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;
823 * After skipping a page migration on a shared page, skip N more numa page
824 * migrations unconditionally. This reduces the number of NUMA migrations
825 * in shared memory workloads, and has the effect of pulling tasks towards
826 * where their memory lives, over pulling the memory towards the task.
828 unsigned int sysctl_numa_balancing_migrate_deferred = 16;
830 static unsigned int task_nr_scan_windows(struct task_struct *p)
832 unsigned long rss = 0;
833 unsigned long nr_scan_pages;
836 * Calculations based on RSS as non-present and empty pages are skipped
837 * by the PTE scanner and NUMA hinting faults should be trapped based
840 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
841 rss = get_mm_rss(p->mm);
845 rss = round_up(rss, nr_scan_pages);
846 return rss / nr_scan_pages;
849 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
850 #define MAX_SCAN_WINDOW 2560
852 static unsigned int task_scan_min(struct task_struct *p)
854 unsigned int scan, floor;
855 unsigned int windows = 1;
857 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
858 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
859 floor = 1000 / windows;
861 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
862 return max_t(unsigned int, floor, scan);
865 static unsigned int task_scan_max(struct task_struct *p)
867 unsigned int smin = task_scan_min(p);
870 /* Watch for min being lower than max due to floor calculations */
871 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
872 return max(smin, smax);
875 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
877 rq->nr_numa_running += (p->numa_preferred_nid != -1);
878 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
881 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
883 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
884 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
890 spinlock_t lock; /* nr_tasks, tasks */
893 struct list_head task_list;
896 unsigned long total_faults;
897 unsigned long faults[0];
900 pid_t task_numa_group_id(struct task_struct *p)
902 return p->numa_group ? p->numa_group->gid : 0;
905 static inline int task_faults_idx(int nid, int priv)
907 return 2 * nid + priv;
910 static inline unsigned long task_faults(struct task_struct *p, int nid)
915 return p->numa_faults[task_faults_idx(nid, 0)] +
916 p->numa_faults[task_faults_idx(nid, 1)];
919 static inline unsigned long group_faults(struct task_struct *p, int nid)
924 return p->numa_group->faults[task_faults_idx(nid, 0)] +
925 p->numa_group->faults[task_faults_idx(nid, 1)];
929 * These return the fraction of accesses done by a particular task, or
930 * task group, on a particular numa node. The group weight is given a
931 * larger multiplier, in order to group tasks together that are almost
932 * evenly spread out between numa nodes.
934 static inline unsigned long task_weight(struct task_struct *p, int nid)
936 unsigned long total_faults;
941 total_faults = p->total_numa_faults;
946 return 1000 * task_faults(p, nid) / total_faults;
949 static inline unsigned long group_weight(struct task_struct *p, int nid)
951 if (!p->numa_group || !p->numa_group->total_faults)
954 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
957 static unsigned long weighted_cpuload(const int cpu);
958 static unsigned long source_load(int cpu, int type);
959 static unsigned long target_load(int cpu, int type);
960 static unsigned long power_of(int cpu);
961 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
963 /* Cached statistics for all CPUs within a node */
965 unsigned long nr_running;
968 /* Total compute capacity of CPUs on a node */
971 /* Approximate capacity in terms of runnable tasks on a node */
972 unsigned long capacity;
977 * XXX borrowed from update_sg_lb_stats
979 static void update_numa_stats(struct numa_stats *ns, int nid)
983 memset(ns, 0, sizeof(*ns));
984 for_each_cpu(cpu, cpumask_of_node(nid)) {
985 struct rq *rq = cpu_rq(cpu);
987 ns->nr_running += rq->nr_running;
988 ns->load += weighted_cpuload(cpu);
989 ns->power += power_of(cpu);
995 * If we raced with hotplug and there are no CPUs left in our mask
996 * the @ns structure is NULL'ed and task_numa_compare() will
997 * not find this node attractive.
999 * We'll either bail at !has_capacity, or we'll detect a huge imbalance
1005 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1006 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1007 ns->has_capacity = (ns->nr_running < ns->capacity);
1010 struct task_numa_env {
1011 struct task_struct *p;
1013 int src_cpu, src_nid;
1014 int dst_cpu, dst_nid;
1016 struct numa_stats src_stats, dst_stats;
1020 struct task_struct *best_task;
1025 static void task_numa_assign(struct task_numa_env *env,
1026 struct task_struct *p, long imp)
1029 put_task_struct(env->best_task);
1034 env->best_imp = imp;
1035 env->best_cpu = env->dst_cpu;
1039 * This checks if the overall compute and NUMA accesses of the system would
1040 * be improved if the source tasks was migrated to the target dst_cpu taking
1041 * into account that it might be best if task running on the dst_cpu should
1042 * be exchanged with the source task
1044 static void task_numa_compare(struct task_numa_env *env,
1045 long taskimp, long groupimp)
1047 struct rq *src_rq = cpu_rq(env->src_cpu);
1048 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1049 struct task_struct *cur;
1050 long dst_load, src_load;
1052 long imp = (groupimp > 0) ? groupimp : taskimp;
1055 cur = ACCESS_ONCE(dst_rq->curr);
1056 if (cur->pid == 0) /* idle */
1060 * "imp" is the fault differential for the source task between the
1061 * source and destination node. Calculate the total differential for
1062 * the source task and potential destination task. The more negative
1063 * the value is, the more rmeote accesses that would be expected to
1064 * be incurred if the tasks were swapped.
1067 /* Skip this swap candidate if cannot move to the source cpu */
1068 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1072 * If dst and source tasks are in the same NUMA group, or not
1073 * in any group then look only at task weights.
1075 if (cur->numa_group == env->p->numa_group) {
1076 imp = taskimp + task_weight(cur, env->src_nid) -
1077 task_weight(cur, env->dst_nid);
1079 * Add some hysteresis to prevent swapping the
1080 * tasks within a group over tiny differences.
1082 if (cur->numa_group)
1086 * Compare the group weights. If a task is all by
1087 * itself (not part of a group), use the task weight
1090 if (env->p->numa_group)
1095 if (cur->numa_group)
1096 imp += group_weight(cur, env->src_nid) -
1097 group_weight(cur, env->dst_nid);
1099 imp += task_weight(cur, env->src_nid) -
1100 task_weight(cur, env->dst_nid);
1104 if (imp < env->best_imp)
1108 /* Is there capacity at our destination? */
1109 if (env->src_stats.has_capacity &&
1110 !env->dst_stats.has_capacity)
1116 /* Balance doesn't matter much if we're running a task per cpu */
1117 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1121 * In the overloaded case, try and keep the load balanced.
1124 dst_load = env->dst_stats.load;
1125 src_load = env->src_stats.load;
1127 /* XXX missing power terms */
1128 load = task_h_load(env->p);
1133 load = task_h_load(cur);
1138 /* make src_load the smaller */
1139 if (dst_load < src_load)
1140 swap(dst_load, src_load);
1142 if (src_load * env->imbalance_pct < dst_load * 100)
1146 task_numa_assign(env, cur, imp);
1151 static void task_numa_find_cpu(struct task_numa_env *env,
1152 long taskimp, long groupimp)
1156 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1157 /* Skip this CPU if the source task cannot migrate */
1158 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1162 task_numa_compare(env, taskimp, groupimp);
1166 static int task_numa_migrate(struct task_struct *p)
1168 struct task_numa_env env = {
1171 .src_cpu = task_cpu(p),
1172 .src_nid = task_node(p),
1174 .imbalance_pct = 112,
1180 struct sched_domain *sd;
1181 unsigned long taskweight, groupweight;
1183 long taskimp, groupimp;
1186 * Pick the lowest SD_NUMA domain, as that would have the smallest
1187 * imbalance and would be the first to start moving tasks about.
1189 * And we want to avoid any moving of tasks about, as that would create
1190 * random movement of tasks -- counter the numa conditions we're trying
1194 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1196 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1200 * Cpusets can break the scheduler domain tree into smaller
1201 * balance domains, some of which do not cross NUMA boundaries.
1202 * Tasks that are "trapped" in such domains cannot be migrated
1203 * elsewhere, so there is no point in (re)trying.
1205 if (unlikely(!sd)) {
1206 p->numa_preferred_nid = task_node(p);
1210 taskweight = task_weight(p, env.src_nid);
1211 groupweight = group_weight(p, env.src_nid);
1212 update_numa_stats(&env.src_stats, env.src_nid);
1213 env.dst_nid = p->numa_preferred_nid;
1214 taskimp = task_weight(p, env.dst_nid) - taskweight;
1215 groupimp = group_weight(p, env.dst_nid) - groupweight;
1216 update_numa_stats(&env.dst_stats, env.dst_nid);
1218 /* If the preferred nid has capacity, try to use it. */
1219 if (env.dst_stats.has_capacity)
1220 task_numa_find_cpu(&env, taskimp, groupimp);
1222 /* No space available on the preferred nid. Look elsewhere. */
1223 if (env.best_cpu == -1) {
1224 for_each_online_node(nid) {
1225 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1228 /* Only consider nodes where both task and groups benefit */
1229 taskimp = task_weight(p, nid) - taskweight;
1230 groupimp = group_weight(p, nid) - groupweight;
1231 if (taskimp < 0 && groupimp < 0)
1235 update_numa_stats(&env.dst_stats, env.dst_nid);
1236 task_numa_find_cpu(&env, taskimp, groupimp);
1240 /* No better CPU than the current one was found. */
1241 if (env.best_cpu == -1)
1244 sched_setnuma(p, env.dst_nid);
1247 * Reset the scan period if the task is being rescheduled on an
1248 * alternative node to recheck if the tasks is now properly placed.
1250 p->numa_scan_period = task_scan_min(p);
1252 if (env.best_task == NULL) {
1253 ret = migrate_task_to(p, env.best_cpu);
1255 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1259 ret = migrate_swap(p, env.best_task);
1261 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1262 put_task_struct(env.best_task);
1266 /* Attempt to migrate a task to a CPU on the preferred node. */
1267 static void numa_migrate_preferred(struct task_struct *p)
1269 /* This task has no NUMA fault statistics yet */
1270 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1273 /* Periodically retry migrating the task to the preferred node */
1274 p->numa_migrate_retry = jiffies + HZ;
1276 /* Success if task is already running on preferred CPU */
1277 if (task_node(p) == p->numa_preferred_nid)
1280 /* Otherwise, try migrate to a CPU on the preferred node */
1281 task_numa_migrate(p);
1285 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1286 * increments. The more local the fault statistics are, the higher the scan
1287 * period will be for the next scan window. If local/remote ratio is below
1288 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1289 * scan period will decrease
1291 #define NUMA_PERIOD_SLOTS 10
1292 #define NUMA_PERIOD_THRESHOLD 3
1295 * Increase the scan period (slow down scanning) if the majority of
1296 * our memory is already on our local node, or if the majority of
1297 * the page accesses are shared with other processes.
1298 * Otherwise, decrease the scan period.
1300 static void update_task_scan_period(struct task_struct *p,
1301 unsigned long shared, unsigned long private)
1303 unsigned int period_slot;
1307 unsigned long remote = p->numa_faults_locality[0];
1308 unsigned long local = p->numa_faults_locality[1];
1311 * If there were no record hinting faults then either the task is
1312 * completely idle or all activity is areas that are not of interest
1313 * to automatic numa balancing. Scan slower
1315 if (local + shared == 0) {
1316 p->numa_scan_period = min(p->numa_scan_period_max,
1317 p->numa_scan_period << 1);
1319 p->mm->numa_next_scan = jiffies +
1320 msecs_to_jiffies(p->numa_scan_period);
1326 * Prepare to scale scan period relative to the current period.
1327 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1328 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1329 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1331 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1332 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1333 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1334 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1337 diff = slot * period_slot;
1339 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1342 * Scale scan rate increases based on sharing. There is an
1343 * inverse relationship between the degree of sharing and
1344 * the adjustment made to the scanning period. Broadly
1345 * speaking the intent is that there is little point
1346 * scanning faster if shared accesses dominate as it may
1347 * simply bounce migrations uselessly
1349 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1350 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1353 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1354 task_scan_min(p), task_scan_max(p));
1355 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1358 static void task_numa_placement(struct task_struct *p)
1360 int seq, nid, max_nid = -1, max_group_nid = -1;
1361 unsigned long max_faults = 0, max_group_faults = 0;
1362 unsigned long fault_types[2] = { 0, 0 };
1363 spinlock_t *group_lock = NULL;
1365 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1366 if (p->numa_scan_seq == seq)
1368 p->numa_scan_seq = seq;
1369 p->numa_scan_period_max = task_scan_max(p);
1371 /* If the task is part of a group prevent parallel updates to group stats */
1372 if (p->numa_group) {
1373 group_lock = &p->numa_group->lock;
1374 spin_lock(group_lock);
1377 /* Find the node with the highest number of faults */
1378 for_each_online_node(nid) {
1379 unsigned long faults = 0, group_faults = 0;
1382 for (priv = 0; priv < 2; priv++) {
1385 i = task_faults_idx(nid, priv);
1386 diff = -p->numa_faults[i];
1388 /* Decay existing window, copy faults since last scan */
1389 p->numa_faults[i] >>= 1;
1390 p->numa_faults[i] += p->numa_faults_buffer[i];
1391 fault_types[priv] += p->numa_faults_buffer[i];
1392 p->numa_faults_buffer[i] = 0;
1394 faults += p->numa_faults[i];
1395 diff += p->numa_faults[i];
1396 p->total_numa_faults += diff;
1397 if (p->numa_group) {
1398 /* safe because we can only change our own group */
1399 p->numa_group->faults[i] += diff;
1400 p->numa_group->total_faults += diff;
1401 group_faults += p->numa_group->faults[i];
1405 if (faults > max_faults) {
1406 max_faults = faults;
1410 if (group_faults > max_group_faults) {
1411 max_group_faults = group_faults;
1412 max_group_nid = nid;
1416 update_task_scan_period(p, fault_types[0], fault_types[1]);
1418 if (p->numa_group) {
1420 * If the preferred task and group nids are different,
1421 * iterate over the nodes again to find the best place.
1423 if (max_nid != max_group_nid) {
1424 unsigned long weight, max_weight = 0;
1426 for_each_online_node(nid) {
1427 weight = task_weight(p, nid) + group_weight(p, nid);
1428 if (weight > max_weight) {
1429 max_weight = weight;
1435 spin_unlock(group_lock);
1438 /* Preferred node as the node with the most faults */
1439 if (max_faults && max_nid != p->numa_preferred_nid) {
1440 /* Update the preferred nid and migrate task if possible */
1441 sched_setnuma(p, max_nid);
1442 numa_migrate_preferred(p);
1446 static inline int get_numa_group(struct numa_group *grp)
1448 return atomic_inc_not_zero(&grp->refcount);
1451 static inline void put_numa_group(struct numa_group *grp)
1453 if (atomic_dec_and_test(&grp->refcount))
1454 kfree_rcu(grp, rcu);
1457 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1460 struct numa_group *grp, *my_grp;
1461 struct task_struct *tsk;
1463 int cpu = cpupid_to_cpu(cpupid);
1466 if (unlikely(!p->numa_group)) {
1467 unsigned int size = sizeof(struct numa_group) +
1468 2*nr_node_ids*sizeof(unsigned long);
1470 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1474 atomic_set(&grp->refcount, 1);
1475 spin_lock_init(&grp->lock);
1476 INIT_LIST_HEAD(&grp->task_list);
1479 for (i = 0; i < 2*nr_node_ids; i++)
1480 grp->faults[i] = p->numa_faults[i];
1482 grp->total_faults = p->total_numa_faults;
1484 list_add(&p->numa_entry, &grp->task_list);
1486 rcu_assign_pointer(p->numa_group, grp);
1490 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1492 if (!cpupid_match_pid(tsk, cpupid))
1495 grp = rcu_dereference(tsk->numa_group);
1499 my_grp = p->numa_group;
1504 * Only join the other group if its bigger; if we're the bigger group,
1505 * the other task will join us.
1507 if (my_grp->nr_tasks > grp->nr_tasks)
1511 * Tie-break on the grp address.
1513 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1516 /* Always join threads in the same process. */
1517 if (tsk->mm == current->mm)
1520 /* Simple filter to avoid false positives due to PID collisions */
1521 if (flags & TNF_SHARED)
1524 /* Update priv based on whether false sharing was detected */
1527 if (join && !get_numa_group(grp))
1535 double_lock(&my_grp->lock, &grp->lock);
1537 for (i = 0; i < 2*nr_node_ids; i++) {
1538 my_grp->faults[i] -= p->numa_faults[i];
1539 grp->faults[i] += p->numa_faults[i];
1541 my_grp->total_faults -= p->total_numa_faults;
1542 grp->total_faults += p->total_numa_faults;
1544 list_move(&p->numa_entry, &grp->task_list);
1548 spin_unlock(&my_grp->lock);
1549 spin_unlock(&grp->lock);
1551 rcu_assign_pointer(p->numa_group, grp);
1553 put_numa_group(my_grp);
1561 void task_numa_free(struct task_struct *p)
1563 struct numa_group *grp = p->numa_group;
1565 void *numa_faults = p->numa_faults;
1568 spin_lock(&grp->lock);
1569 for (i = 0; i < 2*nr_node_ids; i++)
1570 grp->faults[i] -= p->numa_faults[i];
1571 grp->total_faults -= p->total_numa_faults;
1573 list_del(&p->numa_entry);
1575 spin_unlock(&grp->lock);
1576 rcu_assign_pointer(p->numa_group, NULL);
1577 put_numa_group(grp);
1580 p->numa_faults = NULL;
1581 p->numa_faults_buffer = NULL;
1586 * Got a PROT_NONE fault for a page on @node.
1588 void task_numa_fault(int last_cpupid, int node, int pages, int flags)
1590 struct task_struct *p = current;
1591 bool migrated = flags & TNF_MIGRATED;
1594 if (!numabalancing_enabled)
1597 /* for example, ksmd faulting in a user's mm */
1601 /* Do not worry about placement if exiting */
1602 if (p->state == TASK_DEAD)
1605 /* Allocate buffer to track faults on a per-node basis */
1606 if (unlikely(!p->numa_faults)) {
1607 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1609 /* numa_faults and numa_faults_buffer share the allocation */
1610 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1611 if (!p->numa_faults)
1614 BUG_ON(p->numa_faults_buffer);
1615 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1616 p->total_numa_faults = 0;
1617 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1621 * First accesses are treated as private, otherwise consider accesses
1622 * to be private if the accessing pid has not changed
1624 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1627 priv = cpupid_match_pid(p, last_cpupid);
1628 if (!priv && !(flags & TNF_NO_GROUP))
1629 task_numa_group(p, last_cpupid, flags, &priv);
1632 task_numa_placement(p);
1635 * Retry task to preferred node migration periodically, in case it
1636 * case it previously failed, or the scheduler moved us.
