2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #define WMULT_CONST (~0U)
182 #define WMULT_SHIFT 32
184 static void __update_inv_weight(struct load_weight *lw)
188 if (likely(lw->inv_weight))
191 w = scale_load_down(lw->weight);
193 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
195 else if (unlikely(!w))
196 lw->inv_weight = WMULT_CONST;
198 lw->inv_weight = WMULT_CONST / w;
202 * delta_exec * weight / lw.weight
204 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
207 * we're guaranteed shift stays positive because inv_weight is guaranteed to
208 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
211 * weight/lw.weight <= 1, and therefore our shift will also be positive.
213 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
215 u64 fact = scale_load_down(weight);
216 int shift = WMULT_SHIFT;
218 __update_inv_weight(lw);
220 if (unlikely(fact >> 32)) {
227 /* hint to use a 32x32->64 mul */
228 fact = (u64)(u32)fact * lw->inv_weight;
235 return mul_u64_u32_shr(delta_exec, fact, shift);
239 const struct sched_class fair_sched_class;
241 /**************************************************************
242 * CFS operations on generic schedulable entities:
245 #ifdef CONFIG_FAIR_GROUP_SCHED
247 /* cpu runqueue to which this cfs_rq is attached */
248 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
253 /* An entity is a task if it doesn't "own" a runqueue */
254 #define entity_is_task(se) (!se->my_q)
256 static inline struct task_struct *task_of(struct sched_entity *se)
258 #ifdef CONFIG_SCHED_DEBUG
259 WARN_ON_ONCE(!entity_is_task(se));
261 return container_of(se, struct task_struct, se);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
285 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 if (!cfs_rq->on_list) {
292 * Ensure we either appear before our parent (if already
293 * enqueued) or force our parent to appear after us when it is
294 * enqueued. The fact that we always enqueue bottom-up
295 * reduces this to two cases.
297 if (cfs_rq->tg->parent &&
298 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
299 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
300 &rq_of(cfs_rq)->leaf_cfs_rq_list);
302 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
303 &rq_of(cfs_rq)->leaf_cfs_rq_list);
307 /* We should have no load, but we need to update last_decay. */
308 update_cfs_rq_blocked_load(cfs_rq, 0);
312 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
314 if (cfs_rq->on_list) {
315 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
320 /* Iterate thr' all leaf cfs_rq's on a runqueue */
321 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
322 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
324 /* Do the two (enqueued) entities belong to the same group ? */
326 is_same_group(struct sched_entity *se, struct sched_entity *pse)
328 if (se->cfs_rq == pse->cfs_rq)
334 static inline struct sched_entity *parent_entity(struct sched_entity *se)
339 /* return depth at which a sched entity is present in the hierarchy */
340 static inline int depth_se(struct sched_entity *se)
344 for_each_sched_entity(se)
351 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
353 int se_depth, pse_depth;
356 * preemption test can be made between sibling entities who are in the
357 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
358 * both tasks until we find their ancestors who are siblings of common
362 /* First walk up until both entities are at same depth */
363 se_depth = depth_se(*se);
364 pse_depth = depth_se(*pse);
366 while (se_depth > pse_depth) {
368 *se = parent_entity(*se);
371 while (pse_depth > se_depth) {
373 *pse = parent_entity(*pse);
376 while (!is_same_group(*se, *pse)) {
377 *se = parent_entity(*se);
378 *pse = parent_entity(*pse);
382 #else /* !CONFIG_FAIR_GROUP_SCHED */
384 static inline struct task_struct *task_of(struct sched_entity *se)
386 return container_of(se, struct task_struct, se);
389 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
391 return container_of(cfs_rq, struct rq, cfs);
394 #define entity_is_task(se) 1
396 #define for_each_sched_entity(se) \
397 for (; se; se = NULL)
399 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
401 return &task_rq(p)->cfs;
404 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
406 struct task_struct *p = task_of(se);
407 struct rq *rq = task_rq(p);
412 /* runqueue "owned" by this group */
413 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
418 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
422 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
426 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
427 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
430 is_same_group(struct sched_entity *se, struct sched_entity *pse)
435 static inline struct sched_entity *parent_entity(struct sched_entity *se)
441 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
445 #endif /* CONFIG_FAIR_GROUP_SCHED */
447 static __always_inline
448 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
450 /**************************************************************
451 * Scheduling class tree data structure manipulation methods:
454 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
456 s64 delta = (s64)(vruntime - max_vruntime);
458 max_vruntime = vruntime;
463 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
465 s64 delta = (s64)(vruntime - min_vruntime);
467 min_vruntime = vruntime;
472 static inline int entity_before(struct sched_entity *a,
473 struct sched_entity *b)
475 return (s64)(a->vruntime - b->vruntime) < 0;
478 static void update_min_vruntime(struct cfs_rq *cfs_rq)
480 u64 vruntime = cfs_rq->min_vruntime;
483 vruntime = cfs_rq->curr->vruntime;
485 if (cfs_rq->rb_leftmost) {
486 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
491 vruntime = se->vruntime;
493 vruntime = min_vruntime(vruntime, se->vruntime);
496 /* ensure we never gain time by being placed backwards. */
497 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
500 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
505 * Enqueue an entity into the rb-tree:
507 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
509 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
510 struct rb_node *parent = NULL;
511 struct sched_entity *entry;
515 * Find the right place in the rbtree:
519 entry = rb_entry(parent, struct sched_entity, run_node);
521 * We dont care about collisions. Nodes with
522 * the same key stay together.
524 if (entity_before(se, entry)) {
525 link = &parent->rb_left;
527 link = &parent->rb_right;
533 * Maintain a cache of leftmost tree entries (it is frequently
537 cfs_rq->rb_leftmost = &se->run_node;
539 rb_link_node(&se->run_node, parent, link);
540 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
543 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
545 if (cfs_rq->rb_leftmost == &se->run_node) {
546 struct rb_node *next_node;
548 next_node = rb_next(&se->run_node);
549 cfs_rq->rb_leftmost = next_node;
552 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
555 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
557 struct rb_node *left = cfs_rq->rb_leftmost;
562 return rb_entry(left, struct sched_entity, run_node);
565 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
567 struct rb_node *next = rb_next(&se->run_node);
572 return rb_entry(next, struct sched_entity, run_node);
575 #ifdef CONFIG_SCHED_DEBUG
576 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
578 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
583 return rb_entry(last, struct sched_entity, run_node);
586 /**************************************************************
587 * Scheduling class statistics methods:
590 int sched_proc_update_handler(struct ctl_table *table, int write,
591 void __user *buffer, size_t *lenp,
594 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
595 int factor = get_update_sysctl_factor();
600 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
601 sysctl_sched_min_granularity);
603 #define WRT_SYSCTL(name) \
604 (normalized_sysctl_##name = sysctl_##name / (factor))
605 WRT_SYSCTL(sched_min_granularity);
606 WRT_SYSCTL(sched_latency);
607 WRT_SYSCTL(sched_wakeup_granularity);
617 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
619 if (unlikely(se->load.weight != NICE_0_LOAD))
620 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
626 * The idea is to set a period in which each task runs once.
628 * When there are too many tasks (sched_nr_latency) we have to stretch
629 * this period because otherwise the slices get too small.
631 * p = (nr <= nl) ? l : l*nr/nl
633 static u64 __sched_period(unsigned long nr_running)
635 u64 period = sysctl_sched_latency;
636 unsigned long nr_latency = sched_nr_latency;
638 if (unlikely(nr_running > nr_latency)) {
639 period = sysctl_sched_min_granularity;
640 period *= nr_running;
647 * We calculate the wall-time slice from the period by taking a part
648 * proportional to the weight.
652 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
654 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
656 for_each_sched_entity(se) {
657 struct load_weight *load;
658 struct load_weight lw;
660 cfs_rq = cfs_rq_of(se);
661 load = &cfs_rq->load;
663 if (unlikely(!se->on_rq)) {
666 update_load_add(&lw, se->load.weight);
669 slice = __calc_delta(slice, se->load.weight, load);
675 * We calculate the vruntime slice of a to-be-inserted task.
679 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
681 return calc_delta_fair(sched_slice(cfs_rq, se), se);
685 static unsigned long task_h_load(struct task_struct *p);
687 static inline void __update_task_entity_contrib(struct sched_entity *se);
689 /* Give new task start runnable values to heavy its load in infant time */
690 void init_task_runnable_average(struct task_struct *p)
694 p->se.avg.decay_count = 0;
695 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
696 p->se.avg.runnable_avg_sum = slice;
697 p->se.avg.runnable_avg_period = slice;
698 __update_task_entity_contrib(&p->se);
701 void init_task_runnable_average(struct task_struct *p)
707 * Update the current task's runtime statistics.
709 static void update_curr(struct cfs_rq *cfs_rq)
711 struct sched_entity *curr = cfs_rq->curr;
712 u64 now = rq_clock_task(rq_of(cfs_rq));
718 delta_exec = now - curr->exec_start;
719 if (unlikely((s64)delta_exec <= 0))
722 curr->exec_start = now;
724 schedstat_set(curr->statistics.exec_max,
725 max(delta_exec, curr->statistics.exec_max));
727 curr->sum_exec_runtime += delta_exec;
728 schedstat_add(cfs_rq, exec_clock, delta_exec);
730 curr->vruntime += calc_delta_fair(delta_exec, curr);
731 update_min_vruntime(cfs_rq);
733 if (entity_is_task(curr)) {
734 struct task_struct *curtask = task_of(curr);
736 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
737 cpuacct_charge(curtask, delta_exec);
738 account_group_exec_runtime(curtask, delta_exec);
741 account_cfs_rq_runtime(cfs_rq, delta_exec);
745 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
747 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
751 * Task is being enqueued - update stats:
753 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
756 * Are we enqueueing a waiting task? (for current tasks
757 * a dequeue/enqueue event is a NOP)
759 if (se != cfs_rq->curr)
760 update_stats_wait_start(cfs_rq, se);
764 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
766 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
767 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
768 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
769 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
770 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
771 #ifdef CONFIG_SCHEDSTATS
772 if (entity_is_task(se)) {
773 trace_sched_stat_wait(task_of(se),
774 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
777 schedstat_set(se->statistics.wait_start, 0);
781 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
784 * Mark the end of the wait period if dequeueing a
787 if (se != cfs_rq->curr)
788 update_stats_wait_end(cfs_rq, se);
792 * We are picking a new current task - update its stats:
795 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
798 * We are starting a new run period:
800 se->exec_start = rq_clock_task(rq_of(cfs_rq));
803 /**************************************************
804 * Scheduling class queueing methods:
807 #ifdef CONFIG_NUMA_BALANCING
809 * Approximate time to scan a full NUMA task in ms. The task scan period is
810 * calculated based on the tasks virtual memory size and
811 * numa_balancing_scan_size.
813 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
814 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
816 /* Portion of address space to scan in MB */
817 unsigned int sysctl_numa_balancing_scan_size = 256;
819 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
820 unsigned int sysctl_numa_balancing_scan_delay = 1000;
822 static unsigned int task_nr_scan_windows(struct task_struct *p)
824 unsigned long rss = 0;
825 unsigned long nr_scan_pages;
828 * Calculations based on RSS as non-present and empty pages are skipped
829 * by the PTE scanner and NUMA hinting faults should be trapped based
832 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
833 rss = get_mm_rss(p->mm);
837 rss = round_up(rss, nr_scan_pages);
838 return rss / nr_scan_pages;
841 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
842 #define MAX_SCAN_WINDOW 2560
844 static unsigned int task_scan_min(struct task_struct *p)
846 unsigned int scan, floor;
847 unsigned int windows = 1;
849 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
850 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
851 floor = 1000 / windows;
853 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
854 return max_t(unsigned int, floor, scan);
857 static unsigned int task_scan_max(struct task_struct *p)
859 unsigned int smin = task_scan_min(p);
862 /* Watch for min being lower than max due to floor calculations */
863 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
864 return max(smin, smax);
867 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
869 rq->nr_numa_running += (p->numa_preferred_nid != -1);
870 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
873 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
875 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
876 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
882 spinlock_t lock; /* nr_tasks, tasks */
885 struct list_head task_list;
888 nodemask_t active_nodes;
889 unsigned long total_faults;
890 unsigned long *faults_cpu;
891 unsigned long faults[0];
894 pid_t task_numa_group_id(struct task_struct *p)
896 return p->numa_group ? p->numa_group->gid : 0;
899 static inline int task_faults_idx(int nid, int priv)
901 return 2 * nid + priv;
904 static inline unsigned long task_faults(struct task_struct *p, int nid)
906 if (!p->numa_faults_memory)
909 return p->numa_faults_memory[task_faults_idx(nid, 0)] +
910 p->numa_faults_memory[task_faults_idx(nid, 1)];
913 static inline unsigned long group_faults(struct task_struct *p, int nid)
918 return p->numa_group->faults[task_faults_idx(nid, 0)] +
919 p->numa_group->faults[task_faults_idx(nid, 1)];
922 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
924 return group->faults_cpu[task_faults_idx(nid, 0)] +
925 group->faults_cpu[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;
938 if (!p->numa_faults_memory)
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 int ret = migrate_task_to(p, env.best_cpu);
1257 ret = migrate_swap(p, env.best_task);
1258 put_task_struct(env.best_task);
1262 /* Attempt to migrate a task to a CPU on the preferred node. */
1263 static void numa_migrate_preferred(struct task_struct *p)
1265 /* This task has no NUMA fault statistics yet */
1266 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1269 /* Periodically retry migrating the task to the preferred node */
1270 p->numa_migrate_retry = jiffies + HZ;
1272 /* Success if task is already running on preferred CPU */
1273 if (task_node(p) == p->numa_preferred_nid)
1276 /* Otherwise, try migrate to a CPU on the preferred node */
1277 task_numa_migrate(p);
1281 * Find the nodes on which the workload is actively running. We do this by
1282 * tracking the nodes from which NUMA hinting faults are triggered. This can
1283 * be different from the set of nodes where the workload's memory is currently
1286 * The bitmask is used to make smarter decisions on when to do NUMA page
1287 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1288 * are added when they cause over 6/16 of the maximum number of faults, but
1289 * only removed when they drop below 3/16.
1291 static void update_numa_active_node_mask(struct numa_group *numa_group)
1293 unsigned long faults, max_faults = 0;
1296 for_each_online_node(nid) {
1297 faults = group_faults_cpu(numa_group, nid);
1298 if (faults > max_faults)
1299 max_faults = faults;
1302 for_each_online_node(nid) {
1303 faults = group_faults_cpu(numa_group, nid);
1304 if (!node_isset(nid, numa_group->active_nodes)) {
1305 if (faults > max_faults * 6 / 16)
1306 node_set(nid, numa_group->active_nodes);
1307 } else if (faults < max_faults * 3 / 16)
1308 node_clear(nid, numa_group->active_nodes);
1313 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1314 * increments. The more local the fault statistics are, the higher the scan
1315 * period will be for the next scan window. If local/remote ratio is below
1316 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1317 * scan period will decrease
1319 #define NUMA_PERIOD_SLOTS 10
1320 #define NUMA_PERIOD_THRESHOLD 3
1323 * Increase the scan period (slow down scanning) if the majority of
1324 * our memory is already on our local node, or if the majority of
1325 * the page accesses are shared with other processes.
1326 * Otherwise, decrease the scan period.
1328 static void update_task_scan_period(struct task_struct *p,
1329 unsigned long shared, unsigned long private)
1331 unsigned int period_slot;
1335 unsigned long remote = p->numa_faults_locality[0];
1336 unsigned long local = p->numa_faults_locality[1];
1339 * If there were no record hinting faults then either the task is
1340 * completely idle or all activity is areas that are not of interest
1341 * to automatic numa balancing. Scan slower
1343 if (local + shared == 0) {
1344 p->numa_scan_period = min(p->numa_scan_period_max,
1345 p->numa_scan_period << 1);
1347 p->mm->numa_next_scan = jiffies +
1348 msecs_to_jiffies(p->numa_scan_period);
1354 * Prepare to scale scan period relative to the current period.
1355 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1356 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1357 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1359 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1360 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1361 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1362 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1365 diff = slot * period_slot;
1367 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1370 * Scale scan rate increases based on sharing. There is an
1371 * inverse relationship between the degree of sharing and
1372 * the adjustment made to the scanning period. Broadly
1373 * speaking the intent is that there is little point
1374 * scanning faster if shared accesses dominate as it may
1375 * simply bounce migrations uselessly
1377 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1378 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1381 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1382 task_scan_min(p), task_scan_max(p));
1383 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1386 static void task_numa_placement(struct task_struct *p)
1388 int seq, nid, max_nid = -1, max_group_nid = -1;
1389 unsigned long max_faults = 0, max_group_faults = 0;
1390 unsigned long fault_types[2] = { 0, 0 };
1391 spinlock_t *group_lock = NULL;
1393 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1394 if (p->numa_scan_seq == seq)
1396 p->numa_scan_seq = seq;
1397 p->numa_scan_period_max = task_scan_max(p);
1399 /* If the task is part of a group prevent parallel updates to group stats */
1400 if (p->numa_group) {
1401 group_lock = &p->numa_group->lock;
1402 spin_lock(group_lock);
1405 /* Find the node with the highest number of faults */
1406 for_each_online_node(nid) {
1407 unsigned long faults = 0, group_faults = 0;
1410 for (priv = 0; priv < 2; priv++) {
1413 i = task_faults_idx(nid, priv);
1414 diff = -p->numa_faults_memory[i];
1415 f_diff = -p->numa_faults_cpu[i];
1417 /* Decay existing window, copy faults since last scan */
1418 p->numa_faults_memory[i] >>= 1;
1419 p->numa_faults_memory[i] += p->numa_faults_buffer_memory[i];
1420 fault_types[priv] += p->numa_faults_buffer_memory[i];
1421 p->numa_faults_buffer_memory[i] = 0;
1423 p->numa_faults_cpu[i] >>= 1;
1424 p->numa_faults_cpu[i] += p->numa_faults_buffer_cpu[i];
1425 p->numa_faults_buffer_cpu[i] = 0;
1427 faults += p->numa_faults_memory[i];
1428 diff += p->numa_faults_memory[i];
1429 f_diff += p->numa_faults_cpu[i];
1430 p->total_numa_faults += diff;
1431 if (p->numa_group) {
1432 /* safe because we can only change our own group */
1433 p->numa_group->faults[i] += diff;
1434 p->numa_group->faults_cpu[i] += f_diff;
1435 p->numa_group->total_faults += diff;
1436 group_faults += p->numa_group->faults[i];
1440 if (faults > max_faults) {
1441 max_faults = faults;
1445 if (group_faults > max_group_faults) {
1446 max_group_faults = group_faults;
1447 max_group_nid = nid;
1451 update_task_scan_period(p, fault_types[0], fault_types[1]);
1453 if (p->numa_group) {
1454 update_numa_active_node_mask(p->numa_group);
1456 * If the preferred task and group nids are different,
1457 * iterate over the nodes again to find the best place.
