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;
891 * Faults_cpu is used to decide whether memory should move
892 * towards the CPU. As a consequence, these stats are weighted
893 * more by CPU use than by memory faults.
895 unsigned long *faults_cpu;
896 unsigned long faults[0];
899 /* Shared or private faults. */
900 #define NR_NUMA_HINT_FAULT_TYPES 2
902 /* Memory and CPU locality */
903 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
905 /* Averaged statistics, and temporary buffers. */
906 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
908 pid_t task_numa_group_id(struct task_struct *p)
910 return p->numa_group ? p->numa_group->gid : 0;
913 static inline int task_faults_idx(int nid, int priv)
915 return NR_NUMA_HINT_FAULT_TYPES * nid + priv;
918 static inline unsigned long task_faults(struct task_struct *p, int nid)
920 if (!p->numa_faults_memory)
923 return p->numa_faults_memory[task_faults_idx(nid, 0)] +
924 p->numa_faults_memory[task_faults_idx(nid, 1)];
927 static inline unsigned long group_faults(struct task_struct *p, int nid)
932 return p->numa_group->faults[task_faults_idx(nid, 0)] +
933 p->numa_group->faults[task_faults_idx(nid, 1)];
936 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
938 return group->faults_cpu[task_faults_idx(nid, 0)] +
939 group->faults_cpu[task_faults_idx(nid, 1)];
943 * These return the fraction of accesses done by a particular task, or
944 * task group, on a particular numa node. The group weight is given a
945 * larger multiplier, in order to group tasks together that are almost
946 * evenly spread out between numa nodes.
948 static inline unsigned long task_weight(struct task_struct *p, int nid)
950 unsigned long total_faults;
952 if (!p->numa_faults_memory)
955 total_faults = p->total_numa_faults;
960 return 1000 * task_faults(p, nid) / total_faults;
963 static inline unsigned long group_weight(struct task_struct *p, int nid)
965 if (!p->numa_group || !p->numa_group->total_faults)
968 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
971 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
972 int src_nid, int dst_cpu)
974 struct numa_group *ng = p->numa_group;
975 int dst_nid = cpu_to_node(dst_cpu);
976 int last_cpupid, this_cpupid;
978 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
981 * Multi-stage node selection is used in conjunction with a periodic
982 * migration fault to build a temporal task<->page relation. By using
983 * a two-stage filter we remove short/unlikely relations.
985 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
986 * a task's usage of a particular page (n_p) per total usage of this
987 * page (n_t) (in a given time-span) to a probability.
989 * Our periodic faults will sample this probability and getting the
990 * same result twice in a row, given these samples are fully
991 * independent, is then given by P(n)^2, provided our sample period
992 * is sufficiently short compared to the usage pattern.
994 * This quadric squishes small probabilities, making it less likely we
995 * act on an unlikely task<->page relation.
997 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
998 if (!cpupid_pid_unset(last_cpupid) &&
999 cpupid_to_nid(last_cpupid) != dst_nid)
1002 /* Always allow migrate on private faults */
1003 if (cpupid_match_pid(p, last_cpupid))
1006 /* A shared fault, but p->numa_group has not been set up yet. */
1011 * Do not migrate if the destination is not a node that
1012 * is actively used by this numa group.
1014 if (!node_isset(dst_nid, ng->active_nodes))
1018 * Source is a node that is not actively used by this
1019 * numa group, while the destination is. Migrate.
1021 if (!node_isset(src_nid, ng->active_nodes))
1025 * Both source and destination are nodes in active
1026 * use by this numa group. Maximize memory bandwidth
1027 * by migrating from more heavily used groups, to less
1028 * heavily used ones, spreading the load around.
1029 * Use a 1/4 hysteresis to avoid spurious page movement.
1031 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1034 static unsigned long weighted_cpuload(const int cpu);
1035 static unsigned long source_load(int cpu, int type);
1036 static unsigned long target_load(int cpu, int type);
1037 static unsigned long power_of(int cpu);
1038 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1040 /* Cached statistics for all CPUs within a node */
1042 unsigned long nr_running;
1045 /* Total compute capacity of CPUs on a node */
1046 unsigned long power;
1048 /* Approximate capacity in terms of runnable tasks on a node */
1049 unsigned long capacity;
1054 * XXX borrowed from update_sg_lb_stats
1056 static void update_numa_stats(struct numa_stats *ns, int nid)
1060 memset(ns, 0, sizeof(*ns));
1061 for_each_cpu(cpu, cpumask_of_node(nid)) {
1062 struct rq *rq = cpu_rq(cpu);
1064 ns->nr_running += rq->nr_running;
1065 ns->load += weighted_cpuload(cpu);
1066 ns->power += power_of(cpu);
1072 * If we raced with hotplug and there are no CPUs left in our mask
1073 * the @ns structure is NULL'ed and task_numa_compare() will
1074 * not find this node attractive.
1076 * We'll either bail at !has_capacity, or we'll detect a huge imbalance
1082 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1083 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1084 ns->has_capacity = (ns->nr_running < ns->capacity);
1087 struct task_numa_env {
1088 struct task_struct *p;
1090 int src_cpu, src_nid;
1091 int dst_cpu, dst_nid;
1093 struct numa_stats src_stats, dst_stats;
1097 struct task_struct *best_task;
1102 static void task_numa_assign(struct task_numa_env *env,
1103 struct task_struct *p, long imp)
1106 put_task_struct(env->best_task);
1111 env->best_imp = imp;
1112 env->best_cpu = env->dst_cpu;
1116 * This checks if the overall compute and NUMA accesses of the system would
1117 * be improved if the source tasks was migrated to the target dst_cpu taking
1118 * into account that it might be best if task running on the dst_cpu should
1119 * be exchanged with the source task
1121 static void task_numa_compare(struct task_numa_env *env,
1122 long taskimp, long groupimp)
1124 struct rq *src_rq = cpu_rq(env->src_cpu);
1125 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1126 struct task_struct *cur;
1127 long dst_load, src_load;
1129 long imp = (groupimp > 0) ? groupimp : taskimp;
1132 cur = ACCESS_ONCE(dst_rq->curr);
1133 if (cur->pid == 0) /* idle */
1137 * "imp" is the fault differential for the source task between the
1138 * source and destination node. Calculate the total differential for
1139 * the source task and potential destination task. The more negative
1140 * the value is, the more rmeote accesses that would be expected to
1141 * be incurred if the tasks were swapped.
1144 /* Skip this swap candidate if cannot move to the source cpu */
1145 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1149 * If dst and source tasks are in the same NUMA group, or not
1150 * in any group then look only at task weights.
1152 if (cur->numa_group == env->p->numa_group) {
1153 imp = taskimp + task_weight(cur, env->src_nid) -
1154 task_weight(cur, env->dst_nid);
1156 * Add some hysteresis to prevent swapping the
1157 * tasks within a group over tiny differences.
1159 if (cur->numa_group)
1163 * Compare the group weights. If a task is all by
1164 * itself (not part of a group), use the task weight
1167 if (env->p->numa_group)
1172 if (cur->numa_group)
1173 imp += group_weight(cur, env->src_nid) -
1174 group_weight(cur, env->dst_nid);
1176 imp += task_weight(cur, env->src_nid) -
1177 task_weight(cur, env->dst_nid);
1181 if (imp < env->best_imp)
1185 /* Is there capacity at our destination? */
1186 if (env->src_stats.has_capacity &&
1187 !env->dst_stats.has_capacity)
1193 /* Balance doesn't matter much if we're running a task per cpu */
1194 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1198 * In the overloaded case, try and keep the load balanced.
1201 dst_load = env->dst_stats.load;
1202 src_load = env->src_stats.load;
1204 /* XXX missing power terms */
1205 load = task_h_load(env->p);
1210 load = task_h_load(cur);
1215 /* make src_load the smaller */
1216 if (dst_load < src_load)
1217 swap(dst_load, src_load);
1219 if (src_load * env->imbalance_pct < dst_load * 100)
1223 task_numa_assign(env, cur, imp);
1228 static void task_numa_find_cpu(struct task_numa_env *env,
1229 long taskimp, long groupimp)
1233 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1234 /* Skip this CPU if the source task cannot migrate */
1235 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1239 task_numa_compare(env, taskimp, groupimp);
1243 static int task_numa_migrate(struct task_struct *p)
1245 struct task_numa_env env = {
1248 .src_cpu = task_cpu(p),
1249 .src_nid = task_node(p),
1251 .imbalance_pct = 112,
1257 struct sched_domain *sd;
1258 unsigned long taskweight, groupweight;
1260 long taskimp, groupimp;
1263 * Pick the lowest SD_NUMA domain, as that would have the smallest
1264 * imbalance and would be the first to start moving tasks about.
1266 * And we want to avoid any moving of tasks about, as that would create
1267 * random movement of tasks -- counter the numa conditions we're trying
1271 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1273 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1277 * Cpusets can break the scheduler domain tree into smaller
1278 * balance domains, some of which do not cross NUMA boundaries.
1279 * Tasks that are "trapped" in such domains cannot be migrated
1280 * elsewhere, so there is no point in (re)trying.
1282 if (unlikely(!sd)) {
1283 p->numa_preferred_nid = task_node(p);
1287 taskweight = task_weight(p, env.src_nid);
1288 groupweight = group_weight(p, env.src_nid);
1289 update_numa_stats(&env.src_stats, env.src_nid);
1290 env.dst_nid = p->numa_preferred_nid;
1291 taskimp = task_weight(p, env.dst_nid) - taskweight;
1292 groupimp = group_weight(p, env.dst_nid) - groupweight;
1293 update_numa_stats(&env.dst_stats, env.dst_nid);
1295 /* If the preferred nid has capacity, try to use it. */
1296 if (env.dst_stats.has_capacity)
1297 task_numa_find_cpu(&env, taskimp, groupimp);
1299 /* No space available on the preferred nid. Look elsewhere. */
1300 if (env.best_cpu == -1) {
1301 for_each_online_node(nid) {
1302 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1305 /* Only consider nodes where both task and groups benefit */
1306 taskimp = task_weight(p, nid) - taskweight;
1307 groupimp = group_weight(p, nid) - groupweight;
1308 if (taskimp < 0 && groupimp < 0)
1312 update_numa_stats(&env.dst_stats, env.dst_nid);
1313 task_numa_find_cpu(&env, taskimp, groupimp);
1317 /* No better CPU than the current one was found. */
1318 if (env.best_cpu == -1)
1321 sched_setnuma(p, env.dst_nid);
1324 * Reset the scan period if the task is being rescheduled on an
1325 * alternative node to recheck if the tasks is now properly placed.
1327 p->numa_scan_period = task_scan_min(p);
1329 if (env.best_task == NULL) {
1330 ret = migrate_task_to(p, env.best_cpu);
1332 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1336 ret = migrate_swap(p, env.best_task);
1338 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1339 put_task_struct(env.best_task);
1343 /* Attempt to migrate a task to a CPU on the preferred node. */
1344 static void numa_migrate_preferred(struct task_struct *p)
1346 /* This task has no NUMA fault statistics yet */
1347 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1350 /* Periodically retry migrating the task to the preferred node */
1351 p->numa_migrate_retry = jiffies + HZ;
1353 /* Success if task is already running on preferred CPU */
1354 if (task_node(p) == p->numa_preferred_nid)
1357 /* Otherwise, try migrate to a CPU on the preferred node */
1358 task_numa_migrate(p);
1362 * Find the nodes on which the workload is actively running. We do this by
1363 * tracking the nodes from which NUMA hinting faults are triggered. This can
1364 * be different from the set of nodes where the workload's memory is currently
1367 * The bitmask is used to make smarter decisions on when to do NUMA page
1368 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1369 * are added when they cause over 6/16 of the maximum number of faults, but
1370 * only removed when they drop below 3/16.
1372 static void update_numa_active_node_mask(struct numa_group *numa_group)
1374 unsigned long faults, max_faults = 0;
1377 for_each_online_node(nid) {
1378 faults = group_faults_cpu(numa_group, nid);
1379 if (faults > max_faults)
1380 max_faults = faults;
1383 for_each_online_node(nid) {
1384 faults = group_faults_cpu(numa_group, nid);
1385 if (!node_isset(nid, numa_group->active_nodes)) {
1386 if (faults > max_faults * 6 / 16)
1387 node_set(nid, numa_group->active_nodes);
1388 } else if (faults < max_faults * 3 / 16)
1389 node_clear(nid, numa_group->active_nodes);
1394 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1395 * increments. The more local the fault statistics are, the higher the scan
1396 * period will be for the next scan window. If local/remote ratio is below
1397 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1398 * scan period will decrease
1400 #define NUMA_PERIOD_SLOTS 10
1401 #define NUMA_PERIOD_THRESHOLD 3
1404 * Increase the scan period (slow down scanning) if the majority of
1405 * our memory is already on our local node, or if the majority of
1406 * the page accesses are shared with other processes.
1407 * Otherwise, decrease the scan period.
1409 static void update_task_scan_period(struct task_struct *p,
1410 unsigned long shared, unsigned long private)
1412 unsigned int period_slot;
1416 unsigned long remote = p->numa_faults_locality[0];
1417 unsigned long local = p->numa_faults_locality[1];
1420 * If there were no record hinting faults then either the task is
1421 * completely idle or all activity is areas that are not of interest
1422 * to automatic numa balancing. Scan slower
1424 if (local + shared == 0) {
1425 p->numa_scan_period = min(p->numa_scan_period_max,
1426 p->numa_scan_period << 1);
1428 p->mm->numa_next_scan = jiffies +
1429 msecs_to_jiffies(p->numa_scan_period);
1435 * Prepare to scale scan period relative to the current period.
1436 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1437 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1438 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1440 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1441 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1442 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1443 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1446 diff = slot * period_slot;
1448 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1451 * Scale scan rate increases based on sharing. There is an
1452 * inverse relationship between the degree of sharing and
1453 * the adjustment made to the scanning period. Broadly
1454 * speaking the intent is that there is little point
1455 * scanning faster if shared accesses dominate as it may
1456 * simply bounce migrations uselessly
1458 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1459 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1462 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1463 task_scan_min(p), task_scan_max(p));
1464 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1468 * Get the fraction of time the task has been running since the last
1469 * NUMA placement cycle. The scheduler keeps similar statistics, but
1470 * decays those on a 32ms period, which is orders of magnitude off
1471 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1472 * stats only if the task is so new there are no NUMA statistics yet.
1474 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1476 u64 runtime, delta, now;
1477 /* Use the start of this time slice to avoid calculations. */
1478 now = p->se.exec_start;
1479 runtime = p->se.sum_exec_runtime;
1481 if (p->last_task_numa_placement) {
1482 delta = runtime - p->last_sum_exec_runtime;
1483 *period = now - p->last_task_numa_placement;
1485 delta = p->se.avg.runnable_avg_sum;
1486 *period = p->se.avg.runnable_avg_period;
1489 p->last_sum_exec_runtime = runtime;
1490 p->last_task_numa_placement = now;
1495 static void task_numa_placement(struct task_struct *p)
1497 int seq, nid, max_nid = -1, max_group_nid = -1;
1498 unsigned long max_faults = 0, max_group_faults = 0;
1499 unsigned long fault_types[2] = { 0, 0 };
1500 unsigned long total_faults;
1501 u64 runtime, period;
1502 spinlock_t *group_lock = NULL;
1504 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1505 if (p->numa_scan_seq == seq)
1507 p->numa_scan_seq = seq;
1508 p->numa_scan_period_max = task_scan_max(p);
1510 total_faults = p->numa_faults_locality[0] +
1511 p->numa_faults_locality[1];
1512 runtime = numa_get_avg_runtime(p, &period);
1514 /* If the task is part of a group prevent parallel updates to group stats */
1515 if (p->numa_group) {
1516 group_lock = &p->numa_group->lock;
1517 spin_lock(group_lock);
1520 /* Find the node with the highest number of faults */
1521 for_each_online_node(nid) {
1522 unsigned long faults = 0, group_faults = 0;
1525 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1526 long diff, f_diff, f_weight;
1528 i = task_faults_idx(nid, priv);
1530 /* Decay existing window, copy faults since last scan */
1531 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1532 fault_types[priv] += p->numa_faults_buffer_memory[i];
1533 p->numa_faults_buffer_memory[i] = 0;
1536 * Normalize the faults_from, so all tasks in a group
1537 * count according to CPU use, instead of by the raw
1538 * number of faults. Tasks with little runtime have
1539 * little over-all impact on throughput, and thus their
1540 * faults are less important.
1542 f_weight = div64_u64(runtime << 16, period + 1);
1543 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1545 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1546 p->numa_faults_buffer_cpu[i] = 0;
1548 p->numa_faults_memory[i] += diff;
1549 p->numa_faults_cpu[i] += f_diff;
1550 faults += p->numa_faults_memory[i];
1551 p->total_numa_faults += diff;
1552 if (p->numa_group) {
1553 /* safe because we can only change our own group */
1554 p->numa_group->faults[i] += diff;
1555 p->numa_group->faults_cpu[i] += f_diff;
1556 p->numa_group->total_faults += diff;
1557 group_faults += p->numa_group->faults[i];
1561 if (faults > max_faults) {
1562 max_faults = faults;
1566 if (group_faults > max_group_faults) {
1567 max_group_faults = group_faults;
1568 max_group_nid = nid;
1572 update_task_scan_period(p, fault_types[0], fault_types[1]);
1574 if (p->numa_group) {
1575 update_numa_active_node_mask(p->numa_group);
1577 * If the preferred task and group nids are different,
1578 * iterate over the nodes again to find the best place.
1580 if (max_nid != max_group_nid) {
1581 unsigned long weight, max_weight = 0;
1583 for_each_online_node(nid) {
1584 weight = task_weight(p, nid) + group_weight(p, nid);
1585 if (weight > max_weight) {
1586 max_weight = weight;
1592 spin_unlock(group_lock);
1595 /* Preferred node as the node with the most faults */
1596 if (max_faults && max_nid != p->numa_preferred_nid) {
1597 /* Update the preferred nid and migrate task if possible */
1598 sched_setnuma(p, max_nid);
1599 numa_migrate_preferred(p);
1603 static inline int get_numa_group(struct numa_group *grp)
1605 return atomic_inc_not_zero(&grp->refcount);
1608 static inline void put_numa_group(struct numa_group *grp)
1610 if (atomic_dec_and_test(&grp->refcount))
1611 kfree_rcu(grp, rcu);
1614 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1617 struct numa_group *grp, *my_grp;
1618 struct task_struct *tsk;
1620 int cpu = cpupid_to_cpu(cpupid);
1623 if (unlikely(!p->numa_group)) {
1624 unsigned int size = sizeof(struct numa_group) +
1625 4*nr_node_ids*sizeof(unsigned long);
1627 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1631 atomic_set(&grp->refcount, 1);
1632 spin_lock_init(&grp->lock);
1633 INIT_LIST_HEAD(&grp->task_list);
1635 /* Second half of the array tracks nids where faults happen */
1636 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1639 node_set(task_node(current), grp->active_nodes);
1641 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1642 grp->faults[i] = p->numa_faults_memory[i];
1644 grp->total_faults = p->total_numa_faults;
1646 list_add(&p->numa_entry, &grp->task_list);
1648 rcu_assign_pointer(p->numa_group, grp);
1652 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1654 if (!cpupid_match_pid(tsk, cpupid))
1657 grp = rcu_dereference(tsk->numa_group);
1661 my_grp = p->numa_group;
1666 * Only join the other group if its bigger; if we're the bigger group,
1667 * the other task will join us.
