2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #define WMULT_CONST (~0U)
182 #define WMULT_SHIFT 32
184 static void __update_inv_weight(struct load_weight *lw)
188 if (likely(lw->inv_weight))
191 w = scale_load_down(lw->weight);
193 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
195 else if (unlikely(!w))
196 lw->inv_weight = WMULT_CONST;
198 lw->inv_weight = WMULT_CONST / w;
202 * delta_exec * weight / lw.weight
204 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
207 * we're guaranteed shift stays positive because inv_weight is guaranteed to
208 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
211 * weight/lw.weight <= 1, and therefore our shift will also be positive.
213 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
215 u64 fact = scale_load_down(weight);
216 int shift = WMULT_SHIFT;
218 __update_inv_weight(lw);
220 if (unlikely(fact >> 32)) {
227 /* hint to use a 32x32->64 mul */
228 fact = (u64)(u32)fact * lw->inv_weight;
235 return mul_u64_u32_shr(delta_exec, fact, shift);
239 const struct sched_class fair_sched_class;
241 /**************************************************************
242 * CFS operations on generic schedulable entities:
245 #ifdef CONFIG_FAIR_GROUP_SCHED
247 /* cpu runqueue to which this cfs_rq is attached */
248 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
253 /* An entity is a task if it doesn't "own" a runqueue */
254 #define entity_is_task(se) (!se->my_q)
256 static inline struct task_struct *task_of(struct sched_entity *se)
258 #ifdef CONFIG_SCHED_DEBUG
259 WARN_ON_ONCE(!entity_is_task(se));
261 return container_of(se, struct task_struct, se);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
285 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 if (!cfs_rq->on_list) {
292 * Ensure we either appear before our parent (if already
293 * enqueued) or force our parent to appear after us when it is
294 * enqueued. The fact that we always enqueue bottom-up
295 * reduces this to two cases.
297 if (cfs_rq->tg->parent &&
298 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
299 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
300 &rq_of(cfs_rq)->leaf_cfs_rq_list);
302 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
303 &rq_of(cfs_rq)->leaf_cfs_rq_list);
307 /* We should have no load, but we need to update last_decay. */
308 update_cfs_rq_blocked_load(cfs_rq, 0);
312 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
314 if (cfs_rq->on_list) {
315 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
320 /* Iterate thr' all leaf cfs_rq's on a runqueue */
321 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
322 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
324 /* Do the two (enqueued) entities belong to the same group ? */
325 static inline struct cfs_rq *
326 is_same_group(struct sched_entity *se, struct sched_entity *pse)
328 if (se->cfs_rq == pse->cfs_rq)
334 static inline struct sched_entity *parent_entity(struct sched_entity *se)
340 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
342 int se_depth, pse_depth;
345 * preemption test can be made between sibling entities who are in the
346 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
347 * both tasks until we find their ancestors who are siblings of common
351 /* First walk up until both entities are at same depth */
352 se_depth = (*se)->depth;
353 pse_depth = (*pse)->depth;
355 while (se_depth > pse_depth) {
357 *se = parent_entity(*se);
360 while (pse_depth > se_depth) {
362 *pse = parent_entity(*pse);
365 while (!is_same_group(*se, *pse)) {
366 *se = parent_entity(*se);
367 *pse = parent_entity(*pse);
371 #else /* !CONFIG_FAIR_GROUP_SCHED */
373 static inline struct task_struct *task_of(struct sched_entity *se)
375 return container_of(se, struct task_struct, se);
378 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
380 return container_of(cfs_rq, struct rq, cfs);
383 #define entity_is_task(se) 1
385 #define for_each_sched_entity(se) \
386 for (; se; se = NULL)
388 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
390 return &task_rq(p)->cfs;
393 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
395 struct task_struct *p = task_of(se);
396 struct rq *rq = task_rq(p);
401 /* runqueue "owned" by this group */
402 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
407 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
415 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
416 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
418 static inline struct sched_entity *parent_entity(struct sched_entity *se)
424 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
428 #endif /* CONFIG_FAIR_GROUP_SCHED */
430 static __always_inline
431 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
433 /**************************************************************
434 * Scheduling class tree data structure manipulation methods:
437 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
439 s64 delta = (s64)(vruntime - max_vruntime);
441 max_vruntime = vruntime;
446 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
448 s64 delta = (s64)(vruntime - min_vruntime);
450 min_vruntime = vruntime;
455 static inline int entity_before(struct sched_entity *a,
456 struct sched_entity *b)
458 return (s64)(a->vruntime - b->vruntime) < 0;
461 static void update_min_vruntime(struct cfs_rq *cfs_rq)
463 u64 vruntime = cfs_rq->min_vruntime;
466 vruntime = cfs_rq->curr->vruntime;
468 if (cfs_rq->rb_leftmost) {
469 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
474 vruntime = se->vruntime;
476 vruntime = min_vruntime(vruntime, se->vruntime);
479 /* ensure we never gain time by being placed backwards. */
480 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
483 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
488 * Enqueue an entity into the rb-tree:
490 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
492 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
493 struct rb_node *parent = NULL;
494 struct sched_entity *entry;
498 * Find the right place in the rbtree:
502 entry = rb_entry(parent, struct sched_entity, run_node);
504 * We dont care about collisions. Nodes with
505 * the same key stay together.