1459 if (max_nid != max_group_nid) {
1460 unsigned long weight, max_weight = 0;
1462 for_each_online_node(nid) {
1463 weight = task_weight(p, nid) + group_weight(p, nid);
1464 if (weight > max_weight) {
1465 max_weight = weight;
1471 spin_unlock(group_lock);
1474 /* Preferred node as the node with the most faults */
1475 if (max_faults && max_nid != p->numa_preferred_nid) {
1476 /* Update the preferred nid and migrate task if possible */
1477 sched_setnuma(p, max_nid);
1478 numa_migrate_preferred(p);
1482 static inline int get_numa_group(struct numa_group *grp)
1484 return atomic_inc_not_zero(&grp->refcount);
1487 static inline void put_numa_group(struct numa_group *grp)
1489 if (atomic_dec_and_test(&grp->refcount))
1490 kfree_rcu(grp, rcu);
1493 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1496 struct numa_group *grp, *my_grp;
1497 struct task_struct *tsk;
1499 int cpu = cpupid_to_cpu(cpupid);
1502 if (unlikely(!p->numa_group)) {
1503 unsigned int size = sizeof(struct numa_group) +
1504 4*nr_node_ids*sizeof(unsigned long);
1506 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1510 atomic_set(&grp->refcount, 1);
1511 spin_lock_init(&grp->lock);
1512 INIT_LIST_HEAD(&grp->task_list);
1514 /* Second half of the array tracks nids where faults happen */
1515 grp->faults_cpu = grp->faults + 2 * nr_node_ids;
1517 node_set(task_node(current), grp->active_nodes);
1519 for (i = 0; i < 4*nr_node_ids; i++)
1520 grp->faults[i] = p->numa_faults_memory[i];
1522 grp->total_faults = p->total_numa_faults;
1524 list_add(&p->numa_entry, &grp->task_list);
1526 rcu_assign_pointer(p->numa_group, grp);
1530 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1532 if (!cpupid_match_pid(tsk, cpupid))
1535 grp = rcu_dereference(tsk->numa_group);
1539 my_grp = p->numa_group;
1544 * Only join the other group if its bigger; if we're the bigger group,
1545 * the other task will join us.
1547 if (my_grp->nr_tasks > grp->nr_tasks)
1551 * Tie-break on the grp address.
1553 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1556 /* Always join threads in the same process. */
1557 if (tsk->mm == current->mm)
1560 /* Simple filter to avoid false positives due to PID collisions */
1561 if (flags & TNF_SHARED)
1564 /* Update priv based on whether false sharing was detected */
1567 if (join && !get_numa_group(grp))
1575 double_lock(&my_grp->lock, &grp->lock);
1577 for (i = 0; i < 4*nr_node_ids; i++) {
1578 my_grp->faults[i] -= p->numa_faults_memory[i];
1579 grp->faults[i] += p->numa_faults_memory[i];
1581 my_grp->total_faults -= p->total_numa_faults;
1582 grp->total_faults += p->total_numa_faults;
1584 list_move(&p->numa_entry, &grp->task_list);
1588 spin_unlock(&my_grp->lock);
1589 spin_unlock(&grp->lock);
1591 rcu_assign_pointer(p->numa_group, grp);
1593 put_numa_group(my_grp);
1601 void task_numa_free(struct task_struct *p)
1603 struct numa_group *grp = p->numa_group;
1605 void *numa_faults = p->numa_faults_memory;
1608 spin_lock(&grp->lock);
1609 for (i = 0; i < 4*nr_node_ids; i++)
1610 grp->faults[i] -= p->numa_faults_memory[i];
1611 grp->total_faults -= p->total_numa_faults;
1613 list_del(&p->numa_entry);
1615 spin_unlock(&grp->lock);
1616 rcu_assign_pointer(p->numa_group, NULL);
1617 put_numa_group(grp);
1620 p->numa_faults_memory = NULL;
1621 p->numa_faults_buffer_memory = NULL;
1622 p->numa_faults_cpu= NULL;
1623 p->numa_faults_buffer_cpu = NULL;
1628 * Got a PROT_NONE fault for a page on @node.
1630 void task_numa_fault(int last_cpupid, int node, int pages, int flags)
1632 struct task_struct *p = current;
1633 bool migrated = flags & TNF_MIGRATED;
1634 int this_node = task_node(current);
1637 if (!numabalancing_enabled)
1640 /* for example, ksmd faulting in a user's mm */
1644 /* Do not worry about placement if exiting */
1645 if (p->state == TASK_DEAD)
1648 /* Allocate buffer to track faults on a per-node basis */
1649 if (unlikely(!p->numa_faults_memory)) {
1650 int size = sizeof(*p->numa_faults_memory) * 4 * nr_node_ids;
1652 /* numa_faults and numa_faults_buffer share the allocation */
1653 p->numa_faults_memory = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1654 if (!p->numa_faults_memory)
1657 BUG_ON(p->numa_faults_buffer_memory);
1658 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1659 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1660 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1661 p->total_numa_faults = 0;
1662 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1666 * First accesses are treated as private, otherwise consider accesses
1667 * to be private if the accessing pid has not changed
1669 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1672 priv = cpupid_match_pid(p, last_cpupid);
1673 if (!priv && !(flags & TNF_NO_GROUP))
1674 task_numa_group(p, last_cpupid, flags, &priv);
1677 task_numa_placement(p);
1680 * Retry task to preferred node migration periodically, in case it
1681 * case it previously failed, or the scheduler moved us.
1683 if (time_after(jiffies, p->numa_migrate_retry))
1684 numa_migrate_preferred(p);
1687 p->numa_pages_migrated += pages;
1689 p->numa_faults_buffer_memory[task_faults_idx(node, priv)] += pages;
1690 p->numa_faults_buffer_cpu[task_faults_idx(this_node, priv)] += pages;
1691 p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages;
1694 static void reset_ptenuma_scan(struct task_struct *p)
1696 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1697 p->mm->numa_scan_offset = 0;
1701 * The expensive part of numa migration is done from task_work context.
1702 * Triggered from task_tick_numa().
1704 void task_numa_work(struct callback_head *work)
1706 unsigned long migrate, next_scan, now = jiffies;
1707 struct task_struct *p = current;
1708 struct mm_struct *mm = p->mm;
1709 struct vm_area_struct *vma;
1710 unsigned long start, end;
1711 unsigned long nr_pte_updates = 0;
1714 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1716 work->next = work; /* protect against double add */
1718 * Who cares about NUMA placement when they're dying.
1720 * NOTE: make sure not to dereference p->mm before this check,
1721 * exit_task_work() happens _after_ exit_mm() so we could be called
1722 * without p->mm even though we still had it when we enqueued this
1725 if (p->flags & PF_EXITING)
1728 if (!mm->numa_next_scan) {
1729 mm->numa_next_scan = now +
1730 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1734 * Enforce maximal scan/migration frequency..
1736 migrate = mm->numa_next_scan;
1737 if (time_before(now, migrate))
1740 if (p->numa_scan_period == 0) {
1741 p->numa_scan_period_max = task_scan_max(p);
1742 p->numa_scan_period = task_scan_min(p);
1745 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1746 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1750 * Delay this task enough that another task of this mm will likely win
1751 * the next time around.
1753 p->node_stamp += 2 * TICK_NSEC;
1755 start = mm->numa_scan_offset;
1756 pages = sysctl_numa_balancing_scan_size;
1757 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1761 down_read(&mm->mmap_sem);
1762 vma = find_vma(mm, start);
1764 reset_ptenuma_scan(p);
1768 for (; vma; vma = vma->vm_next) {
1769 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1773 * Shared library pages mapped by multiple processes are not
1774 * migrated as it is expected they are cache replicated. Avoid
1775 * hinting faults in read-only file-backed mappings or the vdso
1776 * as migrating the pages will be of marginal benefit.
1779 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1783 * Skip inaccessible VMAs to avoid any confusion between
1784 * PROT_NONE and NUMA hinting ptes
1786 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1790 start = max(start, vma->vm_start);
1791 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1792 end = min(end, vma->vm_end);
1793 nr_pte_updates += change_prot_numa(vma, start, end);
1796 * Scan sysctl_numa_balancing_scan_size but ensure that
1797 * at least one PTE is updated so that unused virtual
1798 * address space is quickly skipped.
1801 pages -= (end - start) >> PAGE_SHIFT;
1806 } while (end != vma->vm_end);
1811 * It is possible to reach the end of the VMA list but the last few
1812 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1813 * would find the !migratable VMA on the next scan but not reset the
1814 * scanner to the start so check it now.
1817 mm->numa_scan_offset = start;
1819 reset_ptenuma_scan(p);
1820 up_read(&mm->mmap_sem);
1824 * Drive the periodic memory faults..
1826 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1828 struct callback_head *work = &curr->numa_work;
1832 * We don't care about NUMA placement if we don't have memory.
1834 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1838 * Using runtime rather than walltime has the dual advantage that
1839 * we (mostly) drive the selection from busy threads and that the
1840 * task needs to have done some actual work before we bother with
1843 now = curr->se.sum_exec_runtime;
1844 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1846 if (now - curr->node_stamp > period) {
1847 if (!curr->node_stamp)
1848 curr->numa_scan_period = task_scan_min(curr);
1849 curr->node_stamp += period;
1851 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1852 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1853 task_work_add(curr, work, true);
1858 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1862 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1866 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1869 #endif /* CONFIG_NUMA_BALANCING */
1872 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1874 update_load_add(&cfs_rq->load, se->load.weight);
1875 if (!parent_entity(se))
1876 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1878 if (entity_is_task(se)) {
1879 struct rq *rq = rq_of(cfs_rq);
1881 account_numa_enqueue(rq, task_of(se));
1882 list_add(&se->group_node, &rq->cfs_tasks);
1885 cfs_rq->nr_running++;
1889 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1891 update_load_sub(&cfs_rq->load, se->load.weight);
1892 if (!parent_entity(se))
1893 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1894 if (entity_is_task(se)) {
1895 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
1896 list_del_init(&se->group_node);
1898 cfs_rq->nr_running--;
1901 #ifdef CONFIG_FAIR_GROUP_SCHED
1903 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1908 * Use this CPU's actual weight instead of the last load_contribution
1909 * to gain a more accurate current total weight. See
1910 * update_cfs_rq_load_contribution().
1912 tg_weight = atomic_long_read(&tg->load_avg);
1913 tg_weight -= cfs_rq->tg_load_contrib;
1914 tg_weight += cfs_rq->load.weight;
1919 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1921 long tg_weight, load, shares;
1923 tg_weight = calc_tg_weight(tg, cfs_rq);
1924 load = cfs_rq->load.weight;
1926 shares = (tg->shares * load);
1928 shares /= tg_weight;
1930 if (shares < MIN_SHARES)
1931 shares = MIN_SHARES;
1932 if (shares > tg->shares)
1933 shares = tg->shares;
1937 # else /* CONFIG_SMP */
1938 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1942 # endif /* CONFIG_SMP */
1943 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1944 unsigned long weight)
1947 /* commit outstanding execution time */
1948 if (cfs_rq->curr == se)
1949 update_curr(cfs_rq);
1950 account_entity_dequeue(cfs_rq, se);
1953 update_load_set(&se->load, weight);
1956 account_entity_enqueue(cfs_rq, se);
1959 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1961 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1963 struct task_group *tg;
1964 struct sched_entity *se;
1968 se = tg->se[cpu_of(rq_of(cfs_rq))];
1969 if (!se || throttled_hierarchy(cfs_rq))
1972 if (likely(se->load.weight == tg->shares))
1975 shares = calc_cfs_shares(cfs_rq, tg);
1977 reweight_entity(cfs_rq_of(se), se, shares);
1979 #else /* CONFIG_FAIR_GROUP_SCHED */
1980 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1983 #endif /* CONFIG_FAIR_GROUP_SCHED */
1987 * We choose a half-life close to 1 scheduling period.
1988 * Note: The tables below are dependent on this value.
1990 #define LOAD_AVG_PERIOD 32
1991 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1992 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1994 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1995 static const u32 runnable_avg_yN_inv[] = {
1996 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1997 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1998 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1999 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2000 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2001 0x85aac367, 0x82cd8698,
2005 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2006 * over-estimates when re-combining.
2008 static const u32 runnable_avg_yN_sum[] = {
2009 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2010 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2011 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2016 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2018 static __always_inline u64 decay_load(u64 val, u64 n)
2020 unsigned int local_n;
2024 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2027 /* after bounds checking we can collapse to 32-bit */
2031 * As y^PERIOD = 1/2, we can combine
2032 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2033 * With a look-up table which covers k^n (n<PERIOD)
2035 * To achieve constant time decay_load.
2037 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2038 val >>= local_n / LOAD_AVG_PERIOD;
2039 local_n %= LOAD_AVG_PERIOD;
2042 val *= runnable_avg_yN_inv[local_n];
2043 /* We don't use SRR here since we always want to round down. */
2048 * For updates fully spanning n periods, the contribution to runnable
2049 * average will be: \Sum 1024*y^n
2051 * We can compute this reasonably efficiently by combining:
2052 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2054 static u32 __compute_runnable_contrib(u64 n)
2058 if (likely(n <= LOAD_AVG_PERIOD))
2059 return runnable_avg_yN_sum[n];
2060 else if (unlikely(n >= LOAD_AVG_MAX_N))
2061 return LOAD_AVG_MAX;
2063 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2065 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2066 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2068 n -= LOAD_AVG_PERIOD;
2069 } while (n > LOAD_AVG_PERIOD);
2071 contrib = decay_load(contrib, n);
2072 return contrib + runnable_avg_yN_sum[n];
2076 * We can represent the historical contribution to runnable average as the
2077 * coefficients of a geometric series. To do this we sub-divide our runnable
2078 * history into segments of approximately 1ms (1024us); label the segment that
2079 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2081 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2083 * (now) (~1ms ago) (~2ms ago)
2085 * Let u_i denote the fraction of p_i that the entity was runnable.
2087 * We then designate the fractions u_i as our co-efficients, yielding the
2088 * following representation of historical load:
2089 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2091 * We choose y based on the with of a reasonably scheduling period, fixing:
2094 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2095 * approximately half as much as the contribution to load within the last ms
2098 * When a period "rolls over" and we have new u_0`, multiplying the previous
2099 * sum again by y is sufficient to update:
2100 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2101 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2103 static __always_inline int __update_entity_runnable_avg(u64 now,
2104 struct sched_avg *sa,
2108 u32 runnable_contrib;
2109 int delta_w, decayed = 0;
2111 delta = now - sa->last_runnable_update;
2113 * This should only happen when time goes backwards, which it
2114 * unfortunately does during sched clock init when we swap over to TSC.
2116 if ((s64)delta < 0) {
2117 sa->last_runnable_update = now;
2122 * Use 1024ns as the unit of measurement since it's a reasonable
2123 * approximation of 1us and fast to compute.
2128 sa->last_runnable_update = now;
2130 /* delta_w is the amount already accumulated against our next period */
2131 delta_w = sa->runnable_avg_period % 1024;
2132 if (delta + delta_w >= 1024) {
2133 /* period roll-over */
2137 * Now that we know we're crossing a period boundary, figure
2138 * out how much from delta we need to complete the current
2139 * period and accrue it.
2141 delta_w = 1024 - delta_w;
2143 sa->runnable_avg_sum += delta_w;
2144 sa->runnable_avg_period += delta_w;
2148 /* Figure out how many additional periods this update spans */
2149 periods = delta / 1024;
2152 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2154 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2157 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2158 runnable_contrib = __compute_runnable_contrib(periods);
2160 sa->runnable_avg_sum += runnable_contrib;
2161 sa->runnable_avg_period += runnable_contrib;
2164 /* Remainder of delta accrued against u_0` */
2166 sa->runnable_avg_sum += delta;
2167 sa->runnable_avg_period += delta;
2172 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2173 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2175 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2176 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2178 decays -= se->avg.decay_count;
2182 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2183 se->avg.decay_count = 0;
2188 #ifdef CONFIG_FAIR_GROUP_SCHED
2189 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2192 struct task_group *tg = cfs_rq->tg;
2195 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2196 tg_contrib -= cfs_rq->tg_load_contrib;
2198 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2199 atomic_long_add(tg_contrib, &tg->load_avg);
2200 cfs_rq->tg_load_contrib += tg_contrib;
2205 * Aggregate cfs_rq runnable averages into an equivalent task_group
2206 * representation for computing load contributions.
2208 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2209 struct cfs_rq *cfs_rq)
2211 struct task_group *tg = cfs_rq->tg;
2214 /* The fraction of a cpu used by this cfs_rq */
2215 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2216 sa->runnable_avg_period + 1);
2217 contrib -= cfs_rq->tg_runnable_contrib;
2219 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2220 atomic_add(contrib, &tg->runnable_avg);
2221 cfs_rq->tg_runnable_contrib += contrib;
2225 static inline void __update_group_entity_contrib(struct sched_entity *se)
2227 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2228 struct task_group *tg = cfs_rq->tg;
2233 contrib = cfs_rq->tg_load_contrib * tg->shares;
2234 se->avg.load_avg_contrib = div_u64(contrib,
2235 atomic_long_read(&tg->load_avg) + 1);
2238 * For group entities we need to compute a correction term in the case
2239 * that they are consuming <1 cpu so that we would contribute the same
2240 * load as a task of equal weight.
2242 * Explicitly co-ordinating this measurement would be expensive, but
2243 * fortunately the sum of each cpus contribution forms a usable
2244 * lower-bound on the true value.
2246 * Consider the aggregate of 2 contributions. Either they are disjoint
2247 * (and the sum represents true value) or they are disjoint and we are
2248 * understating by the aggregate of their overlap.
2250 * Extending this to N cpus, for a given overlap, the maximum amount we
2251 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2252 * cpus that overlap for this interval and w_i is the interval width.
2254 * On a small machine; the first term is well-bounded which bounds the
2255 * total error since w_i is a subset of the period. Whereas on a
2256 * larger machine, while this first term can be larger, if w_i is the
2257 * of consequential size guaranteed to see n_i*w_i quickly converge to
2258 * our upper bound of 1-cpu.
2260 runnable_avg = atomic_read(&tg->runnable_avg);
2261 if (runnable_avg < NICE_0_LOAD) {
2262 se->avg.load_avg_contrib *= runnable_avg;
2263 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2267 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2268 int force_update) {}
2269 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2270 struct cfs_rq *cfs_rq) {}
2271 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2274 static inline void __update_task_entity_contrib(struct sched_entity *se)
2278 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2279 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2280 contrib /= (se->avg.runnable_avg_period + 1);
2281 se->avg.load_avg_contrib = scale_load(contrib);
2284 /* Compute the current contribution to load_avg by se, return any delta */
2285 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2287 long old_contrib = se->avg.load_avg_contrib;
2289 if (entity_is_task(se)) {
2290 __update_task_entity_contrib(se);
2292 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2293 __update_group_entity_contrib(se);
2296 return se->avg.load_avg_contrib - old_contrib;
2299 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2302 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2303 cfs_rq->blocked_load_avg -= load_contrib;
2305 cfs_rq->blocked_load_avg = 0;
2308 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2310 /* Update a sched_entity's runnable average */
2311 static inline void update_entity_load_avg(struct sched_entity *se,
2314 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2319 * For a group entity we need to use their owned cfs_rq_clock_task() in
2320 * case they are the parent of a throttled hierarchy.