1669 if (my_grp->nr_tasks > grp->nr_tasks)
1673 * Tie-break on the grp address.
1675 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1678 /* Always join threads in the same process. */
1679 if (tsk->mm == current->mm)
1682 /* Simple filter to avoid false positives due to PID collisions */
1683 if (flags & TNF_SHARED)
1686 /* Update priv based on whether false sharing was detected */
1689 if (join && !get_numa_group(grp))
1697 double_lock(&my_grp->lock, &grp->lock);
1699 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1700 my_grp->faults[i] -= p->numa_faults_memory[i];
1701 grp->faults[i] += p->numa_faults_memory[i];
1703 my_grp->total_faults -= p->total_numa_faults;
1704 grp->total_faults += p->total_numa_faults;
1706 list_move(&p->numa_entry, &grp->task_list);
1710 spin_unlock(&my_grp->lock);
1711 spin_unlock(&grp->lock);
1713 rcu_assign_pointer(p->numa_group, grp);
1715 put_numa_group(my_grp);
1723 void task_numa_free(struct task_struct *p)
1725 struct numa_group *grp = p->numa_group;
1727 void *numa_faults = p->numa_faults_memory;
1730 spin_lock(&grp->lock);
1731 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1732 grp->faults[i] -= p->numa_faults_memory[i];
1733 grp->total_faults -= p->total_numa_faults;
1735 list_del(&p->numa_entry);
1737 spin_unlock(&grp->lock);
1738 rcu_assign_pointer(p->numa_group, NULL);
1739 put_numa_group(grp);
1742 p->numa_faults_memory = NULL;
1743 p->numa_faults_buffer_memory = NULL;
1744 p->numa_faults_cpu= NULL;
1745 p->numa_faults_buffer_cpu = NULL;
1750 * Got a PROT_NONE fault for a page on @node.
1752 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1754 struct task_struct *p = current;
1755 bool migrated = flags & TNF_MIGRATED;
1756 int cpu_node = task_node(current);
1759 if (!numabalancing_enabled)
1762 /* for example, ksmd faulting in a user's mm */
1766 /* Do not worry about placement if exiting */
1767 if (p->state == TASK_DEAD)
1770 /* Allocate buffer to track faults on a per-node basis */
1771 if (unlikely(!p->numa_faults_memory)) {
1772 int size = sizeof(*p->numa_faults_memory) *
1773 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1775 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1776 if (!p->numa_faults_memory)
1779 BUG_ON(p->numa_faults_buffer_memory);
1781 * The averaged statistics, shared & private, memory & cpu,
1782 * occupy the first half of the array. The second half of the
1783 * array is for current counters, which are averaged into the
1784 * first set by task_numa_placement.
1786 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1787 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1788 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1789 p->total_numa_faults = 0;
1790 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1794 * First accesses are treated as private, otherwise consider accesses
1795 * to be private if the accessing pid has not changed
1797 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1800 priv = cpupid_match_pid(p, last_cpupid);
1801 if (!priv && !(flags & TNF_NO_GROUP))
1802 task_numa_group(p, last_cpupid, flags, &priv);
1805 task_numa_placement(p);
1808 * Retry task to preferred node migration periodically, in case it
1809 * case it previously failed, or the scheduler moved us.
1811 if (time_after(jiffies, p->numa_migrate_retry))
1812 numa_migrate_preferred(p);
1815 p->numa_pages_migrated += pages;
1817 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1818 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1819 p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages;
1822 static void reset_ptenuma_scan(struct task_struct *p)
1824 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1825 p->mm->numa_scan_offset = 0;
1829 * The expensive part of numa migration is done from task_work context.
1830 * Triggered from task_tick_numa().
1832 void task_numa_work(struct callback_head *work)
1834 unsigned long migrate, next_scan, now = jiffies;
1835 struct task_struct *p = current;
1836 struct mm_struct *mm = p->mm;
1837 struct vm_area_struct *vma;
1838 unsigned long start, end;
1839 unsigned long nr_pte_updates = 0;
1842 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1844 work->next = work; /* protect against double add */
1846 * Who cares about NUMA placement when they're dying.
1848 * NOTE: make sure not to dereference p->mm before this check,
1849 * exit_task_work() happens _after_ exit_mm() so we could be called
1850 * without p->mm even though we still had it when we enqueued this
1853 if (p->flags & PF_EXITING)
1856 if (!mm->numa_next_scan) {
1857 mm->numa_next_scan = now +
1858 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1862 * Enforce maximal scan/migration frequency..
1864 migrate = mm->numa_next_scan;
1865 if (time_before(now, migrate))
1868 if (p->numa_scan_period == 0) {
1869 p->numa_scan_period_max = task_scan_max(p);
1870 p->numa_scan_period = task_scan_min(p);
1873 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1874 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1878 * Delay this task enough that another task of this mm will likely win
1879 * the next time around.
1881 p->node_stamp += 2 * TICK_NSEC;
1883 start = mm->numa_scan_offset;
1884 pages = sysctl_numa_balancing_scan_size;
1885 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1889 down_read(&mm->mmap_sem);
1890 vma = find_vma(mm, start);
1892 reset_ptenuma_scan(p);
1896 for (; vma; vma = vma->vm_next) {
1897 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1901 * Shared library pages mapped by multiple processes are not
1902 * migrated as it is expected they are cache replicated. Avoid
1903 * hinting faults in read-only file-backed mappings or the vdso
1904 * as migrating the pages will be of marginal benefit.
1907 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1911 * Skip inaccessible VMAs to avoid any confusion between
1912 * PROT_NONE and NUMA hinting ptes
1914 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1918 start = max(start, vma->vm_start);
1919 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1920 end = min(end, vma->vm_end);
1921 nr_pte_updates += change_prot_numa(vma, start, end);
1924 * Scan sysctl_numa_balancing_scan_size but ensure that
1925 * at least one PTE is updated so that unused virtual
1926 * address space is quickly skipped.
1929 pages -= (end - start) >> PAGE_SHIFT;
1934 } while (end != vma->vm_end);
1939 * It is possible to reach the end of the VMA list but the last few
1940 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1941 * would find the !migratable VMA on the next scan but not reset the
1942 * scanner to the start so check it now.
1945 mm->numa_scan_offset = start;
1947 reset_ptenuma_scan(p);
1948 up_read(&mm->mmap_sem);
1952 * Drive the periodic memory faults..
1954 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1956 struct callback_head *work = &curr->numa_work;
1960 * We don't care about NUMA placement if we don't have memory.
1962 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1966 * Using runtime rather than walltime has the dual advantage that
1967 * we (mostly) drive the selection from busy threads and that the
1968 * task needs to have done some actual work before we bother with
1971 now = curr->se.sum_exec_runtime;
1972 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1974 if (now - curr->node_stamp > period) {
1975 if (!curr->node_stamp)
1976 curr->numa_scan_period = task_scan_min(curr);
1977 curr->node_stamp += period;
1979 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1980 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1981 task_work_add(curr, work, true);
1986 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1990 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1994 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1997 #endif /* CONFIG_NUMA_BALANCING */
2000 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2002 update_load_add(&cfs_rq->load, se->load.weight);
2003 if (!parent_entity(se))
2004 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2006 if (entity_is_task(se)) {
2007 struct rq *rq = rq_of(cfs_rq);
2009 account_numa_enqueue(rq, task_of(se));
2010 list_add(&se->group_node, &rq->cfs_tasks);
2013 cfs_rq->nr_running++;
2017 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2019 update_load_sub(&cfs_rq->load, se->load.weight);
2020 if (!parent_entity(se))
2021 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2022 if (entity_is_task(se)) {
2023 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2024 list_del_init(&se->group_node);
2026 cfs_rq->nr_running--;
2029 #ifdef CONFIG_FAIR_GROUP_SCHED
2031 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2036 * Use this CPU's actual weight instead of the last load_contribution
2037 * to gain a more accurate current total weight. See
2038 * update_cfs_rq_load_contribution().
2040 tg_weight = atomic_long_read(&tg->load_avg);
2041 tg_weight -= cfs_rq->tg_load_contrib;
2042 tg_weight += cfs_rq->load.weight;
2047 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2049 long tg_weight, load, shares;
2051 tg_weight = calc_tg_weight(tg, cfs_rq);
2052 load = cfs_rq->load.weight;
2054 shares = (tg->shares * load);
2056 shares /= tg_weight;
2058 if (shares < MIN_SHARES)
2059 shares = MIN_SHARES;
2060 if (shares > tg->shares)
2061 shares = tg->shares;
2065 # else /* CONFIG_SMP */
2066 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2070 # endif /* CONFIG_SMP */
2071 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2072 unsigned long weight)
2075 /* commit outstanding execution time */
2076 if (cfs_rq->curr == se)
2077 update_curr(cfs_rq);
2078 account_entity_dequeue(cfs_rq, se);
2081 update_load_set(&se->load, weight);
2084 account_entity_enqueue(cfs_rq, se);
2087 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2089 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2091 struct task_group *tg;
2092 struct sched_entity *se;
2096 se = tg->se[cpu_of(rq_of(cfs_rq))];
2097 if (!se || throttled_hierarchy(cfs_rq))
2100 if (likely(se->load.weight == tg->shares))
2103 shares = calc_cfs_shares(cfs_rq, tg);
2105 reweight_entity(cfs_rq_of(se), se, shares);
2107 #else /* CONFIG_FAIR_GROUP_SCHED */
2108 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2111 #endif /* CONFIG_FAIR_GROUP_SCHED */
2115 * We choose a half-life close to 1 scheduling period.
2116 * Note: The tables below are dependent on this value.
2118 #define LOAD_AVG_PERIOD 32
2119 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2120 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2122 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2123 static const u32 runnable_avg_yN_inv[] = {
2124 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2125 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2126 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2127 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2128 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2129 0x85aac367, 0x82cd8698,
2133 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2134 * over-estimates when re-combining.
2136 static const u32 runnable_avg_yN_sum[] = {
2137 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2138 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2139 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2144 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2146 static __always_inline u64 decay_load(u64 val, u64 n)
2148 unsigned int local_n;
2152 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2155 /* after bounds checking we can collapse to 32-bit */
2159 * As y^PERIOD = 1/2, we can combine
2160 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2161 * With a look-up table which covers k^n (n<PERIOD)
2163 * To achieve constant time decay_load.
2165 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2166 val >>= local_n / LOAD_AVG_PERIOD;
2167 local_n %= LOAD_AVG_PERIOD;
2170 val *= runnable_avg_yN_inv[local_n];
2171 /* We don't use SRR here since we always want to round down. */
2176 * For updates fully spanning n periods, the contribution to runnable
2177 * average will be: \Sum 1024*y^n
2179 * We can compute this reasonably efficiently by combining:
2180 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2182 static u32 __compute_runnable_contrib(u64 n)
2186 if (likely(n <= LOAD_AVG_PERIOD))
2187 return runnable_avg_yN_sum[n];
2188 else if (unlikely(n >= LOAD_AVG_MAX_N))
2189 return LOAD_AVG_MAX;
2191 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2193 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2194 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2196 n -= LOAD_AVG_PERIOD;
2197 } while (n > LOAD_AVG_PERIOD);
2199 contrib = decay_load(contrib, n);
2200 return contrib + runnable_avg_yN_sum[n];
2204 * We can represent the historical contribution to runnable average as the
2205 * coefficients of a geometric series. To do this we sub-divide our runnable
2206 * history into segments of approximately 1ms (1024us); label the segment that
2207 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2209 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2211 * (now) (~1ms ago) (~2ms ago)
2213 * Let u_i denote the fraction of p_i that the entity was runnable.
2215 * We then designate the fractions u_i as our co-efficients, yielding the
2216 * following representation of historical load:
2217 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2219 * We choose y based on the with of a reasonably scheduling period, fixing:
2222 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2223 * approximately half as much as the contribution to load within the last ms
2226 * When a period "rolls over" and we have new u_0`, multiplying the previous
2227 * sum again by y is sufficient to update:
2228 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2229 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2231 static __always_inline int __update_entity_runnable_avg(u64 now,
2232 struct sched_avg *sa,
2236 u32 runnable_contrib;
2237 int delta_w, decayed = 0;
2239 delta = now - sa->last_runnable_update;
2241 * This should only happen when time goes backwards, which it
2242 * unfortunately does during sched clock init when we swap over to TSC.
2244 if ((s64)delta < 0) {
2245 sa->last_runnable_update = now;
2250 * Use 1024ns as the unit of measurement since it's a reasonable
2251 * approximation of 1us and fast to compute.
2256 sa->last_runnable_update = now;
2258 /* delta_w is the amount already accumulated against our next period */
2259 delta_w = sa->runnable_avg_period % 1024;
2260 if (delta + delta_w >= 1024) {
2261 /* period roll-over */
2265 * Now that we know we're crossing a period boundary, figure
2266 * out how much from delta we need to complete the current
2267 * period and accrue it.
2269 delta_w = 1024 - delta_w;
2271 sa->runnable_avg_sum += delta_w;
2272 sa->runnable_avg_period += delta_w;
2276 /* Figure out how many additional periods this update spans */
2277 periods = delta / 1024;
2280 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2282 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2285 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2286 runnable_contrib = __compute_runnable_contrib(periods);
2288 sa->runnable_avg_sum += runnable_contrib;
2289 sa->runnable_avg_period += runnable_contrib;
2292 /* Remainder of delta accrued against u_0` */
2294 sa->runnable_avg_sum += delta;
2295 sa->runnable_avg_period += delta;
2300 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2301 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2303 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2304 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2306 decays -= se->avg.decay_count;
2310 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2311 se->avg.decay_count = 0;
2316 #ifdef CONFIG_FAIR_GROUP_SCHED
2317 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2320 struct task_group *tg = cfs_rq->tg;
2323 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2324 tg_contrib -= cfs_rq->tg_load_contrib;
2326 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2327 atomic_long_add(tg_contrib, &tg->load_avg);
2328 cfs_rq->tg_load_contrib += tg_contrib;
2333 * Aggregate cfs_rq runnable averages into an equivalent task_group
2334 * representation for computing load contributions.
2336 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2337 struct cfs_rq *cfs_rq)
2339 struct task_group *tg = cfs_rq->tg;
2342 /* The fraction of a cpu used by this cfs_rq */
2343 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2344 sa->runnable_avg_period + 1);
2345 contrib -= cfs_rq->tg_runnable_contrib;
2347 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2348 atomic_add(contrib, &tg->runnable_avg);
2349 cfs_rq->tg_runnable_contrib += contrib;
2353 static inline void __update_group_entity_contrib(struct sched_entity *se)
2355 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2356 struct task_group *tg = cfs_rq->tg;
2361 contrib = cfs_rq->tg_load_contrib * tg->shares;
2362 se->avg.load_avg_contrib = div_u64(contrib,
2363 atomic_long_read(&tg->load_avg) + 1);
2366 * For group entities we need to compute a correction term in the case
2367 * that they are consuming <1 cpu so that we would contribute the same
2368 * load as a task of equal weight.
2370 * Explicitly co-ordinating this measurement would be expensive, but
2371 * fortunately the sum of each cpus contribution forms a usable
2372 * lower-bound on the true value.
2374 * Consider the aggregate of 2 contributions. Either they are disjoint
2375 * (and the sum represents true value) or they are disjoint and we are
2376 * understating by the aggregate of their overlap.
2378 * Extending this to N cpus, for a given overlap, the maximum amount we
2379 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2380 * cpus that overlap for this interval and w_i is the interval width.
2382 * On a small machine; the first term is well-bounded which bounds the
2383 * total error since w_i is a subset of the period. Whereas on a
2384 * larger machine, while this first term can be larger, if w_i is the
2385 * of consequential size guaranteed to see n_i*w_i quickly converge to
2386 * our upper bound of 1-cpu.
2388 runnable_avg = atomic_read(&tg->runnable_avg);
2389 if (runnable_avg < NICE_0_LOAD) {
2390 se->avg.load_avg_contrib *= runnable_avg;
2391 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2395 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2396 int force_update) {}
2397 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2398 struct cfs_rq *cfs_rq) {}
2399 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2402 static inline void __update_task_entity_contrib(struct sched_entity *se)
2406 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2407 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2408 contrib /= (se->avg.runnable_avg_period + 1);
2409 se->avg.load_avg_contrib = scale_load(contrib);
2412 /* Compute the current contribution to load_avg by se, return any delta */
2413 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2415 long old_contrib = se->avg.load_avg_contrib;
2417 if (entity_is_task(se)) {
2418 __update_task_entity_contrib(se);
2420 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2421 __update_group_entity_contrib(se);
2424 return se->avg.load_avg_contrib - old_contrib;
2427 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2430 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2431 cfs_rq->blocked_load_avg -= load_contrib;
2433 cfs_rq->blocked_load_avg = 0;
2436 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2438 /* Update a sched_entity's runnable average */
2439 static inline void update_entity_load_avg(struct sched_entity *se,
2442 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2447 * For a group entity we need to use their owned cfs_rq_clock_task() in
2448 * case they are the parent of a throttled hierarchy.
2450 if (entity_is_task(se))
2451 now = cfs_rq_clock_task(cfs_rq);
2453 now = cfs_rq_clock_task(group_cfs_rq(se));
2455 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2458 contrib_delta = __update_entity_load_avg_contrib(se);
2464 cfs_rq->runnable_load_avg += contrib_delta;
2466 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2470 * Decay the load contributed by all blocked children and account this so that
2471 * their contribution may appropriately discounted when they wake up.
2473 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2475 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2478 decays = now - cfs_rq->last_decay;
2479 if (!decays && !force_update)
2482 if (atomic_long_read(&cfs_rq->removed_load)) {
2483 unsigned long removed_load;
2484 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2485 subtract_blocked_load_contrib(cfs_rq, removed_load);
2489 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2491 atomic64_add(decays, &cfs_rq->decay_counter);
2492 cfs_rq->last_decay = now;
2495 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2498 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2500 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2501 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2504 /* Add the load generated by se into cfs_rq's child load-average */
2505 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2506 struct sched_entity *se,
2510 * We track migrations using entity decay_count <= 0, on a wake-up
2511 * migration we use a negative decay count to track the remote decays
2512 * accumulated while sleeping.
