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 power_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 power;
1031 /* Approximate capacity in terms of runnable tasks on a node */
1032 unsigned long 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->power += power_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_capacity, or we'll detect a huge imbalance
1065 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1066 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1067 ns->has_capacity = (ns->nr_running < ns->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 orig_src_load, long orig_dst_load,
1099 long src_load, long dst_load,
1100 struct task_numa_env *env)
1104 /* We care about the slope of the imbalance, not the direction. */
1105 if (dst_load < src_load)
1106 swap(dst_load, src_load);
1108 /* Is the difference below the threshold? */
1109 imb = dst_load * 100 - src_load * env->imbalance_pct;
1114 * The imbalance is above the allowed threshold.
1115 * Compare it with the old imbalance.
1117 if (orig_dst_load < orig_src_load)
1118 swap(orig_dst_load, orig_src_load);
1120 old_imb = orig_dst_load * 100 - orig_src_load * env->imbalance_pct;
1122 /* Would this change make things worse? */
1123 return (old_imb > imb);
1127 * This checks if the overall compute and NUMA accesses of the system would
1128 * be improved if the source tasks was migrated to the target dst_cpu taking
1129 * into account that it might be best if task running on the dst_cpu should
1130 * be exchanged with the source task
1132 static void task_numa_compare(struct task_numa_env *env,
1133 long taskimp, long groupimp)
1135 struct rq *src_rq = cpu_rq(env->src_cpu);
1136 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1137 struct task_struct *cur;
1138 long orig_src_load, src_load;
1139 long orig_dst_load, dst_load;
1141 long imp = (groupimp > 0) ? groupimp : taskimp;
1144 cur = ACCESS_ONCE(dst_rq->curr);
1145 if (cur->pid == 0) /* idle */
1149 * "imp" is the fault differential for the source task between the
1150 * source and destination node. Calculate the total differential for
1151 * the source task and potential destination task. The more negative
1152 * the value is, the more rmeote accesses that would be expected to
1153 * be incurred if the tasks were swapped.
1156 /* Skip this swap candidate if cannot move to the source cpu */
1157 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1161 * If dst and source tasks are in the same NUMA group, or not
1162 * in any group then look only at task weights.
1164 if (cur->numa_group == env->p->numa_group) {
1165 imp = taskimp + task_weight(cur, env->src_nid) -
1166 task_weight(cur, env->dst_nid);
1168 * Add some hysteresis to prevent swapping the
1169 * tasks within a group over tiny differences.
1171 if (cur->numa_group)
1175 * Compare the group weights. If a task is all by
1176 * itself (not part of a group), use the task weight
1179 if (env->p->numa_group)
1184 if (cur->numa_group)
1185 imp += group_weight(cur, env->src_nid) -
1186 group_weight(cur, env->dst_nid);
1188 imp += task_weight(cur, env->src_nid) -
1189 task_weight(cur, env->dst_nid);
1193 if (imp < env->best_imp)
1197 /* Is there capacity at our destination? */
1198 if (env->src_stats.has_capacity &&
1199 !env->dst_stats.has_capacity)
1205 /* Balance doesn't matter much if we're running a task per cpu */
1206 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1210 * In the overloaded case, try and keep the load balanced.
1213 orig_dst_load = env->dst_stats.load;
1214 orig_src_load = env->src_stats.load;
1216 /* XXX missing power terms */
1217 load = task_h_load(env->p);
1218 dst_load = orig_dst_load + load;
1219 src_load = orig_src_load - load;
1222 load = task_h_load(cur);
1227 if (load_too_imbalanced(orig_src_load, orig_dst_load,
1228 src_load, dst_load, env))
1232 task_numa_assign(env, cur, imp);
1237 static void task_numa_find_cpu(struct task_numa_env *env,
1238 long taskimp, long groupimp)
1242 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1243 /* Skip this CPU if the source task cannot migrate */
1244 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1248 task_numa_compare(env, taskimp, groupimp);
1252 static int task_numa_migrate(struct task_struct *p)
1254 struct task_numa_env env = {
1257 .src_cpu = task_cpu(p),
1258 .src_nid = task_node(p),
1260 .imbalance_pct = 112,
1266 struct sched_domain *sd;
1267 unsigned long taskweight, groupweight;
1269 long taskimp, groupimp;
1272 * Pick the lowest SD_NUMA domain, as that would have the smallest
1273 * imbalance and would be the first to start moving tasks about.
1275 * And we want to avoid any moving of tasks about, as that would create
1276 * random movement of tasks -- counter the numa conditions we're trying
1280 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1282 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1286 * Cpusets can break the scheduler domain tree into smaller
1287 * balance domains, some of which do not cross NUMA boundaries.
1288 * Tasks that are "trapped" in such domains cannot be migrated
1289 * elsewhere, so there is no point in (re)trying.
1291 if (unlikely(!sd)) {
1292 p->numa_preferred_nid = task_node(p);
1296 taskweight = task_weight(p, env.src_nid);
1297 groupweight = group_weight(p, env.src_nid);
1298 update_numa_stats(&env.src_stats, env.src_nid);
1299 env.dst_nid = p->numa_preferred_nid;
1300 taskimp = task_weight(p, env.dst_nid) - taskweight;
1301 groupimp = group_weight(p, env.dst_nid) - groupweight;
1302 update_numa_stats(&env.dst_stats, env.dst_nid);
1304 /* If the preferred nid has capacity, try to use it. */
1305 if (env.dst_stats.has_capacity)
1306 task_numa_find_cpu(&env, taskimp, groupimp);
1308 /* No space available on the preferred nid. Look elsewhere. */
1309 if (env.best_cpu == -1) {
1310 for_each_online_node(nid) {
1311 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1314 /* Only consider nodes where both task and groups benefit */
1315 taskimp = task_weight(p, nid) - taskweight;
1316 groupimp = group_weight(p, nid) - groupweight;
1317 if (taskimp < 0 && groupimp < 0)
1321 update_numa_stats(&env.dst_stats, env.dst_nid);
1322 task_numa_find_cpu(&env, taskimp, groupimp);
1326 /* No better CPU than the current one was found. */
1327 if (env.best_cpu == -1)
1331 * If the task is part of a workload that spans multiple NUMA nodes,
1332 * and is migrating into one of the workload's active nodes, remember
1333 * this node as the task's preferred numa node, so the workload can
1335 * A task that migrated to a second choice node will be better off
1336 * trying for a better one later. Do not set the preferred node here.
1338 if (p->numa_group && node_isset(env.dst_nid, p->numa_group->active_nodes))
1339 sched_setnuma(p, env.dst_nid);
1342 * Reset the scan period if the task is being rescheduled on an
1343 * alternative node to recheck if the tasks is now properly placed.
1345 p->numa_scan_period = task_scan_min(p);
1347 if (env.best_task == NULL) {
1348 ret = migrate_task_to(p, env.best_cpu);
1350 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1354 ret = migrate_swap(p, env.best_task);
1356 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1357 put_task_struct(env.best_task);
1361 /* Attempt to migrate a task to a CPU on the preferred node. */
1362 static void numa_migrate_preferred(struct task_struct *p)
1364 unsigned long interval = HZ;
1366 /* This task has no NUMA fault statistics yet */
1367 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1370 /* Periodically retry migrating the task to the preferred node */
1371 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1372 p->numa_migrate_retry = jiffies + interval;
1374 /* Success if task is already running on preferred CPU */
1375 if (task_node(p) == p->numa_preferred_nid)
1378 /* Otherwise, try migrate to a CPU on the preferred node */
1379 task_numa_migrate(p);
1383 * Find the nodes on which the workload is actively running. We do this by
1384 * tracking the nodes from which NUMA hinting faults are triggered. This can
1385 * be different from the set of nodes where the workload's memory is currently
1388 * The bitmask is used to make smarter decisions on when to do NUMA page
1389 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1390 * are added when they cause over 6/16 of the maximum number of faults, but
1391 * only removed when they drop below 3/16.
1393 static void update_numa_active_node_mask(struct numa_group *numa_group)
1395 unsigned long faults, max_faults = 0;
1398 for_each_online_node(nid) {
1399 faults = group_faults_cpu(numa_group, nid);
1400 if (faults > max_faults)
1401 max_faults = faults;
1404 for_each_online_node(nid) {
1405 faults = group_faults_cpu(numa_group, nid);
1406 if (!node_isset(nid, numa_group->active_nodes)) {
1407 if (faults > max_faults * 6 / 16)
1408 node_set(nid, numa_group->active_nodes);
1409 } else if (faults < max_faults * 3 / 16)
1410 node_clear(nid, numa_group->active_nodes);
1415 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1416 * increments. The more local the fault statistics are, the higher the scan
1417 * period will be for the next scan window. If local/remote ratio is below
1418 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1419 * scan period will decrease
1421 #define NUMA_PERIOD_SLOTS 10
1422 #define NUMA_PERIOD_THRESHOLD 3
1425 * Increase the scan period (slow down scanning) if the majority of
1426 * our memory is already on our local node, or if the majority of
1427 * the page accesses are shared with other processes.
1428 * Otherwise, decrease the scan period.
1430 static void update_task_scan_period(struct task_struct *p,
1431 unsigned long shared, unsigned long private)
1433 unsigned int period_slot;
1437 unsigned long remote = p->numa_faults_locality[0];
1438 unsigned long local = p->numa_faults_locality[1];
1441 * If there were no record hinting faults then either the task is
1442 * completely idle or all activity is areas that are not of interest
1443 * to automatic numa balancing. Scan slower
1445 if (local + shared == 0) {
1446 p->numa_scan_period = min(p->numa_scan_period_max,
1447 p->numa_scan_period << 1);
1449 p->mm->numa_next_scan = jiffies +
1450 msecs_to_jiffies(p->numa_scan_period);
1456 * Prepare to scale scan period relative to the current period.
1457 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1458 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1459 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1461 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1462 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1463 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1464 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1467 diff = slot * period_slot;
1469 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1472 * Scale scan rate increases based on sharing. There is an
1473 * inverse relationship between the degree of sharing and
1474 * the adjustment made to the scanning period. Broadly
1475 * speaking the intent is that there is little point
1476 * scanning faster if shared accesses dominate as it may
1477 * simply bounce migrations uselessly
1479 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1480 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1483 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1484 task_scan_min(p), task_scan_max(p));
1485 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1489 * Get the fraction of time the task has been running since the last
1490 * NUMA placement cycle. The scheduler keeps similar statistics, but
1491 * decays those on a 32ms period, which is orders of magnitude off
1492 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1493 * stats only if the task is so new there are no NUMA statistics yet.
1495 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1497 u64 runtime, delta, now;
1498 /* Use the start of this time slice to avoid calculations. */
1499 now = p->se.exec_start;
1500 runtime = p->se.sum_exec_runtime;
1502 if (p->last_task_numa_placement) {
1503 delta = runtime - p->last_sum_exec_runtime;
1504 *period = now - p->last_task_numa_placement;
1506 delta = p->se.avg.runnable_avg_sum;
1507 *period = p->se.avg.runnable_avg_period;
1510 p->last_sum_exec_runtime = runtime;
1511 p->last_task_numa_placement = now;
1516 static void task_numa_placement(struct task_struct *p)
1518 int seq, nid, max_nid = -1, max_group_nid = -1;
1519 unsigned long max_faults = 0, max_group_faults = 0;
1520 unsigned long fault_types[2] = { 0, 0 };
1521 unsigned long total_faults;
1522 u64 runtime, period;
1523 spinlock_t *group_lock = NULL;
1525 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1526 if (p->numa_scan_seq == seq)
1528 p->numa_scan_seq = seq;
1529 p->numa_scan_period_max = task_scan_max(p);
1531 total_faults = p->numa_faults_locality[0] +
1532 p->numa_faults_locality[1];
1533 runtime = numa_get_avg_runtime(p, &period);
1535 /* If the task is part of a group prevent parallel updates to group stats */
1536 if (p->numa_group) {
1537 group_lock = &p->numa_group->lock;
1538 spin_lock_irq(group_lock);
1541 /* Find the node with the highest number of faults */
1542 for_each_online_node(nid) {
1543 unsigned long faults = 0, group_faults = 0;
1546 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1547 long diff, f_diff, f_weight;
1549 i = task_faults_idx(nid, priv);
1551 /* Decay existing window, copy faults since last scan */
1552 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1553 fault_types[priv] += p->numa_faults_buffer_memory[i];
1554 p->numa_faults_buffer_memory[i] = 0;
1557 * Normalize the faults_from, so all tasks in a group
1558 * count according to CPU use, instead of by the raw
1559 * number of faults. Tasks with little runtime have
1560 * little over-all impact on throughput, and thus their
1561 * faults are less important.
1563 f_weight = div64_u64(runtime << 16, period + 1);
1564 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1566 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1567 p->numa_faults_buffer_cpu[i] = 0;
1569 p->numa_faults_memory[i] += diff;
1570 p->numa_faults_cpu[i] += f_diff;
1571 faults += p->numa_faults_memory[i];
1572 p->total_numa_faults += diff;
1573 if (p->numa_group) {
1574 /* safe because we can only change our own group */
1575 p->numa_group->faults[i] += diff;
1576 p->numa_group->faults_cpu[i] += f_diff;
1577 p->numa_group->total_faults += diff;
1578 group_faults += p->numa_group->faults[i];
1582 if (faults > max_faults) {
1583 max_faults = faults;
1587 if (group_faults > max_group_faults) {
1588 max_group_faults = group_faults;
1589 max_group_nid = nid;
1593 update_task_scan_period(p, fault_types[0], fault_types[1]);
1595 if (p->numa_group) {
1596 update_numa_active_node_mask(p->numa_group);
1598 * If the preferred task and group nids are different,
1599 * iterate over the nodes again to find the best place.
1601 if (max_nid != max_group_nid) {
1602 unsigned long weight, max_weight = 0;
1604 for_each_online_node(nid) {
1605 weight = task_weight(p, nid) + group_weight(p, nid);
1606 if (weight > max_weight) {
1607 max_weight = weight;
1613 spin_unlock_irq(group_lock);
1616 /* Preferred node as the node with the most faults */
1617 if (max_faults && max_nid != p->numa_preferred_nid) {
1618 /* Update the preferred nid and migrate task if possible */
1619 sched_setnuma(p, max_nid);
1620 numa_migrate_preferred(p);
1624 static inline int get_numa_group(struct numa_group *grp)
1626 return atomic_inc_not_zero(&grp->refcount);
1629 static inline void put_numa_group(struct numa_group *grp)
1631 if (atomic_dec_and_test(&grp->refcount))
1632 kfree_rcu(grp, rcu);
1635 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1638 struct numa_group *grp, *my_grp;
1639 struct task_struct *tsk;
1641 int cpu = cpupid_to_cpu(cpupid);
1644 if (unlikely(!p->numa_group)) {
1645 unsigned int size = sizeof(struct numa_group) +
1646 4*nr_node_ids*sizeof(unsigned long);
1648 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1652 atomic_set(&grp->refcount, 1);
1653 spin_lock_init(&grp->lock);
1654 INIT_LIST_HEAD(&grp->task_list);
1656 /* Second half of the array tracks nids where faults happen */
1657 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1660 node_set(task_node(current), grp->active_nodes);
1662 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1663 grp->faults[i] = p->numa_faults_memory[i];
1665 grp->total_faults = p->total_numa_faults;
1667 list_add(&p->numa_entry, &grp->task_list);
1669 rcu_assign_pointer(p->numa_group, grp);
1673 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1675 if (!cpupid_match_pid(tsk, cpupid))
1678 grp = rcu_dereference(tsk->numa_group);
1682 my_grp = p->numa_group;
1687 * Only join the other group if its bigger; if we're the bigger group,
1688 * the other task will join us.
1690 if (my_grp->nr_tasks > grp->nr_tasks)
1694 * Tie-break on the grp address.
1696 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1699 /* Always join threads in the same process. */
1700 if (tsk->mm == current->mm)
1703 /* Simple filter to avoid false positives due to PID collisions */
1704 if (flags & TNF_SHARED)
1707 /* Update priv based on whether false sharing was detected */
1710 if (join && !get_numa_group(grp))
1718 BUG_ON(irqs_disabled());
1719 double_lock_irq(&my_grp->lock, &grp->lock);
1721 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1722 my_grp->faults[i] -= p->numa_faults_memory[i];
1723 grp->faults[i] += p->numa_faults_memory[i];
1725 my_grp->total_faults -= p->total_numa_faults;
1726 grp->total_faults += p->total_numa_faults;
1728 list_move(&p->numa_entry, &grp->task_list);
1732 spin_unlock(&my_grp->lock);
1733 spin_unlock_irq(&grp->lock);
1735 rcu_assign_pointer(p->numa_group, grp);
1737 put_numa_group(my_grp);
1745 void task_numa_free(struct task_struct *p)
1747 struct numa_group *grp = p->numa_group;
1749 void *numa_faults = p->numa_faults_memory;
1752 spin_lock_irq(&grp->lock);
1753 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1754 grp->faults[i] -= p->numa_faults_memory[i];
1755 grp->total_faults -= p->total_numa_faults;
1757 list_del(&p->numa_entry);
1759 spin_unlock_irq(&grp->lock);
1760 rcu_assign_pointer(p->numa_group, NULL);
1761 put_numa_group(grp);
1764 p->numa_faults_memory = NULL;
1765 p->numa_faults_buffer_memory = NULL;
1766 p->numa_faults_cpu= NULL;
1767 p->numa_faults_buffer_cpu = NULL;
1772 * Got a PROT_NONE fault for a page on @node.
1774 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1776 struct task_struct *p = current;
1777 bool migrated = flags & TNF_MIGRATED;
1778 int cpu_node = task_node(current);
1779 int local = !!(flags & TNF_FAULT_LOCAL);
1782 if (!numabalancing_enabled)
1785 /* for example, ksmd faulting in a user's mm */
1789 /* Do not worry about placement if exiting */
1790 if (p->state == TASK_DEAD)
1793 /* Allocate buffer to track faults on a per-node basis */
1794 if (unlikely(!p->numa_faults_memory)) {
1795 int size = sizeof(*p->numa_faults_memory) *
1796 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1798 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1799 if (!p->numa_faults_memory)
1802 BUG_ON(p->numa_faults_buffer_memory);
1804 * The averaged statistics, shared & private, memory & cpu,
1805 * occupy the first half of the array. The second half of the
1806 * array is for current counters, which are averaged into the
1807 * first set by task_numa_placement.
1809 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1810 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1811 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1812 p->total_numa_faults = 0;
1813 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1817 * First accesses are treated as private, otherwise consider accesses
1818 * to be private if the accessing pid has not changed
1820 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1823 priv = cpupid_match_pid(p, last_cpupid);
1824 if (!priv && !(flags & TNF_NO_GROUP))
1825 task_numa_group(p, last_cpupid, flags, &priv);
1829 * If a workload spans multiple NUMA nodes, a shared fault that
1830 * occurs wholly within the set of nodes that the workload is
1831 * actively using should be counted as local. This allows the
1832 * scan rate to slow down when a workload has settled down.