1638 if (time_after(jiffies, p->numa_migrate_retry))
1639 numa_migrate_preferred(p);
1642 p->numa_pages_migrated += pages;
1644 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1645 p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages;
1648 static void reset_ptenuma_scan(struct task_struct *p)
1650 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1651 p->mm->numa_scan_offset = 0;
1655 * The expensive part of numa migration is done from task_work context.
1656 * Triggered from task_tick_numa().
1658 void task_numa_work(struct callback_head *work)
1660 unsigned long migrate, next_scan, now = jiffies;
1661 struct task_struct *p = current;
1662 struct mm_struct *mm = p->mm;
1663 struct vm_area_struct *vma;
1664 unsigned long start, end;
1665 unsigned long nr_pte_updates = 0;
1668 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1670 work->next = work; /* protect against double add */
1672 * Who cares about NUMA placement when they're dying.
1674 * NOTE: make sure not to dereference p->mm before this check,
1675 * exit_task_work() happens _after_ exit_mm() so we could be called
1676 * without p->mm even though we still had it when we enqueued this
1679 if (p->flags & PF_EXITING)
1682 if (!mm->numa_next_scan) {
1683 mm->numa_next_scan = now +
1684 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1688 * Enforce maximal scan/migration frequency..
1690 migrate = mm->numa_next_scan;
1691 if (time_before(now, migrate))
1694 if (p->numa_scan_period == 0) {
1695 p->numa_scan_period_max = task_scan_max(p);
1696 p->numa_scan_period = task_scan_min(p);
1699 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1700 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1704 * Delay this task enough that another task of this mm will likely win
1705 * the next time around.
1707 p->node_stamp += 2 * TICK_NSEC;
1709 start = mm->numa_scan_offset;
1710 pages = sysctl_numa_balancing_scan_size;
1711 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1715 down_read(&mm->mmap_sem);
1716 vma = find_vma(mm, start);
1718 reset_ptenuma_scan(p);
1722 for (; vma; vma = vma->vm_next) {
1723 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1727 * Shared library pages mapped by multiple processes are not
1728 * migrated as it is expected they are cache replicated. Avoid
1729 * hinting faults in read-only file-backed mappings or the vdso
1730 * as migrating the pages will be of marginal benefit.
1733 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1737 * Skip inaccessible VMAs to avoid any confusion between
1738 * PROT_NONE and NUMA hinting ptes
1740 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1744 start = max(start, vma->vm_start);
1745 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1746 end = min(end, vma->vm_end);
1747 nr_pte_updates += change_prot_numa(vma, start, end);
1750 * Scan sysctl_numa_balancing_scan_size but ensure that
1751 * at least one PTE is updated so that unused virtual
1752 * address space is quickly skipped.
1755 pages -= (end - start) >> PAGE_SHIFT;
1760 } while (end != vma->vm_end);
1765 * It is possible to reach the end of the VMA list but the last few
1766 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1767 * would find the !migratable VMA on the next scan but not reset the
1768 * scanner to the start so check it now.
1771 mm->numa_scan_offset = start;
1773 reset_ptenuma_scan(p);
1774 up_read(&mm->mmap_sem);
1778 * Drive the periodic memory faults..
1780 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1782 struct callback_head *work = &curr->numa_work;
1786 * We don't care about NUMA placement if we don't have memory.
1788 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1792 * Using runtime rather than walltime has the dual advantage that
1793 * we (mostly) drive the selection from busy threads and that the
1794 * task needs to have done some actual work before we bother with
1797 now = curr->se.sum_exec_runtime;
1798 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1800 if (now - curr->node_stamp > period) {
1801 if (!curr->node_stamp)
1802 curr->numa_scan_period = task_scan_min(curr);
1803 curr->node_stamp += period;
1805 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1806 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1807 task_work_add(curr, work, true);
1812 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1816 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1820 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1823 #endif /* CONFIG_NUMA_BALANCING */
1826 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1828 update_load_add(&cfs_rq->load, se->load.weight);
1829 if (!parent_entity(se))
1830 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1832 if (entity_is_task(se)) {
1833 struct rq *rq = rq_of(cfs_rq);
1835 account_numa_enqueue(rq, task_of(se));
1836 list_add(&se->group_node, &rq->cfs_tasks);
1839 cfs_rq->nr_running++;
1843 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1845 update_load_sub(&cfs_rq->load, se->load.weight);
1846 if (!parent_entity(se))
1847 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1848 if (entity_is_task(se)) {
1849 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
1850 list_del_init(&se->group_node);
1852 cfs_rq->nr_running--;
1855 #ifdef CONFIG_FAIR_GROUP_SCHED
1857 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1862 * Use this CPU's actual weight instead of the last load_contribution
1863 * to gain a more accurate current total weight. See
1864 * update_cfs_rq_load_contribution().
1866 tg_weight = atomic_long_read(&tg->load_avg);
1867 tg_weight -= cfs_rq->tg_load_contrib;
1868 tg_weight += cfs_rq->load.weight;
1873 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1875 long tg_weight, load, shares;
1877 tg_weight = calc_tg_weight(tg, cfs_rq);
1878 load = cfs_rq->load.weight;
1880 shares = (tg->shares * load);
1882 shares /= tg_weight;
1884 if (shares < MIN_SHARES)
1885 shares = MIN_SHARES;
1886 if (shares > tg->shares)
1887 shares = tg->shares;
1891 # else /* CONFIG_SMP */
1892 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1896 # endif /* CONFIG_SMP */
1897 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1898 unsigned long weight)
1901 /* commit outstanding execution time */
1902 if (cfs_rq->curr == se)
1903 update_curr(cfs_rq);
1904 account_entity_dequeue(cfs_rq, se);
1907 update_load_set(&se->load, weight);
1910 account_entity_enqueue(cfs_rq, se);
1913 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1915 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1917 struct task_group *tg;
1918 struct sched_entity *se;
1922 se = tg->se[cpu_of(rq_of(cfs_rq))];
1923 if (!se || throttled_hierarchy(cfs_rq))
1926 if (likely(se->load.weight == tg->shares))
1929 shares = calc_cfs_shares(cfs_rq, tg);
1931 reweight_entity(cfs_rq_of(se), se, shares);
1933 #else /* CONFIG_FAIR_GROUP_SCHED */
1934 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1937 #endif /* CONFIG_FAIR_GROUP_SCHED */
1941 * We choose a half-life close to 1 scheduling period.
1942 * Note: The tables below are dependent on this value.
1944 #define LOAD_AVG_PERIOD 32
1945 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1946 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1948 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1949 static const u32 runnable_avg_yN_inv[] = {
1950 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1951 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1952 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1953 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1954 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1955 0x85aac367, 0x82cd8698,
1959 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1960 * over-estimates when re-combining.
1962 static const u32 runnable_avg_yN_sum[] = {
1963 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1964 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1965 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1970 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1972 static __always_inline u64 decay_load(u64 val, u64 n)
1974 unsigned int local_n;
1978 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1981 /* after bounds checking we can collapse to 32-bit */
1985 * As y^PERIOD = 1/2, we can combine
1986 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1987 * With a look-up table which covers k^n (n<PERIOD)
1989 * To achieve constant time decay_load.
1991 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1992 val >>= local_n / LOAD_AVG_PERIOD;
1993 local_n %= LOAD_AVG_PERIOD;
1996 val *= runnable_avg_yN_inv[local_n];
1997 /* We don't use SRR here since we always want to round down. */
2002 * For updates fully spanning n periods, the contribution to runnable
2003 * average will be: \Sum 1024*y^n
2005 * We can compute this reasonably efficiently by combining:
2006 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2008 static u32 __compute_runnable_contrib(u64 n)
2012 if (likely(n <= LOAD_AVG_PERIOD))
2013 return runnable_avg_yN_sum[n];
2014 else if (unlikely(n >= LOAD_AVG_MAX_N))
2015 return LOAD_AVG_MAX;
2017 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2019 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2020 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2022 n -= LOAD_AVG_PERIOD;
2023 } while (n > LOAD_AVG_PERIOD);
2025 contrib = decay_load(contrib, n);
2026 return contrib + runnable_avg_yN_sum[n];
2030 * We can represent the historical contribution to runnable average as the
2031 * coefficients of a geometric series. To do this we sub-divide our runnable
2032 * history into segments of approximately 1ms (1024us); label the segment that
2033 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2035 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2037 * (now) (~1ms ago) (~2ms ago)
2039 * Let u_i denote the fraction of p_i that the entity was runnable.
2041 * We then designate the fractions u_i as our co-efficients, yielding the
2042 * following representation of historical load:
2043 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2045 * We choose y based on the with of a reasonably scheduling period, fixing:
2048 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2049 * approximately half as much as the contribution to load within the last ms
2052 * When a period "rolls over" and we have new u_0`, multiplying the previous
2053 * sum again by y is sufficient to update:
2054 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2055 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2057 static __always_inline int __update_entity_runnable_avg(u64 now,
2058 struct sched_avg *sa,
2062 u32 runnable_contrib;
2063 int delta_w, decayed = 0;
2065 delta = now - sa->last_runnable_update;
2067 * This should only happen when time goes backwards, which it
2068 * unfortunately does during sched clock init when we swap over to TSC.
2070 if ((s64)delta < 0) {
2071 sa->last_runnable_update = now;
2076 * Use 1024ns as the unit of measurement since it's a reasonable
2077 * approximation of 1us and fast to compute.
2082 sa->last_runnable_update = now;
2084 /* delta_w is the amount already accumulated against our next period */
2085 delta_w = sa->runnable_avg_period % 1024;
2086 if (delta + delta_w >= 1024) {
2087 /* period roll-over */
2091 * Now that we know we're crossing a period boundary, figure
2092 * out how much from delta we need to complete the current
2093 * period and accrue it.
2095 delta_w = 1024 - delta_w;
2097 sa->runnable_avg_sum += delta_w;
2098 sa->runnable_avg_period += delta_w;
2102 /* Figure out how many additional periods this update spans */
2103 periods = delta / 1024;
2106 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2108 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2111 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2112 runnable_contrib = __compute_runnable_contrib(periods);
2114 sa->runnable_avg_sum += runnable_contrib;
2115 sa->runnable_avg_period += runnable_contrib;
2118 /* Remainder of delta accrued against u_0` */
2120 sa->runnable_avg_sum += delta;
2121 sa->runnable_avg_period += delta;
2126 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2127 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2129 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2130 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2132 decays -= se->avg.decay_count;
2136 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2137 se->avg.decay_count = 0;
2142 #ifdef CONFIG_FAIR_GROUP_SCHED
2143 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2146 struct task_group *tg = cfs_rq->tg;
2149 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2150 tg_contrib -= cfs_rq->tg_load_contrib;
2152 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2153 atomic_long_add(tg_contrib, &tg->load_avg);
2154 cfs_rq->tg_load_contrib += tg_contrib;
2159 * Aggregate cfs_rq runnable averages into an equivalent task_group
2160 * representation for computing load contributions.
2162 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2163 struct cfs_rq *cfs_rq)
2165 struct task_group *tg = cfs_rq->tg;
2168 /* The fraction of a cpu used by this cfs_rq */
2169 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2170 sa->runnable_avg_period + 1);
2171 contrib -= cfs_rq->tg_runnable_contrib;
2173 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2174 atomic_add(contrib, &tg->runnable_avg);
2175 cfs_rq->tg_runnable_contrib += contrib;
2179 static inline void __update_group_entity_contrib(struct sched_entity *se)
2181 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2182 struct task_group *tg = cfs_rq->tg;
2187 contrib = cfs_rq->tg_load_contrib * tg->shares;
2188 se->avg.load_avg_contrib = div_u64(contrib,
2189 atomic_long_read(&tg->load_avg) + 1);
2192 * For group entities we need to compute a correction term in the case
2193 * that they are consuming <1 cpu so that we would contribute the same
2194 * load as a task of equal weight.
2196 * Explicitly co-ordinating this measurement would be expensive, but
2197 * fortunately the sum of each cpus contribution forms a usable
2198 * lower-bound on the true value.
2200 * Consider the aggregate of 2 contributions. Either they are disjoint
2201 * (and the sum represents true value) or they are disjoint and we are
2202 * understating by the aggregate of their overlap.
2204 * Extending this to N cpus, for a given overlap, the maximum amount we
2205 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2206 * cpus that overlap for this interval and w_i is the interval width.
2208 * On a small machine; the first term is well-bounded which bounds the
2209 * total error since w_i is a subset of the period. Whereas on a
2210 * larger machine, while this first term can be larger, if w_i is the
2211 * of consequential size guaranteed to see n_i*w_i quickly converge to
2212 * our upper bound of 1-cpu.
2214 runnable_avg = atomic_read(&tg->runnable_avg);
2215 if (runnable_avg < NICE_0_LOAD) {
2216 se->avg.load_avg_contrib *= runnable_avg;
2217 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2221 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2222 int force_update) {}
2223 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2224 struct cfs_rq *cfs_rq) {}
2225 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2228 static inline void __update_task_entity_contrib(struct sched_entity *se)
2232 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2233 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2234 contrib /= (se->avg.runnable_avg_period + 1);
2235 se->avg.load_avg_contrib = scale_load(contrib);
2238 /* Compute the current contribution to load_avg by se, return any delta */
2239 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2241 long old_contrib = se->avg.load_avg_contrib;
2243 if (entity_is_task(se)) {
2244 __update_task_entity_contrib(se);
2246 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2247 __update_group_entity_contrib(se);
2250 return se->avg.load_avg_contrib - old_contrib;
2253 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2256 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2257 cfs_rq->blocked_load_avg -= load_contrib;
2259 cfs_rq->blocked_load_avg = 0;
2262 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2264 /* Update a sched_entity's runnable average */
2265 static inline void update_entity_load_avg(struct sched_entity *se,
2268 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2273 * For a group entity we need to use their owned cfs_rq_clock_task() in
2274 * case they are the parent of a throttled hierarchy.
2276 if (entity_is_task(se))
2277 now = cfs_rq_clock_task(cfs_rq);
2279 now = cfs_rq_clock_task(group_cfs_rq(se));
2281 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2284 contrib_delta = __update_entity_load_avg_contrib(se);
2290 cfs_rq->runnable_load_avg += contrib_delta;
2292 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2296 * Decay the load contributed by all blocked children and account this so that
2297 * their contribution may appropriately discounted when they wake up.
2299 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2301 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2304 decays = now - cfs_rq->last_decay;
2305 if (!decays && !force_update)
2308 if (atomic_long_read(&cfs_rq->removed_load)) {
2309 unsigned long removed_load;
2310 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2311 subtract_blocked_load_contrib(cfs_rq, removed_load);
2315 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2317 atomic64_add(decays, &cfs_rq->decay_counter);
2318 cfs_rq->last_decay = now;
2321 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2324 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2326 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2327 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2330 /* Add the load generated by se into cfs_rq's child load-average */
2331 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2332 struct sched_entity *se,
2336 * We track migrations using entity decay_count <= 0, on a wake-up
2337 * migration we use a negative decay count to track the remote decays
2338 * accumulated while sleeping.
2340 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2341 * are seen by enqueue_entity_load_avg() as a migration with an already
2342 * constructed load_avg_contrib.
2344 if (unlikely(se->avg.decay_count <= 0)) {
2345 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2346 if (se->avg.decay_count) {
2348 * In a wake-up migration we have to approximate the
2349 * time sleeping. This is because we can't synchronize
2350 * clock_task between the two cpus, and it is not
2351 * guaranteed to be read-safe. Instead, we can
2352 * approximate this using our carried decays, which are
2353 * explicitly atomically readable.
2355 se->avg.last_runnable_update -= (-se->avg.decay_count)
2357 update_entity_load_avg(se, 0);
2358 /* Indicate that we're now synchronized and on-rq */
2359 se->avg.decay_count = 0;
2364 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2365 * would have made count negative); we must be careful to avoid
2366 * double-accounting blocked time after synchronizing decays.
2368 se->avg.last_runnable_update += __synchronize_entity_decay(se)
2372 /* migrated tasks did not contribute to our blocked load */
2374 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2375 update_entity_load_avg(se, 0);
2378 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2379 /* we force update consideration on load-balancer moves */
2380 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2384 * Remove se's load from this cfs_rq child load-average, if the entity is
2385 * transitioning to a blocked state we track its projected decay using
2388 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2389 struct sched_entity *se,
2392 update_entity_load_avg(se, 1);
2393 /* we force update consideration on load-balancer moves */
2394 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2396 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2398 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2399 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2400 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2404 * Update the rq's load with the elapsed running time before entering
2405 * idle. if the last scheduled task is not a CFS task, idle_enter will
2406 * be the only way to update the runnable statistic.
2408 void idle_enter_fair(struct rq *this_rq)
2410 update_rq_runnable_avg(this_rq, 1);
2414 * Update the rq's load with the elapsed idle time before a task is
2415 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2416 * be the only way to update the runnable statistic.
