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 ? */
325 static inline struct cfs_rq *
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)
340 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
342 int se_depth, pse_depth;
345 * preemption test can be made between sibling entities who are in the
346 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
347 * both tasks until we find their ancestors who are siblings of common
351 /* First walk up until both entities are at same depth */
352 se_depth = (*se)->depth;
353 pse_depth = (*pse)->depth;
355 while (se_depth > pse_depth) {
357 *se = parent_entity(*se);
360 while (pse_depth > se_depth) {
362 *pse = parent_entity(*pse);
365 while (!is_same_group(*se, *pse)) {
366 *se = parent_entity(*se);
367 *pse = parent_entity(*pse);
371 #else /* !CONFIG_FAIR_GROUP_SCHED */
373 static inline struct task_struct *task_of(struct sched_entity *se)
375 return container_of(se, struct task_struct, se);
378 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
380 return container_of(cfs_rq, struct rq, cfs);
383 #define entity_is_task(se) 1
385 #define for_each_sched_entity(se) \
386 for (; se; se = NULL)
388 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
390 return &task_rq(p)->cfs;
393 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
395 struct task_struct *p = task_of(se);
396 struct rq *rq = task_rq(p);
401 /* runqueue "owned" by this group */
402 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
407 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
415 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
416 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
418 static inline struct sched_entity *parent_entity(struct sched_entity *se)
424 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
428 #endif /* CONFIG_FAIR_GROUP_SCHED */
430 static __always_inline
431 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
433 /**************************************************************
434 * Scheduling class tree data structure manipulation methods:
437 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
439 s64 delta = (s64)(vruntime - max_vruntime);
441 max_vruntime = vruntime;
446 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
448 s64 delta = (s64)(vruntime - min_vruntime);
450 min_vruntime = vruntime;
455 static inline int entity_before(struct sched_entity *a,
456 struct sched_entity *b)
458 return (s64)(a->vruntime - b->vruntime) < 0;
461 static void update_min_vruntime(struct cfs_rq *cfs_rq)
463 u64 vruntime = cfs_rq->min_vruntime;
466 vruntime = cfs_rq->curr->vruntime;
468 if (cfs_rq->rb_leftmost) {
469 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
474 vruntime = se->vruntime;
476 vruntime = min_vruntime(vruntime, se->vruntime);
479 /* ensure we never gain time by being placed backwards. */
480 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
483 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
488 * Enqueue an entity into the rb-tree:
490 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
492 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
493 struct rb_node *parent = NULL;
494 struct sched_entity *entry;
498 * Find the right place in the rbtree:
502 entry = rb_entry(parent, struct sched_entity, run_node);
504 * We dont care about collisions. Nodes with
505 * the same key stay together.
507 if (entity_before(se, entry)) {
508 link = &parent->rb_left;
510 link = &parent->rb_right;
516 * Maintain a cache of leftmost tree entries (it is frequently
520 cfs_rq->rb_leftmost = &se->run_node;
522 rb_link_node(&se->run_node, parent, link);
523 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
526 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
528 if (cfs_rq->rb_leftmost == &se->run_node) {
529 struct rb_node *next_node;
531 next_node = rb_next(&se->run_node);
532 cfs_rq->rb_leftmost = next_node;
535 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
538 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
540 struct rb_node *left = cfs_rq->rb_leftmost;
545 return rb_entry(left, struct sched_entity, run_node);
548 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
550 struct rb_node *next = rb_next(&se->run_node);
555 return rb_entry(next, struct sched_entity, run_node);
558 #ifdef CONFIG_SCHED_DEBUG
559 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
561 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
566 return rb_entry(last, struct sched_entity, run_node);
569 /**************************************************************
570 * Scheduling class statistics methods:
573 int sched_proc_update_handler(struct ctl_table *table, int write,
574 void __user *buffer, size_t *lenp,
577 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
578 int factor = get_update_sysctl_factor();
583 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
584 sysctl_sched_min_granularity);
586 #define WRT_SYSCTL(name) \
587 (normalized_sysctl_##name = sysctl_##name / (factor))
588 WRT_SYSCTL(sched_min_granularity);
589 WRT_SYSCTL(sched_latency);
590 WRT_SYSCTL(sched_wakeup_granularity);
600 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
602 if (unlikely(se->load.weight != NICE_0_LOAD))
603 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
609 * The idea is to set a period in which each task runs once.
611 * When there are too many tasks (sched_nr_latency) we have to stretch
612 * this period because otherwise the slices get too small.
614 * p = (nr <= nl) ? l : l*nr/nl
616 static u64 __sched_period(unsigned long nr_running)
618 u64 period = sysctl_sched_latency;
619 unsigned long nr_latency = sched_nr_latency;
621 if (unlikely(nr_running > nr_latency)) {
622 period = sysctl_sched_min_granularity;
623 period *= nr_running;
630 * We calculate the wall-time slice from the period by taking a part
631 * proportional to the weight.
635 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
637 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
639 for_each_sched_entity(se) {
640 struct load_weight *load;
641 struct load_weight lw;
643 cfs_rq = cfs_rq_of(se);
644 load = &cfs_rq->load;
646 if (unlikely(!se->on_rq)) {
649 update_load_add(&lw, se->load.weight);
652 slice = __calc_delta(slice, se->load.weight, load);
658 * We calculate the vruntime slice of a to-be-inserted task.
662 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
664 return calc_delta_fair(sched_slice(cfs_rq, se), se);
668 static unsigned long task_h_load(struct task_struct *p);
670 static inline void __update_task_entity_contrib(struct sched_entity *se);
672 /* Give new task start runnable values to heavy its load in infant time */
673 void init_task_runnable_average(struct task_struct *p)
677 p->se.avg.decay_count = 0;
678 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
679 p->se.avg.runnable_avg_sum = slice;
680 p->se.avg.runnable_avg_period = slice;
681 __update_task_entity_contrib(&p->se);
684 void init_task_runnable_average(struct task_struct *p)
690 * Update the current task's runtime statistics.
692 static void update_curr(struct cfs_rq *cfs_rq)
694 struct sched_entity *curr = cfs_rq->curr;
695 u64 now = rq_clock_task(rq_of(cfs_rq));
701 delta_exec = now - curr->exec_start;
702 if (unlikely((s64)delta_exec <= 0))
705 curr->exec_start = now;
707 schedstat_set(curr->statistics.exec_max,
708 max(delta_exec, curr->statistics.exec_max));
710 curr->sum_exec_runtime += delta_exec;
711 schedstat_add(cfs_rq, exec_clock, delta_exec);
713 curr->vruntime += calc_delta_fair(delta_exec, curr);
714 update_min_vruntime(cfs_rq);
716 if (entity_is_task(curr)) {
717 struct task_struct *curtask = task_of(curr);
719 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
720 cpuacct_charge(curtask, delta_exec);
721 account_group_exec_runtime(curtask, delta_exec);
724 account_cfs_rq_runtime(cfs_rq, delta_exec);
728 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
730 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
734 * Task is being enqueued - update stats:
736 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
739 * Are we enqueueing a waiting task? (for current tasks
740 * a dequeue/enqueue event is a NOP)
742 if (se != cfs_rq->curr)
743 update_stats_wait_start(cfs_rq, se);
747 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
749 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
750 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
751 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
752 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
753 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
754 #ifdef CONFIG_SCHEDSTATS
755 if (entity_is_task(se)) {
756 trace_sched_stat_wait(task_of(se),
757 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
760 schedstat_set(se->statistics.wait_start, 0);
764 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
767 * Mark the end of the wait period if dequeueing a
770 if (se != cfs_rq->curr)
771 update_stats_wait_end(cfs_rq, se);
775 * We are picking a new current task - update its stats:
778 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
781 * We are starting a new run period:
783 se->exec_start = rq_clock_task(rq_of(cfs_rq));
786 /**************************************************
787 * Scheduling class queueing methods:
790 #ifdef CONFIG_NUMA_BALANCING
792 * Approximate time to scan a full NUMA task in ms. The task scan period is
793 * calculated based on the tasks virtual memory size and
794 * numa_balancing_scan_size.
796 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
797 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
799 /* Portion of address space to scan in MB */
800 unsigned int sysctl_numa_balancing_scan_size = 256;
802 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
803 unsigned int sysctl_numa_balancing_scan_delay = 1000;
805 static unsigned int task_nr_scan_windows(struct task_struct *p)
807 unsigned long rss = 0;
808 unsigned long nr_scan_pages;
811 * Calculations based on RSS as non-present and empty pages are skipped
812 * by the PTE scanner and NUMA hinting faults should be trapped based
815 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
816 rss = get_mm_rss(p->mm);
820 rss = round_up(rss, nr_scan_pages);
821 return rss / nr_scan_pages;
824 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
825 #define MAX_SCAN_WINDOW 2560
827 static unsigned int task_scan_min(struct task_struct *p)
829 unsigned int scan, floor;
830 unsigned int windows = 1;
832 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
833 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
834 floor = 1000 / windows;
836 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
837 return max_t(unsigned int, floor, scan);
840 static unsigned int task_scan_max(struct task_struct *p)
842 unsigned int smin = task_scan_min(p);
845 /* Watch for min being lower than max due to floor calculations */
846 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
847 return max(smin, smax);
850 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
852 rq->nr_numa_running += (p->numa_preferred_nid != -1);
853 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
856 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
858 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
859 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
865 spinlock_t lock; /* nr_tasks, tasks */
868 struct list_head task_list;
871 nodemask_t active_nodes;
872 unsigned long total_faults;
874 * Faults_cpu is used to decide whether memory should move
875 * towards the CPU. As a consequence, these stats are weighted
876 * more by CPU use than by memory faults.
878 unsigned long *faults_cpu;
879 unsigned long faults[0];
882 /* Shared or private faults. */
883 #define NR_NUMA_HINT_FAULT_TYPES 2
885 /* Memory and CPU locality */
886 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
888 /* Averaged statistics, and temporary buffers. */
889 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
891 pid_t task_numa_group_id(struct task_struct *p)
893 return p->numa_group ? p->numa_group->gid : 0;
896 static inline int task_faults_idx(int nid, int priv)
898 return NR_NUMA_HINT_FAULT_TYPES * nid + priv;
901 static inline unsigned long task_faults(struct task_struct *p, int nid)
903 if (!p->numa_faults_memory)
906 return p->numa_faults_memory[task_faults_idx(nid, 0)] +
907 p->numa_faults_memory[task_faults_idx(nid, 1)];
910 static inline unsigned long group_faults(struct task_struct *p, int nid)
915 return p->numa_group->faults[task_faults_idx(nid, 0)] +
916 p->numa_group->faults[task_faults_idx(nid, 1)];
919 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
921 return group->faults_cpu[task_faults_idx(nid, 0)] +
922 group->faults_cpu[task_faults_idx(nid, 1)];
926 * These return the fraction of accesses done by a particular task, or
927 * task group, on a particular numa node. The group weight is given a
928 * larger multiplier, in order to group tasks together that are almost
929 * evenly spread out between numa nodes.
931 static inline unsigned long task_weight(struct task_struct *p, int nid)
933 unsigned long total_faults;
935 if (!p->numa_faults_memory)
938 total_faults = p->total_numa_faults;
943 return 1000 * task_faults(p, nid) / total_faults;
946 static inline unsigned long group_weight(struct task_struct *p, int nid)
948 if (!p->numa_group || !p->numa_group->total_faults)
951 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
954 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
955 int src_nid, int dst_cpu)
957 struct numa_group *ng = p->numa_group;
958 int dst_nid = cpu_to_node(dst_cpu);
959 int last_cpupid, this_cpupid;
961 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
964 * Multi-stage node selection is used in conjunction with a periodic
965 * migration fault to build a temporal task<->page relation. By using
966 * a two-stage filter we remove short/unlikely relations.
968 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
969 * a task's usage of a particular page (n_p) per total usage of this
970 * page (n_t) (in a given time-span) to a probability.
972 * Our periodic faults will sample this probability and getting the
973 * same result twice in a row, given these samples are fully
974 * independent, is then given by P(n)^2, provided our sample period
975 * is sufficiently short compared to the usage pattern.
977 * This quadric squishes small probabilities, making it less likely we
978 * act on an unlikely task<->page relation.
980 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
981 if (!cpupid_pid_unset(last_cpupid) &&
982 cpupid_to_nid(last_cpupid) != dst_nid)
985 /* Always allow migrate on private faults */
986 if (cpupid_match_pid(p, last_cpupid))
989 /* A shared fault, but p->numa_group has not been set up yet. */
994 * Do not migrate if the destination is not a node that
995 * is actively used by this numa group.
997 if (!node_isset(dst_nid, ng->active_nodes))
1001 * Source is a node that is not actively used by this
1002 * numa group, while the destination is. Migrate.
1004 if (!node_isset(src_nid, ng->active_nodes))
1008 * Both source and destination are nodes in active
1009 * use by this numa group. Maximize memory bandwidth
1010 * by migrating from more heavily used groups, to less
1011 * heavily used ones, spreading the load around.
1012 * Use a 1/4 hysteresis to avoid spurious page movement.
1014 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1017 static unsigned long weighted_cpuload(const int cpu);
1018 static unsigned long source_load(int cpu, int type);
1019 static unsigned long target_load(int cpu, int type);
1020 static unsigned long capacity_of(int cpu);
1021 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1023 /* Cached statistics for all CPUs within a node */
1025 unsigned long nr_running;
1028 /* Total compute capacity of CPUs on a node */
1029 unsigned long compute_capacity;
1031 /* Approximate capacity in terms of runnable tasks on a node */
1032 unsigned long task_capacity;
1033 int has_free_capacity;
1037 * XXX borrowed from update_sg_lb_stats
1039 static void update_numa_stats(struct numa_stats *ns, int nid)
1043 memset(ns, 0, sizeof(*ns));
1044 for_each_cpu(cpu, cpumask_of_node(nid)) {
1045 struct rq *rq = cpu_rq(cpu);
1047 ns->nr_running += rq->nr_running;
1048 ns->load += weighted_cpuload(cpu);
1049 ns->compute_capacity += capacity_of(cpu);
1055 * If we raced with hotplug and there are no CPUs left in our mask
1056 * the @ns structure is NULL'ed and task_numa_compare() will
1057 * not find this node attractive.
1059 * We'll either bail at !has_free_capacity, or we'll detect a huge
1060 * imbalance and bail there.
1066 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE);
1067 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1070 struct task_numa_env {
1071 struct task_struct *p;
1073 int src_cpu, src_nid;
1074 int dst_cpu, dst_nid;
1076 struct numa_stats src_stats, dst_stats;
1080 struct task_struct *best_task;
1085 static void task_numa_assign(struct task_numa_env *env,
1086 struct task_struct *p, long imp)
1089 put_task_struct(env->best_task);
1094 env->best_imp = imp;
1095 env->best_cpu = env->dst_cpu;
1098 static bool load_too_imbalanced(long src_load, long dst_load,
1099 struct task_numa_env *env)
1102 long orig_src_load, orig_dst_load;
1103 long src_capacity, dst_capacity;
1106 * The load is corrected for the CPU capacity available on each node.
1109 * ------------ vs ---------
1110 * src_capacity dst_capacity
1112 src_capacity = env->src_stats.compute_capacity;
1113 dst_capacity = env->dst_stats.compute_capacity;
1115 /* We care about the slope of the imbalance, not the direction. */
1116 if (dst_load < src_load)
1117 swap(dst_load, src_load);
1119 /* Is the difference below the threshold? */
1120 imb = dst_load * src_capacity * 100 -
1121 src_load * dst_capacity * env->imbalance_pct;
1126 * The imbalance is above the allowed threshold.
1127 * Compare it with the old imbalance.
1129 orig_src_load = env->src_stats.load;
1130 orig_dst_load = env->dst_stats.load;
1132 if (orig_dst_load < orig_src_load)
1133 swap(orig_dst_load, orig_src_load);
1135 old_imb = orig_dst_load * src_capacity * 100 -
1136 orig_src_load * dst_capacity * env->imbalance_pct;
1138 /* Would this change make things worse? */
1139 return (imb > old_imb);
1143 * This checks if the overall compute and NUMA accesses of the system would
1144 * be improved if the source tasks was migrated to the target dst_cpu taking
1145 * into account that it might be best if task running on the dst_cpu should
1146 * be exchanged with the source task
1148 static void task_numa_compare(struct task_numa_env *env,
1149 long taskimp, long groupimp)
1151 struct rq *src_rq = cpu_rq(env->src_cpu);
1152 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1153 struct task_struct *cur;
1154 struct task_group *tg;
1155 long src_load, dst_load;
1157 long imp = (groupimp > 0) ? groupimp : taskimp;
1160 cur = ACCESS_ONCE(dst_rq->curr);
1161 if (cur->pid == 0) /* idle */
1165 * "imp" is the fault differential for the source task between the
1166 * source and destination node. Calculate the total differential for
1167 * the source task and potential destination task. The more negative
1168 * the value is, the more rmeote accesses that would be expected to
1169 * be incurred if the tasks were swapped.
1172 /* Skip this swap candidate if cannot move to the source cpu */
1173 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1177 * If dst and source tasks are in the same NUMA group, or not
1178 * in any group then look only at task weights.
1180 if (cur->numa_group == env->p->numa_group) {
1181 imp = taskimp + task_weight(cur, env->src_nid) -
1182 task_weight(cur, env->dst_nid);
1184 * Add some hysteresis to prevent swapping the
1185 * tasks within a group over tiny differences.
1187 if (cur->numa_group)
1191 * Compare the group weights. If a task is all by
1192 * itself (not part of a group), use the task weight
1195 if (env->p->numa_group)
1200 if (cur->numa_group)
1201 imp += group_weight(cur, env->src_nid) -
1202 group_weight(cur, env->dst_nid);
1204 imp += task_weight(cur, env->src_nid) -
1205 task_weight(cur, env->dst_nid);
1209 if (imp < env->best_imp)
1213 /* Is there capacity at our destination? */
1214 if (env->src_stats.has_free_capacity &&
1215 !env->dst_stats.has_free_capacity)
1221 /* Balance doesn't matter much if we're running a task per cpu */
1222 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1226 * In the overloaded case, try and keep the load balanced.
1229 src_load = env->src_stats.load;
1230 dst_load = env->dst_stats.load;
1232 /* Calculate the effect of moving env->p from src to dst. */
1233 load = env->p->se.load.weight;
1234 tg = task_group(env->p);
1235 src_load += effective_load(tg, env->src_cpu, -load, -load);
1236 dst_load += effective_load(tg, env->dst_cpu, load, load);
1239 /* Cur moves in the opposite direction. */
1240 load = cur->se.load.weight;
1241 tg = task_group(cur);
1242 src_load += effective_load(tg, env->src_cpu, load, load);
1243 dst_load += effective_load(tg, env->dst_cpu, -load, -load);
1246 if (load_too_imbalanced(src_load, dst_load, env))
1250 task_numa_assign(env, cur, imp);
1255 static void task_numa_find_cpu(struct task_numa_env *env,
1256 long taskimp, long groupimp)
1260 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1261 /* Skip this CPU if the source task cannot migrate */
1262 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1266 task_numa_compare(env, taskimp, groupimp);
1270 static int task_numa_migrate(struct task_struct *p)
1272 struct task_numa_env env = {
1275 .src_cpu = task_cpu(p),
1276 .src_nid = task_node(p),
1278 .imbalance_pct = 112,
1284 struct sched_domain *sd;
1285 unsigned long taskweight, groupweight;
1287 long taskimp, groupimp;
1290 * Pick the lowest SD_NUMA domain, as that would have the smallest
1291 * imbalance and would be the first to start moving tasks about.
1293 * And we want to avoid any moving of tasks about, as that would create
1294 * random movement of tasks -- counter the numa conditions we're trying
1298 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1300 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1304 * Cpusets can break the scheduler domain tree into smaller
1305 * balance domains, some of which do not cross NUMA boundaries.
1306 * Tasks that are "trapped" in such domains cannot be migrated
1307 * elsewhere, so there is no point in (re)trying.
1309 if (unlikely(!sd)) {
1310 p->numa_preferred_nid = task_node(p);
1314 taskweight = task_weight(p, env.src_nid);
1315 groupweight = group_weight(p, env.src_nid);
1316 update_numa_stats(&env.src_stats, env.src_nid);
1317 env.dst_nid = p->numa_preferred_nid;
1318 taskimp = task_weight(p, env.dst_nid) - taskweight;
1319 groupimp = group_weight(p, env.dst_nid) - groupweight;
1320 update_numa_stats(&env.dst_stats, env.dst_nid);
1322 /* Try to find a spot on the preferred nid. */
1323 task_numa_find_cpu(&env, taskimp, groupimp);
1325 /* No space available on the preferred nid. Look elsewhere. */
1326 if (env.best_cpu == -1) {
1327 for_each_online_node(nid) {
1328 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1331 /* Only consider nodes where both task and groups benefit */
1332 taskimp = task_weight(p, nid) - taskweight;
1333 groupimp = group_weight(p, nid) - groupweight;
1334 if (taskimp < 0 && groupimp < 0)
1338 update_numa_stats(&env.dst_stats, env.dst_nid);
1339 task_numa_find_cpu(&env, taskimp, groupimp);
1343 /* No better CPU than the current one was found. */
1344 if (env.best_cpu == -1)
1348 * If the task is part of a workload that spans multiple NUMA nodes,
1349 * and is migrating into one of the workload's active nodes, remember
1350 * this node as the task's preferred numa node, so the workload can
1352 * A task that migrated to a second choice node will be better off
1353 * trying for a better one later. Do not set the preferred node here.