1834 if (!priv && !local && p->numa_group &&
1835 node_isset(cpu_node, p->numa_group->active_nodes) &&
1836 node_isset(mem_node, p->numa_group->active_nodes))
1839 task_numa_placement(p);
1842 * Retry task to preferred node migration periodically, in case it
1843 * case it previously failed, or the scheduler moved us.
1845 if (time_after(jiffies, p->numa_migrate_retry))
1846 numa_migrate_preferred(p);
1849 p->numa_pages_migrated += pages;
1851 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1852 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1853 p->numa_faults_locality[local] += pages;
1856 static void reset_ptenuma_scan(struct task_struct *p)
1858 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1859 p->mm->numa_scan_offset = 0;
1863 * The expensive part of numa migration is done from task_work context.
1864 * Triggered from task_tick_numa().
1866 void task_numa_work(struct callback_head *work)
1868 unsigned long migrate, next_scan, now = jiffies;
1869 struct task_struct *p = current;
1870 struct mm_struct *mm = p->mm;
1871 struct vm_area_struct *vma;
1872 unsigned long start, end;
1873 unsigned long nr_pte_updates = 0;
1876 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1878 work->next = work; /* protect against double add */
1880 * Who cares about NUMA placement when they're dying.
1882 * NOTE: make sure not to dereference p->mm before this check,
1883 * exit_task_work() happens _after_ exit_mm() so we could be called
1884 * without p->mm even though we still had it when we enqueued this
1887 if (p->flags & PF_EXITING)
1890 if (!mm->numa_next_scan) {
1891 mm->numa_next_scan = now +
1892 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1896 * Enforce maximal scan/migration frequency..
1898 migrate = mm->numa_next_scan;
1899 if (time_before(now, migrate))
1902 if (p->numa_scan_period == 0) {
1903 p->numa_scan_period_max = task_scan_max(p);
1904 p->numa_scan_period = task_scan_min(p);
1907 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1908 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1912 * Delay this task enough that another task of this mm will likely win
1913 * the next time around.
1915 p->node_stamp += 2 * TICK_NSEC;
1917 start = mm->numa_scan_offset;
1918 pages = sysctl_numa_balancing_scan_size;
1919 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1923 down_read(&mm->mmap_sem);
1924 vma = find_vma(mm, start);
1926 reset_ptenuma_scan(p);
1930 for (; vma; vma = vma->vm_next) {
1931 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1935 * Shared library pages mapped by multiple processes are not
1936 * migrated as it is expected they are cache replicated. Avoid
1937 * hinting faults in read-only file-backed mappings or the vdso
1938 * as migrating the pages will be of marginal benefit.
1941 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1945 * Skip inaccessible VMAs to avoid any confusion between
1946 * PROT_NONE and NUMA hinting ptes
1948 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1952 start = max(start, vma->vm_start);
1953 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1954 end = min(end, vma->vm_end);
1955 nr_pte_updates += change_prot_numa(vma, start, end);
1958 * Scan sysctl_numa_balancing_scan_size but ensure that
1959 * at least one PTE is updated so that unused virtual
1960 * address space is quickly skipped.
1963 pages -= (end - start) >> PAGE_SHIFT;
1970 } while (end != vma->vm_end);
1975 * It is possible to reach the end of the VMA list but the last few
1976 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1977 * would find the !migratable VMA on the next scan but not reset the
1978 * scanner to the start so check it now.
1981 mm->numa_scan_offset = start;
1983 reset_ptenuma_scan(p);
1984 up_read(&mm->mmap_sem);
1988 * Drive the periodic memory faults..
1990 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1992 struct callback_head *work = &curr->numa_work;
1996 * We don't care about NUMA placement if we don't have memory.
1998 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2002 * Using runtime rather than walltime has the dual advantage that
2003 * we (mostly) drive the selection from busy threads and that the
2004 * task needs to have done some actual work before we bother with
2007 now = curr->se.sum_exec_runtime;
2008 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2010 if (now - curr->node_stamp > period) {
2011 if (!curr->node_stamp)
2012 curr->numa_scan_period = task_scan_min(curr);
2013 curr->node_stamp += period;
2015 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2016 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2017 task_work_add(curr, work, true);
2022 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2026 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2030 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2033 #endif /* CONFIG_NUMA_BALANCING */
2036 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2038 update_load_add(&cfs_rq->load, se->load.weight);
2039 if (!parent_entity(se))
2040 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2042 if (entity_is_task(se)) {
2043 struct rq *rq = rq_of(cfs_rq);
2045 account_numa_enqueue(rq, task_of(se));
2046 list_add(&se->group_node, &rq->cfs_tasks);
2049 cfs_rq->nr_running++;
2053 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2055 update_load_sub(&cfs_rq->load, se->load.weight);
2056 if (!parent_entity(se))
2057 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2058 if (entity_is_task(se)) {
2059 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2060 list_del_init(&se->group_node);
2062 cfs_rq->nr_running--;
2065 #ifdef CONFIG_FAIR_GROUP_SCHED
2067 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2072 * Use this CPU's actual weight instead of the last load_contribution
2073 * to gain a more accurate current total weight. See
2074 * update_cfs_rq_load_contribution().
2076 tg_weight = atomic_long_read(&tg->load_avg);
2077 tg_weight -= cfs_rq->tg_load_contrib;
2078 tg_weight += cfs_rq->load.weight;
2083 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2085 long tg_weight, load, shares;
2087 tg_weight = calc_tg_weight(tg, cfs_rq);
2088 load = cfs_rq->load.weight;
2090 shares = (tg->shares * load);
2092 shares /= tg_weight;
2094 if (shares < MIN_SHARES)
2095 shares = MIN_SHARES;
2096 if (shares > tg->shares)
2097 shares = tg->shares;
2101 # else /* CONFIG_SMP */
2102 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2106 # endif /* CONFIG_SMP */
2107 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2108 unsigned long weight)
2111 /* commit outstanding execution time */
2112 if (cfs_rq->curr == se)
2113 update_curr(cfs_rq);
2114 account_entity_dequeue(cfs_rq, se);
2117 update_load_set(&se->load, weight);
2120 account_entity_enqueue(cfs_rq, se);
2123 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2125 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2127 struct task_group *tg;
2128 struct sched_entity *se;
2132 se = tg->se[cpu_of(rq_of(cfs_rq))];
2133 if (!se || throttled_hierarchy(cfs_rq))
2136 if (likely(se->load.weight == tg->shares))
2139 shares = calc_cfs_shares(cfs_rq, tg);
2141 reweight_entity(cfs_rq_of(se), se, shares);
2143 #else /* CONFIG_FAIR_GROUP_SCHED */
2144 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2147 #endif /* CONFIG_FAIR_GROUP_SCHED */
2151 * We choose a half-life close to 1 scheduling period.
2152 * Note: The tables below are dependent on this value.
2154 #define LOAD_AVG_PERIOD 32
2155 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2156 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2158 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2159 static const u32 runnable_avg_yN_inv[] = {
2160 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2161 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2162 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2163 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2164 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2165 0x85aac367, 0x82cd8698,
2169 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2170 * over-estimates when re-combining.
2172 static const u32 runnable_avg_yN_sum[] = {
2173 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2174 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2175 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2180 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2182 static __always_inline u64 decay_load(u64 val, u64 n)
2184 unsigned int local_n;
2188 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2191 /* after bounds checking we can collapse to 32-bit */
2195 * As y^PERIOD = 1/2, we can combine
2196 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2197 * With a look-up table which covers k^n (n<PERIOD)
2199 * To achieve constant time decay_load.
2201 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2202 val >>= local_n / LOAD_AVG_PERIOD;
2203 local_n %= LOAD_AVG_PERIOD;
2206 val *= runnable_avg_yN_inv[local_n];
2207 /* We don't use SRR here since we always want to round down. */
2212 * For updates fully spanning n periods, the contribution to runnable
2213 * average will be: \Sum 1024*y^n
2215 * We can compute this reasonably efficiently by combining:
2216 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2218 static u32 __compute_runnable_contrib(u64 n)
2222 if (likely(n <= LOAD_AVG_PERIOD))
2223 return runnable_avg_yN_sum[n];
2224 else if (unlikely(n >= LOAD_AVG_MAX_N))
2225 return LOAD_AVG_MAX;
2227 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2229 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2230 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2232 n -= LOAD_AVG_PERIOD;
2233 } while (n > LOAD_AVG_PERIOD);
2235 contrib = decay_load(contrib, n);
2236 return contrib + runnable_avg_yN_sum[n];
2240 * We can represent the historical contribution to runnable average as the
2241 * coefficients of a geometric series. To do this we sub-divide our runnable
2242 * history into segments of approximately 1ms (1024us); label the segment that
2243 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2245 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2247 * (now) (~1ms ago) (~2ms ago)
2249 * Let u_i denote the fraction of p_i that the entity was runnable.
2251 * We then designate the fractions u_i as our co-efficients, yielding the
2252 * following representation of historical load:
2253 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2255 * We choose y based on the with of a reasonably scheduling period, fixing:
2258 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2259 * approximately half as much as the contribution to load within the last ms
2262 * When a period "rolls over" and we have new u_0`, multiplying the previous
2263 * sum again by y is sufficient to update:
2264 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2265 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2267 static __always_inline int __update_entity_runnable_avg(u64 now,
2268 struct sched_avg *sa,
2272 u32 runnable_contrib;
2273 int delta_w, decayed = 0;
2275 delta = now - sa->last_runnable_update;
2277 * This should only happen when time goes backwards, which it
2278 * unfortunately does during sched clock init when we swap over to TSC.
2280 if ((s64)delta < 0) {
2281 sa->last_runnable_update = now;
2286 * Use 1024ns as the unit of measurement since it's a reasonable
2287 * approximation of 1us and fast to compute.
2292 sa->last_runnable_update = now;
2294 /* delta_w is the amount already accumulated against our next period */
2295 delta_w = sa->runnable_avg_period % 1024;
2296 if (delta + delta_w >= 1024) {
2297 /* period roll-over */
2301 * Now that we know we're crossing a period boundary, figure
2302 * out how much from delta we need to complete the current
2303 * period and accrue it.
2305 delta_w = 1024 - delta_w;
2307 sa->runnable_avg_sum += delta_w;
2308 sa->runnable_avg_period += delta_w;
2312 /* Figure out how many additional periods this update spans */
2313 periods = delta / 1024;
2316 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2318 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2321 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2322 runnable_contrib = __compute_runnable_contrib(periods);
2324 sa->runnable_avg_sum += runnable_contrib;
2325 sa->runnable_avg_period += runnable_contrib;
2328 /* Remainder of delta accrued against u_0` */
2330 sa->runnable_avg_sum += delta;
2331 sa->runnable_avg_period += delta;
2336 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2337 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2339 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2340 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2342 decays -= se->avg.decay_count;
2346 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2347 se->avg.decay_count = 0;
2352 #ifdef CONFIG_FAIR_GROUP_SCHED
2353 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2356 struct task_group *tg = cfs_rq->tg;
2359 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2360 tg_contrib -= cfs_rq->tg_load_contrib;
2362 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2363 atomic_long_add(tg_contrib, &tg->load_avg);
2364 cfs_rq->tg_load_contrib += tg_contrib;
2369 * Aggregate cfs_rq runnable averages into an equivalent task_group
2370 * representation for computing load contributions.
2372 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2373 struct cfs_rq *cfs_rq)
2375 struct task_group *tg = cfs_rq->tg;
2378 /* The fraction of a cpu used by this cfs_rq */
2379 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2380 sa->runnable_avg_period + 1);
2381 contrib -= cfs_rq->tg_runnable_contrib;
2383 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2384 atomic_add(contrib, &tg->runnable_avg);
2385 cfs_rq->tg_runnable_contrib += contrib;
2389 static inline void __update_group_entity_contrib(struct sched_entity *se)
2391 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2392 struct task_group *tg = cfs_rq->tg;
2397 contrib = cfs_rq->tg_load_contrib * tg->shares;
2398 se->avg.load_avg_contrib = div_u64(contrib,
2399 atomic_long_read(&tg->load_avg) + 1);
2402 * For group entities we need to compute a correction term in the case
2403 * that they are consuming <1 cpu so that we would contribute the same
2404 * load as a task of equal weight.
2406 * Explicitly co-ordinating this measurement would be expensive, but
2407 * fortunately the sum of each cpus contribution forms a usable
2408 * lower-bound on the true value.
2410 * Consider the aggregate of 2 contributions. Either they are disjoint
2411 * (and the sum represents true value) or they are disjoint and we are
2412 * understating by the aggregate of their overlap.
2414 * Extending this to N cpus, for a given overlap, the maximum amount we
2415 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2416 * cpus that overlap for this interval and w_i is the interval width.
2418 * On a small machine; the first term is well-bounded which bounds the
2419 * total error since w_i is a subset of the period. Whereas on a
2420 * larger machine, while this first term can be larger, if w_i is the
2421 * of consequential size guaranteed to see n_i*w_i quickly converge to
2422 * our upper bound of 1-cpu.
2424 runnable_avg = atomic_read(&tg->runnable_avg);
2425 if (runnable_avg < NICE_0_LOAD) {
2426 se->avg.load_avg_contrib *= runnable_avg;
2427 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2431 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2433 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2434 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2436 #else /* CONFIG_FAIR_GROUP_SCHED */
2437 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2438 int force_update) {}
2439 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2440 struct cfs_rq *cfs_rq) {}
2441 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2442 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2443 #endif /* CONFIG_FAIR_GROUP_SCHED */
2445 static inline void __update_task_entity_contrib(struct sched_entity *se)
2449 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2450 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2451 contrib /= (se->avg.runnable_avg_period + 1);
2452 se->avg.load_avg_contrib = scale_load(contrib);
2455 /* Compute the current contribution to load_avg by se, return any delta */
2456 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2458 long old_contrib = se->avg.load_avg_contrib;
2460 if (entity_is_task(se)) {
2461 __update_task_entity_contrib(se);
2463 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2464 __update_group_entity_contrib(se);
2467 return se->avg.load_avg_contrib - old_contrib;
2470 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2473 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2474 cfs_rq->blocked_load_avg -= load_contrib;
2476 cfs_rq->blocked_load_avg = 0;
2479 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2481 /* Update a sched_entity's runnable average */
2482 static inline void update_entity_load_avg(struct sched_entity *se,
2485 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2490 * For a group entity we need to use their owned cfs_rq_clock_task() in
2491 * case they are the parent of a throttled hierarchy.
2493 if (entity_is_task(se))
2494 now = cfs_rq_clock_task(cfs_rq);
2496 now = cfs_rq_clock_task(group_cfs_rq(se));
2498 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2501 contrib_delta = __update_entity_load_avg_contrib(se);
2507 cfs_rq->runnable_load_avg += contrib_delta;
2509 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2513 * Decay the load contributed by all blocked children and account this so that
2514 * their contribution may appropriately discounted when they wake up.
2516 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2518 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2521 decays = now - cfs_rq->last_decay;
2522 if (!decays && !force_update)
2525 if (atomic_long_read(&cfs_rq->removed_load)) {
2526 unsigned long removed_load;
2527 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2528 subtract_blocked_load_contrib(cfs_rq, removed_load);
2532 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2534 atomic64_add(decays, &cfs_rq->decay_counter);
2535 cfs_rq->last_decay = now;
2538 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2541 /* Add the load generated by se into cfs_rq's child load-average */
2542 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2543 struct sched_entity *se,
2547 * We track migrations using entity decay_count <= 0, on a wake-up
2548 * migration we use a negative decay count to track the remote decays
2549 * accumulated while sleeping.
2551 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2552 * are seen by enqueue_entity_load_avg() as a migration with an already
2553 * constructed load_avg_contrib.
2555 if (unlikely(se->avg.decay_count <= 0)) {
2556 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2557 if (se->avg.decay_count) {
2559 * In a wake-up migration we have to approximate the
2560 * time sleeping. This is because we can't synchronize
2561 * clock_task between the two cpus, and it is not
2562 * guaranteed to be read-safe. Instead, we can
2563 * approximate this using our carried decays, which are
2564 * explicitly atomically readable.
2566 se->avg.last_runnable_update -= (-se->avg.decay_count)
2568 update_entity_load_avg(se, 0);
2569 /* Indicate that we're now synchronized and on-rq */
2570 se->avg.decay_count = 0;
2574 __synchronize_entity_decay(se);
2577 /* migrated tasks did not contribute to our blocked load */
2579 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2580 update_entity_load_avg(se, 0);
2583 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2584 /* we force update consideration on load-balancer moves */
2585 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2589 * Remove se's load from this cfs_rq child load-average, if the entity is
2590 * transitioning to a blocked state we track its projected decay using
2593 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2594 struct sched_entity *se,
2597 update_entity_load_avg(se, 1);
2598 /* we force update consideration on load-balancer moves */
2599 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2601 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2603 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2604 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2605 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2609 * Update the rq's load with the elapsed running time before entering
2610 * idle. if the last scheduled task is not a CFS task, idle_enter will
2611 * be the only way to update the runnable statistic.
2613 void idle_enter_fair(struct rq *this_rq)
2615 update_rq_runnable_avg(this_rq, 1);
2619 * Update the rq's load with the elapsed idle time before a task is
2620 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2621 * be the only way to update the runnable statistic.
2623 void idle_exit_fair(struct rq *this_rq)
2625 update_rq_runnable_avg(this_rq, 0);
2628 static int idle_balance(struct rq *this_rq);
2630 #else /* CONFIG_SMP */
2632 static inline void update_entity_load_avg(struct sched_entity *se,
2633 int update_cfs_rq) {}
2634 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2635 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2636 struct sched_entity *se,
2638 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2639 struct sched_entity *se,
2641 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2642 int force_update) {}
2644 static inline int idle_balance(struct rq *rq)
2649 #endif /* CONFIG_SMP */
2651 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2653 #ifdef CONFIG_SCHEDSTATS
2654 struct task_struct *tsk = NULL;
2656 if (entity_is_task(se))
2659 if (se->statistics.sleep_start) {
2660 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2665 if (unlikely(delta > se->statistics.sleep_max))
2666 se->statistics.sleep_max = delta;
2668 se->statistics.sleep_start = 0;
2669 se->statistics.sum_sleep_runtime += delta;
2672 account_scheduler_latency(tsk, delta >> 10, 1);
2673 trace_sched_stat_sleep(tsk, delta);
2676 if (se->statistics.block_start) {
2677 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2682 if (unlikely(delta > se->statistics.block_max))
2683 se->statistics.block_max = delta;
2685 se->statistics.block_start = 0;
2686 se->statistics.sum_sleep_runtime += delta;
2689 if (tsk->in_iowait) {
2690 se->statistics.iowait_sum += delta;
2691 se->statistics.iowait_count++;
2692 trace_sched_stat_iowait(tsk, delta);
2695 trace_sched_stat_blocked(tsk, delta);
2698 * Blocking time is in units of nanosecs, so shift by
2699 * 20 to get a milliseconds-range estimation of the
2700 * amount of time that the task spent sleeping:
2702 if (unlikely(prof_on == SLEEP_PROFILING)) {
2703 profile_hits(SLEEP_PROFILING,
2704 (void *)get_wchan(tsk),
2707 account_scheduler_latency(tsk, delta >> 10, 0);
2713 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2715 #ifdef CONFIG_SCHED_DEBUG
2716 s64 d = se->vruntime - cfs_rq->min_vruntime;
2721 if (d > 3*sysctl_sched_latency)
2722 schedstat_inc(cfs_rq, nr_spread_over);
2727 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2729 u64 vruntime = cfs_rq->min_vruntime;
2732 * The 'current' period is already promised to the current tasks,
2733 * however the extra weight of the new task will slow them down a
2734 * little, place the new task so that it fits in the slot that
2735 * stays open at the end.
