link);
540 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
543 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
545 if (cfs_rq->rb_leftmost == &se->run_node) {
546 struct rb_node *next_node;
548 next_node = rb_next(&se->run_node);
549 cfs_rq->rb_leftmost = next_node;
552 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
555 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
557 struct rb_node *left = cfs_rq->rb_leftmost;
562 return rb_entry(left, struct sched_entity, run_node);
565 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
567 struct rb_node *next = rb_next(&se->run_node);
572 return rb_entry(next, struct sched_entity, run_node);
575 #ifdef CONFIG_SCHED_DEBUG
576 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
578 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
583 return rb_entry(last, struct sched_entity, run_node);
586 /**************************************************************
587 * Scheduling class statistics methods:
590 int sched_proc_update_handler(struct ctl_table *table, int write,
591 void __user *buffer, size_t *lenp,
594 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
595 int factor = get_update_sysctl_factor();
600 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
601 sysctl_sched_min_granularity);
603 #define WRT_SYSCTL(name) \
604 (normalized_sysctl_##name = sysctl_##name / (factor))
605 WRT_SYSCTL(sched_min_granularity);
606 WRT_SYSCTL(sched_latency);
607 WRT_SYSCTL(sched_wakeup_granularity);
617 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
619 if (unlikely(se->load.weight != NICE_0_LOAD))
620 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
626 * The idea is to set a period in which each task runs once.
628 * When there are too many tasks (sched_nr_latency) we have to stretch
629 * this period because otherwise the slices get too small.
631 * p = (nr <= nl) ? l : l*nr/nl
633 static u64 __sched_period(unsigned long nr_running)
635 u64 period = sysctl_sched_latency;
636 unsigned long nr_latency = sched_nr_latency;
638 if (unlikely(nr_running > nr_latency)) {
639 period = sysctl_sched_min_granularity;
640 period *= nr_running;
647 * We calculate the wall-time slice from the period by taking a part
648 * proportional to the weight.
652 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
654 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
656 for_each_sched_entity(se) {
657 struct load_weight *load;
658 struct load_weight lw;
660 cfs_rq = cfs_rq_of(se);
661 load = &cfs_rq->load;
663 if (unlikely(!se->on_rq)) {
666 update_load_add(&lw, se->load.weight);
669 slice = __calc_delta(slice, se->load.weight, load);
675 * We calculate the vruntime slice of a to-be-inserted task.
679 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
681 return calc_delta_fair(sched_slice(cfs_rq, se), se);
685 static unsigned long task_h_load(struct task_struct *p);
687 static inline void __update_task_entity_contrib(struct sched_entity *se);
689 /* Give new task start runnable values to heavy its load in infant time */
690 void init_task_runnable_average(struct task_struct *p)
694 p->se.avg.decay_count = 0;
695 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
696 p->se.avg.runnable_avg_sum = slice;
697 p->se.avg.runnable_avg_period = slice;
698 __update_task_entity_contrib(&p->se);
701 void init_task_runnable_average(struct task_struct *p)
707 * Update the current task's runtime statistics.
709 static void update_curr(struct cfs_rq *cfs_rq)
711 struct sched_entity *curr = cfs_rq->curr;
712 u64 now = rq_clock_task(rq_of(cfs_rq));
718 delta_exec = now - curr->exec_start;
719 if (unlikely((s64)delta_exec <= 0))
722 curr->exec_start = now;
724 schedstat_set(curr->statistics.exec_max,
725 max(delta_exec, curr->statistics.exec_max));
727 curr->sum_exec_runtime += delta_exec;
728 schedstat_add(cfs_rq, exec_clock, delta_exec);
730 curr->vruntime += calc_delta_fair(delta_exec, curr);
731 update_min_vruntime(cfs_rq);
733 if (entity_is_task(curr)) {
734 struct task_struct *curtask = task_of(curr);
736 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
737 cpuacct_charge(curtask, delta_exec);
738 account_group_exec_runtime(curtask, delta_exec);
741 account_cfs_rq_runtime(cfs_rq, delta_exec);
745 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
747 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
751 * Task is being enqueued - update stats:
753 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
756 * Are we enqueueing a waiting task? (for current tasks
757 * a dequeue/enqueue event is a NOP)
759 if (se != cfs_rq->curr)
760 update_stats_wait_start(cfs_rq, se);
764 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
766 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
767 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
768 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
769 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
770 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
771 #ifdef CONFIG_SCHEDSTATS
772 if (entity_is_task(se)) {
773 trace_sched_stat_wait(task_of(se),
774 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
777 schedstat_set(se->statistics.wait_start, 0);
781 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
784 * Mark the end of the wait period if dequeueing a
787 if (se != cfs_rq->curr)
788 update_stats_wait_end(cfs_rq, se);
792 * We are picking a new current task - update its stats:
795 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
798 * We are starting a new run period:
800 se->exec_start = rq_clock_task(rq_of(cfs_rq));
803 /**************************************************
804 * Scheduling class queueing methods:
807 #ifdef CONFIG_NUMA_BALANCING
809 * Approximate time to scan a full NUMA task in ms. The task scan period is
810 * calculated based on the tasks virtual memory size and
811 * numa_balancing_scan_size.
813 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
814 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
816 /* Portion of address space to scan in MB */
817 unsigned int sysctl_numa_balancing_scan_size = 256;
819 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
820 unsigned int sysctl_numa_balancing_scan_delay = 1000;
822 static unsigned int task_nr_scan_windows(struct task_struct *p)
824 unsigned long rss = 0;
825 unsigned long nr_scan_pages;
828 * Calculations based on RSS as non-present and empty pages are skipped
829 * by the PTE scanner and NUMA hinting faults should be trapped based
832 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
833 rss = get_mm_rss(p->mm);
837 rss = round_up(rss, nr_scan_pages);
838 return rss / nr_scan_pages;
841 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
842 #define MAX_SCAN_WINDOW 2560
844 static unsigned int task_scan_min(struct task_struct *p)
846 unsigned int scan, floor;
847 unsigned int windows = 1;
849 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
850 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
851 floor = 1000 / windows;
853 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
854 return max_t(unsigned int, floor, scan);
857 static unsigned int task_scan_max(struct task_struct *p)
859 unsigned int smin = task_scan_min(p);
862 /* Watch for min being lower than max due to floor calculations */
863 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
864 return max(smin, smax);
867 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
869 rq->nr_numa_running += (p->numa_preferred_nid != -1);
870 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
873 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
875 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
876 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
882 spinlock_t lock; /* nr_tasks, tasks */
885 struct list_head task_list;
888 unsigned long total_faults;
889 unsigned long faults[0];
892 pid_t task_numa_group_id(struct task_struct *p)
894 return p->numa_group ? p->numa_group->gid : 0;
897 static inline int task_faults_idx(int nid, int priv)
899 return 2 * nid + priv;
902 static inline unsigned long task_faults(struct task_struct *p, int nid)
907 return p->numa_faults[task_faults_idx(nid, 0)] +
908 p->numa_faults[task_faults_idx(nid, 1)];
911 static inline unsigned long group_faults(struct task_struct *p, int nid)
916 return p->numa_group->faults[task_faults_idx(nid, 0)] +
917 p->numa_group->faults[task_faults_idx(nid, 1)];
921 * These return the fraction of accesses done by a particular task, or
922 * task group, on a particular numa node. The group weight is given a
923 * larger multiplier, in order to group tasks together that are almost
924 * evenly spread out between numa nodes.
926 static inline unsigned long task_weight(struct task_struct *p, int nid)
928 unsigned long total_faults;
933 total_faults = p->total_numa_faults;
938 return 1000 * task_faults(p, nid) / total_faults;
941 static inline unsigned long group_weight(struct task_struct *p, int nid)
943 if (!p->numa_group || !p->numa_group->total_faults)
946 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
949 static unsigned long weighted_cpuload(const int cpu);
950 static unsigned long source_load(int cpu, int type);
951 static unsigned long target_load(int cpu, int type);
952 static unsigned long power_of(int cpu);
953 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
955 /* Cached statistics for all CPUs within a node */
957 unsigned long nr_running;
960 /* Total compute capacity of CPUs on a node */
963 /* Approximate capacity in terms of runnable tasks on a node */
964 unsigned long capacity;
969 * XXX borrowed from update_sg_lb_stats
971 static void update_numa_stats(struct numa_stats *ns, int nid)
975 memset(ns, 0, sizeof(*ns));
976 for_each_cpu(cpu, cpumask_of_node(nid)) {
977 struct rq *rq = cpu_rq(cpu);
979 ns->nr_running += rq->nr_running;
980 ns->load += weighted_cpuload(cpu);
981 ns->power += power_of(cpu);
987 * If we raced with hotplug and there are no CPUs left in our mask
988 * the @ns structure is NULL'ed and task_numa_compare() will
989 * not find this node attractive.
991 * We'll either bail at !has_capacity, or we'll detect a huge imbalance
997 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
998 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
999 ns->has_capacity = (ns->nr_running < ns->capacity);
1002 struct task_numa_env {
1003 struct task_struct *p;
1005 int src_cpu, src_nid;
1006 int dst_cpu, dst_nid;
1008 struct numa_stats src_stats, dst_stats;
1012 struct task_struct *best_task;
1017 static void task_numa_assign(struct task_numa_env *env,
1018 struct task_struct *p, long imp)
1021 put_task_struct(env->best_task);
1026 env->best_imp = imp;
1027 env->best_cpu = env->dst_cpu;
1031 * This checks if the overall compute and NUMA accesses of the system would
1032 * be improved if the source tasks was migrated to the target dst_cpu taking
1033 * into account that it might be best if task running on the dst_cpu should
1034 * be exchanged with the source task
1036 static void task_numa_compare(struct task_numa_env *env,
1037 long taskimp, long groupimp)
1039 struct rq *src_rq = cpu_rq(env->src_cpu);
1040 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1041 struct task_struct *cur;
1042 long dst_load, src_load;
1044 long imp = (groupimp > 0) ? groupimp : taskimp;
1047 cur = ACCESS_ONCE(dst_rq->curr);
1048 if (cur->pid == 0) /* idle */
1052 * "imp" is the fault differential for the source task between the
1053 * source and destination node. Calculate the total differential for
1054 * the source task and potential destination task. The more negative
1055 * the value is, the more rmeote accesses that would be expected to
1056 * be incurred if the tasks were swapped.
1059 /* Skip this swap candidate if cannot move to the source cpu */
1060 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1064 * If dst and source tasks are in the same NUMA group, or not
1065 * in any group then look only at task weights.
1067 if (cur->numa_group == env->p->numa_group) {
1068 imp = taskimp + task_weight(cur, env->src_nid) -
1069 task_weight(cur, env->dst_nid);
1071 * Add some hysteresis to prevent swapping the
1072 * tasks within a group over tiny differences.
1074 if (cur->numa_group)
1078 * Compare the group weights. If a task is all by
1079 * itself (not part of a group), use the task weight
1082 if (env->p->numa_group)
1087 if (cur->numa_group)
1088 imp += group_weight(cur, env->src_nid) -
1089 group_weight(cur, env->dst_nid);
1091 imp += task_weight(cur, env->src_nid) -
1092 task_weight(cur, env->dst_nid);
1096 if (imp < env->best_imp)
1100 /* Is there capacity at our destination? */
1101 if (env->src_stats.has_capacity &&
1102 !env->dst_stats.has_capacity)
1108 /* Balance doesn't matter much if we're running a task per cpu */
1109 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1113 * In the overloaded case, try and keep the load balanced.
1116 dst_load = env->dst_stats.load;
1117 src_load = env->src_stats.load;
1119 /* XXX missing power terms */
1120 load = task_h_load(env->p);
1125 load = task_h_load(cur);
1130 /* make src_load the smaller */
1131 if (dst_load < src_load)
1132 swap(dst_load, src_load);
1134 if (src_load * env->imbalance_pct < dst_load * 100)
1138 task_numa_assign(env, cur, imp);
1143 static void task_numa_find_cpu(struct task_numa_env *env,
1144 long taskimp, long groupimp)
1148 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1149 /* Skip this CPU if the source task cannot migrate */
1150 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1154 task_numa_compare(env, taskimp, groupimp);
1158 static int task_numa_migrate(struct task_struct *p)
1160 struct task_numa_env env = {
1163 .src_cpu = task_cpu(p),
1164 .src_nid = task_node(p),
1166 .imbalance_pct = 112,
1172 struct sched_domain *sd;
1173 unsigned long taskweight, groupweight;
1175 long taskimp, groupimp;
1178 * Pick the lowest SD_NUMA domain, as that would have the smallest
1179 * imbalance and would be the first to start moving tasks about.
1181 * And we want to avoid any moving of tasks about, as that would create
1182 * random movement of tasks -- counter the numa conditions we're trying
1186 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1188 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1192 * Cpusets can break the scheduler domain tree into smaller
1193 * balance domains, some of which do not cross NUMA boundaries.
1194 * Tasks that are "trapped" in such domains cannot be migrated
1195 * elsewhere, so there is no point in (re)trying.
1197 if (unlikely(!sd)) {
1198 p->numa_preferred_nid = task_node(p);
1202 taskweight = task_weight(p, env.src_nid);
1203 groupweight = group_weight(p, env.src_nid);
1204 update_numa_stats(&env.src_stats, env.src_nid);
1205 env.dst_nid = p->numa_preferred_nid;
1206 taskimp = task_weight(p, env.dst_nid) - taskweight;
1207 groupimp = group_weight(p, env.dst_nid) - groupweight;
1208 update_numa_stats(&env.dst_stats, env.dst_nid);
1210 /* If the preferred nid has capacity, try to use it. */
1211 if (env.dst_stats.has_capacity)
1212 task_numa_find_cpu(&env, taskimp, groupimp);
1214 /* No space available on the preferred nid. Look elsewhere. */
1215 if (env.best_cpu == -1) {
1216 for_each_online_node(nid) {
1217 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1220 /* Only consider nodes where both task and groups benefit */
1221 taskimp = task_weight(p, nid) - taskweight;
1222 groupimp = group_weight(p, nid) - groupweight;
1223 if (taskimp < 0 && groupimp < 0)
1227 update_numa_stats(&env.dst_stats, env.dst_nid);
1228 task_numa_find_cpu(&env, taskimp, groupimp);
1232 /* No better CPU than the current one was found. */
1233 if (env.best_cpu == -1)
1236 sched_setnuma(p, env.dst_nid);
1239 * Reset the scan period if the task is being rescheduled on an
1240 * alternative node to recheck if the tasks is now properly placed.
1242 p->numa_scan_period = task_scan_min(p);
1244 if (env.best_task == NULL) {
1245 int ret = migrate_task_to(p, env.best_cpu);
1249 ret = migrate_swap(p, env.best_task);
1250 put_task_struct(env.best_task);
1254 /* Attempt to migrate a task to a CPU on the preferred node. */
1255 static void numa_migrate_preferred(struct task_struct *p)
1257 /* This task has no NUMA fault statistics yet */
1258 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1261 /* Periodically retry migrating the task to the preferred node */
1262 p->numa_migrate_retry = jiffies + HZ;
1264 /* Success if task is already running on preferred CPU */
1265 if (task_node(p) == p->numa_preferred_nid)
1268 /* Otherwise, try migrate to a CPU on the preferred node */
1269 task_numa_migrate(p);
1273 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1274 * increments. The more local the fault statistics are, the higher the scan
1275 * period will be for the next scan window. If local/remote ratio is below
1276 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1277 * scan period will decrease
1279 #define NUMA_PERIOD_SLOTS 10
1280 #define NUMA_PERIOD_THRESHOLD 3
1283 * Increase the scan period (slow down scanning) if the majority of
1284 * our memory is already on our local node, or if the majority of
1285 * the page accesses are shared with other processes.
1286 * Otherwise, decrease the scan period.
1288 static void update_task_scan_period(struct task_struct *p,
1289 unsigned long shared, unsigned long private)
1291 unsigned int period_slot;
1295 unsigned long remote = p->numa_faults_locality[0];
1296 unsigned long local = p->numa_faults_locality[1];
1299 * If there were no record hinting faults then either the task is
1300 * completely idle or all activity is areas that are not of interest
1301 * to automatic numa balancing. Scan slower
1303 if (local + shared == 0) {
1304 p->numa_scan_period = min(p->numa_scan_period_max,
1305 p->numa_scan_period << 1);
1307 p->mm->numa_next_scan = jiffies +
1308 msecs_to_jiffies(p->numa_scan_period);
1314 * Prepare to scale scan period relative to the current period.
1315 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1316 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1317 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1319 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1320 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1321 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1322 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1325 diff = slot * period_slot;
1327 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1330 * Scale scan rate increases based on sharing. There is an
1331 * inverse relationship between the degree of sharing and
1332 * the adjustment made to the scanning period. Broadly
1333 * speaking the intent is that there is little point
1334 * scanning faster if shared accesses dominate as it may
1335 * simply bounce migrations uselessly
1337 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1338 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1341 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1342 task_scan_min(p), task_scan_max(p));
1343 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1346 static void task_numa_placement(struct task_struct *p)
1348 int seq, nid, max_nid = -1, max_group_nid = -1;
1349 unsigned long max_faults = 0, max_group_faults = 0;
1350 unsigned long fault_types[2] = { 0, 0 };
1351 spinlock_t *group_lock = NULL;
1353 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1354 if (p->numa_scan_seq == seq)
1356 p->numa_scan_seq = seq;
1357 p->numa_scan_period_max = task_scan_max(p);
1359 /* If the task is part of a group prevent parallel updates to group stats */
1360 if (p->numa_group) {
1361 group_lock = &p->numa_group->lock;
1362 spin_lock(group_lock);
1365 /* Find the node with the highest number of faults */
1366 for_each_online_node(nid) {
1367 unsigned long faults = 0, group_faults = 0;
1370 for (priv = 0; priv < 2; priv++) {
1373 i = task_faults_idx(nid, priv);
1374 diff = -p->numa_faults[i];
1376 /* Decay existing window, copy faults since last scan */
1377 p->numa_faults[i] >>= 1;
1378 p->numa_faults[i] += p->numa_faults_buffer[i];
1379 fault_types[priv] += p->numa_faults_buffer[i];
1380 p->numa_faults_buffer[i] = 0;
1382 faults += p->numa_faults[i];
1383 diff += p->numa_faults[i];
1384 p->total_numa_faults += diff;
1385 if (p->numa_group) {
1386 /* safe because we can only change our own group */
1387 p->numa_group->faults[i] += diff;
1388 p->numa_group->total_faults += diff;
1389 group_faults += p->numa_group->faults[i];
1393 if (faults > max_faults) {
1394 max_faults = faults;
1398 if (group_faults > max_group_faults) {
1399 max_group_faults = group_faults;
1400 max_group_nid = nid;
1404 update_task_scan_period(p, fault_types[0], fault_types[1]);
1406 if (p->numa_group) {
1408 * If the preferred task and group nids are different,
1409 * iterate over the nodes again to find the best place.