507 if (entity_before(se, entry)) {
508 link = &parent->rb_left;
510 link = &parent->rb_right;
516 * Maintain a cache of leftmost tree entries (it is frequently
520 cfs_rq->rb_leftmost = &se->run_node;
522 rb_link_node(&se->run_node, parent, link);
523 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
526 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
528 if (cfs_rq->rb_leftmost == &se->run_node) {
529 struct rb_node *next_node;
531 next_node = rb_next(&se->run_node);
532 cfs_rq->rb_leftmost = next_node;
535 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
538 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
540 struct rb_node *left = cfs_rq->rb_leftmost;
545 return rb_entry(left, struct sched_entity, run_node);
548 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
550 struct rb_node *next = rb_next(&se->run_node);
555 return rb_entry(next, struct sched_entity, run_node);
558 #ifdef CONFIG_SCHED_DEBUG
559 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
561 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
566 return rb_entry(last, struct sched_entity, run_node);
569 /**************************************************************
570 * Scheduling class statistics methods:
573 int sched_proc_update_handler(struct ctl_table *table, int write,
574 void __user *buffer, size_t *lenp,
577 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
578 int factor = get_update_sysctl_factor();
583 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
584 sysctl_sched_min_granularity);
586 #define WRT_SYSCTL(name) \
587 (normalized_sysctl_##name = sysctl_##name / (factor))
588 WRT_SYSCTL(sched_min_granularity);
589 WRT_SYSCTL(sched_latency);
590 WRT_SYSCTL(sched_wakeup_granularity);
600 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
602 if (unlikely(se->load.weight != NICE_0_LOAD))
603 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
609 * The idea is to set a period in which each task runs once.
611 * When there are too many tasks (sched_nr_latency) we have to stretch
612 * this period because otherwise the slices get too small.
614 * p = (nr <= nl) ? l : l*nr/nl
616 static u64 __sched_period(unsigned long nr_running)
618 u64 period = sysctl_sched_latency;
619 unsigned long nr_latency = sched_nr_latency;
621 if (unlikely(nr_running > nr_latency)) {
622 period = sysctl_sched_min_granularity;
623 period *= nr_running;
630 * We calculate the wall-time slice from the period by taking a part
631 * proportional to the weight.
635 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
637 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
639 for_each_sched_entity(se) {
640 struct load_weight *load;
641 struct load_weight lw;
643 cfs_rq = cfs_rq_of(se);
644 load = &cfs_rq->load;
646 if (unlikely(!se->on_rq)) {
649 update_load_add(&lw, se->load.weight);
652 slice = __calc_delta(slice, se->load.weight, load);
658 * We calculate the vruntime slice of a to-be-inserted task.
662 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
664 return calc_delta_fair(sched_slice(cfs_rq, se), se);
668 static unsigned long task_h_load(struct task_struct *p);
670 static inline void __update_task_entity_contrib(struct sched_entity *se);
672 /* Give new task start runnable values to heavy its load in infant time */
673 void init_task_runnable_average(struct task_struct *p)
677 p->se.avg.decay_count = 0;
678 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
679 p->se.avg.runnable_avg_sum = slice;
680 p->se.avg.runnable_avg_period = slice;
681 __update_task_entity_contrib(&p->se);
684 void init_task_runnable_average(struct task_struct *p)
690 * Update the current task's runtime statistics.