2737 if (initial && sched_feat(START_DEBIT))
2738 vruntime += sched_vslice(cfs_rq, se);
2740 /* sleeps up to a single latency don't count. */
2742 unsigned long thresh = sysctl_sched_latency;
2745 * Halve their sleep time's effect, to allow
2746 * for a gentler effect of sleepers:
2748 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2754 /* ensure we never gain time by being placed backwards. */
2755 se->vruntime = max_vruntime(se->vruntime, vruntime);
2758 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2761 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2764 * Update the normalized vruntime before updating min_vruntime
2765 * through calling update_curr().
2767 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2768 se->vruntime += cfs_rq->min_vruntime;
2771 * Update run-time statistics of the 'current'.
2773 update_curr(cfs_rq);
2774 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2775 account_entity_enqueue(cfs_rq, se);
2776 update_cfs_shares(cfs_rq);
2778 if (flags & ENQUEUE_WAKEUP) {
2779 place_entity(cfs_rq, se, 0);
2780 enqueue_sleeper(cfs_rq, se);
2783 update_stats_enqueue(cfs_rq, se);
2784 check_spread(cfs_rq, se);
2785 if (se != cfs_rq->curr)
2786 __enqueue_entity(cfs_rq, se);
2789 if (cfs_rq->nr_running == 1) {
2790 list_add_leaf_cfs_rq(cfs_rq);
2791 check_enqueue_throttle(cfs_rq);
2795 static void __clear_buddies_last(struct sched_entity *se)
2797 for_each_sched_entity(se) {
2798 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2799 if (cfs_rq->last != se)
2802 cfs_rq->last = NULL;
2806 static void __clear_buddies_next(struct sched_entity *se)
2808 for_each_sched_entity(se) {
2809 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2810 if (cfs_rq->next != se)
2813 cfs_rq->next = NULL;
2817 static void __clear_buddies_skip(struct sched_entity *se)
2819 for_each_sched_entity(se) {
2820 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2821 if (cfs_rq->skip != se)
2824 cfs_rq->skip = NULL;
2828 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2830 if (cfs_rq->last == se)
2831 __clear_buddies_last(se);
2833 if (cfs_rq->next == se)
2834 __clear_buddies_next(se);
2836 if (cfs_rq->skip == se)
2837 __clear_buddies_skip(se);
2840 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2843 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2846 * Update run-time statistics of the 'current'.
2848 update_curr(cfs_rq);
2849 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2851 update_stats_dequeue(cfs_rq, se);
2852 if (flags & DEQUEUE_SLEEP) {
2853 #ifdef CONFIG_SCHEDSTATS
2854 if (entity_is_task(se)) {
2855 struct task_struct *tsk = task_of(se);
2857 if (tsk->state & TASK_INTERRUPTIBLE)
2858 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2859 if (tsk->state & TASK_UNINTERRUPTIBLE)
2860 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2865 clear_buddies(cfs_rq, se);
2867 if (se != cfs_rq->curr)
2868 __dequeue_entity(cfs_rq, se);
2870 account_entity_dequeue(cfs_rq, se);
2873 * Normalize the entity after updating the min_vruntime because the
2874 * update can refer to the ->curr item and we need to reflect this
2875 * movement in our normalized position.
2877 if (!(flags & DEQUEUE_SLEEP))
2878 se->vruntime -= cfs_rq->min_vruntime;
2880 /* return excess runtime on last dequeue */
2881 return_cfs_rq_runtime(cfs_rq);
2883 update_min_vruntime(cfs_rq);
2884 update_cfs_shares(cfs_rq);
2888 * Preempt the current task with a newly woken task if needed:
2891 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2893 unsigned long ideal_runtime, delta_exec;
2894 struct sched_entity *se;
2897 ideal_runtime = sched_slice(cfs_rq, curr);
2898 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2899 if (delta_exec > ideal_runtime) {
2900 resched_task(rq_of(cfs_rq)->curr);
2902 * The current task ran long enough, ensure it doesn't get
2903 * re-elected due to buddy favours.
2905 clear_buddies(cfs_rq, curr);
2910 * Ensure that a task that missed wakeup preemption by a
2911 * narrow margin doesn't have to wait for a full slice.
2912 * This also mitigates buddy induced latencies under load.
2914 if (delta_exec < sysctl_sched_min_granularity)
2917 se = __pick_first_entity(cfs_rq);
2918 delta = curr->vruntime - se->vruntime;
2923 if (delta > ideal_runtime)
2924 resched_task(rq_of(cfs_rq)->curr);
2928 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2930 /* 'current' is not kept within the tree. */
2933 * Any task has to be enqueued before it get to execute on
2934 * a CPU. So account for the time it spent waiting on the
2937 update_stats_wait_end(cfs_rq, se);
2938 __dequeue_entity(cfs_rq, se);
2941 update_stats_curr_start(cfs_rq, se);
2943 #ifdef CONFIG_SCHEDSTATS
2945 * Track our maximum slice length, if the CPU's load is at
2946 * least twice that of our own weight (i.e. dont track it
2947 * when there are only lesser-weight tasks around):
2949 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2950 se->statistics.slice_max = max(se->statistics.slice_max,
2951 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2954 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2958 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2961 * Pick the next process, keeping these things in mind, in this order:
2962 * 1) keep things fair between processes/task groups
2963 * 2) pick the "next" process, since someone really wants that to run
2964 * 3) pick the "last" process, for cache locality
2965 * 4) do not run the "skip" process, if something else is available
2967 static struct sched_entity *
2968 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2970 struct sched_entity *left = __pick_first_entity(cfs_rq);
2971 struct sched_entity *se;
2974 * If curr is set we have to see if its left of the leftmost entity
2975 * still in the tree, provided there was anything in the tree at all.
2977 if (!left || (curr && entity_before(curr, left)))
2980 se = left; /* ideally we run the leftmost entity */
2983 * Avoid running the skip buddy, if running something else can
2984 * be done without getting too unfair.
2986 if (cfs_rq->skip == se) {
2987 struct sched_entity *second;
2990 second = __pick_first_entity(cfs_rq);
2992 second = __pick_next_entity(se);
2993 if (!second || (curr && entity_before(curr, second)))
2997 if (second && wakeup_preempt_entity(second, left) < 1)
3002 * Prefer last buddy, try to return the CPU to a preempted task.
3004 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3008 * Someone really wants this to run. If it's not unfair, run it.
3010 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3013 clear_buddies(cfs_rq, se);
3018 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3020 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3023 * If still on the runqueue then deactivate_task()
3024 * was not called and update_curr() has to be done:
3027 update_curr(cfs_rq);
3029 /* throttle cfs_rqs exceeding runtime */
3030 check_cfs_rq_runtime(cfs_rq);
3032 check_spread(cfs_rq, prev);
3034 update_stats_wait_start(cfs_rq, prev);
3035 /* Put 'current' back into the tree. */
3036 __enqueue_entity(cfs_rq, prev);
3037 /* in !on_rq case, update occurred at dequeue */
3038 update_entity_load_avg(prev, 1);
3040 cfs_rq->curr = NULL;
3044 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3047 * Update run-time statistics of the 'current'.
3049 update_curr(cfs_rq);
3052 * Ensure that runnable average is periodically updated.
3054 update_entity_load_avg(curr, 1);
3055 update_cfs_rq_blocked_load(cfs_rq, 1);
3056 update_cfs_shares(cfs_rq);
3058 #ifdef CONFIG_SCHED_HRTICK
3060 * queued ticks are scheduled to match the slice, so don't bother
3061 * validating it and just reschedule.
3064 resched_task(rq_of(cfs_rq)->curr);
3068 * don't let the period tick interfere with the hrtick preemption
3070 if (!sched_feat(DOUBLE_TICK) &&
3071 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3075 if (cfs_rq->nr_running > 1)
3076 check_preempt_tick(cfs_rq, curr);
3080 /**************************************************
3081 * CFS bandwidth control machinery
3084 #ifdef CONFIG_CFS_BANDWIDTH
3086 #ifdef HAVE_JUMP_LABEL
3087 static struct static_key __cfs_bandwidth_used;
3089 static inline bool cfs_bandwidth_used(void)
3091 return static_key_false(&__cfs_bandwidth_used);
3094 void cfs_bandwidth_usage_inc(void)
3096 static_key_slow_inc(&__cfs_bandwidth_used);
3099 void cfs_bandwidth_usage_dec(void)
3101 static_key_slow_dec(&__cfs_bandwidth_used);
3103 #else /* HAVE_JUMP_LABEL */
3104 static bool cfs_bandwidth_used(void)
3109 void cfs_bandwidth_usage_inc(void) {}
3110 void cfs_bandwidth_usage_dec(void) {}
3111 #endif /* HAVE_JUMP_LABEL */
3114 * default period for cfs group bandwidth.
3115 * default: 0.1s, units: nanoseconds
3117 static inline u64 default_cfs_period(void)
3119 return 100000000ULL;
3122 static inline u64 sched_cfs_bandwidth_slice(void)
3124 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3128 * Replenish runtime according to assigned quota and update expiration time.
3129 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3130 * additional synchronization around rq->lock.
3132 * requires cfs_b->lock
3134 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3138 if (cfs_b->quota == RUNTIME_INF)
3141 now = sched_clock_cpu(smp_processor_id());
3142 cfs_b->runtime = cfs_b->quota;
3143 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3146 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3148 return &tg->cfs_bandwidth;
3151 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3152 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3154 if (unlikely(cfs_rq->throttle_count))
3155 return cfs_rq->throttled_clock_task;
3157 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3160 /* returns 0 on failure to allocate runtime */
3161 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3163 struct task_group *tg = cfs_rq->tg;
3164 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3165 u64 amount = 0, min_amount, expires;
3167 /* note: this is a positive sum as runtime_remaining <= 0 */
3168 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3170 raw_spin_lock(&cfs_b->lock);
3171 if (cfs_b->quota == RUNTIME_INF)
3172 amount = min_amount;
3175 * If the bandwidth pool has become inactive, then at least one
3176 * period must have elapsed since the last consumption.
3177 * Refresh the global state and ensure bandwidth timer becomes
3180 if (!cfs_b->timer_active) {
3181 __refill_cfs_bandwidth_runtime(cfs_b);
3182 __start_cfs_bandwidth(cfs_b);
3185 if (cfs_b->runtime > 0) {
3186 amount = min(cfs_b->runtime, min_amount);
3187 cfs_b->runtime -= amount;
3191 expires = cfs_b->runtime_expires;
3192 raw_spin_unlock(&cfs_b->lock);
3194 cfs_rq->runtime_remaining += amount;
3196 * we may have advanced our local expiration to account for allowed
3197 * spread between our sched_clock and the one on which runtime was
3200 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3201 cfs_rq->runtime_expires = expires;
3203 return cfs_rq->runtime_remaining > 0;
3207 * Note: This depends on the synchronization provided by sched_clock and the
3208 * fact that rq->clock snapshots this value.
3210 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3212 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3214 /* if the deadline is ahead of our clock, nothing to do */
3215 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3218 if (cfs_rq->runtime_remaining < 0)
3222 * If the local deadline has passed we have to consider the
3223 * possibility that our sched_clock is 'fast' and the global deadline
3224 * has not truly expired.
3226 * Fortunately we can check determine whether this the case by checking
3227 * whether the global deadline has advanced. It is valid to compare
3228 * cfs_b->runtime_expires without any locks since we only care about
3229 * exact equality, so a partial write will still work.
3232 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3233 /* extend local deadline, drift is bounded above by 2 ticks */
3234 cfs_rq->runtime_expires += TICK_NSEC;
3236 /* global deadline is ahead, expiration has passed */
3237 cfs_rq->runtime_remaining = 0;
3241 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3243 /* dock delta_exec before expiring quota (as it could span periods) */
3244 cfs_rq->runtime_remaining -= delta_exec;
3245 expire_cfs_rq_runtime(cfs_rq);
3247 if (likely(cfs_rq->runtime_remaining > 0))
3251 * if we're unable to extend our runtime we resched so that the active
3252 * hierarchy can be throttled
3254 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3255 resched_task(rq_of(cfs_rq)->curr);
3258 static __always_inline
3259 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3261 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3264 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3267 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3269 return cfs_bandwidth_used() && cfs_rq->throttled;
3272 /* check whether cfs_rq, or any parent, is throttled */
3273 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3275 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3279 * Ensure that neither of the group entities corresponding to src_cpu or
3280 * dest_cpu are members of a throttled hierarchy when performing group
3281 * load-balance operations.
3283 static inline int throttled_lb_pair(struct task_group *tg,
3284 int src_cpu, int dest_cpu)
3286 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3288 src_cfs_rq = tg->cfs_rq[src_cpu];
3289 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3291 return throttled_hierarchy(src_cfs_rq) ||
3292 throttled_hierarchy(dest_cfs_rq);
3295 /* updated child weight may affect parent so we have to do this bottom up */
3296 static int tg_unthrottle_up(struct task_group *tg, void *data)
3298 struct rq *rq = data;
3299 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3301 cfs_rq->throttle_count--;
3303 if (!cfs_rq->throttle_count) {
3304 /* adjust cfs_rq_clock_task() */
3305 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3306 cfs_rq->throttled_clock_task;
3313 static int tg_throttle_down(struct task_group *tg, void *data)
3315 struct rq *rq = data;
3316 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3318 /* group is entering throttled state, stop time */
3319 if (!cfs_rq->throttle_count)
3320 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3321 cfs_rq->throttle_count++;
3326 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3328 struct rq *rq = rq_of(cfs_rq);
3329 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3330 struct sched_entity *se;
3331 long task_delta, dequeue = 1;
3333 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3335 /* freeze hierarchy runnable averages while throttled */
3337 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3340 task_delta = cfs_rq->h_nr_running;
3341 for_each_sched_entity(se) {
3342 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3343 /* throttled entity or throttle-on-deactivate */
3348 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3349 qcfs_rq->h_nr_running -= task_delta;
3351 if (qcfs_rq->load.weight)
3356 sub_nr_running(rq, task_delta);
3358 cfs_rq->throttled = 1;
3359 cfs_rq->throttled_clock = rq_clock(rq);
3360 raw_spin_lock(&cfs_b->lock);
3361 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3362 if (!cfs_b->timer_active)
3363 __start_cfs_bandwidth(cfs_b);
3364 raw_spin_unlock(&cfs_b->lock);
3367 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3369 struct rq *rq = rq_of(cfs_rq);
3370 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3371 struct sched_entity *se;
3375 se = cfs_rq->tg->se[cpu_of(rq)];
3377 cfs_rq->throttled = 0;
3379 update_rq_clock(rq);
3381 raw_spin_lock(&cfs_b->lock);
3382 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3383 list_del_rcu(&cfs_rq->throttled_list);
3384 raw_spin_unlock(&cfs_b->lock);
3386 /* update hierarchical throttle state */
3387 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3389 if (!cfs_rq->load.weight)
3392 task_delta = cfs_rq->h_nr_running;
3393 for_each_sched_entity(se) {
3397 cfs_rq = cfs_rq_of(se);
3399 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3400 cfs_rq->h_nr_running += task_delta;
3402 if (cfs_rq_throttled(cfs_rq))
3407 add_nr_running(rq, task_delta);
3409 /* determine whether we need to wake up potentially idle cpu */
3410 if (rq->curr == rq->idle && rq->cfs.nr_running)
3411 resched_task(rq->curr);
3414 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3415 u64 remaining, u64 expires)
3417 struct cfs_rq *cfs_rq;
3418 u64 runtime = remaining;
3421 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3423 struct rq *rq = rq_of(cfs_rq);
3425 raw_spin_lock(&rq->lock);
3426 if (!cfs_rq_throttled(cfs_rq))
3429 runtime = -cfs_rq->runtime_remaining + 1;
3430 if (runtime > remaining)
3431 runtime = remaining;
3432 remaining -= runtime;
3434 cfs_rq->runtime_remaining += runtime;
3435 cfs_rq->runtime_expires = expires;
3437 /* we check whether we're throttled above */
3438 if (cfs_rq->runtime_remaining > 0)
3439 unthrottle_cfs_rq(cfs_rq);
3442 raw_spin_unlock(&rq->lock);
3453 * Responsible for refilling a task_group's bandwidth and unthrottling its
3454 * cfs_rqs as appropriate. If there has been no activity within the last
3455 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3456 * used to track this state.
3458 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3460 u64 runtime, runtime_expires;
3463 /* no need to continue the timer with no bandwidth constraint */
3464 if (cfs_b->quota == RUNTIME_INF)
3465 goto out_deactivate;
3467 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3468 cfs_b->nr_periods += overrun;
3471 * idle depends on !throttled (for the case of a large deficit), and if
3472 * we're going inactive then everything else can be deferred
3474 if (cfs_b->idle && !throttled)
3475 goto out_deactivate;
3478 * if we have relooped after returning idle once, we need to update our
3479 * status as actually running, so that other cpus doing
3480 * __start_cfs_bandwidth will stop trying to cancel us.
3482 cfs_b->timer_active = 1;
3484 __refill_cfs_bandwidth_runtime(cfs_b);
3487 /* mark as potentially idle for the upcoming period */
3492 /* account preceding periods in which throttling occurred */
3493 cfs_b->nr_throttled += overrun;
3496 * There are throttled entities so we must first use the new bandwidth
3497 * to unthrottle them before making it generally available. This
3498 * ensures that all existing debts will be paid before a new cfs_rq is
3501 runtime = cfs_b->runtime;
3502 runtime_expires = cfs_b->runtime_expires;
3506 * This check is repeated as we are holding onto the new bandwidth
3507 * while we unthrottle. This can potentially race with an unthrottled
3508 * group trying to acquire new bandwidth from the global pool.
3510 while (throttled && runtime > 0) {
3511 raw_spin_unlock(&cfs_b->lock);
3512 /* we can't nest cfs_b->lock while distributing bandwidth */
3513 runtime = distribute_cfs_runtime(cfs_b, runtime,
3515 raw_spin_lock(&cfs_b->lock);
3517 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3520 /* return (any) remaining runtime */
3521 cfs_b->runtime = runtime;
3523 * While we are ensured activity in the period following an
3524 * unthrottle, this also covers the case in which the new bandwidth is
3525 * insufficient to cover the existing bandwidth deficit. (Forcing the
3526 * timer to remain active while there are any throttled entities.)