1411 if (max_nid != max_group_nid) {
1412 unsigned long weight, max_weight = 0;
1414 for_each_online_node(nid) {
1415 weight = task_weight(p, nid) + group_weight(p, nid);
1416 if (weight > max_weight) {
1417 max_weight = weight;
1423 spin_unlock(group_lock);
1426 /* Preferred node as the node with the most faults */
1427 if (max_faults && max_nid != p->numa_preferred_nid) {
1428 /* Update the preferred nid and migrate task if possible */
1429 sched_setnuma(p, max_nid);
1430 numa_migrate_preferred(p);
1434 static inline int get_numa_group(struct numa_group *grp)
1436 return atomic_inc_not_zero(&grp->refcount);
1439 static inline void put_numa_group(struct numa_group *grp)
1441 if (atomic_dec_and_test(&grp->refcount))
1442 kfree_rcu(grp, rcu);
1445 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1448 struct numa_group *grp, *my_grp;
1449 struct task_struct *tsk;
1451 int cpu = cpupid_to_cpu(cpupid);
1454 if (unlikely(!p->numa_group)) {
1455 unsigned int size = sizeof(struct numa_group) +
1456 2*nr_node_ids*sizeof(unsigned long);
1458 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1462 atomic_set(&grp->refcount, 1);
1463 spin_lock_init(&grp->lock);
1464 INIT_LIST_HEAD(&grp->task_list);
1467 for (i = 0; i < 2*nr_node_ids; i++)
1468 grp->faults[i] = p->numa_faults[i];
1470 grp->total_faults = p->total_numa_faults;
1472 list_add(&p->numa_entry, &grp->task_list);
1474 rcu_assign_pointer(p->numa_group, grp);
1478 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1480 if (!cpupid_match_pid(tsk, cpupid))
1483 grp = rcu_dereference(tsk->numa_group);
1487 my_grp = p->numa_group;
1492 * Only join the other group if its bigger; if we're the bigger group,
1493 * the other task will join us.
1495 if (my_grp->nr_tasks > grp->nr_tasks)
1499 * Tie-break on the grp address.
1501 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1504 /* Always join threads in the same process. */
1505 if (tsk->mm == current->mm)
1508 /* Simple filter to avoid false positives due to PID collisions */
1509 if (flags & TNF_SHARED)
1512 /* Update priv based on whether false sharing was detected */
1515 if (join && !get_numa_group(grp))
1523 double_lock(&my_grp->lock, &grp->lock);
1525 for (i = 0; i < 2*nr_node_ids; i++) {
1526 my_grp->faults[i] -= p->numa_faults[i];
1527 grp->faults[i] += p->numa_faults[i];
1529 my_grp->total_faults -= p->total_numa_faults;
1530 grp->total_faults += p->total_numa_faults;
1532 list_move(&p->numa_entry, &grp->task_list);
1536 spin_unlock(&my_grp->lock);
1537 spin_unlock(&grp->lock);
1539 rcu_assign_pointer(p->numa_group, grp);
1541 put_numa_group(my_grp);
1549 void task_numa_free(struct task_struct *p)
1551 struct numa_group *grp = p->numa_group;
1553 void *numa_faults = p->numa_faults;
1556 spin_lock(&grp->lock);
1557 for (i = 0; i < 2*nr_node_ids; i++)
1558 grp->faults[i] -= p->numa_faults[i];
1559 grp->total_faults -= p->total_numa_faults;
1561 list_del(&p->numa_entry);
1563 spin_unlock(&grp->lock);
1564 rcu_assign_pointer(p->numa_group, NULL);
1565 put_numa_group(grp);
1568 p->numa_faults = NULL;
1569 p->numa_faults_buffer = NULL;
1574 * Got a PROT_NONE fault for a page on @node.
1576 void task_numa_fault(int last_cpupid, int node, int pages, int flags)
1578 struct task_struct *p = current;
1579 bool migrated = flags & TNF_MIGRATED;
1582 if (!numabalancing_enabled)
1585 /* for example, ksmd faulting in a user's mm */
1589 /* Do not worry about placement if exiting */
1590 if (p->state == TASK_DEAD)
1593 /* Allocate buffer to track faults on a per-node basis */
1594 if (unlikely(!p->numa_faults)) {
1595 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1597 /* numa_faults and numa_faults_buffer share the allocation */
1598 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1599 if (!p->numa_faults)
1602 BUG_ON(p->numa_faults_buffer);
1603 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1604 p->total_numa_faults = 0;
1605 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1609 * First accesses are treated as private, otherwise consider accesses
1610 * to be private if the accessing pid has not changed
1612 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1615 priv = cpupid_match_pid(p, last_cpupid);
1616 if (!priv && !(flags & TNF_NO_GROUP))
1617 task_numa_group(p, last_cpupid, flags, &priv);
1620 task_numa_placement(p);
1623 * Retry task to preferred node migration periodically, in case it
1624 * case it previously failed, or the scheduler moved us.
1626 if (time_after(jiffies, p->numa_migrate_retry))
1627 numa_migrate_preferred(p);
1630 p->numa_pages_migrated += pages;
1632 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1633 p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages;
1636 static void reset_ptenuma_scan(struct task_struct *p)
1638 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1639 p->mm->numa_scan_offset = 0;
1643 * The expensive part of numa migration is done from task_work context.
1644 * Triggered from task_tick_numa().
1646 void task_numa_work(struct callback_head *work)
1648 unsigned long migrate, next_scan, now = jiffies;
1649 struct task_struct *p = current;
1650 struct mm_struct *mm = p->mm;
1651 struct vm_area_struct *vma;
1652 unsigned long start, end;
1653 unsigned long nr_pte_updates = 0;
1656 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1658 work->next = work; /* protect against double add */
1660 * Who cares about NUMA placement when they're dying.
1662 * NOTE: make sure not to dereference p->mm before this check,
1663 * exit_task_work() happens _after_ exit_mm() so we could be called
1664 * without p->mm even though we still had it when we enqueued this
1667 if (p->flags & PF_EXITING)
1670 if (!mm->numa_next_scan) {
1671 mm->numa_next_scan = now +
1672 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1676 * Enforce maximal scan/migration frequency..
1678 migrate = mm->numa_next_scan;
1679 if (time_before(now, migrate))
1682 if (p->numa_scan_period == 0) {
1683 p->numa_scan_period_max = task_scan_max(p);
1684 p->numa_scan_period = task_scan_min(p);
1687 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1688 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1692 * Delay this task enough that another task of this mm will likely win
1693 * the next time around.
1695 p->node_stamp += 2 * TICK_NSEC;
1697 start = mm->numa_scan_offset;
1698 pages = sysctl_numa_balancing_scan_size;
1699 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1703 down_read(&mm->mmap_sem);
1704 vma = find_vma(mm, start);
1706 reset_ptenuma_scan(p);
1710 for (; vma; vma = vma->vm_next) {
1711 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1715 * Shared library pages mapped by multiple processes are not
1716 * migrated as it is expected they are cache replicated. Avoid
1717 * hinting faults in read-only file-backed mappings or the vdso
1718 * as migrating the pages will be of marginal benefit.
1721 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1725 * Skip inaccessible VMAs to avoid any confusion between
1726 * PROT_NONE and NUMA hinting ptes
1728 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1732 start = max(start, vma->vm_start);
1733 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1734 end = min(end, vma->vm_end);
1735 nr_pte_updates += change_prot_numa(vma, start, end);
1738 * Scan sysctl_numa_balancing_scan_size but ensure that
1739 * at least one PTE is updated so that unused virtual
1740 * address space is quickly skipped.
1743 pages -= (end - start) >> PAGE_SHIFT;
1748 } while (end != vma->vm_end);
1753 * It is possible to reach the end of the VMA list but the last few
1754 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1755 * would find the !migratable VMA on the next scan but not reset the
1756 * scanner to the start so check it now.
1759 mm->numa_scan_offset = start;
1761 reset_ptenuma_scan(p);
1762 up_read(&mm->mmap_sem);
1766 * Drive the periodic memory faults..
1768 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1770 struct callback_head *work = &curr->numa_work;
1774 * We don't care about NUMA placement if we don't have memory.
1776 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1780 * Using runtime rather than walltime has the dual advantage that
1781 * we (mostly) drive the selection from busy threads and that the
1782 * task needs to have done some actual work before we bother with
1785 now = curr->se.sum_exec_runtime;
1786 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1788 if (now - curr->node_stamp > period) {
1789 if (!curr->node_stamp)
1790 curr->numa_scan_period = task_scan_min(curr);
1791 curr->node_stamp += period;
1793 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1794 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1795 task_work_add(curr, work, true);
1800 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1804 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1808 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1811 #endif /* CONFIG_NUMA_BALANCING */
1814 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1816 update_load_add(&cfs_rq->load, se->load.weight);
1817 if (!parent_entity(se))
1818 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1820 if (entity_is_task(se)) {
1821 struct rq *rq = rq_of(cfs_rq);
1823 account_numa_enqueue(rq, task_of(se));
1824 list_add(&se->group_node, &rq->cfs_tasks);
1827 cfs_rq->nr_running++;
1831 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1833 update_load_sub(&cfs_rq->load, se->load.weight);
1834 if (!parent_entity(se))
1835 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1836 if (entity_is_task(se)) {
1837 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
1838 list_del_init(&se->group_node);
1840 cfs_rq->nr_running--;
1843 #ifdef CONFIG_FAIR_GROUP_SCHED
1845 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1850 * Use this CPU's actual weight instead of the last load_contribution
1851 * to gain a more accurate current total weight. See
1852 * update_cfs_rq_load_contribution().
1854 tg_weight = atomic_long_read(&tg->load_avg);
1855 tg_weight -= cfs_rq->tg_load_contrib;
1856 tg_weight += cfs_rq->load.weight;
1861 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1863 long tg_weight, load, shares;
1865 tg_weight = calc_tg_weight(tg, cfs_rq);
1866 load = cfs_rq->load.weight;
1868 shares = (tg->shares * load);
1870 shares /= tg_weight;
1872 if (shares < MIN_SHARES)
1873 shares = MIN_SHARES;
1874 if (shares > tg->shares)
1875 shares = tg->shares;
1879 # else /* CONFIG_SMP */
1880 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1884 # endif /* CONFIG_SMP */
1885 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1886 unsigned long weight)
1889 /* commit outstanding execution time */
1890 if (cfs_rq->curr == se)
1891 update_curr(cfs_rq);
1892 account_entity_dequeue(cfs_rq, se);
1895 update_load_set(&se->load, weight);
1898 account_entity_enqueue(cfs_rq, se);
1901 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1903 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1905 struct task_group *tg;
1906 struct sched_entity *se;
1910 se = tg->se[cpu_of(rq_of(cfs_rq))];
1911 if (!se || throttled_hierarchy(cfs_rq))
1914 if (likely(se->load.weight == tg->shares))
1917 shares = calc_cfs_shares(cfs_rq, tg);
1919 reweight_entity(cfs_rq_of(se), se, shares);
1921 #else /* CONFIG_FAIR_GROUP_SCHED */
1922 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1925 #endif /* CONFIG_FAIR_GROUP_SCHED */
1929 * We choose a half-life close to 1 scheduling period.
1930 * Note: The tables below are dependent on this value.
1932 #define LOAD_AVG_PERIOD 32
1933 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1934 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1936 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1937 static const u32 runnable_avg_yN_inv[] = {
1938 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1939 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1940 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1941 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1942 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1943 0x85aac367, 0x82cd8698,
1947 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1948 * over-estimates when re-combining.
1950 static const u32 runnable_avg_yN_sum[] = {
1951 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1952 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1953 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1958 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1960 static __always_inline u64 decay_load(u64 val, u64 n)
1962 unsigned int local_n;
1966 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1969 /* after bounds checking we can collapse to 32-bit */
1973 * As y^PERIOD = 1/2, we can combine
1974 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1975 * With a look-up table which covers k^n (n<PERIOD)
1977 * To achieve constant time decay_load.
1979 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1980 val >>= local_n / LOAD_AVG_PERIOD;
1981 local_n %= LOAD_AVG_PERIOD;
1984 val *= runnable_avg_yN_inv[local_n];
1985 /* We don't use SRR here since we always want to round down. */
1990 * For updates fully spanning n periods, the contribution to runnable
1991 * average will be: \Sum 1024*y^n
1993 * We can compute this reasonably efficiently by combining:
1994 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1996 static u32 __compute_runnable_contrib(u64 n)
2000 if (likely(n <= LOAD_AVG_PERIOD))
2001 return runnable_avg_yN_sum[n];
2002 else if (unlikely(n >= LOAD_AVG_MAX_N))
2003 return LOAD_AVG_MAX;
2005 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2007 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2008 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2010 n -= LOAD_AVG_PERIOD;
2011 } while (n > LOAD_AVG_PERIOD);
2013 contrib = decay_load(contrib, n);
2014 return contrib + runnable_avg_yN_sum[n];
2018 * We can represent the historical contribution to runnable average as the
2019 * coefficients of a geometric series. To do this we sub-divide our runnable
2020 * history into segments of approximately 1ms (1024us); label the segment that
2021 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2023 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2025 * (now) (~1ms ago) (~2ms ago)
2027 * Let u_i denote the fraction of p_i that the entity was runnable.
2029 * We then designate the fractions u_i as our co-efficients, yielding the
2030 * following representation of historical load:
2031 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2033 * We choose y based on the with of a reasonably scheduling period, fixing:
2036 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2037 * approximately half as much as the contribution to load within the last ms
2040 * When a period "rolls over" and we have new u_0`, multiplying the previous
2041 * sum again by y is sufficient to update:
2042 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2043 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2045 static __always_inline int __update_entity_runnable_avg(u64 now,
2046 struct sched_avg *sa,
2050 u32 runnable_contrib;
2051 int delta_w, decayed = 0;
2053 delta = now - sa->last_runnable_update;
2055 * This should only happen when time goes backwards, which it
2056 * unfortunately does during sched clock init when we swap over to TSC.
2058 if ((s64)delta < 0) {
2059 sa->last_runnable_update = now;
2064 * Use 1024ns as the unit of measurement since it's a reasonable
2065 * approximation of 1us and fast to compute.
2070 sa->last_runnable_update = now;
2072 /* delta_w is the amount already accumulated against our next period */
2073 delta_w = sa->runnable_avg_period % 1024;
2074 if (delta + delta_w >= 1024) {
2075 /* period roll-over */
2079 * Now that we know we're crossing a period boundary, figure
2080 * out how much from delta we need to complete the current
2081 * period and accrue it.
2083 delta_w = 1024 - delta_w;
2085 sa->runnable_avg_sum += delta_w;
2086 sa->runnable_avg_period += delta_w;
2090 /* Figure out how many additional periods this update spans */
2091 periods = delta / 1024;
2094 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2096 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2099 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2100 runnable_contrib = __compute_runnable_contrib(periods);
2102 sa->runnable_avg_sum += runnable_contrib;
2103 sa->runnable_avg_period += runnable_contrib;
2106 /* Remainder of delta accrued against u_0` */
2108 sa->runnable_avg_sum += delta;
2109 sa->runnable_avg_period += delta;
2114 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2115 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2117 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2118 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2120 decays -= se->avg.decay_count;
2124 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2125 se->avg.decay_count = 0;
2130 #ifdef CONFIG_FAIR_GROUP_SCHED
2131 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2134 struct task_group *tg = cfs_rq->tg;
2137 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2138 tg_contrib -= cfs_rq->tg_load_contrib;
2140 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2141 atomic_long_add(tg_contrib, &tg->load_avg);
2142 cfs_rq->tg_load_contrib += tg_contrib;
2147 * Aggregate cfs_rq runnable averages into an equivalent task_group
2148 * representation for computing load contributions.
2150 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2151 struct cfs_rq *cfs_rq)
2153 struct task_group *tg = cfs_rq->tg;
2156 /* The fraction of a cpu used by this cfs_rq */
2157 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2158 sa->runnable_avg_period + 1);
2159 contrib -= cfs_rq->tg_runnable_contrib;
2161 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2162 atomic_add(contrib, &tg->runnable_avg);
2163 cfs_rq->tg_runnable_contrib += contrib;
2167 static inline void __update_group_entity_contrib(struct sched_entity *se)
2169 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2170 struct task_group *tg = cfs_rq->tg;
2175 contrib = cfs_rq->tg_load_contrib * tg->shares;
2176 se->avg.load_avg_contrib = div_u64(contrib,
2177 atomic_long_read(&tg->load_avg) + 1);
2180 * For group entities we need to compute a correction term in the case
2181 * that they are consuming <1 cpu so that we would contribute the same
2182 * load as a task of equal weight.
2184 * Explicitly co-ordinating this measurement would be expensive, but
2185 * fortunately the sum of each cpus contribution forms a usable
2186 * lower-bound on the true value.
2188 * Consider the aggregate of 2 contributions. Either they are disjoint
2189 * (and the sum represents true value) or they are disjoint and we are
2190 * understating by the aggregate of their overlap.
2192 * Extending this to N cpus, for a given overlap, the maximum amount we
2193 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2194 * cpus that overlap for this interval and w_i is the interval width.
2196 * On a small machine; the first term is well-bounded which bounds the
2197 * total error since w_i is a subset of the period. Whereas on a
2198 * larger machine, while this first term can be larger, if w_i is the
2199 * of consequential size guaranteed to see n_i*w_i quickly converge to
2200 * our upper bound of 1-cpu.
2202 runnable_avg = atomic_read(&tg->runnable_avg);
2203 if (runnable_avg < NICE_0_LOAD) {
2204 se->avg.load_avg_contrib *= runnable_avg;
2205 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2209 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2210 int force_update) {}
2211 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2212 struct cfs_rq *cfs_rq) {}
2213 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2216 static inline void __update_task_entity_contrib(struct sched_entity *se)
2220 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2221 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2222 contrib /= (se->avg.runnable_avg_period + 1);
2223 se->avg.load_avg_contrib = scale_load(contrib);
2226 /* Compute the current contribution to load_avg by se, return any delta */
2227 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2229 long old_contrib = se->avg.load_avg_contrib;
2231 if (entity_is_task(se)) {
2232 __update_task_entity_contrib(se);
2234 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2235 __update_group_entity_contrib(se);
2238 return se->avg.load_avg_contrib - old_contrib;
2241 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2244 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2245 cfs_rq->blocked_load_avg -= load_contrib;
2247 cfs_rq->blocked_load_avg = 0;
2250 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2252 /* Update a sched_entity's runnable average */
2253 static inline void update_entity_load_avg(struct sched_entity *se,
2256 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2261 * For a group entity we need to use their owned cfs_rq_clock_task() in
2262 * case they are the parent of a throttled hierarchy.