2514 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2515 * are seen by enqueue_entity_load_avg() as a migration with an already
2516 * constructed load_avg_contrib.
2518 if (unlikely(se->avg.decay_count <= 0)) {
2519 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2520 if (se->avg.decay_count) {
2522 * In a wake-up migration we have to approximate the
2523 * time sleeping. This is because we can't synchronize
2524 * clock_task between the two cpus, and it is not
2525 * guaranteed to be read-safe. Instead, we can
2526 * approximate this using our carried decays, which are
2527 * explicitly atomically readable.
2529 se->avg.last_runnable_update -= (-se->avg.decay_count)
2531 update_entity_load_avg(se, 0);
2532 /* Indicate that we're now synchronized and on-rq */
2533 se->avg.decay_count = 0;
2537 __synchronize_entity_decay(se);
2540 /* migrated tasks did not contribute to our blocked load */
2542 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2543 update_entity_load_avg(se, 0);
2546 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2547 /* we force update consideration on load-balancer moves */
2548 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2552 * Remove se's load from this cfs_rq child load-average, if the entity is
2553 * transitioning to a blocked state we track its projected decay using
2556 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2557 struct sched_entity *se,
2560 update_entity_load_avg(se, 1);
2561 /* we force update consideration on load-balancer moves */
2562 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2564 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2566 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2567 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2568 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2572 * Update the rq's load with the elapsed running time before entering
2573 * idle. if the last scheduled task is not a CFS task, idle_enter will
2574 * be the only way to update the runnable statistic.
2576 void idle_enter_fair(struct rq *this_rq)
2578 update_rq_runnable_avg(this_rq, 1);
2582 * Update the rq's load with the elapsed idle time before a task is
2583 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2584 * be the only way to update the runnable statistic.
2586 void idle_exit_fair(struct rq *this_rq)
2588 update_rq_runnable_avg(this_rq, 0);
2592 static inline void update_entity_load_avg(struct sched_entity *se,
2593 int update_cfs_rq) {}
2594 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2595 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2596 struct sched_entity *se,
2598 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2599 struct sched_entity *se,
2601 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2602 int force_update) {}
2605 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2607 #ifdef CONFIG_SCHEDSTATS
2608 struct task_struct *tsk = NULL;
2610 if (entity_is_task(se))
2613 if (se->statistics.sleep_start) {
2614 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2619 if (unlikely(delta > se->statistics.sleep_max))
2620 se->statistics.sleep_max = delta;
2622 se->statistics.sleep_start = 0;
2623 se->statistics.sum_sleep_runtime += delta;
2626 account_scheduler_latency(tsk, delta >> 10, 1);
2627 trace_sched_stat_sleep(tsk, delta);
2630 if (se->statistics.block_start) {
2631 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2636 if (unlikely(delta > se->statistics.block_max))
2637 se->statistics.block_max = delta;
2639 se->statistics.block_start = 0;
2640 se->statistics.sum_sleep_runtime += delta;
2643 if (tsk->in_iowait) {
2644 se->statistics.iowait_sum += delta;
2645 se->statistics.iowait_count++;
2646 trace_sched_stat_iowait(tsk, delta);
2649 trace_sched_stat_blocked(tsk, delta);
2652 * Blocking time is in units of nanosecs, so shift by
2653 * 20 to get a milliseconds-range estimation of the
2654 * amount of time that the task spent sleeping:
2656 if (unlikely(prof_on == SLEEP_PROFILING)) {
2657 profile_hits(SLEEP_PROFILING,
2658 (void *)get_wchan(tsk),
2661 account_scheduler_latency(tsk, delta >> 10, 0);
2667 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2669 #ifdef CONFIG_SCHED_DEBUG
2670 s64 d = se->vruntime - cfs_rq->min_vruntime;
2675 if (d > 3*sysctl_sched_latency)
2676 schedstat_inc(cfs_rq, nr_spread_over);
2681 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2683 u64 vruntime = cfs_rq->min_vruntime;
2686 * The 'current' period is already promised to the current tasks,
2687 * however the extra weight of the new task will slow them down a
2688 * little, place the new task so that it fits in the slot that
2689 * stays open at the end.
2691 if (initial && sched_feat(START_DEBIT))
2692 vruntime += sched_vslice(cfs_rq, se);
2694 /* sleeps up to a single latency don't count. */
2696 unsigned long thresh = sysctl_sched_latency;
2699 * Halve their sleep time's effect, to allow
2700 * for a gentler effect of sleepers:
2702 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2708 /* ensure we never gain time by being placed backwards. */
2709 se->vruntime = max_vruntime(se->vruntime, vruntime);
2712 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2715 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2718 * Update the normalized vruntime before updating min_vruntime
2719 * through calling update_curr().
2721 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2722 se->vruntime += cfs_rq->min_vruntime;
2725 * Update run-time statistics of the 'current'.
2727 update_curr(cfs_rq);
2728 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2729 account_entity_enqueue(cfs_rq, se);
2730 update_cfs_shares(cfs_rq);
2732 if (flags & ENQUEUE_WAKEUP) {
2733 place_entity(cfs_rq, se, 0);
2734 enqueue_sleeper(cfs_rq, se);
2737 update_stats_enqueue(cfs_rq, se);
2738 check_spread(cfs_rq, se);
2739 if (se != cfs_rq->curr)
2740 __enqueue_entity(cfs_rq, se);
2743 if (cfs_rq->nr_running == 1) {
2744 list_add_leaf_cfs_rq(cfs_rq);
2745 check_enqueue_throttle(cfs_rq);
2749 static void __clear_buddies_last(struct sched_entity *se)
2751 for_each_sched_entity(se) {
2752 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2753 if (cfs_rq->last == se)
2754 cfs_rq->last = NULL;
2760 static void __clear_buddies_next(struct sched_entity *se)
2762 for_each_sched_entity(se) {
2763 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2764 if (cfs_rq->next == se)
2765 cfs_rq->next = NULL;
2771 static void __clear_buddies_skip(struct sched_entity *se)
2773 for_each_sched_entity(se) {
2774 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2775 if (cfs_rq->skip == se)
2776 cfs_rq->skip = NULL;
2782 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2784 if (cfs_rq->last == se)
2785 __clear_buddies_last(se);
2787 if (cfs_rq->next == se)
2788 __clear_buddies_next(se);
2790 if (cfs_rq->skip == se)
2791 __clear_buddies_skip(se);
2794 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2797 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2800 * Update run-time statistics of the 'current'.
2802 update_curr(cfs_rq);
2803 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2805 update_stats_dequeue(cfs_rq, se);
2806 if (flags & DEQUEUE_SLEEP) {
2807 #ifdef CONFIG_SCHEDSTATS
2808 if (entity_is_task(se)) {
2809 struct task_struct *tsk = task_of(se);
2811 if (tsk->state & TASK_INTERRUPTIBLE)
2812 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2813 if (tsk->state & TASK_UNINTERRUPTIBLE)
2814 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2819 clear_buddies(cfs_rq, se);
2821 if (se != cfs_rq->curr)
2822 __dequeue_entity(cfs_rq, se);
2824 account_entity_dequeue(cfs_rq, se);
2827 * Normalize the entity after updating the min_vruntime because the
2828 * update can refer to the ->curr item and we need to reflect this
2829 * movement in our normalized position.
2831 if (!(flags & DEQUEUE_SLEEP))
2832 se->vruntime -= cfs_rq->min_vruntime;
2834 /* return excess runtime on last dequeue */
2835 return_cfs_rq_runtime(cfs_rq);
2837 update_min_vruntime(cfs_rq);
2838 update_cfs_shares(cfs_rq);
2842 * Preempt the current task with a newly woken task if needed:
2845 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2847 unsigned long ideal_runtime, delta_exec;
2848 struct sched_entity *se;
2851 ideal_runtime = sched_slice(cfs_rq, curr);
2852 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2853 if (delta_exec > ideal_runtime) {
2854 resched_task(rq_of(cfs_rq)->curr);
2856 * The current task ran long enough, ensure it doesn't get
2857 * re-elected due to buddy favours.
2859 clear_buddies(cfs_rq, curr);
2864 * Ensure that a task that missed wakeup preemption by a
2865 * narrow margin doesn't have to wait for a full slice.
2866 * This also mitigates buddy induced latencies under load.
2868 if (delta_exec < sysctl_sched_min_granularity)
2871 se = __pick_first_entity(cfs_rq);
2872 delta = curr->vruntime - se->vruntime;
2877 if (delta > ideal_runtime)
2878 resched_task(rq_of(cfs_rq)->curr);
2882 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2884 /* 'current' is not kept within the tree. */
2887 * Any task has to be enqueued before it get to execute on
2888 * a CPU. So account for the time it spent waiting on the
2891 update_stats_wait_end(cfs_rq, se);
2892 __dequeue_entity(cfs_rq, se);
2895 update_stats_curr_start(cfs_rq, se);
2897 #ifdef CONFIG_SCHEDSTATS
2899 * Track our maximum slice length, if the CPU's load is at
2900 * least twice that of our own weight (i.e. dont track it
2901 * when there are only lesser-weight tasks around):
2903 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2904 se->statistics.slice_max = max(se->statistics.slice_max,
2905 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2908 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2912 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2915 * Pick the next process, keeping these things in mind, in this order:
2916 * 1) keep things fair between processes/task groups
2917 * 2) pick the "next" process, since someone really wants that to run
2918 * 3) pick the "last" process, for cache locality
2919 * 4) do not run the "skip" process, if something else is available
2921 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2923 struct sched_entity *se = __pick_first_entity(cfs_rq);
2924 struct sched_entity *left = se;
2927 * Avoid running the skip buddy, if running something else can
2928 * be done without getting too unfair.
2930 if (cfs_rq->skip == se) {
2931 struct sched_entity *second = __pick_next_entity(se);
2932 if (second && wakeup_preempt_entity(second, left) < 1)
2937 * Prefer last buddy, try to return the CPU to a preempted task.
2939 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2943 * Someone really wants this to run. If it's not unfair, run it.
2945 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2948 clear_buddies(cfs_rq, se);
2953 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2955 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2958 * If still on the runqueue then deactivate_task()
2959 * was not called and update_curr() has to be done:
2962 update_curr(cfs_rq);
2964 /* throttle cfs_rqs exceeding runtime */
2965 check_cfs_rq_runtime(cfs_rq);
2967 check_spread(cfs_rq, prev);
2969 update_stats_wait_start(cfs_rq, prev);
2970 /* Put 'current' back into the tree. */
2971 __enqueue_entity(cfs_rq, prev);
2972 /* in !on_rq case, update occurred at dequeue */
2973 update_entity_load_avg(prev, 1);
2975 cfs_rq->curr = NULL;
2979 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2982 * Update run-time statistics of the 'current'.
2984 update_curr(cfs_rq);
2987 * Ensure that runnable average is periodically updated.
2989 update_entity_load_avg(curr, 1);
2990 update_cfs_rq_blocked_load(cfs_rq, 1);
2991 update_cfs_shares(cfs_rq);
2993 #ifdef CONFIG_SCHED_HRTICK
2995 * queued ticks are scheduled to match the slice, so don't bother
2996 * validating it and just reschedule.
2999 resched_task(rq_of(cfs_rq)->curr);
3003 * don't let the period tick interfere with the hrtick preemption
3005 if (!sched_feat(DOUBLE_TICK) &&
3006 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3010 if (cfs_rq->nr_running > 1)
3011 check_preempt_tick(cfs_rq, curr);
3015 /**************************************************
3016 * CFS bandwidth control machinery
3019 #ifdef CONFIG_CFS_BANDWIDTH
3021 #ifdef HAVE_JUMP_LABEL
3022 static struct static_key __cfs_bandwidth_used;
3024 static inline bool cfs_bandwidth_used(void)
3026 return static_key_false(&__cfs_bandwidth_used);
3029 void cfs_bandwidth_usage_inc(void)
3031 static_key_slow_inc(&__cfs_bandwidth_used);
3034 void cfs_bandwidth_usage_dec(void)
3036 static_key_slow_dec(&__cfs_bandwidth_used);
3038 #else /* HAVE_JUMP_LABEL */
3039 static bool cfs_bandwidth_used(void)
3044 void cfs_bandwidth_usage_inc(void) {}
3045 void cfs_bandwidth_usage_dec(void) {}
3046 #endif /* HAVE_JUMP_LABEL */
3049 * default period for cfs group bandwidth.
3050 * default: 0.1s, units: nanoseconds
3052 static inline u64 default_cfs_period(void)
3054 return 100000000ULL;
3057 static inline u64 sched_cfs_bandwidth_slice(void)
3059 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3063 * Replenish runtime according to assigned quota and update expiration time.
3064 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3065 * additional synchronization around rq->lock.
3067 * requires cfs_b->lock
3069 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3073 if (cfs_b->quota == RUNTIME_INF)
3076 now = sched_clock_cpu(smp_processor_id());
3077 cfs_b->runtime = cfs_b->quota;
3078 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3081 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3083 return &tg->cfs_bandwidth;
3086 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3087 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3089 if (unlikely(cfs_rq->throttle_count))
3090 return cfs_rq->throttled_clock_task;
3092 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3095 /* returns 0 on failure to allocate runtime */
3096 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3098 struct task_group *tg = cfs_rq->tg;
3099 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3100 u64 amount = 0, min_amount, expires;
3102 /* note: this is a positive sum as runtime_remaining <= 0 */
3103 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3105 raw_spin_lock(&cfs_b->lock);
3106 if (cfs_b->quota == RUNTIME_INF)
3107 amount = min_amount;
3110 * If the bandwidth pool has become inactive, then at least one
3111 * period must have elapsed since the last consumption.
3112 * Refresh the global state and ensure bandwidth timer becomes
3115 if (!cfs_b->timer_active) {
3116 __refill_cfs_bandwidth_runtime(cfs_b);
3117 __start_cfs_bandwidth(cfs_b);
3120 if (cfs_b->runtime > 0) {
3121 amount = min(cfs_b->runtime, min_amount);
3122 cfs_b->runtime -= amount;
3126 expires = cfs_b->runtime_expires;
3127 raw_spin_unlock(&cfs_b->lock);
3129 cfs_rq->runtime_remaining += amount;
3131 * we may have advanced our local expiration to account for allowed
3132 * spread between our sched_clock and the one on which runtime was
3135 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3136 cfs_rq->runtime_expires = expires;
3138 return cfs_rq->runtime_remaining > 0;
3142 * Note: This depends on the synchronization provided by sched_clock and the
3143 * fact that rq->clock snapshots this value.
3145 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3147 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3149 /* if the deadline is ahead of our clock, nothing to do */
3150 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3153 if (cfs_rq->runtime_remaining < 0)
3157 * If the local deadline has passed we have to consider the
3158 * possibility that our sched_clock is 'fast' and the global deadline
3159 * has not truly expired.
3161 * Fortunately we can check determine whether this the case by checking
3162 * whether the global deadline has advanced.
3165 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
3166 /* extend local deadline, drift is bounded above by 2 ticks */
3167 cfs_rq->runtime_expires += TICK_NSEC;
3169 /* global deadline is ahead, expiration has passed */
3170 cfs_rq->runtime_remaining = 0;
3174 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3176 /* dock delta_exec before expiring quota (as it could span periods) */
3177 cfs_rq->runtime_remaining -= delta_exec;
3178 expire_cfs_rq_runtime(cfs_rq);
3180 if (likely(cfs_rq->runtime_remaining > 0))
3184 * if we're unable to extend our runtime we resched so that the active
3185 * hierarchy can be throttled
3187 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3188 resched_task(rq_of(cfs_rq)->curr);
3191 static __always_inline
3192 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3194 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3197 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3200 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3202 return cfs_bandwidth_used() && cfs_rq->throttled;
3205 /* check whether cfs_rq, or any parent, is throttled */
3206 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3208 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3212 * Ensure that neither of the group entities corresponding to src_cpu or
3213 * dest_cpu are members of a throttled hierarchy when performing group
3214 * load-balance operations.
3216 static inline int throttled_lb_pair(struct task_group *tg,
3217 int src_cpu, int dest_cpu)
3219 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3221 src_cfs_rq = tg->cfs_rq[src_cpu];
3222 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3224 return throttled_hierarchy(src_cfs_rq) ||
3225 throttled_hierarchy(dest_cfs_rq);
3228 /* updated child weight may affect parent so we have to do this bottom up */
3229 static int tg_unthrottle_up(struct task_group *tg, void *data)
3231 struct rq *rq = data;
3232 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3234 cfs_rq->throttle_count--;
3236 if (!cfs_rq->throttle_count) {
3237 /* adjust cfs_rq_clock_task() */
3238 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3239 cfs_rq->throttled_clock_task;
3246 static int tg_throttle_down(struct task_group *tg, void *data)
3248 struct rq *rq = data;
3249 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3251 /* group is entering throttled state, stop time */
3252 if (!cfs_rq->throttle_count)
3253 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3254 cfs_rq->throttle_count++;
3259 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3261 struct rq *rq = rq_of(cfs_rq);
3262 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3263 struct sched_entity *se;
3264 long task_delta, dequeue = 1;
3266 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3268 /* freeze hierarchy runnable averages while throttled */
3270 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3273 task_delta = cfs_rq->h_nr_running;
3274 for_each_sched_entity(se) {
3275 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3276 /* throttled entity or throttle-on-deactivate */
3281 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3282 qcfs_rq->h_nr_running -= task_delta;
3284 if (qcfs_rq->load.weight)
3289 rq->nr_running -= task_delta;
3291 cfs_rq->throttled = 1;
3292 cfs_rq->throttled_clock = rq_clock(rq);
3293 raw_spin_lock(&cfs_b->lock);
3294 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3295 if (!cfs_b->timer_active)
3296 __start_cfs_bandwidth(cfs_b);
3297 raw_spin_unlock(&cfs_b->lock);
3300 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3302 struct rq *rq = rq_of(cfs_rq);
3303 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3304 struct sched_entity *se;
3308 se = cfs_rq->tg->se[cpu_of(rq)];
3310 cfs_rq->throttled = 0;
3312 update_rq_clock(rq);
3314 raw_spin_lock(&cfs_b->lock);
3315 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3316 list_del_rcu(&cfs_rq->throttled_list);
3317 raw_spin_unlock(&cfs_b->lock);
3319 /* update hierarchical throttle state */
3320 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3322 if (!cfs_rq->load.weight)
3325 task_delta = cfs_rq->h_nr_running;
3326 for_each_sched_entity(se) {
3330 cfs_rq = cfs_rq_of(se);
3332 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3333 cfs_rq->h_nr_running += task_delta;
3335 if (cfs_rq_throttled(cfs_rq))
3340 rq->nr_running += task_delta;
3342 /* determine whether we need to wake up potentially idle cpu */
3343 if (rq->curr == rq->idle && rq->cfs.nr_running)
3344 resched_task(rq->curr);
3347 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3348 u64 remaining, u64 expires)
3350 struct cfs_rq *cfs_rq;
3351 u64 runtime = remaining;
3354 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3356 struct rq *rq = rq_of(cfs_rq);
3358 raw_spin_lock(&rq->lock);
3359 if (!cfs_rq_throttled(cfs_rq))
3362 runtime = -cfs_rq->runtime_remaining + 1;
3363 if (runtime > remaining)
3364 runtime = remaining;
3365 remaining -= runtime;
3367 cfs_rq->runtime_remaining += runtime;
3368 cfs_rq->runtime_expires = expires;
3370 /* we check whether we're throttled above */
3371 if (cfs_rq->runtime_remaining > 0)
3372 unthrottle_cfs_rq(cfs_rq);
3375 raw_spin_unlock(&rq->lock);
3386 * Responsible for refilling a task_group's bandwidth and unthrottling its
3387 * cfs_rqs as appropriate. If there has been no activity within the last
3388 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3389 * used to track this state.