2418 void idle_exit_fair(struct rq *this_rq)
2420 update_rq_runnable_avg(this_rq, 0);
2424 static inline void update_entity_load_avg(struct sched_entity *se,
2425 int update_cfs_rq) {}
2426 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2427 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2428 struct sched_entity *se,
2430 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2431 struct sched_entity *se,
2433 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2434 int force_update) {}
2437 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2439 #ifdef CONFIG_SCHEDSTATS
2440 struct task_struct *tsk = NULL;
2442 if (entity_is_task(se))
2445 if (se->statistics.sleep_start) {
2446 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2451 if (unlikely(delta > se->statistics.sleep_max))
2452 se->statistics.sleep_max = delta;
2454 se->statistics.sleep_start = 0;
2455 se->statistics.sum_sleep_runtime += delta;
2458 account_scheduler_latency(tsk, delta >> 10, 1);
2459 trace_sched_stat_sleep(tsk, delta);
2462 if (se->statistics.block_start) {
2463 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2468 if (unlikely(delta > se->statistics.block_max))
2469 se->statistics.block_max = delta;
2471 se->statistics.block_start = 0;
2472 se->statistics.sum_sleep_runtime += delta;
2475 if (tsk->in_iowait) {
2476 se->statistics.iowait_sum += delta;
2477 se->statistics.iowait_count++;
2478 trace_sched_stat_iowait(tsk, delta);
2481 trace_sched_stat_blocked(tsk, delta);
2484 * Blocking time is in units of nanosecs, so shift by
2485 * 20 to get a milliseconds-range estimation of the
2486 * amount of time that the task spent sleeping:
2488 if (unlikely(prof_on == SLEEP_PROFILING)) {
2489 profile_hits(SLEEP_PROFILING,
2490 (void *)get_wchan(tsk),
2493 account_scheduler_latency(tsk, delta >> 10, 0);
2499 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2501 #ifdef CONFIG_SCHED_DEBUG
2502 s64 d = se->vruntime - cfs_rq->min_vruntime;
2507 if (d > 3*sysctl_sched_latency)
2508 schedstat_inc(cfs_rq, nr_spread_over);
2513 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2515 u64 vruntime = cfs_rq->min_vruntime;
2518 * The 'current' period is already promised to the current tasks,
2519 * however the extra weight of the new task will slow them down a
2520 * little, place the new task so that it fits in the slot that
2521 * stays open at the end.
2523 if (initial && sched_feat(START_DEBIT))
2524 vruntime += sched_vslice(cfs_rq, se);
2526 /* sleeps up to a single latency don't count. */
2528 unsigned long thresh = sysctl_sched_latency;
2531 * Halve their sleep time's effect, to allow
2532 * for a gentler effect of sleepers:
2534 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2540 /* ensure we never gain time by being placed backwards. */
2541 se->vruntime = max_vruntime(se->vruntime, vruntime);
2544 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2547 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2550 * Update the normalized vruntime before updating min_vruntime
2551 * through calling update_curr().
2553 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2554 se->vruntime += cfs_rq->min_vruntime;
2557 * Update run-time statistics of the 'current'.
2559 update_curr(cfs_rq);
2560 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2561 account_entity_enqueue(cfs_rq, se);
2562 update_cfs_shares(cfs_rq);
2564 if (flags & ENQUEUE_WAKEUP) {
2565 place_entity(cfs_rq, se, 0);
2566 enqueue_sleeper(cfs_rq, se);
2569 update_stats_enqueue(cfs_rq, se);
2570 check_spread(cfs_rq, se);
2571 if (se != cfs_rq->curr)
2572 __enqueue_entity(cfs_rq, se);
2575 if (cfs_rq->nr_running == 1) {
2576 list_add_leaf_cfs_rq(cfs_rq);
2577 check_enqueue_throttle(cfs_rq);
2581 static void __clear_buddies_last(struct sched_entity *se)
2583 for_each_sched_entity(se) {
2584 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2585 if (cfs_rq->last == se)
2586 cfs_rq->last = NULL;
2592 static void __clear_buddies_next(struct sched_entity *se)
2594 for_each_sched_entity(se) {
2595 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2596 if (cfs_rq->next == se)
2597 cfs_rq->next = NULL;
2603 static void __clear_buddies_skip(struct sched_entity *se)
2605 for_each_sched_entity(se) {
2606 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2607 if (cfs_rq->skip == se)
2608 cfs_rq->skip = NULL;
2614 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2616 if (cfs_rq->last == se)
2617 __clear_buddies_last(se);
2619 if (cfs_rq->next == se)
2620 __clear_buddies_next(se);
2622 if (cfs_rq->skip == se)
2623 __clear_buddies_skip(se);
2626 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2629 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2632 * Update run-time statistics of the 'current'.
2634 update_curr(cfs_rq);
2635 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2637 update_stats_dequeue(cfs_rq, se);
2638 if (flags & DEQUEUE_SLEEP) {
2639 #ifdef CONFIG_SCHEDSTATS
2640 if (entity_is_task(se)) {
2641 struct task_struct *tsk = task_of(se);
2643 if (tsk->state & TASK_INTERRUPTIBLE)
2644 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2645 if (tsk->state & TASK_UNINTERRUPTIBLE)
2646 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2651 clear_buddies(cfs_rq, se);
2653 if (se != cfs_rq->curr)
2654 __dequeue_entity(cfs_rq, se);
2656 account_entity_dequeue(cfs_rq, se);
2659 * Normalize the entity after updating the min_vruntime because the
2660 * update can refer to the ->curr item and we need to reflect this
2661 * movement in our normalized position.
2663 if (!(flags & DEQUEUE_SLEEP))
2664 se->vruntime -= cfs_rq->min_vruntime;
2666 /* return excess runtime on last dequeue */
2667 return_cfs_rq_runtime(cfs_rq);
2669 update_min_vruntime(cfs_rq);
2670 update_cfs_shares(cfs_rq);
2674 * Preempt the current task with a newly woken task if needed:
2677 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2679 unsigned long ideal_runtime, delta_exec;
2680 struct sched_entity *se;
2683 ideal_runtime = sched_slice(cfs_rq, curr);
2684 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2685 if (delta_exec > ideal_runtime) {
2686 resched_task(rq_of(cfs_rq)->curr);
2688 * The current task ran long enough, ensure it doesn't get
2689 * re-elected due to buddy favours.
2691 clear_buddies(cfs_rq, curr);
2696 * Ensure that a task that missed wakeup preemption by a
2697 * narrow margin doesn't have to wait for a full slice.
2698 * This also mitigates buddy induced latencies under load.
2700 if (delta_exec < sysctl_sched_min_granularity)
2703 se = __pick_first_entity(cfs_rq);
2704 delta = curr->vruntime - se->vruntime;
2709 if (delta > ideal_runtime)
2710 resched_task(rq_of(cfs_rq)->curr);
2714 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2716 /* 'current' is not kept within the tree. */
2719 * Any task has to be enqueued before it get to execute on
2720 * a CPU. So account for the time it spent waiting on the
2723 update_stats_wait_end(cfs_rq, se);
2724 __dequeue_entity(cfs_rq, se);
2727 update_stats_curr_start(cfs_rq, se);
2729 #ifdef CONFIG_SCHEDSTATS
2731 * Track our maximum slice length, if the CPU's load is at
2732 * least twice that of our own weight (i.e. dont track it
2733 * when there are only lesser-weight tasks around):
2735 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2736 se->statistics.slice_max = max(se->statistics.slice_max,
2737 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2740 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2744 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2747 * Pick the next process, keeping these things in mind, in this order:
2748 * 1) keep things fair between processes/task groups
2749 * 2) pick the "next" process, since someone really wants that to run
2750 * 3) pick the "last" process, for cache locality
2751 * 4) do not run the "skip" process, if something else is available
2753 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2755 struct sched_entity *se = __pick_first_entity(cfs_rq);
2756 struct sched_entity *left = se;
2759 * Avoid running the skip buddy, if running something else can
2760 * be done without getting too unfair.
2762 if (cfs_rq->skip == se) {
2763 struct sched_entity *second = __pick_next_entity(se);
2764 if (second && wakeup_preempt_entity(second, left) < 1)
2769 * Prefer last buddy, try to return the CPU to a preempted task.
2771 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2775 * Someone really wants this to run. If it's not unfair, run it.
2777 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2780 clear_buddies(cfs_rq, se);
2785 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2787 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2790 * If still on the runqueue then deactivate_task()
2791 * was not called and update_curr() has to be done:
2794 update_curr(cfs_rq);
2796 /* throttle cfs_rqs exceeding runtime */
2797 check_cfs_rq_runtime(cfs_rq);
2799 check_spread(cfs_rq, prev);
2801 update_stats_wait_start(cfs_rq, prev);
2802 /* Put 'current' back into the tree. */
2803 __enqueue_entity(cfs_rq, prev);
2804 /* in !on_rq case, update occurred at dequeue */
2805 update_entity_load_avg(prev, 1);
2807 cfs_rq->curr = NULL;
2811 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2814 * Update run-time statistics of the 'current'.
2816 update_curr(cfs_rq);
2819 * Ensure that runnable average is periodically updated.
2821 update_entity_load_avg(curr, 1);
2822 update_cfs_rq_blocked_load(cfs_rq, 1);
2823 update_cfs_shares(cfs_rq);
2825 #ifdef CONFIG_SCHED_HRTICK
2827 * queued ticks are scheduled to match the slice, so don't bother
2828 * validating it and just reschedule.
2831 resched_task(rq_of(cfs_rq)->curr);
2835 * don't let the period tick interfere with the hrtick preemption
2837 if (!sched_feat(DOUBLE_TICK) &&
2838 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2842 if (cfs_rq->nr_running > 1)
2843 check_preempt_tick(cfs_rq, curr);
2847 /**************************************************
2848 * CFS bandwidth control machinery
2851 #ifdef CONFIG_CFS_BANDWIDTH
2853 #ifdef HAVE_JUMP_LABEL
2854 static struct static_key __cfs_bandwidth_used;
2856 static inline bool cfs_bandwidth_used(void)
2858 return static_key_false(&__cfs_bandwidth_used);
2861 void cfs_bandwidth_usage_inc(void)
2863 static_key_slow_inc(&__cfs_bandwidth_used);
2866 void cfs_bandwidth_usage_dec(void)
2868 static_key_slow_dec(&__cfs_bandwidth_used);
2870 #else /* HAVE_JUMP_LABEL */
2871 static bool cfs_bandwidth_used(void)
2876 void cfs_bandwidth_usage_inc(void) {}
2877 void cfs_bandwidth_usage_dec(void) {}
2878 #endif /* HAVE_JUMP_LABEL */
2881 * default period for cfs group bandwidth.
2882 * default: 0.1s, units: nanoseconds
2884 static inline u64 default_cfs_period(void)
2886 return 100000000ULL;
2889 static inline u64 sched_cfs_bandwidth_slice(void)
2891 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2895 * Replenish runtime according to assigned quota and update expiration time.
2896 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2897 * additional synchronization around rq->lock.
2899 * requires cfs_b->lock
2901 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2905 if (cfs_b->quota == RUNTIME_INF)
2908 now = sched_clock_cpu(smp_processor_id());
2909 cfs_b->runtime = cfs_b->quota;
2910 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2913 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2915 return &tg->cfs_bandwidth;
2918 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2919 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2921 if (unlikely(cfs_rq->throttle_count))
2922 return cfs_rq->throttled_clock_task;
2924 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2927 /* returns 0 on failure to allocate runtime */
2928 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2930 struct task_group *tg = cfs_rq->tg;
2931 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2932 u64 amount = 0, min_amount, expires;
2934 /* note: this is a positive sum as runtime_remaining <= 0 */
2935 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2937 raw_spin_lock(&cfs_b->lock);
2938 if (cfs_b->quota == RUNTIME_INF)
2939 amount = min_amount;
2942 * If the bandwidth pool has become inactive, then at least one
2943 * period must have elapsed since the last consumption.
2944 * Refresh the global state and ensure bandwidth timer becomes
2947 if (!cfs_b->timer_active) {
2948 __refill_cfs_bandwidth_runtime(cfs_b);
2949 __start_cfs_bandwidth(cfs_b);
2952 if (cfs_b->runtime > 0) {
2953 amount = min(cfs_b->runtime, min_amount);
2954 cfs_b->runtime -= amount;
2958 expires = cfs_b->runtime_expires;
2959 raw_spin_unlock(&cfs_b->lock);
2961 cfs_rq->runtime_remaining += amount;
2963 * we may have advanced our local expiration to account for allowed
2964 * spread between our sched_clock and the one on which runtime was
2967 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2968 cfs_rq->runtime_expires = expires;
2970 return cfs_rq->runtime_remaining > 0;
2974 * Note: This depends on the synchronization provided by sched_clock and the
2975 * fact that rq->clock snapshots this value.
2977 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2979 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2981 /* if the deadline is ahead of our clock, nothing to do */
2982 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2985 if (cfs_rq->runtime_remaining < 0)
2989 * If the local deadline has passed we have to consider the
2990 * possibility that our sched_clock is 'fast' and the global deadline
2991 * has not truly expired.
2993 * Fortunately we can check determine whether this the case by checking
2994 * whether the global deadline has advanced.
2997 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2998 /* extend local deadline, drift is bounded above by 2 ticks */
2999 cfs_rq->runtime_expires += TICK_NSEC;
3001 /* global deadline is ahead, expiration has passed */
3002 cfs_rq->runtime_remaining = 0;
3006 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3008 /* dock delta_exec before expiring quota (as it could span periods) */
3009 cfs_rq->runtime_remaining -= delta_exec;
3010 expire_cfs_rq_runtime(cfs_rq);
3012 if (likely(cfs_rq->runtime_remaining > 0))
3016 * if we're unable to extend our runtime we resched so that the active
3017 * hierarchy can be throttled
3019 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3020 resched_task(rq_of(cfs_rq)->curr);
3023 static __always_inline
3024 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3026 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3029 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3032 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3034 return cfs_bandwidth_used() && cfs_rq->throttled;
3037 /* check whether cfs_rq, or any parent, is throttled */
3038 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3040 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3044 * Ensure that neither of the group entities corresponding to src_cpu or
3045 * dest_cpu are members of a throttled hierarchy when performing group
3046 * load-balance operations.
3048 static inline int throttled_lb_pair(struct task_group *tg,
3049 int src_cpu, int dest_cpu)
3051 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3053 src_cfs_rq = tg->cfs_rq[src_cpu];
3054 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3056 return throttled_hierarchy(src_cfs_rq) ||
3057 throttled_hierarchy(dest_cfs_rq);
3060 /* updated child weight may affect parent so we have to do this bottom up */
3061 static int tg_unthrottle_up(struct task_group *tg, void *data)
3063 struct rq *rq = data;
3064 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3066 cfs_rq->throttle_count--;
3068 if (!cfs_rq->throttle_count) {
3069 /* adjust cfs_rq_clock_task() */
3070 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3071 cfs_rq->throttled_clock_task;
3078 static int tg_throttle_down(struct task_group *tg, void *data)
3080 struct rq *rq = data;
3081 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3083 /* group is entering throttled state, stop time */
3084 if (!cfs_rq->throttle_count)
3085 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3086 cfs_rq->throttle_count++;
3091 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3093 struct rq *rq = rq_of(cfs_rq);
3094 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3095 struct sched_entity *se;
3096 long task_delta, dequeue = 1;
3098 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3100 /* freeze hierarchy runnable averages while throttled */
3102 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3105 task_delta = cfs_rq->h_nr_running;
3106 for_each_sched_entity(se) {
3107 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3108 /* throttled entity or throttle-on-deactivate */
3113 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3114 qcfs_rq->h_nr_running -= task_delta;
3116 if (qcfs_rq->load.weight)
3121 rq->nr_running -= task_delta;
3123 cfs_rq->throttled = 1;
3124 cfs_rq->throttled_clock = rq_clock(rq);
3125 raw_spin_lock(&cfs_b->lock);
3126 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3127 if (!cfs_b->timer_active)
3128 __start_cfs_bandwidth(cfs_b);
3129 raw_spin_unlock(&cfs_b->lock);
3132 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3134 struct rq *rq = rq_of(cfs_rq);
3135 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3136 struct sched_entity *se;
3140 se = cfs_rq->tg->se[cpu_of(rq)];
3142 cfs_rq->throttled = 0;
3144 update_rq_clock(rq);
3146 raw_spin_lock(&cfs_b->lock);
3147 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3148 list_del_rcu(&cfs_rq->throttled_list);
3149 raw_spin_unlock(&cfs_b->lock);
3151 /* update hierarchical throttle state */
3152 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3154 if (!cfs_rq->load.weight)
3157 task_delta = cfs_rq->h_nr_running;
3158 for_each_sched_entity(se) {
3162 cfs_rq = cfs_rq_of(se);
3164 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3165 cfs_rq->h_nr_running += task_delta;
3167 if (cfs_rq_throttled(cfs_rq))
3172 rq->nr_running += task_delta;
3174 /* determine whether we need to wake up potentially idle cpu */
3175 if (rq->curr == rq->idle && rq->cfs.nr_running)
3176 resched_task(rq->curr);
3179 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3180 u64 remaining, u64 expires)
3182 struct cfs_rq *cfs_rq;
3183 u64 runtime = remaining;
3186 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3188 struct rq *rq = rq_of(cfs_rq);
3190 raw_spin_lock(&rq->lock);
3191 if (!cfs_rq_throttled(cfs_rq))
3194 runtime = -cfs_rq->runtime_remaining + 1;
3195 if (runtime > remaining)
3196 runtime = remaining;
3197 remaining -= runtime;
3199 cfs_rq->runtime_remaining += runtime;
3200 cfs_rq->runtime_expires = expires;
3202 /* we check whether we're throttled above */
3203 if (cfs_rq->runtime_remaining > 0)
3204 unthrottle_cfs_rq(cfs_rq);
3207 raw_spin_unlock(&rq->lock);
3218 * Responsible for refilling a task_group's bandwidth and unthrottling its
3219 * cfs_rqs as appropriate. If there has been no activity within the last
3220 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3221 * used to track this state.
3223 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3225 u64 runtime, runtime_expires;
3226 int idle = 1, throttled;
3228 raw_spin_lock(&cfs_b->lock);
3229 /* no need to continue the timer with no bandwidth constraint */
3230 if (cfs_b->quota == RUNTIME_INF)
3233 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3234 /* idle depends on !throttled (for the case of a large deficit) */
3235 idle = cfs_b->idle && !throttled;
3236 cfs_b->nr_periods += overrun;
3238 /* if we're going inactive then everything else can be deferred */
3243 * if we have relooped after returning idle once, we need to update our
3244 * status as actually running, so that other cpus doing
3245 * __start_cfs_bandwidth will stop trying to cancel us.
3247 cfs_b->timer_active = 1;
3249 __refill_cfs_bandwidth_runtime(cfs_b);
3252 /* mark as potentially idle for the upcoming period */
3257 /* account preceding periods in which throttling occurred */
3258 cfs_b->nr_throttled += overrun;
3261 * There are throttled entities so we must first use the new bandwidth
3262 * to unthrottle them before making it generally available. This
3263 * ensures that all existing debts will be paid before a new cfs_rq is
3266 runtime = cfs_b->runtime;
3267 runtime_expires = cfs_b->runtime_expires;
3271 * This check is repeated as we are holding onto the new bandwidth
3272 * while we unthrottle. This can potentially race with an unthrottled
3273 * group trying to acquire new bandwidth from the global pool.