2264 if (entity_is_task(se))
2265 now = cfs_rq_clock_task(cfs_rq);
2267 now = cfs_rq_clock_task(group_cfs_rq(se));
2269 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2272 contrib_delta = __update_entity_load_avg_contrib(se);
2278 cfs_rq->runnable_load_avg += contrib_delta;
2280 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2284 * Decay the load contributed by all blocked children and account this so that
2285 * their contribution may appropriately discounted when they wake up.
2287 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2289 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2292 decays = now - cfs_rq->last_decay;
2293 if (!decays && !force_update)
2296 if (atomic_long_read(&cfs_rq->removed_load)) {
2297 unsigned long removed_load;
2298 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2299 subtract_blocked_load_contrib(cfs_rq, removed_load);
2303 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2305 atomic64_add(decays, &cfs_rq->decay_counter);
2306 cfs_rq->last_decay = now;
2309 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2312 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2314 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2315 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2318 /* Add the load generated by se into cfs_rq's child load-average */
2319 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2320 struct sched_entity *se,
2324 * We track migrations using entity decay_count <= 0, on a wake-up
2325 * migration we use a negative decay count to track the remote decays
2326 * accumulated while sleeping.
2328 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2329 * are seen by enqueue_entity_load_avg() as a migration with an already
2330 * constructed load_avg_contrib.
2332 if (unlikely(se->avg.decay_count <= 0)) {
2333 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2334 if (se->avg.decay_count) {
2336 * In a wake-up migration we have to approximate the
2337 * time sleeping. This is because we can't synchronize
2338 * clock_task between the two cpus, and it is not
2339 * guaranteed to be read-safe. Instead, we can
2340 * approximate this using our carried decays, which are
2341 * explicitly atomically readable.
2343 se->avg.last_runnable_update -= (-se->avg.decay_count)
2345 update_entity_load_avg(se, 0);
2346 /* Indicate that we're now synchronized and on-rq */
2347 se->avg.decay_count = 0;
2351 __synchronize_entity_decay(se);
2354 /* migrated tasks did not contribute to our blocked load */
2356 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2357 update_entity_load_avg(se, 0);
2360 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2361 /* we force update consideration on load-balancer moves */
2362 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2366 * Remove se's load from this cfs_rq child load-average, if the entity is
2367 * transitioning to a blocked state we track its projected decay using
2370 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2371 struct sched_entity *se,
2374 update_entity_load_avg(se, 1);
2375 /* we force update consideration on load-balancer moves */
2376 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2378 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2380 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2381 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2382 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2386 * Update the rq's load with the elapsed running time before entering
2387 * idle. if the last scheduled task is not a CFS task, idle_enter will
2388 * be the only way to update the runnable statistic.
2390 void idle_enter_fair(struct rq *this_rq)
2392 update_rq_runnable_avg(this_rq, 1);
2396 * Update the rq's load with the elapsed idle time before a task is
2397 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2398 * be the only way to update the runnable statistic.
2400 void idle_exit_fair(struct rq *this_rq)
2402 update_rq_runnable_avg(this_rq, 0);
2406 static inline void update_entity_load_avg(struct sched_entity *se,
2407 int update_cfs_rq) {}
2408 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2409 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2410 struct sched_entity *se,
2412 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2413 struct sched_entity *se,
2415 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2416 int force_update) {}
2419 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2421 #ifdef CONFIG_SCHEDSTATS
2422 struct task_struct *tsk = NULL;
2424 if (entity_is_task(se))
2427 if (se->statistics.sleep_start) {
2428 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2433 if (unlikely(delta > se->statistics.sleep_max))
2434 se->statistics.sleep_max = delta;
2436 se->statistics.sleep_start = 0;
2437 se->statistics.sum_sleep_runtime += delta;
2440 account_scheduler_latency(tsk, delta >> 10, 1);
2441 trace_sched_stat_sleep(tsk, delta);
2444 if (se->statistics.block_start) {
2445 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2450 if (unlikely(delta > se->statistics.block_max))
2451 se->statistics.block_max = delta;
2453 se->statistics.block_start = 0;
2454 se->statistics.sum_sleep_runtime += delta;
2457 if (tsk->in_iowait) {
2458 se->statistics.iowait_sum += delta;
2459 se->statistics.iowait_count++;
2460 trace_sched_stat_iowait(tsk, delta);
2463 trace_sched_stat_blocked(tsk, delta);
2466 * Blocking time is in units of nanosecs, so shift by
2467 * 20 to get a milliseconds-range estimation of the
2468 * amount of time that the task spent sleeping:
2470 if (unlikely(prof_on == SLEEP_PROFILING)) {
2471 profile_hits(SLEEP_PROFILING,
2472 (void *)get_wchan(tsk),
2475 account_scheduler_latency(tsk, delta >> 10, 0);
2481 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2483 #ifdef CONFIG_SCHED_DEBUG
2484 s64 d = se->vruntime - cfs_rq->min_vruntime;
2489 if (d > 3*sysctl_sched_latency)
2490 schedstat_inc(cfs_rq, nr_spread_over);
2495 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2497 u64 vruntime = cfs_rq->min_vruntime;
2500 * The 'current' period is already promised to the current tasks,
2501 * however the extra weight of the new task will slow them down a
2502 * little, place the new task so that it fits in the slot that
2503 * stays open at the end.
2505 if (initial && sched_feat(START_DEBIT))
2506 vruntime += sched_vslice(cfs_rq, se);
2508 /* sleeps up to a single latency don't count. */
2510 unsigned long thresh = sysctl_sched_latency;
2513 * Halve their sleep time's effect, to allow
2514 * for a gentler effect of sleepers:
2516 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2522 /* ensure we never gain time by being placed backwards. */
2523 se->vruntime = max_vruntime(se->vruntime, vruntime);
2526 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2529 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2532 * Update the normalized vruntime before updating min_vruntime
2533 * through calling update_curr().
2535 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2536 se->vruntime += cfs_rq->min_vruntime;
2539 * Update run-time statistics of the 'current'.
2541 update_curr(cfs_rq);
2542 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2543 account_entity_enqueue(cfs_rq, se);
2544 update_cfs_shares(cfs_rq);
2546 if (flags & ENQUEUE_WAKEUP) {
2547 place_entity(cfs_rq, se, 0);
2548 enqueue_sleeper(cfs_rq, se);
2551 update_stats_enqueue(cfs_rq, se);
2552 check_spread(cfs_rq, se);
2553 if (se != cfs_rq->curr)
2554 __enqueue_entity(cfs_rq, se);
2557 if (cfs_rq->nr_running == 1) {
2558 list_add_leaf_cfs_rq(cfs_rq);
2559 check_enqueue_throttle(cfs_rq);
2563 static void __clear_buddies_last(struct sched_entity *se)
2565 for_each_sched_entity(se) {
2566 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2567 if (cfs_rq->last == se)
2568 cfs_rq->last = NULL;
2574 static void __clear_buddies_next(struct sched_entity *se)
2576 for_each_sched_entity(se) {
2577 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2578 if (cfs_rq->next == se)
2579 cfs_rq->next = NULL;
2585 static void __clear_buddies_skip(struct sched_entity *se)
2587 for_each_sched_entity(se) {
2588 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2589 if (cfs_rq->skip == se)
2590 cfs_rq->skip = NULL;
2596 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2598 if (cfs_rq->last == se)
2599 __clear_buddies_last(se);
2601 if (cfs_rq->next == se)
2602 __clear_buddies_next(se);
2604 if (cfs_rq->skip == se)
2605 __clear_buddies_skip(se);
2608 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2611 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2614 * Update run-time statistics of the 'current'.
2616 update_curr(cfs_rq);
2617 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2619 update_stats_dequeue(cfs_rq, se);
2620 if (flags & DEQUEUE_SLEEP) {
2621 #ifdef CONFIG_SCHEDSTATS
2622 if (entity_is_task(se)) {
2623 struct task_struct *tsk = task_of(se);
2625 if (tsk->state & TASK_INTERRUPTIBLE)
2626 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2627 if (tsk->state & TASK_UNINTERRUPTIBLE)
2628 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2633 clear_buddies(cfs_rq, se);
2635 if (se != cfs_rq->curr)
2636 __dequeue_entity(cfs_rq, se);
2638 account_entity_dequeue(cfs_rq, se);
2641 * Normalize the entity after updating the min_vruntime because the
2642 * update can refer to the ->curr item and we need to reflect this
2643 * movement in our normalized position.
2645 if (!(flags & DEQUEUE_SLEEP))
2646 se->vruntime -= cfs_rq->min_vruntime;
2648 /* return excess runtime on last dequeue */
2649 return_cfs_rq_runtime(cfs_rq);
2651 update_min_vruntime(cfs_rq);
2652 update_cfs_shares(cfs_rq);
2656 * Preempt the current task with a newly woken task if needed:
2659 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2661 unsigned long ideal_runtime, delta_exec;
2662 struct sched_entity *se;
2665 ideal_runtime = sched_slice(cfs_rq, curr);
2666 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2667 if (delta_exec > ideal_runtime) {
2668 resched_task(rq_of(cfs_rq)->curr);
2670 * The current task ran long enough, ensure it doesn't get
2671 * re-elected due to buddy favours.
2673 clear_buddies(cfs_rq, curr);
2678 * Ensure that a task that missed wakeup preemption by a
2679 * narrow margin doesn't have to wait for a full slice.
2680 * This also mitigates buddy induced latencies under load.
2682 if (delta_exec < sysctl_sched_min_granularity)
2685 se = __pick_first_entity(cfs_rq);
2686 delta = curr->vruntime - se->vruntime;
2691 if (delta > ideal_runtime)
2692 resched_task(rq_of(cfs_rq)->curr);
2696 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2698 /* 'current' is not kept within the tree. */
2701 * Any task has to be enqueued before it get to execute on
2702 * a CPU. So account for the time it spent waiting on the
2705 update_stats_wait_end(cfs_rq, se);
2706 __dequeue_entity(cfs_rq, se);
2709 update_stats_curr_start(cfs_rq, se);
2711 #ifdef CONFIG_SCHEDSTATS
2713 * Track our maximum slice length, if the CPU's load is at
2714 * least twice that of our own weight (i.e. dont track it
2715 * when there are only lesser-weight tasks around):
2717 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2718 se->statistics.slice_max = max(se->statistics.slice_max,
2719 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2722 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2726 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2729 * Pick the next process, keeping these things in mind, in this order:
2730 * 1) keep things fair between processes/task groups
2731 * 2) pick the "next" process, since someone really wants that to run
2732 * 3) pick the "last" process, for cache locality
2733 * 4) do not run the "skip" process, if something else is available
2735 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2737 struct sched_entity *se = __pick_first_entity(cfs_rq);
2738 struct sched_entity *left = se;
2741 * Avoid running the skip buddy, if running something else can
2742 * be done without getting too unfair.
2744 if (cfs_rq->skip == se) {
2745 struct sched_entity *second = __pick_next_entity(se);
2746 if (second && wakeup_preempt_entity(second, left) < 1)
2751 * Prefer last buddy, try to return the CPU to a preempted task.
2753 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2757 * Someone really wants this to run. If it's not unfair, run it.
2759 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2762 clear_buddies(cfs_rq, se);
2767 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2769 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2772 * If still on the runqueue then deactivate_task()
2773 * was not called and update_curr() has to be done:
2776 update_curr(cfs_rq);
2778 /* throttle cfs_rqs exceeding runtime */
2779 check_cfs_rq_runtime(cfs_rq);
2781 check_spread(cfs_rq, prev);
2783 update_stats_wait_start(cfs_rq, prev);
2784 /* Put 'current' back into the tree. */
2785 __enqueue_entity(cfs_rq, prev);
2786 /* in !on_rq case, update occurred at dequeue */
2787 update_entity_load_avg(prev, 1);
2789 cfs_rq->curr = NULL;
2793 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2796 * Update run-time statistics of the 'current'.
2798 update_curr(cfs_rq);
2801 * Ensure that runnable average is periodically updated.
2803 update_entity_load_avg(curr, 1);
2804 update_cfs_rq_blocked_load(cfs_rq, 1);
2805 update_cfs_shares(cfs_rq);
2807 #ifdef CONFIG_SCHED_HRTICK
2809 * queued ticks are scheduled to match the slice, so don't bother
2810 * validating it and just reschedule.
2813 resched_task(rq_of(cfs_rq)->curr);
2817 * don't let the period tick interfere with the hrtick preemption
2819 if (!sched_feat(DOUBLE_TICK) &&
2820 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2824 if (cfs_rq->nr_running > 1)
2825 check_preempt_tick(cfs_rq, curr);
2829 /**************************************************
2830 * CFS bandwidth control machinery
2833 #ifdef CONFIG_CFS_BANDWIDTH
2835 #ifdef HAVE_JUMP_LABEL
2836 static struct static_key __cfs_bandwidth_used;
2838 static inline bool cfs_bandwidth_used(void)
2840 return static_key_false(&__cfs_bandwidth_used);
2843 void cfs_bandwidth_usage_inc(void)
2845 static_key_slow_inc(&__cfs_bandwidth_used);
2848 void cfs_bandwidth_usage_dec(void)
2850 static_key_slow_dec(&__cfs_bandwidth_used);
2852 #else /* HAVE_JUMP_LABEL */
2853 static bool cfs_bandwidth_used(void)
2858 void cfs_bandwidth_usage_inc(void) {}
2859 void cfs_bandwidth_usage_dec(void) {}
2860 #endif /* HAVE_JUMP_LABEL */
2863 * default period for cfs group bandwidth.
2864 * default: 0.1s, units: nanoseconds
2866 static inline u64 default_cfs_period(void)
2868 return 100000000ULL;
2871 static inline u64 sched_cfs_bandwidth_slice(void)
2873 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2877 * Replenish runtime according to assigned quota and update expiration time.
2878 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2879 * additional synchronization around rq->lock.
2881 * requires cfs_b->lock
2883 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2887 if (cfs_b->quota == RUNTIME_INF)
2890 now = sched_clock_cpu(smp_processor_id());
2891 cfs_b->runtime = cfs_b->quota;
2892 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2895 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2897 return &tg->cfs_bandwidth;
2900 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2901 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2903 if (unlikely(cfs_rq->throttle_count))
2904 return cfs_rq->throttled_clock_task;
2906 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2909 /* returns 0 on failure to allocate runtime */
2910 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2912 struct task_group *tg = cfs_rq->tg;
2913 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2914 u64 amount = 0, min_amount, expires;
2916 /* note: this is a positive sum as runtime_remaining <= 0 */
2917 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2919 raw_spin_lock(&cfs_b->lock);
2920 if (cfs_b->quota == RUNTIME_INF)
2921 amount = min_amount;
2924 * If the bandwidth pool has become inactive, then at least one
2925 * period must have elapsed since the last consumption.
2926 * Refresh the global state and ensure bandwidth timer becomes
2929 if (!cfs_b->timer_active) {
2930 __refill_cfs_bandwidth_runtime(cfs_b);
2931 __start_cfs_bandwidth(cfs_b);
2934 if (cfs_b->runtime > 0) {
2935 amount = min(cfs_b->runtime, min_amount);
2936 cfs_b->runtime -= amount;
2940 expires = cfs_b->runtime_expires;
2941 raw_spin_unlock(&cfs_b->lock);
2943 cfs_rq->runtime_remaining += amount;
2945 * we may have advanced our local expiration to account for allowed
2946 * spread between our sched_clock and the one on which runtime was
2949 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2950 cfs_rq->runtime_expires = expires;
2952 return cfs_rq->runtime_remaining > 0;
2956 * Note: This depends on the synchronization provided by sched_clock and the
2957 * fact that rq->clock snapshots this value.
2959 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2961 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2963 /* if the deadline is ahead of our clock, nothing to do */
2964 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2967 if (cfs_rq->runtime_remaining < 0)
2971 * If the local deadline has passed we have to consider the
2972 * possibility that our sched_clock is 'fast' and the global deadline
2973 * has not truly expired.
2975 * Fortunately we can check determine whether this the case by checking
2976 * whether the global deadline has advanced.
2979 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2980 /* extend local deadline, drift is bounded above by 2 ticks */
2981 cfs_rq->runtime_expires += TICK_NSEC;
2983 /* global deadline is ahead, expiration has passed */
2984 cfs_rq->runtime_remaining = 0;
2988 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
2990 /* dock delta_exec before expiring quota (as it could span periods) */
2991 cfs_rq->runtime_remaining -= delta_exec;
2992 expire_cfs_rq_runtime(cfs_rq);
2994 if (likely(cfs_rq->runtime_remaining > 0))
2998 * if we're unable to extend our runtime we resched so that the active
2999 * hierarchy can be throttled
3001 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3002 resched_task(rq_of(cfs_rq)->curr);
3005 static __always_inline
3006 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3008 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3011 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3014 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3016 return cfs_bandwidth_used() && cfs_rq->throttled;
3019 /* check whether cfs_rq, or any parent, is throttled */
3020 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3022 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3026 * Ensure that neither of the group entities corresponding to src_cpu or
3027 * dest_cpu are members of a throttled hierarchy when performing group
3028 * load-balance operations.