1355 if (p->numa_group && node_isset(env.dst_nid, p->numa_group->active_nodes))
1356 sched_setnuma(p, env.dst_nid);
1359 * Reset the scan period if the task is being rescheduled on an
1360 * alternative node to recheck if the tasks is now properly placed.
1362 p->numa_scan_period = task_scan_min(p);
1364 if (env.best_task == NULL) {
1365 ret = migrate_task_to(p, env.best_cpu);
1367 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1371 ret = migrate_swap(p, env.best_task);
1373 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1374 put_task_struct(env.best_task);
1378 /* Attempt to migrate a task to a CPU on the preferred node. */
1379 static void numa_migrate_preferred(struct task_struct *p)
1381 unsigned long interval = HZ;
1383 /* This task has no NUMA fault statistics yet */
1384 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1387 /* Periodically retry migrating the task to the preferred node */
1388 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1389 p->numa_migrate_retry = jiffies + interval;
1391 /* Success if task is already running on preferred CPU */
1392 if (task_node(p) == p->numa_preferred_nid)
1395 /* Otherwise, try migrate to a CPU on the preferred node */
1396 task_numa_migrate(p);
1400 * Find the nodes on which the workload is actively running. We do this by
1401 * tracking the nodes from which NUMA hinting faults are triggered. This can
1402 * be different from the set of nodes where the workload's memory is currently
1405 * The bitmask is used to make smarter decisions on when to do NUMA page
1406 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1407 * are added when they cause over 6/16 of the maximum number of faults, but
1408 * only removed when they drop below 3/16.
1410 static void update_numa_active_node_mask(struct numa_group *numa_group)
1412 unsigned long faults, max_faults = 0;
1415 for_each_online_node(nid) {
1416 faults = group_faults_cpu(numa_group, nid);
1417 if (faults > max_faults)
1418 max_faults = faults;
1421 for_each_online_node(nid) {
1422 faults = group_faults_cpu(numa_group, nid);
1423 if (!node_isset(nid, numa_group->active_nodes)) {
1424 if (faults > max_faults * 6 / 16)
1425 node_set(nid, numa_group->active_nodes);
1426 } else if (faults < max_faults * 3 / 16)
1427 node_clear(nid, numa_group->active_nodes);
1432 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1433 * increments. The more local the fault statistics are, the higher the scan
1434 * period will be for the next scan window. If local/remote ratio is below
1435 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1436 * scan period will decrease
1438 #define NUMA_PERIOD_SLOTS 10
1439 #define NUMA_PERIOD_THRESHOLD 3
1442 * Increase the scan period (slow down scanning) if the majority of
1443 * our memory is already on our local node, or if the majority of
1444 * the page accesses are shared with other processes.
1445 * Otherwise, decrease the scan period.
1447 static void update_task_scan_period(struct task_struct *p,
1448 unsigned long shared, unsigned long private)
1450 unsigned int period_slot;
1454 unsigned long remote = p->numa_faults_locality[0];
1455 unsigned long local = p->numa_faults_locality[1];
1458 * If there were no record hinting faults then either the task is
1459 * completely idle or all activity is areas that are not of interest
1460 * to automatic numa balancing. Scan slower
1462 if (local + shared == 0) {
1463 p->numa_scan_period = min(p->numa_scan_period_max,
1464 p->numa_scan_period << 1);
1466 p->mm->numa_next_scan = jiffies +
1467 msecs_to_jiffies(p->numa_scan_period);
1473 * Prepare to scale scan period relative to the current period.
1474 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1475 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1476 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1478 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1479 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1480 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1481 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1484 diff = slot * period_slot;
1486 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1489 * Scale scan rate increases based on sharing. There is an
1490 * inverse relationship between the degree of sharing and
1491 * the adjustment made to the scanning period. Broadly
1492 * speaking the intent is that there is little point
1493 * scanning faster if shared accesses dominate as it may
1494 * simply bounce migrations uselessly
1496 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1497 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1500 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1501 task_scan_min(p), task_scan_max(p));
1502 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1506 * Get the fraction of time the task has been running since the last
1507 * NUMA placement cycle. The scheduler keeps similar statistics, but
1508 * decays those on a 32ms period, which is orders of magnitude off
1509 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1510 * stats only if the task is so new there are no NUMA statistics yet.
1512 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1514 u64 runtime, delta, now;
1515 /* Use the start of this time slice to avoid calculations. */
1516 now = p->se.exec_start;
1517 runtime = p->se.sum_exec_runtime;
1519 if (p->last_task_numa_placement) {
1520 delta = runtime - p->last_sum_exec_runtime;
1521 *period = now - p->last_task_numa_placement;
1523 delta = p->se.avg.runnable_avg_sum;
1524 *period = p->se.avg.runnable_avg_period;
1527 p->last_sum_exec_runtime = runtime;
1528 p->last_task_numa_placement = now;
1533 static void task_numa_placement(struct task_struct *p)
1535 int seq, nid, max_nid = -1, max_group_nid = -1;
1536 unsigned long max_faults = 0, max_group_faults = 0;
1537 unsigned long fault_types[2] = { 0, 0 };
1538 unsigned long total_faults;
1539 u64 runtime, period;
1540 spinlock_t *group_lock = NULL;
1542 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1543 if (p->numa_scan_seq == seq)
1545 p->numa_scan_seq = seq;
1546 p->numa_scan_period_max = task_scan_max(p);
1548 total_faults = p->numa_faults_locality[0] +
1549 p->numa_faults_locality[1];
1550 runtime = numa_get_avg_runtime(p, &period);
1552 /* If the task is part of a group prevent parallel updates to group stats */
1553 if (p->numa_group) {
1554 group_lock = &p->numa_group->lock;
1555 spin_lock_irq(group_lock);
1558 /* Find the node with the highest number of faults */
1559 for_each_online_node(nid) {
1560 unsigned long faults = 0, group_faults = 0;
1563 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1564 long diff, f_diff, f_weight;
1566 i = task_faults_idx(nid, priv);
1568 /* Decay existing window, copy faults since last scan */
1569 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1570 fault_types[priv] += p->numa_faults_buffer_memory[i];
1571 p->numa_faults_buffer_memory[i] = 0;
1574 * Normalize the faults_from, so all tasks in a group
1575 * count according to CPU use, instead of by the raw
1576 * number of faults. Tasks with little runtime have
1577 * little over-all impact on throughput, and thus their
1578 * faults are less important.
1580 f_weight = div64_u64(runtime << 16, period + 1);
1581 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1583 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1584 p->numa_faults_buffer_cpu[i] = 0;
1586 p->numa_faults_memory[i] += diff;
1587 p->numa_faults_cpu[i] += f_diff;
1588 faults += p->numa_faults_memory[i];
1589 p->total_numa_faults += diff;
1590 if (p->numa_group) {
1591 /* safe because we can only change our own group */
1592 p->numa_group->faults[i] += diff;
1593 p->numa_group->faults_cpu[i] += f_diff;
1594 p->numa_group->total_faults += diff;
1595 group_faults += p->numa_group->faults[i];
1599 if (faults > max_faults) {
1600 max_faults = faults;
1604 if (group_faults > max_group_faults) {
1605 max_group_faults = group_faults;
1606 max_group_nid = nid;
1610 update_task_scan_period(p, fault_types[0], fault_types[1]);
1612 if (p->numa_group) {
1613 update_numa_active_node_mask(p->numa_group);
1614 spin_unlock_irq(group_lock);
1615 max_nid = max_group_nid;
1619 /* Set the new preferred node */
1620 if (max_nid != p->numa_preferred_nid)
1621 sched_setnuma(p, max_nid);
1623 if (task_node(p) != p->numa_preferred_nid)
1624 numa_migrate_preferred(p);
1628 static inline int get_numa_group(struct numa_group *grp)
1630 return atomic_inc_not_zero(&grp->refcount);
1633 static inline void put_numa_group(struct numa_group *grp)
1635 if (atomic_dec_and_test(&grp->refcount))
1636 kfree_rcu(grp, rcu);
1639 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1642 struct numa_group *grp, *my_grp;
1643 struct task_struct *tsk;
1645 int cpu = cpupid_to_cpu(cpupid);
1648 if (unlikely(!p->numa_group)) {
1649 unsigned int size = sizeof(struct numa_group) +
1650 4*nr_node_ids*sizeof(unsigned long);
1652 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1656 atomic_set(&grp->refcount, 1);
1657 spin_lock_init(&grp->lock);
1658 INIT_LIST_HEAD(&grp->task_list);
1660 /* Second half of the array tracks nids where faults happen */
1661 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1664 node_set(task_node(current), grp->active_nodes);
1666 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1667 grp->faults[i] = p->numa_faults_memory[i];
1669 grp->total_faults = p->total_numa_faults;
1671 list_add(&p->numa_entry, &grp->task_list);
1673 rcu_assign_pointer(p->numa_group, grp);
1677 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1679 if (!cpupid_match_pid(tsk, cpupid))
1682 grp = rcu_dereference(tsk->numa_group);
1686 my_grp = p->numa_group;
1691 * Only join the other group if its bigger; if we're the bigger group,
1692 * the other task will join us.
1694 if (my_grp->nr_tasks > grp->nr_tasks)
1698 * Tie-break on the grp address.
1700 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1703 /* Always join threads in the same process. */
1704 if (tsk->mm == current->mm)
1707 /* Simple filter to avoid false positives due to PID collisions */
1708 if (flags & TNF_SHARED)
1711 /* Update priv based on whether false sharing was detected */
1714 if (join && !get_numa_group(grp))
1722 BUG_ON(irqs_disabled());
1723 double_lock_irq(&my_grp->lock, &grp->lock);
1725 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1726 my_grp->faults[i] -= p->numa_faults_memory[i];
1727 grp->faults[i] += p->numa_faults_memory[i];
1729 my_grp->total_faults -= p->total_numa_faults;
1730 grp->total_faults += p->total_numa_faults;
1732 list_move(&p->numa_entry, &grp->task_list);
1736 spin_unlock(&my_grp->lock);
1737 spin_unlock_irq(&grp->lock);
1739 rcu_assign_pointer(p->numa_group, grp);
1741 put_numa_group(my_grp);
1749 void task_numa_free(struct task_struct *p)
1751 struct numa_group *grp = p->numa_group;
1752 void *numa_faults = p->numa_faults_memory;
1753 unsigned long flags;
1757 spin_lock_irqsave(&grp->lock, flags);
1758 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1759 grp->faults[i] -= p->numa_faults_memory[i];
1760 grp->total_faults -= p->total_numa_faults;
1762 list_del(&p->numa_entry);
1764 spin_unlock_irqrestore(&grp->lock, flags);
1765 rcu_assign_pointer(p->numa_group, NULL);
1766 put_numa_group(grp);
1769 p->numa_faults_memory = NULL;
1770 p->numa_faults_buffer_memory = NULL;
1771 p->numa_faults_cpu= NULL;
1772 p->numa_faults_buffer_cpu = NULL;
1777 * Got a PROT_NONE fault for a page on @node.
1779 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1781 struct task_struct *p = current;
1782 bool migrated = flags & TNF_MIGRATED;
1783 int cpu_node = task_node(current);
1784 int local = !!(flags & TNF_FAULT_LOCAL);
1787 if (!numabalancing_enabled)
1790 /* for example, ksmd faulting in a user's mm */
1794 /* Do not worry about placement if exiting */
1795 if (p->state == TASK_DEAD)
1798 /* Allocate buffer to track faults on a per-node basis */
1799 if (unlikely(!p->numa_faults_memory)) {
1800 int size = sizeof(*p->numa_faults_memory) *
1801 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1803 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1804 if (!p->numa_faults_memory)
1807 BUG_ON(p->numa_faults_buffer_memory);
1809 * The averaged statistics, shared & private, memory & cpu,
1810 * occupy the first half of the array. The second half of the
1811 * array is for current counters, which are averaged into the
1812 * first set by task_numa_placement.
1814 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1815 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1816 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1817 p->total_numa_faults = 0;
1818 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1822 * First accesses are treated as private, otherwise consider accesses
1823 * to be private if the accessing pid has not changed
1825 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1828 priv = cpupid_match_pid(p, last_cpupid);
1829 if (!priv && !(flags & TNF_NO_GROUP))
1830 task_numa_group(p, last_cpupid, flags, &priv);
1834 * If a workload spans multiple NUMA nodes, a shared fault that
1835 * occurs wholly within the set of nodes that the workload is
1836 * actively using should be counted as local. This allows the
1837 * scan rate to slow down when a workload has settled down.
1839 if (!priv && !local && p->numa_group &&
1840 node_isset(cpu_node, p->numa_group->active_nodes) &&
1841 node_isset(mem_node, p->numa_group->active_nodes))
1844 task_numa_placement(p);
1847 * Retry task to preferred node migration periodically, in case it
1848 * case it previously failed, or the scheduler moved us.
1850 if (time_after(jiffies, p->numa_migrate_retry))
1851 numa_migrate_preferred(p);
1854 p->numa_pages_migrated += pages;
1856 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1857 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1858 p->numa_faults_locality[local] += pages;
1861 static void reset_ptenuma_scan(struct task_struct *p)
1863 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1864 p->mm->numa_scan_offset = 0;
1868 * The expensive part of numa migration is done from task_work context.
1869 * Triggered from task_tick_numa().
1871 void task_numa_work(struct callback_head *work)
1873 unsigned long migrate, next_scan, now = jiffies;
1874 struct task_struct *p = current;
1875 struct mm_struct *mm = p->mm;
1876 struct vm_area_struct *vma;
1877 unsigned long start, end;
1878 unsigned long nr_pte_updates = 0;
1881 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1883 work->next = work; /* protect against double add */
1885 * Who cares about NUMA placement when they're dying.
1887 * NOTE: make sure not to dereference p->mm before this check,
1888 * exit_task_work() happens _after_ exit_mm() so we could be called
1889 * without p->mm even though we still had it when we enqueued this
1892 if (p->flags & PF_EXITING)
1895 if (!mm->numa_next_scan) {
1896 mm->numa_next_scan = now +
1897 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1901 * Enforce maximal scan/migration frequency..
1903 migrate = mm->numa_next_scan;
1904 if (time_before(now, migrate))
1907 if (p->numa_scan_period == 0) {
1908 p->numa_scan_period_max = task_scan_max(p);
1909 p->numa_scan_period = task_scan_min(p);
1912 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1913 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1917 * Delay this task enough that another task of this mm will likely win
1918 * the next time around.
1920 p->node_stamp += 2 * TICK_NSEC;
1922 start = mm->numa_scan_offset;
1923 pages = sysctl_numa_balancing_scan_size;
1924 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1928 down_read(&mm->mmap_sem);
1929 vma = find_vma(mm, start);
1931 reset_ptenuma_scan(p);
1935 for (; vma; vma = vma->vm_next) {
1936 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1940 * Shared library pages mapped by multiple processes are not
1941 * migrated as it is expected they are cache replicated. Avoid
1942 * hinting faults in read-only file-backed mappings or the vdso
1943 * as migrating the pages will be of marginal benefit.
1946 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1950 * Skip inaccessible VMAs to avoid any confusion between
1951 * PROT_NONE and NUMA hinting ptes
1953 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1957 start = max(start, vma->vm_start);
1958 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1959 end = min(end, vma->vm_end);
1960 nr_pte_updates += change_prot_numa(vma, start, end);
1963 * Scan sysctl_numa_balancing_scan_size but ensure that
1964 * at least one PTE is updated so that unused virtual
1965 * address space is quickly skipped.
1968 pages -= (end - start) >> PAGE_SHIFT;
1975 } while (end != vma->vm_end);
1980 * It is possible to reach the end of the VMA list but the last few
1981 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1982 * would find the !migratable VMA on the next scan but not reset the
1983 * scanner to the start so check it now.
1986 mm->numa_scan_offset = start;
1988 reset_ptenuma_scan(p);
1989 up_read(&mm->mmap_sem);
1993 * Drive the periodic memory faults..
1995 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1997 struct callback_head *work = &curr->numa_work;
2001 * We don't care about NUMA placement if we don't have memory.
2003 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2007 * Using runtime rather than walltime has the dual advantage that
2008 * we (mostly) drive the selection from busy threads and that the
2009 * task needs to have done some actual work before we bother with
2012 now = curr->se.sum_exec_runtime;
2013 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2015 if (now - curr->node_stamp > period) {
2016 if (!curr->node_stamp)
2017 curr->numa_scan_period = task_scan_min(curr);
2018 curr->node_stamp += period;
2020 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2021 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2022 task_work_add(curr, work, true);
2027 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2031 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2035 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2038 #endif /* CONFIG_NUMA_BALANCING */
2041 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2043 update_load_add(&cfs_rq->load, se->load.weight);
2044 if (!parent_entity(se))
2045 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2047 if (entity_is_task(se)) {
2048 struct rq *rq = rq_of(cfs_rq);
2050 account_numa_enqueue(rq, task_of(se));
2051 list_add(&se->group_node, &rq->cfs_tasks);
2054 cfs_rq->nr_running++;
2058 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2060 update_load_sub(&cfs_rq->load, se->load.weight);
2061 if (!parent_entity(se))
2062 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2063 if (entity_is_task(se)) {
2064 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2065 list_del_init(&se->group_node);
2067 cfs_rq->nr_running--;
2070 #ifdef CONFIG_FAIR_GROUP_SCHED
2072 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2077 * Use this CPU's actual weight instead of the last load_contribution
2078 * to gain a more accurate current total weight. See
2079 * update_cfs_rq_load_contribution().
2081 tg_weight = atomic_long_read(&tg->load_avg);
2082 tg_weight -= cfs_rq->tg_load_contrib;
2083 tg_weight += cfs_rq->load.weight;
2088 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2090 long tg_weight, load, shares;
2092 tg_weight = calc_tg_weight(tg, cfs_rq);
2093 load = cfs_rq->load.weight;
2095 shares = (tg->shares * load);
2097 shares /= tg_weight;
2099 if (shares < MIN_SHARES)
2100 shares = MIN_SHARES;
2101 if (shares > tg->shares)
2102 shares = tg->shares;
2106 # else /* CONFIG_SMP */
2107 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2111 # endif /* CONFIG_SMP */
2112 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2113 unsigned long weight)
2116 /* commit outstanding execution time */
2117 if (cfs_rq->curr == se)
2118 update_curr(cfs_rq);
2119 account_entity_dequeue(cfs_rq, se);
2122 update_load_set(&se->load, weight);
2125 account_entity_enqueue(cfs_rq, se);
2128 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2130 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2132 struct task_group *tg;
2133 struct sched_entity *se;
2137 se = tg->se[cpu_of(rq_of(cfs_rq))];
2138 if (!se || throttled_hierarchy(cfs_rq))
2141 if (likely(se->load.weight == tg->shares))
2144 shares = calc_cfs_shares(cfs_rq, tg);
2146 reweight_entity(cfs_rq_of(se), se, shares);
2148 #else /* CONFIG_FAIR_GROUP_SCHED */
2149 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2152 #endif /* CONFIG_FAIR_GROUP_SCHED */
2156 * We choose a half-life close to 1 scheduling period.
2157 * Note: The tables below are dependent on this value.
2159 #define LOAD_AVG_PERIOD 32
2160 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2161 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2163 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2164 static const u32 runnable_avg_yN_inv[] = {
2165 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2166 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2167 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2168 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2169 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2170 0x85aac367, 0x82cd8698,
2174 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2175 * over-estimates when re-combining.
2177 static const u32 runnable_avg_yN_sum[] = {
2178 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2179 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2180 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2185 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2187 static __always_inline u64 decay_load(u64 val, u64 n)
2189 unsigned int local_n;
2193 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2196 /* after bounds checking we can collapse to 32-bit */
2200 * As y^PERIOD = 1/2, we can combine
2201 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2202 * With a look-up table which covers k^n (n<PERIOD)
2204 * To achieve constant time decay_load.