3275 while (throttled && runtime > 0) {
3276 raw_spin_unlock(&cfs_b->lock);
3277 /* we can't nest cfs_b->lock while distributing bandwidth */
3278 runtime = distribute_cfs_runtime(cfs_b, runtime,
3280 raw_spin_lock(&cfs_b->lock);
3282 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3285 /* return (any) remaining runtime */
3286 cfs_b->runtime = runtime;
3288 * While we are ensured activity in the period following an
3289 * unthrottle, this also covers the case in which the new bandwidth is
3290 * insufficient to cover the existing bandwidth deficit. (Forcing the
3291 * timer to remain active while there are any throttled entities.)
3296 cfs_b->timer_active = 0;
3297 raw_spin_unlock(&cfs_b->lock);
3302 /* a cfs_rq won't donate quota below this amount */
3303 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3304 /* minimum remaining period time to redistribute slack quota */
3305 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3306 /* how long we wait to gather additional slack before distributing */
3307 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3310 * Are we near the end of the current quota period?
3312 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3313 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3314 * migrate_hrtimers, base is never cleared, so we are fine.
3316 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3318 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3321 /* if the call-back is running a quota refresh is already occurring */
3322 if (hrtimer_callback_running(refresh_timer))
3325 /* is a quota refresh about to occur? */
3326 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3327 if (remaining < min_expire)
3333 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3335 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3337 /* if there's a quota refresh soon don't bother with slack */
3338 if (runtime_refresh_within(cfs_b, min_left))
3341 start_bandwidth_timer(&cfs_b->slack_timer,
3342 ns_to_ktime(cfs_bandwidth_slack_period));
3345 /* we know any runtime found here is valid as update_curr() precedes return */
3346 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3348 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3349 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3351 if (slack_runtime <= 0)
3354 raw_spin_lock(&cfs_b->lock);
3355 if (cfs_b->quota != RUNTIME_INF &&
3356 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3357 cfs_b->runtime += slack_runtime;
3359 /* we are under rq->lock, defer unthrottling using a timer */
3360 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3361 !list_empty(&cfs_b->throttled_cfs_rq))
3362 start_cfs_slack_bandwidth(cfs_b);
3364 raw_spin_unlock(&cfs_b->lock);
3366 /* even if it's not valid for return we don't want to try again */
3367 cfs_rq->runtime_remaining -= slack_runtime;
3370 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3372 if (!cfs_bandwidth_used())
3375 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3378 __return_cfs_rq_runtime(cfs_rq);
3382 * This is done with a timer (instead of inline with bandwidth return) since
3383 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3385 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3387 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3390 /* confirm we're still not at a refresh boundary */
3391 raw_spin_lock(&cfs_b->lock);
3392 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3393 raw_spin_unlock(&cfs_b->lock);
3397 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3398 runtime = cfs_b->runtime;
3401 expires = cfs_b->runtime_expires;
3402 raw_spin_unlock(&cfs_b->lock);
3407 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3409 raw_spin_lock(&cfs_b->lock);
3410 if (expires == cfs_b->runtime_expires)
3411 cfs_b->runtime = runtime;
3412 raw_spin_unlock(&cfs_b->lock);
3416 * When a group wakes up we want to make sure that its quota is not already
3417 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3418 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3420 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3422 if (!cfs_bandwidth_used())
3425 /* an active group must be handled by the update_curr()->put() path */
3426 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3429 /* ensure the group is not already throttled */
3430 if (cfs_rq_throttled(cfs_rq))
3433 /* update runtime allocation */
3434 account_cfs_rq_runtime(cfs_rq, 0);
3435 if (cfs_rq->runtime_remaining <= 0)
3436 throttle_cfs_rq(cfs_rq);
3439 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3440 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3442 if (!cfs_bandwidth_used())
3445 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3449 * it's possible for a throttled entity to be forced into a running
3450 * state (e.g. set_curr_task), in this case we're finished.
3452 if (cfs_rq_throttled(cfs_rq))
3455 throttle_cfs_rq(cfs_rq);
3458 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3460 struct cfs_bandwidth *cfs_b =
3461 container_of(timer, struct cfs_bandwidth, slack_timer);
3462 do_sched_cfs_slack_timer(cfs_b);
3464 return HRTIMER_NORESTART;
3467 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3469 struct cfs_bandwidth *cfs_b =
3470 container_of(timer, struct cfs_bandwidth, period_timer);
3476 now = hrtimer_cb_get_time(timer);
3477 overrun = hrtimer_forward(timer, now, cfs_b->period);
3482 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3485 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3488 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3490 raw_spin_lock_init(&cfs_b->lock);
3492 cfs_b->quota = RUNTIME_INF;
3493 cfs_b->period = ns_to_ktime(default_cfs_period());
3495 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3496 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3497 cfs_b->period_timer.function = sched_cfs_period_timer;
3498 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3499 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3502 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3504 cfs_rq->runtime_enabled = 0;
3505 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3508 /* requires cfs_b->lock, may release to reprogram timer */
3509 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3512 * The timer may be active because we're trying to set a new bandwidth
3513 * period or because we're racing with the tear-down path
3514 * (timer_active==0 becomes visible before the hrtimer call-back
3515 * terminates). In either case we ensure that it's re-programmed
3517 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3518 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3519 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3520 raw_spin_unlock(&cfs_b->lock);
3522 raw_spin_lock(&cfs_b->lock);
3523 /* if someone else restarted the timer then we're done */
3524 if (cfs_b->timer_active)
3528 cfs_b->timer_active = 1;
3529 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3532 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3534 hrtimer_cancel(&cfs_b->period_timer);
3535 hrtimer_cancel(&cfs_b->slack_timer);
3538 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3540 struct cfs_rq *cfs_rq;
3542 for_each_leaf_cfs_rq(rq, cfs_rq) {
3543 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3545 if (!cfs_rq->runtime_enabled)
3549 * clock_task is not advancing so we just need to make sure
3550 * there's some valid quota amount
3552 cfs_rq->runtime_remaining = cfs_b->quota;
3553 if (cfs_rq_throttled(cfs_rq))
3554 unthrottle_cfs_rq(cfs_rq);
3558 #else /* CONFIG_CFS_BANDWIDTH */
3559 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3561 return rq_clock_task(rq_of(cfs_rq));
3564 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3565 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3566 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3567 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3569 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3574 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3579 static inline int throttled_lb_pair(struct task_group *tg,
3580 int src_cpu, int dest_cpu)
3585 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3587 #ifdef CONFIG_FAIR_GROUP_SCHED
3588 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3591 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3595 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3596 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3598 #endif /* CONFIG_CFS_BANDWIDTH */
3600 /**************************************************
3601 * CFS operations on tasks:
3604 #ifdef CONFIG_SCHED_HRTICK
3605 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3607 struct sched_entity *se = &p->se;
3608 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3610 WARN_ON(task_rq(p) != rq);
3612 if (cfs_rq->nr_running > 1) {
3613 u64 slice = sched_slice(cfs_rq, se);
3614 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3615 s64 delta = slice - ran;
3624 * Don't schedule slices shorter than 10000ns, that just
3625 * doesn't make sense. Rely on vruntime for fairness.
3628 delta = max_t(s64, 10000LL, delta);
3630 hrtick_start(rq, delta);
3635 * called from enqueue/dequeue and updates the hrtick when the
3636 * current task is from our class and nr_running is low enough
3639 static void hrtick_update(struct rq *rq)
3641 struct task_struct *curr = rq->curr;
3643 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3646 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3647 hrtick_start_fair(rq, curr);
3649 #else /* !CONFIG_SCHED_HRTICK */
3651 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3655 static inline void hrtick_update(struct rq *rq)
3661 * The enqueue_task method is called before nr_running is
3662 * increased. Here we update the fair scheduling stats and
3663 * then put the task into the rbtree:
3666 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3668 struct cfs_rq *cfs_rq;
3669 struct sched_entity *se = &p->se;
3671 for_each_sched_entity(se) {
3674 cfs_rq = cfs_rq_of(se);
3675 enqueue_entity(cfs_rq, se, flags);
3678 * end evaluation on encountering a throttled cfs_rq
3680 * note: in the case of encountering a throttled cfs_rq we will
3681 * post the final h_nr_running increment below.
3683 if (cfs_rq_throttled(cfs_rq))
3685 cfs_rq->h_nr_running++;
3687 flags = ENQUEUE_WAKEUP;
3690 for_each_sched_entity(se) {
3691 cfs_rq = cfs_rq_of(se);
3692 cfs_rq->h_nr_running++;
3694 if (cfs_rq_throttled(cfs_rq))
3697 update_cfs_shares(cfs_rq);
3698 update_entity_load_avg(se, 1);
3702 update_rq_runnable_avg(rq, rq->nr_running);
3708 static void set_next_buddy(struct sched_entity *se);
3711 * The dequeue_task method is called before nr_running is
3712 * decreased. We remove the task from the rbtree and
3713 * update the fair scheduling stats:
3715 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3717 struct cfs_rq *cfs_rq;
3718 struct sched_entity *se = &p->se;
3719 int task_sleep = flags & DEQUEUE_SLEEP;
3721 for_each_sched_entity(se) {
3722 cfs_rq = cfs_rq_of(se);
3723 dequeue_entity(cfs_rq, se, flags);
3726 * end evaluation on encountering a throttled cfs_rq
3728 * note: in the case of encountering a throttled cfs_rq we will
3729 * post the final h_nr_running decrement below.
3731 if (cfs_rq_throttled(cfs_rq))
3733 cfs_rq->h_nr_running--;
3735 /* Don't dequeue parent if it has other entities besides us */
3736 if (cfs_rq->load.weight) {
3738 * Bias pick_next to pick a task from this cfs_rq, as
3739 * p is sleeping when it is within its sched_slice.
3741 if (task_sleep && parent_entity(se))
3742 set_next_buddy(parent_entity(se));
3744 /* avoid re-evaluating load for this entity */
3745 se = parent_entity(se);
3748 flags |= DEQUEUE_SLEEP;
3751 for_each_sched_entity(se) {
3752 cfs_rq = cfs_rq_of(se);
3753 cfs_rq->h_nr_running--;
3755 if (cfs_rq_throttled(cfs_rq))
3758 update_cfs_shares(cfs_rq);
3759 update_entity_load_avg(se, 1);
3764 update_rq_runnable_avg(rq, 1);
3770 /* Used instead of source_load when we know the type == 0 */
3771 static unsigned long weighted_cpuload(const int cpu)
3773 return cpu_rq(cpu)->cfs.runnable_load_avg;
3777 * Return a low guess at the load of a migration-source cpu weighted
3778 * according to the scheduling class and "nice" value.
3780 * We want to under-estimate the load of migration sources, to
3781 * balance conservatively.
3783 static unsigned long source_load(int cpu, int type)
3785 struct rq *rq = cpu_rq(cpu);
3786 unsigned long total = weighted_cpuload(cpu);
3788 if (type == 0 || !sched_feat(LB_BIAS))
3791 return min(rq->cpu_load[type-1], total);
3795 * Return a high guess at the load of a migration-target cpu weighted
3796 * according to the scheduling class and "nice" value.
3798 static unsigned long target_load(int cpu, int type)
3800 struct rq *rq = cpu_rq(cpu);
3801 unsigned long total = weighted_cpuload(cpu);
3803 if (type == 0 || !sched_feat(LB_BIAS))
3806 return max(rq->cpu_load[type-1], total);
3809 static unsigned long power_of(int cpu)
3811 return cpu_rq(cpu)->cpu_power;
3814 static unsigned long cpu_avg_load_per_task(int cpu)
3816 struct rq *rq = cpu_rq(cpu);
3817 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3818 unsigned long load_avg = rq->cfs.runnable_load_avg;
3821 return load_avg / nr_running;
3826 static void record_wakee(struct task_struct *p)
3829 * Rough decay (wiping) for cost saving, don't worry
3830 * about the boundary, really active task won't care
3833 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3834 current->wakee_flips = 0;
3835 current->wakee_flip_decay_ts = jiffies;
3838 if (current->last_wakee != p) {
3839 current->last_wakee = p;
3840 current->wakee_flips++;
3844 static void task_waking_fair(struct task_struct *p)
3846 struct sched_entity *se = &p->se;
3847 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3850 #ifndef CONFIG_64BIT
3851 u64 min_vruntime_copy;
3854 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3856 min_vruntime = cfs_rq->min_vruntime;
3857 } while (min_vruntime != min_vruntime_copy);
3859 min_vruntime = cfs_rq->min_vruntime;
3862 se->vruntime -= min_vruntime;
3866 #ifdef CONFIG_FAIR_GROUP_SCHED
3868 * effective_load() calculates the load change as seen from the root_task_group
3870 * Adding load to a group doesn't make a group heavier, but can cause movement
3871 * of group shares between cpus. Assuming the shares were perfectly aligned one
3872 * can calculate the shift in shares.
3874 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3875 * on this @cpu and results in a total addition (subtraction) of @wg to the
3876 * total group weight.
3878 * Given a runqueue weight distribution (rw_i) we can compute a shares
3879 * distribution (s_i) using:
3881 * s_i = rw_i / \Sum rw_j (1)
3883 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3884 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3885 * shares distribution (s_i):
3887 * rw_i = { 2, 4, 1, 0 }
3888 * s_i = { 2/7, 4/7, 1/7, 0 }
3890 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3891 * task used to run on and the CPU the waker is running on), we need to
3892 * compute the effect of waking a task on either CPU and, in case of a sync
3893 * wakeup, compute the effect of the current task going to sleep.
3895 * So for a change of @wl to the local @cpu with an overall group weight change
3896 * of @wl we can compute the new shares distribution (s'_i) using:
3898 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3900 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3901 * differences in waking a task to CPU 0. The additional task changes the
3902 * weight and shares distributions like:
3904 * rw'_i = { 3, 4, 1, 0 }
3905 * s'_i = { 3/8, 4/8, 1/8, 0 }
3907 * We can then compute the difference in effective weight by using:
3909 * dw_i = S * (s'_i - s_i) (3)
3911 * Where 'S' is the group weight as seen by its parent.
3913 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3914 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3915 * 4/7) times the weight of the group.
3917 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3919 struct sched_entity *se = tg->se[cpu];
3921 if (!tg->parent) /* the trivial, non-cgroup case */
3924 for_each_sched_entity(se) {
3930 * W = @wg + \Sum rw_j
3932 W = wg + calc_tg_weight(tg, se->my_q);
3937 w = se->my_q->load.weight + wl;
3940 * wl = S * s'_i; see (2)
3943 wl = (w * tg->shares) / W;
3948 * Per the above, wl is the new se->load.weight value; since
3949 * those are clipped to [MIN_SHARES, ...) do so now. See
3950 * calc_cfs_shares().
3952 if (wl < MIN_SHARES)
3956 * wl = dw_i = S * (s'_i - s_i); see (3)
3958 wl -= se->load.weight;
3961 * Recursively apply this logic to all parent groups to compute
3962 * the final effective load change on the root group. Since
3963 * only the @tg group gets extra weight, all parent groups can
3964 * only redistribute existing shares. @wl is the shift in shares
3965 * resulting from this level per the above.
3974 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3981 static int wake_wide(struct task_struct *p)
3983 int factor = this_cpu_read(sd_llc_size);
3986 * Yeah, it's the switching-frequency, could means many wakee or
3987 * rapidly switch, use factor here will just help to automatically
3988 * adjust the loose-degree, so bigger node will lead to more pull.
3990 if (p->wakee_flips > factor) {
3992 * wakee is somewhat hot, it needs certain amount of cpu
3993 * resource, so if waker is far more hot, prefer to leave
3996 if (current->wakee_flips > (factor * p->wakee_flips))
4003 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4005 s64 this_load, load;
4006 int idx, this_cpu, prev_cpu;
4007 unsigned long tl_per_task;
4008 struct task_group *tg;
4009 unsigned long weight;
4013 * If we wake multiple tasks be careful to not bounce
4014 * ourselves around too much.
4020 this_cpu = smp_processor_id();
4021 prev_cpu = task_cpu(p);
4022 load = source_load(prev_cpu, idx);
4023 this_load = target_load(this_cpu, idx);
4026 * If sync wakeup then subtract the (maximum possible)
4027 * effect of the currently running task from the load
4028 * of the current CPU:
4031 tg = task_group(current);
4032 weight = current->se.load.weight;
4034 this_load += effective_load(tg, this_cpu, -weight, -weight);
4035 load += effective_load(tg, prev_cpu, 0, -weight);
4039 weight = p->se.load.weight;
4042 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4043 * due to the sync cause above having dropped this_load to 0, we'll
4044 * always have an imbalance, but there's really nothing you can do
4045 * about that, so that's good too.
4047 * Otherwise check if either cpus are near enough in load to allow this
4048 * task to be woken on this_cpu.
4050 if (this_load > 0) {
4051 s64 this_eff_load, prev_eff_load;
4053 this_eff_load = 100;
4054 this_eff_load *= power_of(prev_cpu);
4055 this_eff_load *= this_load +
4056 effective_load(tg, this_cpu, weight, weight);
4058 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4059 prev_eff_load *= power_of(this_cpu);
4060 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4062 balanced = this_eff_load <= prev_eff_load;
4067 * If the currently running task will sleep within
4068 * a reasonable amount of time then attract this newly
4071 if (sync && balanced)
4074 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4075 tl_per_task = cpu_avg_load_per_task(this_cpu);
4078 (this_load <= load &&
4079 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4081 * This domain has SD_WAKE_AFFINE and
4082 * p is cache cold in this domain, and
4083 * there is no bad imbalance.