3030 static inline int throttled_lb_pair(struct task_group *tg,
3031 int src_cpu, int dest_cpu)
3033 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3035 src_cfs_rq = tg->cfs_rq[src_cpu];
3036 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3038 return throttled_hierarchy(src_cfs_rq) ||
3039 throttled_hierarchy(dest_cfs_rq);
3042 /* updated child weight may affect parent so we have to do this bottom up */
3043 static int tg_unthrottle_up(struct task_group *tg, void *data)
3045 struct rq *rq = data;
3046 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3048 cfs_rq->throttle_count--;
3050 if (!cfs_rq->throttle_count) {
3051 /* adjust cfs_rq_clock_task() */
3052 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3053 cfs_rq->throttled_clock_task;
3060 static int tg_throttle_down(struct task_group *tg, void *data)
3062 struct rq *rq = data;
3063 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3065 /* group is entering throttled state, stop time */
3066 if (!cfs_rq->throttle_count)
3067 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3068 cfs_rq->throttle_count++;
3073 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3075 struct rq *rq = rq_of(cfs_rq);
3076 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3077 struct sched_entity *se;
3078 long task_delta, dequeue = 1;
3080 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3082 /* freeze hierarchy runnable averages while throttled */
3084 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3087 task_delta = cfs_rq->h_nr_running;
3088 for_each_sched_entity(se) {
3089 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3090 /* throttled entity or throttle-on-deactivate */
3095 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3096 qcfs_rq->h_nr_running -= task_delta;
3098 if (qcfs_rq->load.weight)
3103 rq->nr_running -= task_delta;
3105 cfs_rq->throttled = 1;
3106 cfs_rq->throttled_clock = rq_clock(rq);
3107 raw_spin_lock(&cfs_b->lock);
3108 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3109 if (!cfs_b->timer_active)
3110 __start_cfs_bandwidth(cfs_b);
3111 raw_spin_unlock(&cfs_b->lock);
3114 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3116 struct rq *rq = rq_of(cfs_rq);
3117 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3118 struct sched_entity *se;
3122 se = cfs_rq->tg->se[cpu_of(rq)];
3124 cfs_rq->throttled = 0;
3126 update_rq_clock(rq);
3128 raw_spin_lock(&cfs_b->lock);
3129 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3130 list_del_rcu(&cfs_rq->throttled_list);
3131 raw_spin_unlock(&cfs_b->lock);
3133 /* update hierarchical throttle state */
3134 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3136 if (!cfs_rq->load.weight)
3139 task_delta = cfs_rq->h_nr_running;
3140 for_each_sched_entity(se) {
3144 cfs_rq = cfs_rq_of(se);
3146 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3147 cfs_rq->h_nr_running += task_delta;
3149 if (cfs_rq_throttled(cfs_rq))
3154 rq->nr_running += task_delta;
3156 /* determine whether we need to wake up potentially idle cpu */
3157 if (rq->curr == rq->idle && rq->cfs.nr_running)
3158 resched_task(rq->curr);
3161 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3162 u64 remaining, u64 expires)
3164 struct cfs_rq *cfs_rq;
3165 u64 runtime = remaining;
3168 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3170 struct rq *rq = rq_of(cfs_rq);
3172 raw_spin_lock(&rq->lock);
3173 if (!cfs_rq_throttled(cfs_rq))
3176 runtime = -cfs_rq->runtime_remaining + 1;
3177 if (runtime > remaining)
3178 runtime = remaining;
3179 remaining -= runtime;
3181 cfs_rq->runtime_remaining += runtime;
3182 cfs_rq->runtime_expires = expires;
3184 /* we check whether we're throttled above */
3185 if (cfs_rq->runtime_remaining > 0)
3186 unthrottle_cfs_rq(cfs_rq);
3189 raw_spin_unlock(&rq->lock);
3200 * Responsible for refilling a task_group's bandwidth and unthrottling its
3201 * cfs_rqs as appropriate. If there has been no activity within the last
3202 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3203 * used to track this state.
3205 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3207 u64 runtime, runtime_expires;
3208 int idle = 1, throttled;
3210 raw_spin_lock(&cfs_b->lock);
3211 /* no need to continue the timer with no bandwidth constraint */
3212 if (cfs_b->quota == RUNTIME_INF)
3215 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3216 /* idle depends on !throttled (for the case of a large deficit) */
3217 idle = cfs_b->idle && !throttled;
3218 cfs_b->nr_periods += overrun;
3220 /* if we're going inactive then everything else can be deferred */
3225 * if we have relooped after returning idle once, we need to update our
3226 * status as actually running, so that other cpus doing
3227 * __start_cfs_bandwidth will stop trying to cancel us.
3229 cfs_b->timer_active = 1;
3231 __refill_cfs_bandwidth_runtime(cfs_b);
3234 /* mark as potentially idle for the upcoming period */
3239 /* account preceding periods in which throttling occurred */
3240 cfs_b->nr_throttled += overrun;
3243 * There are throttled entities so we must first use the new bandwidth
3244 * to unthrottle them before making it generally available. This
3245 * ensures that all existing debts will be paid before a new cfs_rq is
3248 runtime = cfs_b->runtime;
3249 runtime_expires = cfs_b->runtime_expires;
3253 * This check is repeated as we are holding onto the new bandwidth
3254 * while we unthrottle. This can potentially race with an unthrottled
3255 * group trying to acquire new bandwidth from the global pool.
3257 while (throttled && runtime > 0) {
3258 raw_spin_unlock(&cfs_b->lock);
3259 /* we can't nest cfs_b->lock while distributing bandwidth */
3260 runtime = distribute_cfs_runtime(cfs_b, runtime,
3262 raw_spin_lock(&cfs_b->lock);
3264 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3267 /* return (any) remaining runtime */
3268 cfs_b->runtime = runtime;
3270 * While we are ensured activity in the period following an
3271 * unthrottle, this also covers the case in which the new bandwidth is
3272 * insufficient to cover the existing bandwidth deficit. (Forcing the
3273 * timer to remain active while there are any throttled entities.)
3278 cfs_b->timer_active = 0;
3279 raw_spin_unlock(&cfs_b->lock);
3284 /* a cfs_rq won't donate quota below this amount */
3285 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3286 /* minimum remaining period time to redistribute slack quota */
3287 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3288 /* how long we wait to gather additional slack before distributing */
3289 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3292 * Are we near the end of the current quota period?
3294 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3295 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3296 * migrate_hrtimers, base is never cleared, so we are fine.
3298 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3300 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3303 /* if the call-back is running a quota refresh is already occurring */
3304 if (hrtimer_callback_running(refresh_timer))
3307 /* is a quota refresh about to occur? */
3308 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3309 if (remaining < min_expire)
3315 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3317 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3319 /* if there's a quota refresh soon don't bother with slack */
3320 if (runtime_refresh_within(cfs_b, min_left))
3323 start_bandwidth_timer(&cfs_b->slack_timer,
3324 ns_to_ktime(cfs_bandwidth_slack_period));
3327 /* we know any runtime found here is valid as update_curr() precedes return */
3328 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3330 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3331 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3333 if (slack_runtime <= 0)
3336 raw_spin_lock(&cfs_b->lock);
3337 if (cfs_b->quota != RUNTIME_INF &&
3338 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3339 cfs_b->runtime += slack_runtime;
3341 /* we are under rq->lock, defer unthrottling using a timer */
3342 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3343 !list_empty(&cfs_b->throttled_cfs_rq))
3344 start_cfs_slack_bandwidth(cfs_b);
3346 raw_spin_unlock(&cfs_b->lock);
3348 /* even if it's not valid for return we don't want to try again */
3349 cfs_rq->runtime_remaining -= slack_runtime;
3352 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3354 if (!cfs_bandwidth_used())
3357 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3360 __return_cfs_rq_runtime(cfs_rq);
3364 * This is done with a timer (instead of inline with bandwidth return) since
3365 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3367 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3369 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3372 /* confirm we're still not at a refresh boundary */
3373 raw_spin_lock(&cfs_b->lock);
3374 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3375 raw_spin_unlock(&cfs_b->lock);
3379 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3380 runtime = cfs_b->runtime;
3383 expires = cfs_b->runtime_expires;
3384 raw_spin_unlock(&cfs_b->lock);
3389 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3391 raw_spin_lock(&cfs_b->lock);
3392 if (expires == cfs_b->runtime_expires)
3393 cfs_b->runtime = runtime;
3394 raw_spin_unlock(&cfs_b->lock);
3398 * When a group wakes up we want to make sure that its quota is not already
3399 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3400 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3402 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3404 if (!cfs_bandwidth_used())
3407 /* an active group must be handled by the update_curr()->put() path */
3408 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3411 /* ensure the group is not already throttled */
3412 if (cfs_rq_throttled(cfs_rq))
3415 /* update runtime allocation */
3416 account_cfs_rq_runtime(cfs_rq, 0);
3417 if (cfs_rq->runtime_remaining <= 0)
3418 throttle_cfs_rq(cfs_rq);
3421 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3422 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3424 if (!cfs_bandwidth_used())
3427 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3431 * it's possible for a throttled entity to be forced into a running
3432 * state (e.g. set_curr_task), in this case we're finished.
3434 if (cfs_rq_throttled(cfs_rq))
3437 throttle_cfs_rq(cfs_rq);
3440 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3442 struct cfs_bandwidth *cfs_b =
3443 container_of(timer, struct cfs_bandwidth, slack_timer);
3444 do_sched_cfs_slack_timer(cfs_b);
3446 return HRTIMER_NORESTART;
3449 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3451 struct cfs_bandwidth *cfs_b =
3452 container_of(timer, struct cfs_bandwidth, period_timer);
3458 now = hrtimer_cb_get_time(timer);
3459 overrun = hrtimer_forward(timer, now, cfs_b->period);
3464 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3467 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3470 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3472 raw_spin_lock_init(&cfs_b->lock);
3474 cfs_b->quota = RUNTIME_INF;
3475 cfs_b->period = ns_to_ktime(default_cfs_period());
3477 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3478 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3479 cfs_b->period_timer.function = sched_cfs_period_timer;
3480 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3481 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3484 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3486 cfs_rq->runtime_enabled = 0;
3487 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3490 /* requires cfs_b->lock, may release to reprogram timer */
3491 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3494 * The timer may be active because we're trying to set a new bandwidth
3495 * period or because we're racing with the tear-down path
3496 * (timer_active==0 becomes visible before the hrtimer call-back
3497 * terminates). In either case we ensure that it's re-programmed
3499 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3500 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3501 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3502 raw_spin_unlock(&cfs_b->lock);
3504 raw_spin_lock(&cfs_b->lock);
3505 /* if someone else restarted the timer then we're done */
3506 if (cfs_b->timer_active)
3510 cfs_b->timer_active = 1;
3511 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3514 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3516 hrtimer_cancel(&cfs_b->period_timer);
3517 hrtimer_cancel(&cfs_b->slack_timer);
3520 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3522 struct cfs_rq *cfs_rq;
3524 for_each_leaf_cfs_rq(rq, cfs_rq) {
3525 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3527 if (!cfs_rq->runtime_enabled)
3531 * clock_task is not advancing so we just need to make sure
3532 * there's some valid quota amount
3534 cfs_rq->runtime_remaining = cfs_b->quota;
3535 if (cfs_rq_throttled(cfs_rq))
3536 unthrottle_cfs_rq(cfs_rq);
3540 #else /* CONFIG_CFS_BANDWIDTH */
3541 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3543 return rq_clock_task(rq_of(cfs_rq));
3546 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3547 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3548 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3549 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3551 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3556 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3561 static inline int throttled_lb_pair(struct task_group *tg,
3562 int src_cpu, int dest_cpu)
3567 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3569 #ifdef CONFIG_FAIR_GROUP_SCHED
3570 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3573 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3577 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3578 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3580 #endif /* CONFIG_CFS_BANDWIDTH */
3582 /**************************************************
3583 * CFS operations on tasks:
3586 #ifdef CONFIG_SCHED_HRTICK
3587 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3589 struct sched_entity *se = &p->se;
3590 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3592 WARN_ON(task_rq(p) != rq);
3594 if (cfs_rq->nr_running > 1) {
3595 u64 slice = sched_slice(cfs_rq, se);
3596 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3597 s64 delta = slice - ran;
3606 * Don't schedule slices shorter than 10000ns, that just
3607 * doesn't make sense. Rely on vruntime for fairness.
3610 delta = max_t(s64, 10000LL, delta);
3612 hrtick_start(rq, delta);
3617 * called from enqueue/dequeue and updates the hrtick when the
3618 * current task is from our class and nr_running is low enough
3621 static void hrtick_update(struct rq *rq)
3623 struct task_struct *curr = rq->curr;
3625 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3628 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3629 hrtick_start_fair(rq, curr);
3631 #else /* !CONFIG_SCHED_HRTICK */
3633 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3637 static inline void hrtick_update(struct rq *rq)
3643 * The enqueue_task method is called before nr_running is
3644 * increased. Here we update the fair scheduling stats and
3645 * then put the task into the rbtree:
3648 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3650 struct cfs_rq *cfs_rq;
3651 struct sched_entity *se = &p->se;
3653 for_each_sched_entity(se) {
3656 cfs_rq = cfs_rq_of(se);
3657 enqueue_entity(cfs_rq, se, flags);
3660 * end evaluation on encountering a throttled cfs_rq
3662 * note: in the case of encountering a throttled cfs_rq we will
3663 * post the final h_nr_running increment below.
3665 if (cfs_rq_throttled(cfs_rq))
3667 cfs_rq->h_nr_running++;
3669 flags = ENQUEUE_WAKEUP;
3672 for_each_sched_entity(se) {
3673 cfs_rq = cfs_rq_of(se);
3674 cfs_rq->h_nr_running++;
3676 if (cfs_rq_throttled(cfs_rq))
3679 update_cfs_shares(cfs_rq);
3680 update_entity_load_avg(se, 1);
3684 update_rq_runnable_avg(rq, rq->nr_running);
3690 static void set_next_buddy(struct sched_entity *se);
3693 * The dequeue_task method is called before nr_running is
3694 * decreased. We remove the task from the rbtree and
3695 * update the fair scheduling stats:
3697 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3699 struct cfs_rq *cfs_rq;
3700 struct sched_entity *se = &p->se;
3701 int task_sleep = flags & DEQUEUE_SLEEP;
3703 for_each_sched_entity(se) {
3704 cfs_rq = cfs_rq_of(se);
3705 dequeue_entity(cfs_rq, se, flags);
3708 * end evaluation on encountering a throttled cfs_rq
3710 * note: in the case of encountering a throttled cfs_rq we will
3711 * post the final h_nr_running decrement below.
3713 if (cfs_rq_throttled(cfs_rq))
3715 cfs_rq->h_nr_running--;
3717 /* Don't dequeue parent if it has other entities besides us */
3718 if (cfs_rq->load.weight) {
3720 * Bias pick_next to pick a task from this cfs_rq, as
3721 * p is sleeping when it is within its sched_slice.
3723 if (task_sleep && parent_entity(se))
3724 set_next_buddy(parent_entity(se));
3726 /* avoid re-evaluating load for this entity */
3727 se = parent_entity(se);
3730 flags |= DEQUEUE_SLEEP;
3733 for_each_sched_entity(se) {
3734 cfs_rq = cfs_rq_of(se);
3735 cfs_rq->h_nr_running--;
3737 if (cfs_rq_throttled(cfs_rq))
3740 update_cfs_shares(cfs_rq);
3741 update_entity_load_avg(se, 1);
3746 update_rq_runnable_avg(rq, 1);
3752 /* Used instead of source_load when we know the type == 0 */
3753 static unsigned long weighted_cpuload(const int cpu)
3755 return cpu_rq(cpu)->cfs.runnable_load_avg;
3759 * Return a low guess at the load of a migration-source cpu weighted
3760 * according to the scheduling class and "nice" value.
3762 * We want to under-estimate the load of migration sources, to
3763 * balance conservatively.
3765 static unsigned long source_load(int cpu, int type)
3767 struct rq *rq = cpu_rq(cpu);
3768 unsigned long total = weighted_cpuload(cpu);
3770 if (type == 0 || !sched_feat(LB_BIAS))
3773 return min(rq->cpu_load[type-1], total);
3777 * Return a high guess at the load of a migration-target cpu weighted
3778 * according to the scheduling class and "nice" value.
3780 static unsigned long target_load(int cpu, int type)
3782 struct rq *rq = cpu_rq(cpu);
3783 unsigned long total = weighted_cpuload(cpu);
3785 if (type == 0 || !sched_feat(LB_BIAS))
3788 return max(rq->cpu_load[type-1], total);
3791 static unsigned long power_of(int cpu)
3793 return cpu_rq(cpu)->cpu_power;
3796 static unsigned long cpu_avg_load_per_task(int cpu)
3798 struct rq *rq = cpu_rq(cpu);
3799 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3800 unsigned long load_avg = rq->cfs.runnable_load_avg;
3803 return load_avg / nr_running;
3808 static void record_wakee(struct task_struct *p)
3811 * Rough decay (wiping) for cost saving, don't worry
3812 * about the boundary, really active task won't care
3815 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3816 current->wakee_flips = 0;
3817 current->wakee_flip_decay_ts = jiffies;
3820 if (current->last_wakee != p) {
3821 current->last_wakee = p;
3822 current->wakee_flips++;
3826 static void task_waking_fair(struct task_struct *p)
3828 struct sched_entity *se = &p->se;
3829 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3832 #ifndef CONFIG_64BIT
3833 u64 min_vruntime_copy;
3836 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3838 min_vruntime = cfs_rq->min_vruntime;
3839 } while (min_vruntime != min_vruntime_copy);
3841 min_vruntime = cfs_rq->min_vruntime;
3844 se->vruntime -= min_vruntime;
3848 #ifdef CONFIG_FAIR_GROUP_SCHED
3850 * effective_load() calculates the load change as seen from the root_task_group
3852 * Adding load to a group doesn't make a group heavier, but can cause movement
3853 * of group shares between cpus. Assuming the shares were perfectly aligned one
3854 * can calculate the shift in shares.
3856 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3857 * on this @cpu and results in a total addition (subtraction) of @wg to the
3858 * total group weight.
3860 * Given a runqueue weight distribution (rw_i) we can compute a shares
3861 * distribution (s_i) using:
3863 * s_i = rw_i / \Sum rw_j (1)
3865 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3866 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3867 * shares distribution (s_i):
3869 * rw_i = { 2, 4, 1, 0 }
3870 * s_i = { 2/7, 4/7, 1/7, 0 }
3872 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3873 * task used to run on and the CPU the waker is running on), we need to
3874 * compute the effect of waking a task on either CPU and, in case of a sync
3875 * wakeup, compute the effect of the current task going to sleep.
3877 * So for a change of @wl to the local @cpu with an overall group weight change
3878 * of @wl we can compute the new shares distribution (s'_i) using:
3880 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3882 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3883 * differences in waking a task to CPU 0. The additional task changes the
3884 * weight and shares distributions like:
3886 * rw'_i = { 3, 4, 1, 0 }
3887 * s'_i = { 3/8, 4/8, 1/8, 0 }
3889 * We can then compute the difference in effective weight by using:
3891 * dw_i = S * (s'_i - s_i) (3)
3893 * Where 'S' is the group weight as seen by its parent.
3895 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3896 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3897 * 4/7) times the weight of the group.
3899 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3901 struct sched_entity *se = tg->se[cpu];
3903 if (!tg->parent) /* the trivial, non-cgroup case */
3906 for_each_sched_entity(se) {
3912 * W = @wg + \Sum rw_j
3914 W = wg + calc_tg_weight(tg, se->my_q);
3919 w = se->my_q->load.weight + wl;
3922 * wl = S * s'_i; see (2)
3925 wl = (w * tg->shares) / W;
3930 * Per the above, wl is the new se->load.weight value; since
3931 * those are clipped to [MIN_SHARES, ...) do so now. See
3932 * calc_cfs_shares().
3934 if (wl < MIN_SHARES)
3938 * wl = dw_i = S * (s'_i - s_i); see (3)
3940 wl -= se->load.weight;
3943 * Recursively apply this logic to all parent groups to compute
3944 * the final effective load change on the root group. Since
3945 * only the @tg group gets extra weight, all parent groups can
3946 * only redistribute existing shares. @wl is the shift in shares
3947 * resulting from this level per the above.
3956 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3963 static int wake_wide(struct task_struct *p)
3965 int factor = this_cpu_read(sd_llc_size);
3968 * Yeah, it's the switching-frequency, could means many wakee or
3969 * rapidly switch, use factor here will just help to automatically
3970 * adjust the loose-degree, so bigger node will lead to more pull.