3533 cfs_b->timer_active = 0;
3537 /* a cfs_rq won't donate quota below this amount */
3538 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3539 /* minimum remaining period time to redistribute slack quota */
3540 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3541 /* how long we wait to gather additional slack before distributing */
3542 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3545 * Are we near the end of the current quota period?
3547 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3548 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3549 * migrate_hrtimers, base is never cleared, so we are fine.
3551 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3553 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3556 /* if the call-back is running a quota refresh is already occurring */
3557 if (hrtimer_callback_running(refresh_timer))
3560 /* is a quota refresh about to occur? */
3561 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3562 if (remaining < min_expire)
3568 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3570 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3572 /* if there's a quota refresh soon don't bother with slack */
3573 if (runtime_refresh_within(cfs_b, min_left))
3576 start_bandwidth_timer(&cfs_b->slack_timer,
3577 ns_to_ktime(cfs_bandwidth_slack_period));
3580 /* we know any runtime found here is valid as update_curr() precedes return */
3581 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3583 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3584 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3586 if (slack_runtime <= 0)
3589 raw_spin_lock(&cfs_b->lock);
3590 if (cfs_b->quota != RUNTIME_INF &&
3591 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3592 cfs_b->runtime += slack_runtime;
3594 /* we are under rq->lock, defer unthrottling using a timer */
3595 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3596 !list_empty(&cfs_b->throttled_cfs_rq))
3597 start_cfs_slack_bandwidth(cfs_b);
3599 raw_spin_unlock(&cfs_b->lock);
3601 /* even if it's not valid for return we don't want to try again */
3602 cfs_rq->runtime_remaining -= slack_runtime;
3605 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3607 if (!cfs_bandwidth_used())
3610 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3613 __return_cfs_rq_runtime(cfs_rq);
3617 * This is done with a timer (instead of inline with bandwidth return) since
3618 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3620 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3622 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3625 /* confirm we're still not at a refresh boundary */
3626 raw_spin_lock(&cfs_b->lock);
3627 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3628 raw_spin_unlock(&cfs_b->lock);
3632 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3633 runtime = cfs_b->runtime;
3636 expires = cfs_b->runtime_expires;
3637 raw_spin_unlock(&cfs_b->lock);
3642 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3644 raw_spin_lock(&cfs_b->lock);
3645 if (expires == cfs_b->runtime_expires)
3646 cfs_b->runtime = runtime;
3647 raw_spin_unlock(&cfs_b->lock);
3651 * When a group wakes up we want to make sure that its quota is not already
3652 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3653 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3655 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3657 if (!cfs_bandwidth_used())
3660 /* an active group must be handled by the update_curr()->put() path */
3661 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3664 /* ensure the group is not already throttled */
3665 if (cfs_rq_throttled(cfs_rq))
3668 /* update runtime allocation */
3669 account_cfs_rq_runtime(cfs_rq, 0);
3670 if (cfs_rq->runtime_remaining <= 0)
3671 throttle_cfs_rq(cfs_rq);
3674 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3675 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3677 if (!cfs_bandwidth_used())
3680 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3684 * it's possible for a throttled entity to be forced into a running
3685 * state (e.g. set_curr_task), in this case we're finished.
3687 if (cfs_rq_throttled(cfs_rq))
3690 throttle_cfs_rq(cfs_rq);
3694 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3696 struct cfs_bandwidth *cfs_b =
3697 container_of(timer, struct cfs_bandwidth, slack_timer);
3698 do_sched_cfs_slack_timer(cfs_b);
3700 return HRTIMER_NORESTART;
3703 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3705 struct cfs_bandwidth *cfs_b =
3706 container_of(timer, struct cfs_bandwidth, period_timer);
3711 raw_spin_lock(&cfs_b->lock);
3713 now = hrtimer_cb_get_time(timer);
3714 overrun = hrtimer_forward(timer, now, cfs_b->period);
3719 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3721 raw_spin_unlock(&cfs_b->lock);
3723 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3726 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3728 raw_spin_lock_init(&cfs_b->lock);
3730 cfs_b->quota = RUNTIME_INF;
3731 cfs_b->period = ns_to_ktime(default_cfs_period());
3733 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3734 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3735 cfs_b->period_timer.function = sched_cfs_period_timer;
3736 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3737 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3740 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3742 cfs_rq->runtime_enabled = 0;
3743 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3746 /* requires cfs_b->lock, may release to reprogram timer */
3747 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3750 * The timer may be active because we're trying to set a new bandwidth
3751 * period or because we're racing with the tear-down path
3752 * (timer_active==0 becomes visible before the hrtimer call-back
3753 * terminates). In either case we ensure that it's re-programmed
3755 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3756 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3757 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3758 raw_spin_unlock(&cfs_b->lock);
3760 raw_spin_lock(&cfs_b->lock);
3761 /* if someone else restarted the timer then we're done */
3762 if (cfs_b->timer_active)
3766 cfs_b->timer_active = 1;
3767 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3770 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3772 hrtimer_cancel(&cfs_b->period_timer);
3773 hrtimer_cancel(&cfs_b->slack_timer);
3776 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3778 struct cfs_rq *cfs_rq;
3780 for_each_leaf_cfs_rq(rq, cfs_rq) {
3781 if (!cfs_rq->runtime_enabled)
3785 * clock_task is not advancing so we just need to make sure
3786 * there's some valid quota amount
3788 cfs_rq->runtime_remaining = 1;
3789 if (cfs_rq_throttled(cfs_rq))
3790 unthrottle_cfs_rq(cfs_rq);
3794 #else /* CONFIG_CFS_BANDWIDTH */
3795 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3797 return rq_clock_task(rq_of(cfs_rq));
3800 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3801 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3802 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3803 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3805 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3810 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3815 static inline int throttled_lb_pair(struct task_group *tg,
3816 int src_cpu, int dest_cpu)
3821 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3823 #ifdef CONFIG_FAIR_GROUP_SCHED
3824 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3827 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3831 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3832 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3834 #endif /* CONFIG_CFS_BANDWIDTH */
3836 /**************************************************
3837 * CFS operations on tasks:
3840 #ifdef CONFIG_SCHED_HRTICK
3841 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3843 struct sched_entity *se = &p->se;
3844 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3846 WARN_ON(task_rq(p) != rq);
3848 if (cfs_rq->nr_running > 1) {
3849 u64 slice = sched_slice(cfs_rq, se);
3850 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3851 s64 delta = slice - ran;
3860 * Don't schedule slices shorter than 10000ns, that just
3861 * doesn't make sense. Rely on vruntime for fairness.
3864 delta = max_t(s64, 10000LL, delta);
3866 hrtick_start(rq, delta);
3871 * called from enqueue/dequeue and updates the hrtick when the
3872 * current task is from our class and nr_running is low enough
3875 static void hrtick_update(struct rq *rq)
3877 struct task_struct *curr = rq->curr;
3879 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3882 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3883 hrtick_start_fair(rq, curr);
3885 #else /* !CONFIG_SCHED_HRTICK */
3887 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3891 static inline void hrtick_update(struct rq *rq)
3897 * The enqueue_task method is called before nr_running is
3898 * increased. Here we update the fair scheduling stats and
3899 * then put the task into the rbtree:
3902 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3904 struct cfs_rq *cfs_rq;
3905 struct sched_entity *se = &p->se;
3907 for_each_sched_entity(se) {
3910 cfs_rq = cfs_rq_of(se);
3911 enqueue_entity(cfs_rq, se, flags);
3914 * end evaluation on encountering a throttled cfs_rq
3916 * note: in the case of encountering a throttled cfs_rq we will
3917 * post the final h_nr_running increment below.
3919 if (cfs_rq_throttled(cfs_rq))
3921 cfs_rq->h_nr_running++;
3923 flags = ENQUEUE_WAKEUP;
3926 for_each_sched_entity(se) {
3927 cfs_rq = cfs_rq_of(se);
3928 cfs_rq->h_nr_running++;
3930 if (cfs_rq_throttled(cfs_rq))
3933 update_cfs_shares(cfs_rq);
3934 update_entity_load_avg(se, 1);
3938 update_rq_runnable_avg(rq, rq->nr_running);
3939 add_nr_running(rq, 1);
3944 static void set_next_buddy(struct sched_entity *se);
3947 * The dequeue_task method is called before nr_running is
3948 * decreased. We remove the task from the rbtree and
3949 * update the fair scheduling stats:
3951 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3953 struct cfs_rq *cfs_rq;
3954 struct sched_entity *se = &p->se;
3955 int task_sleep = flags & DEQUEUE_SLEEP;
3957 for_each_sched_entity(se) {
3958 cfs_rq = cfs_rq_of(se);
3959 dequeue_entity(cfs_rq, se, flags);
3962 * end evaluation on encountering a throttled cfs_rq
3964 * note: in the case of encountering a throttled cfs_rq we will
3965 * post the final h_nr_running decrement below.
3967 if (cfs_rq_throttled(cfs_rq))
3969 cfs_rq->h_nr_running--;
3971 /* Don't dequeue parent if it has other entities besides us */
3972 if (cfs_rq->load.weight) {
3974 * Bias pick_next to pick a task from this cfs_rq, as
3975 * p is sleeping when it is within its sched_slice.
3977 if (task_sleep && parent_entity(se))
3978 set_next_buddy(parent_entity(se));
3980 /* avoid re-evaluating load for this entity */
3981 se = parent_entity(se);
3984 flags |= DEQUEUE_SLEEP;
3987 for_each_sched_entity(se) {
3988 cfs_rq = cfs_rq_of(se);
3989 cfs_rq->h_nr_running--;
3991 if (cfs_rq_throttled(cfs_rq))
3994 update_cfs_shares(cfs_rq);
3995 update_entity_load_avg(se, 1);
3999 sub_nr_running(rq, 1);
4000 update_rq_runnable_avg(rq, 1);
4006 /* Used instead of source_load when we know the type == 0 */
4007 static unsigned long weighted_cpuload(const int cpu)
4009 return cpu_rq(cpu)->cfs.runnable_load_avg;
4013 * Return a low guess at the load of a migration-source cpu weighted
4014 * according to the scheduling class and "nice" value.
4016 * We want to under-estimate the load of migration sources, to
4017 * balance conservatively.
4019 static unsigned long source_load(int cpu, int type)
4021 struct rq *rq = cpu_rq(cpu);
4022 unsigned long total = weighted_cpuload(cpu);
4024 if (type == 0 || !sched_feat(LB_BIAS))
4027 return min(rq->cpu_load[type-1], total);
4031 * Return a high guess at the load of a migration-target cpu weighted
4032 * according to the scheduling class and "nice" value.
4034 static unsigned long target_load(int cpu, int type)
4036 struct rq *rq = cpu_rq(cpu);
4037 unsigned long total = weighted_cpuload(cpu);
4039 if (type == 0 || !sched_feat(LB_BIAS))
4042 return max(rq->cpu_load[type-1], total);
4045 static unsigned long power_of(int cpu)
4047 return cpu_rq(cpu)->cpu_power;
4050 static unsigned long cpu_avg_load_per_task(int cpu)
4052 struct rq *rq = cpu_rq(cpu);
4053 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
4054 unsigned long load_avg = rq->cfs.runnable_load_avg;
4057 return load_avg / nr_running;
4062 static void record_wakee(struct task_struct *p)
4065 * Rough decay (wiping) for cost saving, don't worry
4066 * about the boundary, really active task won't care
4069 if (jiffies > current->wakee_flip_decay_ts + HZ) {
4070 current->wakee_flips >>= 1;
4071 current->wakee_flip_decay_ts = jiffies;
4074 if (current->last_wakee != p) {
4075 current->last_wakee = p;
4076 current->wakee_flips++;
4080 static void task_waking_fair(struct task_struct *p)
4082 struct sched_entity *se = &p->se;
4083 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4086 #ifndef CONFIG_64BIT
4087 u64 min_vruntime_copy;
4090 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4092 min_vruntime = cfs_rq->min_vruntime;
4093 } while (min_vruntime != min_vruntime_copy);
4095 min_vruntime = cfs_rq->min_vruntime;
4098 se->vruntime -= min_vruntime;
4102 #ifdef CONFIG_FAIR_GROUP_SCHED
4104 * effective_load() calculates the load change as seen from the root_task_group
4106 * Adding load to a group doesn't make a group heavier, but can cause movement
4107 * of group shares between cpus. Assuming the shares were perfectly aligned one
4108 * can calculate the shift in shares.
4110 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4111 * on this @cpu and results in a total addition (subtraction) of @wg to the
4112 * total group weight.
4114 * Given a runqueue weight distribution (rw_i) we can compute a shares
4115 * distribution (s_i) using:
4117 * s_i = rw_i / \Sum rw_j (1)
4119 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4120 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4121 * shares distribution (s_i):
4123 * rw_i = { 2, 4, 1, 0 }
4124 * s_i = { 2/7, 4/7, 1/7, 0 }
4126 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4127 * task used to run on and the CPU the waker is running on), we need to
4128 * compute the effect of waking a task on either CPU and, in case of a sync
4129 * wakeup, compute the effect of the current task going to sleep.
4131 * So for a change of @wl to the local @cpu with an overall group weight change
4132 * of @wl we can compute the new shares distribution (s'_i) using:
4134 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4136 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4137 * differences in waking a task to CPU 0. The additional task changes the
4138 * weight and shares distributions like:
4140 * rw'_i = { 3, 4, 1, 0 }
4141 * s'_i = { 3/8, 4/8, 1/8, 0 }
4143 * We can then compute the difference in effective weight by using:
4145 * dw_i = S * (s'_i - s_i) (3)
4147 * Where 'S' is the group weight as seen by its parent.
4149 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4150 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4151 * 4/7) times the weight of the group.
4153 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4155 struct sched_entity *se = tg->se[cpu];
4157 if (!tg->parent) /* the trivial, non-cgroup case */
4160 for_each_sched_entity(se) {
4166 * W = @wg + \Sum rw_j
4168 W = wg + calc_tg_weight(tg, se->my_q);
4173 w = se->my_q->load.weight + wl;
4176 * wl = S * s'_i; see (2)
4179 wl = (w * tg->shares) / W;
4184 * Per the above, wl is the new se->load.weight value; since
4185 * those are clipped to [MIN_SHARES, ...) do so now. See
4186 * calc_cfs_shares().
4188 if (wl < MIN_SHARES)
4192 * wl = dw_i = S * (s'_i - s_i); see (3)
4194 wl -= se->load.weight;
4197 * Recursively apply this logic to all parent groups to compute
4198 * the final effective load change on the root group. Since
4199 * only the @tg group gets extra weight, all parent groups can
4200 * only redistribute existing shares. @wl is the shift in shares
4201 * resulting from this level per the above.
4210 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4217 static int wake_wide(struct task_struct *p)
4219 int factor = this_cpu_read(sd_llc_size);
4222 * Yeah, it's the switching-frequency, could means many wakee or
4223 * rapidly switch, use factor here will just help to automatically
4224 * adjust the loose-degree, so bigger node will lead to more pull.
4226 if (p->wakee_flips > factor) {
4228 * wakee is somewhat hot, it needs certain amount of cpu
4229 * resource, so if waker is far more hot, prefer to leave
4232 if (current->wakee_flips > (factor * p->wakee_flips))
4239 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4241 s64 this_load, load;
4242 int idx, this_cpu, prev_cpu;
4243 unsigned long tl_per_task;
4244 struct task_group *tg;
4245 unsigned long weight;
4249 * If we wake multiple tasks be careful to not bounce
4250 * ourselves around too much.
4256 this_cpu = smp_processor_id();
4257 prev_cpu = task_cpu(p);
4258 load = source_load(prev_cpu, idx);
4259 this_load = target_load(this_cpu, idx);
4262 * If sync wakeup then subtract the (maximum possible)
4263 * effect of the currently running task from the load
4264 * of the current CPU:
4267 tg = task_group(current);
4268 weight = current->se.load.weight;
4270 this_load += effective_load(tg, this_cpu, -weight, -weight);
4271 load += effective_load(tg, prev_cpu, 0, -weight);
4275 weight = p->se.load.weight;
4278 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4279 * due to the sync cause above having dropped this_load to 0, we'll
4280 * always have an imbalance, but there's really nothing you can do
4281 * about that, so that's good too.
4283 * Otherwise check if either cpus are near enough in load to allow this
4284 * task to be woken on this_cpu.
4286 if (this_load > 0) {
4287 s64 this_eff_load, prev_eff_load;
4289 this_eff_load = 100;
4290 this_eff_load *= power_of(prev_cpu);
4291 this_eff_load *= this_load +
4292 effective_load(tg, this_cpu, weight, weight);
4294 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4295 prev_eff_load *= power_of(this_cpu);
4296 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4298 balanced = this_eff_load <= prev_eff_load;
4303 * If the currently running task will sleep within
4304 * a reasonable amount of time then attract this newly
4307 if (sync && balanced)
4310 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4311 tl_per_task = cpu_avg_load_per_task(this_cpu);
4314 (this_load <= load &&
4315 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4317 * This domain has SD_WAKE_AFFINE and
4318 * p is cache cold in this domain, and
4319 * there is no bad imbalance.
4321 schedstat_inc(sd, ttwu_move_affine);
4322 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4330 * find_idlest_group finds and returns the least busy CPU group within the
4333 static struct sched_group *
4334 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4335 int this_cpu, int sd_flag)
4337 struct sched_group *idlest = NULL, *group = sd->groups;
4338 unsigned long min_load = ULONG_MAX, this_load = 0;
4339 int load_idx = sd->forkexec_idx;
4340 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4342 if (sd_flag & SD_BALANCE_WAKE)
4343 load_idx = sd->wake_idx;
4346 unsigned long load, avg_load;
4350 /* Skip over this group if it has no CPUs allowed */
4351 if (!cpumask_intersects(sched_group_cpus(group),
4352 tsk_cpus_allowed(p)))
4355 local_group = cpumask_test_cpu(this_cpu,
4356 sched_group_cpus(group));
4358 /* Tally up the load of all CPUs in the group */
4361 for_each_cpu(i, sched_group_cpus(group)) {
4362 /* Bias balancing toward cpus of our domain */
4364 load = source_load(i, load_idx);
4366 load = target_load(i, load_idx);
4371 /* Adjust by relative CPU power of the group */
4372 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4375 this_load = avg_load;
4376 } else if (avg_load < min_load) {
4377 min_load = avg_load;
4380 } while (group = group->next, group != sd->groups);
4382 if (!idlest || 100*this_load < imbalance*min_load)
4388 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4391 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4393 unsigned long load, min_load = ULONG_MAX;
4397 /* Traverse only the allowed CPUs */
4398 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4399 load = weighted_cpuload(i);
4401 if (load < min_load || (load == min_load && i == this_cpu)) {
4411 * Try and locate an idle CPU in the sched_domain.
4413 static int select_idle_sibling(struct task_struct *p, int target)
4415 struct sched_domain *sd;
4416 struct sched_group *sg;
4417 int i = task_cpu(p);
4419 if (idle_cpu(target))
4423 * If the prevous cpu is cache affine and idle, don't be stupid.