2322 if (entity_is_task(se))
2323 now = cfs_rq_clock_task(cfs_rq);
2325 now = cfs_rq_clock_task(group_cfs_rq(se));
2327 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2330 contrib_delta = __update_entity_load_avg_contrib(se);
2336 cfs_rq->runnable_load_avg += contrib_delta;
2338 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2342 * Decay the load contributed by all blocked children and account this so that
2343 * their contribution may appropriately discounted when they wake up.
2345 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2347 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2350 decays = now - cfs_rq->last_decay;
2351 if (!decays && !force_update)
2354 if (atomic_long_read(&cfs_rq->removed_load)) {
2355 unsigned long removed_load;
2356 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2357 subtract_blocked_load_contrib(cfs_rq, removed_load);
2361 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2363 atomic64_add(decays, &cfs_rq->decay_counter);
2364 cfs_rq->last_decay = now;
2367 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2370 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2372 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2373 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2376 /* Add the load generated by se into cfs_rq's child load-average */
2377 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2378 struct sched_entity *se,
2382 * We track migrations using entity decay_count <= 0, on a wake-up
2383 * migration we use a negative decay count to track the remote decays
2384 * accumulated while sleeping.
2386 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2387 * are seen by enqueue_entity_load_avg() as a migration with an already
2388 * constructed load_avg_contrib.
2390 if (unlikely(se->avg.decay_count <= 0)) {
2391 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2392 if (se->avg.decay_count) {
2394 * In a wake-up migration we have to approximate the
2395 * time sleeping. This is because we can't synchronize
2396 * clock_task between the two cpus, and it is not
2397 * guaranteed to be read-safe. Instead, we can
2398 * approximate this using our carried decays, which are
2399 * explicitly atomically readable.
2401 se->avg.last_runnable_update -= (-se->avg.decay_count)
2403 update_entity_load_avg(se, 0);
2404 /* Indicate that we're now synchronized and on-rq */
2405 se->avg.decay_count = 0;
2409 __synchronize_entity_decay(se);
2412 /* migrated tasks did not contribute to our blocked load */
2414 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2415 update_entity_load_avg(se, 0);
2418 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2419 /* we force update consideration on load-balancer moves */
2420 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2424 * Remove se's load from this cfs_rq child load-average, if the entity is
2425 * transitioning to a blocked state we track its projected decay using
2428 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2429 struct sched_entity *se,
2432 update_entity_load_avg(se, 1);
2433 /* we force update consideration on load-balancer moves */
2434 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2436 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2438 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2439 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2440 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2444 * Update the rq's load with the elapsed running time before entering
2445 * idle. if the last scheduled task is not a CFS task, idle_enter will
2446 * be the only way to update the runnable statistic.
2448 void idle_enter_fair(struct rq *this_rq)
2450 update_rq_runnable_avg(this_rq, 1);
2454 * Update the rq's load with the elapsed idle time before a task is
2455 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2456 * be the only way to update the runnable statistic.
2458 void idle_exit_fair(struct rq *this_rq)
2460 update_rq_runnable_avg(this_rq, 0);
2464 static inline void update_entity_load_avg(struct sched_entity *se,
2465 int update_cfs_rq) {}
2466 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2467 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2468 struct sched_entity *se,
2470 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2471 struct sched_entity *se,
2473 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2474 int force_update) {}
2477 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2479 #ifdef CONFIG_SCHEDSTATS
2480 struct task_struct *tsk = NULL;
2482 if (entity_is_task(se))
2485 if (se->statistics.sleep_start) {
2486 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2491 if (unlikely(delta > se->statistics.sleep_max))
2492 se->statistics.sleep_max = delta;
2494 se->statistics.sleep_start = 0;
2495 se->statistics.sum_sleep_runtime += delta;
2498 account_scheduler_latency(tsk, delta >> 10, 1);
2499 trace_sched_stat_sleep(tsk, delta);
2502 if (se->statistics.block_start) {
2503 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2508 if (unlikely(delta > se->statistics.block_max))
2509 se->statistics.block_max = delta;
2511 se->statistics.block_start = 0;
2512 se->statistics.sum_sleep_runtime += delta;
2515 if (tsk->in_iowait) {
2516 se->statistics.iowait_sum += delta;
2517 se->statistics.iowait_count++;
2518 trace_sched_stat_iowait(tsk, delta);
2521 trace_sched_stat_blocked(tsk, delta);
2524 * Blocking time is in units of nanosecs, so shift by
2525 * 20 to get a milliseconds-range estimation of the
2526 * amount of time that the task spent sleeping:
2528 if (unlikely(prof_on == SLEEP_PROFILING)) {
2529 profile_hits(SLEEP_PROFILING,
2530 (void *)get_wchan(tsk),
2533 account_scheduler_latency(tsk, delta >> 10, 0);
2539 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2541 #ifdef CONFIG_SCHED_DEBUG
2542 s64 d = se->vruntime - cfs_rq->min_vruntime;
2547 if (d > 3*sysctl_sched_latency)
2548 schedstat_inc(cfs_rq, nr_spread_over);
2553 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2555 u64 vruntime = cfs_rq->min_vruntime;
2558 * The 'current' period is already promised to the current tasks,
2559 * however the extra weight of the new task will slow them down a
2560 * little, place the new task so that it fits in the slot that
2561 * stays open at the end.
2563 if (initial && sched_feat(START_DEBIT))
2564 vruntime += sched_vslice(cfs_rq, se);
2566 /* sleeps up to a single latency don't count. */
2568 unsigned long thresh = sysctl_sched_latency;
2571 * Halve their sleep time's effect, to allow
2572 * for a gentler effect of sleepers:
2574 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2580 /* ensure we never gain time by being placed backwards. */
2581 se->vruntime = max_vruntime(se->vruntime, vruntime);
2584 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2587 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2590 * Update the normalized vruntime before updating min_vruntime
2591 * through calling update_curr().
2593 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2594 se->vruntime += cfs_rq->min_vruntime;
2597 * Update run-time statistics of the 'current'.
2599 update_curr(cfs_rq);
2600 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2601 account_entity_enqueue(cfs_rq, se);
2602 update_cfs_shares(cfs_rq);
2604 if (flags & ENQUEUE_WAKEUP) {
2605 place_entity(cfs_rq, se, 0);
2606 enqueue_sleeper(cfs_rq, se);
2609 update_stats_enqueue(cfs_rq, se);
2610 check_spread(cfs_rq, se);
2611 if (se != cfs_rq->curr)
2612 __enqueue_entity(cfs_rq, se);
2615 if (cfs_rq->nr_running == 1) {
2616 list_add_leaf_cfs_rq(cfs_rq);
2617 check_enqueue_throttle(cfs_rq);
2621 static void __clear_buddies_last(struct sched_entity *se)
2623 for_each_sched_entity(se) {
2624 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2625 if (cfs_rq->last == se)
2626 cfs_rq->last = NULL;
2632 static void __clear_buddies_next(struct sched_entity *se)
2634 for_each_sched_entity(se) {
2635 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2636 if (cfs_rq->next == se)
2637 cfs_rq->next = NULL;
2643 static void __clear_buddies_skip(struct sched_entity *se)
2645 for_each_sched_entity(se) {
2646 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2647 if (cfs_rq->skip == se)
2648 cfs_rq->skip = NULL;
2654 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2656 if (cfs_rq->last == se)
2657 __clear_buddies_last(se);
2659 if (cfs_rq->next == se)
2660 __clear_buddies_next(se);
2662 if (cfs_rq->skip == se)
2663 __clear_buddies_skip(se);
2666 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2669 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2672 * Update run-time statistics of the 'current'.
2674 update_curr(cfs_rq);
2675 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2677 update_stats_dequeue(cfs_rq, se);
2678 if (flags & DEQUEUE_SLEEP) {
2679 #ifdef CONFIG_SCHEDSTATS
2680 if (entity_is_task(se)) {
2681 struct task_struct *tsk = task_of(se);
2683 if (tsk->state & TASK_INTERRUPTIBLE)
2684 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2685 if (tsk->state & TASK_UNINTERRUPTIBLE)
2686 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2691 clear_buddies(cfs_rq, se);
2693 if (se != cfs_rq->curr)
2694 __dequeue_entity(cfs_rq, se);
2696 account_entity_dequeue(cfs_rq, se);
2699 * Normalize the entity after updating the min_vruntime because the
2700 * update can refer to the ->curr item and we need to reflect this
2701 * movement in our normalized position.
2703 if (!(flags & DEQUEUE_SLEEP))
2704 se->vruntime -= cfs_rq->min_vruntime;
2706 /* return excess runtime on last dequeue */
2707 return_cfs_rq_runtime(cfs_rq);
2709 update_min_vruntime(cfs_rq);
2710 update_cfs_shares(cfs_rq);
2714 * Preempt the current task with a newly woken task if needed:
2717 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2719 unsigned long ideal_runtime, delta_exec;
2720 struct sched_entity *se;
2723 ideal_runtime = sched_slice(cfs_rq, curr);
2724 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2725 if (delta_exec > ideal_runtime) {
2726 resched_task(rq_of(cfs_rq)->curr);
2728 * The current task ran long enough, ensure it doesn't get
2729 * re-elected due to buddy favours.
2731 clear_buddies(cfs_rq, curr);
2736 * Ensure that a task that missed wakeup preemption by a
2737 * narrow margin doesn't have to wait for a full slice.
2738 * This also mitigates buddy induced latencies under load.
2740 if (delta_exec < sysctl_sched_min_granularity)
2743 se = __pick_first_entity(cfs_rq);
2744 delta = curr->vruntime - se->vruntime;
2749 if (delta > ideal_runtime)
2750 resched_task(rq_of(cfs_rq)->curr);
2754 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2756 /* 'current' is not kept within the tree. */
2759 * Any task has to be enqueued before it get to execute on
2760 * a CPU. So account for the time it spent waiting on the
2763 update_stats_wait_end(cfs_rq, se);
2764 __dequeue_entity(cfs_rq, se);
2767 update_stats_curr_start(cfs_rq, se);
2769 #ifdef CONFIG_SCHEDSTATS
2771 * Track our maximum slice length, if the CPU's load is at
2772 * least twice that of our own weight (i.e. dont track it
2773 * when there are only lesser-weight tasks around):
2775 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2776 se->statistics.slice_max = max(se->statistics.slice_max,
2777 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2780 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2784 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2787 * Pick the next process, keeping these things in mind, in this order:
2788 * 1) keep things fair between processes/task groups
2789 * 2) pick the "next" process, since someone really wants that to run
2790 * 3) pick the "last" process, for cache locality
2791 * 4) do not run the "skip" process, if something else is available
2793 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2795 struct sched_entity *se = __pick_first_entity(cfs_rq);
2796 struct sched_entity *left = se;
2799 * Avoid running the skip buddy, if running something else can
2800 * be done without getting too unfair.
2802 if (cfs_rq->skip == se) {
2803 struct sched_entity *second = __pick_next_entity(se);
2804 if (second && wakeup_preempt_entity(second, left) < 1)
2809 * Prefer last buddy, try to return the CPU to a preempted task.
2811 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2815 * Someone really wants this to run. If it's not unfair, run it.
2817 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2820 clear_buddies(cfs_rq, se);
2825 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2827 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2830 * If still on the runqueue then deactivate_task()
2831 * was not called and update_curr() has to be done:
2834 update_curr(cfs_rq);
2836 /* throttle cfs_rqs exceeding runtime */
2837 check_cfs_rq_runtime(cfs_rq);
2839 check_spread(cfs_rq, prev);
2841 update_stats_wait_start(cfs_rq, prev);
2842 /* Put 'current' back into the tree. */
2843 __enqueue_entity(cfs_rq, prev);
2844 /* in !on_rq case, update occurred at dequeue */
2845 update_entity_load_avg(prev, 1);
2847 cfs_rq->curr = NULL;
2851 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2854 * Update run-time statistics of the 'current'.
2856 update_curr(cfs_rq);
2859 * Ensure that runnable average is periodically updated.
2861 update_entity_load_avg(curr, 1);
2862 update_cfs_rq_blocked_load(cfs_rq, 1);
2863 update_cfs_shares(cfs_rq);
2865 #ifdef CONFIG_SCHED_HRTICK
2867 * queued ticks are scheduled to match the slice, so don't bother
2868 * validating it and just reschedule.
2871 resched_task(rq_of(cfs_rq)->curr);
2875 * don't let the period tick interfere with the hrtick preemption
2877 if (!sched_feat(DOUBLE_TICK) &&
2878 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2882 if (cfs_rq->nr_running > 1)
2883 check_preempt_tick(cfs_rq, curr);
2887 /**************************************************
2888 * CFS bandwidth control machinery
2891 #ifdef CONFIG_CFS_BANDWIDTH
2893 #ifdef HAVE_JUMP_LABEL
2894 static struct static_key __cfs_bandwidth_used;
2896 static inline bool cfs_bandwidth_used(void)
2898 return static_key_false(&__cfs_bandwidth_used);
2901 void cfs_bandwidth_usage_inc(void)
2903 static_key_slow_inc(&__cfs_bandwidth_used);
2906 void cfs_bandwidth_usage_dec(void)
2908 static_key_slow_dec(&__cfs_bandwidth_used);
2910 #else /* HAVE_JUMP_LABEL */
2911 static bool cfs_bandwidth_used(void)
2916 void cfs_bandwidth_usage_inc(void) {}
2917 void cfs_bandwidth_usage_dec(void) {}
2918 #endif /* HAVE_JUMP_LABEL */
2921 * default period for cfs group bandwidth.
2922 * default: 0.1s, units: nanoseconds
2924 static inline u64 default_cfs_period(void)
2926 return 100000000ULL;
2929 static inline u64 sched_cfs_bandwidth_slice(void)
2931 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2935 * Replenish runtime according to assigned quota and update expiration time.
2936 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2937 * additional synchronization around rq->lock.
2939 * requires cfs_b->lock
2941 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2945 if (cfs_b->quota == RUNTIME_INF)
2948 now = sched_clock_cpu(smp_processor_id());
2949 cfs_b->runtime = cfs_b->quota;
2950 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2953 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2955 return &tg->cfs_bandwidth;
2958 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2959 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2961 if (unlikely(cfs_rq->throttle_count))
2962 return cfs_rq->throttled_clock_task;
2964 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2967 /* returns 0 on failure to allocate runtime */
2968 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2970 struct task_group *tg = cfs_rq->tg;
2971 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2972 u64 amount = 0, min_amount, expires;
2974 /* note: this is a positive sum as runtime_remaining <= 0 */
2975 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2977 raw_spin_lock(&cfs_b->lock);
2978 if (cfs_b->quota == RUNTIME_INF)
2979 amount = min_amount;
2982 * If the bandwidth pool has become inactive, then at least one
2983 * period must have elapsed since the last consumption.
2984 * Refresh the global state and ensure bandwidth timer becomes
2987 if (!cfs_b->timer_active) {
2988 __refill_cfs_bandwidth_runtime(cfs_b);
2989 __start_cfs_bandwidth(cfs_b);
2992 if (cfs_b->runtime > 0) {
2993 amount = min(cfs_b->runtime, min_amount);
2994 cfs_b->runtime -= amount;
2998 expires = cfs_b->runtime_expires;
2999 raw_spin_unlock(&cfs_b->lock);
3001 cfs_rq->runtime_remaining += amount;
3003 * we may have advanced our local expiration to account for allowed
3004 * spread between our sched_clock and the one on which runtime was
3007 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3008 cfs_rq->runtime_expires = expires;
3010 return cfs_rq->runtime_remaining > 0;
3014 * Note: This depends on the synchronization provided by sched_clock and the
3015 * fact that rq->clock snapshots this value.
3017 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3019 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3021 /* if the deadline is ahead of our clock, nothing to do */
3022 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3025 if (cfs_rq->runtime_remaining < 0)
3029 * If the local deadline has passed we have to consider the
3030 * possibility that our sched_clock is 'fast' and the global deadline
3031 * has not truly expired.
3033 * Fortunately we can check determine whether this the case by checking
3034 * whether the global deadline has advanced.
3037 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
3038 /* extend local deadline, drift is bounded above by 2 ticks */
3039 cfs_rq->runtime_expires += TICK_NSEC;
3041 /* global deadline is ahead, expiration has passed */
3042 cfs_rq->runtime_remaining = 0;
3046 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3048 /* dock delta_exec before expiring quota (as it could span periods) */
3049 cfs_rq->runtime_remaining -= delta_exec;
3050 expire_cfs_rq_runtime(cfs_rq);
3052 if (likely(cfs_rq->runtime_remaining > 0))
3056 * if we're unable to extend our runtime we resched so that the active
3057 * hierarchy can be throttled
3059 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3060 resched_task(rq_of(cfs_rq)->curr);
3063 static __always_inline
3064 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3066 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3069 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3072 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3074 return cfs_bandwidth_used() && cfs_rq->throttled;
3077 /* check whether cfs_rq, or any parent, is throttled */
3078 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3080 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3084 * Ensure that neither of the group entities corresponding to src_cpu or
3085 * dest_cpu are members of a throttled hierarchy when performing group
3086 * load-balance operations.