3391 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3393 u64 runtime, runtime_expires;
3394 int idle = 1, throttled;
3396 raw_spin_lock(&cfs_b->lock);
3397 /* no need to continue the timer with no bandwidth constraint */
3398 if (cfs_b->quota == RUNTIME_INF)
3401 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3402 /* idle depends on !throttled (for the case of a large deficit) */
3403 idle = cfs_b->idle && !throttled;
3404 cfs_b->nr_periods += overrun;
3406 /* if we're going inactive then everything else can be deferred */
3411 * if we have relooped after returning idle once, we need to update our
3412 * status as actually running, so that other cpus doing
3413 * __start_cfs_bandwidth will stop trying to cancel us.
3415 cfs_b->timer_active = 1;
3417 __refill_cfs_bandwidth_runtime(cfs_b);
3420 /* mark as potentially idle for the upcoming period */
3425 /* account preceding periods in which throttling occurred */
3426 cfs_b->nr_throttled += overrun;
3429 * There are throttled entities so we must first use the new bandwidth
3430 * to unthrottle them before making it generally available. This
3431 * ensures that all existing debts will be paid before a new cfs_rq is
3434 runtime = cfs_b->runtime;
3435 runtime_expires = cfs_b->runtime_expires;
3439 * This check is repeated as we are holding onto the new bandwidth
3440 * while we unthrottle. This can potentially race with an unthrottled
3441 * group trying to acquire new bandwidth from the global pool.
3443 while (throttled && runtime > 0) {
3444 raw_spin_unlock(&cfs_b->lock);
3445 /* we can't nest cfs_b->lock while distributing bandwidth */
3446 runtime = distribute_cfs_runtime(cfs_b, runtime,
3448 raw_spin_lock(&cfs_b->lock);
3450 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3453 /* return (any) remaining runtime */
3454 cfs_b->runtime = runtime;
3456 * While we are ensured activity in the period following an
3457 * unthrottle, this also covers the case in which the new bandwidth is
3458 * insufficient to cover the existing bandwidth deficit. (Forcing the
3459 * timer to remain active while there are any throttled entities.)
3464 cfs_b->timer_active = 0;
3465 raw_spin_unlock(&cfs_b->lock);
3470 /* a cfs_rq won't donate quota below this amount */
3471 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3472 /* minimum remaining period time to redistribute slack quota */
3473 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3474 /* how long we wait to gather additional slack before distributing */
3475 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3478 * Are we near the end of the current quota period?
3480 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3481 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3482 * migrate_hrtimers, base is never cleared, so we are fine.
3484 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3486 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3489 /* if the call-back is running a quota refresh is already occurring */
3490 if (hrtimer_callback_running(refresh_timer))
3493 /* is a quota refresh about to occur? */
3494 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3495 if (remaining < min_expire)
3501 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3503 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3505 /* if there's a quota refresh soon don't bother with slack */
3506 if (runtime_refresh_within(cfs_b, min_left))
3509 start_bandwidth_timer(&cfs_b->slack_timer,
3510 ns_to_ktime(cfs_bandwidth_slack_period));
3513 /* we know any runtime found here is valid as update_curr() precedes return */
3514 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3516 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3517 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3519 if (slack_runtime <= 0)
3522 raw_spin_lock(&cfs_b->lock);
3523 if (cfs_b->quota != RUNTIME_INF &&
3524 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3525 cfs_b->runtime += slack_runtime;
3527 /* we are under rq->lock, defer unthrottling using a timer */
3528 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3529 !list_empty(&cfs_b->throttled_cfs_rq))
3530 start_cfs_slack_bandwidth(cfs_b);
3532 raw_spin_unlock(&cfs_b->lock);
3534 /* even if it's not valid for return we don't want to try again */
3535 cfs_rq->runtime_remaining -= slack_runtime;
3538 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3540 if (!cfs_bandwidth_used())
3543 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3546 __return_cfs_rq_runtime(cfs_rq);
3550 * This is done with a timer (instead of inline with bandwidth return) since
3551 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3553 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3555 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3558 /* confirm we're still not at a refresh boundary */
3559 raw_spin_lock(&cfs_b->lock);
3560 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3561 raw_spin_unlock(&cfs_b->lock);
3565 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3566 runtime = cfs_b->runtime;
3569 expires = cfs_b->runtime_expires;
3570 raw_spin_unlock(&cfs_b->lock);
3575 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3577 raw_spin_lock(&cfs_b->lock);
3578 if (expires == cfs_b->runtime_expires)
3579 cfs_b->runtime = runtime;
3580 raw_spin_unlock(&cfs_b->lock);
3584 * When a group wakes up we want to make sure that its quota is not already
3585 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3586 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3588 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3590 if (!cfs_bandwidth_used())
3593 /* an active group must be handled by the update_curr()->put() path */
3594 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3597 /* ensure the group is not already throttled */
3598 if (cfs_rq_throttled(cfs_rq))
3601 /* update runtime allocation */
3602 account_cfs_rq_runtime(cfs_rq, 0);
3603 if (cfs_rq->runtime_remaining <= 0)
3604 throttle_cfs_rq(cfs_rq);
3607 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3608 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3610 if (!cfs_bandwidth_used())
3613 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3617 * it's possible for a throttled entity to be forced into a running
3618 * state (e.g. set_curr_task), in this case we're finished.
3620 if (cfs_rq_throttled(cfs_rq))
3623 throttle_cfs_rq(cfs_rq);
3626 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3628 struct cfs_bandwidth *cfs_b =
3629 container_of(timer, struct cfs_bandwidth, slack_timer);
3630 do_sched_cfs_slack_timer(cfs_b);
3632 return HRTIMER_NORESTART;
3635 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3637 struct cfs_bandwidth *cfs_b =
3638 container_of(timer, struct cfs_bandwidth, period_timer);
3644 now = hrtimer_cb_get_time(timer);
3645 overrun = hrtimer_forward(timer, now, cfs_b->period);
3650 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3653 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3656 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3658 raw_spin_lock_init(&cfs_b->lock);
3660 cfs_b->quota = RUNTIME_INF;
3661 cfs_b->period = ns_to_ktime(default_cfs_period());
3663 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3664 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3665 cfs_b->period_timer.function = sched_cfs_period_timer;
3666 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3667 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3670 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3672 cfs_rq->runtime_enabled = 0;
3673 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3676 /* requires cfs_b->lock, may release to reprogram timer */
3677 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3680 * The timer may be active because we're trying to set a new bandwidth
3681 * period or because we're racing with the tear-down path
3682 * (timer_active==0 becomes visible before the hrtimer call-back
3683 * terminates). In either case we ensure that it's re-programmed
3685 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3686 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3687 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3688 raw_spin_unlock(&cfs_b->lock);
3690 raw_spin_lock(&cfs_b->lock);
3691 /* if someone else restarted the timer then we're done */
3692 if (cfs_b->timer_active)
3696 cfs_b->timer_active = 1;
3697 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3700 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3702 hrtimer_cancel(&cfs_b->period_timer);
3703 hrtimer_cancel(&cfs_b->slack_timer);
3706 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3708 struct cfs_rq *cfs_rq;
3710 for_each_leaf_cfs_rq(rq, cfs_rq) {
3711 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3713 if (!cfs_rq->runtime_enabled)
3717 * clock_task is not advancing so we just need to make sure
3718 * there's some valid quota amount
3720 cfs_rq->runtime_remaining = cfs_b->quota;
3721 if (cfs_rq_throttled(cfs_rq))
3722 unthrottle_cfs_rq(cfs_rq);
3726 #else /* CONFIG_CFS_BANDWIDTH */
3727 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3729 return rq_clock_task(rq_of(cfs_rq));
3732 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3733 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3734 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3735 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3737 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3742 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3747 static inline int throttled_lb_pair(struct task_group *tg,
3748 int src_cpu, int dest_cpu)
3753 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3755 #ifdef CONFIG_FAIR_GROUP_SCHED
3756 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3759 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3763 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3764 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3766 #endif /* CONFIG_CFS_BANDWIDTH */
3768 /**************************************************
3769 * CFS operations on tasks:
3772 #ifdef CONFIG_SCHED_HRTICK
3773 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3775 struct sched_entity *se = &p->se;
3776 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3778 WARN_ON(task_rq(p) != rq);
3780 if (cfs_rq->nr_running > 1) {
3781 u64 slice = sched_slice(cfs_rq, se);
3782 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3783 s64 delta = slice - ran;
3792 * Don't schedule slices shorter than 10000ns, that just
3793 * doesn't make sense. Rely on vruntime for fairness.
3796 delta = max_t(s64, 10000LL, delta);
3798 hrtick_start(rq, delta);
3803 * called from enqueue/dequeue and updates the hrtick when the
3804 * current task is from our class and nr_running is low enough
3807 static void hrtick_update(struct rq *rq)
3809 struct task_struct *curr = rq->curr;
3811 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3814 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3815 hrtick_start_fair(rq, curr);
3817 #else /* !CONFIG_SCHED_HRTICK */
3819 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3823 static inline void hrtick_update(struct rq *rq)
3829 * The enqueue_task method is called before nr_running is
3830 * increased. Here we update the fair scheduling stats and
3831 * then put the task into the rbtree:
3834 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3836 struct cfs_rq *cfs_rq;
3837 struct sched_entity *se = &p->se;
3839 for_each_sched_entity(se) {
3842 cfs_rq = cfs_rq_of(se);
3843 enqueue_entity(cfs_rq, se, flags);
3846 * end evaluation on encountering a throttled cfs_rq
3848 * note: in the case of encountering a throttled cfs_rq we will
3849 * post the final h_nr_running increment below.
3851 if (cfs_rq_throttled(cfs_rq))
3853 cfs_rq->h_nr_running++;
3855 flags = ENQUEUE_WAKEUP;
3858 for_each_sched_entity(se) {
3859 cfs_rq = cfs_rq_of(se);
3860 cfs_rq->h_nr_running++;
3862 if (cfs_rq_throttled(cfs_rq))
3865 update_cfs_shares(cfs_rq);
3866 update_entity_load_avg(se, 1);
3870 update_rq_runnable_avg(rq, rq->nr_running);
3876 static void set_next_buddy(struct sched_entity *se);
3879 * The dequeue_task method is called before nr_running is
3880 * decreased. We remove the task from the rbtree and
3881 * update the fair scheduling stats:
3883 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3885 struct cfs_rq *cfs_rq;
3886 struct sched_entity *se = &p->se;
3887 int task_sleep = flags & DEQUEUE_SLEEP;
3889 for_each_sched_entity(se) {
3890 cfs_rq = cfs_rq_of(se);
3891 dequeue_entity(cfs_rq, se, flags);
3894 * end evaluation on encountering a throttled cfs_rq
3896 * note: in the case of encountering a throttled cfs_rq we will
3897 * post the final h_nr_running decrement below.
3899 if (cfs_rq_throttled(cfs_rq))
3901 cfs_rq->h_nr_running--;
3903 /* Don't dequeue parent if it has other entities besides us */
3904 if (cfs_rq->load.weight) {
3906 * Bias pick_next to pick a task from this cfs_rq, as
3907 * p is sleeping when it is within its sched_slice.
3909 if (task_sleep && parent_entity(se))
3910 set_next_buddy(parent_entity(se));
3912 /* avoid re-evaluating load for this entity */
3913 se = parent_entity(se);
3916 flags |= DEQUEUE_SLEEP;
3919 for_each_sched_entity(se) {
3920 cfs_rq = cfs_rq_of(se);
3921 cfs_rq->h_nr_running--;
3923 if (cfs_rq_throttled(cfs_rq))
3926 update_cfs_shares(cfs_rq);
3927 update_entity_load_avg(se, 1);
3932 update_rq_runnable_avg(rq, 1);
3938 /* Used instead of source_load when we know the type == 0 */
3939 static unsigned long weighted_cpuload(const int cpu)
3941 return cpu_rq(cpu)->cfs.runnable_load_avg;
3945 * Return a low guess at the load of a migration-source cpu weighted
3946 * according to the scheduling class and "nice" value.
3948 * We want to under-estimate the load of migration sources, to
3949 * balance conservatively.
3951 static unsigned long source_load(int cpu, int type)
3953 struct rq *rq = cpu_rq(cpu);
3954 unsigned long total = weighted_cpuload(cpu);
3956 if (type == 0 || !sched_feat(LB_BIAS))
3959 return min(rq->cpu_load[type-1], total);
3963 * Return a high guess at the load of a migration-target cpu weighted
3964 * according to the scheduling class and "nice" value.
3966 static unsigned long target_load(int cpu, int type)
3968 struct rq *rq = cpu_rq(cpu);
3969 unsigned long total = weighted_cpuload(cpu);
3971 if (type == 0 || !sched_feat(LB_BIAS))
3974 return max(rq->cpu_load[type-1], total);
3977 static unsigned long power_of(int cpu)
3979 return cpu_rq(cpu)->cpu_power;
3982 static unsigned long cpu_avg_load_per_task(int cpu)
3984 struct rq *rq = cpu_rq(cpu);
3985 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3986 unsigned long load_avg = rq->cfs.runnable_load_avg;
3989 return load_avg / nr_running;
3994 static void record_wakee(struct task_struct *p)
3997 * Rough decay (wiping) for cost saving, don't worry
3998 * about the boundary, really active task won't care
4001 if (jiffies > current->wakee_flip_decay_ts + HZ) {
4002 current->wakee_flips = 0;
4003 current->wakee_flip_decay_ts = jiffies;
4006 if (current->last_wakee != p) {
4007 current->last_wakee = p;
4008 current->wakee_flips++;
4012 static void task_waking_fair(struct task_struct *p)
4014 struct sched_entity *se = &p->se;
4015 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4018 #ifndef CONFIG_64BIT
4019 u64 min_vruntime_copy;
4022 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4024 min_vruntime = cfs_rq->min_vruntime;
4025 } while (min_vruntime != min_vruntime_copy);
4027 min_vruntime = cfs_rq->min_vruntime;
4030 se->vruntime -= min_vruntime;
4034 #ifdef CONFIG_FAIR_GROUP_SCHED
4036 * effective_load() calculates the load change as seen from the root_task_group
4038 * Adding load to a group doesn't make a group heavier, but can cause movement
4039 * of group shares between cpus. Assuming the shares were perfectly aligned one
4040 * can calculate the shift in shares.
4042 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4043 * on this @cpu and results in a total addition (subtraction) of @wg to the
4044 * total group weight.
4046 * Given a runqueue weight distribution (rw_i) we can compute a shares
4047 * distribution (s_i) using:
4049 * s_i = rw_i / \Sum rw_j (1)
4051 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4052 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4053 * shares distribution (s_i):
4055 * rw_i = { 2, 4, 1, 0 }
4056 * s_i = { 2/7, 4/7, 1/7, 0 }
4058 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4059 * task used to run on and the CPU the waker is running on), we need to
4060 * compute the effect of waking a task on either CPU and, in case of a sync
4061 * wakeup, compute the effect of the current task going to sleep.
4063 * So for a change of @wl to the local @cpu with an overall group weight change
4064 * of @wl we can compute the new shares distribution (s'_i) using:
4066 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4068 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4069 * differences in waking a task to CPU 0. The additional task changes the
4070 * weight and shares distributions like:
4072 * rw'_i = { 3, 4, 1, 0 }
4073 * s'_i = { 3/8, 4/8, 1/8, 0 }
4075 * We can then compute the difference in effective weight by using:
4077 * dw_i = S * (s'_i - s_i) (3)
4079 * Where 'S' is the group weight as seen by its parent.
4081 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4082 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4083 * 4/7) times the weight of the group.
4085 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4087 struct sched_entity *se = tg->se[cpu];
4089 if (!tg->parent) /* the trivial, non-cgroup case */
4092 for_each_sched_entity(se) {
4098 * W = @wg + \Sum rw_j
4100 W = wg + calc_tg_weight(tg, se->my_q);
4105 w = se->my_q->load.weight + wl;
4108 * wl = S * s'_i; see (2)
4111 wl = (w * tg->shares) / W;
4116 * Per the above, wl is the new se->load.weight value; since
4117 * those are clipped to [MIN_SHARES, ...) do so now. See
4118 * calc_cfs_shares().
4120 if (wl < MIN_SHARES)
4124 * wl = dw_i = S * (s'_i - s_i); see (3)
4126 wl -= se->load.weight;
4129 * Recursively apply this logic to all parent groups to compute
4130 * the final effective load change on the root group. Since
4131 * only the @tg group gets extra weight, all parent groups can
4132 * only redistribute existing shares. @wl is the shift in shares
4133 * resulting from this level per the above.
4142 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4149 static int wake_wide(struct task_struct *p)
4151 int factor = this_cpu_read(sd_llc_size);
4154 * Yeah, it's the switching-frequency, could means many wakee or
4155 * rapidly switch, use factor here will just help to automatically
4156 * adjust the loose-degree, so bigger node will lead to more pull.
4158 if (p->wakee_flips > factor) {
4160 * wakee is somewhat hot, it needs certain amount of cpu
4161 * resource, so if waker is far more hot, prefer to leave
4164 if (current->wakee_flips > (factor * p->wakee_flips))
4171 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4173 s64 this_load, load;
4174 int idx, this_cpu, prev_cpu;
4175 unsigned long tl_per_task;
4176 struct task_group *tg;
4177 unsigned long weight;
4181 * If we wake multiple tasks be careful to not bounce
4182 * ourselves around too much.