2206 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2207 val >>= local_n / LOAD_AVG_PERIOD;
2208 local_n %= LOAD_AVG_PERIOD;
2211 val *= runnable_avg_yN_inv[local_n];
2212 /* We don't use SRR here since we always want to round down. */
2217 * For updates fully spanning n periods, the contribution to runnable
2218 * average will be: \Sum 1024*y^n
2220 * We can compute this reasonably efficiently by combining:
2221 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2223 static u32 __compute_runnable_contrib(u64 n)
2227 if (likely(n <= LOAD_AVG_PERIOD))
2228 return runnable_avg_yN_sum[n];
2229 else if (unlikely(n >= LOAD_AVG_MAX_N))
2230 return LOAD_AVG_MAX;
2232 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2234 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2235 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2237 n -= LOAD_AVG_PERIOD;
2238 } while (n > LOAD_AVG_PERIOD);
2240 contrib = decay_load(contrib, n);
2241 return contrib + runnable_avg_yN_sum[n];
2245 * We can represent the historical contribution to runnable average as the
2246 * coefficients of a geometric series. To do this we sub-divide our runnable
2247 * history into segments of approximately 1ms (1024us); label the segment that
2248 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2250 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2252 * (now) (~1ms ago) (~2ms ago)
2254 * Let u_i denote the fraction of p_i that the entity was runnable.
2256 * We then designate the fractions u_i as our co-efficients, yielding the
2257 * following representation of historical load:
2258 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2260 * We choose y based on the with of a reasonably scheduling period, fixing:
2263 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2264 * approximately half as much as the contribution to load within the last ms
2267 * When a period "rolls over" and we have new u_0`, multiplying the previous
2268 * sum again by y is sufficient to update:
2269 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2270 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2272 static __always_inline int __update_entity_runnable_avg(u64 now,
2273 struct sched_avg *sa,
2277 u32 runnable_contrib;
2278 int delta_w, decayed = 0;
2280 delta = now - sa->last_runnable_update;
2282 * This should only happen when time goes backwards, which it
2283 * unfortunately does during sched clock init when we swap over to TSC.
2285 if ((s64)delta < 0) {
2286 sa->last_runnable_update = now;
2291 * Use 1024ns as the unit of measurement since it's a reasonable
2292 * approximation of 1us and fast to compute.
2297 sa->last_runnable_update = now;
2299 /* delta_w is the amount already accumulated against our next period */
2300 delta_w = sa->runnable_avg_period % 1024;
2301 if (delta + delta_w >= 1024) {
2302 /* period roll-over */
2306 * Now that we know we're crossing a period boundary, figure
2307 * out how much from delta we need to complete the current
2308 * period and accrue it.
2310 delta_w = 1024 - delta_w;
2312 sa->runnable_avg_sum += delta_w;
2313 sa->runnable_avg_period += delta_w;
2317 /* Figure out how many additional periods this update spans */
2318 periods = delta / 1024;
2321 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2323 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2326 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2327 runnable_contrib = __compute_runnable_contrib(periods);
2329 sa->runnable_avg_sum += runnable_contrib;
2330 sa->runnable_avg_period += runnable_contrib;
2333 /* Remainder of delta accrued against u_0` */
2335 sa->runnable_avg_sum += delta;
2336 sa->runnable_avg_period += delta;
2341 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2342 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2344 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2345 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2347 decays -= se->avg.decay_count;
2351 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2352 se->avg.decay_count = 0;
2357 #ifdef CONFIG_FAIR_GROUP_SCHED
2358 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2361 struct task_group *tg = cfs_rq->tg;
2364 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2365 tg_contrib -= cfs_rq->tg_load_contrib;
2367 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2368 atomic_long_add(tg_contrib, &tg->load_avg);
2369 cfs_rq->tg_load_contrib += tg_contrib;
2374 * Aggregate cfs_rq runnable averages into an equivalent task_group
2375 * representation for computing load contributions.
2377 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2378 struct cfs_rq *cfs_rq)
2380 struct task_group *tg = cfs_rq->tg;
2383 /* The fraction of a cpu used by this cfs_rq */
2384 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2385 sa->runnable_avg_period + 1);
2386 contrib -= cfs_rq->tg_runnable_contrib;
2388 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2389 atomic_add(contrib, &tg->runnable_avg);
2390 cfs_rq->tg_runnable_contrib += contrib;
2394 static inline void __update_group_entity_contrib(struct sched_entity *se)
2396 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2397 struct task_group *tg = cfs_rq->tg;
2402 contrib = cfs_rq->tg_load_contrib * tg->shares;
2403 se->avg.load_avg_contrib = div_u64(contrib,
2404 atomic_long_read(&tg->load_avg) + 1);
2407 * For group entities we need to compute a correction term in the case
2408 * that they are consuming <1 cpu so that we would contribute the same
2409 * load as a task of equal weight.
2411 * Explicitly co-ordinating this measurement would be expensive, but
2412 * fortunately the sum of each cpus contribution forms a usable
2413 * lower-bound on the true value.
2415 * Consider the aggregate of 2 contributions. Either they are disjoint
2416 * (and the sum represents true value) or they are disjoint and we are
2417 * understating by the aggregate of their overlap.
2419 * Extending this to N cpus, for a given overlap, the maximum amount we
2420 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2421 * cpus that overlap for this interval and w_i is the interval width.
2423 * On a small machine; the first term is well-bounded which bounds the
2424 * total error since w_i is a subset of the period. Whereas on a
2425 * larger machine, while this first term can be larger, if w_i is the
2426 * of consequential size guaranteed to see n_i*w_i quickly converge to
2427 * our upper bound of 1-cpu.
2429 runnable_avg = atomic_read(&tg->runnable_avg);
2430 if (runnable_avg < NICE_0_LOAD) {
2431 se->avg.load_avg_contrib *= runnable_avg;
2432 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2436 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2438 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2439 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2441 #else /* CONFIG_FAIR_GROUP_SCHED */
2442 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2443 int force_update) {}
2444 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2445 struct cfs_rq *cfs_rq) {}
2446 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2447 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2448 #endif /* CONFIG_FAIR_GROUP_SCHED */
2450 static inline void __update_task_entity_contrib(struct sched_entity *se)
2454 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2455 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2456 contrib /= (se->avg.runnable_avg_period + 1);
2457 se->avg.load_avg_contrib = scale_load(contrib);
2460 /* Compute the current contribution to load_avg by se, return any delta */
2461 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2463 long old_contrib = se->avg.load_avg_contrib;
2465 if (entity_is_task(se)) {
2466 __update_task_entity_contrib(se);
2468 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2469 __update_group_entity_contrib(se);
2472 return se->avg.load_avg_contrib - old_contrib;
2475 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2478 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2479 cfs_rq->blocked_load_avg -= load_contrib;
2481 cfs_rq->blocked_load_avg = 0;
2484 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2486 /* Update a sched_entity's runnable average */
2487 static inline void update_entity_load_avg(struct sched_entity *se,
2490 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2495 * For a group entity we need to use their owned cfs_rq_clock_task() in
2496 * case they are the parent of a throttled hierarchy.
2498 if (entity_is_task(se))
2499 now = cfs_rq_clock_task(cfs_rq);
2501 now = cfs_rq_clock_task(group_cfs_rq(se));
2503 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2506 contrib_delta = __update_entity_load_avg_contrib(se);
2512 cfs_rq->runnable_load_avg += contrib_delta;
2514 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2518 * Decay the load contributed by all blocked children and account this so that
2519 * their contribution may appropriately discounted when they wake up.
2521 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2523 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2526 decays = now - cfs_rq->last_decay;
2527 if (!decays && !force_update)
2530 if (atomic_long_read(&cfs_rq->removed_load)) {
2531 unsigned long removed_load;
2532 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2533 subtract_blocked_load_contrib(cfs_rq, removed_load);
2537 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2539 atomic64_add(decays, &cfs_rq->decay_counter);
2540 cfs_rq->last_decay = now;
2543 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2546 /* Add the load generated by se into cfs_rq's child load-average */
2547 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2548 struct sched_entity *se,
2552 * We track migrations using entity decay_count <= 0, on a wake-up
2553 * migration we use a negative decay count to track the remote decays
2554 * accumulated while sleeping.
2556 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2557 * are seen by enqueue_entity_load_avg() as a migration with an already
2558 * constructed load_avg_contrib.
2560 if (unlikely(se->avg.decay_count <= 0)) {
2561 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2562 if (se->avg.decay_count) {
2564 * In a wake-up migration we have to approximate the
2565 * time sleeping. This is because we can't synchronize
2566 * clock_task between the two cpus, and it is not
2567 * guaranteed to be read-safe. Instead, we can
2568 * approximate this using our carried decays, which are
2569 * explicitly atomically readable.
2571 se->avg.last_runnable_update -= (-se->avg.decay_count)
2573 update_entity_load_avg(se, 0);
2574 /* Indicate that we're now synchronized and on-rq */
2575 se->avg.decay_count = 0;
2579 __synchronize_entity_decay(se);
2582 /* migrated tasks did not contribute to our blocked load */
2584 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2585 update_entity_load_avg(se, 0);
2588 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2589 /* we force update consideration on load-balancer moves */
2590 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2594 * Remove se's load from this cfs_rq child load-average, if the entity is
2595 * transitioning to a blocked state we track its projected decay using
2598 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2599 struct sched_entity *se,
2602 update_entity_load_avg(se, 1);
2603 /* we force update consideration on load-balancer moves */
2604 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2606 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2608 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2609 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2610 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2614 * Update the rq's load with the elapsed running time before entering
2615 * idle. if the last scheduled task is not a CFS task, idle_enter will
2616 * be the only way to update the runnable statistic.
2618 void idle_enter_fair(struct rq *this_rq)
2620 update_rq_runnable_avg(this_rq, 1);
2624 * Update the rq's load with the elapsed idle time before a task is
2625 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2626 * be the only way to update the runnable statistic.
2628 void idle_exit_fair(struct rq *this_rq)
2630 update_rq_runnable_avg(this_rq, 0);
2633 static int idle_balance(struct rq *this_rq);
2635 #else /* CONFIG_SMP */
2637 static inline void update_entity_load_avg(struct sched_entity *se,
2638 int update_cfs_rq) {}
2639 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2640 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2641 struct sched_entity *se,
2643 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2644 struct sched_entity *se,
2646 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2647 int force_update) {}
2649 static inline int idle_balance(struct rq *rq)
2654 #endif /* CONFIG_SMP */
2656 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2658 #ifdef CONFIG_SCHEDSTATS
2659 struct task_struct *tsk = NULL;
2661 if (entity_is_task(se))
2664 if (se->statistics.sleep_start) {
2665 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2670 if (unlikely(delta > se->statistics.sleep_max))
2671 se->statistics.sleep_max = delta;
2673 se->statistics.sleep_start = 0;
2674 se->statistics.sum_sleep_runtime += delta;
2677 account_scheduler_latency(tsk, delta >> 10, 1);
2678 trace_sched_stat_sleep(tsk, delta);
2681 if (se->statistics.block_start) {
2682 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2687 if (unlikely(delta > se->statistics.block_max))
2688 se->statistics.block_max = delta;
2690 se->statistics.block_start = 0;
2691 se->statistics.sum_sleep_runtime += delta;
2694 if (tsk->in_iowait) {
2695 se->statistics.iowait_sum += delta;
2696 se->statistics.iowait_count++;
2697 trace_sched_stat_iowait(tsk, delta);
2700 trace_sched_stat_blocked(tsk, delta);
2703 * Blocking time is in units of nanosecs, so shift by
2704 * 20 to get a milliseconds-range estimation of the
2705 * amount of time that the task spent sleeping:
2707 if (unlikely(prof_on == SLEEP_PROFILING)) {
2708 profile_hits(SLEEP_PROFILING,
2709 (void *)get_wchan(tsk),
2712 account_scheduler_latency(tsk, delta >> 10, 0);
2718 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2720 #ifdef CONFIG_SCHED_DEBUG
2721 s64 d = se->vruntime - cfs_rq->min_vruntime;
2726 if (d > 3*sysctl_sched_latency)
2727 schedstat_inc(cfs_rq, nr_spread_over);
2732 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2734 u64 vruntime = cfs_rq->min_vruntime;
2737 * The 'current' period is already promised to the current tasks,
2738 * however the extra weight of the new task will slow them down a
2739 * little, place the new task so that it fits in the slot that
2740 * stays open at the end.
2742 if (initial && sched_feat(START_DEBIT))
2743 vruntime += sched_vslice(cfs_rq, se);
2745 /* sleeps up to a single latency don't count. */
2747 unsigned long thresh = sysctl_sched_latency;
2750 * Halve their sleep time's effect, to allow
2751 * for a gentler effect of sleepers:
2753 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2759 /* ensure we never gain time by being placed backwards. */
2760 se->vruntime = max_vruntime(se->vruntime, vruntime);
2763 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2766 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2769 * Update the normalized vruntime before updating min_vruntime
2770 * through calling update_curr().
2772 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2773 se->vruntime += cfs_rq->min_vruntime;
2776 * Update run-time statistics of the 'current'.
2778 update_curr(cfs_rq);
2779 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2780 account_entity_enqueue(cfs_rq, se);
2781 update_cfs_shares(cfs_rq);
2783 if (flags & ENQUEUE_WAKEUP) {
2784 place_entity(cfs_rq, se, 0);
2785 enqueue_sleeper(cfs_rq, se);
2788 update_stats_enqueue(cfs_rq, se);
2789 check_spread(cfs_rq, se);
2790 if (se != cfs_rq->curr)
2791 __enqueue_entity(cfs_rq, se);
2794 if (cfs_rq->nr_running == 1) {
2795 list_add_leaf_cfs_rq(cfs_rq);
2796 check_enqueue_throttle(cfs_rq);
2800 static void __clear_buddies_last(struct sched_entity *se)
2802 for_each_sched_entity(se) {
2803 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2804 if (cfs_rq->last != se)
2807 cfs_rq->last = NULL;
2811 static void __clear_buddies_next(struct sched_entity *se)
2813 for_each_sched_entity(se) {
2814 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2815 if (cfs_rq->next != se)
2818 cfs_rq->next = NULL;
2822 static void __clear_buddies_skip(struct sched_entity *se)
2824 for_each_sched_entity(se) {
2825 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2826 if (cfs_rq->skip != se)
2829 cfs_rq->skip = NULL;
2833 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2835 if (cfs_rq->last == se)
2836 __clear_buddies_last(se);
2838 if (cfs_rq->next == se)
2839 __clear_buddies_next(se);
2841 if (cfs_rq->skip == se)
2842 __clear_buddies_skip(se);
2845 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2848 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2851 * Update run-time statistics of the 'current'.
2853 update_curr(cfs_rq);
2854 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2856 update_stats_dequeue(cfs_rq, se);
2857 if (flags & DEQUEUE_SLEEP) {
2858 #ifdef CONFIG_SCHEDSTATS
2859 if (entity_is_task(se)) {
2860 struct task_struct *tsk = task_of(se);
2862 if (tsk->state & TASK_INTERRUPTIBLE)
2863 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2864 if (tsk->state & TASK_UNINTERRUPTIBLE)
2865 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2870 clear_buddies(cfs_rq, se);
2872 if (se != cfs_rq->curr)
2873 __dequeue_entity(cfs_rq, se);
2875 account_entity_dequeue(cfs_rq, se);
2878 * Normalize the entity after updating the min_vruntime because the
2879 * update can refer to the ->curr item and we need to reflect this
2880 * movement in our normalized position.
2882 if (!(flags & DEQUEUE_SLEEP))
2883 se->vruntime -= cfs_rq->min_vruntime;
2885 /* return excess runtime on last dequeue */
2886 return_cfs_rq_runtime(cfs_rq);
2888 update_min_vruntime(cfs_rq);
2889 update_cfs_shares(cfs_rq);
2893 * Preempt the current task with a newly woken task if needed:
2896 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2898 unsigned long ideal_runtime, delta_exec;
2899 struct sched_entity *se;
2902 ideal_runtime = sched_slice(cfs_rq, curr);
2903 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2904 if (delta_exec > ideal_runtime) {
2905 resched_task(rq_of(cfs_rq)->curr);
2907 * The current task ran long enough, ensure it doesn't get
2908 * re-elected due to buddy favours.
2910 clear_buddies(cfs_rq, curr);
2915 * Ensure that a task that missed wakeup preemption by a
2916 * narrow margin doesn't have to wait for a full slice.
2917 * This also mitigates buddy induced latencies under load.
2919 if (delta_exec < sysctl_sched_min_granularity)
2922 se = __pick_first_entity(cfs_rq);
2923 delta = curr->vruntime - se->vruntime;
2928 if (delta > ideal_runtime)
2929 resched_task(rq_of(cfs_rq)->curr);
2933 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2935 /* 'current' is not kept within the tree. */
2938 * Any task has to be enqueued before it get to execute on
2939 * a CPU. So account for the time it spent waiting on the
2942 update_stats_wait_end(cfs_rq, se);
2943 __dequeue_entity(cfs_rq, se);
2946 update_stats_curr_start(cfs_rq, se);
2948 #ifdef CONFIG_SCHEDSTATS
2950 * Track our maximum slice length, if the CPU's load is at
2951 * least twice that of our own weight (i.e. dont track it
2952 * when there are only lesser-weight tasks around):
2954 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2955 se->statistics.slice_max = max(se->statistics.slice_max,
2956 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2959 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2963 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2966 * Pick the next process, keeping these things in mind, in this order:
2967 * 1) keep things fair between processes/task groups
2968 * 2) pick the "next" process, since someone really wants that to run
2969 * 3) pick the "last" process, for cache locality
2970 * 4) do not run the "skip" process, if something else is available
2972 static struct sched_entity *
2973 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2975 struct sched_entity *left = __pick_first_entity(cfs_rq);
2976 struct sched_entity *se;
2979 * If curr is set we have to see if its left of the leftmost entity
2980 * still in the tree, provided there was anything in the tree at all.
2982 if (!left || (curr && entity_before(curr, left)))
2985 se = left; /* ideally we run the leftmost entity */
2988 * Avoid running the skip buddy, if running something else can
2989 * be done without getting too unfair.
2991 if (cfs_rq->skip == se) {
2992 struct sched_entity *second;
2995 second = __pick_first_entity(cfs_rq);
2997 second = __pick_next_entity(se);
2998 if (!second || (curr && entity_before(curr, second)))
3002 if (second && wakeup_preempt_entity(second, left) < 1)
3007 * Prefer last buddy, try to return the CPU to a preempted task.
3009 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3013 * Someone really wants this to run. If it's not unfair, run it.
3015 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3018 clear_buddies(cfs_rq, se);
3023 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3025 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3028 * If still on the runqueue then deactivate_task()
3029 * was not called and update_curr() has to be done:
3032 update_curr(cfs_rq);
3034 /* throttle cfs_rqs exceeding runtime */
3035 check_cfs_rq_runtime(cfs_rq);
3037 check_spread(cfs_rq, prev);
3039 update_stats_wait_start(cfs_rq, prev);
3040 /* Put 'current' back into the tree. */
3041 __enqueue_entity(cfs_rq, prev);
3042 /* in !on_rq case, update occurred at dequeue */
3043 update_entity_load_avg(prev, 1);
3045 cfs_rq->curr = NULL;
3049 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3052 * Update run-time statistics of the 'current'.
3054 update_curr(cfs_rq);
3057 * Ensure that runnable average is periodically updated.
3059 update_entity_load_avg(curr, 1);
3060 update_cfs_rq_blocked_load(cfs_rq, 1);
3061 update_cfs_shares(cfs_rq);
3063 #ifdef CONFIG_SCHED_HRTICK
3065 * queued ticks are scheduled to match the slice, so don't bother
3066 * validating it and just reschedule.
3069 resched_task(rq_of(cfs_rq)->curr);
3073 * don't let the period tick interfere with the hrtick preemption
3075 if (!sched_feat(DOUBLE_TICK) &&
3076 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3080 if (cfs_rq->nr_running > 1)
3081 check_preempt_tick(cfs_rq, curr);
3085 /**************************************************
3086 * CFS bandwidth control machinery
3089 #ifdef CONFIG_CFS_BANDWIDTH
3091 #ifdef HAVE_JUMP_LABEL
3092 static struct static_key __cfs_bandwidth_used;
3094 static inline bool cfs_bandwidth_used(void)
3096 return static_key_false(&__cfs_bandwidth_used);
3099 void cfs_bandwidth_usage_inc(void)
3101 static_key_slow_inc(&__cfs_bandwidth_used);
3104 void cfs_bandwidth_usage_dec(void)
3106 static_key_slow_dec(&__cfs_bandwidth_used);
3108 #else /* HAVE_JUMP_LABEL */
3109 static bool cfs_bandwidth_used(void)
3114 void cfs_bandwidth_usage_inc(void) {}
3115 void cfs_bandwidth_usage_dec(void) {}
3116 #endif /* HAVE_JUMP_LABEL */
3119 * default period for cfs group bandwidth.
3120 * default: 0.1s, units: nanoseconds
3122 static inline u64 default_cfs_period(void)
3124 return 100000000ULL;
3127 static inline u64 sched_cfs_bandwidth_slice(void)
3129 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3133 * Replenish runtime according to assigned quota and update expiration time.
3134 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3135 * additional synchronization around rq->lock.