692 static void update_curr(struct cfs_rq *cfs_rq)
694 struct sched_entity *curr = cfs_rq->curr;
695 u64 now = rq_clock_task(rq_of(cfs_rq));
701 delta_exec = now - curr->exec_start;
702 if (unlikely((s64)delta_exec <= 0))
705 curr->exec_start = now;
707 schedstat_set(curr->statistics.exec_max,
708 max(delta_exec, curr->statistics.exec_max));
710 curr->sum_exec_runtime += delta_exec;
711 schedstat_add(cfs_rq, exec_clock, delta_exec);
713 curr->vruntime += calc_delta_fair(delta_exec, curr);
714 update_min_vruntime(cfs_rq);
716 if (entity_is_task(curr)) {
717 struct task_struct *curtask = task_of(curr);
719 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
720 cpuacct_charge(curtask, delta_exec);
721 account_group_exec_runtime(curtask, delta_exec);
724 account_cfs_rq_runtime(cfs_rq, delta_exec);
728 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
730 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
734 * Task is being enqueued - update stats:
736 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
739 * Are we enqueueing a waiting task? (for current tasks
740 * a dequeue/enqueue event is a NOP)
742 if (se != cfs_rq->curr)
743 update_stats_wait_start(cfs_rq, se);
747 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
749 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
750 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
751 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
752 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
753 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
754 #ifdef CONFIG_SCHEDSTATS
755 if (entity_is_task(se)) {
756 trace_sched_stat_wait(task_of(se),
757 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
760 schedstat_set(se->statistics.wait_start, 0);
764 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
767 * Mark the end of the wait period if dequeueing a
770 if (se != cfs_rq->curr)
771 update_stats_wait_end(cfs_rq, se);
775 * We are picking a new current task - update its stats:
778 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
781 * We are starting a new run period:
783 se->exec_start = rq_clock_task(rq_of(cfs_rq));
786 /**************************************************
787 * Scheduling class queueing methods:
790 #ifdef CONFIG_NUMA_BALANCING
792 * Approximate time to scan a full NUMA task in ms. The task scan period is
793 * calculated based on the tasks virtual memory size and
794 * numa_balancing_scan_size.
796 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
797 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
799 /* Portion of address space to scan in MB */
800 unsigned int sysctl_numa_balancing_scan_size = 256;
802 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
803 unsigned int sysctl_numa_balancing_scan_delay = 1000;
805 static unsigned int task_nr_scan_windows(struct task_struct *p)
807 unsigned long rss = 0;
808 unsigned long nr_scan_pages;
811 * Calculations based on RSS as non-present and empty pages are skipped
812 * by the PTE scanner and NUMA hinting faults should be trapped based
815 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
816 rss = get_mm_rss(p->mm);
820 rss = round_up(rss, nr_scan_pages);
821 return rss / nr_scan_pages;
824 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
825 #define MAX_SCAN_WINDOW 2560
827 static unsigned int task_scan_min(struct task_struct *p)
829 unsigned int scan, floor;
830 unsigned int windows = 1;
832 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
833 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
834 floor = 1000 / windows;
836 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
837 return max_t(unsigned int, floor, scan);
840 static unsigned int task_scan_max(struct task_struct *p)
842 unsigned int smin = task_scan_min(p);
845 /* Watch for min being lower than max due to floor calculations */
846 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
847 return max(smin, smax);
850 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
852 rq->nr_numa_running += (p->numa_preferred_nid != -1);
853 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
856 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
858 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
859 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
865 spinlock_t lock; /* nr_tasks, tasks */
868 struct list_head task_list;
871 nodemask_t active_nodes;
872 unsigned long total_faults;
874 * Faults_cpu is used to decide whether memory should move
875 * towards the CPU. As a consequence, these stats are weighted
876 * more by CPU use than by memory faults.
878 unsigned long *faults_cpu;
879 unsigned long faults[0];
882 /* Shared or private faults. */
883 #define NR_NUMA_HINT_FAULT_TYPES 2
885 /* Memory and CPU locality */
886 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
888 /* Averaged statistics, and temporary buffers. */
889 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
891 pid_t task_numa_group_id(struct task_struct *p)
893 return p->numa_group ? p->numa_group->gid : 0;
896 static inline int task_faults_idx(int nid, int priv)
898 return NR_NUMA_HINT_FAULT_TYPES * nid + priv;
901 static inline unsigned long task_faults(struct task_struct *p, int nid)
903 if (!p->numa_faults_memory)
906 return p->numa_faults_memory[task_faults_idx(nid, 0)] +
907 p->numa_faults_memory[task_faults_idx(nid, 1)];
910 static inline unsigned long group_faults(struct task_struct *p, int nid)
915 return p->numa_group->faults[task_faults_idx(nid, 0)] +
916 p->numa_group->faults[task_faults_idx(nid, 1)];
919 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
921 return group->faults_cpu[task_faults_idx(nid, 0)] +
922 group->faults_cpu[task_faults_idx(nid, 1)];
926 * These return the fraction of accesses done by a particular task, or
927 * task group, on a particular numa node. The group weight is given a
928 * larger multiplier, in order to group tasks together that are almost
929 * evenly spread out between numa nodes.