4188 this_cpu = smp_processor_id();
4189 prev_cpu = task_cpu(p);
4190 load = source_load(prev_cpu, idx);
4191 this_load = target_load(this_cpu, idx);
4194 * If sync wakeup then subtract the (maximum possible)
4195 * effect of the currently running task from the load
4196 * of the current CPU:
4199 tg = task_group(current);
4200 weight = current->se.load.weight;
4202 this_load += effective_load(tg, this_cpu, -weight, -weight);
4203 load += effective_load(tg, prev_cpu, 0, -weight);
4207 weight = p->se.load.weight;
4210 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4211 * due to the sync cause above having dropped this_load to 0, we'll
4212 * always have an imbalance, but there's really nothing you can do
4213 * about that, so that's good too.
4215 * Otherwise check if either cpus are near enough in load to allow this
4216 * task to be woken on this_cpu.
4218 if (this_load > 0) {
4219 s64 this_eff_load, prev_eff_load;
4221 this_eff_load = 100;
4222 this_eff_load *= power_of(prev_cpu);
4223 this_eff_load *= this_load +
4224 effective_load(tg, this_cpu, weight, weight);
4226 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4227 prev_eff_load *= power_of(this_cpu);
4228 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4230 balanced = this_eff_load <= prev_eff_load;
4235 * If the currently running task will sleep within
4236 * a reasonable amount of time then attract this newly
4239 if (sync && balanced)
4242 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4243 tl_per_task = cpu_avg_load_per_task(this_cpu);
4246 (this_load <= load &&
4247 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4249 * This domain has SD_WAKE_AFFINE and
4250 * p is cache cold in this domain, and
4251 * there is no bad imbalance.
4253 schedstat_inc(sd, ttwu_move_affine);
4254 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4262 * find_idlest_group finds and returns the least busy CPU group within the
4265 static struct sched_group *
4266 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4267 int this_cpu, int sd_flag)
4269 struct sched_group *idlest = NULL, *group = sd->groups;
4270 unsigned long min_load = ULONG_MAX, this_load = 0;
4271 int load_idx = sd->forkexec_idx;
4272 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4274 if (sd_flag & SD_BALANCE_WAKE)
4275 load_idx = sd->wake_idx;
4278 unsigned long load, avg_load;
4282 /* Skip over this group if it has no CPUs allowed */
4283 if (!cpumask_intersects(sched_group_cpus(group),
4284 tsk_cpus_allowed(p)))
4287 local_group = cpumask_test_cpu(this_cpu,
4288 sched_group_cpus(group));
4290 /* Tally up the load of all CPUs in the group */
4293 for_each_cpu(i, sched_group_cpus(group)) {
4294 /* Bias balancing toward cpus of our domain */
4296 load = source_load(i, load_idx);
4298 load = target_load(i, load_idx);
4303 /* Adjust by relative CPU power of the group */
4304 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4307 this_load = avg_load;
4308 } else if (avg_load < min_load) {
4309 min_load = avg_load;
4312 } while (group = group->next, group != sd->groups);
4314 if (!idlest || 100*this_load < imbalance*min_load)
4320 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4323 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4325 unsigned long load, min_load = ULONG_MAX;
4329 /* Traverse only the allowed CPUs */
4330 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4331 load = weighted_cpuload(i);
4333 if (load < min_load || (load == min_load && i == this_cpu)) {
4343 * Try and locate an idle CPU in the sched_domain.
4345 static int select_idle_sibling(struct task_struct *p, int target)
4347 struct sched_domain *sd;
4348 struct sched_group *sg;
4349 int i = task_cpu(p);
4351 if (idle_cpu(target))
4355 * If the prevous cpu is cache affine and idle, don't be stupid.
4357 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4361 * Otherwise, iterate the domains and find an elegible idle cpu.
4363 sd = rcu_dereference(per_cpu(sd_llc, target));
4364 for_each_lower_domain(sd) {
4367 if (!cpumask_intersects(sched_group_cpus(sg),
4368 tsk_cpus_allowed(p)))
4371 for_each_cpu(i, sched_group_cpus(sg)) {
4372 if (i == target || !idle_cpu(i))
4376 target = cpumask_first_and(sched_group_cpus(sg),
4377 tsk_cpus_allowed(p));
4381 } while (sg != sd->groups);
4388 * sched_balance_self: balance the current task (running on cpu) in domains
4389 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4392 * Balance, ie. select the least loaded group.
4394 * Returns the target CPU number, or the same CPU if no balancing is needed.
4396 * preempt must be disabled.
4399 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4401 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4402 int cpu = smp_processor_id();
4404 int want_affine = 0;
4405 int sync = wake_flags & WF_SYNC;
4407 if (p->nr_cpus_allowed == 1)
4410 if (sd_flag & SD_BALANCE_WAKE) {
4411 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4417 for_each_domain(cpu, tmp) {
4418 if (!(tmp->flags & SD_LOAD_BALANCE))
4422 * If both cpu and prev_cpu are part of this domain,
4423 * cpu is a valid SD_WAKE_AFFINE target.
4425 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4426 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4431 if (tmp->flags & sd_flag)
4436 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4439 new_cpu = select_idle_sibling(p, prev_cpu);
4444 struct sched_group *group;
4447 if (!(sd->flags & sd_flag)) {
4452 group = find_idlest_group(sd, p, cpu, sd_flag);
4458 new_cpu = find_idlest_cpu(group, p, cpu);
4459 if (new_cpu == -1 || new_cpu == cpu) {
4460 /* Now try balancing at a lower domain level of cpu */
4465 /* Now try balancing at a lower domain level of new_cpu */
4467 weight = sd->span_weight;
4469 for_each_domain(cpu, tmp) {
4470 if (weight <= tmp->span_weight)
4472 if (tmp->flags & sd_flag)
4475 /* while loop will break here if sd == NULL */
4484 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4485 * cfs_rq_of(p) references at time of call are still valid and identify the
4486 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4487 * other assumptions, including the state of rq->lock, should be made.
4490 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4492 struct sched_entity *se = &p->se;
4493 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4496 * Load tracking: accumulate removed load so that it can be processed
4497 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4498 * to blocked load iff they have a positive decay-count. It can never
4499 * be negative here since on-rq tasks have decay-count == 0.
4501 if (se->avg.decay_count) {
4502 se->avg.decay_count = -__synchronize_entity_decay(se);
4503 atomic_long_add(se->avg.load_avg_contrib,
4504 &cfs_rq->removed_load);
4507 #endif /* CONFIG_SMP */
4509 static unsigned long
4510 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4512 unsigned long gran = sysctl_sched_wakeup_granularity;
4515 * Since its curr running now, convert the gran from real-time
4516 * to virtual-time in his units.
4518 * By using 'se' instead of 'curr' we penalize light tasks, so
4519 * they get preempted easier. That is, if 'se' < 'curr' then
4520 * the resulting gran will be larger, therefore penalizing the
4521 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4522 * be smaller, again penalizing the lighter task.
4524 * This is especially important for buddies when the leftmost
4525 * task is higher priority than the buddy.
4527 return calc_delta_fair(gran, se);
4531 * Should 'se' preempt 'curr'.
4545 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4547 s64 gran, vdiff = curr->vruntime - se->vruntime;
4552 gran = wakeup_gran(curr, se);
4559 static void set_last_buddy(struct sched_entity *se)
4561 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4564 for_each_sched_entity(se)
4565 cfs_rq_of(se)->last = se;
4568 static void set_next_buddy(struct sched_entity *se)
4570 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4573 for_each_sched_entity(se)
4574 cfs_rq_of(se)->next = se;
4577 static void set_skip_buddy(struct sched_entity *se)
4579 for_each_sched_entity(se)
4580 cfs_rq_of(se)->skip = se;
4584 * Preempt the current task with a newly woken task if needed:
4586 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4588 struct task_struct *curr = rq->curr;
4589 struct sched_entity *se = &curr->se, *pse = &p->se;
4590 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4591 int scale = cfs_rq->nr_running >= sched_nr_latency;
4592 int next_buddy_marked = 0;
4594 if (unlikely(se == pse))
4598 * This is possible from callers such as move_task(), in which we
4599 * unconditionally check_prempt_curr() after an enqueue (which may have
4600 * lead to a throttle). This both saves work and prevents false
4601 * next-buddy nomination below.
4603 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4606 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4607 set_next_buddy(pse);
4608 next_buddy_marked = 1;
4612 * We can come here with TIF_NEED_RESCHED already set from new task
4615 * Note: this also catches the edge-case of curr being in a throttled
4616 * group (e.g. via set_curr_task), since update_curr() (in the
4617 * enqueue of curr) will have resulted in resched being set. This
4618 * prevents us from potentially nominating it as a false LAST_BUDDY
4621 if (test_tsk_need_resched(curr))
4624 /* Idle tasks are by definition preempted by non-idle tasks. */
4625 if (unlikely(curr->policy == SCHED_IDLE) &&
4626 likely(p->policy != SCHED_IDLE))
4630 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4631 * is driven by the tick):
4633 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4636 find_matching_se(&se, &pse);
4637 update_curr(cfs_rq_of(se));
4639 if (wakeup_preempt_entity(se, pse) == 1) {
4641 * Bias pick_next to pick the sched entity that is
4642 * triggering this preemption.
4644 if (!next_buddy_marked)
4645 set_next_buddy(pse);
4654 * Only set the backward buddy when the current task is still
4655 * on the rq. This can happen when a wakeup gets interleaved
4656 * with schedule on the ->pre_schedule() or idle_balance()
4657 * point, either of which can * drop the rq lock.
4659 * Also, during early boot the idle thread is in the fair class,
4660 * for obvious reasons its a bad idea to schedule back to it.
4662 if (unlikely(!se->on_rq || curr == rq->idle))
4665 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4669 static struct task_struct *pick_next_task_fair(struct rq *rq)
4671 struct task_struct *p;
4672 struct cfs_rq *cfs_rq = &rq->cfs;
4673 struct sched_entity *se;
4675 if (!cfs_rq->nr_running)
4679 se = pick_next_entity(cfs_rq);
4680 set_next_entity(cfs_rq, se);
4681 cfs_rq = group_cfs_rq(se);
4685 if (hrtick_enabled(rq))
4686 hrtick_start_fair(rq, p);
4692 * Account for a descheduled task:
4694 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4696 struct sched_entity *se = &prev->se;
4697 struct cfs_rq *cfs_rq;
4699 for_each_sched_entity(se) {
4700 cfs_rq = cfs_rq_of(se);
4701 put_prev_entity(cfs_rq, se);
4706 * sched_yield() is very simple
4708 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4710 static void yield_task_fair(struct rq *rq)
4712 struct task_struct *curr = rq->curr;
4713 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4714 struct sched_entity *se = &curr->se;
4717 * Are we the only task in the tree?
4719 if (unlikely(rq->nr_running == 1))
4722 clear_buddies(cfs_rq, se);
4724 if (curr->policy != SCHED_BATCH) {
4725 update_rq_clock(rq);
4727 * Update run-time statistics of the 'current'.
4729 update_curr(cfs_rq);
4731 * Tell update_rq_clock() that we've just updated,
4732 * so we don't do microscopic update in schedule()
4733 * and double the fastpath cost.
4735 rq->skip_clock_update = 1;
4741 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4743 struct sched_entity *se = &p->se;
4745 /* throttled hierarchies are not runnable */
4746 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4749 /* Tell the scheduler that we'd really like pse to run next. */
4752 yield_task_fair(rq);
4758 /**************************************************
4759 * Fair scheduling class load-balancing methods.
4763 * The purpose of load-balancing is to achieve the same basic fairness the
4764 * per-cpu scheduler provides, namely provide a proportional amount of compute
4765 * time to each task. This is expressed in the following equation:
4767 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4769 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4770 * W_i,0 is defined as:
4772 * W_i,0 = \Sum_j w_i,j (2)
4774 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4775 * is derived from the nice value as per prio_to_weight[].
4777 * The weight average is an exponential decay average of the instantaneous
4780 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4782 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4783 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4784 * can also include other factors [XXX].
4786 * To achieve this balance we define a measure of imbalance which follows
4787 * directly from (1):
4789 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4791 * We them move tasks around to minimize the imbalance. In the continuous
4792 * function space it is obvious this converges, in the discrete case we get
4793 * a few fun cases generally called infeasible weight scenarios.
4796 * - infeasible weights;
4797 * - local vs global optima in the discrete case. ]
4802 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4803 * for all i,j solution, we create a tree of cpus that follows the hardware
4804 * topology where each level pairs two lower groups (or better). This results
4805 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4806 * tree to only the first of the previous level and we decrease the frequency
4807 * of load-balance at each level inv. proportional to the number of cpus in
4813 * \Sum { --- * --- * 2^i } = O(n) (5)
4815 * `- size of each group
4816 * | | `- number of cpus doing load-balance
4818 * `- sum over all levels
4820 * Coupled with a limit on how many tasks we can migrate every balance pass,
4821 * this makes (5) the runtime complexity of the balancer.
4823 * An important property here is that each CPU is still (indirectly) connected
4824 * to every other cpu in at most O(log n) steps:
4826 * The adjacency matrix of the resulting graph is given by:
4829 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4832 * And you'll find that:
4834 * A^(log_2 n)_i,j != 0 for all i,j (7)
4836 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4837 * The task movement gives a factor of O(m), giving a convergence complexity
4840 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4845 * In order to avoid CPUs going idle while there's still work to do, new idle
4846 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4847 * tree itself instead of relying on other CPUs to bring it work.
4849 * This adds some complexity to both (5) and (8) but it reduces the total idle
4857 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4860 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4865 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4867 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4869 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4872 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4873 * rewrite all of this once again.]
4876 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4878 enum fbq_type { regular, remote, all };
4880 #define LBF_ALL_PINNED 0x01
4881 #define LBF_NEED_BREAK 0x02
4882 #define LBF_DST_PINNED 0x04
4883 #define LBF_SOME_PINNED 0x08
4886 struct sched_domain *sd;
4894 struct cpumask *dst_grpmask;
4896 enum cpu_idle_type idle;
4898 /* The set of CPUs under consideration for load-balancing */
4899 struct cpumask *cpus;
4904 unsigned int loop_break;
4905 unsigned int loop_max;
4907 enum fbq_type fbq_type;
4911 * move_task - move a task from one runqueue to another runqueue.
4912 * Both runqueues must be locked.
4914 static void move_task(struct task_struct *p, struct lb_env *env)
4916 deactivate_task(env->src_rq, p, 0);
4917 set_task_cpu(p, env->dst_cpu);
4918 activate_task(env->dst_rq, p, 0);
4919 check_preempt_curr(env->dst_rq, p, 0);
4923 * Is this task likely cache-hot:
4926 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4930 if (p->sched_class != &fair_sched_class)
4933 if (unlikely(p->policy == SCHED_IDLE))
4937 * Buddy candidates are cache hot:
4939 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4940 (&p->se == cfs_rq_of(&p->se)->next ||
4941 &p->se == cfs_rq_of(&p->se)->last))
4944 if (sysctl_sched_migration_cost == -1)
4946 if (sysctl_sched_migration_cost == 0)
4949 delta = now - p->se.exec_start;
4951 return delta < (s64)sysctl_sched_migration_cost;
4954 #ifdef CONFIG_NUMA_BALANCING
4955 /* Returns true if the destination node has incurred more faults */
4956 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4958 int src_nid, dst_nid;
4960 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
4961 !(env->sd->flags & SD_NUMA)) {
4965 src_nid = cpu_to_node(env->src_cpu);
4966 dst_nid = cpu_to_node(env->dst_cpu);
4968 if (src_nid == dst_nid)
4971 /* Always encourage migration to the preferred node. */
4972 if (dst_nid == p->numa_preferred_nid)
4975 /* If both task and group weight improve, this move is a winner. */
4976 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4977 group_weight(p, dst_nid) > group_weight(p, src_nid))
4984 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4986 int src_nid, dst_nid;
4988 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4991 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
4994 src_nid = cpu_to_node(env->src_cpu);
4995 dst_nid = cpu_to_node(env->dst_cpu);
4997 if (src_nid == dst_nid)
5000 /* Migrating away from the preferred node is always bad. */
5001 if (src_nid == p->numa_preferred_nid)
5004 /* If either task or group weight get worse, don't do it. */
5005 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
5006 group_weight(p, dst_nid) < group_weight(p, src_nid))
5013 static inline bool migrate_improves_locality(struct task_struct *p,
5019 static inline bool migrate_degrades_locality(struct task_struct *p,
5027 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5030 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5032 int tsk_cache_hot = 0;
5034 * We do not migrate tasks that are:
5035 * 1) throttled_lb_pair, or
5036 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5037 * 3) running (obviously), or
5038 * 4) are cache-hot on their current CPU.
5040 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5043 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5046 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5048 env->flags |= LBF_SOME_PINNED;
5051 * Remember if this task can be migrated to any other cpu in
5052 * our sched_group. We may want to revisit it if we couldn't
5053 * meet load balance goals by pulling other tasks on src_cpu.
5055 * Also avoid computing new_dst_cpu if we have already computed
5056 * one in current iteration.
5058 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5061 /* Prevent to re-select dst_cpu via env's cpus */
5062 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5063 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5064 env->flags |= LBF_DST_PINNED;
5065 env->new_dst_cpu = cpu;
5073 /* Record that we found atleast one task that could run on dst_cpu */
5074 env->flags &= ~LBF_ALL_PINNED;
5076 if (task_running(env->src_rq, p)) {
5077 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5082 * Aggressive migration if:
5083 * 1) destination numa is preferred
5084 * 2) task is cache cold, or
5085 * 3) too many balance attempts have failed.
5087 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
5089 tsk_cache_hot = migrate_degrades_locality(p, env);
5091 if (migrate_improves_locality(p, env)) {
5092 #ifdef CONFIG_SCHEDSTATS
5093 if (tsk_cache_hot) {
5094 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5095 schedstat_inc(p, se.statistics.nr_forced_migrations);
5101 if (!tsk_cache_hot ||
5102 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5104 if (tsk_cache_hot) {
5105 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5106 schedstat_inc(p, se.statistics.nr_forced_migrations);
5112 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5117 * move_one_task tries to move exactly one task from busiest to this_rq, as
5118 * part of active balancing operations within "domain".
5119 * Returns 1 if successful and 0 otherwise.
5121 * Called with both runqueues locked.
5123 static int move_one_task(struct lb_env *env)
5125 struct task_struct *p, *n;
5127 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5128 if (!can_migrate_task(p, env))
5133 * Right now, this is only the second place move_task()
5134 * is called, so we can safely collect move_task()
5135 * stats here rather than inside move_task().
5137 schedstat_inc(env->sd, lb_gained[env->idle]);
5143 static const unsigned int sched_nr_migrate_break = 32;
5146 * move_tasks tries to move up to imbalance weighted load from busiest to
5147 * this_rq, as part of a balancing operation within domain "sd".
5148 * Returns 1 if successful and 0 otherwise.