3088 static inline int throttled_lb_pair(struct task_group *tg,
3089 int src_cpu, int dest_cpu)
3091 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3093 src_cfs_rq = tg->cfs_rq[src_cpu];
3094 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3096 return throttled_hierarchy(src_cfs_rq) ||
3097 throttled_hierarchy(dest_cfs_rq);
3100 /* updated child weight may affect parent so we have to do this bottom up */
3101 static int tg_unthrottle_up(struct task_group *tg, void *data)
3103 struct rq *rq = data;
3104 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3106 cfs_rq->throttle_count--;
3108 if (!cfs_rq->throttle_count) {
3109 /* adjust cfs_rq_clock_task() */
3110 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3111 cfs_rq->throttled_clock_task;
3118 static int tg_throttle_down(struct task_group *tg, void *data)
3120 struct rq *rq = data;
3121 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3123 /* group is entering throttled state, stop time */
3124 if (!cfs_rq->throttle_count)
3125 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3126 cfs_rq->throttle_count++;
3131 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3133 struct rq *rq = rq_of(cfs_rq);
3134 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3135 struct sched_entity *se;
3136 long task_delta, dequeue = 1;
3138 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3140 /* freeze hierarchy runnable averages while throttled */
3142 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3145 task_delta = cfs_rq->h_nr_running;
3146 for_each_sched_entity(se) {
3147 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3148 /* throttled entity or throttle-on-deactivate */
3153 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3154 qcfs_rq->h_nr_running -= task_delta;
3156 if (qcfs_rq->load.weight)
3161 rq->nr_running -= task_delta;
3163 cfs_rq->throttled = 1;
3164 cfs_rq->throttled_clock = rq_clock(rq);
3165 raw_spin_lock(&cfs_b->lock);
3166 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3167 if (!cfs_b->timer_active)
3168 __start_cfs_bandwidth(cfs_b);
3169 raw_spin_unlock(&cfs_b->lock);
3172 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3174 struct rq *rq = rq_of(cfs_rq);
3175 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3176 struct sched_entity *se;
3180 se = cfs_rq->tg->se[cpu_of(rq)];
3182 cfs_rq->throttled = 0;
3184 update_rq_clock(rq);
3186 raw_spin_lock(&cfs_b->lock);
3187 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3188 list_del_rcu(&cfs_rq->throttled_list);
3189 raw_spin_unlock(&cfs_b->lock);
3191 /* update hierarchical throttle state */
3192 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3194 if (!cfs_rq->load.weight)
3197 task_delta = cfs_rq->h_nr_running;
3198 for_each_sched_entity(se) {
3202 cfs_rq = cfs_rq_of(se);
3204 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3205 cfs_rq->h_nr_running += task_delta;
3207 if (cfs_rq_throttled(cfs_rq))
3212 rq->nr_running += task_delta;
3214 /* determine whether we need to wake up potentially idle cpu */
3215 if (rq->curr == rq->idle && rq->cfs.nr_running)
3216 resched_task(rq->curr);
3219 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3220 u64 remaining, u64 expires)
3222 struct cfs_rq *cfs_rq;
3223 u64 runtime = remaining;
3226 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3228 struct rq *rq = rq_of(cfs_rq);
3230 raw_spin_lock(&rq->lock);
3231 if (!cfs_rq_throttled(cfs_rq))
3234 runtime = -cfs_rq->runtime_remaining + 1;
3235 if (runtime > remaining)
3236 runtime = remaining;
3237 remaining -= runtime;
3239 cfs_rq->runtime_remaining += runtime;
3240 cfs_rq->runtime_expires = expires;
3242 /* we check whether we're throttled above */
3243 if (cfs_rq->runtime_remaining > 0)
3244 unthrottle_cfs_rq(cfs_rq);
3247 raw_spin_unlock(&rq->lock);
3258 * Responsible for refilling a task_group's bandwidth and unthrottling its
3259 * cfs_rqs as appropriate. If there has been no activity within the last
3260 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3261 * used to track this state.
3263 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3265 u64 runtime, runtime_expires;
3266 int idle = 1, throttled;
3268 raw_spin_lock(&cfs_b->lock);
3269 /* no need to continue the timer with no bandwidth constraint */
3270 if (cfs_b->quota == RUNTIME_INF)
3273 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3274 /* idle depends on !throttled (for the case of a large deficit) */
3275 idle = cfs_b->idle && !throttled;
3276 cfs_b->nr_periods += overrun;
3278 /* if we're going inactive then everything else can be deferred */
3283 * if we have relooped after returning idle once, we need to update our
3284 * status as actually running, so that other cpus doing
3285 * __start_cfs_bandwidth will stop trying to cancel us.
3287 cfs_b->timer_active = 1;
3289 __refill_cfs_bandwidth_runtime(cfs_b);
3292 /* mark as potentially idle for the upcoming period */
3297 /* account preceding periods in which throttling occurred */
3298 cfs_b->nr_throttled += overrun;
3301 * There are throttled entities so we must first use the new bandwidth
3302 * to unthrottle them before making it generally available. This
3303 * ensures that all existing debts will be paid before a new cfs_rq is
3306 runtime = cfs_b->runtime;
3307 runtime_expires = cfs_b->runtime_expires;
3311 * This check is repeated as we are holding onto the new bandwidth
3312 * while we unthrottle. This can potentially race with an unthrottled
3313 * group trying to acquire new bandwidth from the global pool.
3315 while (throttled && runtime > 0) {
3316 raw_spin_unlock(&cfs_b->lock);
3317 /* we can't nest cfs_b->lock while distributing bandwidth */
3318 runtime = distribute_cfs_runtime(cfs_b, runtime,
3320 raw_spin_lock(&cfs_b->lock);
3322 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3325 /* return (any) remaining runtime */
3326 cfs_b->runtime = runtime;
3328 * While we are ensured activity in the period following an
3329 * unthrottle, this also covers the case in which the new bandwidth is
3330 * insufficient to cover the existing bandwidth deficit. (Forcing the
3331 * timer to remain active while there are any throttled entities.)
3336 cfs_b->timer_active = 0;
3337 raw_spin_unlock(&cfs_b->lock);
3342 /* a cfs_rq won't donate quota below this amount */
3343 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3344 /* minimum remaining period time to redistribute slack quota */
3345 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3346 /* how long we wait to gather additional slack before distributing */
3347 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3350 * Are we near the end of the current quota period?
3352 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3353 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3354 * migrate_hrtimers, base is never cleared, so we are fine.
3356 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3358 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3361 /* if the call-back is running a quota refresh is already occurring */
3362 if (hrtimer_callback_running(refresh_timer))
3365 /* is a quota refresh about to occur? */
3366 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3367 if (remaining < min_expire)
3373 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3375 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3377 /* if there's a quota refresh soon don't bother with slack */
3378 if (runtime_refresh_within(cfs_b, min_left))
3381 start_bandwidth_timer(&cfs_b->slack_timer,
3382 ns_to_ktime(cfs_bandwidth_slack_period));
3385 /* we know any runtime found here is valid as update_curr() precedes return */
3386 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3388 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3389 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3391 if (slack_runtime <= 0)
3394 raw_spin_lock(&cfs_b->lock);
3395 if (cfs_b->quota != RUNTIME_INF &&
3396 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3397 cfs_b->runtime += slack_runtime;
3399 /* we are under rq->lock, defer unthrottling using a timer */
3400 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3401 !list_empty(&cfs_b->throttled_cfs_rq))
3402 start_cfs_slack_bandwidth(cfs_b);
3404 raw_spin_unlock(&cfs_b->lock);
3406 /* even if it's not valid for return we don't want to try again */
3407 cfs_rq->runtime_remaining -= slack_runtime;
3410 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3412 if (!cfs_bandwidth_used())
3415 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3418 __return_cfs_rq_runtime(cfs_rq);
3422 * This is done with a timer (instead of inline with bandwidth return) since
3423 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3425 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3427 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3430 /* confirm we're still not at a refresh boundary */
3431 raw_spin_lock(&cfs_b->lock);
3432 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3433 raw_spin_unlock(&cfs_b->lock);
3437 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3438 runtime = cfs_b->runtime;
3441 expires = cfs_b->runtime_expires;
3442 raw_spin_unlock(&cfs_b->lock);
3447 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3449 raw_spin_lock(&cfs_b->lock);
3450 if (expires == cfs_b->runtime_expires)
3451 cfs_b->runtime = runtime;
3452 raw_spin_unlock(&cfs_b->lock);
3456 * When a group wakes up we want to make sure that its quota is not already
3457 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3458 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3460 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3462 if (!cfs_bandwidth_used())
3465 /* an active group must be handled by the update_curr()->put() path */
3466 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3469 /* ensure the group is not already throttled */
3470 if (cfs_rq_throttled(cfs_rq))
3473 /* update runtime allocation */
3474 account_cfs_rq_runtime(cfs_rq, 0);
3475 if (cfs_rq->runtime_remaining <= 0)
3476 throttle_cfs_rq(cfs_rq);
3479 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3480 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3482 if (!cfs_bandwidth_used())
3485 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3489 * it's possible for a throttled entity to be forced into a running
3490 * state (e.g. set_curr_task), in this case we're finished.
3492 if (cfs_rq_throttled(cfs_rq))
3495 throttle_cfs_rq(cfs_rq);
3498 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3500 struct cfs_bandwidth *cfs_b =
3501 container_of(timer, struct cfs_bandwidth, slack_timer);
3502 do_sched_cfs_slack_timer(cfs_b);
3504 return HRTIMER_NORESTART;
3507 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3509 struct cfs_bandwidth *cfs_b =
3510 container_of(timer, struct cfs_bandwidth, period_timer);
3516 now = hrtimer_cb_get_time(timer);
3517 overrun = hrtimer_forward(timer, now, cfs_b->period);
3522 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3525 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3528 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3530 raw_spin_lock_init(&cfs_b->lock);
3532 cfs_b->quota = RUNTIME_INF;
3533 cfs_b->period = ns_to_ktime(default_cfs_period());
3535 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3536 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3537 cfs_b->period_timer.function = sched_cfs_period_timer;
3538 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3539 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3542 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3544 cfs_rq->runtime_enabled = 0;
3545 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3548 /* requires cfs_b->lock, may release to reprogram timer */
3549 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3552 * The timer may be active because we're trying to set a new bandwidth
3553 * period or because we're racing with the tear-down path
3554 * (timer_active==0 becomes visible before the hrtimer call-back
3555 * terminates). In either case we ensure that it's re-programmed
3557 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3558 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3559 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3560 raw_spin_unlock(&cfs_b->lock);
3562 raw_spin_lock(&cfs_b->lock);
3563 /* if someone else restarted the timer then we're done */
3564 if (cfs_b->timer_active)
3568 cfs_b->timer_active = 1;
3569 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3572 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3574 hrtimer_cancel(&cfs_b->period_timer);
3575 hrtimer_cancel(&cfs_b->slack_timer);
3578 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3580 struct cfs_rq *cfs_rq;
3582 for_each_leaf_cfs_rq(rq, cfs_rq) {
3583 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3585 if (!cfs_rq->runtime_enabled)
3589 * clock_task is not advancing so we just need to make sure
3590 * there's some valid quota amount
3592 cfs_rq->runtime_remaining = cfs_b->quota;
3593 if (cfs_rq_throttled(cfs_rq))
3594 unthrottle_cfs_rq(cfs_rq);
3598 #else /* CONFIG_CFS_BANDWIDTH */
3599 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3601 return rq_clock_task(rq_of(cfs_rq));
3604 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3605 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3606 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3607 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3609 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3614 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3619 static inline int throttled_lb_pair(struct task_group *tg,
3620 int src_cpu, int dest_cpu)
3625 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3627 #ifdef CONFIG_FAIR_GROUP_SCHED
3628 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3631 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3635 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3636 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3638 #endif /* CONFIG_CFS_BANDWIDTH */
3640 /**************************************************
3641 * CFS operations on tasks:
3644 #ifdef CONFIG_SCHED_HRTICK
3645 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3647 struct sched_entity *se = &p->se;
3648 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3650 WARN_ON(task_rq(p) != rq);
3652 if (cfs_rq->nr_running > 1) {
3653 u64 slice = sched_slice(cfs_rq, se);
3654 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3655 s64 delta = slice - ran;
3664 * Don't schedule slices shorter than 10000ns, that just
3665 * doesn't make sense. Rely on vruntime for fairness.
3668 delta = max_t(s64, 10000LL, delta);
3670 hrtick_start(rq, delta);
3675 * called from enqueue/dequeue and updates the hrtick when the
3676 * current task is from our class and nr_running is low enough
3679 static void hrtick_update(struct rq *rq)
3681 struct task_struct *curr = rq->curr;
3683 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3686 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3687 hrtick_start_fair(rq, curr);
3689 #else /* !CONFIG_SCHED_HRTICK */
3691 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3695 static inline void hrtick_update(struct rq *rq)
3701 * The enqueue_task method is called before nr_running is
3702 * increased. Here we update the fair scheduling stats and
3703 * then put the task into the rbtree:
3706 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3708 struct cfs_rq *cfs_rq;
3709 struct sched_entity *se = &p->se;
3711 for_each_sched_entity(se) {
3714 cfs_rq = cfs_rq_of(se);
3715 enqueue_entity(cfs_rq, se, flags);
3718 * end evaluation on encountering a throttled cfs_rq
3720 * note: in the case of encountering a throttled cfs_rq we will
3721 * post the final h_nr_running increment below.
3723 if (cfs_rq_throttled(cfs_rq))
3725 cfs_rq->h_nr_running++;
3727 flags = ENQUEUE_WAKEUP;
3730 for_each_sched_entity(se) {
3731 cfs_rq = cfs_rq_of(se);
3732 cfs_rq->h_nr_running++;
3734 if (cfs_rq_throttled(cfs_rq))
3737 update_cfs_shares(cfs_rq);
3738 update_entity_load_avg(se, 1);
3742 update_rq_runnable_avg(rq, rq->nr_running);
3748 static void set_next_buddy(struct sched_entity *se);
3751 * The dequeue_task method is called before nr_running is
3752 * decreased. We remove the task from the rbtree and
3753 * update the fair scheduling stats:
3755 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3757 struct cfs_rq *cfs_rq;
3758 struct sched_entity *se = &p->se;
3759 int task_sleep = flags & DEQUEUE_SLEEP;
3761 for_each_sched_entity(se) {
3762 cfs_rq = cfs_rq_of(se);
3763 dequeue_entity(cfs_rq, se, flags);
3766 * end evaluation on encountering a throttled cfs_rq
3768 * note: in the case of encountering a throttled cfs_rq we will
3769 * post the final h_nr_running decrement below.
3771 if (cfs_rq_throttled(cfs_rq))
3773 cfs_rq->h_nr_running--;
3775 /* Don't dequeue parent if it has other entities besides us */
3776 if (cfs_rq->load.weight) {
3778 * Bias pick_next to pick a task from this cfs_rq, as
3779 * p is sleeping when it is within its sched_slice.
3781 if (task_sleep && parent_entity(se))
3782 set_next_buddy(parent_entity(se));
3784 /* avoid re-evaluating load for this entity */
3785 se = parent_entity(se);
3788 flags |= DEQUEUE_SLEEP;
3791 for_each_sched_entity(se) {
3792 cfs_rq = cfs_rq_of(se);
3793 cfs_rq->h_nr_running--;
3795 if (cfs_rq_throttled(cfs_rq))
3798 update_cfs_shares(cfs_rq);
3799 update_entity_load_avg(se, 1);
3804 update_rq_runnable_avg(rq, 1);
3810 /* Used instead of source_load when we know the type == 0 */
3811 static unsigned long weighted_cpuload(const int cpu)
3813 return cpu_rq(cpu)->cfs.runnable_load_avg;
3817 * Return a low guess at the load of a migration-source cpu weighted
3818 * according to the scheduling class and "nice" value.
3820 * We want to under-estimate the load of migration sources, to
3821 * balance conservatively.
3823 static unsigned long source_load(int cpu, int type)
3825 struct rq *rq = cpu_rq(cpu);
3826 unsigned long total = weighted_cpuload(cpu);
3828 if (type == 0 || !sched_feat(LB_BIAS))
3831 return min(rq->cpu_load[type-1], total);
3835 * Return a high guess at the load of a migration-target cpu weighted
3836 * according to the scheduling class and "nice" value.
3838 static unsigned long target_load(int cpu, int type)
3840 struct rq *rq = cpu_rq(cpu);
3841 unsigned long total = weighted_cpuload(cpu);
3843 if (type == 0 || !sched_feat(LB_BIAS))
3846 return max(rq->cpu_load[type-1], total);
3849 static unsigned long power_of(int cpu)
3851 return cpu_rq(cpu)->cpu_power;
3854 static unsigned long cpu_avg_load_per_task(int cpu)
3856 struct rq *rq = cpu_rq(cpu);
3857 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3858 unsigned long load_avg = rq->cfs.runnable_load_avg;
3861 return load_avg / nr_running;
3866 static void record_wakee(struct task_struct *p)
3869 * Rough decay (wiping) for cost saving, don't worry
3870 * about the boundary, really active task won't care
3873 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3874 current->wakee_flips = 0;
3875 current->wakee_flip_decay_ts = jiffies;
3878 if (current->last_wakee != p) {
3879 current->last_wakee = p;
3880 current->wakee_flips++;
3884 static void task_waking_fair(struct task_struct *p)
3886 struct sched_entity *se = &p->se;
3887 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3890 #ifndef CONFIG_64BIT
3891 u64 min_vruntime_copy;
3894 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3896 min_vruntime = cfs_rq->min_vruntime;
3897 } while (min_vruntime != min_vruntime_copy);
3899 min_vruntime = cfs_rq->min_vruntime;
3902 se->vruntime -= min_vruntime;
3906 #ifdef CONFIG_FAIR_GROUP_SCHED
3908 * effective_load() calculates the load change as seen from the root_task_group
3910 * Adding load to a group doesn't make a group heavier, but can cause movement
3911 * of group shares between cpus. Assuming the shares were perfectly aligned one
3912 * can calculate the shift in shares.
3914 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3915 * on this @cpu and results in a total addition (subtraction) of @wg to the
3916 * total group weight.
3918 * Given a runqueue weight distribution (rw_i) we can compute a shares
3919 * distribution (s_i) using:
3921 * s_i = rw_i / \Sum rw_j (1)
3923 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3924 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3925 * shares distribution (s_i):
3927 * rw_i = { 2, 4, 1, 0 }
3928 * s_i = { 2/7, 4/7, 1/7, 0 }
3930 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3931 * task used to run on and the CPU the waker is running on), we need to
3932 * compute the effect of waking a task on either CPU and, in case of a sync
3933 * wakeup, compute the effect of the current task going to sleep.
3935 * So for a change of @wl to the local @cpu with an overall group weight change
3936 * of @wl we can compute the new shares distribution (s'_i) using:
3938 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3940 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3941 * differences in waking a task to CPU 0. The additional task changes the
3942 * weight and shares distributions like:
3944 * rw'_i = { 3, 4, 1, 0 }
3945 * s'_i = { 3/8, 4/8, 1/8, 0 }
3947 * We can then compute the difference in effective weight by using:
3949 * dw_i = S * (s'_i - s_i) (3)
3951 * Where 'S' is the group weight as seen by its parent.
3953 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3954 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3955 * 4/7) times the weight of the group.
3957 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3959 struct sched_entity *se = tg->se[cpu];
3961 if (!tg->parent) /* the trivial, non-cgroup case */
3964 for_each_sched_entity(se) {
3970 * W = @wg + \Sum rw_j
3972 W = wg + calc_tg_weight(tg, se->my_q);
3977 w = se->my_q->load.weight + wl;
3980 * wl = S * s'_i; see (2)
3983 wl = (w * tg->shares) / W;
3988 * Per the above, wl is the new se->load.weight value; since
3989 * those are clipped to [MIN_SHARES, ...) do so now. See
3990 * calc_cfs_shares().
3992 if (wl < MIN_SHARES)
3996 * wl = dw_i = S * (s'_i - s_i); see (3)
3998 wl -= se->load.weight;
4001 * Recursively apply this logic to all parent groups to compute
4002 * the final effective load change on the root group. Since
4003 * only the @tg group gets extra weight, all parent groups can
4004 * only redistribute existing shares. @wl is the shift in shares
4005 * resulting from this level per the above.