3137 * requires cfs_b->lock
3139 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3143 if (cfs_b->quota == RUNTIME_INF)
3146 now = sched_clock_cpu(smp_processor_id());
3147 cfs_b->runtime = cfs_b->quota;
3148 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3151 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3153 return &tg->cfs_bandwidth;
3156 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3157 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3159 if (unlikely(cfs_rq->throttle_count))
3160 return cfs_rq->throttled_clock_task;
3162 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3165 /* returns 0 on failure to allocate runtime */
3166 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3168 struct task_group *tg = cfs_rq->tg;
3169 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3170 u64 amount = 0, min_amount, expires;
3172 /* note: this is a positive sum as runtime_remaining <= 0 */
3173 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3175 raw_spin_lock(&cfs_b->lock);
3176 if (cfs_b->quota == RUNTIME_INF)
3177 amount = min_amount;
3180 * If the bandwidth pool has become inactive, then at least one
3181 * period must have elapsed since the last consumption.
3182 * Refresh the global state and ensure bandwidth timer becomes
3185 if (!cfs_b->timer_active) {
3186 __refill_cfs_bandwidth_runtime(cfs_b);
3187 __start_cfs_bandwidth(cfs_b, false);
3190 if (cfs_b->runtime > 0) {
3191 amount = min(cfs_b->runtime, min_amount);
3192 cfs_b->runtime -= amount;
3196 expires = cfs_b->runtime_expires;
3197 raw_spin_unlock(&cfs_b->lock);
3199 cfs_rq->runtime_remaining += amount;
3201 * we may have advanced our local expiration to account for allowed
3202 * spread between our sched_clock and the one on which runtime was
3205 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3206 cfs_rq->runtime_expires = expires;
3208 return cfs_rq->runtime_remaining > 0;
3212 * Note: This depends on the synchronization provided by sched_clock and the
3213 * fact that rq->clock snapshots this value.
3215 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3217 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3219 /* if the deadline is ahead of our clock, nothing to do */
3220 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3223 if (cfs_rq->runtime_remaining < 0)
3227 * If the local deadline has passed we have to consider the
3228 * possibility that our sched_clock is 'fast' and the global deadline
3229 * has not truly expired.
3231 * Fortunately we can check determine whether this the case by checking
3232 * whether the global deadline has advanced. It is valid to compare
3233 * cfs_b->runtime_expires without any locks since we only care about
3234 * exact equality, so a partial write will still work.
3237 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3238 /* extend local deadline, drift is bounded above by 2 ticks */
3239 cfs_rq->runtime_expires += TICK_NSEC;
3241 /* global deadline is ahead, expiration has passed */
3242 cfs_rq->runtime_remaining = 0;
3246 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3248 /* dock delta_exec before expiring quota (as it could span periods) */
3249 cfs_rq->runtime_remaining -= delta_exec;
3250 expire_cfs_rq_runtime(cfs_rq);
3252 if (likely(cfs_rq->runtime_remaining > 0))
3256 * if we're unable to extend our runtime we resched so that the active
3257 * hierarchy can be throttled
3259 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3260 resched_task(rq_of(cfs_rq)->curr);
3263 static __always_inline
3264 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3266 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3269 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3272 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3274 return cfs_bandwidth_used() && cfs_rq->throttled;
3277 /* check whether cfs_rq, or any parent, is throttled */
3278 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3280 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3284 * Ensure that neither of the group entities corresponding to src_cpu or
3285 * dest_cpu are members of a throttled hierarchy when performing group
3286 * load-balance operations.
3288 static inline int throttled_lb_pair(struct task_group *tg,
3289 int src_cpu, int dest_cpu)
3291 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3293 src_cfs_rq = tg->cfs_rq[src_cpu];
3294 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3296 return throttled_hierarchy(src_cfs_rq) ||
3297 throttled_hierarchy(dest_cfs_rq);
3300 /* updated child weight may affect parent so we have to do this bottom up */
3301 static int tg_unthrottle_up(struct task_group *tg, void *data)
3303 struct rq *rq = data;
3304 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3306 cfs_rq->throttle_count--;
3308 if (!cfs_rq->throttle_count) {
3309 /* adjust cfs_rq_clock_task() */
3310 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3311 cfs_rq->throttled_clock_task;
3318 static int tg_throttle_down(struct task_group *tg, void *data)
3320 struct rq *rq = data;
3321 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3323 /* group is entering throttled state, stop time */
3324 if (!cfs_rq->throttle_count)
3325 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3326 cfs_rq->throttle_count++;
3331 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3333 struct rq *rq = rq_of(cfs_rq);
3334 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3335 struct sched_entity *se;
3336 long task_delta, dequeue = 1;
3338 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3340 /* freeze hierarchy runnable averages while throttled */
3342 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3345 task_delta = cfs_rq->h_nr_running;
3346 for_each_sched_entity(se) {
3347 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3348 /* throttled entity or throttle-on-deactivate */
3353 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3354 qcfs_rq->h_nr_running -= task_delta;
3356 if (qcfs_rq->load.weight)
3361 sub_nr_running(rq, task_delta);
3363 cfs_rq->throttled = 1;
3364 cfs_rq->throttled_clock = rq_clock(rq);
3365 raw_spin_lock(&cfs_b->lock);
3367 * Add to the _head_ of the list, so that an already-started
3368 * distribute_cfs_runtime will not see us
3370 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3371 if (!cfs_b->timer_active)
3372 __start_cfs_bandwidth(cfs_b, false);
3373 raw_spin_unlock(&cfs_b->lock);
3376 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3378 struct rq *rq = rq_of(cfs_rq);
3379 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3380 struct sched_entity *se;
3384 se = cfs_rq->tg->se[cpu_of(rq)];
3386 cfs_rq->throttled = 0;
3388 update_rq_clock(rq);
3390 raw_spin_lock(&cfs_b->lock);
3391 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3392 list_del_rcu(&cfs_rq->throttled_list);
3393 raw_spin_unlock(&cfs_b->lock);
3395 /* update hierarchical throttle state */
3396 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3398 if (!cfs_rq->load.weight)
3401 task_delta = cfs_rq->h_nr_running;
3402 for_each_sched_entity(se) {
3406 cfs_rq = cfs_rq_of(se);
3408 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3409 cfs_rq->h_nr_running += task_delta;
3411 if (cfs_rq_throttled(cfs_rq))
3416 add_nr_running(rq, task_delta);
3418 /* determine whether we need to wake up potentially idle cpu */
3419 if (rq->curr == rq->idle && rq->cfs.nr_running)
3420 resched_task(rq->curr);
3423 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3424 u64 remaining, u64 expires)
3426 struct cfs_rq *cfs_rq;
3428 u64 starting_runtime = remaining;
3431 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3433 struct rq *rq = rq_of(cfs_rq);
3435 raw_spin_lock(&rq->lock);
3436 if (!cfs_rq_throttled(cfs_rq))
3439 runtime = -cfs_rq->runtime_remaining + 1;
3440 if (runtime > remaining)
3441 runtime = remaining;
3442 remaining -= runtime;
3444 cfs_rq->runtime_remaining += runtime;
3445 cfs_rq->runtime_expires = expires;
3447 /* we check whether we're throttled above */
3448 if (cfs_rq->runtime_remaining > 0)
3449 unthrottle_cfs_rq(cfs_rq);
3452 raw_spin_unlock(&rq->lock);
3459 return starting_runtime - remaining;
3463 * Responsible for refilling a task_group's bandwidth and unthrottling its
3464 * cfs_rqs as appropriate. If there has been no activity within the last
3465 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3466 * used to track this state.
3468 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3470 u64 runtime, runtime_expires;
3473 /* no need to continue the timer with no bandwidth constraint */
3474 if (cfs_b->quota == RUNTIME_INF)
3475 goto out_deactivate;
3477 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3478 cfs_b->nr_periods += overrun;
3481 * idle depends on !throttled (for the case of a large deficit), and if
3482 * we're going inactive then everything else can be deferred
3484 if (cfs_b->idle && !throttled)
3485 goto out_deactivate;
3488 * if we have relooped after returning idle once, we need to update our
3489 * status as actually running, so that other cpus doing
3490 * __start_cfs_bandwidth will stop trying to cancel us.
3492 cfs_b->timer_active = 1;
3494 __refill_cfs_bandwidth_runtime(cfs_b);
3497 /* mark as potentially idle for the upcoming period */
3502 /* account preceding periods in which throttling occurred */
3503 cfs_b->nr_throttled += overrun;
3505 runtime_expires = cfs_b->runtime_expires;
3508 * This check is repeated as we are holding onto the new bandwidth while
3509 * we unthrottle. This can potentially race with an unthrottled group
3510 * trying to acquire new bandwidth from the global pool. This can result
3511 * in us over-using our runtime if it is all used during this loop, but
3512 * only by limited amounts in that extreme case.
3514 while (throttled && cfs_b->runtime > 0) {
3515 runtime = cfs_b->runtime;
3516 raw_spin_unlock(&cfs_b->lock);
3517 /* we can't nest cfs_b->lock while distributing bandwidth */
3518 runtime = distribute_cfs_runtime(cfs_b, runtime,
3520 raw_spin_lock(&cfs_b->lock);
3522 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3524 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3528 * While we are ensured activity in the period following an
3529 * unthrottle, this also covers the case in which the new bandwidth is
3530 * insufficient to cover the existing bandwidth deficit. (Forcing the
3531 * timer to remain active while there are any throttled entities.)
3538 cfs_b->timer_active = 0;
3542 /* a cfs_rq won't donate quota below this amount */
3543 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3544 /* minimum remaining period time to redistribute slack quota */
3545 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3546 /* how long we wait to gather additional slack before distributing */
3547 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3550 * Are we near the end of the current quota period?
3552 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3553 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3554 * migrate_hrtimers, base is never cleared, so we are fine.
3556 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3558 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3561 /* if the call-back is running a quota refresh is already occurring */
3562 if (hrtimer_callback_running(refresh_timer))
3565 /* is a quota refresh about to occur? */
3566 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3567 if (remaining < min_expire)
3573 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3575 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3577 /* if there's a quota refresh soon don't bother with slack */
3578 if (runtime_refresh_within(cfs_b, min_left))
3581 start_bandwidth_timer(&cfs_b->slack_timer,
3582 ns_to_ktime(cfs_bandwidth_slack_period));
3585 /* we know any runtime found here is valid as update_curr() precedes return */
3586 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3588 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3589 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3591 if (slack_runtime <= 0)
3594 raw_spin_lock(&cfs_b->lock);
3595 if (cfs_b->quota != RUNTIME_INF &&
3596 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3597 cfs_b->runtime += slack_runtime;
3599 /* we are under rq->lock, defer unthrottling using a timer */
3600 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3601 !list_empty(&cfs_b->throttled_cfs_rq))
3602 start_cfs_slack_bandwidth(cfs_b);
3604 raw_spin_unlock(&cfs_b->lock);
3606 /* even if it's not valid for return we don't want to try again */
3607 cfs_rq->runtime_remaining -= slack_runtime;
3610 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3612 if (!cfs_bandwidth_used())
3615 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3618 __return_cfs_rq_runtime(cfs_rq);
3622 * This is done with a timer (instead of inline with bandwidth return) since
3623 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3625 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3627 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3630 /* confirm we're still not at a refresh boundary */
3631 raw_spin_lock(&cfs_b->lock);
3632 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3633 raw_spin_unlock(&cfs_b->lock);
3637 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3638 runtime = cfs_b->runtime;
3640 expires = cfs_b->runtime_expires;
3641 raw_spin_unlock(&cfs_b->lock);
3646 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3648 raw_spin_lock(&cfs_b->lock);
3649 if (expires == cfs_b->runtime_expires)
3650 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3651 raw_spin_unlock(&cfs_b->lock);
3655 * When a group wakes up we want to make sure that its quota is not already
3656 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3657 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3659 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3661 if (!cfs_bandwidth_used())
3664 /* an active group must be handled by the update_curr()->put() path */
3665 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3668 /* ensure the group is not already throttled */
3669 if (cfs_rq_throttled(cfs_rq))
3672 /* update runtime allocation */
3673 account_cfs_rq_runtime(cfs_rq, 0);
3674 if (cfs_rq->runtime_remaining <= 0)
3675 throttle_cfs_rq(cfs_rq);
3678 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3679 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3681 if (!cfs_bandwidth_used())
3684 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3688 * it's possible for a throttled entity to be forced into a running
3689 * state (e.g. set_curr_task), in this case we're finished.
3691 if (cfs_rq_throttled(cfs_rq))
3694 throttle_cfs_rq(cfs_rq);
3698 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3700 struct cfs_bandwidth *cfs_b =
3701 container_of(timer, struct cfs_bandwidth, slack_timer);
3702 do_sched_cfs_slack_timer(cfs_b);
3704 return HRTIMER_NORESTART;
3707 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3709 struct cfs_bandwidth *cfs_b =
3710 container_of(timer, struct cfs_bandwidth, period_timer);
3715 raw_spin_lock(&cfs_b->lock);
3717 now = hrtimer_cb_get_time(timer);
3718 overrun = hrtimer_forward(timer, now, cfs_b->period);
3723 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3725 raw_spin_unlock(&cfs_b->lock);
3727 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3730 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3732 raw_spin_lock_init(&cfs_b->lock);
3734 cfs_b->quota = RUNTIME_INF;
3735 cfs_b->period = ns_to_ktime(default_cfs_period());
3737 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3738 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3739 cfs_b->period_timer.function = sched_cfs_period_timer;
3740 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3741 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3744 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3746 cfs_rq->runtime_enabled = 0;
3747 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3750 /* requires cfs_b->lock, may release to reprogram timer */
3751 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force)
3754 * The timer may be active because we're trying to set a new bandwidth
3755 * period or because we're racing with the tear-down path
3756 * (timer_active==0 becomes visible before the hrtimer call-back
3757 * terminates). In either case we ensure that it's re-programmed
3759 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3760 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3761 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3762 raw_spin_unlock(&cfs_b->lock);
3764 raw_spin_lock(&cfs_b->lock);
3765 /* if someone else restarted the timer then we're done */
3766 if (!force && cfs_b->timer_active)
3770 cfs_b->timer_active = 1;
3771 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3774 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3776 hrtimer_cancel(&cfs_b->period_timer);
3777 hrtimer_cancel(&cfs_b->slack_timer);
3780 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3782 struct cfs_rq *cfs_rq;
3784 for_each_leaf_cfs_rq(rq, cfs_rq) {
3785 if (!cfs_rq->runtime_enabled)
3789 * clock_task is not advancing so we just need to make sure
3790 * there's some valid quota amount
3792 cfs_rq->runtime_remaining = 1;
3793 if (cfs_rq_throttled(cfs_rq))
3794 unthrottle_cfs_rq(cfs_rq);
3798 #else /* CONFIG_CFS_BANDWIDTH */
3799 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3801 return rq_clock_task(rq_of(cfs_rq));
3804 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3805 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3806 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3807 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3809 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3814 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3819 static inline int throttled_lb_pair(struct task_group *tg,
3820 int src_cpu, int dest_cpu)
3825 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3827 #ifdef CONFIG_FAIR_GROUP_SCHED
3828 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3831 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3835 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3836 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3838 #endif /* CONFIG_CFS_BANDWIDTH */
3840 /**************************************************
3841 * CFS operations on tasks:
3844 #ifdef CONFIG_SCHED_HRTICK
3845 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3847 struct sched_entity *se = &p->se;
3848 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3850 WARN_ON(task_rq(p) != rq);
3852 if (cfs_rq->nr_running > 1) {
3853 u64 slice = sched_slice(cfs_rq, se);
3854 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3855 s64 delta = slice - ran;
3864 * Don't schedule slices shorter than 10000ns, that just
3865 * doesn't make sense. Rely on vruntime for fairness.
3868 delta = max_t(s64, 10000LL, delta);
3870 hrtick_start(rq, delta);
3875 * called from enqueue/dequeue and updates the hrtick when the
3876 * current task is from our class and nr_running is low enough
3879 static void hrtick_update(struct rq *rq)
3881 struct task_struct *curr = rq->curr;
3883 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3886 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3887 hrtick_start_fair(rq, curr);
3889 #else /* !CONFIG_SCHED_HRTICK */
3891 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3895 static inline void hrtick_update(struct rq *rq)
3901 * The enqueue_task method is called before nr_running is
3902 * increased. Here we update the fair scheduling stats and
3903 * then put the task into the rbtree:
3906 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3908 struct cfs_rq *cfs_rq;
3909 struct sched_entity *se = &p->se;
3911 for_each_sched_entity(se) {
3914 cfs_rq = cfs_rq_of(se);
3915 enqueue_entity(cfs_rq, se, flags);
3918 * end evaluation on encountering a throttled cfs_rq
3920 * note: in the case of encountering a throttled cfs_rq we will
3921 * post the final h_nr_running increment below.
3923 if (cfs_rq_throttled(cfs_rq))
3925 cfs_rq->h_nr_running++;
3927 flags = ENQUEUE_WAKEUP;
3930 for_each_sched_entity(se) {
3931 cfs_rq = cfs_rq_of(se);
3932 cfs_rq->h_nr_running++;
3934 if (cfs_rq_throttled(cfs_rq))
3937 update_cfs_shares(cfs_rq);
3938 update_entity_load_avg(se, 1);
3942 update_rq_runnable_avg(rq, rq->nr_running);
3943 add_nr_running(rq, 1);
3948 static void set_next_buddy(struct sched_entity *se);
3951 * The dequeue_task method is called before nr_running is
3952 * decreased. We remove the task from the rbtree and
3953 * update the fair scheduling stats:
3955 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3957 struct cfs_rq *cfs_rq;
3958 struct sched_entity *se = &p->se;
3959 int task_sleep = flags & DEQUEUE_SLEEP;
3961 for_each_sched_entity(se) {
3962 cfs_rq = cfs_rq_of(se);
3963 dequeue_entity(cfs_rq, se, flags);
3966 * end evaluation on encountering a throttled cfs_rq
3968 * note: in the case of encountering a throttled cfs_rq we will
3969 * post the final h_nr_running decrement below.
3971 if (cfs_rq_throttled(cfs_rq))
3973 cfs_rq->h_nr_running--;
3975 /* Don't dequeue parent if it has other entities besides us */
3976 if (cfs_rq->load.weight) {
3978 * Bias pick_next to pick a task from this cfs_rq, as
3979 * p is sleeping when it is within its sched_slice.
3981 if (task_sleep && parent_entity(se))
3982 set_next_buddy(parent_entity(se));
3984 /* avoid re-evaluating load for this entity */
3985 se = parent_entity(se);
3988 flags |= DEQUEUE_SLEEP;
3991 for_each_sched_entity(se) {
3992 cfs_rq = cfs_rq_of(se);
3993 cfs_rq->h_nr_running--;
3995 if (cfs_rq_throttled(cfs_rq))
3998 update_cfs_shares(cfs_rq);
3999 update_entity_load_avg(se, 1);
4003 sub_nr_running(rq, 1);
4004 update_rq_runnable_avg(rq, 1);
4010 /* Used instead of source_load when we know the type == 0 */
4011 static unsigned long weighted_cpuload(const int cpu)
4013 return cpu_rq(cpu)->cfs.runnable_load_avg;
4017 * Return a low guess at the load of a migration-source cpu weighted
4018 * according to the scheduling class and "nice" value.
4020 * We want to under-estimate the load of migration sources, to
4021 * balance conservatively.
4023 static unsigned long source_load(int cpu, int type)
4025 struct rq *rq = cpu_rq(cpu);
4026 unsigned long total = weighted_cpuload(cpu);
4028 if (type == 0 || !sched_feat(LB_BIAS))
4031 return min(rq->cpu_load[type-1], total);
4035 * Return a high guess at the load of a migration-target cpu weighted
4036 * according to the scheduling class and "nice" value.
4038 static unsigned long target_load(int cpu, int type)
4040 struct rq *rq = cpu_rq(cpu);
4041 unsigned long total = weighted_cpuload(cpu);
4043 if (type == 0 || !sched_feat(LB_BIAS))
4046 return max(rq->cpu_load[type-1], total);
4049 static unsigned long capacity_of(int cpu)
4051 return cpu_rq(cpu)->cpu_capacity;
4054 static unsigned long cpu_avg_load_per_task(int cpu)
4056 struct rq *rq = cpu_rq(cpu);
4057 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
4058 unsigned long load_avg = rq->cfs.runnable_load_avg;
4061 return load_avg / nr_running;
4066 static void record_wakee(struct task_struct *p)
4069 * Rough decay (wiping) for cost saving, don't worry
4070 * about the boundary, really active task won't care
4073 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4074 current->wakee_flips >>= 1;
4075 current->wakee_flip_decay_ts = jiffies;
4078 if (current->last_wakee != p) {
4079 current->last_wakee = p;
4080 current->wakee_flips++;
4084 static void task_waking_fair(struct task_struct *p)
4086 struct sched_entity *se = &p->se;
4087 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4090 #ifndef CONFIG_64BIT
4091 u64 min_vruntime_copy;
4094 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4096 min_vruntime = cfs_rq->min_vruntime;
4097 } while (min_vruntime != min_vruntime_copy);
4099 min_vruntime = cfs_rq->min_vruntime;
4102 se->vruntime -= min_vruntime;
4106 #ifdef CONFIG_FAIR_GROUP_SCHED
4108 * effective_load() calculates the load change as seen from the root_task_group
4110 * Adding load to a group doesn't make a group heavier, but can cause movement
4111 * of group shares between cpus. Assuming the shares were perfectly aligned one
4112 * can calculate the shift in shares.