5150 * Called with both runqueues locked.
5152 static int move_tasks(struct lb_env *env)
5154 struct list_head *tasks = &env->src_rq->cfs_tasks;
5155 struct task_struct *p;
5159 if (env->imbalance <= 0)
5162 while (!list_empty(tasks)) {
5163 p = list_first_entry(tasks, struct task_struct, se.group_node);
5166 /* We've more or less seen every task there is, call it quits */
5167 if (env->loop > env->loop_max)
5170 /* take a breather every nr_migrate tasks */
5171 if (env->loop > env->loop_break) {
5172 env->loop_break += sched_nr_migrate_break;
5173 env->flags |= LBF_NEED_BREAK;
5177 if (!can_migrate_task(p, env))
5180 load = task_h_load(p);
5182 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5185 if ((load / 2) > env->imbalance)
5190 env->imbalance -= load;
5192 #ifdef CONFIG_PREEMPT
5194 * NEWIDLE balancing is a source of latency, so preemptible
5195 * kernels will stop after the first task is pulled to minimize
5196 * the critical section.
5198 if (env->idle == CPU_NEWLY_IDLE)
5203 * We only want to steal up to the prescribed amount of
5206 if (env->imbalance <= 0)
5211 list_move_tail(&p->se.group_node, tasks);
5215 * Right now, this is one of only two places move_task() is called,
5216 * so we can safely collect move_task() stats here rather than
5217 * inside move_task().
5219 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5224 #ifdef CONFIG_FAIR_GROUP_SCHED
5226 * update tg->load_weight by folding this cpu's load_avg
5228 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5230 struct sched_entity *se = tg->se[cpu];
5231 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5233 /* throttled entities do not contribute to load */
5234 if (throttled_hierarchy(cfs_rq))
5237 update_cfs_rq_blocked_load(cfs_rq, 1);
5240 update_entity_load_avg(se, 1);
5242 * We pivot on our runnable average having decayed to zero for
5243 * list removal. This generally implies that all our children
5244 * have also been removed (modulo rounding error or bandwidth
5245 * control); however, such cases are rare and we can fix these
5248 * TODO: fix up out-of-order children on enqueue.
5250 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5251 list_del_leaf_cfs_rq(cfs_rq);
5253 struct rq *rq = rq_of(cfs_rq);
5254 update_rq_runnable_avg(rq, rq->nr_running);
5258 static void update_blocked_averages(int cpu)
5260 struct rq *rq = cpu_rq(cpu);
5261 struct cfs_rq *cfs_rq;
5262 unsigned long flags;
5264 raw_spin_lock_irqsave(&rq->lock, flags);
5265 update_rq_clock(rq);
5267 * Iterates the task_group tree in a bottom up fashion, see
5268 * list_add_leaf_cfs_rq() for details.
5270 for_each_leaf_cfs_rq(rq, cfs_rq) {
5272 * Note: We may want to consider periodically releasing
5273 * rq->lock about these updates so that creating many task
5274 * groups does not result in continually extending hold time.
5276 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5279 raw_spin_unlock_irqrestore(&rq->lock, flags);
5283 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5284 * This needs to be done in a top-down fashion because the load of a child
5285 * group is a fraction of its parents load.
5287 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5289 struct rq *rq = rq_of(cfs_rq);
5290 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5291 unsigned long now = jiffies;
5294 if (cfs_rq->last_h_load_update == now)
5297 cfs_rq->h_load_next = NULL;
5298 for_each_sched_entity(se) {
5299 cfs_rq = cfs_rq_of(se);
5300 cfs_rq->h_load_next = se;
5301 if (cfs_rq->last_h_load_update == now)
5306 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5307 cfs_rq->last_h_load_update = now;
5310 while ((se = cfs_rq->h_load_next) != NULL) {
5311 load = cfs_rq->h_load;
5312 load = div64_ul(load * se->avg.load_avg_contrib,
5313 cfs_rq->runnable_load_avg + 1);
5314 cfs_rq = group_cfs_rq(se);
5315 cfs_rq->h_load = load;
5316 cfs_rq->last_h_load_update = now;
5320 static unsigned long task_h_load(struct task_struct *p)
5322 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5324 update_cfs_rq_h_load(cfs_rq);
5325 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5326 cfs_rq->runnable_load_avg + 1);
5329 static inline void update_blocked_averages(int cpu)
5333 static unsigned long task_h_load(struct task_struct *p)
5335 return p->se.avg.load_avg_contrib;
5339 /********** Helpers for find_busiest_group ************************/
5341 * sg_lb_stats - stats of a sched_group required for load_balancing
5343 struct sg_lb_stats {
5344 unsigned long avg_load; /*Avg load across the CPUs of the group */
5345 unsigned long group_load; /* Total load over the CPUs of the group */
5346 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5347 unsigned long load_per_task;
5348 unsigned long group_power;
5349 unsigned int sum_nr_running; /* Nr tasks running in the group */
5350 unsigned int group_capacity;
5351 unsigned int idle_cpus;
5352 unsigned int group_weight;
5353 int group_imb; /* Is there an imbalance in the group ? */
5354 int group_has_capacity; /* Is there extra capacity in the group? */
5355 #ifdef CONFIG_NUMA_BALANCING
5356 unsigned int nr_numa_running;
5357 unsigned int nr_preferred_running;
5362 * sd_lb_stats - Structure to store the statistics of a sched_domain
5363 * during load balancing.
5365 struct sd_lb_stats {
5366 struct sched_group *busiest; /* Busiest group in this sd */
5367 struct sched_group *local; /* Local group in this sd */
5368 unsigned long total_load; /* Total load of all groups in sd */
5369 unsigned long total_pwr; /* Total power of all groups in sd */
5370 unsigned long avg_load; /* Average load across all groups in sd */
5372 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5373 struct sg_lb_stats local_stat; /* Statistics of the local group */
5376 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5379 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5380 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5381 * We must however clear busiest_stat::avg_load because
5382 * update_sd_pick_busiest() reads this before assignment.
5384 *sds = (struct sd_lb_stats){
5396 * get_sd_load_idx - Obtain the load index for a given sched domain.
5397 * @sd: The sched_domain whose load_idx is to be obtained.
5398 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5400 * Return: The load index.
5402 static inline int get_sd_load_idx(struct sched_domain *sd,
5403 enum cpu_idle_type idle)
5409 load_idx = sd->busy_idx;
5412 case CPU_NEWLY_IDLE:
5413 load_idx = sd->newidle_idx;
5416 load_idx = sd->idle_idx;
5423 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5425 return SCHED_POWER_SCALE;
5428 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5430 return default_scale_freq_power(sd, cpu);
5433 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5435 unsigned long weight = sd->span_weight;
5436 unsigned long smt_gain = sd->smt_gain;
5443 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5445 return default_scale_smt_power(sd, cpu);
5448 static unsigned long scale_rt_power(int cpu)
5450 struct rq *rq = cpu_rq(cpu);
5451 u64 total, available, age_stamp, avg;
5454 * Since we're reading these variables without serialization make sure
5455 * we read them once before doing sanity checks on them.
5457 age_stamp = ACCESS_ONCE(rq->age_stamp);
5458 avg = ACCESS_ONCE(rq->rt_avg);
5460 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5462 if (unlikely(total < avg)) {
5463 /* Ensures that power won't end up being negative */
5466 available = total - avg;
5469 if (unlikely((s64)total < SCHED_POWER_SCALE))
5470 total = SCHED_POWER_SCALE;
5472 total >>= SCHED_POWER_SHIFT;
5474 return div_u64(available, total);
5477 static void update_cpu_power(struct sched_domain *sd, int cpu)
5479 unsigned long weight = sd->span_weight;
5480 unsigned long power = SCHED_POWER_SCALE;
5481 struct sched_group *sdg = sd->groups;
5483 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5484 if (sched_feat(ARCH_POWER))
5485 power *= arch_scale_smt_power(sd, cpu);
5487 power *= default_scale_smt_power(sd, cpu);
5489 power >>= SCHED_POWER_SHIFT;
5492 sdg->sgp->power_orig = power;
5494 if (sched_feat(ARCH_POWER))
5495 power *= arch_scale_freq_power(sd, cpu);
5497 power *= default_scale_freq_power(sd, cpu);
5499 power >>= SCHED_POWER_SHIFT;
5501 power *= scale_rt_power(cpu);
5502 power >>= SCHED_POWER_SHIFT;
5507 cpu_rq(cpu)->cpu_power = power;
5508 sdg->sgp->power = power;
5511 void update_group_power(struct sched_domain *sd, int cpu)
5513 struct sched_domain *child = sd->child;
5514 struct sched_group *group, *sdg = sd->groups;
5515 unsigned long power, power_orig;
5516 unsigned long interval;
5518 interval = msecs_to_jiffies(sd->balance_interval);
5519 interval = clamp(interval, 1UL, max_load_balance_interval);
5520 sdg->sgp->next_update = jiffies + interval;
5523 update_cpu_power(sd, cpu);
5527 power_orig = power = 0;
5529 if (child->flags & SD_OVERLAP) {
5531 * SD_OVERLAP domains cannot assume that child groups
5532 * span the current group.
5535 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5536 struct sched_group_power *sgp;
5537 struct rq *rq = cpu_rq(cpu);
5540 * build_sched_domains() -> init_sched_groups_power()
5541 * gets here before we've attached the domains to the
5544 * Use power_of(), which is set irrespective of domains
5545 * in update_cpu_power().
5547 * This avoids power/power_orig from being 0 and
5548 * causing divide-by-zero issues on boot.
5550 * Runtime updates will correct power_orig.
5552 if (unlikely(!rq->sd)) {
5553 power_orig += power_of(cpu);
5554 power += power_of(cpu);
5558 sgp = rq->sd->groups->sgp;
5559 power_orig += sgp->power_orig;
5560 power += sgp->power;
5564 * !SD_OVERLAP domains can assume that child groups
5565 * span the current group.
5568 group = child->groups;
5570 power_orig += group->sgp->power_orig;
5571 power += group->sgp->power;
5572 group = group->next;
5573 } while (group != child->groups);
5576 sdg->sgp->power_orig = power_orig;
5577 sdg->sgp->power = power;
5581 * Try and fix up capacity for tiny siblings, this is needed when
5582 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5583 * which on its own isn't powerful enough.
5585 * See update_sd_pick_busiest() and check_asym_packing().
5588 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5591 * Only siblings can have significantly less than SCHED_POWER_SCALE
5593 if (!(sd->flags & SD_SHARE_CPUPOWER))
5597 * If ~90% of the cpu_power is still there, we're good.
5599 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5606 * Group imbalance indicates (and tries to solve) the problem where balancing
5607 * groups is inadequate due to tsk_cpus_allowed() constraints.
5609 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5610 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5613 * { 0 1 2 3 } { 4 5 6 7 }
5616 * If we were to balance group-wise we'd place two tasks in the first group and
5617 * two tasks in the second group. Clearly this is undesired as it will overload
5618 * cpu 3 and leave one of the cpus in the second group unused.
5620 * The current solution to this issue is detecting the skew in the first group
5621 * by noticing the lower domain failed to reach balance and had difficulty
5622 * moving tasks due to affinity constraints.
5624 * When this is so detected; this group becomes a candidate for busiest; see
5625 * update_sd_pick_busiest(). And calculate_imbalance() and
5626 * find_busiest_group() avoid some of the usual balance conditions to allow it
5627 * to create an effective group imbalance.
5629 * This is a somewhat tricky proposition since the next run might not find the
5630 * group imbalance and decide the groups need to be balanced again. A most
5631 * subtle and fragile situation.
5634 static inline int sg_imbalanced(struct sched_group *group)
5636 return group->sgp->imbalance;
5640 * Compute the group capacity.
5642 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5643 * first dividing out the smt factor and computing the actual number of cores
5644 * and limit power unit capacity with that.
5646 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5648 unsigned int capacity, smt, cpus;
5649 unsigned int power, power_orig;
5651 power = group->sgp->power;
5652 power_orig = group->sgp->power_orig;
5653 cpus = group->group_weight;
5655 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5656 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5657 capacity = cpus / smt; /* cores */
5659 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5661 capacity = fix_small_capacity(env->sd, group);
5667 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5668 * @env: The load balancing environment.
5669 * @group: sched_group whose statistics are to be updated.
5670 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5671 * @local_group: Does group contain this_cpu.
5672 * @sgs: variable to hold the statistics for this group.
5674 static inline void update_sg_lb_stats(struct lb_env *env,
5675 struct sched_group *group, int load_idx,
5676 int local_group, struct sg_lb_stats *sgs)
5681 memset(sgs, 0, sizeof(*sgs));
5683 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5684 struct rq *rq = cpu_rq(i);
5686 /* Bias balancing toward cpus of our domain */
5688 load = target_load(i, load_idx);
5690 load = source_load(i, load_idx);
5692 sgs->group_load += load;
5693 sgs->sum_nr_running += rq->nr_running;
5694 #ifdef CONFIG_NUMA_BALANCING
5695 sgs->nr_numa_running += rq->nr_numa_running;
5696 sgs->nr_preferred_running += rq->nr_preferred_running;
5698 sgs->sum_weighted_load += weighted_cpuload(i);
5703 /* Adjust by relative CPU power of the group */
5704 sgs->group_power = group->sgp->power;
5705 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5707 if (sgs->sum_nr_running)
5708 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5710 sgs->group_weight = group->group_weight;
5712 sgs->group_imb = sg_imbalanced(group);
5713 sgs->group_capacity = sg_capacity(env, group);
5715 if (sgs->group_capacity > sgs->sum_nr_running)
5716 sgs->group_has_capacity = 1;
5720 * update_sd_pick_busiest - return 1 on busiest group
5721 * @env: The load balancing environment.
5722 * @sds: sched_domain statistics
5723 * @sg: sched_group candidate to be checked for being the busiest
5724 * @sgs: sched_group statistics
5726 * Determine if @sg is a busier group than the previously selected
5729 * Return: %true if @sg is a busier group than the previously selected
5730 * busiest group. %false otherwise.
5732 static bool update_sd_pick_busiest(struct lb_env *env,
5733 struct sd_lb_stats *sds,
5734 struct sched_group *sg,
5735 struct sg_lb_stats *sgs)
5737 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5740 if (sgs->sum_nr_running > sgs->group_capacity)
5747 * ASYM_PACKING needs to move all the work to the lowest
5748 * numbered CPUs in the group, therefore mark all groups
5749 * higher than ourself as busy.
5751 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5752 env->dst_cpu < group_first_cpu(sg)) {
5756 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5763 #ifdef CONFIG_NUMA_BALANCING
5764 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5766 if (sgs->sum_nr_running > sgs->nr_numa_running)
5768 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5773 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5775 if (rq->nr_running > rq->nr_numa_running)
5777 if (rq->nr_running > rq->nr_preferred_running)
5782 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5787 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5791 #endif /* CONFIG_NUMA_BALANCING */
5794 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5795 * @env: The load balancing environment.
5796 * @sds: variable to hold the statistics for this sched_domain.
5798 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5800 struct sched_domain *child = env->sd->child;
5801 struct sched_group *sg = env->sd->groups;
5802 struct sg_lb_stats tmp_sgs;
5803 int load_idx, prefer_sibling = 0;
5805 if (child && child->flags & SD_PREFER_SIBLING)
5808 load_idx = get_sd_load_idx(env->sd, env->idle);
5811 struct sg_lb_stats *sgs = &tmp_sgs;
5814 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5817 sgs = &sds->local_stat;
5819 if (env->idle != CPU_NEWLY_IDLE ||
5820 time_after_eq(jiffies, sg->sgp->next_update))
5821 update_group_power(env->sd, env->dst_cpu);
5824 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5830 * In case the child domain prefers tasks go to siblings
5831 * first, lower the sg capacity to one so that we'll try
5832 * and move all the excess tasks away. We lower the capacity
5833 * of a group only if the local group has the capacity to fit
5834 * these excess tasks, i.e. nr_running < group_capacity. The
5835 * extra check prevents the case where you always pull from the
5836 * heaviest group when it is already under-utilized (possible
5837 * with a large weight task outweighs the tasks on the system).
5839 if (prefer_sibling && sds->local &&
5840 sds->local_stat.group_has_capacity)
5841 sgs->group_capacity = min(sgs->group_capacity, 1U);
5843 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5845 sds->busiest_stat = *sgs;
5849 /* Now, start updating sd_lb_stats */
5850 sds->total_load += sgs->group_load;
5851 sds->total_pwr += sgs->group_power;
5854 } while (sg != env->sd->groups);
5856 if (env->sd->flags & SD_NUMA)
5857 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5861 * check_asym_packing - Check to see if the group is packed into the
5864 * This is primarily intended to used at the sibling level. Some
5865 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5866 * case of POWER7, it can move to lower SMT modes only when higher
5867 * threads are idle. When in lower SMT modes, the threads will
5868 * perform better since they share less core resources. Hence when we
5869 * have idle threads, we want them to be the higher ones.
5871 * This packing function is run on idle threads. It checks to see if
5872 * the busiest CPU in this domain (core in the P7 case) has a higher
5873 * CPU number than the packing function is being run on. Here we are
5874 * assuming lower CPU number will be equivalent to lower a SMT thread
5877 * Return: 1 when packing is required and a task should be moved to
5878 * this CPU. The amount of the imbalance is returned in *imbalance.
5880 * @env: The load balancing environment.
5881 * @sds: Statistics of the sched_domain which is to be packed
5883 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5887 if (!(env->sd->flags & SD_ASYM_PACKING))
5893 busiest_cpu = group_first_cpu(sds->busiest);
5894 if (env->dst_cpu > busiest_cpu)
5897 env->imbalance = DIV_ROUND_CLOSEST(
5898 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5905 * fix_small_imbalance - Calculate the minor imbalance that exists
5906 * amongst the groups of a sched_domain, during
5908 * @env: The load balancing environment.