4014 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4021 static int wake_wide(struct task_struct *p)
4023 int factor = this_cpu_read(sd_llc_size);
4026 * Yeah, it's the switching-frequency, could means many wakee or
4027 * rapidly switch, use factor here will just help to automatically
4028 * adjust the loose-degree, so bigger node will lead to more pull.
4030 if (p->wakee_flips > factor) {
4032 * wakee is somewhat hot, it needs certain amount of cpu
4033 * resource, so if waker is far more hot, prefer to leave
4036 if (current->wakee_flips > (factor * p->wakee_flips))
4043 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4045 s64 this_load, load;
4046 int idx, this_cpu, prev_cpu;
4047 unsigned long tl_per_task;
4048 struct task_group *tg;
4049 unsigned long weight;
4053 * If we wake multiple tasks be careful to not bounce
4054 * ourselves around too much.
4060 this_cpu = smp_processor_id();
4061 prev_cpu = task_cpu(p);
4062 load = source_load(prev_cpu, idx);
4063 this_load = target_load(this_cpu, idx);
4066 * If sync wakeup then subtract the (maximum possible)
4067 * effect of the currently running task from the load
4068 * of the current CPU:
4071 tg = task_group(current);
4072 weight = current->se.load.weight;
4074 this_load += effective_load(tg, this_cpu, -weight, -weight);
4075 load += effective_load(tg, prev_cpu, 0, -weight);
4079 weight = p->se.load.weight;
4082 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4083 * due to the sync cause above having dropped this_load to 0, we'll
4084 * always have an imbalance, but there's really nothing you can do
4085 * about that, so that's good too.
4087 * Otherwise check if either cpus are near enough in load to allow this
4088 * task to be woken on this_cpu.
4090 if (this_load > 0) {
4091 s64 this_eff_load, prev_eff_load;
4093 this_eff_load = 100;
4094 this_eff_load *= power_of(prev_cpu);
4095 this_eff_load *= this_load +
4096 effective_load(tg, this_cpu, weight, weight);
4098 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4099 prev_eff_load *= power_of(this_cpu);
4100 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4102 balanced = this_eff_load <= prev_eff_load;
4107 * If the currently running task will sleep within
4108 * a reasonable amount of time then attract this newly
4111 if (sync && balanced)
4114 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4115 tl_per_task = cpu_avg_load_per_task(this_cpu);
4118 (this_load <= load &&
4119 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4121 * This domain has SD_WAKE_AFFINE and
4122 * p is cache cold in this domain, and
4123 * there is no bad imbalance.
4125 schedstat_inc(sd, ttwu_move_affine);
4126 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4134 * find_idlest_group finds and returns the least busy CPU group within the
4137 static struct sched_group *
4138 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4139 int this_cpu, int sd_flag)
4141 struct sched_group *idlest = NULL, *group = sd->groups;
4142 unsigned long min_load = ULONG_MAX, this_load = 0;
4143 int load_idx = sd->forkexec_idx;
4144 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4146 if (sd_flag & SD_BALANCE_WAKE)
4147 load_idx = sd->wake_idx;
4150 unsigned long load, avg_load;
4154 /* Skip over this group if it has no CPUs allowed */
4155 if (!cpumask_intersects(sched_group_cpus(group),
4156 tsk_cpus_allowed(p)))
4159 local_group = cpumask_test_cpu(this_cpu,
4160 sched_group_cpus(group));
4162 /* Tally up the load of all CPUs in the group */
4165 for_each_cpu(i, sched_group_cpus(group)) {
4166 /* Bias balancing toward cpus of our domain */
4168 load = source_load(i, load_idx);
4170 load = target_load(i, load_idx);
4175 /* Adjust by relative CPU power of the group */
4176 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4179 this_load = avg_load;
4180 } else if (avg_load < min_load) {
4181 min_load = avg_load;
4184 } while (group = group->next, group != sd->groups);
4186 if (!idlest || 100*this_load < imbalance*min_load)
4192 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4195 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4197 unsigned long load, min_load = ULONG_MAX;
4201 /* Traverse only the allowed CPUs */
4202 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4203 load = weighted_cpuload(i);
4205 if (load < min_load || (load == min_load && i == this_cpu)) {
4215 * Try and locate an idle CPU in the sched_domain.
4217 static int select_idle_sibling(struct task_struct *p, int target)
4219 struct sched_domain *sd;
4220 struct sched_group *sg;
4221 int i = task_cpu(p);
4223 if (idle_cpu(target))
4227 * If the prevous cpu is cache affine and idle, don't be stupid.
4229 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4233 * Otherwise, iterate the domains and find an elegible idle cpu.
4235 sd = rcu_dereference(per_cpu(sd_llc, target));
4236 for_each_lower_domain(sd) {
4239 if (!cpumask_intersects(sched_group_cpus(sg),
4240 tsk_cpus_allowed(p)))
4243 for_each_cpu(i, sched_group_cpus(sg)) {
4244 if (i == target || !idle_cpu(i))
4248 target = cpumask_first_and(sched_group_cpus(sg),
4249 tsk_cpus_allowed(p));
4253 } while (sg != sd->groups);
4260 * sched_balance_self: balance the current task (running on cpu) in domains
4261 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4264 * Balance, ie. select the least loaded group.
4266 * Returns the target CPU number, or the same CPU if no balancing is needed.
4268 * preempt must be disabled.
4271 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4273 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4274 int cpu = smp_processor_id();
4276 int want_affine = 0;
4277 int sync = wake_flags & WF_SYNC;
4279 if (p->nr_cpus_allowed == 1)
4282 if (sd_flag & SD_BALANCE_WAKE) {
4283 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4289 for_each_domain(cpu, tmp) {
4290 if (!(tmp->flags & SD_LOAD_BALANCE))
4294 * If both cpu and prev_cpu are part of this domain,
4295 * cpu is a valid SD_WAKE_AFFINE target.
4297 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4298 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4303 if (tmp->flags & sd_flag)
4308 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4311 new_cpu = select_idle_sibling(p, prev_cpu);
4316 struct sched_group *group;
4319 if (!(sd->flags & sd_flag)) {
4324 group = find_idlest_group(sd, p, cpu, sd_flag);
4330 new_cpu = find_idlest_cpu(group, p, cpu);
4331 if (new_cpu == -1 || new_cpu == cpu) {
4332 /* Now try balancing at a lower domain level of cpu */
4337 /* Now try balancing at a lower domain level of new_cpu */
4339 weight = sd->span_weight;
4341 for_each_domain(cpu, tmp) {
4342 if (weight <= tmp->span_weight)
4344 if (tmp->flags & sd_flag)
4347 /* while loop will break here if sd == NULL */
4356 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4357 * cfs_rq_of(p) references at time of call are still valid and identify the
4358 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4359 * other assumptions, including the state of rq->lock, should be made.
4362 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4364 struct sched_entity *se = &p->se;
4365 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4368 * Load tracking: accumulate removed load so that it can be processed
4369 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4370 * to blocked load iff they have a positive decay-count. It can never
4371 * be negative here since on-rq tasks have decay-count == 0.
4373 if (se->avg.decay_count) {
4374 se->avg.decay_count = -__synchronize_entity_decay(se);
4375 atomic_long_add(se->avg.load_avg_contrib,
4376 &cfs_rq->removed_load);
4379 #endif /* CONFIG_SMP */
4381 static unsigned long
4382 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4384 unsigned long gran = sysctl_sched_wakeup_granularity;
4387 * Since its curr running now, convert the gran from real-time
4388 * to virtual-time in his units.
4390 * By using 'se' instead of 'curr' we penalize light tasks, so
4391 * they get preempted easier. That is, if 'se' < 'curr' then
4392 * the resulting gran will be larger, therefore penalizing the
4393 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4394 * be smaller, again penalizing the lighter task.
4396 * This is especially important for buddies when the leftmost
4397 * task is higher priority than the buddy.
4399 return calc_delta_fair(gran, se);
4403 * Should 'se' preempt 'curr'.
4417 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4419 s64 gran, vdiff = curr->vruntime - se->vruntime;
4424 gran = wakeup_gran(curr, se);
4431 static void set_last_buddy(struct sched_entity *se)
4433 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4436 for_each_sched_entity(se)
4437 cfs_rq_of(se)->last = se;
4440 static void set_next_buddy(struct sched_entity *se)
4442 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4445 for_each_sched_entity(se)
4446 cfs_rq_of(se)->next = se;
4449 static void set_skip_buddy(struct sched_entity *se)
4451 for_each_sched_entity(se)
4452 cfs_rq_of(se)->skip = se;
4456 * Preempt the current task with a newly woken task if needed:
4458 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4460 struct task_struct *curr = rq->curr;
4461 struct sched_entity *se = &curr->se, *pse = &p->se;
4462 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4463 int scale = cfs_rq->nr_running >= sched_nr_latency;
4464 int next_buddy_marked = 0;
4466 if (unlikely(se == pse))
4470 * This is possible from callers such as move_task(), in which we
4471 * unconditionally check_prempt_curr() after an enqueue (which may have
4472 * lead to a throttle). This both saves work and prevents false
4473 * next-buddy nomination below.
4475 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4478 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4479 set_next_buddy(pse);
4480 next_buddy_marked = 1;
4484 * We can come here with TIF_NEED_RESCHED already set from new task
4487 * Note: this also catches the edge-case of curr being in a throttled
4488 * group (e.g. via set_curr_task), since update_curr() (in the
4489 * enqueue of curr) will have resulted in resched being set. This
4490 * prevents us from potentially nominating it as a false LAST_BUDDY
4493 if (test_tsk_need_resched(curr))
4496 /* Idle tasks are by definition preempted by non-idle tasks. */
4497 if (unlikely(curr->policy == SCHED_IDLE) &&
4498 likely(p->policy != SCHED_IDLE))
4502 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4503 * is driven by the tick):
4505 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4508 find_matching_se(&se, &pse);
4509 update_curr(cfs_rq_of(se));
4511 if (wakeup_preempt_entity(se, pse) == 1) {
4513 * Bias pick_next to pick the sched entity that is
4514 * triggering this preemption.
4516 if (!next_buddy_marked)
4517 set_next_buddy(pse);
4526 * Only set the backward buddy when the current task is still
4527 * on the rq. This can happen when a wakeup gets interleaved
4528 * with schedule on the ->pre_schedule() or idle_balance()
4529 * point, either of which can * drop the rq lock.
4531 * Also, during early boot the idle thread is in the fair class,
4532 * for obvious reasons its a bad idea to schedule back to it.
4534 if (unlikely(!se->on_rq || curr == rq->idle))
4537 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4541 static struct task_struct *pick_next_task_fair(struct rq *rq)
4543 struct task_struct *p;
4544 struct cfs_rq *cfs_rq = &rq->cfs;
4545 struct sched_entity *se;
4547 if (!cfs_rq->nr_running)
4551 se = pick_next_entity(cfs_rq);
4552 set_next_entity(cfs_rq, se);
4553 cfs_rq = group_cfs_rq(se);
4557 if (hrtick_enabled(rq))
4558 hrtick_start_fair(rq, p);
4564 * Account for a descheduled task:
4566 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4568 struct sched_entity *se = &prev->se;
4569 struct cfs_rq *cfs_rq;
4571 for_each_sched_entity(se) {
4572 cfs_rq = cfs_rq_of(se);
4573 put_prev_entity(cfs_rq, se);
4578 * sched_yield() is very simple
4580 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4582 static void yield_task_fair(struct rq *rq)
4584 struct task_struct *curr = rq->curr;
4585 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4586 struct sched_entity *se = &curr->se;
4589 * Are we the only task in the tree?
4591 if (unlikely(rq->nr_running == 1))
4594 clear_buddies(cfs_rq, se);
4596 if (curr->policy != SCHED_BATCH) {
4597 update_rq_clock(rq);
4599 * Update run-time statistics of the 'current'.
4601 update_curr(cfs_rq);
4603 * Tell update_rq_clock() that we've just updated,
4604 * so we don't do microscopic update in schedule()
4605 * and double the fastpath cost.
4607 rq->skip_clock_update = 1;
4613 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4615 struct sched_entity *se = &p->se;
4617 /* throttled hierarchies are not runnable */
4618 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4621 /* Tell the scheduler that we'd really like pse to run next. */
4624 yield_task_fair(rq);
4630 /**************************************************
4631 * Fair scheduling class load-balancing methods.
4635 * The purpose of load-balancing is to achieve the same basic fairness the
4636 * per-cpu scheduler provides, namely provide a proportional amount of compute
4637 * time to each task. This is expressed in the following equation:
4639 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4641 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4642 * W_i,0 is defined as:
4644 * W_i,0 = \Sum_j w_i,j (2)
4646 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4647 * is derived from the nice value as per prio_to_weight[].
4649 * The weight average is an exponential decay average of the instantaneous
4652 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4654 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4655 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4656 * can also include other factors [XXX].
4658 * To achieve this balance we define a measure of imbalance which follows
4659 * directly from (1):
4661 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4663 * We them move tasks around to minimize the imbalance. In the continuous
4664 * function space it is obvious this converges, in the discrete case we get
4665 * a few fun cases generally called infeasible weight scenarios.
4668 * - infeasible weights;
4669 * - local vs global optima in the discrete case. ]
4674 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4675 * for all i,j solution, we create a tree of cpus that follows the hardware
4676 * topology where each level pairs two lower groups (or better). This results
4677 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4678 * tree to only the first of the previous level and we decrease the frequency
4679 * of load-balance at each level inv. proportional to the number of cpus in
4685 * \Sum { --- * --- * 2^i } = O(n) (5)
4687 * `- size of each group
4688 * | | `- number of cpus doing load-balance
4690 * `- sum over all levels
4692 * Coupled with a limit on how many tasks we can migrate every balance pass,
4693 * this makes (5) the runtime complexity of the balancer.
4695 * An important property here is that each CPU is still (indirectly) connected
4696 * to every other cpu in at most O(log n) steps:
4698 * The adjacency matrix of the resulting graph is given by:
4701 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4704 * And you'll find that:
4706 * A^(log_2 n)_i,j != 0 for all i,j (7)
4708 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4709 * The task movement gives a factor of O(m), giving a convergence complexity
4712 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4717 * In order to avoid CPUs going idle while there's still work to do, new idle
4718 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4719 * tree itself instead of relying on other CPUs to bring it work.
4721 * This adds some complexity to both (5) and (8) but it reduces the total idle
4729 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4732 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4737 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4739 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4741 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4744 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4745 * rewrite all of this once again.]
4748 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4750 enum fbq_type { regular, remote, all };
4752 #define LBF_ALL_PINNED 0x01
4753 #define LBF_NEED_BREAK 0x02
4754 #define LBF_DST_PINNED 0x04
4755 #define LBF_SOME_PINNED 0x08
4758 struct sched_domain *sd;
4766 struct cpumask *dst_grpmask;
4768 enum cpu_idle_type idle;
4770 /* The set of CPUs under consideration for load-balancing */
4771 struct cpumask *cpus;
4776 unsigned int loop_break;
4777 unsigned int loop_max;
4779 enum fbq_type fbq_type;
4783 * move_task - move a task from one runqueue to another runqueue.
4784 * Both runqueues must be locked.
4786 static void move_task(struct task_struct *p, struct lb_env *env)
4788 deactivate_task(env->src_rq, p, 0);
4789 set_task_cpu(p, env->dst_cpu);
4790 activate_task(env->dst_rq, p, 0);
4791 check_preempt_curr(env->dst_rq, p, 0);
4795 * Is this task likely cache-hot:
4798 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4802 if (p->sched_class != &fair_sched_class)
4805 if (unlikely(p->policy == SCHED_IDLE))
4809 * Buddy candidates are cache hot:
4811 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4812 (&p->se == cfs_rq_of(&p->se)->next ||
4813 &p->se == cfs_rq_of(&p->se)->last))
4816 if (sysctl_sched_migration_cost == -1)
4818 if (sysctl_sched_migration_cost == 0)
4821 delta = now - p->se.exec_start;
4823 return delta < (s64)sysctl_sched_migration_cost;
4826 #ifdef CONFIG_NUMA_BALANCING
4827 /* Returns true if the destination node has incurred more faults */
4828 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4830 int src_nid, dst_nid;
4832 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
4833 !(env->sd->flags & SD_NUMA)) {
4837 src_nid = cpu_to_node(env->src_cpu);
4838 dst_nid = cpu_to_node(env->dst_cpu);
4840 if (src_nid == dst_nid)
4843 /* Always encourage migration to the preferred node. */
4844 if (dst_nid == p->numa_preferred_nid)
4847 /* If both task and group weight improve, this move is a winner. */
4848 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4849 group_weight(p, dst_nid) > group_weight(p, src_nid))
4856 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4858 int src_nid, dst_nid;
4860 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4863 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
4866 src_nid = cpu_to_node(env->src_cpu);
4867 dst_nid = cpu_to_node(env->dst_cpu);
4869 if (src_nid == dst_nid)
4872 /* Migrating away from the preferred node is always bad. */
4873 if (src_nid == p->numa_preferred_nid)
4876 /* If either task or group weight get worse, don't do it. */
4877 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
4878 group_weight(p, dst_nid) < group_weight(p, src_nid))
4885 static inline bool migrate_improves_locality(struct task_struct *p,
4891 static inline bool migrate_degrades_locality(struct task_struct *p,
4899 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4902 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4904 int tsk_cache_hot = 0;
4906 * We do not migrate tasks that are:
4907 * 1) throttled_lb_pair, or
4908 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4909 * 3) running (obviously), or
4910 * 4) are cache-hot on their current CPU.
4912 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4915 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4918 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4920 env->flags |= LBF_SOME_PINNED;
4923 * Remember if this task can be migrated to any other cpu in
4924 * our sched_group. We may want to revisit it if we couldn't
4925 * meet load balance goals by pulling other tasks on src_cpu.
4927 * Also avoid computing new_dst_cpu if we have already computed
4928 * one in current iteration.
4930 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4933 /* Prevent to re-select dst_cpu via env's cpus */
4934 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4935 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4936 env->flags |= LBF_DST_PINNED;
4937 env->new_dst_cpu = cpu;
4945 /* Record that we found atleast one task that could run on dst_cpu */
4946 env->flags &= ~LBF_ALL_PINNED;
4948 if (task_running(env->src_rq, p)) {
4949 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4954 * Aggressive migration if:
4955 * 1) destination numa is preferred
4956 * 2) task is cache cold, or
4957 * 3) too many balance attempts have failed.
4959 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4961 tsk_cache_hot = migrate_degrades_locality(p, env);
4963 if (migrate_improves_locality(p, env)) {
4964 #ifdef CONFIG_SCHEDSTATS
4965 if (tsk_cache_hot) {
4966 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4967 schedstat_inc(p, se.statistics.nr_forced_migrations);
4973 if (!tsk_cache_hot ||
4974 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4976 if (tsk_cache_hot) {
4977 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4978 schedstat_inc(p, se.statistics.nr_forced_migrations);
4984 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4989 * move_one_task tries to move exactly one task from busiest to this_rq, as
4990 * part of active balancing operations within "domain".