4425 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4429 * Otherwise, iterate the domains and find an elegible idle cpu.
4431 sd = rcu_dereference(per_cpu(sd_llc, target));
4432 for_each_lower_domain(sd) {
4435 if (!cpumask_intersects(sched_group_cpus(sg),
4436 tsk_cpus_allowed(p)))
4439 for_each_cpu(i, sched_group_cpus(sg)) {
4440 if (i == target || !idle_cpu(i))
4444 target = cpumask_first_and(sched_group_cpus(sg),
4445 tsk_cpus_allowed(p));
4449 } while (sg != sd->groups);
4456 * select_task_rq_fair: Select target runqueue for the waking task in domains
4457 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4458 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4460 * Balances load by selecting the idlest cpu in the idlest group, or under
4461 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4463 * Returns the target cpu number.
4465 * preempt must be disabled.
4468 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4470 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4471 int cpu = smp_processor_id();
4473 int want_affine = 0;
4474 int sync = wake_flags & WF_SYNC;
4476 if (p->nr_cpus_allowed == 1)
4479 if (sd_flag & SD_BALANCE_WAKE) {
4480 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4486 for_each_domain(cpu, tmp) {
4487 if (!(tmp->flags & SD_LOAD_BALANCE))
4491 * If both cpu and prev_cpu are part of this domain,
4492 * cpu is a valid SD_WAKE_AFFINE target.
4494 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4495 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4500 if (tmp->flags & sd_flag)
4504 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4507 if (sd_flag & SD_BALANCE_WAKE) {
4508 new_cpu = select_idle_sibling(p, prev_cpu);
4513 struct sched_group *group;
4516 if (!(sd->flags & sd_flag)) {
4521 group = find_idlest_group(sd, p, cpu, sd_flag);
4527 new_cpu = find_idlest_cpu(group, p, cpu);
4528 if (new_cpu == -1 || new_cpu == cpu) {
4529 /* Now try balancing at a lower domain level of cpu */
4534 /* Now try balancing at a lower domain level of new_cpu */
4536 weight = sd->span_weight;
4538 for_each_domain(cpu, tmp) {
4539 if (weight <= tmp->span_weight)
4541 if (tmp->flags & sd_flag)
4544 /* while loop will break here if sd == NULL */
4553 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4554 * cfs_rq_of(p) references at time of call are still valid and identify the
4555 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4556 * other assumptions, including the state of rq->lock, should be made.
4559 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4561 struct sched_entity *se = &p->se;
4562 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4565 * Load tracking: accumulate removed load so that it can be processed
4566 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4567 * to blocked load iff they have a positive decay-count. It can never
4568 * be negative here since on-rq tasks have decay-count == 0.
4570 if (se->avg.decay_count) {
4571 se->avg.decay_count = -__synchronize_entity_decay(se);
4572 atomic_long_add(se->avg.load_avg_contrib,
4573 &cfs_rq->removed_load);
4576 /* We have migrated, no longer consider this task hot */
4579 #endif /* CONFIG_SMP */
4581 static unsigned long
4582 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4584 unsigned long gran = sysctl_sched_wakeup_granularity;
4587 * Since its curr running now, convert the gran from real-time
4588 * to virtual-time in his units.
4590 * By using 'se' instead of 'curr' we penalize light tasks, so
4591 * they get preempted easier. That is, if 'se' < 'curr' then
4592 * the resulting gran will be larger, therefore penalizing the
4593 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4594 * be smaller, again penalizing the lighter task.
4596 * This is especially important for buddies when the leftmost
4597 * task is higher priority than the buddy.
4599 return calc_delta_fair(gran, se);
4603 * Should 'se' preempt 'curr'.
4617 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4619 s64 gran, vdiff = curr->vruntime - se->vruntime;
4624 gran = wakeup_gran(curr, se);
4631 static void set_last_buddy(struct sched_entity *se)
4633 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4636 for_each_sched_entity(se)
4637 cfs_rq_of(se)->last = se;
4640 static void set_next_buddy(struct sched_entity *se)
4642 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4645 for_each_sched_entity(se)
4646 cfs_rq_of(se)->next = se;
4649 static void set_skip_buddy(struct sched_entity *se)
4651 for_each_sched_entity(se)
4652 cfs_rq_of(se)->skip = se;
4656 * Preempt the current task with a newly woken task if needed:
4658 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4660 struct task_struct *curr = rq->curr;
4661 struct sched_entity *se = &curr->se, *pse = &p->se;
4662 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4663 int scale = cfs_rq->nr_running >= sched_nr_latency;
4664 int next_buddy_marked = 0;
4666 if (unlikely(se == pse))
4670 * This is possible from callers such as move_task(), in which we
4671 * unconditionally check_prempt_curr() after an enqueue (which may have
4672 * lead to a throttle). This both saves work and prevents false
4673 * next-buddy nomination below.
4675 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4678 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4679 set_next_buddy(pse);
4680 next_buddy_marked = 1;
4684 * We can come here with TIF_NEED_RESCHED already set from new task
4687 * Note: this also catches the edge-case of curr being in a throttled
4688 * group (e.g. via set_curr_task), since update_curr() (in the
4689 * enqueue of curr) will have resulted in resched being set. This
4690 * prevents us from potentially nominating it as a false LAST_BUDDY
4693 if (test_tsk_need_resched(curr))
4696 /* Idle tasks are by definition preempted by non-idle tasks. */
4697 if (unlikely(curr->policy == SCHED_IDLE) &&
4698 likely(p->policy != SCHED_IDLE))
4702 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4703 * is driven by the tick):
4705 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4708 find_matching_se(&se, &pse);
4709 update_curr(cfs_rq_of(se));
4711 if (wakeup_preempt_entity(se, pse) == 1) {
4713 * Bias pick_next to pick the sched entity that is
4714 * triggering this preemption.
4716 if (!next_buddy_marked)
4717 set_next_buddy(pse);
4726 * Only set the backward buddy when the current task is still
4727 * on the rq. This can happen when a wakeup gets interleaved
4728 * with schedule on the ->pre_schedule() or idle_balance()
4729 * point, either of which can * drop the rq lock.
4731 * Also, during early boot the idle thread is in the fair class,
4732 * for obvious reasons its a bad idea to schedule back to it.
4734 if (unlikely(!se->on_rq || curr == rq->idle))
4737 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4741 static struct task_struct *
4742 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4744 struct cfs_rq *cfs_rq = &rq->cfs;
4745 struct sched_entity *se;
4746 struct task_struct *p;
4750 #ifdef CONFIG_FAIR_GROUP_SCHED
4751 if (!cfs_rq->nr_running)
4754 if (prev->sched_class != &fair_sched_class)
4758 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4759 * likely that a next task is from the same cgroup as the current.
4761 * Therefore attempt to avoid putting and setting the entire cgroup
4762 * hierarchy, only change the part that actually changes.
4766 struct sched_entity *curr = cfs_rq->curr;
4769 * Since we got here without doing put_prev_entity() we also
4770 * have to consider cfs_rq->curr. If it is still a runnable
4771 * entity, update_curr() will update its vruntime, otherwise
4772 * forget we've ever seen it.
4774 if (curr && curr->on_rq)
4775 update_curr(cfs_rq);
4780 * This call to check_cfs_rq_runtime() will do the throttle and
4781 * dequeue its entity in the parent(s). Therefore the 'simple'
4782 * nr_running test will indeed be correct.
4784 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4787 se = pick_next_entity(cfs_rq, curr);
4788 cfs_rq = group_cfs_rq(se);
4794 * Since we haven't yet done put_prev_entity and if the selected task
4795 * is a different task than we started out with, try and touch the
4796 * least amount of cfs_rqs.
4799 struct sched_entity *pse = &prev->se;
4801 while (!(cfs_rq = is_same_group(se, pse))) {
4802 int se_depth = se->depth;
4803 int pse_depth = pse->depth;
4805 if (se_depth <= pse_depth) {
4806 put_prev_entity(cfs_rq_of(pse), pse);
4807 pse = parent_entity(pse);
4809 if (se_depth >= pse_depth) {
4810 set_next_entity(cfs_rq_of(se), se);
4811 se = parent_entity(se);
4815 put_prev_entity(cfs_rq, pse);
4816 set_next_entity(cfs_rq, se);
4819 if (hrtick_enabled(rq))
4820 hrtick_start_fair(rq, p);
4827 if (!cfs_rq->nr_running)
4830 put_prev_task(rq, prev);
4833 se = pick_next_entity(cfs_rq, NULL);
4834 set_next_entity(cfs_rq, se);
4835 cfs_rq = group_cfs_rq(se);
4840 if (hrtick_enabled(rq))
4841 hrtick_start_fair(rq, p);
4846 new_tasks = idle_balance(rq);
4848 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4849 * possible for any higher priority task to appear. In that case we
4850 * must re-start the pick_next_entity() loop.
4862 * Account for a descheduled task:
4864 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4866 struct sched_entity *se = &prev->se;
4867 struct cfs_rq *cfs_rq;
4869 for_each_sched_entity(se) {
4870 cfs_rq = cfs_rq_of(se);
4871 put_prev_entity(cfs_rq, se);
4876 * sched_yield() is very simple
4878 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4880 static void yield_task_fair(struct rq *rq)
4882 struct task_struct *curr = rq->curr;
4883 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4884 struct sched_entity *se = &curr->se;
4887 * Are we the only task in the tree?
4889 if (unlikely(rq->nr_running == 1))
4892 clear_buddies(cfs_rq, se);
4894 if (curr->policy != SCHED_BATCH) {
4895 update_rq_clock(rq);
4897 * Update run-time statistics of the 'current'.
4899 update_curr(cfs_rq);
4901 * Tell update_rq_clock() that we've just updated,
4902 * so we don't do microscopic update in schedule()
4903 * and double the fastpath cost.
4905 rq->skip_clock_update = 1;
4911 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4913 struct sched_entity *se = &p->se;
4915 /* throttled hierarchies are not runnable */
4916 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4919 /* Tell the scheduler that we'd really like pse to run next. */
4922 yield_task_fair(rq);
4928 /**************************************************
4929 * Fair scheduling class load-balancing methods.
4933 * The purpose of load-balancing is to achieve the same basic fairness the
4934 * per-cpu scheduler provides, namely provide a proportional amount of compute
4935 * time to each task. This is expressed in the following equation:
4937 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4939 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4940 * W_i,0 is defined as:
4942 * W_i,0 = \Sum_j w_i,j (2)
4944 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4945 * is derived from the nice value as per prio_to_weight[].
4947 * The weight average is an exponential decay average of the instantaneous
4950 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4952 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4953 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4954 * can also include other factors [XXX].
4956 * To achieve this balance we define a measure of imbalance which follows
4957 * directly from (1):
4959 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4961 * We them move tasks around to minimize the imbalance. In the continuous
4962 * function space it is obvious this converges, in the discrete case we get
4963 * a few fun cases generally called infeasible weight scenarios.
4966 * - infeasible weights;
4967 * - local vs global optima in the discrete case. ]
4972 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4973 * for all i,j solution, we create a tree of cpus that follows the hardware
4974 * topology where each level pairs two lower groups (or better). This results
4975 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4976 * tree to only the first of the previous level and we decrease the frequency
4977 * of load-balance at each level inv. proportional to the number of cpus in
4983 * \Sum { --- * --- * 2^i } = O(n) (5)
4985 * `- size of each group
4986 * | | `- number of cpus doing load-balance
4988 * `- sum over all levels
4990 * Coupled with a limit on how many tasks we can migrate every balance pass,
4991 * this makes (5) the runtime complexity of the balancer.
4993 * An important property here is that each CPU is still (indirectly) connected
4994 * to every other cpu in at most O(log n) steps:
4996 * The adjacency matrix of the resulting graph is given by:
4999 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5002 * And you'll find that:
5004 * A^(log_2 n)_i,j != 0 for all i,j (7)
5006 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5007 * The task movement gives a factor of O(m), giving a convergence complexity
5010 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5015 * In order to avoid CPUs going idle while there's still work to do, new idle
5016 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5017 * tree itself instead of relying on other CPUs to bring it work.
5019 * This adds some complexity to both (5) and (8) but it reduces the total idle
5027 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5030 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5035 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5037 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5039 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5042 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5043 * rewrite all of this once again.]
5046 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5048 enum fbq_type { regular, remote, all };
5050 #define LBF_ALL_PINNED 0x01
5051 #define LBF_NEED_BREAK 0x02
5052 #define LBF_DST_PINNED 0x04
5053 #define LBF_SOME_PINNED 0x08
5056 struct sched_domain *sd;
5064 struct cpumask *dst_grpmask;
5066 enum cpu_idle_type idle;
5068 /* The set of CPUs under consideration for load-balancing */
5069 struct cpumask *cpus;
5074 unsigned int loop_break;
5075 unsigned int loop_max;
5077 enum fbq_type fbq_type;
5081 * move_task - move a task from one runqueue to another runqueue.
5082 * Both runqueues must be locked.
5084 static void move_task(struct task_struct *p, struct lb_env *env)
5086 deactivate_task(env->src_rq, p, 0);
5087 set_task_cpu(p, env->dst_cpu);
5088 activate_task(env->dst_rq, p, 0);
5089 check_preempt_curr(env->dst_rq, p, 0);
5093 * Is this task likely cache-hot:
5096 task_hot(struct task_struct *p, u64 now)
5100 if (p->sched_class != &fair_sched_class)
5103 if (unlikely(p->policy == SCHED_IDLE))
5107 * Buddy candidates are cache hot:
5109 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
5110 (&p->se == cfs_rq_of(&p->se)->next ||
5111 &p->se == cfs_rq_of(&p->se)->last))
5114 if (sysctl_sched_migration_cost == -1)
5116 if (sysctl_sched_migration_cost == 0)
5119 delta = now - p->se.exec_start;
5121 return delta < (s64)sysctl_sched_migration_cost;
5124 #ifdef CONFIG_NUMA_BALANCING
5125 /* Returns true if the destination node has incurred more faults */
5126 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5128 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5129 int src_nid, dst_nid;
5131 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5132 !(env->sd->flags & SD_NUMA)) {
5136 src_nid = cpu_to_node(env->src_cpu);
5137 dst_nid = cpu_to_node(env->dst_cpu);
5139 if (src_nid == dst_nid)
5143 /* Task is already in the group's interleave set. */
5144 if (node_isset(src_nid, numa_group->active_nodes))
5147 /* Task is moving into the group's interleave set. */
5148 if (node_isset(dst_nid, numa_group->active_nodes))
5151 return group_faults(p, dst_nid) > group_faults(p, src_nid);
5154 /* Encourage migration to the preferred node. */
5155 if (dst_nid == p->numa_preferred_nid)
5158 return task_faults(p, dst_nid) > task_faults(p, src_nid);
5162 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5164 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5165 int src_nid, dst_nid;
5167 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5170 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5173 src_nid = cpu_to_node(env->src_cpu);
5174 dst_nid = cpu_to_node(env->dst_cpu);
5176 if (src_nid == dst_nid)
5180 /* Task is moving within/into the group's interleave set. */
5181 if (node_isset(dst_nid, numa_group->active_nodes))
5184 /* Task is moving out of the group's interleave set. */
5185 if (node_isset(src_nid, numa_group->active_nodes))
5188 return group_faults(p, dst_nid) < group_faults(p, src_nid);
5191 /* Migrating away from the preferred node is always bad. */
5192 if (src_nid == p->numa_preferred_nid)
5195 return task_faults(p, dst_nid) < task_faults(p, src_nid);
5199 static inline bool migrate_improves_locality(struct task_struct *p,
5205 static inline bool migrate_degrades_locality(struct task_struct *p,
5213 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5216 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5218 int tsk_cache_hot = 0;
5220 * We do not migrate tasks that are:
5221 * 1) throttled_lb_pair, or
5222 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5223 * 3) running (obviously), or
5224 * 4) are cache-hot on their current CPU.
5226 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5229 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5232 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5234 env->flags |= LBF_SOME_PINNED;
5237 * Remember if this task can be migrated to any other cpu in
5238 * our sched_group. We may want to revisit it if we couldn't
5239 * meet load balance goals by pulling other tasks on src_cpu.
5241 * Also avoid computing new_dst_cpu if we have already computed
5242 * one in current iteration.
5244 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5247 /* Prevent to re-select dst_cpu via env's cpus */
5248 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5249 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5250 env->flags |= LBF_DST_PINNED;
5251 env->new_dst_cpu = cpu;
5259 /* Record that we found atleast one task that could run on dst_cpu */
5260 env->flags &= ~LBF_ALL_PINNED;
5262 if (task_running(env->src_rq, p)) {
5263 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5268 * Aggressive migration if:
5269 * 1) destination numa is preferred
5270 * 2) task is cache cold, or
5271 * 3) too many balance attempts have failed.
5273 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq));
5275 tsk_cache_hot = migrate_degrades_locality(p, env);
5277 if (migrate_improves_locality(p, env)) {
5278 #ifdef CONFIG_SCHEDSTATS
5279 if (tsk_cache_hot) {
5280 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5281 schedstat_inc(p, se.statistics.nr_forced_migrations);
5287 if (!tsk_cache_hot ||
5288 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5290 if (tsk_cache_hot) {
5291 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5292 schedstat_inc(p, se.statistics.nr_forced_migrations);
5298 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5303 * move_one_task tries to move exactly one task from busiest to this_rq, as
5304 * part of active balancing operations within "domain".
5305 * Returns 1 if successful and 0 otherwise.
5307 * Called with both runqueues locked.
5309 static int move_one_task(struct lb_env *env)
5311 struct task_struct *p, *n;
5313 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5314 if (!can_migrate_task(p, env))
5319 * Right now, this is only the second place move_task()
5320 * is called, so we can safely collect move_task()
5321 * stats here rather than inside move_task().
5323 schedstat_inc(env->sd, lb_gained[env->idle]);
5329 static const unsigned int sched_nr_migrate_break = 32;
5332 * move_tasks tries to move up to imbalance weighted load from busiest to
5333 * this_rq, as part of a balancing operation within domain "sd".
5334 * Returns 1 if successful and 0 otherwise.
5336 * Called with both runqueues locked.
5338 static int move_tasks(struct lb_env *env)
5340 struct list_head *tasks = &env->src_rq->cfs_tasks;
5341 struct task_struct *p;
5345 if (env->imbalance <= 0)
5348 while (!list_empty(tasks)) {
5349 p = list_first_entry(tasks, struct task_struct, se.group_node);
5352 /* We've more or less seen every task there is, call it quits */
5353 if (env->loop > env->loop_max)
5356 /* take a breather every nr_migrate tasks */
5357 if (env->loop > env->loop_break) {
5358 env->loop_break += sched_nr_migrate_break;
5359 env->flags |= LBF_NEED_BREAK;
5363 if (!can_migrate_task(p, env))
5366 load = task_h_load(p);
5368 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5371 if ((load / 2) > env->imbalance)
5376 env->imbalance -= load;
5378 #ifdef CONFIG_PREEMPT
5380 * NEWIDLE balancing is a source of latency, so preemptible
5381 * kernels will stop after the first task is pulled to minimize
5382 * the critical section.