4114 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4115 * on this @cpu and results in a total addition (subtraction) of @wg to the
4116 * total group weight.
4118 * Given a runqueue weight distribution (rw_i) we can compute a shares
4119 * distribution (s_i) using:
4121 * s_i = rw_i / \Sum rw_j (1)
4123 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4124 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4125 * shares distribution (s_i):
4127 * rw_i = { 2, 4, 1, 0 }
4128 * s_i = { 2/7, 4/7, 1/7, 0 }
4130 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4131 * task used to run on and the CPU the waker is running on), we need to
4132 * compute the effect of waking a task on either CPU and, in case of a sync
4133 * wakeup, compute the effect of the current task going to sleep.
4135 * So for a change of @wl to the local @cpu with an overall group weight change
4136 * of @wl we can compute the new shares distribution (s'_i) using:
4138 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4140 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4141 * differences in waking a task to CPU 0. The additional task changes the
4142 * weight and shares distributions like:
4144 * rw'_i = { 3, 4, 1, 0 }
4145 * s'_i = { 3/8, 4/8, 1/8, 0 }
4147 * We can then compute the difference in effective weight by using:
4149 * dw_i = S * (s'_i - s_i) (3)
4151 * Where 'S' is the group weight as seen by its parent.
4153 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4154 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4155 * 4/7) times the weight of the group.
4157 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4159 struct sched_entity *se = tg->se[cpu];
4161 if (!tg->parent) /* the trivial, non-cgroup case */
4164 for_each_sched_entity(se) {
4170 * W = @wg + \Sum rw_j
4172 W = wg + calc_tg_weight(tg, se->my_q);
4177 w = se->my_q->load.weight + wl;
4180 * wl = S * s'_i; see (2)
4183 wl = (w * tg->shares) / W;
4188 * Per the above, wl is the new se->load.weight value; since
4189 * those are clipped to [MIN_SHARES, ...) do so now. See
4190 * calc_cfs_shares().
4192 if (wl < MIN_SHARES)
4196 * wl = dw_i = S * (s'_i - s_i); see (3)
4198 wl -= se->load.weight;
4201 * Recursively apply this logic to all parent groups to compute
4202 * the final effective load change on the root group. Since
4203 * only the @tg group gets extra weight, all parent groups can
4204 * only redistribute existing shares. @wl is the shift in shares
4205 * resulting from this level per the above.
4214 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4221 static int wake_wide(struct task_struct *p)
4223 int factor = this_cpu_read(sd_llc_size);
4226 * Yeah, it's the switching-frequency, could means many wakee or
4227 * rapidly switch, use factor here will just help to automatically
4228 * adjust the loose-degree, so bigger node will lead to more pull.
4230 if (p->wakee_flips > factor) {
4232 * wakee is somewhat hot, it needs certain amount of cpu
4233 * resource, so if waker is far more hot, prefer to leave
4236 if (current->wakee_flips > (factor * p->wakee_flips))
4243 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4245 s64 this_load, load;
4246 int idx, this_cpu, prev_cpu;
4247 unsigned long tl_per_task;
4248 struct task_group *tg;
4249 unsigned long weight;
4253 * If we wake multiple tasks be careful to not bounce
4254 * ourselves around too much.
4260 this_cpu = smp_processor_id();
4261 prev_cpu = task_cpu(p);
4262 load = source_load(prev_cpu, idx);
4263 this_load = target_load(this_cpu, idx);
4266 * If sync wakeup then subtract the (maximum possible)
4267 * effect of the currently running task from the load
4268 * of the current CPU:
4271 tg = task_group(current);
4272 weight = current->se.load.weight;
4274 this_load += effective_load(tg, this_cpu, -weight, -weight);
4275 load += effective_load(tg, prev_cpu, 0, -weight);
4279 weight = p->se.load.weight;
4282 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4283 * due to the sync cause above having dropped this_load to 0, we'll
4284 * always have an imbalance, but there's really nothing you can do
4285 * about that, so that's good too.
4287 * Otherwise check if either cpus are near enough in load to allow this
4288 * task to be woken on this_cpu.
4290 if (this_load > 0) {
4291 s64 this_eff_load, prev_eff_load;
4293 this_eff_load = 100;
4294 this_eff_load *= capacity_of(prev_cpu);
4295 this_eff_load *= this_load +
4296 effective_load(tg, this_cpu, weight, weight);
4298 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4299 prev_eff_load *= capacity_of(this_cpu);
4300 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4302 balanced = this_eff_load <= prev_eff_load;
4307 * If the currently running task will sleep within
4308 * a reasonable amount of time then attract this newly
4311 if (sync && balanced)
4314 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4315 tl_per_task = cpu_avg_load_per_task(this_cpu);
4318 (this_load <= load &&
4319 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4321 * This domain has SD_WAKE_AFFINE and
4322 * p is cache cold in this domain, and
4323 * there is no bad imbalance.
4325 schedstat_inc(sd, ttwu_move_affine);
4326 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4334 * find_idlest_group finds and returns the least busy CPU group within the
4337 static struct sched_group *
4338 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4339 int this_cpu, int sd_flag)
4341 struct sched_group *idlest = NULL, *group = sd->groups;
4342 unsigned long min_load = ULONG_MAX, this_load = 0;
4343 int load_idx = sd->forkexec_idx;
4344 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4346 if (sd_flag & SD_BALANCE_WAKE)
4347 load_idx = sd->wake_idx;
4350 unsigned long load, avg_load;
4354 /* Skip over this group if it has no CPUs allowed */
4355 if (!cpumask_intersects(sched_group_cpus(group),
4356 tsk_cpus_allowed(p)))
4359 local_group = cpumask_test_cpu(this_cpu,
4360 sched_group_cpus(group));
4362 /* Tally up the load of all CPUs in the group */
4365 for_each_cpu(i, sched_group_cpus(group)) {
4366 /* Bias balancing toward cpus of our domain */
4368 load = source_load(i, load_idx);
4370 load = target_load(i, load_idx);
4375 /* Adjust by relative CPU capacity of the group */
4376 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4379 this_load = avg_load;
4380 } else if (avg_load < min_load) {
4381 min_load = avg_load;
4384 } while (group = group->next, group != sd->groups);
4386 if (!idlest || 100*this_load < imbalance*min_load)
4392 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4395 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4397 unsigned long load, min_load = ULONG_MAX;
4401 /* Traverse only the allowed CPUs */
4402 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4403 load = weighted_cpuload(i);
4405 if (load < min_load || (load == min_load && i == this_cpu)) {
4415 * Try and locate an idle CPU in the sched_domain.
4417 static int select_idle_sibling(struct task_struct *p, int target)
4419 struct sched_domain *sd;
4420 struct sched_group *sg;
4421 int i = task_cpu(p);
4423 if (idle_cpu(target))
4427 * If the prevous cpu is cache affine and idle, don't be stupid.
4429 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4433 * Otherwise, iterate the domains and find an elegible idle cpu.
4435 sd = rcu_dereference(per_cpu(sd_llc, target));
4436 for_each_lower_domain(sd) {
4439 if (!cpumask_intersects(sched_group_cpus(sg),
4440 tsk_cpus_allowed(p)))
4443 for_each_cpu(i, sched_group_cpus(sg)) {
4444 if (i == target || !idle_cpu(i))
4448 target = cpumask_first_and(sched_group_cpus(sg),
4449 tsk_cpus_allowed(p));
4453 } while (sg != sd->groups);
4460 * select_task_rq_fair: Select target runqueue for the waking task in domains
4461 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4462 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4464 * Balances load by selecting the idlest cpu in the idlest group, or under
4465 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4467 * Returns the target cpu number.
4469 * preempt must be disabled.
4472 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4474 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4475 int cpu = smp_processor_id();
4477 int want_affine = 0;
4478 int sync = wake_flags & WF_SYNC;
4480 if (p->nr_cpus_allowed == 1)
4483 if (sd_flag & SD_BALANCE_WAKE) {
4484 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4490 for_each_domain(cpu, tmp) {
4491 if (!(tmp->flags & SD_LOAD_BALANCE))
4495 * If both cpu and prev_cpu are part of this domain,
4496 * cpu is a valid SD_WAKE_AFFINE target.
4498 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4499 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4504 if (tmp->flags & sd_flag)
4508 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4511 if (sd_flag & SD_BALANCE_WAKE) {
4512 new_cpu = select_idle_sibling(p, prev_cpu);
4517 struct sched_group *group;
4520 if (!(sd->flags & sd_flag)) {
4525 group = find_idlest_group(sd, p, cpu, sd_flag);
4531 new_cpu = find_idlest_cpu(group, p, cpu);
4532 if (new_cpu == -1 || new_cpu == cpu) {
4533 /* Now try balancing at a lower domain level of cpu */
4538 /* Now try balancing at a lower domain level of new_cpu */
4540 weight = sd->span_weight;
4542 for_each_domain(cpu, tmp) {
4543 if (weight <= tmp->span_weight)
4545 if (tmp->flags & sd_flag)
4548 /* while loop will break here if sd == NULL */
4557 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4558 * cfs_rq_of(p) references at time of call are still valid and identify the
4559 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4560 * other assumptions, including the state of rq->lock, should be made.
4563 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4565 struct sched_entity *se = &p->se;
4566 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4569 * Load tracking: accumulate removed load so that it can be processed
4570 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4571 * to blocked load iff they have a positive decay-count. It can never
4572 * be negative here since on-rq tasks have decay-count == 0.
4574 if (se->avg.decay_count) {
4575 se->avg.decay_count = -__synchronize_entity_decay(se);
4576 atomic_long_add(se->avg.load_avg_contrib,
4577 &cfs_rq->removed_load);
4580 /* We have migrated, no longer consider this task hot */
4583 #endif /* CONFIG_SMP */
4585 static unsigned long
4586 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4588 unsigned long gran = sysctl_sched_wakeup_granularity;
4591 * Since its curr running now, convert the gran from real-time
4592 * to virtual-time in his units.
4594 * By using 'se' instead of 'curr' we penalize light tasks, so
4595 * they get preempted easier. That is, if 'se' < 'curr' then
4596 * the resulting gran will be larger, therefore penalizing the
4597 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4598 * be smaller, again penalizing the lighter task.
4600 * This is especially important for buddies when the leftmost
4601 * task is higher priority than the buddy.
4603 return calc_delta_fair(gran, se);
4607 * Should 'se' preempt 'curr'.
4621 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4623 s64 gran, vdiff = curr->vruntime - se->vruntime;
4628 gran = wakeup_gran(curr, se);
4635 static void set_last_buddy(struct sched_entity *se)
4637 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4640 for_each_sched_entity(se)
4641 cfs_rq_of(se)->last = se;
4644 static void set_next_buddy(struct sched_entity *se)
4646 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4649 for_each_sched_entity(se)
4650 cfs_rq_of(se)->next = se;
4653 static void set_skip_buddy(struct sched_entity *se)
4655 for_each_sched_entity(se)
4656 cfs_rq_of(se)->skip = se;
4660 * Preempt the current task with a newly woken task if needed:
4662 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4664 struct task_struct *curr = rq->curr;
4665 struct sched_entity *se = &curr->se, *pse = &p->se;
4666 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4667 int scale = cfs_rq->nr_running >= sched_nr_latency;
4668 int next_buddy_marked = 0;
4670 if (unlikely(se == pse))
4674 * This is possible from callers such as move_task(), in which we
4675 * unconditionally check_prempt_curr() after an enqueue (which may have
4676 * lead to a throttle). This both saves work and prevents false
4677 * next-buddy nomination below.
4679 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4682 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4683 set_next_buddy(pse);
4684 next_buddy_marked = 1;
4688 * We can come here with TIF_NEED_RESCHED already set from new task
4691 * Note: this also catches the edge-case of curr being in a throttled
4692 * group (e.g. via set_curr_task), since update_curr() (in the
4693 * enqueue of curr) will have resulted in resched being set. This
4694 * prevents us from potentially nominating it as a false LAST_BUDDY
4697 if (test_tsk_need_resched(curr))
4700 /* Idle tasks are by definition preempted by non-idle tasks. */
4701 if (unlikely(curr->policy == SCHED_IDLE) &&
4702 likely(p->policy != SCHED_IDLE))
4706 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4707 * is driven by the tick):
4709 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4712 find_matching_se(&se, &pse);
4713 update_curr(cfs_rq_of(se));
4715 if (wakeup_preempt_entity(se, pse) == 1) {
4717 * Bias pick_next to pick the sched entity that is
4718 * triggering this preemption.
4720 if (!next_buddy_marked)
4721 set_next_buddy(pse);
4730 * Only set the backward buddy when the current task is still
4731 * on the rq. This can happen when a wakeup gets interleaved
4732 * with schedule on the ->pre_schedule() or idle_balance()
4733 * point, either of which can * drop the rq lock.
4735 * Also, during early boot the idle thread is in the fair class,
4736 * for obvious reasons its a bad idea to schedule back to it.
4738 if (unlikely(!se->on_rq || curr == rq->idle))
4741 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4745 static struct task_struct *
4746 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4748 struct cfs_rq *cfs_rq = &rq->cfs;
4749 struct sched_entity *se;
4750 struct task_struct *p;
4754 #ifdef CONFIG_FAIR_GROUP_SCHED
4755 if (!cfs_rq->nr_running)
4758 if (prev->sched_class != &fair_sched_class)
4762 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4763 * likely that a next task is from the same cgroup as the current.
4765 * Therefore attempt to avoid putting and setting the entire cgroup
4766 * hierarchy, only change the part that actually changes.
4770 struct sched_entity *curr = cfs_rq->curr;
4773 * Since we got here without doing put_prev_entity() we also
4774 * have to consider cfs_rq->curr. If it is still a runnable
4775 * entity, update_curr() will update its vruntime, otherwise
4776 * forget we've ever seen it.
4778 if (curr && curr->on_rq)
4779 update_curr(cfs_rq);
4784 * This call to check_cfs_rq_runtime() will do the throttle and
4785 * dequeue its entity in the parent(s). Therefore the 'simple'
4786 * nr_running test will indeed be correct.
4788 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4791 se = pick_next_entity(cfs_rq, curr);
4792 cfs_rq = group_cfs_rq(se);
4798 * Since we haven't yet done put_prev_entity and if the selected task
4799 * is a different task than we started out with, try and touch the
4800 * least amount of cfs_rqs.
4803 struct sched_entity *pse = &prev->se;
4805 while (!(cfs_rq = is_same_group(se, pse))) {
4806 int se_depth = se->depth;
4807 int pse_depth = pse->depth;
4809 if (se_depth <= pse_depth) {
4810 put_prev_entity(cfs_rq_of(pse), pse);
4811 pse = parent_entity(pse);
4813 if (se_depth >= pse_depth) {
4814 set_next_entity(cfs_rq_of(se), se);
4815 se = parent_entity(se);
4819 put_prev_entity(cfs_rq, pse);
4820 set_next_entity(cfs_rq, se);
4823 if (hrtick_enabled(rq))
4824 hrtick_start_fair(rq, p);
4831 if (!cfs_rq->nr_running)
4834 put_prev_task(rq, prev);
4837 se = pick_next_entity(cfs_rq, NULL);
4838 set_next_entity(cfs_rq, se);
4839 cfs_rq = group_cfs_rq(se);
4844 if (hrtick_enabled(rq))
4845 hrtick_start_fair(rq, p);
4850 new_tasks = idle_balance(rq);
4852 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4853 * possible for any higher priority task to appear. In that case we
4854 * must re-start the pick_next_entity() loop.
4866 * Account for a descheduled task:
4868 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4870 struct sched_entity *se = &prev->se;
4871 struct cfs_rq *cfs_rq;
4873 for_each_sched_entity(se) {
4874 cfs_rq = cfs_rq_of(se);
4875 put_prev_entity(cfs_rq, se);
4880 * sched_yield() is very simple
4882 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4884 static void yield_task_fair(struct rq *rq)
4886 struct task_struct *curr = rq->curr;
4887 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4888 struct sched_entity *se = &curr->se;
4891 * Are we the only task in the tree?
4893 if (unlikely(rq->nr_running == 1))
4896 clear_buddies(cfs_rq, se);
4898 if (curr->policy != SCHED_BATCH) {
4899 update_rq_clock(rq);
4901 * Update run-time statistics of the 'current'.
4903 update_curr(cfs_rq);
4905 * Tell update_rq_clock() that we've just updated,
4906 * so we don't do microscopic update in schedule()
4907 * and double the fastpath cost.
4909 rq->skip_clock_update = 1;
4915 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4917 struct sched_entity *se = &p->se;
4919 /* throttled hierarchies are not runnable */
4920 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4923 /* Tell the scheduler that we'd really like pse to run next. */
4926 yield_task_fair(rq);
4932 /**************************************************
4933 * Fair scheduling class load-balancing methods.
4937 * The purpose of load-balancing is to achieve the same basic fairness the
4938 * per-cpu scheduler provides, namely provide a proportional amount of compute
4939 * time to each task. This is expressed in the following equation:
4941 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4943 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4944 * W_i,0 is defined as:
4946 * W_i,0 = \Sum_j w_i,j (2)
4948 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4949 * is derived from the nice value as per prio_to_weight[].
4951 * The weight average is an exponential decay average of the instantaneous
4954 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4956 * C_i is the compute capacity of cpu i, typically it is the
4957 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4958 * can also include other factors [XXX].
4960 * To achieve this balance we define a measure of imbalance which follows
4961 * directly from (1):
4963 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
4965 * We them move tasks around to minimize the imbalance. In the continuous
4966 * function space it is obvious this converges, in the discrete case we get
4967 * a few fun cases generally called infeasible weight scenarios.
4970 * - infeasible weights;
4971 * - local vs global optima in the discrete case. ]
4976 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4977 * for all i,j solution, we create a tree of cpus that follows the hardware
4978 * topology where each level pairs two lower groups (or better). This results
4979 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4980 * tree to only the first of the previous level and we decrease the frequency
4981 * of load-balance at each level inv. proportional to the number of cpus in
4987 * \Sum { --- * --- * 2^i } = O(n) (5)
4989 * `- size of each group
4990 * | | `- number of cpus doing load-balance
4992 * `- sum over all levels
4994 * Coupled with a limit on how many tasks we can migrate every balance pass,
4995 * this makes (5) the runtime complexity of the balancer.
4997 * An important property here is that each CPU is still (indirectly) connected
4998 * to every other cpu in at most O(log n) steps:
5000 * The adjacency matrix of the resulting graph is given by:
5003 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5006 * And you'll find that:
5008 * A^(log_2 n)_i,j != 0 for all i,j (7)
5010 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5011 * The task movement gives a factor of O(m), giving a convergence complexity
5014 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5019 * In order to avoid CPUs going idle while there's still work to do, new idle
5020 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5021 * tree itself instead of relying on other CPUs to bring it work.
5023 * This adds some complexity to both (5) and (8) but it reduces the total idle
5031 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5034 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5039 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5041 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5043 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5046 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5047 * rewrite all of this once again.]
5050 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5052 enum fbq_type { regular, remote, all };
5054 #define LBF_ALL_PINNED 0x01
5055 #define LBF_NEED_BREAK 0x02
5056 #define LBF_DST_PINNED 0x04
5057 #define LBF_SOME_PINNED 0x08
5060 struct sched_domain *sd;
5068 struct cpumask *dst_grpmask;
5070 enum cpu_idle_type idle;
5072 /* The set of CPUs under consideration for load-balancing */
5073 struct cpumask *cpus;
5078 unsigned int loop_break;
5079 unsigned int loop_max;
5081 enum fbq_type fbq_type;
5085 * move_task - move a task from one runqueue to another runqueue.
5086 * Both runqueues must be locked.