4085 schedstat_inc(sd, ttwu_move_affine);
4086 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4094 * find_idlest_group finds and returns the least busy CPU group within the
4097 static struct sched_group *
4098 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4099 int this_cpu, int sd_flag)
4101 struct sched_group *idlest = NULL, *group = sd->groups;
4102 unsigned long min_load = ULONG_MAX, this_load = 0;
4103 int load_idx = sd->forkexec_idx;
4104 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4106 if (sd_flag & SD_BALANCE_WAKE)
4107 load_idx = sd->wake_idx;
4110 unsigned long load, avg_load;
4114 /* Skip over this group if it has no CPUs allowed */
4115 if (!cpumask_intersects(sched_group_cpus(group),
4116 tsk_cpus_allowed(p)))
4119 local_group = cpumask_test_cpu(this_cpu,
4120 sched_group_cpus(group));
4122 /* Tally up the load of all CPUs in the group */
4125 for_each_cpu(i, sched_group_cpus(group)) {
4126 /* Bias balancing toward cpus of our domain */
4128 load = source_load(i, load_idx);
4130 load = target_load(i, load_idx);
4135 /* Adjust by relative CPU power of the group */
4136 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4139 this_load = avg_load;
4140 } else if (avg_load < min_load) {
4141 min_load = avg_load;
4144 } while (group = group->next, group != sd->groups);
4146 if (!idlest || 100*this_load < imbalance*min_load)
4152 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4155 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4157 unsigned long load, min_load = ULONG_MAX;
4161 /* Traverse only the allowed CPUs */
4162 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4163 load = weighted_cpuload(i);
4165 if (load < min_load || (load == min_load && i == this_cpu)) {
4175 * Try and locate an idle CPU in the sched_domain.
4177 static int select_idle_sibling(struct task_struct *p, int target)
4179 struct sched_domain *sd;
4180 struct sched_group *sg;
4181 int i = task_cpu(p);
4183 if (idle_cpu(target))
4187 * If the prevous cpu is cache affine and idle, don't be stupid.
4189 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4193 * Otherwise, iterate the domains and find an elegible idle cpu.
4195 sd = rcu_dereference(per_cpu(sd_llc, target));
4196 for_each_lower_domain(sd) {
4199 if (!cpumask_intersects(sched_group_cpus(sg),
4200 tsk_cpus_allowed(p)))
4203 for_each_cpu(i, sched_group_cpus(sg)) {
4204 if (i == target || !idle_cpu(i))
4208 target = cpumask_first_and(sched_group_cpus(sg),
4209 tsk_cpus_allowed(p));
4213 } while (sg != sd->groups);
4220 * sched_balance_self: balance the current task (running on cpu) in domains
4221 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4224 * Balance, ie. select the least loaded group.
4226 * Returns the target CPU number, or the same CPU if no balancing is needed.
4228 * preempt must be disabled.
4231 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4233 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4234 int cpu = smp_processor_id();
4236 int want_affine = 0;
4237 int sync = wake_flags & WF_SYNC;
4239 if (p->nr_cpus_allowed == 1)
4242 if (sd_flag & SD_BALANCE_WAKE) {
4243 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4249 for_each_domain(cpu, tmp) {
4250 if (!(tmp->flags & SD_LOAD_BALANCE))
4254 * If both cpu and prev_cpu are part of this domain,
4255 * cpu is a valid SD_WAKE_AFFINE target.
4257 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4258 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4263 if (tmp->flags & sd_flag)
4268 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4271 new_cpu = select_idle_sibling(p, prev_cpu);
4276 struct sched_group *group;
4279 if (!(sd->flags & sd_flag)) {
4284 group = find_idlest_group(sd, p, cpu, sd_flag);
4290 new_cpu = find_idlest_cpu(group, p, cpu);
4291 if (new_cpu == -1 || new_cpu == cpu) {
4292 /* Now try balancing at a lower domain level of cpu */
4297 /* Now try balancing at a lower domain level of new_cpu */
4299 weight = sd->span_weight;
4301 for_each_domain(cpu, tmp) {
4302 if (weight <= tmp->span_weight)
4304 if (tmp->flags & sd_flag)
4307 /* while loop will break here if sd == NULL */
4316 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4317 * cfs_rq_of(p) references at time of call are still valid and identify the
4318 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4319 * other assumptions, including the state of rq->lock, should be made.
4322 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4324 struct sched_entity *se = &p->se;
4325 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4328 * Load tracking: accumulate removed load so that it can be processed
4329 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4330 * to blocked load iff they have a positive decay-count. It can never
4331 * be negative here since on-rq tasks have decay-count == 0.
4333 if (se->avg.decay_count) {
4334 se->avg.decay_count = -__synchronize_entity_decay(se);
4335 atomic_long_add(se->avg.load_avg_contrib,
4336 &cfs_rq->removed_load);
4339 #endif /* CONFIG_SMP */
4341 static unsigned long
4342 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4344 unsigned long gran = sysctl_sched_wakeup_granularity;
4347 * Since its curr running now, convert the gran from real-time
4348 * to virtual-time in his units.
4350 * By using 'se' instead of 'curr' we penalize light tasks, so
4351 * they get preempted easier. That is, if 'se' < 'curr' then
4352 * the resulting gran will be larger, therefore penalizing the
4353 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4354 * be smaller, again penalizing the lighter task.
4356 * This is especially important for buddies when the leftmost
4357 * task is higher priority than the buddy.
4359 return calc_delta_fair(gran, se);
4363 * Should 'se' preempt 'curr'.
4377 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4379 s64 gran, vdiff = curr->vruntime - se->vruntime;
4384 gran = wakeup_gran(curr, se);
4391 static void set_last_buddy(struct sched_entity *se)
4393 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4396 for_each_sched_entity(se)
4397 cfs_rq_of(se)->last = se;
4400 static void set_next_buddy(struct sched_entity *se)
4402 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4405 for_each_sched_entity(se)
4406 cfs_rq_of(se)->next = se;
4409 static void set_skip_buddy(struct sched_entity *se)
4411 for_each_sched_entity(se)
4412 cfs_rq_of(se)->skip = se;
4416 * Preempt the current task with a newly woken task if needed:
4418 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4420 struct task_struct *curr = rq->curr;
4421 struct sched_entity *se = &curr->se, *pse = &p->se;
4422 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4423 int scale = cfs_rq->nr_running >= sched_nr_latency;
4424 int next_buddy_marked = 0;
4426 if (unlikely(se == pse))
4430 * This is possible from callers such as move_task(), in which we
4431 * unconditionally check_prempt_curr() after an enqueue (which may have
4432 * lead to a throttle). This both saves work and prevents false
4433 * next-buddy nomination below.
4435 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4438 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4439 set_next_buddy(pse);
4440 next_buddy_marked = 1;
4444 * We can come here with TIF_NEED_RESCHED already set from new task
4447 * Note: this also catches the edge-case of curr being in a throttled
4448 * group (e.g. via set_curr_task), since update_curr() (in the
4449 * enqueue of curr) will have resulted in resched being set. This
4450 * prevents us from potentially nominating it as a false LAST_BUDDY
4453 if (test_tsk_need_resched(curr))
4456 /* Idle tasks are by definition preempted by non-idle tasks. */
4457 if (unlikely(curr->policy == SCHED_IDLE) &&
4458 likely(p->policy != SCHED_IDLE))
4462 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4463 * is driven by the tick):
4465 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4468 find_matching_se(&se, &pse);
4469 update_curr(cfs_rq_of(se));
4471 if (wakeup_preempt_entity(se, pse) == 1) {
4473 * Bias pick_next to pick the sched entity that is
4474 * triggering this preemption.
4476 if (!next_buddy_marked)
4477 set_next_buddy(pse);
4486 * Only set the backward buddy when the current task is still
4487 * on the rq. This can happen when a wakeup gets interleaved
4488 * with schedule on the ->pre_schedule() or idle_balance()
4489 * point, either of which can * drop the rq lock.
4491 * Also, during early boot the idle thread is in the fair class,
4492 * for obvious reasons its a bad idea to schedule back to it.
4494 if (unlikely(!se->on_rq || curr == rq->idle))
4497 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4501 static struct task_struct *pick_next_task_fair(struct rq *rq)
4503 struct task_struct *p;
4504 struct cfs_rq *cfs_rq = &rq->cfs;
4505 struct sched_entity *se;
4507 if (!cfs_rq->nr_running)
4511 se = pick_next_entity(cfs_rq);
4512 set_next_entity(cfs_rq, se);
4513 cfs_rq = group_cfs_rq(se);
4517 if (hrtick_enabled(rq))
4518 hrtick_start_fair(rq, p);
4524 * Account for a descheduled task:
4526 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4528 struct sched_entity *se = &prev->se;
4529 struct cfs_rq *cfs_rq;
4531 for_each_sched_entity(se) {
4532 cfs_rq = cfs_rq_of(se);
4533 put_prev_entity(cfs_rq, se);
4538 * sched_yield() is very simple
4540 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4542 static void yield_task_fair(struct rq *rq)
4544 struct task_struct *curr = rq->curr;
4545 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4546 struct sched_entity *se = &curr->se;
4549 * Are we the only task in the tree?
4551 if (unlikely(rq->nr_running == 1))
4554 clear_buddies(cfs_rq, se);
4556 if (curr->policy != SCHED_BATCH) {
4557 update_rq_clock(rq);
4559 * Update run-time statistics of the 'current'.
4561 update_curr(cfs_rq);
4563 * Tell update_rq_clock() that we've just updated,
4564 * so we don't do microscopic update in schedule()
4565 * and double the fastpath cost.
4567 rq->skip_clock_update = 1;
4573 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4575 struct sched_entity *se = &p->se;
4577 /* throttled hierarchies are not runnable */
4578 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4581 /* Tell the scheduler that we'd really like pse to run next. */
4584 yield_task_fair(rq);
4590 /**************************************************
4591 * Fair scheduling class load-balancing methods.
4595 * The purpose of load-balancing is to achieve the same basic fairness the
4596 * per-cpu scheduler provides, namely provide a proportional amount of compute
4597 * time to each task. This is expressed in the following equation:
4599 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4601 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4602 * W_i,0 is defined as:
4604 * W_i,0 = \Sum_j w_i,j (2)
4606 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4607 * is derived from the nice value as per prio_to_weight[].
4609 * The weight average is an exponential decay average of the instantaneous
4612 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4614 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4615 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4616 * can also include other factors [XXX].
4618 * To achieve this balance we define a measure of imbalance which follows
4619 * directly from (1):
4621 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4623 * We them move tasks around to minimize the imbalance. In the continuous
4624 * function space it is obvious this converges, in the discrete case we get
4625 * a few fun cases generally called infeasible weight scenarios.
4628 * - infeasible weights;
4629 * - local vs global optima in the discrete case. ]
4634 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4635 * for all i,j solution, we create a tree of cpus that follows the hardware
4636 * topology where each level pairs two lower groups (or better). This results
4637 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4638 * tree to only the first of the previous level and we decrease the frequency
4639 * of load-balance at each level inv. proportional to the number of cpus in
4645 * \Sum { --- * --- * 2^i } = O(n) (5)
4647 * `- size of each group
4648 * | | `- number of cpus doing load-balance
4650 * `- sum over all levels
4652 * Coupled with a limit on how many tasks we can migrate every balance pass,
4653 * this makes (5) the runtime complexity of the balancer.
4655 * An important property here is that each CPU is still (indirectly) connected
4656 * to every other cpu in at most O(log n) steps:
4658 * The adjacency matrix of the resulting graph is given by:
4661 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4664 * And you'll find that:
4666 * A^(log_2 n)_i,j != 0 for all i,j (7)
4668 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4669 * The task movement gives a factor of O(m), giving a convergence complexity
4672 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4677 * In order to avoid CPUs going idle while there's still work to do, new idle
4678 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4679 * tree itself instead of relying on other CPUs to bring it work.
4681 * This adds some complexity to both (5) and (8) but it reduces the total idle
4689 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4692 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4697 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4699 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4701 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4704 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4705 * rewrite all of this once again.]
4708 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4710 enum fbq_type { regular, remote, all };
4712 #define LBF_ALL_PINNED 0x01
4713 #define LBF_NEED_BREAK 0x02
4714 #define LBF_DST_PINNED 0x04
4715 #define LBF_SOME_PINNED 0x08
4718 struct sched_domain *sd;
4726 struct cpumask *dst_grpmask;
4728 enum cpu_idle_type idle;
4730 /* The set of CPUs under consideration for load-balancing */
4731 struct cpumask *cpus;
4736 unsigned int loop_break;
4737 unsigned int loop_max;
4739 enum fbq_type fbq_type;
4743 * move_task - move a task from one runqueue to another runqueue.
4744 * Both runqueues must be locked.
4746 static void move_task(struct task_struct *p, struct lb_env *env)
4748 deactivate_task(env->src_rq, p, 0);
4749 set_task_cpu(p, env->dst_cpu);
4750 activate_task(env->dst_rq, p, 0);
4751 check_preempt_curr(env->dst_rq, p, 0);
4755 * Is this task likely cache-hot:
4758 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4762 if (p->sched_class != &fair_sched_class)
4765 if (unlikely(p->policy == SCHED_IDLE))
4769 * Buddy candidates are cache hot:
4771 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4772 (&p->se == cfs_rq_of(&p->se)->next ||
4773 &p->se == cfs_rq_of(&p->se)->last))
4776 if (sysctl_sched_migration_cost == -1)
4778 if (sysctl_sched_migration_cost == 0)
4781 delta = now - p->se.exec_start;
4783 return delta < (s64)sysctl_sched_migration_cost;
4786 #ifdef CONFIG_NUMA_BALANCING
4787 /* Returns true if the destination node has incurred more faults */
4788 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4790 int src_nid, dst_nid;
4792 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4793 !(env->sd->flags & SD_NUMA)) {
4797 src_nid = cpu_to_node(env->src_cpu);
4798 dst_nid = cpu_to_node(env->dst_cpu);
4800 if (src_nid == dst_nid)
4803 /* Always encourage migration to the preferred node. */
4804 if (dst_nid == p->numa_preferred_nid)
4807 /* If both task and group weight improve, this move is a winner. */
4808 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4809 group_weight(p, dst_nid) > group_weight(p, src_nid))
4816 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4818 int src_nid, dst_nid;
4820 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4823 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4826 src_nid = cpu_to_node(env->src_cpu);
4827 dst_nid = cpu_to_node(env->dst_cpu);
4829 if (src_nid == dst_nid)
4832 /* Migrating away from the preferred node is always bad. */
4833 if (src_nid == p->numa_preferred_nid)
4836 /* If either task or group weight get worse, don't do it. */
4837 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
4838 group_weight(p, dst_nid) < group_weight(p, src_nid))
4845 static inline bool migrate_improves_locality(struct task_struct *p,
4851 static inline bool migrate_degrades_locality(struct task_struct *p,
4859 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4862 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4864 int tsk_cache_hot = 0;
4866 * We do not migrate tasks that are:
4867 * 1) throttled_lb_pair, or
4868 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4869 * 3) running (obviously), or
4870 * 4) are cache-hot on their current CPU.
4872 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4875 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4878 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4880 env->flags |= LBF_SOME_PINNED;
4883 * Remember if this task can be migrated to any other cpu in
4884 * our sched_group. We may want to revisit it if we couldn't
4885 * meet load balance goals by pulling other tasks on src_cpu.
4887 * Also avoid computing new_dst_cpu if we have already computed
4888 * one in current iteration.
4890 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4893 /* Prevent to re-select dst_cpu via env's cpus */
4894 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4895 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4896 env->flags |= LBF_DST_PINNED;
4897 env->new_dst_cpu = cpu;
4905 /* Record that we found atleast one task that could run on dst_cpu */
4906 env->flags &= ~LBF_ALL_PINNED;
4908 if (task_running(env->src_rq, p)) {
4909 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4914 * Aggressive migration if:
4915 * 1) destination numa is preferred
4916 * 2) task is cache cold, or
4917 * 3) too many balance attempts have failed.
4919 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4921 tsk_cache_hot = migrate_degrades_locality(p, env);
4923 if (migrate_improves_locality(p, env)) {
4924 #ifdef CONFIG_SCHEDSTATS
4925 if (tsk_cache_hot) {
4926 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4927 schedstat_inc(p, se.statistics.nr_forced_migrations);
4933 if (!tsk_cache_hot ||
4934 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4936 if (tsk_cache_hot) {
4937 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4938 schedstat_inc(p, se.statistics.nr_forced_migrations);
4944 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4949 * move_one_task tries to move exactly one task from busiest to this_rq, as
4950 * part of active balancing operations within "domain".
4951 * Returns 1 if successful and 0 otherwise.
4953 * Called with both runqueues locked.
4955 static int move_one_task(struct lb_env *env)
4957 struct task_struct *p, *n;
4959 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4960 if (!can_migrate_task(p, env))
4965 * Right now, this is only the second place move_task()
4966 * is called, so we can safely collect move_task()
4967 * stats here rather than inside move_task().
4969 schedstat_inc(env->sd, lb_gained[env->idle]);
4975 static const unsigned int sched_nr_migrate_break = 32;
4978 * move_tasks tries to move up to imbalance weighted load from busiest to
4979 * this_rq, as part of a balancing operation within domain "sd".
4980 * Returns 1 if successful and 0 otherwise.
4982 * Called with both runqueues locked.
4984 static int move_tasks(struct lb_env *env)
4986 struct list_head *tasks = &env->src_rq->cfs_tasks;
4987 struct task_struct *p;
4991 if (env->imbalance <= 0)
4994 while (!list_empty(tasks)) {
4995 p = list_first_entry(tasks, struct task_struct, se.group_node);
4998 /* We've more or less seen every task there is, call it quits */
4999 if (env->loop > env->loop_max)
5002 /* take a breather every nr_migrate tasks */
5003 if (env->loop > env->loop_break) {
5004 env->loop_break += sched_nr_migrate_break;
5005 env->flags |= LBF_NEED_BREAK;
5009 if (!can_migrate_task(p, env))
5012 load = task_h_load(p);
5014 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5017 if ((load / 2) > env->imbalance)
5022 env->imbalance -= load;
5024 #ifdef CONFIG_PREEMPT
5026 * NEWIDLE balancing is a source of latency, so preemptible
5027 * kernels will stop after the first task is pulled to minimize
5028 * the critical section.
5030 if (env->idle == CPU_NEWLY_IDLE)
5035 * We only want to steal up to the prescribed amount of
5038 if (env->imbalance <= 0)
5043 list_move_tail(&p->se.group_node, tasks);
5047 * Right now, this is one of only two places move_task() is called,
5048 * so we can safely collect move_task() stats here rather than
5049 * inside move_task().