4991 * Returns 1 if successful and 0 otherwise.
4993 * Called with both runqueues locked.
4995 static int move_one_task(struct lb_env *env)
4997 struct task_struct *p, *n;
4999 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5000 if (!can_migrate_task(p, env))
5005 * Right now, this is only the second place move_task()
5006 * is called, so we can safely collect move_task()
5007 * stats here rather than inside move_task().
5009 schedstat_inc(env->sd, lb_gained[env->idle]);
5015 static const unsigned int sched_nr_migrate_break = 32;
5018 * move_tasks tries to move up to imbalance weighted load from busiest to
5019 * this_rq, as part of a balancing operation within domain "sd".
5020 * Returns 1 if successful and 0 otherwise.
5022 * Called with both runqueues locked.
5024 static int move_tasks(struct lb_env *env)
5026 struct list_head *tasks = &env->src_rq->cfs_tasks;
5027 struct task_struct *p;
5031 if (env->imbalance <= 0)
5034 while (!list_empty(tasks)) {
5035 p = list_first_entry(tasks, struct task_struct, se.group_node);
5038 /* We've more or less seen every task there is, call it quits */
5039 if (env->loop > env->loop_max)
5042 /* take a breather every nr_migrate tasks */
5043 if (env->loop > env->loop_break) {
5044 env->loop_break += sched_nr_migrate_break;
5045 env->flags |= LBF_NEED_BREAK;
5049 if (!can_migrate_task(p, env))
5052 load = task_h_load(p);
5054 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5057 if ((load / 2) > env->imbalance)
5062 env->imbalance -= load;
5064 #ifdef CONFIG_PREEMPT
5066 * NEWIDLE balancing is a source of latency, so preemptible
5067 * kernels will stop after the first task is pulled to minimize
5068 * the critical section.
5070 if (env->idle == CPU_NEWLY_IDLE)
5075 * We only want to steal up to the prescribed amount of
5078 if (env->imbalance <= 0)
5083 list_move_tail(&p->se.group_node, tasks);
5087 * Right now, this is one of only two places move_task() is called,
5088 * so we can safely collect move_task() stats here rather than
5089 * inside move_task().
5091 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5096 #ifdef CONFIG_FAIR_GROUP_SCHED
5098 * update tg->load_weight by folding this cpu's load_avg
5100 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5102 struct sched_entity *se = tg->se[cpu];
5103 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5105 /* throttled entities do not contribute to load */
5106 if (throttled_hierarchy(cfs_rq))
5109 update_cfs_rq_blocked_load(cfs_rq, 1);
5112 update_entity_load_avg(se, 1);
5114 * We pivot on our runnable average having decayed to zero for
5115 * list removal. This generally implies that all our children
5116 * have also been removed (modulo rounding error or bandwidth
5117 * control); however, such cases are rare and we can fix these
5120 * TODO: fix up out-of-order children on enqueue.
5122 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5123 list_del_leaf_cfs_rq(cfs_rq);
5125 struct rq *rq = rq_of(cfs_rq);
5126 update_rq_runnable_avg(rq, rq->nr_running);
5130 static void update_blocked_averages(int cpu)
5132 struct rq *rq = cpu_rq(cpu);
5133 struct cfs_rq *cfs_rq;
5134 unsigned long flags;
5136 raw_spin_lock_irqsave(&rq->lock, flags);
5137 update_rq_clock(rq);
5139 * Iterates the task_group tree in a bottom up fashion, see
5140 * list_add_leaf_cfs_rq() for details.
5142 for_each_leaf_cfs_rq(rq, cfs_rq) {
5144 * Note: We may want to consider periodically releasing
5145 * rq->lock about these updates so that creating many task
5146 * groups does not result in continually extending hold time.
5148 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5151 raw_spin_unlock_irqrestore(&rq->lock, flags);
5155 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5156 * This needs to be done in a top-down fashion because the load of a child
5157 * group is a fraction of its parents load.
5159 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5161 struct rq *rq = rq_of(cfs_rq);
5162 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5163 unsigned long now = jiffies;
5166 if (cfs_rq->last_h_load_update == now)
5169 cfs_rq->h_load_next = NULL;
5170 for_each_sched_entity(se) {
5171 cfs_rq = cfs_rq_of(se);
5172 cfs_rq->h_load_next = se;
5173 if (cfs_rq->last_h_load_update == now)
5178 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5179 cfs_rq->last_h_load_update = now;
5182 while ((se = cfs_rq->h_load_next) != NULL) {
5183 load = cfs_rq->h_load;
5184 load = div64_ul(load * se->avg.load_avg_contrib,
5185 cfs_rq->runnable_load_avg + 1);
5186 cfs_rq = group_cfs_rq(se);
5187 cfs_rq->h_load = load;
5188 cfs_rq->last_h_load_update = now;
5192 static unsigned long task_h_load(struct task_struct *p)
5194 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5196 update_cfs_rq_h_load(cfs_rq);
5197 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5198 cfs_rq->runnable_load_avg + 1);
5201 static inline void update_blocked_averages(int cpu)
5205 static unsigned long task_h_load(struct task_struct *p)
5207 return p->se.avg.load_avg_contrib;
5211 /********** Helpers for find_busiest_group ************************/
5213 * sg_lb_stats - stats of a sched_group required for load_balancing
5215 struct sg_lb_stats {
5216 unsigned long avg_load; /*Avg load across the CPUs of the group */
5217 unsigned long group_load; /* Total load over the CPUs of the group */
5218 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5219 unsigned long load_per_task;
5220 unsigned long group_power;
5221 unsigned int sum_nr_running; /* Nr tasks running in the group */
5222 unsigned int group_capacity;
5223 unsigned int idle_cpus;
5224 unsigned int group_weight;
5225 int group_imb; /* Is there an imbalance in the group ? */
5226 int group_has_capacity; /* Is there extra capacity in the group? */
5227 #ifdef CONFIG_NUMA_BALANCING
5228 unsigned int nr_numa_running;
5229 unsigned int nr_preferred_running;
5234 * sd_lb_stats - Structure to store the statistics of a sched_domain
5235 * during load balancing.
5237 struct sd_lb_stats {
5238 struct sched_group *busiest; /* Busiest group in this sd */
5239 struct sched_group *local; /* Local group in this sd */
5240 unsigned long total_load; /* Total load of all groups in sd */
5241 unsigned long total_pwr; /* Total power of all groups in sd */
5242 unsigned long avg_load; /* Average load across all groups in sd */
5244 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5245 struct sg_lb_stats local_stat; /* Statistics of the local group */
5248 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5251 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5252 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5253 * We must however clear busiest_stat::avg_load because
5254 * update_sd_pick_busiest() reads this before assignment.
5256 *sds = (struct sd_lb_stats){
5268 * get_sd_load_idx - Obtain the load index for a given sched domain.
5269 * @sd: The sched_domain whose load_idx is to be obtained.
5270 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5272 * Return: The load index.
5274 static inline int get_sd_load_idx(struct sched_domain *sd,
5275 enum cpu_idle_type idle)
5281 load_idx = sd->busy_idx;
5284 case CPU_NEWLY_IDLE:
5285 load_idx = sd->newidle_idx;
5288 load_idx = sd->idle_idx;
5295 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5297 return SCHED_POWER_SCALE;
5300 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5302 return default_scale_freq_power(sd, cpu);
5305 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5307 unsigned long weight = sd->span_weight;
5308 unsigned long smt_gain = sd->smt_gain;
5315 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5317 return default_scale_smt_power(sd, cpu);
5320 static unsigned long scale_rt_power(int cpu)
5322 struct rq *rq = cpu_rq(cpu);
5323 u64 total, available, age_stamp, avg;
5326 * Since we're reading these variables without serialization make sure
5327 * we read them once before doing sanity checks on them.
5329 age_stamp = ACCESS_ONCE(rq->age_stamp);
5330 avg = ACCESS_ONCE(rq->rt_avg);
5332 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5334 if (unlikely(total < avg)) {
5335 /* Ensures that power won't end up being negative */
5338 available = total - avg;
5341 if (unlikely((s64)total < SCHED_POWER_SCALE))
5342 total = SCHED_POWER_SCALE;
5344 total >>= SCHED_POWER_SHIFT;
5346 return div_u64(available, total);
5349 static void update_cpu_power(struct sched_domain *sd, int cpu)
5351 unsigned long weight = sd->span_weight;
5352 unsigned long power = SCHED_POWER_SCALE;
5353 struct sched_group *sdg = sd->groups;
5355 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5356 if (sched_feat(ARCH_POWER))
5357 power *= arch_scale_smt_power(sd, cpu);
5359 power *= default_scale_smt_power(sd, cpu);
5361 power >>= SCHED_POWER_SHIFT;
5364 sdg->sgp->power_orig = power;
5366 if (sched_feat(ARCH_POWER))
5367 power *= arch_scale_freq_power(sd, cpu);
5369 power *= default_scale_freq_power(sd, cpu);
5371 power >>= SCHED_POWER_SHIFT;
5373 power *= scale_rt_power(cpu);
5374 power >>= SCHED_POWER_SHIFT;
5379 cpu_rq(cpu)->cpu_power = power;
5380 sdg->sgp->power = power;
5383 void update_group_power(struct sched_domain *sd, int cpu)
5385 struct sched_domain *child = sd->child;
5386 struct sched_group *group, *sdg = sd->groups;
5387 unsigned long power, power_orig;
5388 unsigned long interval;
5390 interval = msecs_to_jiffies(sd->balance_interval);
5391 interval = clamp(interval, 1UL, max_load_balance_interval);
5392 sdg->sgp->next_update = jiffies + interval;
5395 update_cpu_power(sd, cpu);
5399 power_orig = power = 0;
5401 if (child->flags & SD_OVERLAP) {
5403 * SD_OVERLAP domains cannot assume that child groups
5404 * span the current group.
5407 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5408 struct sched_group_power *sgp;
5409 struct rq *rq = cpu_rq(cpu);
5412 * build_sched_domains() -> init_sched_groups_power()
5413 * gets here before we've attached the domains to the
5416 * Use power_of(), which is set irrespective of domains
5417 * in update_cpu_power().
5419 * This avoids power/power_orig from being 0 and
5420 * causing divide-by-zero issues on boot.
5422 * Runtime updates will correct power_orig.
5424 if (unlikely(!rq->sd)) {
5425 power_orig += power_of(cpu);
5426 power += power_of(cpu);
5430 sgp = rq->sd->groups->sgp;
5431 power_orig += sgp->power_orig;
5432 power += sgp->power;
5436 * !SD_OVERLAP domains can assume that child groups
5437 * span the current group.
5440 group = child->groups;
5442 power_orig += group->sgp->power_orig;
5443 power += group->sgp->power;
5444 group = group->next;
5445 } while (group != child->groups);
5448 sdg->sgp->power_orig = power_orig;
5449 sdg->sgp->power = power;
5453 * Try and fix up capacity for tiny siblings, this is needed when
5454 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5455 * which on its own isn't powerful enough.
5457 * See update_sd_pick_busiest() and check_asym_packing().
5460 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5463 * Only siblings can have significantly less than SCHED_POWER_SCALE
5465 if (!(sd->flags & SD_SHARE_CPUPOWER))
5469 * If ~90% of the cpu_power is still there, we're good.
5471 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5478 * Group imbalance indicates (and tries to solve) the problem where balancing
5479 * groups is inadequate due to tsk_cpus_allowed() constraints.
5481 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5482 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5485 * { 0 1 2 3 } { 4 5 6 7 }
5488 * If we were to balance group-wise we'd place two tasks in the first group and
5489 * two tasks in the second group. Clearly this is undesired as it will overload
5490 * cpu 3 and leave one of the cpus in the second group unused.
5492 * The current solution to this issue is detecting the skew in the first group
5493 * by noticing the lower domain failed to reach balance and had difficulty
5494 * moving tasks due to affinity constraints.
5496 * When this is so detected; this group becomes a candidate for busiest; see
5497 * update_sd_pick_busiest(). And calculate_imbalance() and
5498 * find_busiest_group() avoid some of the usual balance conditions to allow it
5499 * to create an effective group imbalance.
5501 * This is a somewhat tricky proposition since the next run might not find the
5502 * group imbalance and decide the groups need to be balanced again. A most
5503 * subtle and fragile situation.
5506 static inline int sg_imbalanced(struct sched_group *group)
5508 return group->sgp->imbalance;
5512 * Compute the group capacity.
5514 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5515 * first dividing out the smt factor and computing the actual number of cores
5516 * and limit power unit capacity with that.
5518 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5520 unsigned int capacity, smt, cpus;
5521 unsigned int power, power_orig;
5523 power = group->sgp->power;
5524 power_orig = group->sgp->power_orig;
5525 cpus = group->group_weight;
5527 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5528 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5529 capacity = cpus / smt; /* cores */
5531 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5533 capacity = fix_small_capacity(env->sd, group);
5539 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5540 * @env: The load balancing environment.
5541 * @group: sched_group whose statistics are to be updated.
5542 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5543 * @local_group: Does group contain this_cpu.
5544 * @sgs: variable to hold the statistics for this group.
5546 static inline void update_sg_lb_stats(struct lb_env *env,
5547 struct sched_group *group, int load_idx,
5548 int local_group, struct sg_lb_stats *sgs)
5553 memset(sgs, 0, sizeof(*sgs));
5555 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5556 struct rq *rq = cpu_rq(i);
5558 /* Bias balancing toward cpus of our domain */
5560 load = target_load(i, load_idx);
5562 load = source_load(i, load_idx);
5564 sgs->group_load += load;
5565 sgs->sum_nr_running += rq->nr_running;
5566 #ifdef CONFIG_NUMA_BALANCING
5567 sgs->nr_numa_running += rq->nr_numa_running;
5568 sgs->nr_preferred_running += rq->nr_preferred_running;
5570 sgs->sum_weighted_load += weighted_cpuload(i);
5575 /* Adjust by relative CPU power of the group */
5576 sgs->group_power = group->sgp->power;
5577 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5579 if (sgs->sum_nr_running)
5580 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5582 sgs->group_weight = group->group_weight;
5584 sgs->group_imb = sg_imbalanced(group);
5585 sgs->group_capacity = sg_capacity(env, group);
5587 if (sgs->group_capacity > sgs->sum_nr_running)
5588 sgs->group_has_capacity = 1;
5592 * update_sd_pick_busiest - return 1 on busiest group
5593 * @env: The load balancing environment.
5594 * @sds: sched_domain statistics
5595 * @sg: sched_group candidate to be checked for being the busiest
5596 * @sgs: sched_group statistics
5598 * Determine if @sg is a busier group than the previously selected
5601 * Return: %true if @sg is a busier group than the previously selected
5602 * busiest group. %false otherwise.
5604 static bool update_sd_pick_busiest(struct lb_env *env,
5605 struct sd_lb_stats *sds,
5606 struct sched_group *sg,
5607 struct sg_lb_stats *sgs)
5609 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5612 if (sgs->sum_nr_running > sgs->group_capacity)
5619 * ASYM_PACKING needs to move all the work to the lowest
5620 * numbered CPUs in the group, therefore mark all groups
5621 * higher than ourself as busy.
5623 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5624 env->dst_cpu < group_first_cpu(sg)) {
5628 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5635 #ifdef CONFIG_NUMA_BALANCING
5636 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5638 if (sgs->sum_nr_running > sgs->nr_numa_running)
5640 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5645 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5647 if (rq->nr_running > rq->nr_numa_running)
5649 if (rq->nr_running > rq->nr_preferred_running)
5654 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5659 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5663 #endif /* CONFIG_NUMA_BALANCING */
5666 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5667 * @env: The load balancing environment.
5668 * @sds: variable to hold the statistics for this sched_domain.
5670 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5672 struct sched_domain *child = env->sd->child;
5673 struct sched_group *sg = env->sd->groups;
5674 struct sg_lb_stats tmp_sgs;
5675 int load_idx, prefer_sibling = 0;
5677 if (child && child->flags & SD_PREFER_SIBLING)
5680 load_idx = get_sd_load_idx(env->sd, env->idle);
5683 struct sg_lb_stats *sgs = &tmp_sgs;
5686 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5689 sgs = &sds->local_stat;
5691 if (env->idle != CPU_NEWLY_IDLE ||
5692 time_after_eq(jiffies, sg->sgp->next_update))
5693 update_group_power(env->sd, env->dst_cpu);
5696 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5702 * In case the child domain prefers tasks go to siblings
5703 * first, lower the sg capacity to one so that we'll try
5704 * and move all the excess tasks away. We lower the capacity
5705 * of a group only if the local group has the capacity to fit
5706 * these excess tasks, i.e. nr_running < group_capacity. The
5707 * extra check prevents the case where you always pull from the
5708 * heaviest group when it is already under-utilized (possible
5709 * with a large weight task outweighs the tasks on the system).
5711 if (prefer_sibling && sds->local &&
5712 sds->local_stat.group_has_capacity)
5713 sgs->group_capacity = min(sgs->group_capacity, 1U);
5715 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5717 sds->busiest_stat = *sgs;
5721 /* Now, start updating sd_lb_stats */
5722 sds->total_load += sgs->group_load;
5723 sds->total_pwr += sgs->group_power;
5726 } while (sg != env->sd->groups);
5728 if (env->sd->flags & SD_NUMA)
5729 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5733 * check_asym_packing - Check to see if the group is packed into the
5736 * This is primarily intended to used at the sibling level. Some
5737 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5738 * case of POWER7, it can move to lower SMT modes only when higher
5739 * threads are idle. When in lower SMT modes, the threads will
5740 * perform better since they share less core resources. Hence when we
5741 * have idle threads, we want them to be the higher ones.
5743 * This packing function is run on idle threads. It checks to see if
5744 * the busiest CPU in this domain (core in the P7 case) has a higher
5745 * CPU number than the packing function is being run on. Here we are
5746 * assuming lower CPU number will be equivalent to lower a SMT thread
5749 * Return: 1 when packing is required and a task should be moved to
5750 * this CPU. The amount of the imbalance is returned in *imbalance.
5752 * @env: The load balancing environment.
5753 * @sds: Statistics of the sched_domain which is to be packed
5755 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5759 if (!(env->sd->flags & SD_ASYM_PACKING))
5765 busiest_cpu = group_first_cpu(sds->busiest);
5766 if (env->dst_cpu > busiest_cpu)
5769 env->imbalance = DIV_ROUND_CLOSEST(
5770 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5777 * fix_small_imbalance - Calculate the minor imbalance that exists
5778 * amongst the groups of a sched_domain, during
5780 * @env: The load balancing environment.