5088 static void move_task(struct task_struct *p, struct lb_env *env)
5090 deactivate_task(env->src_rq, p, 0);
5091 set_task_cpu(p, env->dst_cpu);
5092 activate_task(env->dst_rq, p, 0);
5093 check_preempt_curr(env->dst_rq, p, 0);
5097 * Is this task likely cache-hot:
5099 static int task_hot(struct task_struct *p, struct lb_env *env)
5103 if (p->sched_class != &fair_sched_class)
5106 if (unlikely(p->policy == SCHED_IDLE))
5110 * Buddy candidates are cache hot:
5112 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5113 (&p->se == cfs_rq_of(&p->se)->next ||
5114 &p->se == cfs_rq_of(&p->se)->last))
5117 if (sysctl_sched_migration_cost == -1)
5119 if (sysctl_sched_migration_cost == 0)
5122 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5124 return delta < (s64)sysctl_sched_migration_cost;
5127 #ifdef CONFIG_NUMA_BALANCING
5128 /* Returns true if the destination node has incurred more faults */
5129 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5131 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5132 int src_nid, dst_nid;
5134 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5135 !(env->sd->flags & SD_NUMA)) {
5139 src_nid = cpu_to_node(env->src_cpu);
5140 dst_nid = cpu_to_node(env->dst_cpu);
5142 if (src_nid == dst_nid)
5146 /* Task is already in the group's interleave set. */
5147 if (node_isset(src_nid, numa_group->active_nodes))
5150 /* Task is moving into the group's interleave set. */
5151 if (node_isset(dst_nid, numa_group->active_nodes))
5154 return group_faults(p, dst_nid) > group_faults(p, src_nid);
5157 /* Encourage migration to the preferred node. */
5158 if (dst_nid == p->numa_preferred_nid)
5161 return task_faults(p, dst_nid) > task_faults(p, src_nid);
5165 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5167 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5168 int src_nid, dst_nid;
5170 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5173 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5176 src_nid = cpu_to_node(env->src_cpu);
5177 dst_nid = cpu_to_node(env->dst_cpu);
5179 if (src_nid == dst_nid)
5183 /* Task is moving within/into the group's interleave set. */
5184 if (node_isset(dst_nid, numa_group->active_nodes))
5187 /* Task is moving out of the group's interleave set. */
5188 if (node_isset(src_nid, numa_group->active_nodes))
5191 return group_faults(p, dst_nid) < group_faults(p, src_nid);
5194 /* Migrating away from the preferred node is always bad. */
5195 if (src_nid == p->numa_preferred_nid)
5198 return task_faults(p, dst_nid) < task_faults(p, src_nid);
5202 static inline bool migrate_improves_locality(struct task_struct *p,
5208 static inline bool migrate_degrades_locality(struct task_struct *p,
5216 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5219 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5221 int tsk_cache_hot = 0;
5223 * We do not migrate tasks that are:
5224 * 1) throttled_lb_pair, or
5225 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5226 * 3) running (obviously), or
5227 * 4) are cache-hot on their current CPU.
5229 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5232 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5235 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5237 env->flags |= LBF_SOME_PINNED;
5240 * Remember if this task can be migrated to any other cpu in
5241 * our sched_group. We may want to revisit it if we couldn't
5242 * meet load balance goals by pulling other tasks on src_cpu.
5244 * Also avoid computing new_dst_cpu if we have already computed
5245 * one in current iteration.
5247 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5250 /* Prevent to re-select dst_cpu via env's cpus */
5251 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5252 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5253 env->flags |= LBF_DST_PINNED;
5254 env->new_dst_cpu = cpu;
5262 /* Record that we found atleast one task that could run on dst_cpu */
5263 env->flags &= ~LBF_ALL_PINNED;
5265 if (task_running(env->src_rq, p)) {
5266 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5271 * Aggressive migration if:
5272 * 1) destination numa is preferred
5273 * 2) task is cache cold, or
5274 * 3) too many balance attempts have failed.
5276 tsk_cache_hot = task_hot(p, env);
5278 tsk_cache_hot = migrate_degrades_locality(p, env);
5280 if (migrate_improves_locality(p, env)) {
5281 #ifdef CONFIG_SCHEDSTATS
5282 if (tsk_cache_hot) {
5283 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5284 schedstat_inc(p, se.statistics.nr_forced_migrations);
5290 if (!tsk_cache_hot ||
5291 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5293 if (tsk_cache_hot) {
5294 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5295 schedstat_inc(p, se.statistics.nr_forced_migrations);
5301 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5306 * move_one_task tries to move exactly one task from busiest to this_rq, as
5307 * part of active balancing operations within "domain".
5308 * Returns 1 if successful and 0 otherwise.
5310 * Called with both runqueues locked.
5312 static int move_one_task(struct lb_env *env)
5314 struct task_struct *p, *n;
5316 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5317 if (!can_migrate_task(p, env))
5322 * Right now, this is only the second place move_task()
5323 * is called, so we can safely collect move_task()
5324 * stats here rather than inside move_task().
5326 schedstat_inc(env->sd, lb_gained[env->idle]);
5332 static const unsigned int sched_nr_migrate_break = 32;
5335 * move_tasks tries to move up to imbalance weighted load from busiest to
5336 * this_rq, as part of a balancing operation within domain "sd".
5337 * Returns 1 if successful and 0 otherwise.
5339 * Called with both runqueues locked.
5341 static int move_tasks(struct lb_env *env)
5343 struct list_head *tasks = &env->src_rq->cfs_tasks;
5344 struct task_struct *p;
5348 if (env->imbalance <= 0)
5351 while (!list_empty(tasks)) {
5352 p = list_first_entry(tasks, struct task_struct, se.group_node);
5355 /* We've more or less seen every task there is, call it quits */
5356 if (env->loop > env->loop_max)
5359 /* take a breather every nr_migrate tasks */
5360 if (env->loop > env->loop_break) {
5361 env->loop_break += sched_nr_migrate_break;
5362 env->flags |= LBF_NEED_BREAK;
5366 if (!can_migrate_task(p, env))
5369 load = task_h_load(p);
5371 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5374 if ((load / 2) > env->imbalance)
5379 env->imbalance -= load;
5381 #ifdef CONFIG_PREEMPT
5383 * NEWIDLE balancing is a source of latency, so preemptible
5384 * kernels will stop after the first task is pulled to minimize
5385 * the critical section.
5387 if (env->idle == CPU_NEWLY_IDLE)
5392 * We only want to steal up to the prescribed amount of
5395 if (env->imbalance <= 0)
5400 list_move_tail(&p->se.group_node, tasks);
5404 * Right now, this is one of only two places move_task() is called,
5405 * so we can safely collect move_task() stats here rather than
5406 * inside move_task().
5408 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5413 #ifdef CONFIG_FAIR_GROUP_SCHED
5415 * update tg->load_weight by folding this cpu's load_avg
5417 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5419 struct sched_entity *se = tg->se[cpu];
5420 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5422 /* throttled entities do not contribute to load */
5423 if (throttled_hierarchy(cfs_rq))
5426 update_cfs_rq_blocked_load(cfs_rq, 1);
5429 update_entity_load_avg(se, 1);
5431 * We pivot on our runnable average having decayed to zero for
5432 * list removal. This generally implies that all our children
5433 * have also been removed (modulo rounding error or bandwidth
5434 * control); however, such cases are rare and we can fix these
5437 * TODO: fix up out-of-order children on enqueue.
5439 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5440 list_del_leaf_cfs_rq(cfs_rq);
5442 struct rq *rq = rq_of(cfs_rq);
5443 update_rq_runnable_avg(rq, rq->nr_running);
5447 static void update_blocked_averages(int cpu)
5449 struct rq *rq = cpu_rq(cpu);
5450 struct cfs_rq *cfs_rq;
5451 unsigned long flags;
5453 raw_spin_lock_irqsave(&rq->lock, flags);
5454 update_rq_clock(rq);
5456 * Iterates the task_group tree in a bottom up fashion, see
5457 * list_add_leaf_cfs_rq() for details.
5459 for_each_leaf_cfs_rq(rq, cfs_rq) {
5461 * Note: We may want to consider periodically releasing
5462 * rq->lock about these updates so that creating many task
5463 * groups does not result in continually extending hold time.
5465 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5468 raw_spin_unlock_irqrestore(&rq->lock, flags);
5472 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5473 * This needs to be done in a top-down fashion because the load of a child
5474 * group is a fraction of its parents load.
5476 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5478 struct rq *rq = rq_of(cfs_rq);
5479 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5480 unsigned long now = jiffies;
5483 if (cfs_rq->last_h_load_update == now)
5486 cfs_rq->h_load_next = NULL;
5487 for_each_sched_entity(se) {
5488 cfs_rq = cfs_rq_of(se);
5489 cfs_rq->h_load_next = se;
5490 if (cfs_rq->last_h_load_update == now)
5495 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5496 cfs_rq->last_h_load_update = now;
5499 while ((se = cfs_rq->h_load_next) != NULL) {
5500 load = cfs_rq->h_load;
5501 load = div64_ul(load * se->avg.load_avg_contrib,
5502 cfs_rq->runnable_load_avg + 1);
5503 cfs_rq = group_cfs_rq(se);
5504 cfs_rq->h_load = load;
5505 cfs_rq->last_h_load_update = now;
5509 static unsigned long task_h_load(struct task_struct *p)
5511 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5513 update_cfs_rq_h_load(cfs_rq);
5514 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5515 cfs_rq->runnable_load_avg + 1);
5518 static inline void update_blocked_averages(int cpu)
5522 static unsigned long task_h_load(struct task_struct *p)
5524 return p->se.avg.load_avg_contrib;
5528 /********** Helpers for find_busiest_group ************************/
5530 * sg_lb_stats - stats of a sched_group required for load_balancing
5532 struct sg_lb_stats {
5533 unsigned long avg_load; /*Avg load across the CPUs of the group */
5534 unsigned long group_load; /* Total load over the CPUs of the group */
5535 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5536 unsigned long load_per_task;
5537 unsigned long group_capacity;
5538 unsigned int sum_nr_running; /* Nr tasks running in the group */
5539 unsigned int group_capacity_factor;
5540 unsigned int idle_cpus;
5541 unsigned int group_weight;
5542 int group_imb; /* Is there an imbalance in the group ? */
5543 int group_has_free_capacity;
5544 #ifdef CONFIG_NUMA_BALANCING
5545 unsigned int nr_numa_running;
5546 unsigned int nr_preferred_running;
5551 * sd_lb_stats - Structure to store the statistics of a sched_domain
5552 * during load balancing.
5554 struct sd_lb_stats {
5555 struct sched_group *busiest; /* Busiest group in this sd */
5556 struct sched_group *local; /* Local group in this sd */
5557 unsigned long total_load; /* Total load of all groups in sd */
5558 unsigned long total_capacity; /* Total capacity of all groups in sd */
5559 unsigned long avg_load; /* Average load across all groups in sd */
5561 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5562 struct sg_lb_stats local_stat; /* Statistics of the local group */
5565 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5568 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5569 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5570 * We must however clear busiest_stat::avg_load because
5571 * update_sd_pick_busiest() reads this before assignment.
5573 *sds = (struct sd_lb_stats){
5577 .total_capacity = 0UL,
5585 * get_sd_load_idx - Obtain the load index for a given sched domain.
5586 * @sd: The sched_domain whose load_idx is to be obtained.
5587 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5589 * Return: The load index.
5591 static inline int get_sd_load_idx(struct sched_domain *sd,
5592 enum cpu_idle_type idle)
5598 load_idx = sd->busy_idx;
5601 case CPU_NEWLY_IDLE:
5602 load_idx = sd->newidle_idx;
5605 load_idx = sd->idle_idx;
5612 static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu)
5614 return SCHED_CAPACITY_SCALE;
5617 unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu)
5619 return default_scale_capacity(sd, cpu);
5622 static unsigned long default_scale_smt_capacity(struct sched_domain *sd, int cpu)
5624 unsigned long weight = sd->span_weight;
5625 unsigned long smt_gain = sd->smt_gain;
5632 unsigned long __weak arch_scale_smt_capacity(struct sched_domain *sd, int cpu)
5634 return default_scale_smt_capacity(sd, cpu);
5637 static unsigned long scale_rt_capacity(int cpu)
5639 struct rq *rq = cpu_rq(cpu);
5640 u64 total, available, age_stamp, avg;
5644 * Since we're reading these variables without serialization make sure
5645 * we read them once before doing sanity checks on them.
5647 age_stamp = ACCESS_ONCE(rq->age_stamp);
5648 avg = ACCESS_ONCE(rq->rt_avg);
5650 delta = rq_clock(rq) - age_stamp;
5651 if (unlikely(delta < 0))
5654 total = sched_avg_period() + delta;
5656 if (unlikely(total < avg)) {
5657 /* Ensures that capacity won't end up being negative */
5660 available = total - avg;
5663 if (unlikely((s64)total < SCHED_CAPACITY_SCALE))
5664 total = SCHED_CAPACITY_SCALE;
5666 total >>= SCHED_CAPACITY_SHIFT;
5668 return div_u64(available, total);
5671 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
5673 unsigned long weight = sd->span_weight;
5674 unsigned long capacity = SCHED_CAPACITY_SCALE;
5675 struct sched_group *sdg = sd->groups;
5677 if ((sd->flags & SD_SHARE_CPUCAPACITY) && weight > 1) {
5678 if (sched_feat(ARCH_CAPACITY))
5679 capacity *= arch_scale_smt_capacity(sd, cpu);
5681 capacity *= default_scale_smt_capacity(sd, cpu);
5683 capacity >>= SCHED_CAPACITY_SHIFT;
5686 sdg->sgc->capacity_orig = capacity;
5688 if (sched_feat(ARCH_CAPACITY))
5689 capacity *= arch_scale_freq_capacity(sd, cpu);
5691 capacity *= default_scale_capacity(sd, cpu);
5693 capacity >>= SCHED_CAPACITY_SHIFT;
5695 capacity *= scale_rt_capacity(cpu);
5696 capacity >>= SCHED_CAPACITY_SHIFT;
5701 cpu_rq(cpu)->cpu_capacity = capacity;
5702 sdg->sgc->capacity = capacity;
5705 void update_group_capacity(struct sched_domain *sd, int cpu)
5707 struct sched_domain *child = sd->child;
5708 struct sched_group *group, *sdg = sd->groups;
5709 unsigned long capacity, capacity_orig;
5710 unsigned long interval;
5712 interval = msecs_to_jiffies(sd->balance_interval);
5713 interval = clamp(interval, 1UL, max_load_balance_interval);
5714 sdg->sgc->next_update = jiffies + interval;
5717 update_cpu_capacity(sd, cpu);
5721 capacity_orig = capacity = 0;
5723 if (child->flags & SD_OVERLAP) {
5725 * SD_OVERLAP domains cannot assume that child groups
5726 * span the current group.
5729 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5730 struct sched_group_capacity *sgc;
5731 struct rq *rq = cpu_rq(cpu);
5734 * build_sched_domains() -> init_sched_groups_capacity()
5735 * gets here before we've attached the domains to the
5738 * Use capacity_of(), which is set irrespective of domains
5739 * in update_cpu_capacity().
5741 * This avoids capacity/capacity_orig from being 0 and
5742 * causing divide-by-zero issues on boot.
5744 * Runtime updates will correct capacity_orig.
5746 if (unlikely(!rq->sd)) {
5747 capacity_orig += capacity_of(cpu);
5748 capacity += capacity_of(cpu);
5752 sgc = rq->sd->groups->sgc;
5753 capacity_orig += sgc->capacity_orig;
5754 capacity += sgc->capacity;
5758 * !SD_OVERLAP domains can assume that child groups
5759 * span the current group.
5762 group = child->groups;
5764 capacity_orig += group->sgc->capacity_orig;
5765 capacity += group->sgc->capacity;
5766 group = group->next;
5767 } while (group != child->groups);
5770 sdg->sgc->capacity_orig = capacity_orig;
5771 sdg->sgc->capacity = capacity;
5775 * Try and fix up capacity for tiny siblings, this is needed when
5776 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5777 * which on its own isn't powerful enough.
5779 * See update_sd_pick_busiest() and check_asym_packing().
5782 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5785 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5787 if (!(sd->flags & SD_SHARE_CPUCAPACITY))
5791 * If ~90% of the cpu_capacity is still there, we're good.
5793 if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29)
5800 * Group imbalance indicates (and tries to solve) the problem where balancing
5801 * groups is inadequate due to tsk_cpus_allowed() constraints.
5803 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5804 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5807 * { 0 1 2 3 } { 4 5 6 7 }
5810 * If we were to balance group-wise we'd place two tasks in the first group and
5811 * two tasks in the second group. Clearly this is undesired as it will overload
5812 * cpu 3 and leave one of the cpus in the second group unused.
5814 * The current solution to this issue is detecting the skew in the first group
5815 * by noticing the lower domain failed to reach balance and had difficulty
5816 * moving tasks due to affinity constraints.
5818 * When this is so detected; this group becomes a candidate for busiest; see
5819 * update_sd_pick_busiest(). And calculate_imbalance() and
5820 * find_busiest_group() avoid some of the usual balance conditions to allow it
5821 * to create an effective group imbalance.
5823 * This is a somewhat tricky proposition since the next run might not find the
5824 * group imbalance and decide the groups need to be balanced again. A most
5825 * subtle and fragile situation.
5828 static inline int sg_imbalanced(struct sched_group *group)
5830 return group->sgc->imbalance;
5834 * Compute the group capacity factor.
5836 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5837 * first dividing out the smt factor and computing the actual number of cores
5838 * and limit unit capacity with that.
5840 static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group)
5842 unsigned int capacity_factor, smt, cpus;
5843 unsigned int capacity, capacity_orig;
5845 capacity = group->sgc->capacity;
5846 capacity_orig = group->sgc->capacity_orig;
5847 cpus = group->group_weight;
5849 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5850 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig);
5851 capacity_factor = cpus / smt; /* cores */
5853 capacity_factor = min_t(unsigned,
5854 capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE));
5855 if (!capacity_factor)
5856 capacity_factor = fix_small_capacity(env->sd, group);
5858 return capacity_factor;
5862 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5863 * @env: The load balancing environment.
5864 * @group: sched_group whose statistics are to be updated.
5865 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5866 * @local_group: Does group contain this_cpu.
5867 * @sgs: variable to hold the statistics for this group.
5869 static inline void update_sg_lb_stats(struct lb_env *env,
5870 struct sched_group *group, int load_idx,
5871 int local_group, struct sg_lb_stats *sgs,
5877 memset(sgs, 0, sizeof(*sgs));
5879 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5880 struct rq *rq = cpu_rq(i);
5882 /* Bias balancing toward cpus of our domain */
5884 load = target_load(i, load_idx);
5886 load = source_load(i, load_idx);
5888 sgs->group_load += load;
5889 sgs->sum_nr_running += rq->nr_running;
5891 if (rq->nr_running > 1)
5894 #ifdef CONFIG_NUMA_BALANCING
5895 sgs->nr_numa_running += rq->nr_numa_running;
5896 sgs->nr_preferred_running += rq->nr_preferred_running;
5898 sgs->sum_weighted_load += weighted_cpuload(i);
5903 /* Adjust by relative CPU capacity of the group */
5904 sgs->group_capacity = group->sgc->capacity;
5905 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
5907 if (sgs->sum_nr_running)
5908 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5910 sgs->group_weight = group->group_weight;
5912 sgs->group_imb = sg_imbalanced(group);
5913 sgs->group_capacity_factor = sg_capacity_factor(env, group);
5915 if (sgs->group_capacity_factor > sgs->sum_nr_running)
5916 sgs->group_has_free_capacity = 1;
5920 * update_sd_pick_busiest - return 1 on busiest group
5921 * @env: The load balancing environment.
5922 * @sds: sched_domain statistics
5923 * @sg: sched_group candidate to be checked for being the busiest
5924 * @sgs: sched_group statistics
5926 * Determine if @sg is a busier group than the previously selected
5929 * Return: %true if @sg is a busier group than the previously selected
5930 * busiest group. %false otherwise.
5932 static bool update_sd_pick_busiest(struct lb_env *env,
5933 struct sd_lb_stats *sds,
5934 struct sched_group *sg,
5935 struct sg_lb_stats *sgs)
5937 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5940 if (sgs->sum_nr_running > sgs->group_capacity_factor)
5947 * ASYM_PACKING needs to move all the work to the lowest
5948 * numbered CPUs in the group, therefore mark all groups
5949 * higher than ourself as busy.
5951 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5952 env->dst_cpu < group_first_cpu(sg)) {
5956 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5963 #ifdef CONFIG_NUMA_BALANCING
5964 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5966 if (sgs->sum_nr_running > sgs->nr_numa_running)
5968 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5973 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5975 if (rq->nr_running > rq->nr_numa_running)
5977 if (rq->nr_running > rq->nr_preferred_running)
5982 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5987 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5991 #endif /* CONFIG_NUMA_BALANCING */
5994 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5995 * @env: The load balancing environment.
5996 * @sds: variable to hold the statistics for this sched_domain.
5998 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6000 struct sched_domain *child = env->sd->child;
6001 struct sched_group *sg = env->sd->groups;
6002 struct sg_lb_stats tmp_sgs;
6003 int load_idx, prefer_sibling = 0;
6004 bool overload = false;
6006 if (child && child->flags & SD_PREFER_SIBLING)
6009 load_idx = get_sd_load_idx(env->sd, env->idle);
6012 struct sg_lb_stats *sgs = &tmp_sgs;
6015 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6018 sgs = &sds->local_stat;
6020 if (env->idle != CPU_NEWLY_IDLE ||
6021 time_after_eq(jiffies, sg->sgc->next_update))
6022 update_group_capacity(env->sd, env->dst_cpu);
6025 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6032 * In case the child domain prefers tasks go to siblings
6033 * first, lower the sg capacity factor to one so that we'll try
6034 * and move all the excess tasks away. We lower the capacity
6035 * of a group only if the local group has the capacity to fit
6036 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6037 * extra check prevents the case where you always pull from the
6038 * heaviest group when it is already under-utilized (possible
6039 * with a large weight task outweighs the tasks on the system).