3972 if (p->wakee_flips > factor) {
3974 * wakee is somewhat hot, it needs certain amount of cpu
3975 * resource, so if waker is far more hot, prefer to leave
3978 if (current->wakee_flips > (factor * p->wakee_flips))
3985 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3987 s64 this_load, load;
3988 int idx, this_cpu, prev_cpu;
3989 unsigned long tl_per_task;
3990 struct task_group *tg;
3991 unsigned long weight;
3995 * If we wake multiple tasks be careful to not bounce
3996 * ourselves around too much.
4002 this_cpu = smp_processor_id();
4003 prev_cpu = task_cpu(p);
4004 load = source_load(prev_cpu, idx);
4005 this_load = target_load(this_cpu, idx);
4008 * If sync wakeup then subtract the (maximum possible)
4009 * effect of the currently running task from the load
4010 * of the current CPU:
4013 tg = task_group(current);
4014 weight = current->se.load.weight;
4016 this_load += effective_load(tg, this_cpu, -weight, -weight);
4017 load += effective_load(tg, prev_cpu, 0, -weight);
4021 weight = p->se.load.weight;
4024 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4025 * due to the sync cause above having dropped this_load to 0, we'll
4026 * always have an imbalance, but there's really nothing you can do
4027 * about that, so that's good too.
4029 * Otherwise check if either cpus are near enough in load to allow this
4030 * task to be woken on this_cpu.
4032 if (this_load > 0) {
4033 s64 this_eff_load, prev_eff_load;
4035 this_eff_load = 100;
4036 this_eff_load *= power_of(prev_cpu);
4037 this_eff_load *= this_load +
4038 effective_load(tg, this_cpu, weight, weight);
4040 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4041 prev_eff_load *= power_of(this_cpu);
4042 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4044 balanced = this_eff_load <= prev_eff_load;
4049 * If the currently running task will sleep within
4050 * a reasonable amount of time then attract this newly
4053 if (sync && balanced)
4056 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4057 tl_per_task = cpu_avg_load_per_task(this_cpu);
4060 (this_load <= load &&
4061 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4063 * This domain has SD_WAKE_AFFINE and
4064 * p is cache cold in this domain, and
4065 * there is no bad imbalance.
4067 schedstat_inc(sd, ttwu_move_affine);
4068 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4076 * find_idlest_group finds and returns the least busy CPU group within the
4079 static struct sched_group *
4080 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4081 int this_cpu, int sd_flag)
4083 struct sched_group *idlest = NULL, *group = sd->groups;
4084 unsigned long min_load = ULONG_MAX, this_load = 0;
4085 int load_idx = sd->forkexec_idx;
4086 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4088 if (sd_flag & SD_BALANCE_WAKE)
4089 load_idx = sd->wake_idx;
4092 unsigned long load, avg_load;
4096 /* Skip over this group if it has no CPUs allowed */
4097 if (!cpumask_intersects(sched_group_cpus(group),
4098 tsk_cpus_allowed(p)))
4101 local_group = cpumask_test_cpu(this_cpu,
4102 sched_group_cpus(group));
4104 /* Tally up the load of all CPUs in the group */
4107 for_each_cpu(i, sched_group_cpus(group)) {
4108 /* Bias balancing toward cpus of our domain */
4110 load = source_load(i, load_idx);
4112 load = target_load(i, load_idx);
4117 /* Adjust by relative CPU power of the group */
4118 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4121 this_load = avg_load;
4122 } else if (avg_load < min_load) {
4123 min_load = avg_load;
4126 } while (group = group->next, group != sd->groups);
4128 if (!idlest || 100*this_load < imbalance*min_load)
4134 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4137 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4139 unsigned long load, min_load = ULONG_MAX;
4143 /* Traverse only the allowed CPUs */
4144 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4145 load = weighted_cpuload(i);
4147 if (load < min_load || (load == min_load && i == this_cpu)) {
4157 * Try and locate an idle CPU in the sched_domain.
4159 static int select_idle_sibling(struct task_struct *p, int target)
4161 struct sched_domain *sd;
4162 struct sched_group *sg;
4163 int i = task_cpu(p);
4165 if (idle_cpu(target))
4169 * If the prevous cpu is cache affine and idle, don't be stupid.
4171 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4175 * Otherwise, iterate the domains and find an elegible idle cpu.
4177 sd = rcu_dereference(per_cpu(sd_llc, target));
4178 for_each_lower_domain(sd) {
4181 if (!cpumask_intersects(sched_group_cpus(sg),
4182 tsk_cpus_allowed(p)))
4185 for_each_cpu(i, sched_group_cpus(sg)) {
4186 if (i == target || !idle_cpu(i))
4190 target = cpumask_first_and(sched_group_cpus(sg),
4191 tsk_cpus_allowed(p));
4195 } while (sg != sd->groups);
4202 * sched_balance_self: balance the current task (running on cpu) in domains
4203 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4206 * Balance, ie. select the least loaded group.
4208 * Returns the target CPU number, or the same CPU if no balancing is needed.
4210 * preempt must be disabled.
4213 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4215 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4216 int cpu = smp_processor_id();
4218 int want_affine = 0;
4219 int sync = wake_flags & WF_SYNC;
4221 if (p->nr_cpus_allowed == 1)
4224 if (sd_flag & SD_BALANCE_WAKE) {
4225 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4231 for_each_domain(cpu, tmp) {
4232 if (!(tmp->flags & SD_LOAD_BALANCE))
4236 * If both cpu and prev_cpu are part of this domain,
4237 * cpu is a valid SD_WAKE_AFFINE target.
4239 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4240 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4245 if (tmp->flags & sd_flag)
4250 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4253 new_cpu = select_idle_sibling(p, prev_cpu);
4258 struct sched_group *group;
4261 if (!(sd->flags & sd_flag)) {
4266 group = find_idlest_group(sd, p, cpu, sd_flag);
4272 new_cpu = find_idlest_cpu(group, p, cpu);
4273 if (new_cpu == -1 || new_cpu == cpu) {
4274 /* Now try balancing at a lower domain level of cpu */
4279 /* Now try balancing at a lower domain level of new_cpu */
4281 weight = sd->span_weight;
4283 for_each_domain(cpu, tmp) {
4284 if (weight <= tmp->span_weight)
4286 if (tmp->flags & sd_flag)
4289 /* while loop will break here if sd == NULL */
4298 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4299 * cfs_rq_of(p) references at time of call are still valid and identify the
4300 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4301 * other assumptions, including the state of rq->lock, should be made.
4304 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4306 struct sched_entity *se = &p->se;
4307 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4310 * Load tracking: accumulate removed load so that it can be processed
4311 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4312 * to blocked load iff they have a positive decay-count. It can never
4313 * be negative here since on-rq tasks have decay-count == 0.
4315 if (se->avg.decay_count) {
4316 se->avg.decay_count = -__synchronize_entity_decay(se);
4317 atomic_long_add(se->avg.load_avg_contrib,
4318 &cfs_rq->removed_load);
4321 #endif /* CONFIG_SMP */
4323 static unsigned long
4324 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4326 unsigned long gran = sysctl_sched_wakeup_granularity;
4329 * Since its curr running now, convert the gran from real-time
4330 * to virtual-time in his units.
4332 * By using 'se' instead of 'curr' we penalize light tasks, so
4333 * they get preempted easier. That is, if 'se' < 'curr' then
4334 * the resulting gran will be larger, therefore penalizing the
4335 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4336 * be smaller, again penalizing the lighter task.
4338 * This is especially important for buddies when the leftmost
4339 * task is higher priority than the buddy.
4341 return calc_delta_fair(gran, se);
4345 * Should 'se' preempt 'curr'.
4359 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4361 s64 gran, vdiff = curr->vruntime - se->vruntime;
4366 gran = wakeup_gran(curr, se);
4373 static void set_last_buddy(struct sched_entity *se)
4375 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4378 for_each_sched_entity(se)
4379 cfs_rq_of(se)->last = se;
4382 static void set_next_buddy(struct sched_entity *se)
4384 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4387 for_each_sched_entity(se)
4388 cfs_rq_of(se)->next = se;
4391 static void set_skip_buddy(struct sched_entity *se)
4393 for_each_sched_entity(se)
4394 cfs_rq_of(se)->skip = se;
4398 * Preempt the current task with a newly woken task if needed:
4400 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4402 struct task_struct *curr = rq->curr;
4403 struct sched_entity *se = &curr->se, *pse = &p->se;
4404 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4405 int scale = cfs_rq->nr_running >= sched_nr_latency;
4406 int next_buddy_marked = 0;
4408 if (unlikely(se == pse))
4412 * This is possible from callers such as move_task(), in which we
4413 * unconditionally check_prempt_curr() after an enqueue (which may have
4414 * lead to a throttle). This both saves work and prevents false
4415 * next-buddy nomination below.
4417 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4420 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4421 set_next_buddy(pse);
4422 next_buddy_marked = 1;
4426 * We can come here with TIF_NEED_RESCHED already set from new task
4429 * Note: this also catches the edge-case of curr being in a throttled
4430 * group (e.g. via set_curr_task), since update_curr() (in the
4431 * enqueue of curr) will have resulted in resched being set. This
4432 * prevents us from potentially nominating it as a false LAST_BUDDY
4435 if (test_tsk_need_resched(curr))
4438 /* Idle tasks are by definition preempted by non-idle tasks. */
4439 if (unlikely(curr->policy == SCHED_IDLE) &&
4440 likely(p->policy != SCHED_IDLE))
4444 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4445 * is driven by the tick):
4447 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4450 find_matching_se(&se, &pse);
4451 update_curr(cfs_rq_of(se));
4453 if (wakeup_preempt_entity(se, pse) == 1) {
4455 * Bias pick_next to pick the sched entity that is
4456 * triggering this preemption.
4458 if (!next_buddy_marked)
4459 set_next_buddy(pse);
4468 * Only set the backward buddy when the current task is still
4469 * on the rq. This can happen when a wakeup gets interleaved
4470 * with schedule on the ->pre_schedule() or idle_balance()
4471 * point, either of which can * drop the rq lock.
4473 * Also, during early boot the idle thread is in the fair class,
4474 * for obvious reasons its a bad idea to schedule back to it.
4476 if (unlikely(!se->on_rq || curr == rq->idle))
4479 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4483 static struct task_struct *pick_next_task_fair(struct rq *rq)
4485 struct task_struct *p;
4486 struct cfs_rq *cfs_rq = &rq->cfs;
4487 struct sched_entity *se;
4489 if (!cfs_rq->nr_running)
4493 se = pick_next_entity(cfs_rq);
4494 set_next_entity(cfs_rq, se);
4495 cfs_rq = group_cfs_rq(se);
4499 if (hrtick_enabled(rq))
4500 hrtick_start_fair(rq, p);
4506 * Account for a descheduled task:
4508 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4510 struct sched_entity *se = &prev->se;
4511 struct cfs_rq *cfs_rq;
4513 for_each_sched_entity(se) {
4514 cfs_rq = cfs_rq_of(se);
4515 put_prev_entity(cfs_rq, se);
4520 * sched_yield() is very simple
4522 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4524 static void yield_task_fair(struct rq *rq)
4526 struct task_struct *curr = rq->curr;
4527 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4528 struct sched_entity *se = &curr->se;
4531 * Are we the only task in the tree?
4533 if (unlikely(rq->nr_running == 1))
4536 clear_buddies(cfs_rq, se);
4538 if (curr->policy != SCHED_BATCH) {
4539 update_rq_clock(rq);
4541 * Update run-time statistics of the 'current'.
4543 update_curr(cfs_rq);
4545 * Tell update_rq_clock() that we've just updated,
4546 * so we don't do microscopic update in schedule()
4547 * and double the fastpath cost.
4549 rq->skip_clock_update = 1;
4555 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4557 struct sched_entity *se = &p->se;
4559 /* throttled hierarchies are not runnable */
4560 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4563 /* Tell the scheduler that we'd really like pse to run next. */
4566 yield_task_fair(rq);
4572 /**************************************************
4573 * Fair scheduling class load-balancing methods.
4577 * The purpose of load-balancing is to achieve the same basic fairness the
4578 * per-cpu scheduler provides, namely provide a proportional amount of compute
4579 * time to each task. This is expressed in the following equation:
4581 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4583 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4584 * W_i,0 is defined as:
4586 * W_i,0 = \Sum_j w_i,j (2)
4588 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4589 * is derived from the nice value as per prio_to_weight[].
4591 * The weight average is an exponential decay average of the instantaneous
4594 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4596 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4597 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4598 * can also include other factors [XXX].
4600 * To achieve this balance we define a measure of imbalance which follows
4601 * directly from (1):
4603 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4605 * We them move tasks around to minimize the imbalance. In the continuous
4606 * function space it is obvious this converges, in the discrete case we get
4607 * a few fun cases generally called infeasible weight scenarios.
4610 * - infeasible weights;
4611 * - local vs global optima in the discrete case. ]
4616 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4617 * for all i,j solution, we create a tree of cpus that follows the hardware
4618 * topology where each level pairs two lower groups (or better). This results
4619 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4620 * tree to only the first of the previous level and we decrease the frequency
4621 * of load-balance at each level inv. proportional to the number of cpus in
4627 * \Sum { --- * --- * 2^i } = O(n) (5)
4629 * `- size of each group
4630 * | | `- number of cpus doing load-balance
4632 * `- sum over all levels
4634 * Coupled with a limit on how many tasks we can migrate every balance pass,
4635 * this makes (5) the runtime complexity of the balancer.
4637 * An important property here is that each CPU is still (indirectly) connected
4638 * to every other cpu in at most O(log n) steps:
4640 * The adjacency matrix of the resulting graph is given by:
4643 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4646 * And you'll find that:
4648 * A^(log_2 n)_i,j != 0 for all i,j (7)
4650 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4651 * The task movement gives a factor of O(m), giving a convergence complexity
4654 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4659 * In order to avoid CPUs going idle while there's still work to do, new idle
4660 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4661 * tree itself instead of relying on other CPUs to bring it work.
4663 * This adds some complexity to both (5) and (8) but it reduces the total idle
4671 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4674 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4679 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4681 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4683 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4686 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4687 * rewrite all of this once again.]
4690 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4692 enum fbq_type { regular, remote, all };
4694 #define LBF_ALL_PINNED 0x01
4695 #define LBF_NEED_BREAK 0x02
4696 #define LBF_DST_PINNED 0x04
4697 #define LBF_SOME_PINNED 0x08
4700 struct sched_domain *sd;
4708 struct cpumask *dst_grpmask;
4710 enum cpu_idle_type idle;
4712 /* The set of CPUs under consideration for load-balancing */
4713 struct cpumask *cpus;
4718 unsigned int loop_break;
4719 unsigned int loop_max;
4721 enum fbq_type fbq_type;
4725 * move_task - move a task from one runqueue to another runqueue.
4726 * Both runqueues must be locked.
4728 static void move_task(struct task_struct *p, struct lb_env *env)
4730 deactivate_task(env->src_rq, p, 0);
4731 set_task_cpu(p, env->dst_cpu);
4732 activate_task(env->dst_rq, p, 0);
4733 check_preempt_curr(env->dst_rq, p, 0);
4737 * Is this task likely cache-hot:
4740 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4744 if (p->sched_class != &fair_sched_class)
4747 if (unlikely(p->policy == SCHED_IDLE))
4751 * Buddy candidates are cache hot:
4753 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4754 (&p->se == cfs_rq_of(&p->se)->next ||
4755 &p->se == cfs_rq_of(&p->se)->last))
4758 if (sysctl_sched_migration_cost == -1)
4760 if (sysctl_sched_migration_cost == 0)
4763 delta = now - p->se.exec_start;
4765 return delta < (s64)sysctl_sched_migration_cost;
4768 #ifdef CONFIG_NUMA_BALANCING
4769 /* Returns true if the destination node has incurred more faults */
4770 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4772 int src_nid, dst_nid;
4774 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4775 !(env->sd->flags & SD_NUMA)) {
4779 src_nid = cpu_to_node(env->src_cpu);
4780 dst_nid = cpu_to_node(env->dst_cpu);
4782 if (src_nid == dst_nid)
4785 /* Always encourage migration to the preferred node. */
4786 if (dst_nid == p->numa_preferred_nid)
4789 /* If both task and group weight improve, this move is a winner. */
4790 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4791 group_weight(p, dst_nid) > group_weight(p, src_nid))
4798 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4800 int src_nid, dst_nid;
4802 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4805 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4808 src_nid = cpu_to_node(env->src_cpu);
4809 dst_nid = cpu_to_node(env->dst_cpu);
4811 if (src_nid == dst_nid)
4814 /* Migrating away from the preferred node is always bad. */
4815 if (src_nid == p->numa_preferred_nid)
4818 /* If either task or group weight get worse, don't do it. */
4819 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
4820 group_weight(p, dst_nid) < group_weight(p, src_nid))
4827 static inline bool migrate_improves_locality(struct task_struct *p,
4833 static inline bool migrate_degrades_locality(struct task_struct *p,
4841 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4844 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4846 int tsk_cache_hot = 0;
4848 * We do not migrate tasks that are:
4849 * 1) throttled_lb_pair, or
4850 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4851 * 3) running (obviously), or
4852 * 4) are cache-hot on their current CPU.
4854 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4857 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4860 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4862 env->flags |= LBF_SOME_PINNED;
4865 * Remember if this task can be migrated to any other cpu in
4866 * our sched_group. We may want to revisit it if we couldn't
4867 * meet load balance goals by pulling other tasks on src_cpu.
4869 * Also avoid computing new_dst_cpu if we have already computed
4870 * one in current iteration.
4872 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4875 /* Prevent to re-select dst_cpu via env's cpus */
4876 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4877 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4878 env->flags |= LBF_DST_PINNED;
4879 env->new_dst_cpu = cpu;
4887 /* Record that we found atleast one task that could run on dst_cpu */
4888 env->flags &= ~LBF_ALL_PINNED;
4890 if (task_running(env->src_rq, p)) {
4891 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4896 * Aggressive migration if:
4897 * 1) destination numa is preferred
4898 * 2) task is cache cold, or
4899 * 3) too many balance attempts have failed.
4901 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4903 tsk_cache_hot = migrate_degrades_locality(p, env);
4905 if (migrate_improves_locality(p, env)) {
4906 #ifdef CONFIG_SCHEDSTATS
4907 if (tsk_cache_hot) {
4908 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4909 schedstat_inc(p, se.statistics.nr_forced_migrations);
4915 if (!tsk_cache_hot ||
4916 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4918 if (tsk_cache_hot) {
4919 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4920 schedstat_inc(p, se.statistics.nr_forced_migrations);
4926 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4931 * move_one_task tries to move exactly one task from busiest to this_rq, as
4932 * part of active balancing operations within "domain".
4933 * Returns 1 if successful and 0 otherwise.
4935 * Called with both runqueues locked.
4937 static int move_one_task(struct lb_env *env)
4939 struct task_struct *p, *n;
4941 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4942 if (!can_migrate_task(p, env))
4947 * Right now, this is only the second place move_task()
4948 * is called, so we can safely collect move_task()
4949 * stats here rather than inside move_task().