5909 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5912 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5914 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5915 unsigned int imbn = 2;
5916 unsigned long scaled_busy_load_per_task;
5917 struct sg_lb_stats *local, *busiest;
5919 local = &sds->local_stat;
5920 busiest = &sds->busiest_stat;
5922 if (!local->sum_nr_running)
5923 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5924 else if (busiest->load_per_task > local->load_per_task)
5927 scaled_busy_load_per_task =
5928 (busiest->load_per_task * SCHED_POWER_SCALE) /
5929 busiest->group_power;
5931 if (busiest->avg_load + scaled_busy_load_per_task >=
5932 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5933 env->imbalance = busiest->load_per_task;
5938 * OK, we don't have enough imbalance to justify moving tasks,
5939 * however we may be able to increase total CPU power used by
5943 pwr_now += busiest->group_power *
5944 min(busiest->load_per_task, busiest->avg_load);
5945 pwr_now += local->group_power *
5946 min(local->load_per_task, local->avg_load);
5947 pwr_now /= SCHED_POWER_SCALE;
5949 /* Amount of load we'd subtract */
5950 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5951 busiest->group_power;
5952 if (busiest->avg_load > tmp) {
5953 pwr_move += busiest->group_power *
5954 min(busiest->load_per_task,
5955 busiest->avg_load - tmp);
5958 /* Amount of load we'd add */
5959 if (busiest->avg_load * busiest->group_power <
5960 busiest->load_per_task * SCHED_POWER_SCALE) {
5961 tmp = (busiest->avg_load * busiest->group_power) /
5964 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5967 pwr_move += local->group_power *
5968 min(local->load_per_task, local->avg_load + tmp);
5969 pwr_move /= SCHED_POWER_SCALE;
5971 /* Move if we gain throughput */
5972 if (pwr_move > pwr_now)
5973 env->imbalance = busiest->load_per_task;
5977 * calculate_imbalance - Calculate the amount of imbalance present within the
5978 * groups of a given sched_domain during load balance.
5979 * @env: load balance environment
5980 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5982 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5984 unsigned long max_pull, load_above_capacity = ~0UL;
5985 struct sg_lb_stats *local, *busiest;
5987 local = &sds->local_stat;
5988 busiest = &sds->busiest_stat;
5990 if (busiest->group_imb) {
5992 * In the group_imb case we cannot rely on group-wide averages
5993 * to ensure cpu-load equilibrium, look at wider averages. XXX
5995 busiest->load_per_task =
5996 min(busiest->load_per_task, sds->avg_load);
6000 * In the presence of smp nice balancing, certain scenarios can have
6001 * max load less than avg load(as we skip the groups at or below
6002 * its cpu_power, while calculating max_load..)
6004 if (busiest->avg_load <= sds->avg_load ||
6005 local->avg_load >= sds->avg_load) {
6007 return fix_small_imbalance(env, sds);
6010 if (!busiest->group_imb) {
6012 * Don't want to pull so many tasks that a group would go idle.
6013 * Except of course for the group_imb case, since then we might
6014 * have to drop below capacity to reach cpu-load equilibrium.
6016 load_above_capacity =
6017 (busiest->sum_nr_running - busiest->group_capacity);
6019 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
6020 load_above_capacity /= busiest->group_power;
6024 * We're trying to get all the cpus to the average_load, so we don't
6025 * want to push ourselves above the average load, nor do we wish to
6026 * reduce the max loaded cpu below the average load. At the same time,
6027 * we also don't want to reduce the group load below the group capacity
6028 * (so that we can implement power-savings policies etc). Thus we look
6029 * for the minimum possible imbalance.
6031 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6033 /* How much load to actually move to equalise the imbalance */
6034 env->imbalance = min(
6035 max_pull * busiest->group_power,
6036 (sds->avg_load - local->avg_load) * local->group_power
6037 ) / SCHED_POWER_SCALE;
6040 * if *imbalance is less than the average load per runnable task
6041 * there is no guarantee that any tasks will be moved so we'll have
6042 * a think about bumping its value to force at least one task to be
6045 if (env->imbalance < busiest->load_per_task)
6046 return fix_small_imbalance(env, sds);
6049 /******* find_busiest_group() helpers end here *********************/
6052 * find_busiest_group - Returns the busiest group within the sched_domain
6053 * if there is an imbalance. If there isn't an imbalance, and
6054 * the user has opted for power-savings, it returns a group whose
6055 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6056 * such a group exists.
6058 * Also calculates the amount of weighted load which should be moved
6059 * to restore balance.
6061 * @env: The load balancing environment.
6063 * Return: - The busiest group if imbalance exists.
6064 * - If no imbalance and user has opted for power-savings balance,
6065 * return the least loaded group whose CPUs can be
6066 * put to idle by rebalancing its tasks onto our group.
6068 static struct sched_group *find_busiest_group(struct lb_env *env)
6070 struct sg_lb_stats *local, *busiest;
6071 struct sd_lb_stats sds;
6073 init_sd_lb_stats(&sds);
6076 * Compute the various statistics relavent for load balancing at
6079 update_sd_lb_stats(env, &sds);
6080 local = &sds.local_stat;
6081 busiest = &sds.busiest_stat;
6083 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6084 check_asym_packing(env, &sds))
6087 /* There is no busy sibling group to pull tasks from */
6088 if (!sds.busiest || busiest->sum_nr_running == 0)
6091 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
6094 * If the busiest group is imbalanced the below checks don't
6095 * work because they assume all things are equal, which typically
6096 * isn't true due to cpus_allowed constraints and the like.
6098 if (busiest->group_imb)
6101 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6102 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
6103 !busiest->group_has_capacity)
6107 * If the local group is more busy than the selected busiest group
6108 * don't try and pull any tasks.
6110 if (local->avg_load >= busiest->avg_load)
6114 * Don't pull any tasks if this group is already above the domain
6117 if (local->avg_load >= sds.avg_load)
6120 if (env->idle == CPU_IDLE) {
6122 * This cpu is idle. If the busiest group load doesn't
6123 * have more tasks than the number of available cpu's and
6124 * there is no imbalance between this and busiest group
6125 * wrt to idle cpu's, it is balanced.
6127 if ((local->idle_cpus < busiest->idle_cpus) &&
6128 busiest->sum_nr_running <= busiest->group_weight)
6132 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6133 * imbalance_pct to be conservative.
6135 if (100 * busiest->avg_load <=
6136 env->sd->imbalance_pct * local->avg_load)
6141 /* Looks like there is an imbalance. Compute it */
6142 calculate_imbalance(env, &sds);
6151 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6153 static struct rq *find_busiest_queue(struct lb_env *env,
6154 struct sched_group *group)
6156 struct rq *busiest = NULL, *rq;
6157 unsigned long busiest_load = 0, busiest_power = 1;
6160 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6161 unsigned long power, capacity, wl;
6165 rt = fbq_classify_rq(rq);
6168 * We classify groups/runqueues into three groups:
6169 * - regular: there are !numa tasks
6170 * - remote: there are numa tasks that run on the 'wrong' node
6171 * - all: there is no distinction
6173 * In order to avoid migrating ideally placed numa tasks,
6174 * ignore those when there's better options.
6176 * If we ignore the actual busiest queue to migrate another
6177 * task, the next balance pass can still reduce the busiest
6178 * queue by moving tasks around inside the node.
6180 * If we cannot move enough load due to this classification
6181 * the next pass will adjust the group classification and
6182 * allow migration of more tasks.
6184 * Both cases only affect the total convergence complexity.
6186 if (rt > env->fbq_type)
6189 power = power_of(i);
6190 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
6192 capacity = fix_small_capacity(env->sd, group);
6194 wl = weighted_cpuload(i);
6197 * When comparing with imbalance, use weighted_cpuload()
6198 * which is not scaled with the cpu power.
6200 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6204 * For the load comparisons with the other cpu's, consider
6205 * the weighted_cpuload() scaled with the cpu power, so that
6206 * the load can be moved away from the cpu that is potentially
6207 * running at a lower capacity.
6209 * Thus we're looking for max(wl_i / power_i), crosswise
6210 * multiplication to rid ourselves of the division works out
6211 * to: wl_i * power_j > wl_j * power_i; where j is our
6214 if (wl * busiest_power > busiest_load * power) {
6216 busiest_power = power;
6225 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6226 * so long as it is large enough.
6228 #define MAX_PINNED_INTERVAL 512
6230 /* Working cpumask for load_balance and load_balance_newidle. */
6231 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6233 static int need_active_balance(struct lb_env *env)
6235 struct sched_domain *sd = env->sd;
6237 if (env->idle == CPU_NEWLY_IDLE) {
6240 * ASYM_PACKING needs to force migrate tasks from busy but
6241 * higher numbered CPUs in order to pack all tasks in the
6242 * lowest numbered CPUs.
6244 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6248 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6251 static int active_load_balance_cpu_stop(void *data);
6253 static int should_we_balance(struct lb_env *env)
6255 struct sched_group *sg = env->sd->groups;
6256 struct cpumask *sg_cpus, *sg_mask;
6257 int cpu, balance_cpu = -1;
6260 * In the newly idle case, we will allow all the cpu's
6261 * to do the newly idle load balance.
6263 if (env->idle == CPU_NEWLY_IDLE)
6266 sg_cpus = sched_group_cpus(sg);
6267 sg_mask = sched_group_mask(sg);
6268 /* Try to find first idle cpu */
6269 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6270 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6277 if (balance_cpu == -1)
6278 balance_cpu = group_balance_cpu(sg);
6281 * First idle cpu or the first cpu(busiest) in this sched group
6282 * is eligible for doing load balancing at this and above domains.
6284 return balance_cpu == env->dst_cpu;
6288 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6289 * tasks if there is an imbalance.
6291 static int load_balance(int this_cpu, struct rq *this_rq,
6292 struct sched_domain *sd, enum cpu_idle_type idle,
6293 int *continue_balancing)
6295 int ld_moved, cur_ld_moved, active_balance = 0;
6296 struct sched_domain *sd_parent = sd->parent;
6297 struct sched_group *group;
6299 unsigned long flags;
6300 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6302 struct lb_env env = {
6304 .dst_cpu = this_cpu,
6306 .dst_grpmask = sched_group_cpus(sd->groups),
6308 .loop_break = sched_nr_migrate_break,
6314 * For NEWLY_IDLE load_balancing, we don't need to consider
6315 * other cpus in our group
6317 if (idle == CPU_NEWLY_IDLE)
6318 env.dst_grpmask = NULL;
6320 cpumask_copy(cpus, cpu_active_mask);
6322 schedstat_inc(sd, lb_count[idle]);
6325 if (!should_we_balance(&env)) {
6326 *continue_balancing = 0;
6330 group = find_busiest_group(&env);
6332 schedstat_inc(sd, lb_nobusyg[idle]);
6336 busiest = find_busiest_queue(&env, group);
6338 schedstat_inc(sd, lb_nobusyq[idle]);
6342 BUG_ON(busiest == env.dst_rq);
6344 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6347 if (busiest->nr_running > 1) {
6349 * Attempt to move tasks. If find_busiest_group has found
6350 * an imbalance but busiest->nr_running <= 1, the group is
6351 * still unbalanced. ld_moved simply stays zero, so it is
6352 * correctly treated as an imbalance.
6354 env.flags |= LBF_ALL_PINNED;
6355 env.src_cpu = busiest->cpu;
6356 env.src_rq = busiest;
6357 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6360 local_irq_save(flags);
6361 double_rq_lock(env.dst_rq, busiest);
6364 * cur_ld_moved - load moved in current iteration
6365 * ld_moved - cumulative load moved across iterations
6367 cur_ld_moved = move_tasks(&env);
6368 ld_moved += cur_ld_moved;
6369 double_rq_unlock(env.dst_rq, busiest);
6370 local_irq_restore(flags);
6373 * some other cpu did the load balance for us.
6375 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6376 resched_cpu(env.dst_cpu);
6378 if (env.flags & LBF_NEED_BREAK) {
6379 env.flags &= ~LBF_NEED_BREAK;
6384 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6385 * us and move them to an alternate dst_cpu in our sched_group
6386 * where they can run. The upper limit on how many times we
6387 * iterate on same src_cpu is dependent on number of cpus in our
6390 * This changes load balance semantics a bit on who can move
6391 * load to a given_cpu. In addition to the given_cpu itself
6392 * (or a ilb_cpu acting on its behalf where given_cpu is
6393 * nohz-idle), we now have balance_cpu in a position to move
6394 * load to given_cpu. In rare situations, this may cause
6395 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6396 * _independently_ and at _same_ time to move some load to
6397 * given_cpu) causing exceess load to be moved to given_cpu.
6398 * This however should not happen so much in practice and
6399 * moreover subsequent load balance cycles should correct the
6400 * excess load moved.
6402 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6404 /* Prevent to re-select dst_cpu via env's cpus */
6405 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6407 env.dst_rq = cpu_rq(env.new_dst_cpu);
6408 env.dst_cpu = env.new_dst_cpu;
6409 env.flags &= ~LBF_DST_PINNED;
6411 env.loop_break = sched_nr_migrate_break;
6414 * Go back to "more_balance" rather than "redo" since we
6415 * need to continue with same src_cpu.
6421 * We failed to reach balance because of affinity.
6424 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6426 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6427 *group_imbalance = 1;
6428 } else if (*group_imbalance)
6429 *group_imbalance = 0;
6432 /* All tasks on this runqueue were pinned by CPU affinity */
6433 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6434 cpumask_clear_cpu(cpu_of(busiest), cpus);
6435 if (!cpumask_empty(cpus)) {
6437 env.loop_break = sched_nr_migrate_break;
6445 schedstat_inc(sd, lb_failed[idle]);
6447 * Increment the failure counter only on periodic balance.
6448 * We do not want newidle balance, which can be very
6449 * frequent, pollute the failure counter causing
6450 * excessive cache_hot migrations and active balances.
6452 if (idle != CPU_NEWLY_IDLE)
6453 sd->nr_balance_failed++;
6455 if (need_active_balance(&env)) {
6456 raw_spin_lock_irqsave(&busiest->lock, flags);
6458 /* don't kick the active_load_balance_cpu_stop,
6459 * if the curr task on busiest cpu can't be
6462 if (!cpumask_test_cpu(this_cpu,
6463 tsk_cpus_allowed(busiest->curr))) {
6464 raw_spin_unlock_irqrestore(&busiest->lock,
6466 env.flags |= LBF_ALL_PINNED;
6467 goto out_one_pinned;
6471 * ->active_balance synchronizes accesses to
6472 * ->active_balance_work. Once set, it's cleared
6473 * only after active load balance is finished.
6475 if (!busiest->active_balance) {
6476 busiest->active_balance = 1;
6477 busiest->push_cpu = this_cpu;
6480 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6482 if (active_balance) {
6483 stop_one_cpu_nowait(cpu_of(busiest),
6484 active_load_balance_cpu_stop, busiest,
6485 &busiest->active_balance_work);
6489 * We've kicked active balancing, reset the failure
6492 sd->nr_balance_failed = sd->cache_nice_tries+1;
6495 sd->nr_balance_failed = 0;
6497 if (likely(!active_balance)) {
6498 /* We were unbalanced, so reset the balancing interval */
6499 sd->balance_interval = sd->min_interval;
6502 * If we've begun active balancing, start to back off. This
6503 * case may not be covered by the all_pinned logic if there
6504 * is only 1 task on the busy runqueue (because we don't call
6507 if (sd->balance_interval < sd->max_interval)
6508 sd->balance_interval *= 2;
6514 schedstat_inc(sd, lb_balanced[idle]);
6516 sd->nr_balance_failed = 0;
6519 /* tune up the balancing interval */
6520 if (((env.flags & LBF_ALL_PINNED) &&
6521 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6522 (sd->balance_interval < sd->max_interval))
6523 sd->balance_interval *= 2;
6531 * idle_balance is called by schedule() if this_cpu is about to become
6532 * idle. Attempts to pull tasks from other CPUs.
6534 int idle_balance(struct rq *this_rq)
6536 struct sched_domain *sd;
6537 int pulled_task = 0;
6538 unsigned long next_balance = jiffies + HZ;
6540 int this_cpu = this_rq->cpu;
6542 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6546 * Drop the rq->lock, but keep IRQ/preempt disabled.
6548 raw_spin_unlock(&this_rq->lock);
6550 update_blocked_averages(this_cpu);
6552 for_each_domain(this_cpu, sd) {
6553 unsigned long interval;
6554 int continue_balancing = 1;
6555 u64 t0, domain_cost;
6557 if (!(sd->flags & SD_LOAD_BALANCE))
6560 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6563 if (sd->flags & SD_BALANCE_NEWIDLE) {
6564 t0 = sched_clock_cpu(this_cpu);
6566 /* If we've pulled tasks over stop searching: */
6567 pulled_task = load_balance(this_cpu, this_rq,
6569 &continue_balancing);
6571 domain_cost = sched_clock_cpu(this_cpu) - t0;
6572 if (domain_cost > sd->max_newidle_lb_cost)
6573 sd->max_newidle_lb_cost = domain_cost;
6575 curr_cost += domain_cost;
6578 interval = msecs_to_jiffies(sd->balance_interval);
6579 if (time_after(next_balance, sd->last_balance + interval))
6580 next_balance = sd->last_balance + interval;
6586 raw_spin_lock(&this_rq->lock);
6589 * While browsing the domains, we released the rq lock.
6590 * A task could have be enqueued in the meantime
6592 if (this_rq->nr_running && !pulled_task)
6595 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6597 * We are going idle. next_balance may be set based on
6598 * a busy processor. So reset next_balance.
6600 this_rq->next_balance = next_balance;
6603 if (curr_cost > this_rq->max_idle_balance_cost)
6604 this_rq->max_idle_balance_cost = curr_cost;
6610 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6611 * running tasks off the busiest CPU onto idle CPUs. It requires at
6612 * least 1 task to be running on each physical CPU where possible, and
6613 * avoids physical / logical imbalances.
6615 static int active_load_balance_cpu_stop(void *data)
6617 struct rq *busiest_rq = data;
6618 int busiest_cpu = cpu_of(busiest_rq);
6619 int target_cpu = busiest_rq->push_cpu;
6620 struct rq *target_rq = cpu_rq(target_cpu);
6621 struct sched_domain *sd;
6623 raw_spin_lock_irq(&busiest_rq->lock);
6625 /* make sure the requested cpu hasn't gone down in the meantime */
6626 if (unlikely(busiest_cpu != smp_processor_id() ||
6627 !busiest_rq->active_balance))
6630 /* Is there any task to move? */
6631 if (busiest_rq->nr_running <= 1)
6635 * This condition is "impossible", if it occurs
6636 * we need to fix it. Originally reported by
6637 * Bjorn Helgaas on a 128-cpu setup.
6639 BUG_ON(busiest_rq == target_rq);
6641 /* move a task from busiest_rq to target_rq */
6642 double_lock_balance(busiest_rq, target_rq);
6644 /* Search for an sd spanning us and the target CPU. */
6646 for_each_domain(target_cpu, sd) {
6647 if ((sd->flags & SD_LOAD_BALANCE) &&
6648 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6653 struct lb_env env = {
6655 .dst_cpu = target_cpu,
6656 .dst_rq = target_rq,
6657 .src_cpu = busiest_rq->cpu,
6658 .src_rq = busiest_rq,
6662 schedstat_inc(sd, alb_count);
6664 if (move_one_task(&env))
6665 schedstat_inc(sd, alb_pushed);
6667 schedstat_inc(sd, alb_failed);
6670 double_unlock_balance(busiest_rq, target_rq);
6672 busiest_rq->active_balance = 0;
6673 raw_spin_unlock_irq(&busiest_rq->lock);
6677 #ifdef CONFIG_NO_HZ_COMMON
6679 * idle load balancing details
6680 * - When one of the busy CPUs notice that there may be an idle rebalancing
6681 * needed, they will kick the idle load balancer, which then does idle
6682 * load balancing for all the idle CPUs.