6041 if (prefer_sibling && sds->local &&
6042 sds->local_stat.group_has_free_capacity)
6043 sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U);
6045 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6047 sds->busiest_stat = *sgs;
6051 /* Now, start updating sd_lb_stats */
6052 sds->total_load += sgs->group_load;
6053 sds->total_capacity += sgs->group_capacity;
6056 } while (sg != env->sd->groups);
6058 if (env->sd->flags & SD_NUMA)
6059 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6061 if (!env->sd->parent) {
6062 /* update overload indicator if we are at root domain */
6063 if (env->dst_rq->rd->overload != overload)
6064 env->dst_rq->rd->overload = overload;
6070 * check_asym_packing - Check to see if the group is packed into the
6073 * This is primarily intended to used at the sibling level. Some
6074 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6075 * case of POWER7, it can move to lower SMT modes only when higher
6076 * threads are idle. When in lower SMT modes, the threads will
6077 * perform better since they share less core resources. Hence when we
6078 * have idle threads, we want them to be the higher ones.
6080 * This packing function is run on idle threads. It checks to see if
6081 * the busiest CPU in this domain (core in the P7 case) has a higher
6082 * CPU number than the packing function is being run on. Here we are
6083 * assuming lower CPU number will be equivalent to lower a SMT thread
6086 * Return: 1 when packing is required and a task should be moved to
6087 * this CPU. The amount of the imbalance is returned in *imbalance.
6089 * @env: The load balancing environment.
6090 * @sds: Statistics of the sched_domain which is to be packed
6092 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6096 if (!(env->sd->flags & SD_ASYM_PACKING))
6102 busiest_cpu = group_first_cpu(sds->busiest);
6103 if (env->dst_cpu > busiest_cpu)
6106 env->imbalance = DIV_ROUND_CLOSEST(
6107 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6108 SCHED_CAPACITY_SCALE);
6114 * fix_small_imbalance - Calculate the minor imbalance that exists
6115 * amongst the groups of a sched_domain, during
6117 * @env: The load balancing environment.
6118 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6121 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6123 unsigned long tmp, capa_now = 0, capa_move = 0;
6124 unsigned int imbn = 2;
6125 unsigned long scaled_busy_load_per_task;
6126 struct sg_lb_stats *local, *busiest;
6128 local = &sds->local_stat;
6129 busiest = &sds->busiest_stat;
6131 if (!local->sum_nr_running)
6132 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6133 else if (busiest->load_per_task > local->load_per_task)
6136 scaled_busy_load_per_task =
6137 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6138 busiest->group_capacity;
6140 if (busiest->avg_load + scaled_busy_load_per_task >=
6141 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6142 env->imbalance = busiest->load_per_task;
6147 * OK, we don't have enough imbalance to justify moving tasks,
6148 * however we may be able to increase total CPU capacity used by
6152 capa_now += busiest->group_capacity *
6153 min(busiest->load_per_task, busiest->avg_load);
6154 capa_now += local->group_capacity *
6155 min(local->load_per_task, local->avg_load);
6156 capa_now /= SCHED_CAPACITY_SCALE;
6158 /* Amount of load we'd subtract */
6159 if (busiest->avg_load > scaled_busy_load_per_task) {
6160 capa_move += busiest->group_capacity *
6161 min(busiest->load_per_task,
6162 busiest->avg_load - scaled_busy_load_per_task);
6165 /* Amount of load we'd add */
6166 if (busiest->avg_load * busiest->group_capacity <
6167 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6168 tmp = (busiest->avg_load * busiest->group_capacity) /
6169 local->group_capacity;
6171 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6172 local->group_capacity;
6174 capa_move += local->group_capacity *
6175 min(local->load_per_task, local->avg_load + tmp);
6176 capa_move /= SCHED_CAPACITY_SCALE;
6178 /* Move if we gain throughput */
6179 if (capa_move > capa_now)
6180 env->imbalance = busiest->load_per_task;
6184 * calculate_imbalance - Calculate the amount of imbalance present within the
6185 * groups of a given sched_domain during load balance.
6186 * @env: load balance environment
6187 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6189 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6191 unsigned long max_pull, load_above_capacity = ~0UL;
6192 struct sg_lb_stats *local, *busiest;
6194 local = &sds->local_stat;
6195 busiest = &sds->busiest_stat;
6197 if (busiest->group_imb) {
6199 * In the group_imb case we cannot rely on group-wide averages
6200 * to ensure cpu-load equilibrium, look at wider averages. XXX
6202 busiest->load_per_task =
6203 min(busiest->load_per_task, sds->avg_load);
6207 * In the presence of smp nice balancing, certain scenarios can have
6208 * max load less than avg load(as we skip the groups at or below
6209 * its cpu_capacity, while calculating max_load..)
6211 if (busiest->avg_load <= sds->avg_load ||
6212 local->avg_load >= sds->avg_load) {
6214 return fix_small_imbalance(env, sds);
6217 if (!busiest->group_imb) {
6219 * Don't want to pull so many tasks that a group would go idle.
6220 * Except of course for the group_imb case, since then we might
6221 * have to drop below capacity to reach cpu-load equilibrium.
6223 load_above_capacity =
6224 (busiest->sum_nr_running - busiest->group_capacity_factor);
6226 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE);
6227 load_above_capacity /= busiest->group_capacity;
6231 * We're trying to get all the cpus to the average_load, so we don't
6232 * want to push ourselves above the average load, nor do we wish to
6233 * reduce the max loaded cpu below the average load. At the same time,
6234 * we also don't want to reduce the group load below the group capacity
6235 * (so that we can implement power-savings policies etc). Thus we look
6236 * for the minimum possible imbalance.
6238 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6240 /* How much load to actually move to equalise the imbalance */
6241 env->imbalance = min(
6242 max_pull * busiest->group_capacity,
6243 (sds->avg_load - local->avg_load) * local->group_capacity
6244 ) / SCHED_CAPACITY_SCALE;
6247 * if *imbalance is less than the average load per runnable task
6248 * there is no guarantee that any tasks will be moved so we'll have
6249 * a think about bumping its value to force at least one task to be
6252 if (env->imbalance < busiest->load_per_task)
6253 return fix_small_imbalance(env, sds);
6256 /******* find_busiest_group() helpers end here *********************/
6259 * find_busiest_group - Returns the busiest group within the sched_domain
6260 * if there is an imbalance. If there isn't an imbalance, and
6261 * the user has opted for power-savings, it returns a group whose
6262 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6263 * such a group exists.
6265 * Also calculates the amount of weighted load which should be moved
6266 * to restore balance.
6268 * @env: The load balancing environment.
6270 * Return: - The busiest group if imbalance exists.
6271 * - If no imbalance and user has opted for power-savings balance,
6272 * return the least loaded group whose CPUs can be
6273 * put to idle by rebalancing its tasks onto our group.
6275 static struct sched_group *find_busiest_group(struct lb_env *env)
6277 struct sg_lb_stats *local, *busiest;
6278 struct sd_lb_stats sds;
6280 init_sd_lb_stats(&sds);
6283 * Compute the various statistics relavent for load balancing at
6286 update_sd_lb_stats(env, &sds);
6287 local = &sds.local_stat;
6288 busiest = &sds.busiest_stat;
6290 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6291 check_asym_packing(env, &sds))
6294 /* There is no busy sibling group to pull tasks from */
6295 if (!sds.busiest || busiest->sum_nr_running == 0)
6298 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6299 / sds.total_capacity;
6302 * If the busiest group is imbalanced the below checks don't
6303 * work because they assume all things are equal, which typically
6304 * isn't true due to cpus_allowed constraints and the like.
6306 if (busiest->group_imb)
6309 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6310 if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity &&
6311 !busiest->group_has_free_capacity)
6315 * If the local group is more busy than the selected busiest group
6316 * don't try and pull any tasks.
6318 if (local->avg_load >= busiest->avg_load)
6322 * Don't pull any tasks if this group is already above the domain
6325 if (local->avg_load >= sds.avg_load)
6328 if (env->idle == CPU_IDLE) {
6330 * This cpu is idle. If the busiest group load doesn't
6331 * have more tasks than the number of available cpu's and
6332 * there is no imbalance between this and busiest group
6333 * wrt to idle cpu's, it is balanced.
6335 if ((local->idle_cpus < busiest->idle_cpus) &&
6336 busiest->sum_nr_running <= busiest->group_weight)
6340 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6341 * imbalance_pct to be conservative.
6343 if (100 * busiest->avg_load <=
6344 env->sd->imbalance_pct * local->avg_load)
6349 /* Looks like there is an imbalance. Compute it */
6350 calculate_imbalance(env, &sds);
6359 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6361 static struct rq *find_busiest_queue(struct lb_env *env,
6362 struct sched_group *group)
6364 struct rq *busiest = NULL, *rq;
6365 unsigned long busiest_load = 0, busiest_capacity = 1;
6368 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6369 unsigned long capacity, capacity_factor, wl;
6373 rt = fbq_classify_rq(rq);
6376 * We classify groups/runqueues into three groups:
6377 * - regular: there are !numa tasks
6378 * - remote: there are numa tasks that run on the 'wrong' node
6379 * - all: there is no distinction
6381 * In order to avoid migrating ideally placed numa tasks,
6382 * ignore those when there's better options.
6384 * If we ignore the actual busiest queue to migrate another
6385 * task, the next balance pass can still reduce the busiest
6386 * queue by moving tasks around inside the node.
6388 * If we cannot move enough load due to this classification
6389 * the next pass will adjust the group classification and
6390 * allow migration of more tasks.
6392 * Both cases only affect the total convergence complexity.
6394 if (rt > env->fbq_type)
6397 capacity = capacity_of(i);
6398 capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE);
6399 if (!capacity_factor)
6400 capacity_factor = fix_small_capacity(env->sd, group);
6402 wl = weighted_cpuload(i);
6405 * When comparing with imbalance, use weighted_cpuload()
6406 * which is not scaled with the cpu capacity.
6408 if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance)
6412 * For the load comparisons with the other cpu's, consider
6413 * the weighted_cpuload() scaled with the cpu capacity, so
6414 * that the load can be moved away from the cpu that is
6415 * potentially running at a lower capacity.
6417 * Thus we're looking for max(wl_i / capacity_i), crosswise
6418 * multiplication to rid ourselves of the division works out
6419 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6420 * our previous maximum.
6422 if (wl * busiest_capacity > busiest_load * capacity) {
6424 busiest_capacity = capacity;
6433 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6434 * so long as it is large enough.
6436 #define MAX_PINNED_INTERVAL 512
6438 /* Working cpumask for load_balance and load_balance_newidle. */
6439 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6441 static int need_active_balance(struct lb_env *env)
6443 struct sched_domain *sd = env->sd;
6445 if (env->idle == CPU_NEWLY_IDLE) {
6448 * ASYM_PACKING needs to force migrate tasks from busy but
6449 * higher numbered CPUs in order to pack all tasks in the
6450 * lowest numbered CPUs.
6452 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6456 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6459 static int active_load_balance_cpu_stop(void *data);
6461 static int should_we_balance(struct lb_env *env)
6463 struct sched_group *sg = env->sd->groups;
6464 struct cpumask *sg_cpus, *sg_mask;
6465 int cpu, balance_cpu = -1;
6468 * In the newly idle case, we will allow all the cpu's
6469 * to do the newly idle load balance.
6471 if (env->idle == CPU_NEWLY_IDLE)
6474 sg_cpus = sched_group_cpus(sg);
6475 sg_mask = sched_group_mask(sg);
6476 /* Try to find first idle cpu */
6477 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6478 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6485 if (balance_cpu == -1)
6486 balance_cpu = group_balance_cpu(sg);
6489 * First idle cpu or the first cpu(busiest) in this sched group
6490 * is eligible for doing load balancing at this and above domains.
6492 return balance_cpu == env->dst_cpu;
6496 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6497 * tasks if there is an imbalance.
6499 static int load_balance(int this_cpu, struct rq *this_rq,
6500 struct sched_domain *sd, enum cpu_idle_type idle,
6501 int *continue_balancing)
6503 int ld_moved, cur_ld_moved, active_balance = 0;
6504 struct sched_domain *sd_parent = sd->parent;
6505 struct sched_group *group;
6507 unsigned long flags;
6508 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6510 struct lb_env env = {
6512 .dst_cpu = this_cpu,
6514 .dst_grpmask = sched_group_cpus(sd->groups),
6516 .loop_break = sched_nr_migrate_break,
6522 * For NEWLY_IDLE load_balancing, we don't need to consider
6523 * other cpus in our group
6525 if (idle == CPU_NEWLY_IDLE)
6526 env.dst_grpmask = NULL;
6528 cpumask_copy(cpus, cpu_active_mask);
6530 schedstat_inc(sd, lb_count[idle]);
6533 if (!should_we_balance(&env)) {
6534 *continue_balancing = 0;
6538 group = find_busiest_group(&env);
6540 schedstat_inc(sd, lb_nobusyg[idle]);
6544 busiest = find_busiest_queue(&env, group);
6546 schedstat_inc(sd, lb_nobusyq[idle]);
6550 BUG_ON(busiest == env.dst_rq);
6552 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6555 if (busiest->nr_running > 1) {
6557 * Attempt to move tasks. If find_busiest_group has found
6558 * an imbalance but busiest->nr_running <= 1, the group is
6559 * still unbalanced. ld_moved simply stays zero, so it is
6560 * correctly treated as an imbalance.
6562 env.flags |= LBF_ALL_PINNED;
6563 env.src_cpu = busiest->cpu;
6564 env.src_rq = busiest;
6565 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6568 local_irq_save(flags);
6569 double_rq_lock(env.dst_rq, busiest);
6572 * cur_ld_moved - load moved in current iteration
6573 * ld_moved - cumulative load moved across iterations
6575 cur_ld_moved = move_tasks(&env);
6576 ld_moved += cur_ld_moved;
6577 double_rq_unlock(env.dst_rq, busiest);
6578 local_irq_restore(flags);
6581 * some other cpu did the load balance for us.
6583 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6584 resched_cpu(env.dst_cpu);
6586 if (env.flags & LBF_NEED_BREAK) {
6587 env.flags &= ~LBF_NEED_BREAK;
6592 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6593 * us and move them to an alternate dst_cpu in our sched_group
6594 * where they can run. The upper limit on how many times we
6595 * iterate on same src_cpu is dependent on number of cpus in our
6598 * This changes load balance semantics a bit on who can move
6599 * load to a given_cpu. In addition to the given_cpu itself
6600 * (or a ilb_cpu acting on its behalf where given_cpu is
6601 * nohz-idle), we now have balance_cpu in a position to move
6602 * load to given_cpu. In rare situations, this may cause
6603 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6604 * _independently_ and at _same_ time to move some load to
6605 * given_cpu) causing exceess load to be moved to given_cpu.
6606 * This however should not happen so much in practice and
6607 * moreover subsequent load balance cycles should correct the
6608 * excess load moved.
6610 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6612 /* Prevent to re-select dst_cpu via env's cpus */
6613 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6615 env.dst_rq = cpu_rq(env.new_dst_cpu);
6616 env.dst_cpu = env.new_dst_cpu;
6617 env.flags &= ~LBF_DST_PINNED;
6619 env.loop_break = sched_nr_migrate_break;
6622 * Go back to "more_balance" rather than "redo" since we
6623 * need to continue with same src_cpu.
6629 * We failed to reach balance because of affinity.
6632 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
6634 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6635 *group_imbalance = 1;
6636 } else if (*group_imbalance)
6637 *group_imbalance = 0;
6640 /* All tasks on this runqueue were pinned by CPU affinity */
6641 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6642 cpumask_clear_cpu(cpu_of(busiest), cpus);
6643 if (!cpumask_empty(cpus)) {
6645 env.loop_break = sched_nr_migrate_break;
6653 schedstat_inc(sd, lb_failed[idle]);
6655 * Increment the failure counter only on periodic balance.
6656 * We do not want newidle balance, which can be very
6657 * frequent, pollute the failure counter causing
6658 * excessive cache_hot migrations and active balances.
6660 if (idle != CPU_NEWLY_IDLE)
6661 sd->nr_balance_failed++;
6663 if (need_active_balance(&env)) {
6664 raw_spin_lock_irqsave(&busiest->lock, flags);
6666 /* don't kick the active_load_balance_cpu_stop,
6667 * if the curr task on busiest cpu can't be
6670 if (!cpumask_test_cpu(this_cpu,
6671 tsk_cpus_allowed(busiest->curr))) {
6672 raw_spin_unlock_irqrestore(&busiest->lock,
6674 env.flags |= LBF_ALL_PINNED;
6675 goto out_one_pinned;
6679 * ->active_balance synchronizes accesses to
6680 * ->active_balance_work. Once set, it's cleared
6681 * only after active load balance is finished.
6683 if (!busiest->active_balance) {
6684 busiest->active_balance = 1;
6685 busiest->push_cpu = this_cpu;
6688 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6690 if (active_balance) {
6691 stop_one_cpu_nowait(cpu_of(busiest),
6692 active_load_balance_cpu_stop, busiest,
6693 &busiest->active_balance_work);
6697 * We've kicked active balancing, reset the failure
6700 sd->nr_balance_failed = sd->cache_nice_tries+1;
6703 sd->nr_balance_failed = 0;
6705 if (likely(!active_balance)) {
6706 /* We were unbalanced, so reset the balancing interval */
6707 sd->balance_interval = sd->min_interval;
6710 * If we've begun active balancing, start to back off. This
6711 * case may not be covered by the all_pinned logic if there
6712 * is only 1 task on the busy runqueue (because we don't call
6715 if (sd->balance_interval < sd->max_interval)
6716 sd->balance_interval *= 2;
6722 schedstat_inc(sd, lb_balanced[idle]);
6724 sd->nr_balance_failed = 0;
6727 /* tune up the balancing interval */
6728 if (((env.flags & LBF_ALL_PINNED) &&
6729 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6730 (sd->balance_interval < sd->max_interval))
6731 sd->balance_interval *= 2;
6738 static inline unsigned long
6739 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
6741 unsigned long interval = sd->balance_interval;
6744 interval *= sd->busy_factor;
6746 /* scale ms to jiffies */
6747 interval = msecs_to_jiffies(interval);
6748 interval = clamp(interval, 1UL, max_load_balance_interval);
6754 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
6756 unsigned long interval, next;
6758 interval = get_sd_balance_interval(sd, cpu_busy);
6759 next = sd->last_balance + interval;
6761 if (time_after(*next_balance, next))
6762 *next_balance = next;
6766 * idle_balance is called by schedule() if this_cpu is about to become
6767 * idle. Attempts to pull tasks from other CPUs.
6769 static int idle_balance(struct rq *this_rq)
6771 unsigned long next_balance = jiffies + HZ;
6772 int this_cpu = this_rq->cpu;
6773 struct sched_domain *sd;
6774 int pulled_task = 0;
6777 idle_enter_fair(this_rq);
6780 * We must set idle_stamp _before_ calling idle_balance(), such that we
6781 * measure the duration of idle_balance() as idle time.
6783 this_rq->idle_stamp = rq_clock(this_rq);
6785 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
6786 !this_rq->rd->overload) {
6788 sd = rcu_dereference_check_sched_domain(this_rq->sd);
6790 update_next_balance(sd, 0, &next_balance);
6797 * Drop the rq->lock, but keep IRQ/preempt disabled.
6799 raw_spin_unlock(&this_rq->lock);
6801 update_blocked_averages(this_cpu);
6803 for_each_domain(this_cpu, sd) {
6804 int continue_balancing = 1;
6805 u64 t0, domain_cost;
6807 if (!(sd->flags & SD_LOAD_BALANCE))
6810 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
6811 update_next_balance(sd, 0, &next_balance);
6815 if (sd->flags & SD_BALANCE_NEWIDLE) {
6816 t0 = sched_clock_cpu(this_cpu);
6818 pulled_task = load_balance(this_cpu, this_rq,
6820 &continue_balancing);
6822 domain_cost = sched_clock_cpu(this_cpu) - t0;
6823 if (domain_cost > sd->max_newidle_lb_cost)
6824 sd->max_newidle_lb_cost = domain_cost;
6826 curr_cost += domain_cost;
6829 update_next_balance(sd, 0, &next_balance);
6832 * Stop searching for tasks to pull if there are
6833 * now runnable tasks on this rq.
6835 if (pulled_task || this_rq->nr_running > 0)
6840 raw_spin_lock(&this_rq->lock);
6842 if (curr_cost > this_rq->max_idle_balance_cost)
6843 this_rq->max_idle_balance_cost = curr_cost;
6846 * While browsing the domains, we released the rq lock, a task could
6847 * have been enqueued in the meantime. Since we're not going idle,
6848 * pretend we pulled a task.
6850 if (this_rq->cfs.h_nr_running && !pulled_task)
6854 /* Move the next balance forward */
6855 if (time_after(this_rq->next_balance, next_balance))
6856 this_rq->next_balance = next_balance;
6858 /* Is there a task of a high priority class? */
6859 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
6863 idle_exit_fair(this_rq);
6864 this_rq->idle_stamp = 0;
6871 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6872 * running tasks off the busiest CPU onto idle CPUs. It requires at
6873 * least 1 task to be running on each physical CPU where possible, and
6874 * avoids physical / logical imbalances.