5781 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5784 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5786 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5787 unsigned int imbn = 2;
5788 unsigned long scaled_busy_load_per_task;
5789 struct sg_lb_stats *local, *busiest;
5791 local = &sds->local_stat;
5792 busiest = &sds->busiest_stat;
5794 if (!local->sum_nr_running)
5795 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5796 else if (busiest->load_per_task > local->load_per_task)
5799 scaled_busy_load_per_task =
5800 (busiest->load_per_task * SCHED_POWER_SCALE) /
5801 busiest->group_power;
5803 if (busiest->avg_load + scaled_busy_load_per_task >=
5804 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5805 env->imbalance = busiest->load_per_task;
5810 * OK, we don't have enough imbalance to justify moving tasks,
5811 * however we may be able to increase total CPU power used by
5815 pwr_now += busiest->group_power *
5816 min(busiest->load_per_task, busiest->avg_load);
5817 pwr_now += local->group_power *
5818 min(local->load_per_task, local->avg_load);
5819 pwr_now /= SCHED_POWER_SCALE;
5821 /* Amount of load we'd subtract */
5822 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5823 busiest->group_power;
5824 if (busiest->avg_load > tmp) {
5825 pwr_move += busiest->group_power *
5826 min(busiest->load_per_task,
5827 busiest->avg_load - tmp);
5830 /* Amount of load we'd add */
5831 if (busiest->avg_load * busiest->group_power <
5832 busiest->load_per_task * SCHED_POWER_SCALE) {
5833 tmp = (busiest->avg_load * busiest->group_power) /
5836 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5839 pwr_move += local->group_power *
5840 min(local->load_per_task, local->avg_load + tmp);
5841 pwr_move /= SCHED_POWER_SCALE;
5843 /* Move if we gain throughput */
5844 if (pwr_move > pwr_now)
5845 env->imbalance = busiest->load_per_task;
5849 * calculate_imbalance - Calculate the amount of imbalance present within the
5850 * groups of a given sched_domain during load balance.
5851 * @env: load balance environment
5852 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5854 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5856 unsigned long max_pull, load_above_capacity = ~0UL;
5857 struct sg_lb_stats *local, *busiest;
5859 local = &sds->local_stat;
5860 busiest = &sds->busiest_stat;
5862 if (busiest->group_imb) {
5864 * In the group_imb case we cannot rely on group-wide averages
5865 * to ensure cpu-load equilibrium, look at wider averages. XXX
5867 busiest->load_per_task =
5868 min(busiest->load_per_task, sds->avg_load);
5872 * In the presence of smp nice balancing, certain scenarios can have
5873 * max load less than avg load(as we skip the groups at or below
5874 * its cpu_power, while calculating max_load..)
5876 if (busiest->avg_load <= sds->avg_load ||
5877 local->avg_load >= sds->avg_load) {
5879 return fix_small_imbalance(env, sds);
5882 if (!busiest->group_imb) {
5884 * Don't want to pull so many tasks that a group would go idle.
5885 * Except of course for the group_imb case, since then we might
5886 * have to drop below capacity to reach cpu-load equilibrium.
5888 load_above_capacity =
5889 (busiest->sum_nr_running - busiest->group_capacity);
5891 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5892 load_above_capacity /= busiest->group_power;
5896 * We're trying to get all the cpus to the average_load, so we don't
5897 * want to push ourselves above the average load, nor do we wish to
5898 * reduce the max loaded cpu below the average load. At the same time,
5899 * we also don't want to reduce the group load below the group capacity
5900 * (so that we can implement power-savings policies etc). Thus we look
5901 * for the minimum possible imbalance.
5903 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5905 /* How much load to actually move to equalise the imbalance */
5906 env->imbalance = min(
5907 max_pull * busiest->group_power,
5908 (sds->avg_load - local->avg_load) * local->group_power
5909 ) / SCHED_POWER_SCALE;
5912 * if *imbalance is less than the average load per runnable task
5913 * there is no guarantee that any tasks will be moved so we'll have
5914 * a think about bumping its value to force at least one task to be
5917 if (env->imbalance < busiest->load_per_task)
5918 return fix_small_imbalance(env, sds);
5921 /******* find_busiest_group() helpers end here *********************/
5924 * find_busiest_group - Returns the busiest group within the sched_domain
5925 * if there is an imbalance. If there isn't an imbalance, and
5926 * the user has opted for power-savings, it returns a group whose
5927 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5928 * such a group exists.
5930 * Also calculates the amount of weighted load which should be moved
5931 * to restore balance.
5933 * @env: The load balancing environment.
5935 * Return: - The busiest group if imbalance exists.
5936 * - If no imbalance and user has opted for power-savings balance,
5937 * return the least loaded group whose CPUs can be
5938 * put to idle by rebalancing its tasks onto our group.
5940 static struct sched_group *find_busiest_group(struct lb_env *env)
5942 struct sg_lb_stats *local, *busiest;
5943 struct sd_lb_stats sds;
5945 init_sd_lb_stats(&sds);
5948 * Compute the various statistics relavent for load balancing at
5951 update_sd_lb_stats(env, &sds);
5952 local = &sds.local_stat;
5953 busiest = &sds.busiest_stat;
5955 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5956 check_asym_packing(env, &sds))
5959 /* There is no busy sibling group to pull tasks from */
5960 if (!sds.busiest || busiest->sum_nr_running == 0)
5963 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5966 * If the busiest group is imbalanced the below checks don't
5967 * work because they assume all things are equal, which typically
5968 * isn't true due to cpus_allowed constraints and the like.
5970 if (busiest->group_imb)
5973 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5974 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5975 !busiest->group_has_capacity)
5979 * If the local group is more busy than the selected busiest group
5980 * don't try and pull any tasks.
5982 if (local->avg_load >= busiest->avg_load)
5986 * Don't pull any tasks if this group is already above the domain
5989 if (local->avg_load >= sds.avg_load)
5992 if (env->idle == CPU_IDLE) {
5994 * This cpu is idle. If the busiest group load doesn't
5995 * have more tasks than the number of available cpu's and
5996 * there is no imbalance between this and busiest group
5997 * wrt to idle cpu's, it is balanced.
5999 if ((local->idle_cpus < busiest->idle_cpus) &&
6000 busiest->sum_nr_running <= busiest->group_weight)
6004 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6005 * imbalance_pct to be conservative.
6007 if (100 * busiest->avg_load <=
6008 env->sd->imbalance_pct * local->avg_load)
6013 /* Looks like there is an imbalance. Compute it */
6014 calculate_imbalance(env, &sds);
6023 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6025 static struct rq *find_busiest_queue(struct lb_env *env,
6026 struct sched_group *group)
6028 struct rq *busiest = NULL, *rq;
6029 unsigned long busiest_load = 0, busiest_power = 1;
6032 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6033 unsigned long power, capacity, wl;
6037 rt = fbq_classify_rq(rq);
6040 * We classify groups/runqueues into three groups:
6041 * - regular: there are !numa tasks
6042 * - remote: there are numa tasks that run on the 'wrong' node
6043 * - all: there is no distinction
6045 * In order to avoid migrating ideally placed numa tasks,
6046 * ignore those when there's better options.
6048 * If we ignore the actual busiest queue to migrate another
6049 * task, the next balance pass can still reduce the busiest
6050 * queue by moving tasks around inside the node.
6052 * If we cannot move enough load due to this classification
6053 * the next pass will adjust the group classification and
6054 * allow migration of more tasks.
6056 * Both cases only affect the total convergence complexity.
6058 if (rt > env->fbq_type)
6061 power = power_of(i);
6062 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
6064 capacity = fix_small_capacity(env->sd, group);
6066 wl = weighted_cpuload(i);
6069 * When comparing with imbalance, use weighted_cpuload()
6070 * which is not scaled with the cpu power.
6072 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6076 * For the load comparisons with the other cpu's, consider
6077 * the weighted_cpuload() scaled with the cpu power, so that
6078 * the load can be moved away from the cpu that is potentially
6079 * running at a lower capacity.
6081 * Thus we're looking for max(wl_i / power_i), crosswise
6082 * multiplication to rid ourselves of the division works out
6083 * to: wl_i * power_j > wl_j * power_i; where j is our
6086 if (wl * busiest_power > busiest_load * power) {
6088 busiest_power = power;
6097 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6098 * so long as it is large enough.
6100 #define MAX_PINNED_INTERVAL 512
6102 /* Working cpumask for load_balance and load_balance_newidle. */
6103 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6105 static int need_active_balance(struct lb_env *env)
6107 struct sched_domain *sd = env->sd;
6109 if (env->idle == CPU_NEWLY_IDLE) {
6112 * ASYM_PACKING needs to force migrate tasks from busy but
6113 * higher numbered CPUs in order to pack all tasks in the
6114 * lowest numbered CPUs.
6116 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6120 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6123 static int active_load_balance_cpu_stop(void *data);
6125 static int should_we_balance(struct lb_env *env)
6127 struct sched_group *sg = env->sd->groups;
6128 struct cpumask *sg_cpus, *sg_mask;
6129 int cpu, balance_cpu = -1;
6132 * In the newly idle case, we will allow all the cpu's
6133 * to do the newly idle load balance.
6135 if (env->idle == CPU_NEWLY_IDLE)
6138 sg_cpus = sched_group_cpus(sg);
6139 sg_mask = sched_group_mask(sg);
6140 /* Try to find first idle cpu */
6141 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6142 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6149 if (balance_cpu == -1)
6150 balance_cpu = group_balance_cpu(sg);
6153 * First idle cpu or the first cpu(busiest) in this sched group
6154 * is eligible for doing load balancing at this and above domains.
6156 return balance_cpu == env->dst_cpu;
6160 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6161 * tasks if there is an imbalance.
6163 static int load_balance(int this_cpu, struct rq *this_rq,
6164 struct sched_domain *sd, enum cpu_idle_type idle,
6165 int *continue_balancing)
6167 int ld_moved, cur_ld_moved, active_balance = 0;
6168 struct sched_domain *sd_parent = sd->parent;
6169 struct sched_group *group;
6171 unsigned long flags;
6172 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6174 struct lb_env env = {
6176 .dst_cpu = this_cpu,
6178 .dst_grpmask = sched_group_cpus(sd->groups),
6180 .loop_break = sched_nr_migrate_break,
6186 * For NEWLY_IDLE load_balancing, we don't need to consider
6187 * other cpus in our group
6189 if (idle == CPU_NEWLY_IDLE)
6190 env.dst_grpmask = NULL;
6192 cpumask_copy(cpus, cpu_active_mask);
6194 schedstat_inc(sd, lb_count[idle]);
6197 if (!should_we_balance(&env)) {
6198 *continue_balancing = 0;
6202 group = find_busiest_group(&env);
6204 schedstat_inc(sd, lb_nobusyg[idle]);
6208 busiest = find_busiest_queue(&env, group);
6210 schedstat_inc(sd, lb_nobusyq[idle]);
6214 BUG_ON(busiest == env.dst_rq);
6216 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6219 if (busiest->nr_running > 1) {
6221 * Attempt to move tasks. If find_busiest_group has found
6222 * an imbalance but busiest->nr_running <= 1, the group is
6223 * still unbalanced. ld_moved simply stays zero, so it is
6224 * correctly treated as an imbalance.
6226 env.flags |= LBF_ALL_PINNED;
6227 env.src_cpu = busiest->cpu;
6228 env.src_rq = busiest;
6229 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6232 local_irq_save(flags);
6233 double_rq_lock(env.dst_rq, busiest);
6236 * cur_ld_moved - load moved in current iteration
6237 * ld_moved - cumulative load moved across iterations
6239 cur_ld_moved = move_tasks(&env);
6240 ld_moved += cur_ld_moved;
6241 double_rq_unlock(env.dst_rq, busiest);
6242 local_irq_restore(flags);
6245 * some other cpu did the load balance for us.
6247 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6248 resched_cpu(env.dst_cpu);
6250 if (env.flags & LBF_NEED_BREAK) {
6251 env.flags &= ~LBF_NEED_BREAK;
6256 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6257 * us and move them to an alternate dst_cpu in our sched_group
6258 * where they can run. The upper limit on how many times we
6259 * iterate on same src_cpu is dependent on number of cpus in our
6262 * This changes load balance semantics a bit on who can move
6263 * load to a given_cpu. In addition to the given_cpu itself
6264 * (or a ilb_cpu acting on its behalf where given_cpu is
6265 * nohz-idle), we now have balance_cpu in a position to move
6266 * load to given_cpu. In rare situations, this may cause
6267 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6268 * _independently_ and at _same_ time to move some load to
6269 * given_cpu) causing exceess load to be moved to given_cpu.
6270 * This however should not happen so much in practice and
6271 * moreover subsequent load balance cycles should correct the
6272 * excess load moved.
6274 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6276 /* Prevent to re-select dst_cpu via env's cpus */
6277 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6279 env.dst_rq = cpu_rq(env.new_dst_cpu);
6280 env.dst_cpu = env.new_dst_cpu;
6281 env.flags &= ~LBF_DST_PINNED;
6283 env.loop_break = sched_nr_migrate_break;
6286 * Go back to "more_balance" rather than "redo" since we
6287 * need to continue with same src_cpu.
6293 * We failed to reach balance because of affinity.
6296 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6298 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6299 *group_imbalance = 1;
6300 } else if (*group_imbalance)
6301 *group_imbalance = 0;
6304 /* All tasks on this runqueue were pinned by CPU affinity */
6305 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6306 cpumask_clear_cpu(cpu_of(busiest), cpus);
6307 if (!cpumask_empty(cpus)) {
6309 env.loop_break = sched_nr_migrate_break;
6317 schedstat_inc(sd, lb_failed[idle]);
6319 * Increment the failure counter only on periodic balance.
6320 * We do not want newidle balance, which can be very
6321 * frequent, pollute the failure counter causing
6322 * excessive cache_hot migrations and active balances.
6324 if (idle != CPU_NEWLY_IDLE)
6325 sd->nr_balance_failed++;
6327 if (need_active_balance(&env)) {
6328 raw_spin_lock_irqsave(&busiest->lock, flags);
6330 /* don't kick the active_load_balance_cpu_stop,
6331 * if the curr task on busiest cpu can't be
6334 if (!cpumask_test_cpu(this_cpu,
6335 tsk_cpus_allowed(busiest->curr))) {
6336 raw_spin_unlock_irqrestore(&busiest->lock,
6338 env.flags |= LBF_ALL_PINNED;
6339 goto out_one_pinned;
6343 * ->active_balance synchronizes accesses to
6344 * ->active_balance_work. Once set, it's cleared
6345 * only after active load balance is finished.
6347 if (!busiest->active_balance) {
6348 busiest->active_balance = 1;
6349 busiest->push_cpu = this_cpu;
6352 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6354 if (active_balance) {
6355 stop_one_cpu_nowait(cpu_of(busiest),
6356 active_load_balance_cpu_stop, busiest,
6357 &busiest->active_balance_work);
6361 * We've kicked active balancing, reset the failure
6364 sd->nr_balance_failed = sd->cache_nice_tries+1;
6367 sd->nr_balance_failed = 0;
6369 if (likely(!active_balance)) {
6370 /* We were unbalanced, so reset the balancing interval */
6371 sd->balance_interval = sd->min_interval;
6374 * If we've begun active balancing, start to back off. This
6375 * case may not be covered by the all_pinned logic if there
6376 * is only 1 task on the busy runqueue (because we don't call
6379 if (sd->balance_interval < sd->max_interval)
6380 sd->balance_interval *= 2;
6386 schedstat_inc(sd, lb_balanced[idle]);
6388 sd->nr_balance_failed = 0;
6391 /* tune up the balancing interval */
6392 if (((env.flags & LBF_ALL_PINNED) &&
6393 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6394 (sd->balance_interval < sd->max_interval))
6395 sd->balance_interval *= 2;
6403 * idle_balance is called by schedule() if this_cpu is about to become
6404 * idle. Attempts to pull tasks from other CPUs.
6406 void idle_balance(int this_cpu, struct rq *this_rq)
6408 struct sched_domain *sd;
6409 int pulled_task = 0;
6410 unsigned long next_balance = jiffies + HZ;
6413 this_rq->idle_stamp = rq_clock(this_rq);
6415 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6419 * Drop the rq->lock, but keep IRQ/preempt disabled.
6421 raw_spin_unlock(&this_rq->lock);
6423 update_blocked_averages(this_cpu);
6425 for_each_domain(this_cpu, sd) {
6426 unsigned long interval;
6427 int continue_balancing = 1;
6428 u64 t0, domain_cost;
6430 if (!(sd->flags & SD_LOAD_BALANCE))
6433 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6436 if (sd->flags & SD_BALANCE_NEWIDLE) {
6437 t0 = sched_clock_cpu(this_cpu);
6439 /* If we've pulled tasks over stop searching: */
6440 pulled_task = load_balance(this_cpu, this_rq,
6442 &continue_balancing);
6444 domain_cost = sched_clock_cpu(this_cpu) - t0;
6445 if (domain_cost > sd->max_newidle_lb_cost)
6446 sd->max_newidle_lb_cost = domain_cost;
6448 curr_cost += domain_cost;
6451 interval = msecs_to_jiffies(sd->balance_interval);
6452 if (time_after(next_balance, sd->last_balance + interval))
6453 next_balance = sd->last_balance + interval;
6455 this_rq->idle_stamp = 0;
6461 raw_spin_lock(&this_rq->lock);
6463 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6465 * We are going idle. next_balance may be set based on
6466 * a busy processor. So reset next_balance.
6468 this_rq->next_balance = next_balance;
6471 if (curr_cost > this_rq->max_idle_balance_cost)
6472 this_rq->max_idle_balance_cost = curr_cost;
6476 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6477 * running tasks off the busiest CPU onto idle CPUs. It requires at
6478 * least 1 task to be running on each physical CPU where possible, and
6479 * avoids physical / logical imbalances.
6481 static int active_load_balance_cpu_stop(void *data)
6483 struct rq *busiest_rq = data;
6484 int busiest_cpu = cpu_of(busiest_rq);
6485 int target_cpu = busiest_rq->push_cpu;
6486 struct rq *target_rq = cpu_rq(target_cpu);
6487 struct sched_domain *sd;
6489 raw_spin_lock_irq(&busiest_rq->lock);
6491 /* make sure the requested cpu hasn't gone down in the meantime */
6492 if (unlikely(busiest_cpu != smp_processor_id() ||
6493 !busiest_rq->active_balance))
6496 /* Is there any task to move? */
6497 if (busiest_rq->nr_running <= 1)
6501 * This condition is "impossible", if it occurs
6502 * we need to fix it. Originally reported by
6503 * Bjorn Helgaas on a 128-cpu setup.
6505 BUG_ON(busiest_rq == target_rq);
6507 /* move a task from busiest_rq to target_rq */
6508 double_lock_balance(busiest_rq, target_rq);
6510 /* Search for an sd spanning us and the target CPU. */
6512 for_each_domain(target_cpu, sd) {
6513 if ((sd->flags & SD_LOAD_BALANCE) &&
6514 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6519 struct lb_env env = {
6521 .dst_cpu = target_cpu,
6522 .dst_rq = target_rq,
6523 .src_cpu = busiest_rq->cpu,
6524 .src_rq = busiest_rq,
6528 schedstat_inc(sd, alb_count);
6530 if (move_one_task(&env))
6531 schedstat_inc(sd, alb_pushed);
6533 schedstat_inc(sd, alb_failed);
6536 double_unlock_balance(busiest_rq, target_rq);
6538 busiest_rq->active_balance = 0;
6539 raw_spin_unlock_irq(&busiest_rq->lock);
6543 #ifdef CONFIG_NO_HZ_COMMON
6545 * idle load balancing details
6546 * - When one of the busy CPUs notice that there may be an idle rebalancing
6547 * needed, they will kick the idle load balancer, which then does idle
6548 * load balancing for all the idle CPUs.