5051 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5056 #ifdef CONFIG_FAIR_GROUP_SCHED
5058 * update tg->load_weight by folding this cpu's load_avg
5060 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5062 struct sched_entity *se = tg->se[cpu];
5063 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5065 /* throttled entities do not contribute to load */
5066 if (throttled_hierarchy(cfs_rq))
5069 update_cfs_rq_blocked_load(cfs_rq, 1);
5072 update_entity_load_avg(se, 1);
5074 * We pivot on our runnable average having decayed to zero for
5075 * list removal. This generally implies that all our children
5076 * have also been removed (modulo rounding error or bandwidth
5077 * control); however, such cases are rare and we can fix these
5080 * TODO: fix up out-of-order children on enqueue.
5082 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5083 list_del_leaf_cfs_rq(cfs_rq);
5085 struct rq *rq = rq_of(cfs_rq);
5086 update_rq_runnable_avg(rq, rq->nr_running);
5090 static void update_blocked_averages(int cpu)
5092 struct rq *rq = cpu_rq(cpu);
5093 struct cfs_rq *cfs_rq;
5094 unsigned long flags;
5096 raw_spin_lock_irqsave(&rq->lock, flags);
5097 update_rq_clock(rq);
5099 * Iterates the task_group tree in a bottom up fashion, see
5100 * list_add_leaf_cfs_rq() for details.
5102 for_each_leaf_cfs_rq(rq, cfs_rq) {
5104 * Note: We may want to consider periodically releasing
5105 * rq->lock about these updates so that creating many task
5106 * groups does not result in continually extending hold time.
5108 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5111 raw_spin_unlock_irqrestore(&rq->lock, flags);
5115 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5116 * This needs to be done in a top-down fashion because the load of a child
5117 * group is a fraction of its parents load.
5119 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5121 struct rq *rq = rq_of(cfs_rq);
5122 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5123 unsigned long now = jiffies;
5126 if (cfs_rq->last_h_load_update == now)
5129 cfs_rq->h_load_next = NULL;
5130 for_each_sched_entity(se) {
5131 cfs_rq = cfs_rq_of(se);
5132 cfs_rq->h_load_next = se;
5133 if (cfs_rq->last_h_load_update == now)
5138 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5139 cfs_rq->last_h_load_update = now;
5142 while ((se = cfs_rq->h_load_next) != NULL) {
5143 load = cfs_rq->h_load;
5144 load = div64_ul(load * se->avg.load_avg_contrib,
5145 cfs_rq->runnable_load_avg + 1);
5146 cfs_rq = group_cfs_rq(se);
5147 cfs_rq->h_load = load;
5148 cfs_rq->last_h_load_update = now;
5152 static unsigned long task_h_load(struct task_struct *p)
5154 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5156 update_cfs_rq_h_load(cfs_rq);
5157 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5158 cfs_rq->runnable_load_avg + 1);
5161 static inline void update_blocked_averages(int cpu)
5165 static unsigned long task_h_load(struct task_struct *p)
5167 return p->se.avg.load_avg_contrib;
5171 /********** Helpers for find_busiest_group ************************/
5173 * sg_lb_stats - stats of a sched_group required for load_balancing
5175 struct sg_lb_stats {
5176 unsigned long avg_load; /*Avg load across the CPUs of the group */
5177 unsigned long group_load; /* Total load over the CPUs of the group */
5178 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5179 unsigned long load_per_task;
5180 unsigned long group_power;
5181 unsigned int sum_nr_running; /* Nr tasks running in the group */
5182 unsigned int group_capacity;
5183 unsigned int idle_cpus;
5184 unsigned int group_weight;
5185 int group_imb; /* Is there an imbalance in the group ? */
5186 int group_has_capacity; /* Is there extra capacity in the group? */
5187 #ifdef CONFIG_NUMA_BALANCING
5188 unsigned int nr_numa_running;
5189 unsigned int nr_preferred_running;
5194 * sd_lb_stats - Structure to store the statistics of a sched_domain
5195 * during load balancing.
5197 struct sd_lb_stats {
5198 struct sched_group *busiest; /* Busiest group in this sd */
5199 struct sched_group *local; /* Local group in this sd */
5200 unsigned long total_load; /* Total load of all groups in sd */
5201 unsigned long total_pwr; /* Total power of all groups in sd */
5202 unsigned long avg_load; /* Average load across all groups in sd */
5204 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5205 struct sg_lb_stats local_stat; /* Statistics of the local group */
5208 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5211 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5212 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5213 * We must however clear busiest_stat::avg_load because
5214 * update_sd_pick_busiest() reads this before assignment.
5216 *sds = (struct sd_lb_stats){
5228 * get_sd_load_idx - Obtain the load index for a given sched domain.
5229 * @sd: The sched_domain whose load_idx is to be obtained.
5230 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5232 * Return: The load index.
5234 static inline int get_sd_load_idx(struct sched_domain *sd,
5235 enum cpu_idle_type idle)
5241 load_idx = sd->busy_idx;
5244 case CPU_NEWLY_IDLE:
5245 load_idx = sd->newidle_idx;
5248 load_idx = sd->idle_idx;
5255 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5257 return SCHED_POWER_SCALE;
5260 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5262 return default_scale_freq_power(sd, cpu);
5265 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5267 unsigned long weight = sd->span_weight;
5268 unsigned long smt_gain = sd->smt_gain;
5275 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5277 return default_scale_smt_power(sd, cpu);
5280 static unsigned long scale_rt_power(int cpu)
5282 struct rq *rq = cpu_rq(cpu);
5283 u64 total, available, age_stamp, avg;
5286 * Since we're reading these variables without serialization make sure
5287 * we read them once before doing sanity checks on them.
5289 age_stamp = ACCESS_ONCE(rq->age_stamp);
5290 avg = ACCESS_ONCE(rq->rt_avg);
5292 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5294 if (unlikely(total < avg)) {
5295 /* Ensures that power won't end up being negative */
5298 available = total - avg;
5301 if (unlikely((s64)total < SCHED_POWER_SCALE))
5302 total = SCHED_POWER_SCALE;
5304 total >>= SCHED_POWER_SHIFT;
5306 return div_u64(available, total);
5309 static void update_cpu_power(struct sched_domain *sd, int cpu)
5311 unsigned long weight = sd->span_weight;
5312 unsigned long power = SCHED_POWER_SCALE;
5313 struct sched_group *sdg = sd->groups;
5315 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5316 if (sched_feat(ARCH_POWER))
5317 power *= arch_scale_smt_power(sd, cpu);
5319 power *= default_scale_smt_power(sd, cpu);
5321 power >>= SCHED_POWER_SHIFT;
5324 sdg->sgp->power_orig = power;
5326 if (sched_feat(ARCH_POWER))
5327 power *= arch_scale_freq_power(sd, cpu);
5329 power *= default_scale_freq_power(sd, cpu);
5331 power >>= SCHED_POWER_SHIFT;
5333 power *= scale_rt_power(cpu);
5334 power >>= SCHED_POWER_SHIFT;
5339 cpu_rq(cpu)->cpu_power = power;
5340 sdg->sgp->power = power;
5343 void update_group_power(struct sched_domain *sd, int cpu)
5345 struct sched_domain *child = sd->child;
5346 struct sched_group *group, *sdg = sd->groups;
5347 unsigned long power, power_orig;
5348 unsigned long interval;
5350 interval = msecs_to_jiffies(sd->balance_interval);
5351 interval = clamp(interval, 1UL, max_load_balance_interval);
5352 sdg->sgp->next_update = jiffies + interval;
5355 update_cpu_power(sd, cpu);
5359 power_orig = power = 0;
5361 if (child->flags & SD_OVERLAP) {
5363 * SD_OVERLAP domains cannot assume that child groups
5364 * span the current group.
5367 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5368 struct sched_group_power *sgp;
5369 struct rq *rq = cpu_rq(cpu);
5372 * build_sched_domains() -> init_sched_groups_power()
5373 * gets here before we've attached the domains to the
5376 * Use power_of(), which is set irrespective of domains
5377 * in update_cpu_power().
5379 * This avoids power/power_orig from being 0 and
5380 * causing divide-by-zero issues on boot.
5382 * Runtime updates will correct power_orig.
5384 if (unlikely(!rq->sd)) {
5385 power_orig += power_of(cpu);
5386 power += power_of(cpu);
5390 sgp = rq->sd->groups->sgp;
5391 power_orig += sgp->power_orig;
5392 power += sgp->power;
5396 * !SD_OVERLAP domains can assume that child groups
5397 * span the current group.
5400 group = child->groups;
5402 power_orig += group->sgp->power_orig;
5403 power += group->sgp->power;
5404 group = group->next;
5405 } while (group != child->groups);
5408 sdg->sgp->power_orig = power_orig;
5409 sdg->sgp->power = power;
5413 * Try and fix up capacity for tiny siblings, this is needed when
5414 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5415 * which on its own isn't powerful enough.
5417 * See update_sd_pick_busiest() and check_asym_packing().
5420 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5423 * Only siblings can have significantly less than SCHED_POWER_SCALE
5425 if (!(sd->flags & SD_SHARE_CPUPOWER))
5429 * If ~90% of the cpu_power is still there, we're good.
5431 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5438 * Group imbalance indicates (and tries to solve) the problem where balancing
5439 * groups is inadequate due to tsk_cpus_allowed() constraints.
5441 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5442 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5445 * { 0 1 2 3 } { 4 5 6 7 }
5448 * If we were to balance group-wise we'd place two tasks in the first group and
5449 * two tasks in the second group. Clearly this is undesired as it will overload
5450 * cpu 3 and leave one of the cpus in the second group unused.
5452 * The current solution to this issue is detecting the skew in the first group
5453 * by noticing the lower domain failed to reach balance and had difficulty
5454 * moving tasks due to affinity constraints.
5456 * When this is so detected; this group becomes a candidate for busiest; see
5457 * update_sd_pick_busiest(). And calculate_imbalance() and
5458 * find_busiest_group() avoid some of the usual balance conditions to allow it
5459 * to create an effective group imbalance.
5461 * This is a somewhat tricky proposition since the next run might not find the
5462 * group imbalance and decide the groups need to be balanced again. A most
5463 * subtle and fragile situation.
5466 static inline int sg_imbalanced(struct sched_group *group)
5468 return group->sgp->imbalance;
5472 * Compute the group capacity.
5474 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5475 * first dividing out the smt factor and computing the actual number of cores
5476 * and limit power unit capacity with that.
5478 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5480 unsigned int capacity, smt, cpus;
5481 unsigned int power, power_orig;
5483 power = group->sgp->power;
5484 power_orig = group->sgp->power_orig;
5485 cpus = group->group_weight;
5487 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5488 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5489 capacity = cpus / smt; /* cores */
5491 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5493 capacity = fix_small_capacity(env->sd, group);
5499 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5500 * @env: The load balancing environment.
5501 * @group: sched_group whose statistics are to be updated.
5502 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5503 * @local_group: Does group contain this_cpu.
5504 * @sgs: variable to hold the statistics for this group.
5506 static inline void update_sg_lb_stats(struct lb_env *env,
5507 struct sched_group *group, int load_idx,
5508 int local_group, struct sg_lb_stats *sgs)
5513 memset(sgs, 0, sizeof(*sgs));
5515 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5516 struct rq *rq = cpu_rq(i);
5518 /* Bias balancing toward cpus of our domain */
5520 load = target_load(i, load_idx);
5522 load = source_load(i, load_idx);
5524 sgs->group_load += load;
5525 sgs->sum_nr_running += rq->nr_running;
5526 #ifdef CONFIG_NUMA_BALANCING
5527 sgs->nr_numa_running += rq->nr_numa_running;
5528 sgs->nr_preferred_running += rq->nr_preferred_running;
5530 sgs->sum_weighted_load += weighted_cpuload(i);
5535 /* Adjust by relative CPU power of the group */
5536 sgs->group_power = group->sgp->power;
5537 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5539 if (sgs->sum_nr_running)
5540 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5542 sgs->group_weight = group->group_weight;
5544 sgs->group_imb = sg_imbalanced(group);
5545 sgs->group_capacity = sg_capacity(env, group);
5547 if (sgs->group_capacity > sgs->sum_nr_running)
5548 sgs->group_has_capacity = 1;
5552 * update_sd_pick_busiest - return 1 on busiest group
5553 * @env: The load balancing environment.
5554 * @sds: sched_domain statistics
5555 * @sg: sched_group candidate to be checked for being the busiest
5556 * @sgs: sched_group statistics
5558 * Determine if @sg is a busier group than the previously selected
5561 * Return: %true if @sg is a busier group than the previously selected
5562 * busiest group. %false otherwise.
5564 static bool update_sd_pick_busiest(struct lb_env *env,
5565 struct sd_lb_stats *sds,
5566 struct sched_group *sg,
5567 struct sg_lb_stats *sgs)
5569 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5572 if (sgs->sum_nr_running > sgs->group_capacity)
5579 * ASYM_PACKING needs to move all the work to the lowest
5580 * numbered CPUs in the group, therefore mark all groups
5581 * higher than ourself as busy.
5583 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5584 env->dst_cpu < group_first_cpu(sg)) {
5588 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5595 #ifdef CONFIG_NUMA_BALANCING
5596 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5598 if (sgs->sum_nr_running > sgs->nr_numa_running)
5600 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5605 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5607 if (rq->nr_running > rq->nr_numa_running)
5609 if (rq->nr_running > rq->nr_preferred_running)
5614 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5619 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5623 #endif /* CONFIG_NUMA_BALANCING */
5626 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5627 * @env: The load balancing environment.
5628 * @sds: variable to hold the statistics for this sched_domain.
5630 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5632 struct sched_domain *child = env->sd->child;
5633 struct sched_group *sg = env->sd->groups;
5634 struct sg_lb_stats tmp_sgs;
5635 int load_idx, prefer_sibling = 0;
5637 if (child && child->flags & SD_PREFER_SIBLING)
5640 load_idx = get_sd_load_idx(env->sd, env->idle);
5643 struct sg_lb_stats *sgs = &tmp_sgs;
5646 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5649 sgs = &sds->local_stat;
5651 if (env->idle != CPU_NEWLY_IDLE ||
5652 time_after_eq(jiffies, sg->sgp->next_update))
5653 update_group_power(env->sd, env->dst_cpu);
5656 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5662 * In case the child domain prefers tasks go to siblings
5663 * first, lower the sg capacity to one so that we'll try
5664 * and move all the excess tasks away. We lower the capacity
5665 * of a group only if the local group has the capacity to fit
5666 * these excess tasks, i.e. nr_running < group_capacity. The
5667 * extra check prevents the case where you always pull from the
5668 * heaviest group when it is already under-utilized (possible
5669 * with a large weight task outweighs the tasks on the system).
5671 if (prefer_sibling && sds->local &&
5672 sds->local_stat.group_has_capacity)
5673 sgs->group_capacity = min(sgs->group_capacity, 1U);
5675 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5677 sds->busiest_stat = *sgs;
5681 /* Now, start updating sd_lb_stats */
5682 sds->total_load += sgs->group_load;
5683 sds->total_pwr += sgs->group_power;
5686 } while (sg != env->sd->groups);
5688 if (env->sd->flags & SD_NUMA)
5689 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5693 * check_asym_packing - Check to see if the group is packed into the
5696 * This is primarily intended to used at the sibling level. Some
5697 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5698 * case of POWER7, it can move to lower SMT modes only when higher
5699 * threads are idle. When in lower SMT modes, the threads will
5700 * perform better since they share less core resources. Hence when we
5701 * have idle threads, we want them to be the higher ones.
5703 * This packing function is run on idle threads. It checks to see if
5704 * the busiest CPU in this domain (core in the P7 case) has a higher
5705 * CPU number than the packing function is being run on. Here we are
5706 * assuming lower CPU number will be equivalent to lower a SMT thread
5709 * Return: 1 when packing is required and a task should be moved to
5710 * this CPU. The amount of the imbalance is returned in *imbalance.
5712 * @env: The load balancing environment.
5713 * @sds: Statistics of the sched_domain which is to be packed
5715 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5719 if (!(env->sd->flags & SD_ASYM_PACKING))
5725 busiest_cpu = group_first_cpu(sds->busiest);
5726 if (env->dst_cpu > busiest_cpu)
5729 env->imbalance = DIV_ROUND_CLOSEST(
5730 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5737 * fix_small_imbalance - Calculate the minor imbalance that exists
5738 * amongst the groups of a sched_domain, during
5740 * @env: The load balancing environment.
5741 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5744 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5746 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5747 unsigned int imbn = 2;
5748 unsigned long scaled_busy_load_per_task;
5749 struct sg_lb_stats *local, *busiest;
5751 local = &sds->local_stat;
5752 busiest = &sds->busiest_stat;
5754 if (!local->sum_nr_running)
5755 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5756 else if (busiest->load_per_task > local->load_per_task)
5759 scaled_busy_load_per_task =
5760 (busiest->load_per_task * SCHED_POWER_SCALE) /
5761 busiest->group_power;
5763 if (busiest->avg_load + scaled_busy_load_per_task >=
5764 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5765 env->imbalance = busiest->load_per_task;
5770 * OK, we don't have enough imbalance to justify moving tasks,
5771 * however we may be able to increase total CPU power used by
5775 pwr_now += busiest->group_power *
5776 min(busiest->load_per_task, busiest->avg_load);
5777 pwr_now += local->group_power *
5778 min(local->load_per_task, local->avg_load);
5779 pwr_now /= SCHED_POWER_SCALE;
5781 /* Amount of load we'd subtract */
5782 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5783 busiest->group_power;
5784 if (busiest->avg_load > tmp) {
5785 pwr_move += busiest->group_power *
5786 min(busiest->load_per_task,
5787 busiest->avg_load - tmp);
5790 /* Amount of load we'd add */
5791 if (busiest->avg_load * busiest->group_power <
5792 busiest->load_per_task * SCHED_POWER_SCALE) {
5793 tmp = (busiest->avg_load * busiest->group_power) /
5796 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5799 pwr_move += local->group_power *
5800 min(local->load_per_task, local->avg_load + tmp);
5801 pwr_move /= SCHED_POWER_SCALE;
5803 /* Move if we gain throughput */
5804 if (pwr_move > pwr_now)
5805 env->imbalance = busiest->load_per_task;
5809 * calculate_imbalance - Calculate the amount of imbalance present within the
5810 * groups of a given sched_domain during load balance.