6876 static int active_load_balance_cpu_stop(void *data)
6878 struct rq *busiest_rq = data;
6879 int busiest_cpu = cpu_of(busiest_rq);
6880 int target_cpu = busiest_rq->push_cpu;
6881 struct rq *target_rq = cpu_rq(target_cpu);
6882 struct sched_domain *sd;
6884 raw_spin_lock_irq(&busiest_rq->lock);
6886 /* make sure the requested cpu hasn't gone down in the meantime */
6887 if (unlikely(busiest_cpu != smp_processor_id() ||
6888 !busiest_rq->active_balance))
6891 /* Is there any task to move? */
6892 if (busiest_rq->nr_running <= 1)
6896 * This condition is "impossible", if it occurs
6897 * we need to fix it. Originally reported by
6898 * Bjorn Helgaas on a 128-cpu setup.
6900 BUG_ON(busiest_rq == target_rq);
6902 /* move a task from busiest_rq to target_rq */
6903 double_lock_balance(busiest_rq, target_rq);
6905 /* Search for an sd spanning us and the target CPU. */
6907 for_each_domain(target_cpu, sd) {
6908 if ((sd->flags & SD_LOAD_BALANCE) &&
6909 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6914 struct lb_env env = {
6916 .dst_cpu = target_cpu,
6917 .dst_rq = target_rq,
6918 .src_cpu = busiest_rq->cpu,
6919 .src_rq = busiest_rq,
6923 schedstat_inc(sd, alb_count);
6925 if (move_one_task(&env))
6926 schedstat_inc(sd, alb_pushed);
6928 schedstat_inc(sd, alb_failed);
6931 double_unlock_balance(busiest_rq, target_rq);
6933 busiest_rq->active_balance = 0;
6934 raw_spin_unlock_irq(&busiest_rq->lock);
6938 static inline int on_null_domain(struct rq *rq)
6940 return unlikely(!rcu_dereference_sched(rq->sd));
6943 #ifdef CONFIG_NO_HZ_COMMON
6945 * idle load balancing details
6946 * - When one of the busy CPUs notice that there may be an idle rebalancing
6947 * needed, they will kick the idle load balancer, which then does idle
6948 * load balancing for all the idle CPUs.
6951 cpumask_var_t idle_cpus_mask;
6953 unsigned long next_balance; /* in jiffy units */
6954 } nohz ____cacheline_aligned;
6956 static inline int find_new_ilb(void)
6958 int ilb = cpumask_first(nohz.idle_cpus_mask);
6960 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6967 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6968 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6969 * CPU (if there is one).
6971 static void nohz_balancer_kick(void)
6975 nohz.next_balance++;
6977 ilb_cpu = find_new_ilb();
6979 if (ilb_cpu >= nr_cpu_ids)
6982 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6985 * Use smp_send_reschedule() instead of resched_cpu().
6986 * This way we generate a sched IPI on the target cpu which
6987 * is idle. And the softirq performing nohz idle load balance
6988 * will be run before returning from the IPI.
6990 smp_send_reschedule(ilb_cpu);
6994 static inline void nohz_balance_exit_idle(int cpu)
6996 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6998 * Completely isolated CPUs don't ever set, so we must test.
7000 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7001 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7002 atomic_dec(&nohz.nr_cpus);
7004 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7008 static inline void set_cpu_sd_state_busy(void)
7010 struct sched_domain *sd;
7011 int cpu = smp_processor_id();
7014 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7016 if (!sd || !sd->nohz_idle)
7020 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7025 void set_cpu_sd_state_idle(void)
7027 struct sched_domain *sd;
7028 int cpu = smp_processor_id();
7031 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7033 if (!sd || sd->nohz_idle)
7037 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7043 * This routine will record that the cpu is going idle with tick stopped.
7044 * This info will be used in performing idle load balancing in the future.
7046 void nohz_balance_enter_idle(int cpu)
7049 * If this cpu is going down, then nothing needs to be done.
7051 if (!cpu_active(cpu))
7054 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7058 * If we're a completely isolated CPU, we don't play.
7060 if (on_null_domain(cpu_rq(cpu)))
7063 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7064 atomic_inc(&nohz.nr_cpus);
7065 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7068 static int sched_ilb_notifier(struct notifier_block *nfb,
7069 unsigned long action, void *hcpu)
7071 switch (action & ~CPU_TASKS_FROZEN) {
7073 nohz_balance_exit_idle(smp_processor_id());
7081 static DEFINE_SPINLOCK(balancing);
7084 * Scale the max load_balance interval with the number of CPUs in the system.
7085 * This trades load-balance latency on larger machines for less cross talk.
7087 void update_max_interval(void)
7089 max_load_balance_interval = HZ*num_online_cpus()/10;
7093 * It checks each scheduling domain to see if it is due to be balanced,
7094 * and initiates a balancing operation if so.
7096 * Balancing parameters are set up in init_sched_domains.
7098 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7100 int continue_balancing = 1;
7102 unsigned long interval;
7103 struct sched_domain *sd;
7104 /* Earliest time when we have to do rebalance again */
7105 unsigned long next_balance = jiffies + 60*HZ;
7106 int update_next_balance = 0;
7107 int need_serialize, need_decay = 0;
7110 update_blocked_averages(cpu);
7113 for_each_domain(cpu, sd) {
7115 * Decay the newidle max times here because this is a regular
7116 * visit to all the domains. Decay ~1% per second.
7118 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7119 sd->max_newidle_lb_cost =
7120 (sd->max_newidle_lb_cost * 253) / 256;
7121 sd->next_decay_max_lb_cost = jiffies + HZ;
7124 max_cost += sd->max_newidle_lb_cost;
7126 if (!(sd->flags & SD_LOAD_BALANCE))
7130 * Stop the load balance at this level. There is another
7131 * CPU in our sched group which is doing load balancing more
7134 if (!continue_balancing) {
7140 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7142 need_serialize = sd->flags & SD_SERIALIZE;
7143 if (need_serialize) {
7144 if (!spin_trylock(&balancing))
7148 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7149 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7151 * The LBF_DST_PINNED logic could have changed
7152 * env->dst_cpu, so we can't know our idle
7153 * state even if we migrated tasks. Update it.
7155 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7157 sd->last_balance = jiffies;
7158 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7161 spin_unlock(&balancing);
7163 if (time_after(next_balance, sd->last_balance + interval)) {
7164 next_balance = sd->last_balance + interval;
7165 update_next_balance = 1;
7170 * Ensure the rq-wide value also decays but keep it at a
7171 * reasonable floor to avoid funnies with rq->avg_idle.
7173 rq->max_idle_balance_cost =
7174 max((u64)sysctl_sched_migration_cost, max_cost);
7179 * next_balance will be updated only when there is a need.
7180 * When the cpu is attached to null domain for ex, it will not be
7183 if (likely(update_next_balance))
7184 rq->next_balance = next_balance;
7187 #ifdef CONFIG_NO_HZ_COMMON
7189 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7190 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7192 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7194 int this_cpu = this_rq->cpu;
7198 if (idle != CPU_IDLE ||
7199 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7202 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7203 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7207 * If this cpu gets work to do, stop the load balancing
7208 * work being done for other cpus. Next load
7209 * balancing owner will pick it up.
7214 rq = cpu_rq(balance_cpu);
7217 * If time for next balance is due,
7220 if (time_after_eq(jiffies, rq->next_balance)) {
7221 raw_spin_lock_irq(&rq->lock);
7222 update_rq_clock(rq);
7223 update_idle_cpu_load(rq);
7224 raw_spin_unlock_irq(&rq->lock);
7225 rebalance_domains(rq, CPU_IDLE);
7228 if (time_after(this_rq->next_balance, rq->next_balance))
7229 this_rq->next_balance = rq->next_balance;
7231 nohz.next_balance = this_rq->next_balance;
7233 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7237 * Current heuristic for kicking the idle load balancer in the presence
7238 * of an idle cpu is the system.
7239 * - This rq has more than one task.
7240 * - At any scheduler domain level, this cpu's scheduler group has multiple
7241 * busy cpu's exceeding the group's capacity.
7242 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7243 * domain span are idle.
7245 static inline int nohz_kick_needed(struct rq *rq)
7247 unsigned long now = jiffies;
7248 struct sched_domain *sd;
7249 struct sched_group_capacity *sgc;
7250 int nr_busy, cpu = rq->cpu;
7252 if (unlikely(rq->idle_balance))
7256 * We may be recently in ticked or tickless idle mode. At the first
7257 * busy tick after returning from idle, we will update the busy stats.
7259 set_cpu_sd_state_busy();
7260 nohz_balance_exit_idle(cpu);
7263 * None are in tickless mode and hence no need for NOHZ idle load
7266 if (likely(!atomic_read(&nohz.nr_cpus)))
7269 if (time_before(now, nohz.next_balance))
7272 if (rq->nr_running >= 2)
7276 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7279 sgc = sd->groups->sgc;
7280 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7283 goto need_kick_unlock;
7286 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7288 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7289 sched_domain_span(sd)) < cpu))
7290 goto need_kick_unlock;
7301 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7305 * run_rebalance_domains is triggered when needed from the scheduler tick.
7306 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7308 static void run_rebalance_domains(struct softirq_action *h)
7310 struct rq *this_rq = this_rq();
7311 enum cpu_idle_type idle = this_rq->idle_balance ?
7312 CPU_IDLE : CPU_NOT_IDLE;
7314 rebalance_domains(this_rq, idle);
7317 * If this cpu has a pending nohz_balance_kick, then do the
7318 * balancing on behalf of the other idle cpus whose ticks are
7321 nohz_idle_balance(this_rq, idle);
7325 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7327 void trigger_load_balance(struct rq *rq)
7329 /* Don't need to rebalance while attached to NULL domain */
7330 if (unlikely(on_null_domain(rq)))
7333 if (time_after_eq(jiffies, rq->next_balance))
7334 raise_softirq(SCHED_SOFTIRQ);
7335 #ifdef CONFIG_NO_HZ_COMMON
7336 if (nohz_kick_needed(rq))
7337 nohz_balancer_kick();
7341 static void rq_online_fair(struct rq *rq)
7346 static void rq_offline_fair(struct rq *rq)
7350 /* Ensure any throttled groups are reachable by pick_next_task */
7351 unthrottle_offline_cfs_rqs(rq);
7354 #endif /* CONFIG_SMP */
7357 * scheduler tick hitting a task of our scheduling class:
7359 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7361 struct cfs_rq *cfs_rq;
7362 struct sched_entity *se = &curr->se;
7364 for_each_sched_entity(se) {
7365 cfs_rq = cfs_rq_of(se);
7366 entity_tick(cfs_rq, se, queued);
7369 if (numabalancing_enabled)
7370 task_tick_numa(rq, curr);
7372 update_rq_runnable_avg(rq, 1);
7376 * called on fork with the child task as argument from the parent's context
7377 * - child not yet on the tasklist
7378 * - preemption disabled
7380 static void task_fork_fair(struct task_struct *p)
7382 struct cfs_rq *cfs_rq;
7383 struct sched_entity *se = &p->se, *curr;
7384 int this_cpu = smp_processor_id();
7385 struct rq *rq = this_rq();
7386 unsigned long flags;
7388 raw_spin_lock_irqsave(&rq->lock, flags);
7390 update_rq_clock(rq);
7392 cfs_rq = task_cfs_rq(current);
7393 curr = cfs_rq->curr;
7396 * Not only the cpu but also the task_group of the parent might have
7397 * been changed after parent->se.parent,cfs_rq were copied to
7398 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7399 * of child point to valid ones.
7402 __set_task_cpu(p, this_cpu);
7405 update_curr(cfs_rq);
7408 se->vruntime = curr->vruntime;
7409 place_entity(cfs_rq, se, 1);
7411 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7413 * Upon rescheduling, sched_class::put_prev_task() will place
7414 * 'current' within the tree based on its new key value.
7416 swap(curr->vruntime, se->vruntime);
7417 resched_task(rq->curr);
7420 se->vruntime -= cfs_rq->min_vruntime;
7422 raw_spin_unlock_irqrestore(&rq->lock, flags);
7426 * Priority of the task has changed. Check to see if we preempt
7430 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7436 * Reschedule if we are currently running on this runqueue and
7437 * our priority decreased, or if we are not currently running on
7438 * this runqueue and our priority is higher than the current's
7440 if (rq->curr == p) {
7441 if (p->prio > oldprio)
7442 resched_task(rq->curr);
7444 check_preempt_curr(rq, p, 0);
7447 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7449 struct sched_entity *se = &p->se;
7450 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7453 * Ensure the task's vruntime is normalized, so that when it's
7454 * switched back to the fair class the enqueue_entity(.flags=0) will
7455 * do the right thing.
7457 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7458 * have normalized the vruntime, if it's !on_rq, then only when
7459 * the task is sleeping will it still have non-normalized vruntime.
7461 if (!p->on_rq && p->state != TASK_RUNNING) {
7463 * Fix up our vruntime so that the current sleep doesn't
7464 * cause 'unlimited' sleep bonus.
7466 place_entity(cfs_rq, se, 0);
7467 se->vruntime -= cfs_rq->min_vruntime;
7472 * Remove our load from contribution when we leave sched_fair
7473 * and ensure we don't carry in an old decay_count if we
7476 if (se->avg.decay_count) {
7477 __synchronize_entity_decay(se);
7478 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7484 * We switched to the sched_fair class.
7486 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7488 struct sched_entity *se = &p->se;
7489 #ifdef CONFIG_FAIR_GROUP_SCHED
7491 * Since the real-depth could have been changed (only FAIR
7492 * class maintain depth value), reset depth properly.
7494 se->depth = se->parent ? se->parent->depth + 1 : 0;
7500 * We were most likely switched from sched_rt, so
7501 * kick off the schedule if running, otherwise just see
7502 * if we can still preempt the current task.
7505 resched_task(rq->curr);
7507 check_preempt_curr(rq, p, 0);
7510 /* Account for a task changing its policy or group.
7512 * This routine is mostly called to set cfs_rq->curr field when a task
7513 * migrates between groups/classes.
7515 static void set_curr_task_fair(struct rq *rq)
7517 struct sched_entity *se = &rq->curr->se;
7519 for_each_sched_entity(se) {
7520 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7522 set_next_entity(cfs_rq, se);
7523 /* ensure bandwidth has been allocated on our new cfs_rq */
7524 account_cfs_rq_runtime(cfs_rq, 0);
7528 void init_cfs_rq(struct cfs_rq *cfs_rq)
7530 cfs_rq->tasks_timeline = RB_ROOT;
7531 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7532 #ifndef CONFIG_64BIT
7533 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7536 atomic64_set(&cfs_rq->decay_counter, 1);
7537 atomic_long_set(&cfs_rq->removed_load, 0);
7541 #ifdef CONFIG_FAIR_GROUP_SCHED
7542 static void task_move_group_fair(struct task_struct *p, int on_rq)
7544 struct sched_entity *se = &p->se;
7545 struct cfs_rq *cfs_rq;
7548 * If the task was not on the rq at the time of this cgroup movement
7549 * it must have been asleep, sleeping tasks keep their ->vruntime
7550 * absolute on their old rq until wakeup (needed for the fair sleeper
7551 * bonus in place_entity()).
7553 * If it was on the rq, we've just 'preempted' it, which does convert
7554 * ->vruntime to a relative base.
7556 * Make sure both cases convert their relative position when migrating
7557 * to another cgroup's rq. This does somewhat interfere with the
7558 * fair sleeper stuff for the first placement, but who cares.
7561 * When !on_rq, vruntime of the task has usually NOT been normalized.
7562 * But there are some cases where it has already been normalized:
7564 * - Moving a forked child which is waiting for being woken up by
7565 * wake_up_new_task().
7566 * - Moving a task which has been woken up by try_to_wake_up() and
7567 * waiting for actually being woken up by sched_ttwu_pending().
7569 * To prevent boost or penalty in the new cfs_rq caused by delta
7570 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7572 if (!on_rq && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7576 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7577 set_task_rq(p, task_cpu(p));
7578 se->depth = se->parent ? se->parent->depth + 1 : 0;
7580 cfs_rq = cfs_rq_of(se);
7581 se->vruntime += cfs_rq->min_vruntime;
7584 * migrate_task_rq_fair() will have removed our previous
7585 * contribution, but we must synchronize for ongoing future
7588 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7589 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7594 void free_fair_sched_group(struct task_group *tg)
7598 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7600 for_each_possible_cpu(i) {
7602 kfree(tg->cfs_rq[i]);
7611 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7613 struct cfs_rq *cfs_rq;
7614 struct sched_entity *se;
7617 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7620 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7624 tg->shares = NICE_0_LOAD;
7626 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7628 for_each_possible_cpu(i) {
7629 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7630 GFP_KERNEL, cpu_to_node(i));
7634 se = kzalloc_node(sizeof(struct sched_entity),
7635 GFP_KERNEL, cpu_to_node(i));
7639 init_cfs_rq(cfs_rq);
7640 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7651 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7653 struct rq *rq = cpu_rq(cpu);
7654 unsigned long flags;
7657 * Only empty task groups can be destroyed; so we can speculatively
7658 * check on_list without danger of it being re-added.
7660 if (!tg->cfs_rq[cpu]->on_list)
7663 raw_spin_lock_irqsave(&rq->lock, flags);
7664 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7665 raw_spin_unlock_irqrestore(&rq->lock, flags);
7668 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7669 struct sched_entity *se, int cpu,
7670 struct sched_entity *parent)
7672 struct rq *rq = cpu_rq(cpu);
7676 init_cfs_rq_runtime(cfs_rq);
7678 tg->cfs_rq[cpu] = cfs_rq;
7681 /* se could be NULL for root_task_group */
7686 se->cfs_rq = &rq->cfs;
7689 se->cfs_rq = parent->my_q;
7690 se->depth = parent->depth + 1;
7694 /* guarantee group entities always have weight */
7695 update_load_set(&se->load, NICE_0_LOAD);
7696 se->parent = parent;
7699 static DEFINE_MUTEX(shares_mutex);
7701 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7704 unsigned long flags;
7707 * We can't change the weight of the root cgroup.
7712 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7714 mutex_lock(&shares_mutex);
7715 if (tg->shares == shares)
7718 tg->shares = shares;
7719 for_each_possible_cpu(i) {
7720 struct rq *rq = cpu_rq(i);
7721 struct sched_entity *se;
7724 /* Propagate contribution to hierarchy */
7725 raw_spin_lock_irqsave(&rq->lock, flags);
7727 /* Possible calls to update_curr() need rq clock */
7728 update_rq_clock(rq);
7729 for_each_sched_entity(se)
7730 update_cfs_shares(group_cfs_rq(se));
7731 raw_spin_unlock_irqrestore(&rq->lock, flags);
7735 mutex_unlock(&shares_mutex);
7738 #else /* CONFIG_FAIR_GROUP_SCHED */
7740 void free_fair_sched_group(struct task_group *tg) { }
7742 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7747 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7749 #endif /* CONFIG_FAIR_GROUP_SCHED */
7752 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7754 struct sched_entity *se = &task->se;
7755 unsigned int rr_interval = 0;
7758 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7761 if (rq->cfs.load.weight)
7762 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7768 * All the scheduling class methods:
7770 const struct sched_class fair_sched_class = {
7771 .next = &idle_sched_class,
7772 .enqueue_task = enqueue_task_fair,
7773 .dequeue_task = dequeue_task_fair,
7774 .yield_task = yield_task_fair,
7775 .yield_to_task = yield_to_task_fair,
7777 .check_preempt_curr = check_preempt_wakeup,
7779 .pick_next_task = pick_next_task_fair,
7780 .put_prev_task = put_prev_task_fair,
7783 .select_task_rq = select_task_rq_fair,
7784 .migrate_task_rq = migrate_task_rq_fair,
7786 .rq_online = rq_online_fair,
7787 .rq_offline = rq_offline_fair,
7789 .task_waking = task_waking_fair,
7792 .set_curr_task = set_curr_task_fair,
7793 .task_tick = task_tick_fair,
7794 .task_fork = task_fork_fair,
7796 .prio_changed = prio_changed_fair,
7797 .switched_from = switched_from_fair,
7798 .switched_to = switched_to_fair,
7800 .get_rr_interval = get_rr_interval_fair,
7802 #ifdef CONFIG_FAIR_GROUP_SCHED
7803 .task_move_group = task_move_group_fair,
7807 #ifdef CONFIG_SCHED_DEBUG
7808 void print_cfs_stats(struct seq_file *m, int cpu)
7810 struct cfs_rq *cfs_rq;
7813 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7814 print_cfs_rq(m, cpu, cfs_rq);
7819 __init void init_sched_fair_class(void)
7822 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7824 #ifdef CONFIG_NO_HZ_COMMON
7825 nohz.next_balance = jiffies;
7826 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7827 cpu_notifier(sched_ilb_notifier, 0);