4951 schedstat_inc(env->sd, lb_gained[env->idle]);
4957 static const unsigned int sched_nr_migrate_break = 32;
4960 * move_tasks tries to move up to imbalance weighted load from busiest to
4961 * this_rq, as part of a balancing operation within domain "sd".
4962 * Returns 1 if successful and 0 otherwise.
4964 * Called with both runqueues locked.
4966 static int move_tasks(struct lb_env *env)
4968 struct list_head *tasks = &env->src_rq->cfs_tasks;
4969 struct task_struct *p;
4973 if (env->imbalance <= 0)
4976 while (!list_empty(tasks)) {
4977 p = list_first_entry(tasks, struct task_struct, se.group_node);
4980 /* We've more or less seen every task there is, call it quits */
4981 if (env->loop > env->loop_max)
4984 /* take a breather every nr_migrate tasks */
4985 if (env->loop > env->loop_break) {
4986 env->loop_break += sched_nr_migrate_break;
4987 env->flags |= LBF_NEED_BREAK;
4991 if (!can_migrate_task(p, env))
4994 load = task_h_load(p);
4996 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4999 if ((load / 2) > env->imbalance)
5004 env->imbalance -= load;
5006 #ifdef CONFIG_PREEMPT
5008 * NEWIDLE balancing is a source of latency, so preemptible
5009 * kernels will stop after the first task is pulled to minimize
5010 * the critical section.
5012 if (env->idle == CPU_NEWLY_IDLE)
5017 * We only want to steal up to the prescribed amount of
5020 if (env->imbalance <= 0)
5025 list_move_tail(&p->se.group_node, tasks);
5029 * Right now, this is one of only two places move_task() is called,
5030 * so we can safely collect move_task() stats here rather than
5031 * inside move_task().
5033 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5038 #ifdef CONFIG_FAIR_GROUP_SCHED
5040 * update tg->load_weight by folding this cpu's load_avg
5042 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5044 struct sched_entity *se = tg->se[cpu];
5045 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5047 /* throttled entities do not contribute to load */
5048 if (throttled_hierarchy(cfs_rq))
5051 update_cfs_rq_blocked_load(cfs_rq, 1);
5054 update_entity_load_avg(se, 1);
5056 * We pivot on our runnable average having decayed to zero for
5057 * list removal. This generally implies that all our children
5058 * have also been removed (modulo rounding error or bandwidth
5059 * control); however, such cases are rare and we can fix these
5062 * TODO: fix up out-of-order children on enqueue.
5064 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5065 list_del_leaf_cfs_rq(cfs_rq);
5067 struct rq *rq = rq_of(cfs_rq);
5068 update_rq_runnable_avg(rq, rq->nr_running);
5072 static void update_blocked_averages(int cpu)
5074 struct rq *rq = cpu_rq(cpu);
5075 struct cfs_rq *cfs_rq;
5076 unsigned long flags;
5078 raw_spin_lock_irqsave(&rq->lock, flags);
5079 update_rq_clock(rq);
5081 * Iterates the task_group tree in a bottom up fashion, see
5082 * list_add_leaf_cfs_rq() for details.
5084 for_each_leaf_cfs_rq(rq, cfs_rq) {
5086 * Note: We may want to consider periodically releasing
5087 * rq->lock about these updates so that creating many task
5088 * groups does not result in continually extending hold time.
5090 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5093 raw_spin_unlock_irqrestore(&rq->lock, flags);
5097 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5098 * This needs to be done in a top-down fashion because the load of a child
5099 * group is a fraction of its parents load.
5101 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5103 struct rq *rq = rq_of(cfs_rq);
5104 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5105 unsigned long now = jiffies;
5108 if (cfs_rq->last_h_load_update == now)
5111 cfs_rq->h_load_next = NULL;
5112 for_each_sched_entity(se) {
5113 cfs_rq = cfs_rq_of(se);
5114 cfs_rq->h_load_next = se;
5115 if (cfs_rq->last_h_load_update == now)
5120 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5121 cfs_rq->last_h_load_update = now;
5124 while ((se = cfs_rq->h_load_next) != NULL) {
5125 load = cfs_rq->h_load;
5126 load = div64_ul(load * se->avg.load_avg_contrib,
5127 cfs_rq->runnable_load_avg + 1);
5128 cfs_rq = group_cfs_rq(se);
5129 cfs_rq->h_load = load;
5130 cfs_rq->last_h_load_update = now;
5134 static unsigned long task_h_load(struct task_struct *p)
5136 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5138 update_cfs_rq_h_load(cfs_rq);
5139 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5140 cfs_rq->runnable_load_avg + 1);
5143 static inline void update_blocked_averages(int cpu)
5147 static unsigned long task_h_load(struct task_struct *p)
5149 return p->se.avg.load_avg_contrib;
5153 /********** Helpers for find_busiest_group ************************/
5155 * sg_lb_stats - stats of a sched_group required for load_balancing
5157 struct sg_lb_stats {
5158 unsigned long avg_load; /*Avg load across the CPUs of the group */
5159 unsigned long group_load; /* Total load over the CPUs of the group */
5160 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5161 unsigned long load_per_task;
5162 unsigned long group_power;
5163 unsigned int sum_nr_running; /* Nr tasks running in the group */
5164 unsigned int group_capacity;
5165 unsigned int idle_cpus;
5166 unsigned int group_weight;
5167 int group_imb; /* Is there an imbalance in the group ? */
5168 int group_has_capacity; /* Is there extra capacity in the group? */
5169 #ifdef CONFIG_NUMA_BALANCING
5170 unsigned int nr_numa_running;
5171 unsigned int nr_preferred_running;
5176 * sd_lb_stats - Structure to store the statistics of a sched_domain
5177 * during load balancing.
5179 struct sd_lb_stats {
5180 struct sched_group *busiest; /* Busiest group in this sd */
5181 struct sched_group *local; /* Local group in this sd */
5182 unsigned long total_load; /* Total load of all groups in sd */
5183 unsigned long total_pwr; /* Total power of all groups in sd */
5184 unsigned long avg_load; /* Average load across all groups in sd */
5186 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5187 struct sg_lb_stats local_stat; /* Statistics of the local group */
5190 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5193 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5194 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5195 * We must however clear busiest_stat::avg_load because
5196 * update_sd_pick_busiest() reads this before assignment.
5198 *sds = (struct sd_lb_stats){
5210 * get_sd_load_idx - Obtain the load index for a given sched domain.
5211 * @sd: The sched_domain whose load_idx is to be obtained.
5212 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5214 * Return: The load index.
5216 static inline int get_sd_load_idx(struct sched_domain *sd,
5217 enum cpu_idle_type idle)
5223 load_idx = sd->busy_idx;
5226 case CPU_NEWLY_IDLE:
5227 load_idx = sd->newidle_idx;
5230 load_idx = sd->idle_idx;
5237 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5239 return SCHED_POWER_SCALE;
5242 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5244 return default_scale_freq_power(sd, cpu);
5247 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5249 unsigned long weight = sd->span_weight;
5250 unsigned long smt_gain = sd->smt_gain;
5257 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5259 return default_scale_smt_power(sd, cpu);
5262 static unsigned long scale_rt_power(int cpu)
5264 struct rq *rq = cpu_rq(cpu);
5265 u64 total, available, age_stamp, avg;
5268 * Since we're reading these variables without serialization make sure
5269 * we read them once before doing sanity checks on them.
5271 age_stamp = ACCESS_ONCE(rq->age_stamp);
5272 avg = ACCESS_ONCE(rq->rt_avg);
5274 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5276 if (unlikely(total < avg)) {
5277 /* Ensures that power won't end up being negative */
5280 available = total - avg;
5283 if (unlikely((s64)total < SCHED_POWER_SCALE))
5284 total = SCHED_POWER_SCALE;
5286 total >>= SCHED_POWER_SHIFT;
5288 return div_u64(available, total);
5291 static void update_cpu_power(struct sched_domain *sd, int cpu)
5293 unsigned long weight = sd->span_weight;
5294 unsigned long power = SCHED_POWER_SCALE;
5295 struct sched_group *sdg = sd->groups;
5297 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5298 if (sched_feat(ARCH_POWER))
5299 power *= arch_scale_smt_power(sd, cpu);
5301 power *= default_scale_smt_power(sd, cpu);
5303 power >>= SCHED_POWER_SHIFT;
5306 sdg->sgp->power_orig = power;
5308 if (sched_feat(ARCH_POWER))
5309 power *= arch_scale_freq_power(sd, cpu);
5311 power *= default_scale_freq_power(sd, cpu);
5313 power >>= SCHED_POWER_SHIFT;
5315 power *= scale_rt_power(cpu);
5316 power >>= SCHED_POWER_SHIFT;
5321 cpu_rq(cpu)->cpu_power = power;
5322 sdg->sgp->power = power;
5325 void update_group_power(struct sched_domain *sd, int cpu)
5327 struct sched_domain *child = sd->child;
5328 struct sched_group *group, *sdg = sd->groups;
5329 unsigned long power, power_orig;
5330 unsigned long interval;
5332 interval = msecs_to_jiffies(sd->balance_interval);
5333 interval = clamp(interval, 1UL, max_load_balance_interval);
5334 sdg->sgp->next_update = jiffies + interval;
5337 update_cpu_power(sd, cpu);
5341 power_orig = power = 0;
5343 if (child->flags & SD_OVERLAP) {
5345 * SD_OVERLAP domains cannot assume that child groups
5346 * span the current group.
5349 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5350 struct sched_group_power *sgp;
5351 struct rq *rq = cpu_rq(cpu);
5354 * build_sched_domains() -> init_sched_groups_power()
5355 * gets here before we've attached the domains to the
5358 * Use power_of(), which is set irrespective of domains
5359 * in update_cpu_power().
5361 * This avoids power/power_orig from being 0 and
5362 * causing divide-by-zero issues on boot.
5364 * Runtime updates will correct power_orig.
5366 if (unlikely(!rq->sd)) {
5367 power_orig += power_of(cpu);
5368 power += power_of(cpu);
5372 sgp = rq->sd->groups->sgp;
5373 power_orig += sgp->power_orig;
5374 power += sgp->power;
5378 * !SD_OVERLAP domains can assume that child groups
5379 * span the current group.
5382 group = child->groups;
5384 power_orig += group->sgp->power_orig;
5385 power += group->sgp->power;
5386 group = group->next;
5387 } while (group != child->groups);
5390 sdg->sgp->power_orig = power_orig;
5391 sdg->sgp->power = power;
5395 * Try and fix up capacity for tiny siblings, this is needed when
5396 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5397 * which on its own isn't powerful enough.
5399 * See update_sd_pick_busiest() and check_asym_packing().
5402 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5405 * Only siblings can have significantly less than SCHED_POWER_SCALE
5407 if (!(sd->flags & SD_SHARE_CPUPOWER))
5411 * If ~90% of the cpu_power is still there, we're good.
5413 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5420 * Group imbalance indicates (and tries to solve) the problem where balancing
5421 * groups is inadequate due to tsk_cpus_allowed() constraints.
5423 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5424 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5427 * { 0 1 2 3 } { 4 5 6 7 }
5430 * If we were to balance group-wise we'd place two tasks in the first group and
5431 * two tasks in the second group. Clearly this is undesired as it will overload
5432 * cpu 3 and leave one of the cpus in the second group unused.
5434 * The current solution to this issue is detecting the skew in the first group
5435 * by noticing the lower domain failed to reach balance and had difficulty
5436 * moving tasks due to affinity constraints.
5438 * When this is so detected; this group becomes a candidate for busiest; see
5439 * update_sd_pick_busiest(). And calculate_imbalance() and
5440 * find_busiest_group() avoid some of the usual balance conditions to allow it
5441 * to create an effective group imbalance.
5443 * This is a somewhat tricky proposition since the next run might not find the
5444 * group imbalance and decide the groups need to be balanced again. A most
5445 * subtle and fragile situation.
5448 static inline int sg_imbalanced(struct sched_group *group)
5450 return group->sgp->imbalance;
5454 * Compute the group capacity.
5456 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5457 * first dividing out the smt factor and computing the actual number of cores
5458 * and limit power unit capacity with that.
5460 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5462 unsigned int capacity, smt, cpus;
5463 unsigned int power, power_orig;
5465 power = group->sgp->power;
5466 power_orig = group->sgp->power_orig;
5467 cpus = group->group_weight;
5469 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5470 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5471 capacity = cpus / smt; /* cores */
5473 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5475 capacity = fix_small_capacity(env->sd, group);
5481 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5482 * @env: The load balancing environment.
5483 * @group: sched_group whose statistics are to be updated.
5484 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5485 * @local_group: Does group contain this_cpu.
5486 * @sgs: variable to hold the statistics for this group.
5488 static inline void update_sg_lb_stats(struct lb_env *env,
5489 struct sched_group *group, int load_idx,
5490 int local_group, struct sg_lb_stats *sgs)
5495 memset(sgs, 0, sizeof(*sgs));
5497 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5498 struct rq *rq = cpu_rq(i);
5500 /* Bias balancing toward cpus of our domain */
5502 load = target_load(i, load_idx);
5504 load = source_load(i, load_idx);
5506 sgs->group_load += load;
5507 sgs->sum_nr_running += rq->nr_running;
5508 #ifdef CONFIG_NUMA_BALANCING
5509 sgs->nr_numa_running += rq->nr_numa_running;
5510 sgs->nr_preferred_running += rq->nr_preferred_running;
5512 sgs->sum_weighted_load += weighted_cpuload(i);
5517 /* Adjust by relative CPU power of the group */
5518 sgs->group_power = group->sgp->power;
5519 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5521 if (sgs->sum_nr_running)
5522 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5524 sgs->group_weight = group->group_weight;
5526 sgs->group_imb = sg_imbalanced(group);
5527 sgs->group_capacity = sg_capacity(env, group);
5529 if (sgs->group_capacity > sgs->sum_nr_running)
5530 sgs->group_has_capacity = 1;
5534 * update_sd_pick_busiest - return 1 on busiest group
5535 * @env: The load balancing environment.
5536 * @sds: sched_domain statistics
5537 * @sg: sched_group candidate to be checked for being the busiest
5538 * @sgs: sched_group statistics
5540 * Determine if @sg is a busier group than the previously selected
5543 * Return: %true if @sg is a busier group than the previously selected
5544 * busiest group. %false otherwise.
5546 static bool update_sd_pick_busiest(struct lb_env *env,
5547 struct sd_lb_stats *sds,
5548 struct sched_group *sg,
5549 struct sg_lb_stats *sgs)
5551 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5554 if (sgs->sum_nr_running > sgs->group_capacity)
5561 * ASYM_PACKING needs to move all the work to the lowest
5562 * numbered CPUs in the group, therefore mark all groups
5563 * higher than ourself as busy.
5565 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5566 env->dst_cpu < group_first_cpu(sg)) {
5570 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5577 #ifdef CONFIG_NUMA_BALANCING
5578 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5580 if (sgs->sum_nr_running > sgs->nr_numa_running)
5582 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5587 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5589 if (rq->nr_running > rq->nr_numa_running)
5591 if (rq->nr_running > rq->nr_preferred_running)
5596 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5601 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5605 #endif /* CONFIG_NUMA_BALANCING */
5608 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5609 * @env: The load balancing environment.
5610 * @sds: variable to hold the statistics for this sched_domain.
5612 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5614 struct sched_domain *child = env->sd->child;
5615 struct sched_group *sg = env->sd->groups;
5616 struct sg_lb_stats tmp_sgs;
5617 int load_idx, prefer_sibling = 0;
5619 if (child && child->flags & SD_PREFER_SIBLING)
5622 load_idx = get_sd_load_idx(env->sd, env->idle);
5625 struct sg_lb_stats *sgs = &tmp_sgs;
5628 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5631 sgs = &sds->local_stat;
5633 if (env->idle != CPU_NEWLY_IDLE ||
5634 time_after_eq(jiffies, sg->sgp->next_update))
5635 update_group_power(env->sd, env->dst_cpu);
5638 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5644 * In case the child domain prefers tasks go to siblings
5645 * first, lower the sg capacity to one so that we'll try
5646 * and move all the excess tasks away. We lower the capacity
5647 * of a group only if the local group has the capacity to fit
5648 * these excess tasks, i.e. nr_running < group_capacity. The
5649 * extra check prevents the case where you always pull from the
5650 * heaviest group when it is already under-utilized (possible
5651 * with a large weight task outweighs the tasks on the system).
5653 if (prefer_sibling && sds->local &&
5654 sds->local_stat.group_has_capacity)
5655 sgs->group_capacity = min(sgs->group_capacity, 1U);
5657 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5659 sds->busiest_stat = *sgs;
5663 /* Now, start updating sd_lb_stats */
5664 sds->total_load += sgs->group_load;
5665 sds->total_pwr += sgs->group_power;
5668 } while (sg != env->sd->groups);
5670 if (env->sd->flags & SD_NUMA)
5671 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5675 * check_asym_packing - Check to see if the group is packed into the
5678 * This is primarily intended to used at the sibling level. Some
5679 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5680 * case of POWER7, it can move to lower SMT modes only when higher
5681 * threads are idle. When in lower SMT modes, the threads will
5682 * perform better since they share less core resources. Hence when we
5683 * have idle threads, we want them to be the higher ones.
5685 * This packing function is run on idle threads. It checks to see if
5686 * the busiest CPU in this domain (core in the P7 case) has a higher
5687 * CPU number than the packing function is being run on. Here we are
5688 * assuming lower CPU number will be equivalent to lower a SMT thread
5691 * Return: 1 when packing is required and a task should be moved to
5692 * this CPU. The amount of the imbalance is returned in *imbalance.
5694 * @env: The load balancing environment.
5695 * @sds: Statistics of the sched_domain which is to be packed
5697 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5701 if (!(env->sd->flags & SD_ASYM_PACKING))
5707 busiest_cpu = group_first_cpu(sds->busiest);
5708 if (env->dst_cpu > busiest_cpu)
5711 env->imbalance = DIV_ROUND_CLOSEST(
5712 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5719 * fix_small_imbalance - Calculate the minor imbalance that exists
5720 * amongst the groups of a sched_domain, during
5722 * @env: The load balancing environment.