5384 if (env->idle == CPU_NEWLY_IDLE)
5389 * We only want to steal up to the prescribed amount of
5392 if (env->imbalance <= 0)
5397 list_move_tail(&p->se.group_node, tasks);
5401 * Right now, this is one of only two places move_task() is called,
5402 * so we can safely collect move_task() stats here rather than
5403 * inside move_task().
5405 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5410 #ifdef CONFIG_FAIR_GROUP_SCHED
5412 * update tg->load_weight by folding this cpu's load_avg
5414 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5416 struct sched_entity *se = tg->se[cpu];
5417 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5419 /* throttled entities do not contribute to load */
5420 if (throttled_hierarchy(cfs_rq))
5423 update_cfs_rq_blocked_load(cfs_rq, 1);
5426 update_entity_load_avg(se, 1);
5428 * We pivot on our runnable average having decayed to zero for
5429 * list removal. This generally implies that all our children
5430 * have also been removed (modulo rounding error or bandwidth
5431 * control); however, such cases are rare and we can fix these
5434 * TODO: fix up out-of-order children on enqueue.
5436 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5437 list_del_leaf_cfs_rq(cfs_rq);
5439 struct rq *rq = rq_of(cfs_rq);
5440 update_rq_runnable_avg(rq, rq->nr_running);
5444 static void update_blocked_averages(int cpu)
5446 struct rq *rq = cpu_rq(cpu);
5447 struct cfs_rq *cfs_rq;
5448 unsigned long flags;
5450 raw_spin_lock_irqsave(&rq->lock, flags);
5451 update_rq_clock(rq);
5453 * Iterates the task_group tree in a bottom up fashion, see
5454 * list_add_leaf_cfs_rq() for details.
5456 for_each_leaf_cfs_rq(rq, cfs_rq) {
5458 * Note: We may want to consider periodically releasing
5459 * rq->lock about these updates so that creating many task
5460 * groups does not result in continually extending hold time.
5462 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5465 raw_spin_unlock_irqrestore(&rq->lock, flags);
5469 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5470 * This needs to be done in a top-down fashion because the load of a child
5471 * group is a fraction of its parents load.
5473 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5475 struct rq *rq = rq_of(cfs_rq);
5476 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5477 unsigned long now = jiffies;
5480 if (cfs_rq->last_h_load_update == now)
5483 cfs_rq->h_load_next = NULL;
5484 for_each_sched_entity(se) {
5485 cfs_rq = cfs_rq_of(se);
5486 cfs_rq->h_load_next = se;
5487 if (cfs_rq->last_h_load_update == now)
5492 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5493 cfs_rq->last_h_load_update = now;
5496 while ((se = cfs_rq->h_load_next) != NULL) {
5497 load = cfs_rq->h_load;
5498 load = div64_ul(load * se->avg.load_avg_contrib,
5499 cfs_rq->runnable_load_avg + 1);
5500 cfs_rq = group_cfs_rq(se);
5501 cfs_rq->h_load = load;
5502 cfs_rq->last_h_load_update = now;
5506 static unsigned long task_h_load(struct task_struct *p)
5508 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5510 update_cfs_rq_h_load(cfs_rq);
5511 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5512 cfs_rq->runnable_load_avg + 1);
5515 static inline void update_blocked_averages(int cpu)
5519 static unsigned long task_h_load(struct task_struct *p)
5521 return p->se.avg.load_avg_contrib;
5525 /********** Helpers for find_busiest_group ************************/
5527 * sg_lb_stats - stats of a sched_group required for load_balancing
5529 struct sg_lb_stats {
5530 unsigned long avg_load; /*Avg load across the CPUs of the group */
5531 unsigned long group_load; /* Total load over the CPUs of the group */
5532 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5533 unsigned long load_per_task;
5534 unsigned long group_power;
5535 unsigned int sum_nr_running; /* Nr tasks running in the group */
5536 unsigned int group_capacity;
5537 unsigned int idle_cpus;
5538 unsigned int group_weight;
5539 int group_imb; /* Is there an imbalance in the group ? */
5540 int group_has_capacity; /* Is there extra capacity in the group? */
5541 #ifdef CONFIG_NUMA_BALANCING
5542 unsigned int nr_numa_running;
5543 unsigned int nr_preferred_running;
5548 * sd_lb_stats - Structure to store the statistics of a sched_domain
5549 * during load balancing.
5551 struct sd_lb_stats {
5552 struct sched_group *busiest; /* Busiest group in this sd */
5553 struct sched_group *local; /* Local group in this sd */
5554 unsigned long total_load; /* Total load of all groups in sd */
5555 unsigned long total_pwr; /* Total power of all groups in sd */
5556 unsigned long avg_load; /* Average load across all groups in sd */
5558 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5559 struct sg_lb_stats local_stat; /* Statistics of the local group */
5562 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5565 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5566 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5567 * We must however clear busiest_stat::avg_load because
5568 * update_sd_pick_busiest() reads this before assignment.
5570 *sds = (struct sd_lb_stats){
5582 * get_sd_load_idx - Obtain the load index for a given sched domain.
5583 * @sd: The sched_domain whose load_idx is to be obtained.
5584 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5586 * Return: The load index.
5588 static inline int get_sd_load_idx(struct sched_domain *sd,
5589 enum cpu_idle_type idle)
5595 load_idx = sd->busy_idx;
5598 case CPU_NEWLY_IDLE:
5599 load_idx = sd->newidle_idx;
5602 load_idx = sd->idle_idx;
5609 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5611 return SCHED_POWER_SCALE;
5614 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5616 return default_scale_freq_power(sd, cpu);
5619 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5621 unsigned long weight = sd->span_weight;
5622 unsigned long smt_gain = sd->smt_gain;
5629 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5631 return default_scale_smt_power(sd, cpu);
5634 static unsigned long scale_rt_power(int cpu)
5636 struct rq *rq = cpu_rq(cpu);
5637 u64 total, available, age_stamp, avg;
5641 * Since we're reading these variables without serialization make sure
5642 * we read them once before doing sanity checks on them.
5644 age_stamp = ACCESS_ONCE(rq->age_stamp);
5645 avg = ACCESS_ONCE(rq->rt_avg);
5647 delta = rq_clock(rq) - age_stamp;
5648 if (unlikely(delta < 0))
5651 total = sched_avg_period() + delta;
5653 if (unlikely(total < avg)) {
5654 /* Ensures that power won't end up being negative */
5657 available = total - avg;
5660 if (unlikely((s64)total < SCHED_POWER_SCALE))
5661 total = SCHED_POWER_SCALE;
5663 total >>= SCHED_POWER_SHIFT;
5665 return div_u64(available, total);
5668 static void update_cpu_power(struct sched_domain *sd, int cpu)
5670 unsigned long weight = sd->span_weight;
5671 unsigned long power = SCHED_POWER_SCALE;
5672 struct sched_group *sdg = sd->groups;
5674 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5675 if (sched_feat(ARCH_POWER))
5676 power *= arch_scale_smt_power(sd, cpu);
5678 power *= default_scale_smt_power(sd, cpu);
5680 power >>= SCHED_POWER_SHIFT;
5683 sdg->sgp->power_orig = power;
5685 if (sched_feat(ARCH_POWER))
5686 power *= arch_scale_freq_power(sd, cpu);
5688 power *= default_scale_freq_power(sd, cpu);
5690 power >>= SCHED_POWER_SHIFT;
5692 power *= scale_rt_power(cpu);
5693 power >>= SCHED_POWER_SHIFT;
5698 cpu_rq(cpu)->cpu_power = power;
5699 sdg->sgp->power = power;
5702 void update_group_power(struct sched_domain *sd, int cpu)
5704 struct sched_domain *child = sd->child;
5705 struct sched_group *group, *sdg = sd->groups;
5706 unsigned long power, power_orig;
5707 unsigned long interval;
5709 interval = msecs_to_jiffies(sd->balance_interval);
5710 interval = clamp(interval, 1UL, max_load_balance_interval);
5711 sdg->sgp->next_update = jiffies + interval;
5714 update_cpu_power(sd, cpu);
5718 power_orig = power = 0;
5720 if (child->flags & SD_OVERLAP) {
5722 * SD_OVERLAP domains cannot assume that child groups
5723 * span the current group.
5726 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5727 struct sched_group_power *sgp;
5728 struct rq *rq = cpu_rq(cpu);
5731 * build_sched_domains() -> init_sched_groups_power()
5732 * gets here before we've attached the domains to the
5735 * Use power_of(), which is set irrespective of domains
5736 * in update_cpu_power().
5738 * This avoids power/power_orig from being 0 and
5739 * causing divide-by-zero issues on boot.
5741 * Runtime updates will correct power_orig.
5743 if (unlikely(!rq->sd)) {
5744 power_orig += power_of(cpu);
5745 power += power_of(cpu);
5749 sgp = rq->sd->groups->sgp;
5750 power_orig += sgp->power_orig;
5751 power += sgp->power;
5755 * !SD_OVERLAP domains can assume that child groups
5756 * span the current group.
5759 group = child->groups;
5761 power_orig += group->sgp->power_orig;
5762 power += group->sgp->power;
5763 group = group->next;
5764 } while (group != child->groups);
5767 sdg->sgp->power_orig = power_orig;
5768 sdg->sgp->power = power;
5772 * Try and fix up capacity for tiny siblings, this is needed when
5773 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5774 * which on its own isn't powerful enough.
5776 * See update_sd_pick_busiest() and check_asym_packing().
5779 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5782 * Only siblings can have significantly less than SCHED_POWER_SCALE
5784 if (!(sd->flags & SD_SHARE_CPUPOWER))
5788 * If ~90% of the cpu_power is still there, we're good.
5790 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5797 * Group imbalance indicates (and tries to solve) the problem where balancing
5798 * groups is inadequate due to tsk_cpus_allowed() constraints.
5800 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5801 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5804 * { 0 1 2 3 } { 4 5 6 7 }
5807 * If we were to balance group-wise we'd place two tasks in the first group and
5808 * two tasks in the second group. Clearly this is undesired as it will overload
5809 * cpu 3 and leave one of the cpus in the second group unused.
5811 * The current solution to this issue is detecting the skew in the first group
5812 * by noticing the lower domain failed to reach balance and had difficulty
5813 * moving tasks due to affinity constraints.
5815 * When this is so detected; this group becomes a candidate for busiest; see
5816 * update_sd_pick_busiest(). And calculate_imbalance() and
5817 * find_busiest_group() avoid some of the usual balance conditions to allow it
5818 * to create an effective group imbalance.
5820 * This is a somewhat tricky proposition since the next run might not find the
5821 * group imbalance and decide the groups need to be balanced again. A most
5822 * subtle and fragile situation.
5825 static inline int sg_imbalanced(struct sched_group *group)
5827 return group->sgp->imbalance;
5831 * Compute the group capacity.
5833 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5834 * first dividing out the smt factor and computing the actual number of cores
5835 * and limit power unit capacity with that.
5837 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5839 unsigned int capacity, smt, cpus;
5840 unsigned int power, power_orig;
5842 power = group->sgp->power;
5843 power_orig = group->sgp->power_orig;
5844 cpus = group->group_weight;
5846 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5847 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5848 capacity = cpus / smt; /* cores */
5850 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5852 capacity = fix_small_capacity(env->sd, group);
5858 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5859 * @env: The load balancing environment.
5860 * @group: sched_group whose statistics are to be updated.
5861 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5862 * @local_group: Does group contain this_cpu.
5863 * @sgs: variable to hold the statistics for this group.
5865 static inline void update_sg_lb_stats(struct lb_env *env,
5866 struct sched_group *group, int load_idx,
5867 int local_group, struct sg_lb_stats *sgs)
5872 memset(sgs, 0, sizeof(*sgs));
5874 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5875 struct rq *rq = cpu_rq(i);
5877 /* Bias balancing toward cpus of our domain */
5879 load = target_load(i, load_idx);
5881 load = source_load(i, load_idx);
5883 sgs->group_load += load;
5884 sgs->sum_nr_running += rq->nr_running;
5885 #ifdef CONFIG_NUMA_BALANCING
5886 sgs->nr_numa_running += rq->nr_numa_running;
5887 sgs->nr_preferred_running += rq->nr_preferred_running;
5889 sgs->sum_weighted_load += weighted_cpuload(i);
5894 /* Adjust by relative CPU power of the group */
5895 sgs->group_power = group->sgp->power;
5896 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5898 if (sgs->sum_nr_running)
5899 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5901 sgs->group_weight = group->group_weight;
5903 sgs->group_imb = sg_imbalanced(group);
5904 sgs->group_capacity = sg_capacity(env, group);
5906 if (sgs->group_capacity > sgs->sum_nr_running)
5907 sgs->group_has_capacity = 1;
5911 * update_sd_pick_busiest - return 1 on busiest group
5912 * @env: The load balancing environment.
5913 * @sds: sched_domain statistics
5914 * @sg: sched_group candidate to be checked for being the busiest
5915 * @sgs: sched_group statistics
5917 * Determine if @sg is a busier group than the previously selected
5920 * Return: %true if @sg is a busier group than the previously selected
5921 * busiest group. %false otherwise.
5923 static bool update_sd_pick_busiest(struct lb_env *env,
5924 struct sd_lb_stats *sds,
5925 struct sched_group *sg,
5926 struct sg_lb_stats *sgs)
5928 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5931 if (sgs->sum_nr_running > sgs->group_capacity)
5938 * ASYM_PACKING needs to move all the work to the lowest
5939 * numbered CPUs in the group, therefore mark all groups
5940 * higher than ourself as busy.
5942 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5943 env->dst_cpu < group_first_cpu(sg)) {
5947 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5954 #ifdef CONFIG_NUMA_BALANCING
5955 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5957 if (sgs->sum_nr_running > sgs->nr_numa_running)
5959 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5964 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5966 if (rq->nr_running > rq->nr_numa_running)
5968 if (rq->nr_running > rq->nr_preferred_running)
5973 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5978 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5982 #endif /* CONFIG_NUMA_BALANCING */
5985 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5986 * @env: The load balancing environment.
5987 * @sds: variable to hold the statistics for this sched_domain.
5989 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5991 struct sched_domain *child = env->sd->child;
5992 struct sched_group *sg = env->sd->groups;
5993 struct sg_lb_stats tmp_sgs;
5994 int load_idx, prefer_sibling = 0;
5996 if (child && child->flags & SD_PREFER_SIBLING)
5999 load_idx = get_sd_load_idx(env->sd, env->idle);
6002 struct sg_lb_stats *sgs = &tmp_sgs;
6005 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6008 sgs = &sds->local_stat;
6010 if (env->idle != CPU_NEWLY_IDLE ||
6011 time_after_eq(jiffies, sg->sgp->next_update))
6012 update_group_power(env->sd, env->dst_cpu);
6015 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
6021 * In case the child domain prefers tasks go to siblings
6022 * first, lower the sg capacity to one so that we'll try
6023 * and move all the excess tasks away. We lower the capacity
6024 * of a group only if the local group has the capacity to fit
6025 * these excess tasks, i.e. nr_running < group_capacity. The
6026 * extra check prevents the case where you always pull from the
6027 * heaviest group when it is already under-utilized (possible
6028 * with a large weight task outweighs the tasks on the system).
6030 if (prefer_sibling && sds->local &&
6031 sds->local_stat.group_has_capacity)
6032 sgs->group_capacity = min(sgs->group_capacity, 1U);
6034 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6036 sds->busiest_stat = *sgs;
6040 /* Now, start updating sd_lb_stats */
6041 sds->total_load += sgs->group_load;
6042 sds->total_pwr += sgs->group_power;
6045 } while (sg != env->sd->groups);
6047 if (env->sd->flags & SD_NUMA)
6048 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6052 * check_asym_packing - Check to see if the group is packed into the
6055 * This is primarily intended to used at the sibling level. Some
6056 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6057 * case of POWER7, it can move to lower SMT modes only when higher
6058 * threads are idle. When in lower SMT modes, the threads will
6059 * perform better since they share less core resources. Hence when we
6060 * have idle threads, we want them to be the higher ones.
6062 * This packing function is run on idle threads. It checks to see if
6063 * the busiest CPU in this domain (core in the P7 case) has a higher
6064 * CPU number than the packing function is being run on. Here we are
6065 * assuming lower CPU number will be equivalent to lower a SMT thread
6068 * Return: 1 when packing is required and a task should be moved to
6069 * this CPU. The amount of the imbalance is returned in *imbalance.
6071 * @env: The load balancing environment.
6072 * @sds: Statistics of the sched_domain which is to be packed
6074 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6078 if (!(env->sd->flags & SD_ASYM_PACKING))
6084 busiest_cpu = group_first_cpu(sds->busiest);
6085 if (env->dst_cpu > busiest_cpu)
6088 env->imbalance = DIV_ROUND_CLOSEST(
6089 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
6096 * fix_small_imbalance - Calculate the minor imbalance that exists
6097 * amongst the groups of a sched_domain, during
6099 * @env: The load balancing environment.
6100 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6103 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6105 unsigned long tmp, pwr_now = 0, pwr_move = 0;
6106 unsigned int imbn = 2;
6107 unsigned long scaled_busy_load_per_task;
6108 struct sg_lb_stats *local, *busiest;
6110 local = &sds->local_stat;
6111 busiest = &sds->busiest_stat;
6113 if (!local->sum_nr_running)
6114 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6115 else if (busiest->load_per_task > local->load_per_task)
6118 scaled_busy_load_per_task =
6119 (busiest->load_per_task * SCHED_POWER_SCALE) /
6120 busiest->group_power;
6122 if (busiest->avg_load + scaled_busy_load_per_task >=
6123 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6124 env->imbalance = busiest->load_per_task;
6129 * OK, we don't have enough imbalance to justify moving tasks,
6130 * however we may be able to increase total CPU power used by
6134 pwr_now += busiest->group_power *
6135 min(busiest->load_per_task, busiest->avg_load);
6136 pwr_now += local->group_power *
6137 min(local->load_per_task, local->avg_load);
6138 pwr_now /= SCHED_POWER_SCALE;
6140 /* Amount of load we'd subtract */
6141 if (busiest->avg_load > scaled_busy_load_per_task) {
6142 pwr_move += busiest->group_power *
6143 min(busiest->load_per_task,
6144 busiest->avg_load - scaled_busy_load_per_task);
6147 /* Amount of load we'd add */
6148 if (busiest->avg_load * busiest->group_power <
6149 busiest->load_per_task * SCHED_POWER_SCALE) {
6150 tmp = (busiest->avg_load * busiest->group_power) /
6153 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
6156 pwr_move += local->group_power *
6157 min(local->load_per_task, local->avg_load + tmp);
6158 pwr_move /= SCHED_POWER_SCALE;
6160 /* Move if we gain throughput */
6161 if (pwr_move > pwr_now)
6162 env->imbalance = busiest->load_per_task;
6166 * calculate_imbalance - Calculate the amount of imbalance present within the
6167 * groups of a given sched_domain during load balance.