5811 * @env: load balance environment
5812 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5814 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5816 unsigned long max_pull, load_above_capacity = ~0UL;
5817 struct sg_lb_stats *local, *busiest;
5819 local = &sds->local_stat;
5820 busiest = &sds->busiest_stat;
5822 if (busiest->group_imb) {
5824 * In the group_imb case we cannot rely on group-wide averages
5825 * to ensure cpu-load equilibrium, look at wider averages. XXX
5827 busiest->load_per_task =
5828 min(busiest->load_per_task, sds->avg_load);
5832 * In the presence of smp nice balancing, certain scenarios can have
5833 * max load less than avg load(as we skip the groups at or below
5834 * its cpu_power, while calculating max_load..)
5836 if (busiest->avg_load <= sds->avg_load ||
5837 local->avg_load >= sds->avg_load) {
5839 return fix_small_imbalance(env, sds);
5842 if (!busiest->group_imb) {
5844 * Don't want to pull so many tasks that a group would go idle.
5845 * Except of course for the group_imb case, since then we might
5846 * have to drop below capacity to reach cpu-load equilibrium.
5848 load_above_capacity =
5849 (busiest->sum_nr_running - busiest->group_capacity);
5851 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5852 load_above_capacity /= busiest->group_power;
5856 * We're trying to get all the cpus to the average_load, so we don't
5857 * want to push ourselves above the average load, nor do we wish to
5858 * reduce the max loaded cpu below the average load. At the same time,
5859 * we also don't want to reduce the group load below the group capacity
5860 * (so that we can implement power-savings policies etc). Thus we look
5861 * for the minimum possible imbalance.
5863 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5865 /* How much load to actually move to equalise the imbalance */
5866 env->imbalance = min(
5867 max_pull * busiest->group_power,
5868 (sds->avg_load - local->avg_load) * local->group_power
5869 ) / SCHED_POWER_SCALE;
5872 * if *imbalance is less than the average load per runnable task
5873 * there is no guarantee that any tasks will be moved so we'll have
5874 * a think about bumping its value to force at least one task to be
5877 if (env->imbalance < busiest->load_per_task)
5878 return fix_small_imbalance(env, sds);
5881 /******* find_busiest_group() helpers end here *********************/
5884 * find_busiest_group - Returns the busiest group within the sched_domain
5885 * if there is an imbalance. If there isn't an imbalance, and
5886 * the user has opted for power-savings, it returns a group whose
5887 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5888 * such a group exists.
5890 * Also calculates the amount of weighted load which should be moved
5891 * to restore balance.
5893 * @env: The load balancing environment.
5895 * Return: - The busiest group if imbalance exists.
5896 * - If no imbalance and user has opted for power-savings balance,
5897 * return the least loaded group whose CPUs can be
5898 * put to idle by rebalancing its tasks onto our group.
5900 static struct sched_group *find_busiest_group(struct lb_env *env)
5902 struct sg_lb_stats *local, *busiest;
5903 struct sd_lb_stats sds;
5905 init_sd_lb_stats(&sds);
5908 * Compute the various statistics relavent for load balancing at
5911 update_sd_lb_stats(env, &sds);
5912 local = &sds.local_stat;
5913 busiest = &sds.busiest_stat;
5915 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5916 check_asym_packing(env, &sds))
5919 /* There is no busy sibling group to pull tasks from */
5920 if (!sds.busiest || busiest->sum_nr_running == 0)
5923 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5926 * If the busiest group is imbalanced the below checks don't
5927 * work because they assume all things are equal, which typically
5928 * isn't true due to cpus_allowed constraints and the like.
5930 if (busiest->group_imb)
5933 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5934 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5935 !busiest->group_has_capacity)
5939 * If the local group is more busy than the selected busiest group
5940 * don't try and pull any tasks.
5942 if (local->avg_load >= busiest->avg_load)
5946 * Don't pull any tasks if this group is already above the domain
5949 if (local->avg_load >= sds.avg_load)
5952 if (env->idle == CPU_IDLE) {
5954 * This cpu is idle. If the busiest group load doesn't
5955 * have more tasks than the number of available cpu's and
5956 * there is no imbalance between this and busiest group
5957 * wrt to idle cpu's, it is balanced.
5959 if ((local->idle_cpus < busiest->idle_cpus) &&
5960 busiest->sum_nr_running <= busiest->group_weight)
5964 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5965 * imbalance_pct to be conservative.
5967 if (100 * busiest->avg_load <=
5968 env->sd->imbalance_pct * local->avg_load)
5973 /* Looks like there is an imbalance. Compute it */
5974 calculate_imbalance(env, &sds);
5983 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5985 static struct rq *find_busiest_queue(struct lb_env *env,
5986 struct sched_group *group)
5988 struct rq *busiest = NULL, *rq;
5989 unsigned long busiest_load = 0, busiest_power = 1;
5992 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5993 unsigned long power, capacity, wl;
5997 rt = fbq_classify_rq(rq);
6000 * We classify groups/runqueues into three groups:
6001 * - regular: there are !numa tasks
6002 * - remote: there are numa tasks that run on the 'wrong' node
6003 * - all: there is no distinction
6005 * In order to avoid migrating ideally placed numa tasks,
6006 * ignore those when there's better options.
6008 * If we ignore the actual busiest queue to migrate another
6009 * task, the next balance pass can still reduce the busiest
6010 * queue by moving tasks around inside the node.
6012 * If we cannot move enough load due to this classification
6013 * the next pass will adjust the group classification and
6014 * allow migration of more tasks.
6016 * Both cases only affect the total convergence complexity.
6018 if (rt > env->fbq_type)
6021 power = power_of(i);
6022 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
6024 capacity = fix_small_capacity(env->sd, group);
6026 wl = weighted_cpuload(i);
6029 * When comparing with imbalance, use weighted_cpuload()
6030 * which is not scaled with the cpu power.
6032 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6036 * For the load comparisons with the other cpu's, consider
6037 * the weighted_cpuload() scaled with the cpu power, so that
6038 * the load can be moved away from the cpu that is potentially
6039 * running at a lower capacity.
6041 * Thus we're looking for max(wl_i / power_i), crosswise
6042 * multiplication to rid ourselves of the division works out
6043 * to: wl_i * power_j > wl_j * power_i; where j is our
6046 if (wl * busiest_power > busiest_load * power) {
6048 busiest_power = power;
6057 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6058 * so long as it is large enough.
6060 #define MAX_PINNED_INTERVAL 512
6062 /* Working cpumask for load_balance and load_balance_newidle. */
6063 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6065 static int need_active_balance(struct lb_env *env)
6067 struct sched_domain *sd = env->sd;
6069 if (env->idle == CPU_NEWLY_IDLE) {
6072 * ASYM_PACKING needs to force migrate tasks from busy but
6073 * higher numbered CPUs in order to pack all tasks in the
6074 * lowest numbered CPUs.
6076 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6080 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6083 static int active_load_balance_cpu_stop(void *data);
6085 static int should_we_balance(struct lb_env *env)
6087 struct sched_group *sg = env->sd->groups;
6088 struct cpumask *sg_cpus, *sg_mask;
6089 int cpu, balance_cpu = -1;
6092 * In the newly idle case, we will allow all the cpu's
6093 * to do the newly idle load balance.
6095 if (env->idle == CPU_NEWLY_IDLE)
6098 sg_cpus = sched_group_cpus(sg);
6099 sg_mask = sched_group_mask(sg);
6100 /* Try to find first idle cpu */
6101 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6102 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6109 if (balance_cpu == -1)
6110 balance_cpu = group_balance_cpu(sg);
6113 * First idle cpu or the first cpu(busiest) in this sched group
6114 * is eligible for doing load balancing at this and above domains.
6116 return balance_cpu == env->dst_cpu;
6120 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6121 * tasks if there is an imbalance.
6123 static int load_balance(int this_cpu, struct rq *this_rq,
6124 struct sched_domain *sd, enum cpu_idle_type idle,
6125 int *continue_balancing)
6127 int ld_moved, cur_ld_moved, active_balance = 0;
6128 struct sched_domain *sd_parent = sd->parent;
6129 struct sched_group *group;
6131 unsigned long flags;
6132 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6134 struct lb_env env = {
6136 .dst_cpu = this_cpu,
6138 .dst_grpmask = sched_group_cpus(sd->groups),
6140 .loop_break = sched_nr_migrate_break,
6146 * For NEWLY_IDLE load_balancing, we don't need to consider
6147 * other cpus in our group
6149 if (idle == CPU_NEWLY_IDLE)
6150 env.dst_grpmask = NULL;
6152 cpumask_copy(cpus, cpu_active_mask);
6154 schedstat_inc(sd, lb_count[idle]);
6157 if (!should_we_balance(&env)) {
6158 *continue_balancing = 0;
6162 group = find_busiest_group(&env);
6164 schedstat_inc(sd, lb_nobusyg[idle]);
6168 busiest = find_busiest_queue(&env, group);
6170 schedstat_inc(sd, lb_nobusyq[idle]);
6174 BUG_ON(busiest == env.dst_rq);
6176 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6179 if (busiest->nr_running > 1) {
6181 * Attempt to move tasks. If find_busiest_group has found
6182 * an imbalance but busiest->nr_running <= 1, the group is
6183 * still unbalanced. ld_moved simply stays zero, so it is
6184 * correctly treated as an imbalance.
6186 env.flags |= LBF_ALL_PINNED;
6187 env.src_cpu = busiest->cpu;
6188 env.src_rq = busiest;
6189 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6192 local_irq_save(flags);
6193 double_rq_lock(env.dst_rq, busiest);
6196 * cur_ld_moved - load moved in current iteration
6197 * ld_moved - cumulative load moved across iterations
6199 cur_ld_moved = move_tasks(&env);
6200 ld_moved += cur_ld_moved;
6201 double_rq_unlock(env.dst_rq, busiest);
6202 local_irq_restore(flags);
6205 * some other cpu did the load balance for us.
6207 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6208 resched_cpu(env.dst_cpu);
6210 if (env.flags & LBF_NEED_BREAK) {
6211 env.flags &= ~LBF_NEED_BREAK;
6216 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6217 * us and move them to an alternate dst_cpu in our sched_group
6218 * where they can run. The upper limit on how many times we
6219 * iterate on same src_cpu is dependent on number of cpus in our
6222 * This changes load balance semantics a bit on who can move
6223 * load to a given_cpu. In addition to the given_cpu itself
6224 * (or a ilb_cpu acting on its behalf where given_cpu is
6225 * nohz-idle), we now have balance_cpu in a position to move
6226 * load to given_cpu. In rare situations, this may cause
6227 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6228 * _independently_ and at _same_ time to move some load to
6229 * given_cpu) causing exceess load to be moved to given_cpu.
6230 * This however should not happen so much in practice and
6231 * moreover subsequent load balance cycles should correct the
6232 * excess load moved.
6234 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6236 /* Prevent to re-select dst_cpu via env's cpus */
6237 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6239 env.dst_rq = cpu_rq(env.new_dst_cpu);
6240 env.dst_cpu = env.new_dst_cpu;
6241 env.flags &= ~LBF_DST_PINNED;
6243 env.loop_break = sched_nr_migrate_break;
6246 * Go back to "more_balance" rather than "redo" since we
6247 * need to continue with same src_cpu.
6253 * We failed to reach balance because of affinity.
6256 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6258 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6259 *group_imbalance = 1;
6260 } else if (*group_imbalance)
6261 *group_imbalance = 0;
6264 /* All tasks on this runqueue were pinned by CPU affinity */
6265 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6266 cpumask_clear_cpu(cpu_of(busiest), cpus);
6267 if (!cpumask_empty(cpus)) {
6269 env.loop_break = sched_nr_migrate_break;
6277 schedstat_inc(sd, lb_failed[idle]);
6279 * Increment the failure counter only on periodic balance.
6280 * We do not want newidle balance, which can be very
6281 * frequent, pollute the failure counter causing
6282 * excessive cache_hot migrations and active balances.
6284 if (idle != CPU_NEWLY_IDLE)
6285 sd->nr_balance_failed++;
6287 if (need_active_balance(&env)) {
6288 raw_spin_lock_irqsave(&busiest->lock, flags);
6290 /* don't kick the active_load_balance_cpu_stop,
6291 * if the curr task on busiest cpu can't be
6294 if (!cpumask_test_cpu(this_cpu,
6295 tsk_cpus_allowed(busiest->curr))) {
6296 raw_spin_unlock_irqrestore(&busiest->lock,
6298 env.flags |= LBF_ALL_PINNED;
6299 goto out_one_pinned;
6303 * ->active_balance synchronizes accesses to
6304 * ->active_balance_work. Once set, it's cleared
6305 * only after active load balance is finished.
6307 if (!busiest->active_balance) {
6308 busiest->active_balance = 1;
6309 busiest->push_cpu = this_cpu;
6312 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6314 if (active_balance) {
6315 stop_one_cpu_nowait(cpu_of(busiest),
6316 active_load_balance_cpu_stop, busiest,
6317 &busiest->active_balance_work);
6321 * We've kicked active balancing, reset the failure
6324 sd->nr_balance_failed = sd->cache_nice_tries+1;
6327 sd->nr_balance_failed = 0;
6329 if (likely(!active_balance)) {
6330 /* We were unbalanced, so reset the balancing interval */
6331 sd->balance_interval = sd->min_interval;
6334 * If we've begun active balancing, start to back off. This
6335 * case may not be covered by the all_pinned logic if there
6336 * is only 1 task on the busy runqueue (because we don't call
6339 if (sd->balance_interval < sd->max_interval)
6340 sd->balance_interval *= 2;
6346 schedstat_inc(sd, lb_balanced[idle]);
6348 sd->nr_balance_failed = 0;
6351 /* tune up the balancing interval */
6352 if (((env.flags & LBF_ALL_PINNED) &&
6353 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6354 (sd->balance_interval < sd->max_interval))
6355 sd->balance_interval *= 2;
6363 * idle_balance is called by schedule() if this_cpu is about to become
6364 * idle. Attempts to pull tasks from other CPUs.
6366 void idle_balance(int this_cpu, struct rq *this_rq)
6368 struct sched_domain *sd;
6369 int pulled_task = 0;
6370 unsigned long next_balance = jiffies + HZ;
6373 this_rq->idle_stamp = rq_clock(this_rq);
6375 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6379 * Drop the rq->lock, but keep IRQ/preempt disabled.
6381 raw_spin_unlock(&this_rq->lock);
6383 update_blocked_averages(this_cpu);
6385 for_each_domain(this_cpu, sd) {
6386 unsigned long interval;
6387 int continue_balancing = 1;
6388 u64 t0, domain_cost;
6390 if (!(sd->flags & SD_LOAD_BALANCE))
6393 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6396 if (sd->flags & SD_BALANCE_NEWIDLE) {
6397 t0 = sched_clock_cpu(this_cpu);
6399 /* If we've pulled tasks over stop searching: */
6400 pulled_task = load_balance(this_cpu, this_rq,
6402 &continue_balancing);
6404 domain_cost = sched_clock_cpu(this_cpu) - t0;
6405 if (domain_cost > sd->max_newidle_lb_cost)
6406 sd->max_newidle_lb_cost = domain_cost;
6408 curr_cost += domain_cost;
6411 interval = msecs_to_jiffies(sd->balance_interval);
6412 if (time_after(next_balance, sd->last_balance + interval))
6413 next_balance = sd->last_balance + interval;
6415 this_rq->idle_stamp = 0;
6421 raw_spin_lock(&this_rq->lock);
6423 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6425 * We are going idle. next_balance may be set based on
6426 * a busy processor. So reset next_balance.
6428 this_rq->next_balance = next_balance;
6431 if (curr_cost > this_rq->max_idle_balance_cost)
6432 this_rq->max_idle_balance_cost = curr_cost;
6436 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6437 * running tasks off the busiest CPU onto idle CPUs. It requires at
6438 * least 1 task to be running on each physical CPU where possible, and
6439 * avoids physical / logical imbalances.
6441 static int active_load_balance_cpu_stop(void *data)
6443 struct rq *busiest_rq = data;
6444 int busiest_cpu = cpu_of(busiest_rq);
6445 int target_cpu = busiest_rq->push_cpu;
6446 struct rq *target_rq = cpu_rq(target_cpu);
6447 struct sched_domain *sd;
6449 raw_spin_lock_irq(&busiest_rq->lock);
6451 /* make sure the requested cpu hasn't gone down in the meantime */
6452 if (unlikely(busiest_cpu != smp_processor_id() ||
6453 !busiest_rq->active_balance))
6456 /* Is there any task to move? */
6457 if (busiest_rq->nr_running <= 1)
6461 * This condition is "impossible", if it occurs
6462 * we need to fix it. Originally reported by
6463 * Bjorn Helgaas on a 128-cpu setup.
6465 BUG_ON(busiest_rq == target_rq);
6467 /* move a task from busiest_rq to target_rq */
6468 double_lock_balance(busiest_rq, target_rq);
6470 /* Search for an sd spanning us and the target CPU. */
6472 for_each_domain(target_cpu, sd) {
6473 if ((sd->flags & SD_LOAD_BALANCE) &&
6474 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6479 struct lb_env env = {
6481 .dst_cpu = target_cpu,
6482 .dst_rq = target_rq,
6483 .src_cpu = busiest_rq->cpu,
6484 .src_rq = busiest_rq,
6488 schedstat_inc(sd, alb_count);
6490 if (move_one_task(&env))
6491 schedstat_inc(sd, alb_pushed);
6493 schedstat_inc(sd, alb_failed);
6496 double_unlock_balance(busiest_rq, target_rq);
6498 busiest_rq->active_balance = 0;
6499 raw_spin_unlock_irq(&busiest_rq->lock);
6503 #ifdef CONFIG_NO_HZ_COMMON
6505 * idle load balancing details
6506 * - When one of the busy CPUs notice that there may be an idle rebalancing
6507 * needed, they will kick the idle load balancer, which then does idle
6508 * load balancing for all the idle CPUs.
6511 cpumask_var_t idle_cpus_mask;
6513 unsigned long next_balance; /* in jiffy units */
6514 } nohz ____cacheline_aligned;
6516 static inline int find_new_ilb(void)
6518 int ilb = cpumask_first(nohz.idle_cpus_mask);
6520 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6527 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6528 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6529 * CPU (if there is one).