5723 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5726 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5728 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5729 unsigned int imbn = 2;
5730 unsigned long scaled_busy_load_per_task;
5731 struct sg_lb_stats *local, *busiest;
5733 local = &sds->local_stat;
5734 busiest = &sds->busiest_stat;
5736 if (!local->sum_nr_running)
5737 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5738 else if (busiest->load_per_task > local->load_per_task)
5741 scaled_busy_load_per_task =
5742 (busiest->load_per_task * SCHED_POWER_SCALE) /
5743 busiest->group_power;
5745 if (busiest->avg_load + scaled_busy_load_per_task >=
5746 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5747 env->imbalance = busiest->load_per_task;
5752 * OK, we don't have enough imbalance to justify moving tasks,
5753 * however we may be able to increase total CPU power used by
5757 pwr_now += busiest->group_power *
5758 min(busiest->load_per_task, busiest->avg_load);
5759 pwr_now += local->group_power *
5760 min(local->load_per_task, local->avg_load);
5761 pwr_now /= SCHED_POWER_SCALE;
5763 /* Amount of load we'd subtract */
5764 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5765 busiest->group_power;
5766 if (busiest->avg_load > tmp) {
5767 pwr_move += busiest->group_power *
5768 min(busiest->load_per_task,
5769 busiest->avg_load - tmp);
5772 /* Amount of load we'd add */
5773 if (busiest->avg_load * busiest->group_power <
5774 busiest->load_per_task * SCHED_POWER_SCALE) {
5775 tmp = (busiest->avg_load * busiest->group_power) /
5778 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5781 pwr_move += local->group_power *
5782 min(local->load_per_task, local->avg_load + tmp);
5783 pwr_move /= SCHED_POWER_SCALE;
5785 /* Move if we gain throughput */
5786 if (pwr_move > pwr_now)
5787 env->imbalance = busiest->load_per_task;
5791 * calculate_imbalance - Calculate the amount of imbalance present within the
5792 * groups of a given sched_domain during load balance.
5793 * @env: load balance environment
5794 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5796 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5798 unsigned long max_pull, load_above_capacity = ~0UL;
5799 struct sg_lb_stats *local, *busiest;
5801 local = &sds->local_stat;
5802 busiest = &sds->busiest_stat;
5804 if (busiest->group_imb) {
5806 * In the group_imb case we cannot rely on group-wide averages
5807 * to ensure cpu-load equilibrium, look at wider averages. XXX
5809 busiest->load_per_task =
5810 min(busiest->load_per_task, sds->avg_load);
5814 * In the presence of smp nice balancing, certain scenarios can have
5815 * max load less than avg load(as we skip the groups at or below
5816 * its cpu_power, while calculating max_load..)
5818 if (busiest->avg_load <= sds->avg_load ||
5819 local->avg_load >= sds->avg_load) {
5821 return fix_small_imbalance(env, sds);
5824 if (!busiest->group_imb) {
5826 * Don't want to pull so many tasks that a group would go idle.
5827 * Except of course for the group_imb case, since then we might
5828 * have to drop below capacity to reach cpu-load equilibrium.
5830 load_above_capacity =
5831 (busiest->sum_nr_running - busiest->group_capacity);
5833 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5834 load_above_capacity /= busiest->group_power;
5838 * We're trying to get all the cpus to the average_load, so we don't
5839 * want to push ourselves above the average load, nor do we wish to
5840 * reduce the max loaded cpu below the average load. At the same time,
5841 * we also don't want to reduce the group load below the group capacity
5842 * (so that we can implement power-savings policies etc). Thus we look
5843 * for the minimum possible imbalance.
5845 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5847 /* How much load to actually move to equalise the imbalance */
5848 env->imbalance = min(
5849 max_pull * busiest->group_power,
5850 (sds->avg_load - local->avg_load) * local->group_power
5851 ) / SCHED_POWER_SCALE;
5854 * if *imbalance is less than the average load per runnable task
5855 * there is no guarantee that any tasks will be moved so we'll have
5856 * a think about bumping its value to force at least one task to be
5859 if (env->imbalance < busiest->load_per_task)
5860 return fix_small_imbalance(env, sds);
5863 /******* find_busiest_group() helpers end here *********************/
5866 * find_busiest_group - Returns the busiest group within the sched_domain
5867 * if there is an imbalance. If there isn't an imbalance, and
5868 * the user has opted for power-savings, it returns a group whose
5869 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5870 * such a group exists.
5872 * Also calculates the amount of weighted load which should be moved
5873 * to restore balance.
5875 * @env: The load balancing environment.
5877 * Return: - The busiest group if imbalance exists.
5878 * - If no imbalance and user has opted for power-savings balance,
5879 * return the least loaded group whose CPUs can be
5880 * put to idle by rebalancing its tasks onto our group.
5882 static struct sched_group *find_busiest_group(struct lb_env *env)
5884 struct sg_lb_stats *local, *busiest;
5885 struct sd_lb_stats sds;
5887 init_sd_lb_stats(&sds);
5890 * Compute the various statistics relavent for load balancing at
5893 update_sd_lb_stats(env, &sds);
5894 local = &sds.local_stat;
5895 busiest = &sds.busiest_stat;
5897 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5898 check_asym_packing(env, &sds))
5901 /* There is no busy sibling group to pull tasks from */
5902 if (!sds.busiest || busiest->sum_nr_running == 0)
5905 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5908 * If the busiest group is imbalanced the below checks don't
5909 * work because they assume all things are equal, which typically
5910 * isn't true due to cpus_allowed constraints and the like.
5912 if (busiest->group_imb)
5915 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5916 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5917 !busiest->group_has_capacity)
5921 * If the local group is more busy than the selected busiest group
5922 * don't try and pull any tasks.
5924 if (local->avg_load >= busiest->avg_load)
5928 * Don't pull any tasks if this group is already above the domain
5931 if (local->avg_load >= sds.avg_load)
5934 if (env->idle == CPU_IDLE) {
5936 * This cpu is idle. If the busiest group load doesn't
5937 * have more tasks than the number of available cpu's and
5938 * there is no imbalance between this and busiest group
5939 * wrt to idle cpu's, it is balanced.
5941 if ((local->idle_cpus < busiest->idle_cpus) &&
5942 busiest->sum_nr_running <= busiest->group_weight)
5946 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5947 * imbalance_pct to be conservative.
5949 if (100 * busiest->avg_load <=
5950 env->sd->imbalance_pct * local->avg_load)
5955 /* Looks like there is an imbalance. Compute it */
5956 calculate_imbalance(env, &sds);
5965 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5967 static struct rq *find_busiest_queue(struct lb_env *env,
5968 struct sched_group *group)
5970 struct rq *busiest = NULL, *rq;
5971 unsigned long busiest_load = 0, busiest_power = 1;
5974 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5975 unsigned long power, capacity, wl;
5979 rt = fbq_classify_rq(rq);
5982 * We classify groups/runqueues into three groups:
5983 * - regular: there are !numa tasks
5984 * - remote: there are numa tasks that run on the 'wrong' node
5985 * - all: there is no distinction
5987 * In order to avoid migrating ideally placed numa tasks,
5988 * ignore those when there's better options.
5990 * If we ignore the actual busiest queue to migrate another
5991 * task, the next balance pass can still reduce the busiest
5992 * queue by moving tasks around inside the node.
5994 * If we cannot move enough load due to this classification
5995 * the next pass will adjust the group classification and
5996 * allow migration of more tasks.
5998 * Both cases only affect the total convergence complexity.
6000 if (rt > env->fbq_type)
6003 power = power_of(i);
6004 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
6006 capacity = fix_small_capacity(env->sd, group);
6008 wl = weighted_cpuload(i);
6011 * When comparing with imbalance, use weighted_cpuload()
6012 * which is not scaled with the cpu power.
6014 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6018 * For the load comparisons with the other cpu's, consider
6019 * the weighted_cpuload() scaled with the cpu power, so that
6020 * the load can be moved away from the cpu that is potentially
6021 * running at a lower capacity.
6023 * Thus we're looking for max(wl_i / power_i), crosswise
6024 * multiplication to rid ourselves of the division works out
6025 * to: wl_i * power_j > wl_j * power_i; where j is our
6028 if (wl * busiest_power > busiest_load * power) {
6030 busiest_power = power;
6039 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6040 * so long as it is large enough.
6042 #define MAX_PINNED_INTERVAL 512
6044 /* Working cpumask for load_balance and load_balance_newidle. */
6045 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6047 static int need_active_balance(struct lb_env *env)
6049 struct sched_domain *sd = env->sd;
6051 if (env->idle == CPU_NEWLY_IDLE) {
6054 * ASYM_PACKING needs to force migrate tasks from busy but
6055 * higher numbered CPUs in order to pack all tasks in the
6056 * lowest numbered CPUs.
6058 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6062 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6065 static int active_load_balance_cpu_stop(void *data);
6067 static int should_we_balance(struct lb_env *env)
6069 struct sched_group *sg = env->sd->groups;
6070 struct cpumask *sg_cpus, *sg_mask;
6071 int cpu, balance_cpu = -1;
6074 * In the newly idle case, we will allow all the cpu's
6075 * to do the newly idle load balance.
6077 if (env->idle == CPU_NEWLY_IDLE)
6080 sg_cpus = sched_group_cpus(sg);
6081 sg_mask = sched_group_mask(sg);
6082 /* Try to find first idle cpu */
6083 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6084 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6091 if (balance_cpu == -1)
6092 balance_cpu = group_balance_cpu(sg);
6095 * First idle cpu or the first cpu(busiest) in this sched group
6096 * is eligible for doing load balancing at this and above domains.
6098 return balance_cpu == env->dst_cpu;
6102 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6103 * tasks if there is an imbalance.
6105 static int load_balance(int this_cpu, struct rq *this_rq,
6106 struct sched_domain *sd, enum cpu_idle_type idle,
6107 int *continue_balancing)
6109 int ld_moved, cur_ld_moved, active_balance = 0;
6110 struct sched_domain *sd_parent = sd->parent;
6111 struct sched_group *group;
6113 unsigned long flags;
6114 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6116 struct lb_env env = {
6118 .dst_cpu = this_cpu,
6120 .dst_grpmask = sched_group_cpus(sd->groups),
6122 .loop_break = sched_nr_migrate_break,
6128 * For NEWLY_IDLE load_balancing, we don't need to consider
6129 * other cpus in our group
6131 if (idle == CPU_NEWLY_IDLE)
6132 env.dst_grpmask = NULL;
6134 cpumask_copy(cpus, cpu_active_mask);
6136 schedstat_inc(sd, lb_count[idle]);
6139 if (!should_we_balance(&env)) {
6140 *continue_balancing = 0;
6144 group = find_busiest_group(&env);
6146 schedstat_inc(sd, lb_nobusyg[idle]);
6150 busiest = find_busiest_queue(&env, group);
6152 schedstat_inc(sd, lb_nobusyq[idle]);
6156 BUG_ON(busiest == env.dst_rq);
6158 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6161 if (busiest->nr_running > 1) {
6163 * Attempt to move tasks. If find_busiest_group has found
6164 * an imbalance but busiest->nr_running <= 1, the group is
6165 * still unbalanced. ld_moved simply stays zero, so it is
6166 * correctly treated as an imbalance.
6168 env.flags |= LBF_ALL_PINNED;
6169 env.src_cpu = busiest->cpu;
6170 env.src_rq = busiest;
6171 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6174 local_irq_save(flags);
6175 double_rq_lock(env.dst_rq, busiest);
6178 * cur_ld_moved - load moved in current iteration
6179 * ld_moved - cumulative load moved across iterations
6181 cur_ld_moved = move_tasks(&env);
6182 ld_moved += cur_ld_moved;
6183 double_rq_unlock(env.dst_rq, busiest);
6184 local_irq_restore(flags);
6187 * some other cpu did the load balance for us.
6189 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6190 resched_cpu(env.dst_cpu);
6192 if (env.flags & LBF_NEED_BREAK) {
6193 env.flags &= ~LBF_NEED_BREAK;
6198 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6199 * us and move them to an alternate dst_cpu in our sched_group
6200 * where they can run. The upper limit on how many times we
6201 * iterate on same src_cpu is dependent on number of cpus in our
6204 * This changes load balance semantics a bit on who can move
6205 * load to a given_cpu. In addition to the given_cpu itself
6206 * (or a ilb_cpu acting on its behalf where given_cpu is
6207 * nohz-idle), we now have balance_cpu in a position to move
6208 * load to given_cpu. In rare situations, this may cause
6209 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6210 * _independently_ and at _same_ time to move some load to
6211 * given_cpu) causing exceess load to be moved to given_cpu.
6212 * This however should not happen so much in practice and
6213 * moreover subsequent load balance cycles should correct the
6214 * excess load moved.
6216 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6218 /* Prevent to re-select dst_cpu via env's cpus */
6219 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6221 env.dst_rq = cpu_rq(env.new_dst_cpu);
6222 env.dst_cpu = env.new_dst_cpu;
6223 env.flags &= ~LBF_DST_PINNED;
6225 env.loop_break = sched_nr_migrate_break;
6228 * Go back to "more_balance" rather than "redo" since we
6229 * need to continue with same src_cpu.
6235 * We failed to reach balance because of affinity.
6238 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6240 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6241 *group_imbalance = 1;
6242 } else if (*group_imbalance)
6243 *group_imbalance = 0;
6246 /* All tasks on this runqueue were pinned by CPU affinity */
6247 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6248 cpumask_clear_cpu(cpu_of(busiest), cpus);
6249 if (!cpumask_empty(cpus)) {
6251 env.loop_break = sched_nr_migrate_break;
6259 schedstat_inc(sd, lb_failed[idle]);
6261 * Increment the failure counter only on periodic balance.
6262 * We do not want newidle balance, which can be very
6263 * frequent, pollute the failure counter causing
6264 * excessive cache_hot migrations and active balances.
6266 if (idle != CPU_NEWLY_IDLE)
6267 sd->nr_balance_failed++;
6269 if (need_active_balance(&env)) {
6270 raw_spin_lock_irqsave(&busiest->lock, flags);
6272 /* don't kick the active_load_balance_cpu_stop,
6273 * if the curr task on busiest cpu can't be
6276 if (!cpumask_test_cpu(this_cpu,
6277 tsk_cpus_allowed(busiest->curr))) {
6278 raw_spin_unlock_irqrestore(&busiest->lock,
6280 env.flags |= LBF_ALL_PINNED;
6281 goto out_one_pinned;
6285 * ->active_balance synchronizes accesses to
6286 * ->active_balance_work. Once set, it's cleared
6287 * only after active load balance is finished.
6289 if (!busiest->active_balance) {
6290 busiest->active_balance = 1;
6291 busiest->push_cpu = this_cpu;
6294 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6296 if (active_balance) {
6297 stop_one_cpu_nowait(cpu_of(busiest),
6298 active_load_balance_cpu_stop, busiest,
6299 &busiest->active_balance_work);
6303 * We've kicked active balancing, reset the failure
6306 sd->nr_balance_failed = sd->cache_nice_tries+1;
6309 sd->nr_balance_failed = 0;
6311 if (likely(!active_balance)) {
6312 /* We were unbalanced, so reset the balancing interval */
6313 sd->balance_interval = sd->min_interval;
6316 * If we've begun active balancing, start to back off. This
6317 * case may not be covered by the all_pinned logic if there
6318 * is only 1 task on the busy runqueue (because we don't call
6321 if (sd->balance_interval < sd->max_interval)
6322 sd->balance_interval *= 2;
6328 schedstat_inc(sd, lb_balanced[idle]);
6330 sd->nr_balance_failed = 0;
6333 /* tune up the balancing interval */
6334 if (((env.flags & LBF_ALL_PINNED) &&
6335 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6336 (sd->balance_interval < sd->max_interval))
6337 sd->balance_interval *= 2;
6345 * idle_balance is called by schedule() if this_cpu is about to become
6346 * idle. Attempts to pull tasks from other CPUs.
6348 void idle_balance(int this_cpu, struct rq *this_rq)
6350 struct sched_domain *sd;
6351 int pulled_task = 0;
6352 unsigned long next_balance = jiffies + HZ;
6355 this_rq->idle_stamp = rq_clock(this_rq);
6357 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6361 * Drop the rq->lock, but keep IRQ/preempt disabled.
6363 raw_spin_unlock(&this_rq->lock);
6365 update_blocked_averages(this_cpu);
6367 for_each_domain(this_cpu, sd) {
6368 unsigned long interval;
6369 int continue_balancing = 1;
6370 u64 t0, domain_cost;
6372 if (!(sd->flags & SD_LOAD_BALANCE))
6375 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6378 if (sd->flags & SD_BALANCE_NEWIDLE) {
6379 t0 = sched_clock_cpu(this_cpu);
6381 /* If we've pulled tasks over stop searching: */
6382 pulled_task = load_balance(this_cpu, this_rq,
6384 &continue_balancing);
6386 domain_cost = sched_clock_cpu(this_cpu) - t0;
6387 if (domain_cost > sd->max_newidle_lb_cost)
6388 sd->max_newidle_lb_cost = domain_cost;
6390 curr_cost += domain_cost;
6393 interval = msecs_to_jiffies(sd->balance_interval);
6394 if (time_after(next_balance, sd->last_balance + interval))
6395 next_balance = sd->last_balance + interval;
6397 this_rq->idle_stamp = 0;
6403 raw_spin_lock(&this_rq->lock);
6405 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6407 * We are going idle. next_balance may be set based on
6408 * a busy processor. So reset next_balance.
6410 this_rq->next_balance = next_balance;
6413 if (curr_cost > this_rq->max_idle_balance_cost)
6414 this_rq->max_idle_balance_cost = curr_cost;
6418 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6419 * running tasks off the busiest CPU onto idle CPUs. It requires at
6420 * least 1 task to be running on each physical CPU where possible, and
6421 * avoids physical / logical imbalances.
6423 static int active_load_balance_cpu_stop(void *data)
6425 struct rq *busiest_rq = data;
6426 int busiest_cpu = cpu_of(busiest_rq);
6427 int target_cpu = busiest_rq->push_cpu;
6428 struct rq *target_rq = cpu_rq(target_cpu);
6429 struct sched_domain *sd;
6431 raw_spin_lock_irq(&busiest_rq->lock);
6433 /* make sure the requested cpu hasn't gone down in the meantime */
6434 if (unlikely(busiest_cpu != smp_processor_id() ||
6435 !busiest_rq->active_balance))
6438 /* Is there any task to move? */
6439 if (busiest_rq->nr_running <= 1)
6443 * This condition is "impossible", if it occurs
6444 * we need to fix it. Originally reported by
6445 * Bjorn Helgaas on a 128-cpu setup.
6447 BUG_ON(busiest_rq == target_rq);
6449 /* move a task from busiest_rq to target_rq */
6450 double_lock_balance(busiest_rq, target_rq);
6452 /* Search for an sd spanning us and the target CPU. */
6454 for_each_domain(target_cpu, sd) {
6455 if ((sd->flags & SD_LOAD_BALANCE) &&
6456 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6461 struct lb_env env = {
6463 .dst_cpu = target_cpu,
6464 .dst_rq = target_rq,
6465 .src_cpu = busiest_rq->cpu,
6466 .src_rq = busiest_rq,
6470 schedstat_inc(sd, alb_count);
6472 if (move_one_task(&env))
6473 schedstat_inc(sd, alb_pushed);
6475 schedstat_inc(sd, alb_failed);
6478 double_unlock_balance(busiest_rq, target_rq);
6480 busiest_rq->active_balance = 0;
6481 raw_spin_unlock_irq(&busiest_rq->lock);
6485 #ifdef CONFIG_NO_HZ_COMMON
6487 * idle load balancing details
6488 * - When one of the busy CPUs notice that there may be an idle rebalancing
6489 * needed, they will kick the idle load balancer, which then does idle
6490 * load balancing for all the idle CPUs.