6168 * @env: load balance environment
6169 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6171 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6173 unsigned long max_pull, load_above_capacity = ~0UL;
6174 struct sg_lb_stats *local, *busiest;
6176 local = &sds->local_stat;
6177 busiest = &sds->busiest_stat;
6179 if (busiest->group_imb) {
6181 * In the group_imb case we cannot rely on group-wide averages
6182 * to ensure cpu-load equilibrium, look at wider averages. XXX
6184 busiest->load_per_task =
6185 min(busiest->load_per_task, sds->avg_load);
6189 * In the presence of smp nice balancing, certain scenarios can have
6190 * max load less than avg load(as we skip the groups at or below
6191 * its cpu_power, while calculating max_load..)
6193 if (busiest->avg_load <= sds->avg_load ||
6194 local->avg_load >= sds->avg_load) {
6196 return fix_small_imbalance(env, sds);
6199 if (!busiest->group_imb) {
6201 * Don't want to pull so many tasks that a group would go idle.
6202 * Except of course for the group_imb case, since then we might
6203 * have to drop below capacity to reach cpu-load equilibrium.
6205 load_above_capacity =
6206 (busiest->sum_nr_running - busiest->group_capacity);
6208 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
6209 load_above_capacity /= busiest->group_power;
6213 * We're trying to get all the cpus to the average_load, so we don't
6214 * want to push ourselves above the average load, nor do we wish to
6215 * reduce the max loaded cpu below the average load. At the same time,
6216 * we also don't want to reduce the group load below the group capacity
6217 * (so that we can implement power-savings policies etc). Thus we look
6218 * for the minimum possible imbalance.
6220 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6222 /* How much load to actually move to equalise the imbalance */
6223 env->imbalance = min(
6224 max_pull * busiest->group_power,
6225 (sds->avg_load - local->avg_load) * local->group_power
6226 ) / SCHED_POWER_SCALE;
6229 * if *imbalance is less than the average load per runnable task
6230 * there is no guarantee that any tasks will be moved so we'll have
6231 * a think about bumping its value to force at least one task to be
6234 if (env->imbalance < busiest->load_per_task)
6235 return fix_small_imbalance(env, sds);
6238 /******* find_busiest_group() helpers end here *********************/
6241 * find_busiest_group - Returns the busiest group within the sched_domain
6242 * if there is an imbalance. If there isn't an imbalance, and
6243 * the user has opted for power-savings, it returns a group whose
6244 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6245 * such a group exists.
6247 * Also calculates the amount of weighted load which should be moved
6248 * to restore balance.
6250 * @env: The load balancing environment.
6252 * Return: - The busiest group if imbalance exists.
6253 * - If no imbalance and user has opted for power-savings balance,
6254 * return the least loaded group whose CPUs can be
6255 * put to idle by rebalancing its tasks onto our group.
6257 static struct sched_group *find_busiest_group(struct lb_env *env)
6259 struct sg_lb_stats *local, *busiest;
6260 struct sd_lb_stats sds;
6262 init_sd_lb_stats(&sds);
6265 * Compute the various statistics relavent for load balancing at
6268 update_sd_lb_stats(env, &sds);
6269 local = &sds.local_stat;
6270 busiest = &sds.busiest_stat;
6272 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6273 check_asym_packing(env, &sds))
6276 /* There is no busy sibling group to pull tasks from */
6277 if (!sds.busiest || busiest->sum_nr_running == 0)
6280 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
6283 * If the busiest group is imbalanced the below checks don't
6284 * work because they assume all things are equal, which typically
6285 * isn't true due to cpus_allowed constraints and the like.
6287 if (busiest->group_imb)
6290 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6291 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
6292 !busiest->group_has_capacity)
6296 * If the local group is more busy than the selected busiest group
6297 * don't try and pull any tasks.
6299 if (local->avg_load >= busiest->avg_load)
6303 * Don't pull any tasks if this group is already above the domain
6306 if (local->avg_load >= sds.avg_load)
6309 if (env->idle == CPU_IDLE) {
6311 * This cpu is idle. If the busiest group load doesn't
6312 * have more tasks than the number of available cpu's and
6313 * there is no imbalance between this and busiest group
6314 * wrt to idle cpu's, it is balanced.
6316 if ((local->idle_cpus < busiest->idle_cpus) &&
6317 busiest->sum_nr_running <= busiest->group_weight)
6321 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6322 * imbalance_pct to be conservative.
6324 if (100 * busiest->avg_load <=
6325 env->sd->imbalance_pct * local->avg_load)
6330 /* Looks like there is an imbalance. Compute it */
6331 calculate_imbalance(env, &sds);
6340 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6342 static struct rq *find_busiest_queue(struct lb_env *env,
6343 struct sched_group *group)
6345 struct rq *busiest = NULL, *rq;
6346 unsigned long busiest_load = 0, busiest_power = 1;
6349 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6350 unsigned long power, capacity, wl;
6354 rt = fbq_classify_rq(rq);
6357 * We classify groups/runqueues into three groups:
6358 * - regular: there are !numa tasks
6359 * - remote: there are numa tasks that run on the 'wrong' node
6360 * - all: there is no distinction
6362 * In order to avoid migrating ideally placed numa tasks,
6363 * ignore those when there's better options.
6365 * If we ignore the actual busiest queue to migrate another
6366 * task, the next balance pass can still reduce the busiest
6367 * queue by moving tasks around inside the node.
6369 * If we cannot move enough load due to this classification
6370 * the next pass will adjust the group classification and
6371 * allow migration of more tasks.
6373 * Both cases only affect the total convergence complexity.
6375 if (rt > env->fbq_type)
6378 power = power_of(i);
6379 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
6381 capacity = fix_small_capacity(env->sd, group);
6383 wl = weighted_cpuload(i);
6386 * When comparing with imbalance, use weighted_cpuload()
6387 * which is not scaled with the cpu power.
6389 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6393 * For the load comparisons with the other cpu's, consider
6394 * the weighted_cpuload() scaled with the cpu power, so that
6395 * the load can be moved away from the cpu that is potentially
6396 * running at a lower capacity.
6398 * Thus we're looking for max(wl_i / power_i), crosswise
6399 * multiplication to rid ourselves of the division works out
6400 * to: wl_i * power_j > wl_j * power_i; where j is our
6403 if (wl * busiest_power > busiest_load * power) {
6405 busiest_power = power;
6414 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6415 * so long as it is large enough.
6417 #define MAX_PINNED_INTERVAL 512
6419 /* Working cpumask for load_balance and load_balance_newidle. */
6420 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6422 static int need_active_balance(struct lb_env *env)
6424 struct sched_domain *sd = env->sd;
6426 if (env->idle == CPU_NEWLY_IDLE) {
6429 * ASYM_PACKING needs to force migrate tasks from busy but
6430 * higher numbered CPUs in order to pack all tasks in the
6431 * lowest numbered CPUs.
6433 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6437 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6440 static int active_load_balance_cpu_stop(void *data);
6442 static int should_we_balance(struct lb_env *env)
6444 struct sched_group *sg = env->sd->groups;
6445 struct cpumask *sg_cpus, *sg_mask;
6446 int cpu, balance_cpu = -1;
6449 * In the newly idle case, we will allow all the cpu's
6450 * to do the newly idle load balance.
6452 if (env->idle == CPU_NEWLY_IDLE)
6455 sg_cpus = sched_group_cpus(sg);
6456 sg_mask = sched_group_mask(sg);
6457 /* Try to find first idle cpu */
6458 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6459 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6466 if (balance_cpu == -1)
6467 balance_cpu = group_balance_cpu(sg);
6470 * First idle cpu or the first cpu(busiest) in this sched group
6471 * is eligible for doing load balancing at this and above domains.
6473 return balance_cpu == env->dst_cpu;
6477 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6478 * tasks if there is an imbalance.
6480 static int load_balance(int this_cpu, struct rq *this_rq,
6481 struct sched_domain *sd, enum cpu_idle_type idle,
6482 int *continue_balancing)
6484 int ld_moved, cur_ld_moved, active_balance = 0;
6485 struct sched_domain *sd_parent = sd->parent;
6486 struct sched_group *group;
6488 unsigned long flags;
6489 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6491 struct lb_env env = {
6493 .dst_cpu = this_cpu,
6495 .dst_grpmask = sched_group_cpus(sd->groups),
6497 .loop_break = sched_nr_migrate_break,
6503 * For NEWLY_IDLE load_balancing, we don't need to consider
6504 * other cpus in our group
6506 if (idle == CPU_NEWLY_IDLE)
6507 env.dst_grpmask = NULL;
6509 cpumask_copy(cpus, cpu_active_mask);
6511 schedstat_inc(sd, lb_count[idle]);
6514 if (!should_we_balance(&env)) {
6515 *continue_balancing = 0;
6519 group = find_busiest_group(&env);
6521 schedstat_inc(sd, lb_nobusyg[idle]);
6525 busiest = find_busiest_queue(&env, group);
6527 schedstat_inc(sd, lb_nobusyq[idle]);
6531 BUG_ON(busiest == env.dst_rq);
6533 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6536 if (busiest->nr_running > 1) {
6538 * Attempt to move tasks. If find_busiest_group has found
6539 * an imbalance but busiest->nr_running <= 1, the group is
6540 * still unbalanced. ld_moved simply stays zero, so it is
6541 * correctly treated as an imbalance.
6543 env.flags |= LBF_ALL_PINNED;
6544 env.src_cpu = busiest->cpu;
6545 env.src_rq = busiest;
6546 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6549 local_irq_save(flags);
6550 double_rq_lock(env.dst_rq, busiest);
6553 * cur_ld_moved - load moved in current iteration
6554 * ld_moved - cumulative load moved across iterations
6556 cur_ld_moved = move_tasks(&env);
6557 ld_moved += cur_ld_moved;
6558 double_rq_unlock(env.dst_rq, busiest);
6559 local_irq_restore(flags);
6562 * some other cpu did the load balance for us.
6564 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6565 resched_cpu(env.dst_cpu);
6567 if (env.flags & LBF_NEED_BREAK) {
6568 env.flags &= ~LBF_NEED_BREAK;
6573 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6574 * us and move them to an alternate dst_cpu in our sched_group
6575 * where they can run. The upper limit on how many times we
6576 * iterate on same src_cpu is dependent on number of cpus in our
6579 * This changes load balance semantics a bit on who can move
6580 * load to a given_cpu. In addition to the given_cpu itself
6581 * (or a ilb_cpu acting on its behalf where given_cpu is
6582 * nohz-idle), we now have balance_cpu in a position to move
6583 * load to given_cpu. In rare situations, this may cause
6584 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6585 * _independently_ and at _same_ time to move some load to
6586 * given_cpu) causing exceess load to be moved to given_cpu.
6587 * This however should not happen so much in practice and
6588 * moreover subsequent load balance cycles should correct the
6589 * excess load moved.
6591 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6593 /* Prevent to re-select dst_cpu via env's cpus */
6594 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6596 env.dst_rq = cpu_rq(env.new_dst_cpu);
6597 env.dst_cpu = env.new_dst_cpu;
6598 env.flags &= ~LBF_DST_PINNED;
6600 env.loop_break = sched_nr_migrate_break;
6603 * Go back to "more_balance" rather than "redo" since we
6604 * need to continue with same src_cpu.
6610 * We failed to reach balance because of affinity.
6613 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6615 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6616 *group_imbalance = 1;
6617 } else if (*group_imbalance)
6618 *group_imbalance = 0;
6621 /* All tasks on this runqueue were pinned by CPU affinity */
6622 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6623 cpumask_clear_cpu(cpu_of(busiest), cpus);
6624 if (!cpumask_empty(cpus)) {
6626 env.loop_break = sched_nr_migrate_break;
6634 schedstat_inc(sd, lb_failed[idle]);
6636 * Increment the failure counter only on periodic balance.
6637 * We do not want newidle balance, which can be very
6638 * frequent, pollute the failure counter causing
6639 * excessive cache_hot migrations and active balances.
6641 if (idle != CPU_NEWLY_IDLE)
6642 sd->nr_balance_failed++;
6644 if (need_active_balance(&env)) {
6645 raw_spin_lock_irqsave(&busiest->lock, flags);
6647 /* don't kick the active_load_balance_cpu_stop,
6648 * if the curr task on busiest cpu can't be
6651 if (!cpumask_test_cpu(this_cpu,
6652 tsk_cpus_allowed(busiest->curr))) {
6653 raw_spin_unlock_irqrestore(&busiest->lock,
6655 env.flags |= LBF_ALL_PINNED;
6656 goto out_one_pinned;
6660 * ->active_balance synchronizes accesses to
6661 * ->active_balance_work. Once set, it's cleared
6662 * only after active load balance is finished.
6664 if (!busiest->active_balance) {
6665 busiest->active_balance = 1;
6666 busiest->push_cpu = this_cpu;
6669 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6671 if (active_balance) {
6672 stop_one_cpu_nowait(cpu_of(busiest),
6673 active_load_balance_cpu_stop, busiest,
6674 &busiest->active_balance_work);
6678 * We've kicked active balancing, reset the failure
6681 sd->nr_balance_failed = sd->cache_nice_tries+1;
6684 sd->nr_balance_failed = 0;
6686 if (likely(!active_balance)) {
6687 /* We were unbalanced, so reset the balancing interval */
6688 sd->balance_interval = sd->min_interval;
6691 * If we've begun active balancing, start to back off. This
6692 * case may not be covered by the all_pinned logic if there
6693 * is only 1 task on the busy runqueue (because we don't call
6696 if (sd->balance_interval < sd->max_interval)
6697 sd->balance_interval *= 2;
6703 schedstat_inc(sd, lb_balanced[idle]);
6705 sd->nr_balance_failed = 0;
6708 /* tune up the balancing interval */
6709 if (((env.flags & LBF_ALL_PINNED) &&
6710 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6711 (sd->balance_interval < sd->max_interval))
6712 sd->balance_interval *= 2;
6719 static inline unsigned long
6720 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
6722 unsigned long interval = sd->balance_interval;
6725 interval *= sd->busy_factor;
6727 /* scale ms to jiffies */
6728 interval = msecs_to_jiffies(interval);
6729 interval = clamp(interval, 1UL, max_load_balance_interval);
6735 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
6737 unsigned long interval, next;
6739 interval = get_sd_balance_interval(sd, cpu_busy);
6740 next = sd->last_balance + interval;
6742 if (time_after(*next_balance, next))
6743 *next_balance = next;
6747 * idle_balance is called by schedule() if this_cpu is about to become
6748 * idle. Attempts to pull tasks from other CPUs.
6750 static int idle_balance(struct rq *this_rq)
6752 unsigned long next_balance = jiffies + HZ;
6753 int this_cpu = this_rq->cpu;
6754 struct sched_domain *sd;
6755 int pulled_task = 0;
6758 idle_enter_fair(this_rq);
6761 * We must set idle_stamp _before_ calling idle_balance(), such that we
6762 * measure the duration of idle_balance() as idle time.
6764 this_rq->idle_stamp = rq_clock(this_rq);
6766 if (this_rq->avg_idle < sysctl_sched_migration_cost) {
6768 sd = rcu_dereference_check_sched_domain(this_rq->sd);
6770 update_next_balance(sd, 0, &next_balance);
6777 * Drop the rq->lock, but keep IRQ/preempt disabled.
6779 raw_spin_unlock(&this_rq->lock);
6781 update_blocked_averages(this_cpu);
6783 for_each_domain(this_cpu, sd) {
6784 int continue_balancing = 1;
6785 u64 t0, domain_cost;
6787 if (!(sd->flags & SD_LOAD_BALANCE))
6790 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
6791 update_next_balance(sd, 0, &next_balance);
6795 if (sd->flags & SD_BALANCE_NEWIDLE) {
6796 t0 = sched_clock_cpu(this_cpu);
6798 pulled_task = load_balance(this_cpu, this_rq,
6800 &continue_balancing);
6802 domain_cost = sched_clock_cpu(this_cpu) - t0;
6803 if (domain_cost > sd->max_newidle_lb_cost)
6804 sd->max_newidle_lb_cost = domain_cost;
6806 curr_cost += domain_cost;
6809 update_next_balance(sd, 0, &next_balance);
6812 * Stop searching for tasks to pull if there are
6813 * now runnable tasks on this rq.
6815 if (pulled_task || this_rq->nr_running > 0)
6820 raw_spin_lock(&this_rq->lock);
6822 if (curr_cost > this_rq->max_idle_balance_cost)
6823 this_rq->max_idle_balance_cost = curr_cost;
6826 * While browsing the domains, we released the rq lock, a task could
6827 * have been enqueued in the meantime. Since we're not going idle,
6828 * pretend we pulled a task.
6830 if (this_rq->cfs.h_nr_running && !pulled_task)
6834 /* Move the next balance forward */
6835 if (time_after(this_rq->next_balance, next_balance))
6836 this_rq->next_balance = next_balance;
6838 /* Is there a task of a high priority class? */
6839 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
6843 idle_exit_fair(this_rq);
6844 this_rq->idle_stamp = 0;
6851 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6852 * running tasks off the busiest CPU onto idle CPUs. It requires at
6853 * least 1 task to be running on each physical CPU where possible, and
6854 * avoids physical / logical imbalances.
6856 static int active_load_balance_cpu_stop(void *data)
6858 struct rq *busiest_rq = data;
6859 int busiest_cpu = cpu_of(busiest_rq);
6860 int target_cpu = busiest_rq->push_cpu;
6861 struct rq *target_rq = cpu_rq(target_cpu);
6862 struct sched_domain *sd;
6864 raw_spin_lock_irq(&busiest_rq->lock);
6866 /* make sure the requested cpu hasn't gone down in the meantime */
6867 if (unlikely(busiest_cpu != smp_processor_id() ||
6868 !busiest_rq->active_balance))
6871 /* Is there any task to move? */
6872 if (busiest_rq->nr_running <= 1)
6876 * This condition is "impossible", if it occurs
6877 * we need to fix it. Originally reported by
6878 * Bjorn Helgaas on a 128-cpu setup.
6880 BUG_ON(busiest_rq == target_rq);
6882 /* move a task from busiest_rq to target_rq */
6883 double_lock_balance(busiest_rq, target_rq);
6885 /* Search for an sd spanning us and the target CPU. */
6887 for_each_domain(target_cpu, sd) {
6888 if ((sd->flags & SD_LOAD_BALANCE) &&
6889 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6894 struct lb_env env = {
6896 .dst_cpu = target_cpu,
6897 .dst_rq = target_rq,
6898 .src_cpu = busiest_rq->cpu,
6899 .src_rq = busiest_rq,
6903 schedstat_inc(sd, alb_count);
6905 if (move_one_task(&env))
6906 schedstat_inc(sd, alb_pushed);
6908 schedstat_inc(sd, alb_failed);
6911 double_unlock_balance(busiest_rq, target_rq);
6913 busiest_rq->active_balance = 0;
6914 raw_spin_unlock_irq(&busiest_rq->lock);
6918 static inline int on_null_domain(struct rq *rq)
6920 return unlikely(!rcu_dereference_sched(rq->sd));
6923 #ifdef CONFIG_NO_HZ_COMMON
6925 * idle load balancing details
6926 * - When one of the busy CPUs notice that there may be an idle rebalancing
6927 * needed, they will kick the idle load balancer, which then does idle
6928 * load balancing for all the idle CPUs.