6531 static void nohz_balancer_kick(void)
6535 nohz.next_balance++;
6537 ilb_cpu = find_new_ilb();
6539 if (ilb_cpu >= nr_cpu_ids)
6542 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6545 * Use smp_send_reschedule() instead of resched_cpu().
6546 * This way we generate a sched IPI on the target cpu which
6547 * is idle. And the softirq performing nohz idle load balance
6548 * will be run before returning from the IPI.
6550 smp_send_reschedule(ilb_cpu);
6554 static inline void nohz_balance_exit_idle(int cpu)
6556 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6557 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6558 atomic_dec(&nohz.nr_cpus);
6559 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6563 static inline void set_cpu_sd_state_busy(void)
6565 struct sched_domain *sd;
6566 int cpu = smp_processor_id();
6569 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6571 if (!sd || !sd->nohz_idle)
6575 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6580 void set_cpu_sd_state_idle(void)
6582 struct sched_domain *sd;
6583 int cpu = smp_processor_id();
6586 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6588 if (!sd || sd->nohz_idle)
6592 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6598 * This routine will record that the cpu is going idle with tick stopped.
6599 * This info will be used in performing idle load balancing in the future.
6601 void nohz_balance_enter_idle(int cpu)
6604 * If this cpu is going down, then nothing needs to be done.
6606 if (!cpu_active(cpu))
6609 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6612 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6613 atomic_inc(&nohz.nr_cpus);
6614 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6617 static int sched_ilb_notifier(struct notifier_block *nfb,
6618 unsigned long action, void *hcpu)
6620 switch (action & ~CPU_TASKS_FROZEN) {
6622 nohz_balance_exit_idle(smp_processor_id());
6630 static DEFINE_SPINLOCK(balancing);
6633 * Scale the max load_balance interval with the number of CPUs in the system.
6634 * This trades load-balance latency on larger machines for less cross talk.
6636 void update_max_interval(void)
6638 max_load_balance_interval = HZ*num_online_cpus()/10;
6642 * It checks each scheduling domain to see if it is due to be balanced,
6643 * and initiates a balancing operation if so.
6645 * Balancing parameters are set up in init_sched_domains.
6647 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
6649 int continue_balancing = 1;
6651 unsigned long interval;
6652 struct sched_domain *sd;
6653 /* Earliest time when we have to do rebalance again */
6654 unsigned long next_balance = jiffies + 60*HZ;
6655 int update_next_balance = 0;
6656 int need_serialize, need_decay = 0;
6659 update_blocked_averages(cpu);
6662 for_each_domain(cpu, sd) {
6664 * Decay the newidle max times here because this is a regular
6665 * visit to all the domains. Decay ~1% per second.
6667 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6668 sd->max_newidle_lb_cost =
6669 (sd->max_newidle_lb_cost * 253) / 256;
6670 sd->next_decay_max_lb_cost = jiffies + HZ;
6673 max_cost += sd->max_newidle_lb_cost;
6675 if (!(sd->flags & SD_LOAD_BALANCE))
6679 * Stop the load balance at this level. There is another
6680 * CPU in our sched group which is doing load balancing more
6683 if (!continue_balancing) {
6689 interval = sd->balance_interval;
6690 if (idle != CPU_IDLE)
6691 interval *= sd->busy_factor;
6693 /* scale ms to jiffies */
6694 interval = msecs_to_jiffies(interval);
6695 interval = clamp(interval, 1UL, max_load_balance_interval);
6697 need_serialize = sd->flags & SD_SERIALIZE;
6699 if (need_serialize) {
6700 if (!spin_trylock(&balancing))
6704 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6705 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6707 * The LBF_DST_PINNED logic could have changed
6708 * env->dst_cpu, so we can't know our idle
6709 * state even if we migrated tasks. Update it.
6711 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6713 sd->last_balance = jiffies;
6716 spin_unlock(&balancing);
6718 if (time_after(next_balance, sd->last_balance + interval)) {
6719 next_balance = sd->last_balance + interval;
6720 update_next_balance = 1;
6725 * Ensure the rq-wide value also decays but keep it at a
6726 * reasonable floor to avoid funnies with rq->avg_idle.
6728 rq->max_idle_balance_cost =
6729 max((u64)sysctl_sched_migration_cost, max_cost);
6734 * next_balance will be updated only when there is a need.
6735 * When the cpu is attached to null domain for ex, it will not be
6738 if (likely(update_next_balance))
6739 rq->next_balance = next_balance;
6742 #ifdef CONFIG_NO_HZ_COMMON
6744 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6745 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6747 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
6749 int this_cpu = this_rq->cpu;
6753 if (idle != CPU_IDLE ||
6754 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6757 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6758 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6762 * If this cpu gets work to do, stop the load balancing
6763 * work being done for other cpus. Next load
6764 * balancing owner will pick it up.
6769 rq = cpu_rq(balance_cpu);
6771 raw_spin_lock_irq(&rq->lock);
6772 update_rq_clock(rq);
6773 update_idle_cpu_load(rq);
6774 raw_spin_unlock_irq(&rq->lock);
6776 rebalance_domains(rq, CPU_IDLE);
6778 if (time_after(this_rq->next_balance, rq->next_balance))
6779 this_rq->next_balance = rq->next_balance;
6781 nohz.next_balance = this_rq->next_balance;
6783 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6787 * Current heuristic for kicking the idle load balancer in the presence
6788 * of an idle cpu is the system.
6789 * - This rq has more than one task.
6790 * - At any scheduler domain level, this cpu's scheduler group has multiple
6791 * busy cpu's exceeding the group's power.
6792 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6793 * domain span are idle.
6795 static inline int nohz_kick_needed(struct rq *rq)
6797 unsigned long now = jiffies;
6798 struct sched_domain *sd;
6799 struct sched_group_power *sgp;
6800 int nr_busy, cpu = rq->cpu;
6802 if (unlikely(rq->idle_balance))
6806 * We may be recently in ticked or tickless idle mode. At the first
6807 * busy tick after returning from idle, we will update the busy stats.
6809 set_cpu_sd_state_busy();
6810 nohz_balance_exit_idle(cpu);
6813 * None are in tickless mode and hence no need for NOHZ idle load
6816 if (likely(!atomic_read(&nohz.nr_cpus)))
6819 if (time_before(now, nohz.next_balance))
6822 if (rq->nr_running >= 2)
6826 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6829 sgp = sd->groups->sgp;
6830 nr_busy = atomic_read(&sgp->nr_busy_cpus);
6833 goto need_kick_unlock;
6836 sd = rcu_dereference(per_cpu(sd_asym, cpu));
6838 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
6839 sched_domain_span(sd)) < cpu))
6840 goto need_kick_unlock;
6851 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
6855 * run_rebalance_domains is triggered when needed from the scheduler tick.
6856 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6858 static void run_rebalance_domains(struct softirq_action *h)
6860 struct rq *this_rq = this_rq();
6861 enum cpu_idle_type idle = this_rq->idle_balance ?
6862 CPU_IDLE : CPU_NOT_IDLE;
6864 rebalance_domains(this_rq, idle);
6867 * If this cpu has a pending nohz_balance_kick, then do the
6868 * balancing on behalf of the other idle cpus whose ticks are
6871 nohz_idle_balance(this_rq, idle);
6874 static inline int on_null_domain(struct rq *rq)
6876 return !rcu_dereference_sched(rq->sd);
6880 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6882 void trigger_load_balance(struct rq *rq)
6884 /* Don't need to rebalance while attached to NULL domain */
6885 if (unlikely(on_null_domain(rq)))
6888 if (time_after_eq(jiffies, rq->next_balance))
6889 raise_softirq(SCHED_SOFTIRQ);
6890 #ifdef CONFIG_NO_HZ_COMMON
6891 if (nohz_kick_needed(rq))
6892 nohz_balancer_kick();
6896 static void rq_online_fair(struct rq *rq)
6901 static void rq_offline_fair(struct rq *rq)
6905 /* Ensure any throttled groups are reachable by pick_next_task */
6906 unthrottle_offline_cfs_rqs(rq);
6909 #endif /* CONFIG_SMP */
6912 * scheduler tick hitting a task of our scheduling class:
6914 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6916 struct cfs_rq *cfs_rq;
6917 struct sched_entity *se = &curr->se;
6919 for_each_sched_entity(se) {
6920 cfs_rq = cfs_rq_of(se);
6921 entity_tick(cfs_rq, se, queued);
6924 if (numabalancing_enabled)
6925 task_tick_numa(rq, curr);
6927 update_rq_runnable_avg(rq, 1);
6931 * called on fork with the child task as argument from the parent's context
6932 * - child not yet on the tasklist
6933 * - preemption disabled
6935 static void task_fork_fair(struct task_struct *p)
6937 struct cfs_rq *cfs_rq;
6938 struct sched_entity *se = &p->se, *curr;
6939 int this_cpu = smp_processor_id();
6940 struct rq *rq = this_rq();
6941 unsigned long flags;
6943 raw_spin_lock_irqsave(&rq->lock, flags);
6945 update_rq_clock(rq);
6947 cfs_rq = task_cfs_rq(current);
6948 curr = cfs_rq->curr;
6951 * Not only the cpu but also the task_group of the parent might have
6952 * been changed after parent->se.parent,cfs_rq were copied to
6953 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6954 * of child point to valid ones.
6957 __set_task_cpu(p, this_cpu);
6960 update_curr(cfs_rq);
6963 se->vruntime = curr->vruntime;
6964 place_entity(cfs_rq, se, 1);
6966 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6968 * Upon rescheduling, sched_class::put_prev_task() will place
6969 * 'current' within the tree based on its new key value.
6971 swap(curr->vruntime, se->vruntime);
6972 resched_task(rq->curr);
6975 se->vruntime -= cfs_rq->min_vruntime;
6977 raw_spin_unlock_irqrestore(&rq->lock, flags);
6981 * Priority of the task has changed. Check to see if we preempt
6985 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6991 * Reschedule if we are currently running on this runqueue and
6992 * our priority decreased, or if we are not currently running on
6993 * this runqueue and our priority is higher than the current's
6995 if (rq->curr == p) {
6996 if (p->prio > oldprio)
6997 resched_task(rq->curr);
6999 check_preempt_curr(rq, p, 0);
7002 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7004 struct sched_entity *se = &p->se;
7005 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7008 * Ensure the task's vruntime is normalized, so that when its
7009 * switched back to the fair class the enqueue_entity(.flags=0) will
7010 * do the right thing.
7012 * If it was on_rq, then the dequeue_entity(.flags=0) will already
7013 * have normalized the vruntime, if it was !on_rq, then only when
7014 * the task is sleeping will it still have non-normalized vruntime.
7016 if (!se->on_rq && p->state != TASK_RUNNING) {
7018 * Fix up our vruntime so that the current sleep doesn't
7019 * cause 'unlimited' sleep bonus.
7021 place_entity(cfs_rq, se, 0);
7022 se->vruntime -= cfs_rq->min_vruntime;
7027 * Remove our load from contribution when we leave sched_fair
7028 * and ensure we don't carry in an old decay_count if we
7031 if (se->avg.decay_count) {
7032 __synchronize_entity_decay(se);
7033 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7039 * We switched to the sched_fair class.
7041 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7047 * We were most likely switched from sched_rt, so
7048 * kick off the schedule if running, otherwise just see
7049 * if we can still preempt the current task.
7052 resched_task(rq->curr);
7054 check_preempt_curr(rq, p, 0);
7057 /* Account for a task changing its policy or group.
7059 * This routine is mostly called to set cfs_rq->curr field when a task
7060 * migrates between groups/classes.
7062 static void set_curr_task_fair(struct rq *rq)
7064 struct sched_entity *se = &rq->curr->se;
7066 for_each_sched_entity(se) {
7067 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7069 set_next_entity(cfs_rq, se);
7070 /* ensure bandwidth has been allocated on our new cfs_rq */
7071 account_cfs_rq_runtime(cfs_rq, 0);
7075 void init_cfs_rq(struct cfs_rq *cfs_rq)
7077 cfs_rq->tasks_timeline = RB_ROOT;
7078 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7079 #ifndef CONFIG_64BIT
7080 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7083 atomic64_set(&cfs_rq->decay_counter, 1);
7084 atomic_long_set(&cfs_rq->removed_load, 0);
7088 #ifdef CONFIG_FAIR_GROUP_SCHED
7089 static void task_move_group_fair(struct task_struct *p, int on_rq)
7091 struct cfs_rq *cfs_rq;
7093 * If the task was not on the rq at the time of this cgroup movement
7094 * it must have been asleep, sleeping tasks keep their ->vruntime
7095 * absolute on their old rq until wakeup (needed for the fair sleeper
7096 * bonus in place_entity()).
7098 * If it was on the rq, we've just 'preempted' it, which does convert
7099 * ->vruntime to a relative base.
7101 * Make sure both cases convert their relative position when migrating
7102 * to another cgroup's rq. This does somewhat interfere with the
7103 * fair sleeper stuff for the first placement, but who cares.
7106 * When !on_rq, vruntime of the task has usually NOT been normalized.
7107 * But there are some cases where it has already been normalized:
7109 * - Moving a forked child which is waiting for being woken up by
7110 * wake_up_new_task().
7111 * - Moving a task which has been woken up by try_to_wake_up() and
7112 * waiting for actually being woken up by sched_ttwu_pending().
7114 * To prevent boost or penalty in the new cfs_rq caused by delta
7115 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7117 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7121 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7122 set_task_rq(p, task_cpu(p));
7124 cfs_rq = cfs_rq_of(&p->se);
7125 p->se.vruntime += cfs_rq->min_vruntime;
7128 * migrate_task_rq_fair() will have removed our previous
7129 * contribution, but we must synchronize for ongoing future
7132 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7133 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7138 void free_fair_sched_group(struct task_group *tg)
7142 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7144 for_each_possible_cpu(i) {
7146 kfree(tg->cfs_rq[i]);
7155 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7157 struct cfs_rq *cfs_rq;
7158 struct sched_entity *se;
7161 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7164 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7168 tg->shares = NICE_0_LOAD;
7170 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7172 for_each_possible_cpu(i) {
7173 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7174 GFP_KERNEL, cpu_to_node(i));
7178 se = kzalloc_node(sizeof(struct sched_entity),
7179 GFP_KERNEL, cpu_to_node(i));
7183 init_cfs_rq(cfs_rq);
7184 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7195 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7197 struct rq *rq = cpu_rq(cpu);
7198 unsigned long flags;
7201 * Only empty task groups can be destroyed; so we can speculatively
7202 * check on_list without danger of it being re-added.
7204 if (!tg->cfs_rq[cpu]->on_list)
7207 raw_spin_lock_irqsave(&rq->lock, flags);
7208 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7209 raw_spin_unlock_irqrestore(&rq->lock, flags);
7212 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7213 struct sched_entity *se, int cpu,
7214 struct sched_entity *parent)
7216 struct rq *rq = cpu_rq(cpu);
7220 init_cfs_rq_runtime(cfs_rq);
7222 tg->cfs_rq[cpu] = cfs_rq;
7225 /* se could be NULL for root_task_group */
7230 se->cfs_rq = &rq->cfs;
7232 se->cfs_rq = parent->my_q;
7235 /* guarantee group entities always have weight */
7236 update_load_set(&se->load, NICE_0_LOAD);
7237 se->parent = parent;
7240 static DEFINE_MUTEX(shares_mutex);
7242 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7245 unsigned long flags;
7248 * We can't change the weight of the root cgroup.
7253 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7255 mutex_lock(&shares_mutex);
7256 if (tg->shares == shares)
7259 tg->shares = shares;
7260 for_each_possible_cpu(i) {
7261 struct rq *rq = cpu_rq(i);
7262 struct sched_entity *se;
7265 /* Propagate contribution to hierarchy */
7266 raw_spin_lock_irqsave(&rq->lock, flags);
7268 /* Possible calls to update_curr() need rq clock */
7269 update_rq_clock(rq);
7270 for_each_sched_entity(se)
7271 update_cfs_shares(group_cfs_rq(se));
7272 raw_spin_unlock_irqrestore(&rq->lock, flags);
7276 mutex_unlock(&shares_mutex);
7279 #else /* CONFIG_FAIR_GROUP_SCHED */
7281 void free_fair_sched_group(struct task_group *tg) { }
7283 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7288 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7290 #endif /* CONFIG_FAIR_GROUP_SCHED */
7293 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7295 struct sched_entity *se = &task->se;
7296 unsigned int rr_interval = 0;
7299 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7302 if (rq->cfs.load.weight)
7303 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7309 * All the scheduling class methods:
7311 const struct sched_class fair_sched_class = {
7312 .next = &idle_sched_class,
7313 .enqueue_task = enqueue_task_fair,
7314 .dequeue_task = dequeue_task_fair,
7315 .yield_task = yield_task_fair,
7316 .yield_to_task = yield_to_task_fair,
7318 .check_preempt_curr = check_preempt_wakeup,
7320 .pick_next_task = pick_next_task_fair,
7321 .put_prev_task = put_prev_task_fair,
7324 .select_task_rq = select_task_rq_fair,
7325 .migrate_task_rq = migrate_task_rq_fair,
7327 .rq_online = rq_online_fair,
7328 .rq_offline = rq_offline_fair,
7330 .task_waking = task_waking_fair,
7333 .set_curr_task = set_curr_task_fair,
7334 .task_tick = task_tick_fair,
7335 .task_fork = task_fork_fair,
7337 .prio_changed = prio_changed_fair,
7338 .switched_from = switched_from_fair,
7339 .switched_to = switched_to_fair,
7341 .get_rr_interval = get_rr_interval_fair,
7343 #ifdef CONFIG_FAIR_GROUP_SCHED
7344 .task_move_group = task_move_group_fair,
7348 #ifdef CONFIG_SCHED_DEBUG
7349 void print_cfs_stats(struct seq_file *m, int cpu)
7351 struct cfs_rq *cfs_rq;
7354 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7355 print_cfs_rq(m, cpu, cfs_rq);
7360 __init void init_sched_fair_class(void)
7363 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7365 #ifdef CONFIG_NO_HZ_COMMON
7366 nohz.next_balance = jiffies;
7367 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7368 cpu_notifier(sched_ilb_notifier, 0);