6493 cpumask_var_t idle_cpus_mask;
6495 unsigned long next_balance; /* in jiffy units */
6496 } nohz ____cacheline_aligned;
6498 static inline int find_new_ilb(void)
6500 int ilb = cpumask_first(nohz.idle_cpus_mask);
6502 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6509 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6510 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6511 * CPU (if there is one).
6513 static void nohz_balancer_kick(void)
6517 nohz.next_balance++;
6519 ilb_cpu = find_new_ilb();
6521 if (ilb_cpu >= nr_cpu_ids)
6524 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6527 * Use smp_send_reschedule() instead of resched_cpu().
6528 * This way we generate a sched IPI on the target cpu which
6529 * is idle. And the softirq performing nohz idle load balance
6530 * will be run before returning from the IPI.
6532 smp_send_reschedule(ilb_cpu);
6536 static inline void nohz_balance_exit_idle(int cpu)
6538 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6539 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6540 atomic_dec(&nohz.nr_cpus);
6541 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6545 static inline void set_cpu_sd_state_busy(void)
6547 struct sched_domain *sd;
6548 int cpu = smp_processor_id();
6551 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6553 if (!sd || !sd->nohz_idle)
6557 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6562 void set_cpu_sd_state_idle(void)
6564 struct sched_domain *sd;
6565 int cpu = smp_processor_id();
6568 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6570 if (!sd || sd->nohz_idle)
6574 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6580 * This routine will record that the cpu is going idle with tick stopped.
6581 * This info will be used in performing idle load balancing in the future.
6583 void nohz_balance_enter_idle(int cpu)
6586 * If this cpu is going down, then nothing needs to be done.
6588 if (!cpu_active(cpu))
6591 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6594 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6595 atomic_inc(&nohz.nr_cpus);
6596 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6599 static int sched_ilb_notifier(struct notifier_block *nfb,
6600 unsigned long action, void *hcpu)
6602 switch (action & ~CPU_TASKS_FROZEN) {
6604 nohz_balance_exit_idle(smp_processor_id());
6612 static DEFINE_SPINLOCK(balancing);
6615 * Scale the max load_balance interval with the number of CPUs in the system.
6616 * This trades load-balance latency on larger machines for less cross talk.
6618 void update_max_interval(void)
6620 max_load_balance_interval = HZ*num_online_cpus()/10;
6624 * It checks each scheduling domain to see if it is due to be balanced,
6625 * and initiates a balancing operation if so.
6627 * Balancing parameters are set up in init_sched_domains.
6629 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
6631 int continue_balancing = 1;
6633 unsigned long interval;
6634 struct sched_domain *sd;
6635 /* Earliest time when we have to do rebalance again */
6636 unsigned long next_balance = jiffies + 60*HZ;
6637 int update_next_balance = 0;
6638 int need_serialize, need_decay = 0;
6641 update_blocked_averages(cpu);
6644 for_each_domain(cpu, sd) {
6646 * Decay the newidle max times here because this is a regular
6647 * visit to all the domains. Decay ~1% per second.
6649 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6650 sd->max_newidle_lb_cost =
6651 (sd->max_newidle_lb_cost * 253) / 256;
6652 sd->next_decay_max_lb_cost = jiffies + HZ;
6655 max_cost += sd->max_newidle_lb_cost;
6657 if (!(sd->flags & SD_LOAD_BALANCE))
6661 * Stop the load balance at this level. There is another
6662 * CPU in our sched group which is doing load balancing more
6665 if (!continue_balancing) {
6671 interval = sd->balance_interval;
6672 if (idle != CPU_IDLE)
6673 interval *= sd->busy_factor;
6675 /* scale ms to jiffies */
6676 interval = msecs_to_jiffies(interval);
6677 interval = clamp(interval, 1UL, max_load_balance_interval);
6679 need_serialize = sd->flags & SD_SERIALIZE;
6681 if (need_serialize) {
6682 if (!spin_trylock(&balancing))
6686 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6687 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6689 * The LBF_DST_PINNED logic could have changed
6690 * env->dst_cpu, so we can't know our idle
6691 * state even if we migrated tasks. Update it.
6693 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6695 sd->last_balance = jiffies;
6698 spin_unlock(&balancing);
6700 if (time_after(next_balance, sd->last_balance + interval)) {
6701 next_balance = sd->last_balance + interval;
6702 update_next_balance = 1;
6707 * Ensure the rq-wide value also decays but keep it at a
6708 * reasonable floor to avoid funnies with rq->avg_idle.
6710 rq->max_idle_balance_cost =
6711 max((u64)sysctl_sched_migration_cost, max_cost);
6716 * next_balance will be updated only when there is a need.
6717 * When the cpu is attached to null domain for ex, it will not be
6720 if (likely(update_next_balance))
6721 rq->next_balance = next_balance;
6724 #ifdef CONFIG_NO_HZ_COMMON
6726 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6727 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6729 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
6731 int this_cpu = this_rq->cpu;
6735 if (idle != CPU_IDLE ||
6736 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6739 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6740 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6744 * If this cpu gets work to do, stop the load balancing
6745 * work being done for other cpus. Next load
6746 * balancing owner will pick it up.
6751 rq = cpu_rq(balance_cpu);
6753 raw_spin_lock_irq(&rq->lock);
6754 update_rq_clock(rq);
6755 update_idle_cpu_load(rq);
6756 raw_spin_unlock_irq(&rq->lock);
6758 rebalance_domains(rq, CPU_IDLE);
6760 if (time_after(this_rq->next_balance, rq->next_balance))
6761 this_rq->next_balance = rq->next_balance;
6763 nohz.next_balance = this_rq->next_balance;
6765 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6769 * Current heuristic for kicking the idle load balancer in the presence
6770 * of an idle cpu is the system.
6771 * - This rq has more than one task.
6772 * - At any scheduler domain level, this cpu's scheduler group has multiple
6773 * busy cpu's exceeding the group's power.
6774 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6775 * domain span are idle.
6777 static inline int nohz_kick_needed(struct rq *rq)
6779 unsigned long now = jiffies;
6780 struct sched_domain *sd;
6781 struct sched_group_power *sgp;
6782 int nr_busy, cpu = rq->cpu;
6784 if (unlikely(rq->idle_balance))
6788 * We may be recently in ticked or tickless idle mode. At the first
6789 * busy tick after returning from idle, we will update the busy stats.
6791 set_cpu_sd_state_busy();
6792 nohz_balance_exit_idle(cpu);
6795 * None are in tickless mode and hence no need for NOHZ idle load
6798 if (likely(!atomic_read(&nohz.nr_cpus)))
6801 if (time_before(now, nohz.next_balance))
6804 if (rq->nr_running >= 2)
6808 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6811 sgp = sd->groups->sgp;
6812 nr_busy = atomic_read(&sgp->nr_busy_cpus);
6815 goto need_kick_unlock;
6818 sd = rcu_dereference(per_cpu(sd_asym, cpu));
6820 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
6821 sched_domain_span(sd)) < cpu))
6822 goto need_kick_unlock;
6833 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
6837 * run_rebalance_domains is triggered when needed from the scheduler tick.
6838 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6840 static void run_rebalance_domains(struct softirq_action *h)
6842 struct rq *this_rq = this_rq();
6843 enum cpu_idle_type idle = this_rq->idle_balance ?
6844 CPU_IDLE : CPU_NOT_IDLE;
6846 rebalance_domains(this_rq, idle);
6849 * If this cpu has a pending nohz_balance_kick, then do the
6850 * balancing on behalf of the other idle cpus whose ticks are
6853 nohz_idle_balance(this_rq, idle);
6856 static inline int on_null_domain(struct rq *rq)
6858 return !rcu_dereference_sched(rq->sd);
6862 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6864 void trigger_load_balance(struct rq *rq)
6866 /* Don't need to rebalance while attached to NULL domain */
6867 if (unlikely(on_null_domain(rq)))
6870 if (time_after_eq(jiffies, rq->next_balance))
6871 raise_softirq(SCHED_SOFTIRQ);
6872 #ifdef CONFIG_NO_HZ_COMMON
6873 if (nohz_kick_needed(rq))
6874 nohz_balancer_kick();
6878 static void rq_online_fair(struct rq *rq)
6883 static void rq_offline_fair(struct rq *rq)
6887 /* Ensure any throttled groups are reachable by pick_next_task */
6888 unthrottle_offline_cfs_rqs(rq);
6891 #endif /* CONFIG_SMP */
6894 * scheduler tick hitting a task of our scheduling class:
6896 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6898 struct cfs_rq *cfs_rq;
6899 struct sched_entity *se = &curr->se;
6901 for_each_sched_entity(se) {
6902 cfs_rq = cfs_rq_of(se);
6903 entity_tick(cfs_rq, se, queued);
6906 if (numabalancing_enabled)
6907 task_tick_numa(rq, curr);
6909 update_rq_runnable_avg(rq, 1);
6913 * called on fork with the child task as argument from the parent's context
6914 * - child not yet on the tasklist
6915 * - preemption disabled
6917 static void task_fork_fair(struct task_struct *p)
6919 struct cfs_rq *cfs_rq;
6920 struct sched_entity *se = &p->se, *curr;
6921 int this_cpu = smp_processor_id();
6922 struct rq *rq = this_rq();
6923 unsigned long flags;
6925 raw_spin_lock_irqsave(&rq->lock, flags);
6927 update_rq_clock(rq);
6929 cfs_rq = task_cfs_rq(current);
6930 curr = cfs_rq->curr;
6933 * Not only the cpu but also the task_group of the parent might have
6934 * been changed after parent->se.parent,cfs_rq were copied to
6935 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6936 * of child point to valid ones.
6939 __set_task_cpu(p, this_cpu);
6942 update_curr(cfs_rq);
6945 se->vruntime = curr->vruntime;
6946 place_entity(cfs_rq, se, 1);
6948 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6950 * Upon rescheduling, sched_class::put_prev_task() will place
6951 * 'current' within the tree based on its new key value.
6953 swap(curr->vruntime, se->vruntime);
6954 resched_task(rq->curr);
6957 se->vruntime -= cfs_rq->min_vruntime;
6959 raw_spin_unlock_irqrestore(&rq->lock, flags);
6963 * Priority of the task has changed. Check to see if we preempt
6967 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6973 * Reschedule if we are currently running on this runqueue and
6974 * our priority decreased, or if we are not currently running on
6975 * this runqueue and our priority is higher than the current's
6977 if (rq->curr == p) {
6978 if (p->prio > oldprio)
6979 resched_task(rq->curr);
6981 check_preempt_curr(rq, p, 0);
6984 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6986 struct sched_entity *se = &p->se;
6987 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6990 * Ensure the task's vruntime is normalized, so that when its
6991 * switched back to the fair class the enqueue_entity(.flags=0) will
6992 * do the right thing.
6994 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6995 * have normalized the vruntime, if it was !on_rq, then only when
6996 * the task is sleeping will it still have non-normalized vruntime.
6998 if (!se->on_rq && p->state != TASK_RUNNING) {
7000 * Fix up our vruntime so that the current sleep doesn't
7001 * cause 'unlimited' sleep bonus.
7003 place_entity(cfs_rq, se, 0);
7004 se->vruntime -= cfs_rq->min_vruntime;
7009 * Remove our load from contribution when we leave sched_fair
7010 * and ensure we don't carry in an old decay_count if we
7013 if (se->avg.decay_count) {
7014 __synchronize_entity_decay(se);
7015 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7021 * We switched to the sched_fair class.
7023 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7029 * We were most likely switched from sched_rt, so
7030 * kick off the schedule if running, otherwise just see
7031 * if we can still preempt the current task.
7034 resched_task(rq->curr);
7036 check_preempt_curr(rq, p, 0);
7039 /* Account for a task changing its policy or group.
7041 * This routine is mostly called to set cfs_rq->curr field when a task
7042 * migrates between groups/classes.
7044 static void set_curr_task_fair(struct rq *rq)
7046 struct sched_entity *se = &rq->curr->se;
7048 for_each_sched_entity(se) {
7049 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7051 set_next_entity(cfs_rq, se);
7052 /* ensure bandwidth has been allocated on our new cfs_rq */
7053 account_cfs_rq_runtime(cfs_rq, 0);
7057 void init_cfs_rq(struct cfs_rq *cfs_rq)
7059 cfs_rq->tasks_timeline = RB_ROOT;
7060 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7061 #ifndef CONFIG_64BIT
7062 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7065 atomic64_set(&cfs_rq->decay_counter, 1);
7066 atomic_long_set(&cfs_rq->removed_load, 0);
7070 #ifdef CONFIG_FAIR_GROUP_SCHED
7071 static void task_move_group_fair(struct task_struct *p, int on_rq)
7073 struct cfs_rq *cfs_rq;
7075 * If the task was not on the rq at the time of this cgroup movement
7076 * it must have been asleep, sleeping tasks keep their ->vruntime
7077 * absolute on their old rq until wakeup (needed for the fair sleeper
7078 * bonus in place_entity()).
7080 * If it was on the rq, we've just 'preempted' it, which does convert
7081 * ->vruntime to a relative base.
7083 * Make sure both cases convert their relative position when migrating
7084 * to another cgroup's rq. This does somewhat interfere with the
7085 * fair sleeper stuff for the first placement, but who cares.
7088 * When !on_rq, vruntime of the task has usually NOT been normalized.
7089 * But there are some cases where it has already been normalized:
7091 * - Moving a forked child which is waiting for being woken up by
7092 * wake_up_new_task().
7093 * - Moving a task which has been woken up by try_to_wake_up() and
7094 * waiting for actually being woken up by sched_ttwu_pending().
7096 * To prevent boost or penalty in the new cfs_rq caused by delta
7097 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7099 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7103 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7104 set_task_rq(p, task_cpu(p));
7106 cfs_rq = cfs_rq_of(&p->se);
7107 p->se.vruntime += cfs_rq->min_vruntime;
7110 * migrate_task_rq_fair() will have removed our previous
7111 * contribution, but we must synchronize for ongoing future
7114 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7115 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7120 void free_fair_sched_group(struct task_group *tg)
7124 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7126 for_each_possible_cpu(i) {
7128 kfree(tg->cfs_rq[i]);
7137 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7139 struct cfs_rq *cfs_rq;
7140 struct sched_entity *se;
7143 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7146 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7150 tg->shares = NICE_0_LOAD;
7152 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7154 for_each_possible_cpu(i) {
7155 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7156 GFP_KERNEL, cpu_to_node(i));
7160 se = kzalloc_node(sizeof(struct sched_entity),
7161 GFP_KERNEL, cpu_to_node(i));
7165 init_cfs_rq(cfs_rq);
7166 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7177 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7179 struct rq *rq = cpu_rq(cpu);
7180 unsigned long flags;
7183 * Only empty task groups can be destroyed; so we can speculatively
7184 * check on_list without danger of it being re-added.
7186 if (!tg->cfs_rq[cpu]->on_list)
7189 raw_spin_lock_irqsave(&rq->lock, flags);
7190 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7191 raw_spin_unlock_irqrestore(&rq->lock, flags);
7194 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7195 struct sched_entity *se, int cpu,
7196 struct sched_entity *parent)
7198 struct rq *rq = cpu_rq(cpu);
7202 init_cfs_rq_runtime(cfs_rq);
7204 tg->cfs_rq[cpu] = cfs_rq;
7207 /* se could be NULL for root_task_group */
7212 se->cfs_rq = &rq->cfs;
7214 se->cfs_rq = parent->my_q;
7217 /* guarantee group entities always have weight */
7218 update_load_set(&se->load, NICE_0_LOAD);
7219 se->parent = parent;
7222 static DEFINE_MUTEX(shares_mutex);
7224 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7227 unsigned long flags;
7230 * We can't change the weight of the root cgroup.
7235 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7237 mutex_lock(&shares_mutex);
7238 if (tg->shares == shares)
7241 tg->shares = shares;
7242 for_each_possible_cpu(i) {
7243 struct rq *rq = cpu_rq(i);
7244 struct sched_entity *se;
7247 /* Propagate contribution to hierarchy */
7248 raw_spin_lock_irqsave(&rq->lock, flags);
7250 /* Possible calls to update_curr() need rq clock */
7251 update_rq_clock(rq);
7252 for_each_sched_entity(se)
7253 update_cfs_shares(group_cfs_rq(se));
7254 raw_spin_unlock_irqrestore(&rq->lock, flags);
7258 mutex_unlock(&shares_mutex);
7261 #else /* CONFIG_FAIR_GROUP_SCHED */
7263 void free_fair_sched_group(struct task_group *tg) { }
7265 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7270 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7272 #endif /* CONFIG_FAIR_GROUP_SCHED */
7275 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7277 struct sched_entity *se = &task->se;
7278 unsigned int rr_interval = 0;
7281 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7284 if (rq->cfs.load.weight)
7285 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7291 * All the scheduling class methods:
7293 const struct sched_class fair_sched_class = {
7294 .next = &idle_sched_class,
7295 .enqueue_task = enqueue_task_fair,
7296 .dequeue_task = dequeue_task_fair,
7297 .yield_task = yield_task_fair,
7298 .yield_to_task = yield_to_task_fair,
7300 .check_preempt_curr = check_preempt_wakeup,
7302 .pick_next_task = pick_next_task_fair,
7303 .put_prev_task = put_prev_task_fair,
7306 .select_task_rq = select_task_rq_fair,
7307 .migrate_task_rq = migrate_task_rq_fair,
7309 .rq_online = rq_online_fair,
7310 .rq_offline = rq_offline_fair,
7312 .task_waking = task_waking_fair,
7315 .set_curr_task = set_curr_task_fair,
7316 .task_tick = task_tick_fair,
7317 .task_fork = task_fork_fair,
7319 .prio_changed = prio_changed_fair,
7320 .switched_from = switched_from_fair,
7321 .switched_to = switched_to_fair,
7323 .get_rr_interval = get_rr_interval_fair,
7325 #ifdef CONFIG_FAIR_GROUP_SCHED
7326 .task_move_group = task_move_group_fair,
7330 #ifdef CONFIG_SCHED_DEBUG
7331 void print_cfs_stats(struct seq_file *m, int cpu)
7333 struct cfs_rq *cfs_rq;
7336 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7337 print_cfs_rq(m, cpu, cfs_rq);
7342 __init void init_sched_fair_class(void)
7345 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7347 #ifdef CONFIG_NO_HZ_COMMON
7348 nohz.next_balance = jiffies;
7349 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7350 cpu_notifier(sched_ilb_notifier, 0);