6931 cpumask_var_t idle_cpus_mask;
6933 unsigned long next_balance; /* in jiffy units */
6934 } nohz ____cacheline_aligned;
6936 static inline int find_new_ilb(void)
6938 int ilb = cpumask_first(nohz.idle_cpus_mask);
6940 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6947 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6948 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6949 * CPU (if there is one).
6951 static void nohz_balancer_kick(void)
6955 nohz.next_balance++;
6957 ilb_cpu = find_new_ilb();
6959 if (ilb_cpu >= nr_cpu_ids)
6962 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6965 * Use smp_send_reschedule() instead of resched_cpu().
6966 * This way we generate a sched IPI on the target cpu which
6967 * is idle. And the softirq performing nohz idle load balance
6968 * will be run before returning from the IPI.
6970 smp_send_reschedule(ilb_cpu);
6974 static inline void nohz_balance_exit_idle(int cpu)
6976 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6978 * Completely isolated CPUs don't ever set, so we must test.
6980 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
6981 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6982 atomic_dec(&nohz.nr_cpus);
6984 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6988 static inline void set_cpu_sd_state_busy(void)
6990 struct sched_domain *sd;
6991 int cpu = smp_processor_id();
6994 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6996 if (!sd || !sd->nohz_idle)
7000 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
7005 void set_cpu_sd_state_idle(void)
7007 struct sched_domain *sd;
7008 int cpu = smp_processor_id();
7011 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7013 if (!sd || sd->nohz_idle)
7017 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
7023 * This routine will record that the cpu is going idle with tick stopped.
7024 * This info will be used in performing idle load balancing in the future.
7026 void nohz_balance_enter_idle(int cpu)
7029 * If this cpu is going down, then nothing needs to be done.
7031 if (!cpu_active(cpu))
7034 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7038 * If we're a completely isolated CPU, we don't play.
7040 if (on_null_domain(cpu_rq(cpu)))
7043 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7044 atomic_inc(&nohz.nr_cpus);
7045 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7048 static int sched_ilb_notifier(struct notifier_block *nfb,
7049 unsigned long action, void *hcpu)
7051 switch (action & ~CPU_TASKS_FROZEN) {
7053 nohz_balance_exit_idle(smp_processor_id());
7061 static DEFINE_SPINLOCK(balancing);
7064 * Scale the max load_balance interval with the number of CPUs in the system.
7065 * This trades load-balance latency on larger machines for less cross talk.
7067 void update_max_interval(void)
7069 max_load_balance_interval = HZ*num_online_cpus()/10;
7073 * It checks each scheduling domain to see if it is due to be balanced,
7074 * and initiates a balancing operation if so.
7076 * Balancing parameters are set up in init_sched_domains.
7078 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7080 int continue_balancing = 1;
7082 unsigned long interval;
7083 struct sched_domain *sd;
7084 /* Earliest time when we have to do rebalance again */
7085 unsigned long next_balance = jiffies + 60*HZ;
7086 int update_next_balance = 0;
7087 int need_serialize, need_decay = 0;
7090 update_blocked_averages(cpu);
7093 for_each_domain(cpu, sd) {
7095 * Decay the newidle max times here because this is a regular
7096 * visit to all the domains. Decay ~1% per second.
7098 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7099 sd->max_newidle_lb_cost =
7100 (sd->max_newidle_lb_cost * 253) / 256;
7101 sd->next_decay_max_lb_cost = jiffies + HZ;
7104 max_cost += sd->max_newidle_lb_cost;
7106 if (!(sd->flags & SD_LOAD_BALANCE))
7110 * Stop the load balance at this level. There is another
7111 * CPU in our sched group which is doing load balancing more
7114 if (!continue_balancing) {
7120 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7122 need_serialize = sd->flags & SD_SERIALIZE;
7123 if (need_serialize) {
7124 if (!spin_trylock(&balancing))
7128 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7129 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7131 * The LBF_DST_PINNED logic could have changed
7132 * env->dst_cpu, so we can't know our idle
7133 * state even if we migrated tasks. Update it.
7135 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7137 sd->last_balance = jiffies;
7138 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7141 spin_unlock(&balancing);
7143 if (time_after(next_balance, sd->last_balance + interval)) {
7144 next_balance = sd->last_balance + interval;
7145 update_next_balance = 1;
7150 * Ensure the rq-wide value also decays but keep it at a
7151 * reasonable floor to avoid funnies with rq->avg_idle.
7153 rq->max_idle_balance_cost =
7154 max((u64)sysctl_sched_migration_cost, max_cost);
7159 * next_balance will be updated only when there is a need.
7160 * When the cpu is attached to null domain for ex, it will not be
7163 if (likely(update_next_balance))
7164 rq->next_balance = next_balance;
7167 #ifdef CONFIG_NO_HZ_COMMON
7169 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7170 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7172 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7174 int this_cpu = this_rq->cpu;
7178 if (idle != CPU_IDLE ||
7179 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7182 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7183 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7187 * If this cpu gets work to do, stop the load balancing
7188 * work being done for other cpus. Next load
7189 * balancing owner will pick it up.
7194 rq = cpu_rq(balance_cpu);
7196 raw_spin_lock_irq(&rq->lock);
7197 update_rq_clock(rq);
7198 update_idle_cpu_load(rq);
7199 raw_spin_unlock_irq(&rq->lock);
7201 rebalance_domains(rq, CPU_IDLE);
7203 if (time_after(this_rq->next_balance, rq->next_balance))
7204 this_rq->next_balance = rq->next_balance;
7206 nohz.next_balance = this_rq->next_balance;
7208 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7212 * Current heuristic for kicking the idle load balancer in the presence
7213 * of an idle cpu is the system.
7214 * - This rq has more than one task.
7215 * - At any scheduler domain level, this cpu's scheduler group has multiple
7216 * busy cpu's exceeding the group's power.
7217 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7218 * domain span are idle.
7220 static inline int nohz_kick_needed(struct rq *rq)
7222 unsigned long now = jiffies;
7223 struct sched_domain *sd;
7224 struct sched_group_power *sgp;
7225 int nr_busy, cpu = rq->cpu;
7227 if (unlikely(rq->idle_balance))
7231 * We may be recently in ticked or tickless idle mode. At the first
7232 * busy tick after returning from idle, we will update the busy stats.
7234 set_cpu_sd_state_busy();
7235 nohz_balance_exit_idle(cpu);
7238 * None are in tickless mode and hence no need for NOHZ idle load
7241 if (likely(!atomic_read(&nohz.nr_cpus)))
7244 if (time_before(now, nohz.next_balance))
7247 if (rq->nr_running >= 2)
7251 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7254 sgp = sd->groups->sgp;
7255 nr_busy = atomic_read(&sgp->nr_busy_cpus);
7258 goto need_kick_unlock;
7261 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7263 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7264 sched_domain_span(sd)) < cpu))
7265 goto need_kick_unlock;
7276 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7280 * run_rebalance_domains is triggered when needed from the scheduler tick.
7281 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7283 static void run_rebalance_domains(struct softirq_action *h)
7285 struct rq *this_rq = this_rq();
7286 enum cpu_idle_type idle = this_rq->idle_balance ?
7287 CPU_IDLE : CPU_NOT_IDLE;
7289 rebalance_domains(this_rq, idle);
7292 * If this cpu has a pending nohz_balance_kick, then do the
7293 * balancing on behalf of the other idle cpus whose ticks are
7296 nohz_idle_balance(this_rq, idle);
7300 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7302 void trigger_load_balance(struct rq *rq)
7304 /* Don't need to rebalance while attached to NULL domain */
7305 if (unlikely(on_null_domain(rq)))
7308 if (time_after_eq(jiffies, rq->next_balance))
7309 raise_softirq(SCHED_SOFTIRQ);
7310 #ifdef CONFIG_NO_HZ_COMMON
7311 if (nohz_kick_needed(rq))
7312 nohz_balancer_kick();
7316 static void rq_online_fair(struct rq *rq)
7321 static void rq_offline_fair(struct rq *rq)
7325 /* Ensure any throttled groups are reachable by pick_next_task */
7326 unthrottle_offline_cfs_rqs(rq);
7329 #endif /* CONFIG_SMP */
7332 * scheduler tick hitting a task of our scheduling class:
7334 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7336 struct cfs_rq *cfs_rq;
7337 struct sched_entity *se = &curr->se;
7339 for_each_sched_entity(se) {
7340 cfs_rq = cfs_rq_of(se);
7341 entity_tick(cfs_rq, se, queued);
7344 if (numabalancing_enabled)
7345 task_tick_numa(rq, curr);
7347 update_rq_runnable_avg(rq, 1);
7351 * called on fork with the child task as argument from the parent's context
7352 * - child not yet on the tasklist
7353 * - preemption disabled
7355 static void task_fork_fair(struct task_struct *p)
7357 struct cfs_rq *cfs_rq;
7358 struct sched_entity *se = &p->se, *curr;
7359 int this_cpu = smp_processor_id();
7360 struct rq *rq = this_rq();
7361 unsigned long flags;
7363 raw_spin_lock_irqsave(&rq->lock, flags);
7365 update_rq_clock(rq);
7367 cfs_rq = task_cfs_rq(current);
7368 curr = cfs_rq->curr;
7371 * Not only the cpu but also the task_group of the parent might have
7372 * been changed after parent->se.parent,cfs_rq were copied to
7373 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7374 * of child point to valid ones.
7377 __set_task_cpu(p, this_cpu);
7380 update_curr(cfs_rq);
7383 se->vruntime = curr->vruntime;
7384 place_entity(cfs_rq, se, 1);
7386 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7388 * Upon rescheduling, sched_class::put_prev_task() will place
7389 * 'current' within the tree based on its new key value.
7391 swap(curr->vruntime, se->vruntime);
7392 resched_task(rq->curr);
7395 se->vruntime -= cfs_rq->min_vruntime;
7397 raw_spin_unlock_irqrestore(&rq->lock, flags);
7401 * Priority of the task has changed. Check to see if we preempt
7405 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7411 * Reschedule if we are currently running on this runqueue and
7412 * our priority decreased, or if we are not currently running on
7413 * this runqueue and our priority is higher than the current's
7415 if (rq->curr == p) {
7416 if (p->prio > oldprio)
7417 resched_task(rq->curr);
7419 check_preempt_curr(rq, p, 0);
7422 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7424 struct sched_entity *se = &p->se;
7425 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7428 * Ensure the task's vruntime is normalized, so that when it's
7429 * switched back to the fair class the enqueue_entity(.flags=0) will
7430 * do the right thing.
7432 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7433 * have normalized the vruntime, if it's !on_rq, then only when
7434 * the task is sleeping will it still have non-normalized vruntime.
7436 if (!p->on_rq && p->state != TASK_RUNNING) {
7438 * Fix up our vruntime so that the current sleep doesn't
7439 * cause 'unlimited' sleep bonus.
7441 place_entity(cfs_rq, se, 0);
7442 se->vruntime -= cfs_rq->min_vruntime;
7447 * Remove our load from contribution when we leave sched_fair
7448 * and ensure we don't carry in an old decay_count if we
7451 if (se->avg.decay_count) {
7452 __synchronize_entity_decay(se);
7453 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7459 * We switched to the sched_fair class.
7461 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7463 struct sched_entity *se = &p->se;
7464 #ifdef CONFIG_FAIR_GROUP_SCHED
7466 * Since the real-depth could have been changed (only FAIR
7467 * class maintain depth value), reset depth properly.
7469 se->depth = se->parent ? se->parent->depth + 1 : 0;
7475 * We were most likely switched from sched_rt, so
7476 * kick off the schedule if running, otherwise just see
7477 * if we can still preempt the current task.
7480 resched_task(rq->curr);
7482 check_preempt_curr(rq, p, 0);
7485 /* Account for a task changing its policy or group.
7487 * This routine is mostly called to set cfs_rq->curr field when a task
7488 * migrates between groups/classes.
7490 static void set_curr_task_fair(struct rq *rq)
7492 struct sched_entity *se = &rq->curr->se;
7494 for_each_sched_entity(se) {
7495 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7497 set_next_entity(cfs_rq, se);
7498 /* ensure bandwidth has been allocated on our new cfs_rq */
7499 account_cfs_rq_runtime(cfs_rq, 0);
7503 void init_cfs_rq(struct cfs_rq *cfs_rq)
7505 cfs_rq->tasks_timeline = RB_ROOT;
7506 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7507 #ifndef CONFIG_64BIT
7508 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7511 atomic64_set(&cfs_rq->decay_counter, 1);
7512 atomic_long_set(&cfs_rq->removed_load, 0);
7516 #ifdef CONFIG_FAIR_GROUP_SCHED
7517 static void task_move_group_fair(struct task_struct *p, int on_rq)
7519 struct sched_entity *se = &p->se;
7520 struct cfs_rq *cfs_rq;
7523 * If the task was not on the rq at the time of this cgroup movement
7524 * it must have been asleep, sleeping tasks keep their ->vruntime
7525 * absolute on their old rq until wakeup (needed for the fair sleeper
7526 * bonus in place_entity()).
7528 * If it was on the rq, we've just 'preempted' it, which does convert
7529 * ->vruntime to a relative base.
7531 * Make sure both cases convert their relative position when migrating
7532 * to another cgroup's rq. This does somewhat interfere with the
7533 * fair sleeper stuff for the first placement, but who cares.
7536 * When !on_rq, vruntime of the task has usually NOT been normalized.
7537 * But there are some cases where it has already been normalized:
7539 * - Moving a forked child which is waiting for being woken up by
7540 * wake_up_new_task().
7541 * - Moving a task which has been woken up by try_to_wake_up() and
7542 * waiting for actually being woken up by sched_ttwu_pending().
7544 * To prevent boost or penalty in the new cfs_rq caused by delta
7545 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7547 if (!on_rq && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7551 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7552 set_task_rq(p, task_cpu(p));
7553 se->depth = se->parent ? se->parent->depth + 1 : 0;
7555 cfs_rq = cfs_rq_of(se);
7556 se->vruntime += cfs_rq->min_vruntime;
7559 * migrate_task_rq_fair() will have removed our previous
7560 * contribution, but we must synchronize for ongoing future
7563 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7564 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7569 void free_fair_sched_group(struct task_group *tg)
7573 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7575 for_each_possible_cpu(i) {
7577 kfree(tg->cfs_rq[i]);
7586 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7588 struct cfs_rq *cfs_rq;
7589 struct sched_entity *se;
7592 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7595 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7599 tg->shares = NICE_0_LOAD;
7601 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7603 for_each_possible_cpu(i) {
7604 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7605 GFP_KERNEL, cpu_to_node(i));
7609 se = kzalloc_node(sizeof(struct sched_entity),
7610 GFP_KERNEL, cpu_to_node(i));
7614 init_cfs_rq(cfs_rq);
7615 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7626 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7628 struct rq *rq = cpu_rq(cpu);
7629 unsigned long flags;
7632 * Only empty task groups can be destroyed; so we can speculatively
7633 * check on_list without danger of it being re-added.
7635 if (!tg->cfs_rq[cpu]->on_list)
7638 raw_spin_lock_irqsave(&rq->lock, flags);
7639 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7640 raw_spin_unlock_irqrestore(&rq->lock, flags);
7643 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7644 struct sched_entity *se, int cpu,
7645 struct sched_entity *parent)
7647 struct rq *rq = cpu_rq(cpu);
7651 init_cfs_rq_runtime(cfs_rq);
7653 tg->cfs_rq[cpu] = cfs_rq;
7656 /* se could be NULL for root_task_group */
7661 se->cfs_rq = &rq->cfs;
7664 se->cfs_rq = parent->my_q;
7665 se->depth = parent->depth + 1;
7669 /* guarantee group entities always have weight */
7670 update_load_set(&se->load, NICE_0_LOAD);
7671 se->parent = parent;
7674 static DEFINE_MUTEX(shares_mutex);
7676 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7679 unsigned long flags;
7682 * We can't change the weight of the root cgroup.
7687 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7689 mutex_lock(&shares_mutex);
7690 if (tg->shares == shares)
7693 tg->shares = shares;
7694 for_each_possible_cpu(i) {
7695 struct rq *rq = cpu_rq(i);
7696 struct sched_entity *se;
7699 /* Propagate contribution to hierarchy */
7700 raw_spin_lock_irqsave(&rq->lock, flags);
7702 /* Possible calls to update_curr() need rq clock */
7703 update_rq_clock(rq);
7704 for_each_sched_entity(se)
7705 update_cfs_shares(group_cfs_rq(se));
7706 raw_spin_unlock_irqrestore(&rq->lock, flags);
7710 mutex_unlock(&shares_mutex);
7713 #else /* CONFIG_FAIR_GROUP_SCHED */
7715 void free_fair_sched_group(struct task_group *tg) { }
7717 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7722 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7724 #endif /* CONFIG_FAIR_GROUP_SCHED */
7727 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7729 struct sched_entity *se = &task->se;
7730 unsigned int rr_interval = 0;
7733 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7736 if (rq->cfs.load.weight)
7737 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7743 * All the scheduling class methods:
7745 const struct sched_class fair_sched_class = {
7746 .next = &idle_sched_class,
7747 .enqueue_task = enqueue_task_fair,
7748 .dequeue_task = dequeue_task_fair,
7749 .yield_task = yield_task_fair,
7750 .yield_to_task = yield_to_task_fair,
7752 .check_preempt_curr = check_preempt_wakeup,
7754 .pick_next_task = pick_next_task_fair,
7755 .put_prev_task = put_prev_task_fair,
7758 .select_task_rq = select_task_rq_fair,
7759 .migrate_task_rq = migrate_task_rq_fair,
7761 .rq_online = rq_online_fair,
7762 .rq_offline = rq_offline_fair,
7764 .task_waking = task_waking_fair,
7767 .set_curr_task = set_curr_task_fair,
7768 .task_tick = task_tick_fair,
7769 .task_fork = task_fork_fair,
7771 .prio_changed = prio_changed_fair,
7772 .switched_from = switched_from_fair,
7773 .switched_to = switched_to_fair,
7775 .get_rr_interval = get_rr_interval_fair,
7777 #ifdef CONFIG_FAIR_GROUP_SCHED
7778 .task_move_group = task_move_group_fair,
7782 #ifdef CONFIG_SCHED_DEBUG
7783 void print_cfs_stats(struct seq_file *m, int cpu)
7785 struct cfs_rq *cfs_rq;
7788 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7789 print_cfs_rq(m, cpu, cfs_rq);
7794 __init void init_sched_fair_class(void)
7797 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7799 #ifdef CONFIG_NO_HZ_COMMON
7800 nohz.next_balance = jiffies;
7801 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7802 cpu_notifier(sched_ilb_notifier, 0);