6551 cpumask_var_t idle_cpus_mask;
6553 unsigned long next_balance; /* in jiffy units */
6554 } nohz ____cacheline_aligned;
6556 static inline int find_new_ilb(void)
6558 int ilb = cpumask_first(nohz.idle_cpus_mask);
6560 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6567 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6568 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6569 * CPU (if there is one).
6571 static void nohz_balancer_kick(void)
6575 nohz.next_balance++;
6577 ilb_cpu = find_new_ilb();
6579 if (ilb_cpu >= nr_cpu_ids)
6582 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6585 * Use smp_send_reschedule() instead of resched_cpu().
6586 * This way we generate a sched IPI on the target cpu which
6587 * is idle. And the softirq performing nohz idle load balance
6588 * will be run before returning from the IPI.
6590 smp_send_reschedule(ilb_cpu);
6594 static inline void nohz_balance_exit_idle(int cpu)
6596 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6597 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6598 atomic_dec(&nohz.nr_cpus);
6599 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6603 static inline void set_cpu_sd_state_busy(void)
6605 struct sched_domain *sd;
6606 int cpu = smp_processor_id();
6609 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6611 if (!sd || !sd->nohz_idle)
6615 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6620 void set_cpu_sd_state_idle(void)
6622 struct sched_domain *sd;
6623 int cpu = smp_processor_id();
6626 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6628 if (!sd || sd->nohz_idle)
6632 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6638 * This routine will record that the cpu is going idle with tick stopped.
6639 * This info will be used in performing idle load balancing in the future.
6641 void nohz_balance_enter_idle(int cpu)
6644 * If this cpu is going down, then nothing needs to be done.
6646 if (!cpu_active(cpu))
6649 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6652 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6653 atomic_inc(&nohz.nr_cpus);
6654 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6657 static int sched_ilb_notifier(struct notifier_block *nfb,
6658 unsigned long action, void *hcpu)
6660 switch (action & ~CPU_TASKS_FROZEN) {
6662 nohz_balance_exit_idle(smp_processor_id());
6670 static DEFINE_SPINLOCK(balancing);
6673 * Scale the max load_balance interval with the number of CPUs in the system.
6674 * This trades load-balance latency on larger machines for less cross talk.
6676 void update_max_interval(void)
6678 max_load_balance_interval = HZ*num_online_cpus()/10;
6682 * It checks each scheduling domain to see if it is due to be balanced,
6683 * and initiates a balancing operation if so.
6685 * Balancing parameters are set up in init_sched_domains.
6687 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
6689 int continue_balancing = 1;
6691 unsigned long interval;
6692 struct sched_domain *sd;
6693 /* Earliest time when we have to do rebalance again */
6694 unsigned long next_balance = jiffies + 60*HZ;
6695 int update_next_balance = 0;
6696 int need_serialize, need_decay = 0;
6699 update_blocked_averages(cpu);
6702 for_each_domain(cpu, sd) {
6704 * Decay the newidle max times here because this is a regular
6705 * visit to all the domains. Decay ~1% per second.
6707 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6708 sd->max_newidle_lb_cost =
6709 (sd->max_newidle_lb_cost * 253) / 256;
6710 sd->next_decay_max_lb_cost = jiffies + HZ;
6713 max_cost += sd->max_newidle_lb_cost;
6715 if (!(sd->flags & SD_LOAD_BALANCE))
6719 * Stop the load balance at this level. There is another
6720 * CPU in our sched group which is doing load balancing more
6723 if (!continue_balancing) {
6729 interval = sd->balance_interval;
6730 if (idle != CPU_IDLE)
6731 interval *= sd->busy_factor;
6733 /* scale ms to jiffies */
6734 interval = msecs_to_jiffies(interval);
6735 interval = clamp(interval, 1UL, max_load_balance_interval);
6737 need_serialize = sd->flags & SD_SERIALIZE;
6739 if (need_serialize) {
6740 if (!spin_trylock(&balancing))
6744 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6745 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6747 * The LBF_DST_PINNED logic could have changed
6748 * env->dst_cpu, so we can't know our idle
6749 * state even if we migrated tasks. Update it.
6751 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6753 sd->last_balance = jiffies;
6756 spin_unlock(&balancing);
6758 if (time_after(next_balance, sd->last_balance + interval)) {
6759 next_balance = sd->last_balance + interval;
6760 update_next_balance = 1;
6765 * Ensure the rq-wide value also decays but keep it at a
6766 * reasonable floor to avoid funnies with rq->avg_idle.
6768 rq->max_idle_balance_cost =
6769 max((u64)sysctl_sched_migration_cost, max_cost);
6774 * next_balance will be updated only when there is a need.
6775 * When the cpu is attached to null domain for ex, it will not be
6778 if (likely(update_next_balance))
6779 rq->next_balance = next_balance;
6782 #ifdef CONFIG_NO_HZ_COMMON
6784 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6785 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6787 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
6789 int this_cpu = this_rq->cpu;
6793 if (idle != CPU_IDLE ||
6794 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6797 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6798 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6802 * If this cpu gets work to do, stop the load balancing
6803 * work being done for other cpus. Next load
6804 * balancing owner will pick it up.
6809 rq = cpu_rq(balance_cpu);
6811 raw_spin_lock_irq(&rq->lock);
6812 update_rq_clock(rq);
6813 update_idle_cpu_load(rq);
6814 raw_spin_unlock_irq(&rq->lock);
6816 rebalance_domains(rq, CPU_IDLE);
6818 if (time_after(this_rq->next_balance, rq->next_balance))
6819 this_rq->next_balance = rq->next_balance;
6821 nohz.next_balance = this_rq->next_balance;
6823 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6827 * Current heuristic for kicking the idle load balancer in the presence
6828 * of an idle cpu is the system.
6829 * - This rq has more than one task.
6830 * - At any scheduler domain level, this cpu's scheduler group has multiple
6831 * busy cpu's exceeding the group's power.
6832 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6833 * domain span are idle.
6835 static inline int nohz_kick_needed(struct rq *rq)
6837 unsigned long now = jiffies;
6838 struct sched_domain *sd;
6839 struct sched_group_power *sgp;
6840 int nr_busy, cpu = rq->cpu;
6842 if (unlikely(rq->idle_balance))
6846 * We may be recently in ticked or tickless idle mode. At the first
6847 * busy tick after returning from idle, we will update the busy stats.
6849 set_cpu_sd_state_busy();
6850 nohz_balance_exit_idle(cpu);
6853 * None are in tickless mode and hence no need for NOHZ idle load
6856 if (likely(!atomic_read(&nohz.nr_cpus)))
6859 if (time_before(now, nohz.next_balance))
6862 if (rq->nr_running >= 2)
6866 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6869 sgp = sd->groups->sgp;
6870 nr_busy = atomic_read(&sgp->nr_busy_cpus);
6873 goto need_kick_unlock;
6876 sd = rcu_dereference(per_cpu(sd_asym, cpu));
6878 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
6879 sched_domain_span(sd)) < cpu))
6880 goto need_kick_unlock;
6891 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
6895 * run_rebalance_domains is triggered when needed from the scheduler tick.
6896 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6898 static void run_rebalance_domains(struct softirq_action *h)
6900 struct rq *this_rq = this_rq();
6901 enum cpu_idle_type idle = this_rq->idle_balance ?
6902 CPU_IDLE : CPU_NOT_IDLE;
6904 rebalance_domains(this_rq, idle);
6907 * If this cpu has a pending nohz_balance_kick, then do the
6908 * balancing on behalf of the other idle cpus whose ticks are
6911 nohz_idle_balance(this_rq, idle);
6914 static inline int on_null_domain(struct rq *rq)
6916 return !rcu_dereference_sched(rq->sd);
6920 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6922 void trigger_load_balance(struct rq *rq)
6924 /* Don't need to rebalance while attached to NULL domain */
6925 if (unlikely(on_null_domain(rq)))
6928 if (time_after_eq(jiffies, rq->next_balance))
6929 raise_softirq(SCHED_SOFTIRQ);
6930 #ifdef CONFIG_NO_HZ_COMMON
6931 if (nohz_kick_needed(rq))
6932 nohz_balancer_kick();
6936 static void rq_online_fair(struct rq *rq)
6941 static void rq_offline_fair(struct rq *rq)
6945 /* Ensure any throttled groups are reachable by pick_next_task */
6946 unthrottle_offline_cfs_rqs(rq);
6949 #endif /* CONFIG_SMP */
6952 * scheduler tick hitting a task of our scheduling class:
6954 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6956 struct cfs_rq *cfs_rq;
6957 struct sched_entity *se = &curr->se;
6959 for_each_sched_entity(se) {
6960 cfs_rq = cfs_rq_of(se);
6961 entity_tick(cfs_rq, se, queued);
6964 if (numabalancing_enabled)
6965 task_tick_numa(rq, curr);
6967 update_rq_runnable_avg(rq, 1);
6971 * called on fork with the child task as argument from the parent's context
6972 * - child not yet on the tasklist
6973 * - preemption disabled
6975 static void task_fork_fair(struct task_struct *p)
6977 struct cfs_rq *cfs_rq;
6978 struct sched_entity *se = &p->se, *curr;
6979 int this_cpu = smp_processor_id();
6980 struct rq *rq = this_rq();
6981 unsigned long flags;
6983 raw_spin_lock_irqsave(&rq->lock, flags);
6985 update_rq_clock(rq);
6987 cfs_rq = task_cfs_rq(current);
6988 curr = cfs_rq->curr;
6991 * Not only the cpu but also the task_group of the parent might have
6992 * been changed after parent->se.parent,cfs_rq were copied to
6993 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6994 * of child point to valid ones.
6997 __set_task_cpu(p, this_cpu);
7000 update_curr(cfs_rq);
7003 se->vruntime = curr->vruntime;
7004 place_entity(cfs_rq, se, 1);
7006 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7008 * Upon rescheduling, sched_class::put_prev_task() will place
7009 * 'current' within the tree based on its new key value.
7011 swap(curr->vruntime, se->vruntime);
7012 resched_task(rq->curr);
7015 se->vruntime -= cfs_rq->min_vruntime;
7017 raw_spin_unlock_irqrestore(&rq->lock, flags);
7021 * Priority of the task has changed. Check to see if we preempt
7025 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7031 * Reschedule if we are currently running on this runqueue and
7032 * our priority decreased, or if we are not currently running on
7033 * this runqueue and our priority is higher than the current's
7035 if (rq->curr == p) {
7036 if (p->prio > oldprio)
7037 resched_task(rq->curr);
7039 check_preempt_curr(rq, p, 0);
7042 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7044 struct sched_entity *se = &p->se;
7045 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7048 * Ensure the task's vruntime is normalized, so that when its
7049 * switched back to the fair class the enqueue_entity(.flags=0) will
7050 * do the right thing.
7052 * If it was on_rq, then the dequeue_entity(.flags=0) will already
7053 * have normalized the vruntime, if it was !on_rq, then only when
7054 * the task is sleeping will it still have non-normalized vruntime.
7056 if (!se->on_rq && p->state != TASK_RUNNING) {
7058 * Fix up our vruntime so that the current sleep doesn't
7059 * cause 'unlimited' sleep bonus.
7061 place_entity(cfs_rq, se, 0);
7062 se->vruntime -= cfs_rq->min_vruntime;
7067 * Remove our load from contribution when we leave sched_fair
7068 * and ensure we don't carry in an old decay_count if we
7071 if (se->avg.decay_count) {
7072 __synchronize_entity_decay(se);
7073 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7079 * We switched to the sched_fair class.
7081 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7087 * We were most likely switched from sched_rt, so
7088 * kick off the schedule if running, otherwise just see
7089 * if we can still preempt the current task.
7092 resched_task(rq->curr);
7094 check_preempt_curr(rq, p, 0);
7097 /* Account for a task changing its policy or group.
7099 * This routine is mostly called to set cfs_rq->curr field when a task
7100 * migrates between groups/classes.
7102 static void set_curr_task_fair(struct rq *rq)
7104 struct sched_entity *se = &rq->curr->se;
7106 for_each_sched_entity(se) {
7107 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7109 set_next_entity(cfs_rq, se);
7110 /* ensure bandwidth has been allocated on our new cfs_rq */
7111 account_cfs_rq_runtime(cfs_rq, 0);
7115 void init_cfs_rq(struct cfs_rq *cfs_rq)
7117 cfs_rq->tasks_timeline = RB_ROOT;
7118 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7119 #ifndef CONFIG_64BIT
7120 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7123 atomic64_set(&cfs_rq->decay_counter, 1);
7124 atomic_long_set(&cfs_rq->removed_load, 0);
7128 #ifdef CONFIG_FAIR_GROUP_SCHED
7129 static void task_move_group_fair(struct task_struct *p, int on_rq)
7131 struct cfs_rq *cfs_rq;
7133 * If the task was not on the rq at the time of this cgroup movement
7134 * it must have been asleep, sleeping tasks keep their ->vruntime
7135 * absolute on their old rq until wakeup (needed for the fair sleeper
7136 * bonus in place_entity()).
7138 * If it was on the rq, we've just 'preempted' it, which does convert
7139 * ->vruntime to a relative base.
7141 * Make sure both cases convert their relative position when migrating
7142 * to another cgroup's rq. This does somewhat interfere with the
7143 * fair sleeper stuff for the first placement, but who cares.
7146 * When !on_rq, vruntime of the task has usually NOT been normalized.
7147 * But there are some cases where it has already been normalized:
7149 * - Moving a forked child which is waiting for being woken up by
7150 * wake_up_new_task().
7151 * - Moving a task which has been woken up by try_to_wake_up() and
7152 * waiting for actually being woken up by sched_ttwu_pending().
7154 * To prevent boost or penalty in the new cfs_rq caused by delta
7155 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7157 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7161 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7162 set_task_rq(p, task_cpu(p));
7164 cfs_rq = cfs_rq_of(&p->se);
7165 p->se.vruntime += cfs_rq->min_vruntime;
7168 * migrate_task_rq_fair() will have removed our previous
7169 * contribution, but we must synchronize for ongoing future
7172 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7173 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7178 void free_fair_sched_group(struct task_group *tg)
7182 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7184 for_each_possible_cpu(i) {
7186 kfree(tg->cfs_rq[i]);
7195 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7197 struct cfs_rq *cfs_rq;
7198 struct sched_entity *se;
7201 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7204 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7208 tg->shares = NICE_0_LOAD;
7210 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7212 for_each_possible_cpu(i) {
7213 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7214 GFP_KERNEL, cpu_to_node(i));
7218 se = kzalloc_node(sizeof(struct sched_entity),
7219 GFP_KERNEL, cpu_to_node(i));
7223 init_cfs_rq(cfs_rq);
7224 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7235 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7237 struct rq *rq = cpu_rq(cpu);
7238 unsigned long flags;
7241 * Only empty task groups can be destroyed; so we can speculatively
7242 * check on_list without danger of it being re-added.
7244 if (!tg->cfs_rq[cpu]->on_list)
7247 raw_spin_lock_irqsave(&rq->lock, flags);
7248 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7249 raw_spin_unlock_irqrestore(&rq->lock, flags);
7252 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7253 struct sched_entity *se, int cpu,
7254 struct sched_entity *parent)
7256 struct rq *rq = cpu_rq(cpu);
7260 init_cfs_rq_runtime(cfs_rq);
7262 tg->cfs_rq[cpu] = cfs_rq;
7265 /* se could be NULL for root_task_group */
7270 se->cfs_rq = &rq->cfs;
7272 se->cfs_rq = parent->my_q;
7275 /* guarantee group entities always have weight */
7276 update_load_set(&se->load, NICE_0_LOAD);
7277 se->parent = parent;
7280 static DEFINE_MUTEX(shares_mutex);
7282 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7285 unsigned long flags;
7288 * We can't change the weight of the root cgroup.
7293 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7295 mutex_lock(&shares_mutex);
7296 if (tg->shares == shares)
7299 tg->shares = shares;
7300 for_each_possible_cpu(i) {
7301 struct rq *rq = cpu_rq(i);
7302 struct sched_entity *se;
7305 /* Propagate contribution to hierarchy */
7306 raw_spin_lock_irqsave(&rq->lock, flags);
7308 /* Possible calls to update_curr() need rq clock */
7309 update_rq_clock(rq);
7310 for_each_sched_entity(se)
7311 update_cfs_shares(group_cfs_rq(se));
7312 raw_spin_unlock_irqrestore(&rq->lock, flags);
7316 mutex_unlock(&shares_mutex);
7319 #else /* CONFIG_FAIR_GROUP_SCHED */
7321 void free_fair_sched_group(struct task_group *tg) { }
7323 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7328 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7330 #endif /* CONFIG_FAIR_GROUP_SCHED */
7333 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7335 struct sched_entity *se = &task->se;
7336 unsigned int rr_interval = 0;
7339 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7342 if (rq->cfs.load.weight)
7343 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7349 * All the scheduling class methods:
7351 const struct sched_class fair_sched_class = {
7352 .next = &idle_sched_class,
7353 .enqueue_task = enqueue_task_fair,
7354 .dequeue_task = dequeue_task_fair,
7355 .yield_task = yield_task_fair,
7356 .yield_to_task = yield_to_task_fair,
7358 .check_preempt_curr = check_preempt_wakeup,
7360 .pick_next_task = pick_next_task_fair,
7361 .put_prev_task = put_prev_task_fair,
7364 .select_task_rq = select_task_rq_fair,
7365 .migrate_task_rq = migrate_task_rq_fair,
7367 .rq_online = rq_online_fair,
7368 .rq_offline = rq_offline_fair,
7370 .task_waking = task_waking_fair,
7373 .set_curr_task = set_curr_task_fair,
7374 .task_tick = task_tick_fair,
7375 .task_fork = task_fork_fair,
7377 .prio_changed = prio_changed_fair,
7378 .switched_from = switched_from_fair,
7379 .switched_to = switched_to_fair,
7381 .get_rr_interval = get_rr_interval_fair,
7383 #ifdef CONFIG_FAIR_GROUP_SCHED
7384 .task_move_group = task_move_group_fair,
7388 #ifdef CONFIG_SCHED_DEBUG
7389 void print_cfs_stats(struct seq_file *m, int cpu)
7391 struct cfs_rq *cfs_rq;
7394 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7395 print_cfs_rq(m, cpu, cfs_rq);
7400 __init void init_sched_fair_class(void)
7403 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7405 #ifdef CONFIG_NO_HZ_COMMON
7406 nohz.next_balance = jiffies;
7407 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7408 cpu_notifier(sched_ilb_notifier, 0);