6685 cpumask_var_t idle_cpus_mask;
6687 unsigned long next_balance; /* in jiffy units */
6688 } nohz ____cacheline_aligned;
6690 static inline int find_new_ilb(void)
6692 int ilb = cpumask_first(nohz.idle_cpus_mask);
6694 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6701 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6702 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6703 * CPU (if there is one).
6705 static void nohz_balancer_kick(void)
6709 nohz.next_balance++;
6711 ilb_cpu = find_new_ilb();
6713 if (ilb_cpu >= nr_cpu_ids)
6716 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6719 * Use smp_send_reschedule() instead of resched_cpu().
6720 * This way we generate a sched IPI on the target cpu which
6721 * is idle. And the softirq performing nohz idle load balance
6722 * will be run before returning from the IPI.
6724 smp_send_reschedule(ilb_cpu);
6728 static inline void nohz_balance_exit_idle(int cpu)
6730 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6731 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6732 atomic_dec(&nohz.nr_cpus);
6733 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6737 static inline void set_cpu_sd_state_busy(void)
6739 struct sched_domain *sd;
6740 int cpu = smp_processor_id();
6743 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6745 if (!sd || !sd->nohz_idle)
6749 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6754 void set_cpu_sd_state_idle(void)
6756 struct sched_domain *sd;
6757 int cpu = smp_processor_id();
6760 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6762 if (!sd || sd->nohz_idle)
6766 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6772 * This routine will record that the cpu is going idle with tick stopped.
6773 * This info will be used in performing idle load balancing in the future.
6775 void nohz_balance_enter_idle(int cpu)
6778 * If this cpu is going down, then nothing needs to be done.
6780 if (!cpu_active(cpu))
6783 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6786 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6787 atomic_inc(&nohz.nr_cpus);
6788 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6791 static int sched_ilb_notifier(struct notifier_block *nfb,
6792 unsigned long action, void *hcpu)
6794 switch (action & ~CPU_TASKS_FROZEN) {
6796 nohz_balance_exit_idle(smp_processor_id());
6804 static DEFINE_SPINLOCK(balancing);
6807 * Scale the max load_balance interval with the number of CPUs in the system.
6808 * This trades load-balance latency on larger machines for less cross talk.
6810 void update_max_interval(void)
6812 max_load_balance_interval = HZ*num_online_cpus()/10;
6816 * It checks each scheduling domain to see if it is due to be balanced,
6817 * and initiates a balancing operation if so.
6819 * Balancing parameters are set up in init_sched_domains.
6821 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
6823 int continue_balancing = 1;
6825 unsigned long interval;
6826 struct sched_domain *sd;
6827 /* Earliest time when we have to do rebalance again */
6828 unsigned long next_balance = jiffies + 60*HZ;
6829 int update_next_balance = 0;
6830 int need_serialize, need_decay = 0;
6833 update_blocked_averages(cpu);
6836 for_each_domain(cpu, sd) {
6838 * Decay the newidle max times here because this is a regular
6839 * visit to all the domains. Decay ~1% per second.
6841 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6842 sd->max_newidle_lb_cost =
6843 (sd->max_newidle_lb_cost * 253) / 256;
6844 sd->next_decay_max_lb_cost = jiffies + HZ;
6847 max_cost += sd->max_newidle_lb_cost;
6849 if (!(sd->flags & SD_LOAD_BALANCE))
6853 * Stop the load balance at this level. There is another
6854 * CPU in our sched group which is doing load balancing more
6857 if (!continue_balancing) {
6863 interval = sd->balance_interval;
6864 if (idle != CPU_IDLE)
6865 interval *= sd->busy_factor;
6867 /* scale ms to jiffies */
6868 interval = msecs_to_jiffies(interval);
6869 interval = clamp(interval, 1UL, max_load_balance_interval);
6871 need_serialize = sd->flags & SD_SERIALIZE;
6873 if (need_serialize) {
6874 if (!spin_trylock(&balancing))
6878 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6879 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6881 * The LBF_DST_PINNED logic could have changed
6882 * env->dst_cpu, so we can't know our idle
6883 * state even if we migrated tasks. Update it.
6885 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6887 sd->last_balance = jiffies;
6890 spin_unlock(&balancing);
6892 if (time_after(next_balance, sd->last_balance + interval)) {
6893 next_balance = sd->last_balance + interval;
6894 update_next_balance = 1;
6899 * Ensure the rq-wide value also decays but keep it at a
6900 * reasonable floor to avoid funnies with rq->avg_idle.
6902 rq->max_idle_balance_cost =
6903 max((u64)sysctl_sched_migration_cost, max_cost);
6908 * next_balance will be updated only when there is a need.
6909 * When the cpu is attached to null domain for ex, it will not be
6912 if (likely(update_next_balance))
6913 rq->next_balance = next_balance;
6916 #ifdef CONFIG_NO_HZ_COMMON
6918 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6919 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6921 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
6923 int this_cpu = this_rq->cpu;
6927 if (idle != CPU_IDLE ||
6928 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6931 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6932 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6936 * If this cpu gets work to do, stop the load balancing
6937 * work being done for other cpus. Next load
6938 * balancing owner will pick it up.
6943 rq = cpu_rq(balance_cpu);
6945 raw_spin_lock_irq(&rq->lock);
6946 update_rq_clock(rq);
6947 update_idle_cpu_load(rq);
6948 raw_spin_unlock_irq(&rq->lock);
6950 rebalance_domains(rq, CPU_IDLE);
6952 if (time_after(this_rq->next_balance, rq->next_balance))
6953 this_rq->next_balance = rq->next_balance;
6955 nohz.next_balance = this_rq->next_balance;
6957 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6961 * Current heuristic for kicking the idle load balancer in the presence
6962 * of an idle cpu is the system.
6963 * - This rq has more than one task.
6964 * - At any scheduler domain level, this cpu's scheduler group has multiple
6965 * busy cpu's exceeding the group's power.
6966 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6967 * domain span are idle.
6969 static inline int nohz_kick_needed(struct rq *rq)
6971 unsigned long now = jiffies;
6972 struct sched_domain *sd;
6973 struct sched_group_power *sgp;
6974 int nr_busy, cpu = rq->cpu;
6976 if (unlikely(rq->idle_balance))
6980 * We may be recently in ticked or tickless idle mode. At the first
6981 * busy tick after returning from idle, we will update the busy stats.
6983 set_cpu_sd_state_busy();
6984 nohz_balance_exit_idle(cpu);
6987 * None are in tickless mode and hence no need for NOHZ idle load
6990 if (likely(!atomic_read(&nohz.nr_cpus)))
6993 if (time_before(now, nohz.next_balance))
6996 if (rq->nr_running >= 2)
7000 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7003 sgp = sd->groups->sgp;
7004 nr_busy = atomic_read(&sgp->nr_busy_cpus);
7007 goto need_kick_unlock;
7010 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7012 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7013 sched_domain_span(sd)) < cpu))
7014 goto need_kick_unlock;
7025 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7029 * run_rebalance_domains is triggered when needed from the scheduler tick.
7030 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7032 static void run_rebalance_domains(struct softirq_action *h)
7034 struct rq *this_rq = this_rq();
7035 enum cpu_idle_type idle = this_rq->idle_balance ?
7036 CPU_IDLE : CPU_NOT_IDLE;
7038 rebalance_domains(this_rq, idle);
7041 * If this cpu has a pending nohz_balance_kick, then do the
7042 * balancing on behalf of the other idle cpus whose ticks are
7045 nohz_idle_balance(this_rq, idle);
7048 static inline int on_null_domain(struct rq *rq)
7050 return !rcu_dereference_sched(rq->sd);
7054 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7056 void trigger_load_balance(struct rq *rq)
7058 /* Don't need to rebalance while attached to NULL domain */
7059 if (unlikely(on_null_domain(rq)))
7062 if (time_after_eq(jiffies, rq->next_balance))
7063 raise_softirq(SCHED_SOFTIRQ);
7064 #ifdef CONFIG_NO_HZ_COMMON
7065 if (nohz_kick_needed(rq))
7066 nohz_balancer_kick();
7070 static void rq_online_fair(struct rq *rq)
7075 static void rq_offline_fair(struct rq *rq)
7079 /* Ensure any throttled groups are reachable by pick_next_task */
7080 unthrottle_offline_cfs_rqs(rq);
7083 #endif /* CONFIG_SMP */
7086 * scheduler tick hitting a task of our scheduling class:
7088 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7090 struct cfs_rq *cfs_rq;
7091 struct sched_entity *se = &curr->se;
7093 for_each_sched_entity(se) {
7094 cfs_rq = cfs_rq_of(se);
7095 entity_tick(cfs_rq, se, queued);
7098 if (numabalancing_enabled)
7099 task_tick_numa(rq, curr);
7101 update_rq_runnable_avg(rq, 1);
7105 * called on fork with the child task as argument from the parent's context
7106 * - child not yet on the tasklist
7107 * - preemption disabled
7109 static void task_fork_fair(struct task_struct *p)
7111 struct cfs_rq *cfs_rq;
7112 struct sched_entity *se = &p->se, *curr;
7113 int this_cpu = smp_processor_id();
7114 struct rq *rq = this_rq();
7115 unsigned long flags;
7117 raw_spin_lock_irqsave(&rq->lock, flags);
7119 update_rq_clock(rq);
7121 cfs_rq = task_cfs_rq(current);
7122 curr = cfs_rq->curr;
7125 * Not only the cpu but also the task_group of the parent might have
7126 * been changed after parent->se.parent,cfs_rq were copied to
7127 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7128 * of child point to valid ones.
7131 __set_task_cpu(p, this_cpu);
7134 update_curr(cfs_rq);
7137 se->vruntime = curr->vruntime;
7138 place_entity(cfs_rq, se, 1);
7140 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7142 * Upon rescheduling, sched_class::put_prev_task() will place
7143 * 'current' within the tree based on its new key value.
7145 swap(curr->vruntime, se->vruntime);
7146 resched_task(rq->curr);
7149 se->vruntime -= cfs_rq->min_vruntime;
7151 raw_spin_unlock_irqrestore(&rq->lock, flags);
7155 * Priority of the task has changed. Check to see if we preempt
7159 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7165 * Reschedule if we are currently running on this runqueue and
7166 * our priority decreased, or if we are not currently running on
7167 * this runqueue and our priority is higher than the current's
7169 if (rq->curr == p) {
7170 if (p->prio > oldprio)
7171 resched_task(rq->curr);
7173 check_preempt_curr(rq, p, 0);
7176 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7178 struct sched_entity *se = &p->se;
7179 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7182 * Ensure the task's vruntime is normalized, so that when its
7183 * switched back to the fair class the enqueue_entity(.flags=0) will
7184 * do the right thing.
7186 * If it was on_rq, then the dequeue_entity(.flags=0) will already
7187 * have normalized the vruntime, if it was !on_rq, then only when
7188 * the task is sleeping will it still have non-normalized vruntime.
7190 if (!se->on_rq && p->state != TASK_RUNNING) {
7192 * Fix up our vruntime so that the current sleep doesn't
7193 * cause 'unlimited' sleep bonus.
7195 place_entity(cfs_rq, se, 0);
7196 se->vruntime -= cfs_rq->min_vruntime;
7201 * Remove our load from contribution when we leave sched_fair
7202 * and ensure we don't carry in an old decay_count if we
7205 if (se->avg.decay_count) {
7206 __synchronize_entity_decay(se);
7207 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7213 * We switched to the sched_fair class.
7215 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7221 * We were most likely switched from sched_rt, so
7222 * kick off the schedule if running, otherwise just see
7223 * if we can still preempt the current task.
7226 resched_task(rq->curr);
7228 check_preempt_curr(rq, p, 0);
7231 /* Account for a task changing its policy or group.
7233 * This routine is mostly called to set cfs_rq->curr field when a task
7234 * migrates between groups/classes.
7236 static void set_curr_task_fair(struct rq *rq)
7238 struct sched_entity *se = &rq->curr->se;
7240 for_each_sched_entity(se) {
7241 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7243 set_next_entity(cfs_rq, se);
7244 /* ensure bandwidth has been allocated on our new cfs_rq */
7245 account_cfs_rq_runtime(cfs_rq, 0);
7249 void init_cfs_rq(struct cfs_rq *cfs_rq)
7251 cfs_rq->tasks_timeline = RB_ROOT;
7252 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7253 #ifndef CONFIG_64BIT
7254 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7257 atomic64_set(&cfs_rq->decay_counter, 1);
7258 atomic_long_set(&cfs_rq->removed_load, 0);
7262 #ifdef CONFIG_FAIR_GROUP_SCHED
7263 static void task_move_group_fair(struct task_struct *p, int on_rq)
7265 struct cfs_rq *cfs_rq;
7267 * If the task was not on the rq at the time of this cgroup movement
7268 * it must have been asleep, sleeping tasks keep their ->vruntime
7269 * absolute on their old rq until wakeup (needed for the fair sleeper
7270 * bonus in place_entity()).
7272 * If it was on the rq, we've just 'preempted' it, which does convert
7273 * ->vruntime to a relative base.
7275 * Make sure both cases convert their relative position when migrating
7276 * to another cgroup's rq. This does somewhat interfere with the
7277 * fair sleeper stuff for the first placement, but who cares.
7280 * When !on_rq, vruntime of the task has usually NOT been normalized.
7281 * But there are some cases where it has already been normalized:
7283 * - Moving a forked child which is waiting for being woken up by
7284 * wake_up_new_task().
7285 * - Moving a task which has been woken up by try_to_wake_up() and
7286 * waiting for actually being woken up by sched_ttwu_pending().
7288 * To prevent boost or penalty in the new cfs_rq caused by delta
7289 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7291 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7295 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7296 set_task_rq(p, task_cpu(p));
7298 cfs_rq = cfs_rq_of(&p->se);
7299 p->se.vruntime += cfs_rq->min_vruntime;
7302 * migrate_task_rq_fair() will have removed our previous
7303 * contribution, but we must synchronize for ongoing future
7306 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7307 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7312 void free_fair_sched_group(struct task_group *tg)
7316 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7318 for_each_possible_cpu(i) {
7320 kfree(tg->cfs_rq[i]);
7329 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7331 struct cfs_rq *cfs_rq;
7332 struct sched_entity *se;
7335 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7338 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7342 tg->shares = NICE_0_LOAD;
7344 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7346 for_each_possible_cpu(i) {
7347 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7348 GFP_KERNEL, cpu_to_node(i));
7352 se = kzalloc_node(sizeof(struct sched_entity),
7353 GFP_KERNEL, cpu_to_node(i));
7357 init_cfs_rq(cfs_rq);
7358 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7369 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7371 struct rq *rq = cpu_rq(cpu);
7372 unsigned long flags;
7375 * Only empty task groups can be destroyed; so we can speculatively
7376 * check on_list without danger of it being re-added.
7378 if (!tg->cfs_rq[cpu]->on_list)
7381 raw_spin_lock_irqsave(&rq->lock, flags);
7382 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7383 raw_spin_unlock_irqrestore(&rq->lock, flags);
7386 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7387 struct sched_entity *se, int cpu,
7388 struct sched_entity *parent)
7390 struct rq *rq = cpu_rq(cpu);
7394 init_cfs_rq_runtime(cfs_rq);
7396 tg->cfs_rq[cpu] = cfs_rq;
7399 /* se could be NULL for root_task_group */
7404 se->cfs_rq = &rq->cfs;
7406 se->cfs_rq = parent->my_q;
7409 /* guarantee group entities always have weight */
7410 update_load_set(&se->load, NICE_0_LOAD);
7411 se->parent = parent;
7414 static DEFINE_MUTEX(shares_mutex);
7416 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7419 unsigned long flags;
7422 * We can't change the weight of the root cgroup.
7427 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7429 mutex_lock(&shares_mutex);
7430 if (tg->shares == shares)
7433 tg->shares = shares;
7434 for_each_possible_cpu(i) {
7435 struct rq *rq = cpu_rq(i);
7436 struct sched_entity *se;
7439 /* Propagate contribution to hierarchy */
7440 raw_spin_lock_irqsave(&rq->lock, flags);
7442 /* Possible calls to update_curr() need rq clock */
7443 update_rq_clock(rq);
7444 for_each_sched_entity(se)
7445 update_cfs_shares(group_cfs_rq(se));
7446 raw_spin_unlock_irqrestore(&rq->lock, flags);
7450 mutex_unlock(&shares_mutex);
7453 #else /* CONFIG_FAIR_GROUP_SCHED */
7455 void free_fair_sched_group(struct task_group *tg) { }
7457 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7462 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7464 #endif /* CONFIG_FAIR_GROUP_SCHED */
7467 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7469 struct sched_entity *se = &task->se;
7470 unsigned int rr_interval = 0;
7473 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7476 if (rq->cfs.load.weight)
7477 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7483 * All the scheduling class methods:
7485 const struct sched_class fair_sched_class = {
7486 .next = &idle_sched_class,
7487 .enqueue_task = enqueue_task_fair,
7488 .dequeue_task = dequeue_task_fair,
7489 .yield_task = yield_task_fair,
7490 .yield_to_task = yield_to_task_fair,
7492 .check_preempt_curr = check_preempt_wakeup,
7494 .pick_next_task = pick_next_task_fair,
7495 .put_prev_task = put_prev_task_fair,
7498 .select_task_rq = select_task_rq_fair,
7499 .migrate_task_rq = migrate_task_rq_fair,
7501 .rq_online = rq_online_fair,
7502 .rq_offline = rq_offline_fair,
7504 .task_waking = task_waking_fair,
7507 .set_curr_task = set_curr_task_fair,
7508 .task_tick = task_tick_fair,
7509 .task_fork = task_fork_fair,
7511 .prio_changed = prio_changed_fair,
7512 .switched_from = switched_from_fair,
7513 .switched_to = switched_to_fair,
7515 .get_rr_interval = get_rr_interval_fair,
7517 #ifdef CONFIG_FAIR_GROUP_SCHED
7518 .task_move_group = task_move_group_fair,
7522 #ifdef CONFIG_SCHED_DEBUG
7523 void print_cfs_stats(struct seq_file *m, int cpu)
7525 struct cfs_rq *cfs_rq;
7528 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7529 print_cfs_rq(m, cpu, cfs_rq);
7534 __init void init_sched_fair_class(void)
7537 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7539 #ifdef CONFIG_NO_HZ_COMMON
7540 nohz.next_balance = jiffies;
7541 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7542 cpu_notifier(sched_ilb_notifier, 0);