931 static inline unsigned long task_weight(struct task_struct *p, int nid)
933 unsigned long total_faults;
935 if (!p->numa_faults_memory)
938 total_faults = p->total_numa_faults;
943 return 1000 * task_faults(p, nid) / total_faults;
946 static inline unsigned long group_weight(struct task_struct *p, int nid)
948 if (!p->numa_group || !p->numa_group->total_faults)
951 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
954 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
955 int src_nid, int dst_cpu)
957 struct numa_group *ng = p->numa_group;
958 int dst_nid = cpu_to_node(dst_cpu);
959 int last_cpupid, this_cpupid;
961 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
964 * Multi-stage node selection is used in conjunction with a periodic
965 * migration fault to build a temporal task<->page relation. By using
966 * a two-stage filter we remove short/unlikely relations.
968 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
969 * a task's usage of a particular page (n_p) per total usage of this
970 * page (n_t) (in a given time-span) to a probability.
972 * Our periodic faults will sample this probability and getting the
973 * same result twice in a row, given these samples are fully
974 * independent, is then given by P(n)^2, provided our sample period
975 * is sufficiently short compared to the usage pattern.
977 * This quadric squishes small probabilities, making it less likely we
978 * act on an unlikely task<->page relation.
980 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
981 if (!cpupid_pid_unset(last_cpupid) &&
982 cpupid_to_nid(last_cpupid) != dst_nid)
985 /* Always allow migrate on private faults */
986 if (cpupid_match_pid(p, last_cpupid))
989 /* A shared fault, but p->numa_group has not been set up yet. */
994 * Do not migrate if the destination is not a node that
995 * is actively used by this numa group.
997 if (!node_isset(dst_nid, ng->active_nodes))
1001 * Source is a node that is not actively used by this
1002 * numa group, while the destination is. Migrate.
1004 if (!node_isset(src_nid, ng->active_nodes))
1008 * Both source and destination are nodes in active
1009 * use by this numa group. Maximize memory bandwidth
1010 * by migrating from more heavily used groups, to less
1011 * heavily used ones, spreading the load around.
1012 * Use a 1/4 hysteresis to avoid spurious page movement.
1014 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1017 static unsigned long weighted_cpuload(const int cpu);
1018 static unsigned long source_load(int cpu, int type);
1019 static unsigned long target_load(int cpu, int type);
1020 static unsigned long capacity_of(int cpu);
1021 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1023 /* Cached statistics for all CPUs within a node */
1025 unsigned long nr_running;
1028 /* Total compute capacity of CPUs on a node */
1029 unsigned long compute_capacity;
1031 /* Approximate capacity in terms of runnable tasks on a node */
1032 unsigned long task_capacity;
1033 int has_free_capacity;
1037 * XXX borrowed from update_sg_lb_stats
1039 static void update_numa_stats(struct numa_stats *ns, int nid)
1041 int smt, cpu, cpus = 0;
1042 unsigned long capacity;
1044 memset(ns, 0, sizeof(*ns));
1045 for_each_cpu(cpu, cpumask_of_node(nid)) {
1046 struct rq *rq = cpu_rq(cpu);
1048 ns->nr_running += rq->nr_running;
1049 ns->load += weighted_cpuload(cpu);
1050 ns->compute_capacity += capacity_of(cpu);
1056 * If we raced with hotplug and there are no CPUs left in our mask
1057 * the @ns structure is NULL'ed and task_numa_compare() will
1058 * not find this node attractive.
1060 * We'll either bail at !has_free_capacity, or we'll detect a huge
1061 * imbalance and bail there.
1066 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1067 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1068 capacity = cpus / smt; /* cores */
1070 ns->task_capacity = min_t(unsigned, capacity,
1071 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1072 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1075 struct task_numa_env {
1076 struct task_struct *p;
1078 int src_cpu, src_nid;
1079 int dst_cpu, dst_nid;
1081 struct numa_stats src_stats, dst_stats;
1085 struct task_struct *best_task;
1090 static void task_numa_assign(struct task_numa_env *env,
1091 struct task_struct *p, long imp)
1094 put_task_struct(env->best_task);
1099 env->best_imp = imp;
1100 env->best_cpu = env->dst_cpu;
1103 static bool load_too_imbalanced(long src_load, long dst_load,
1104 struct task_numa_env *env)
1107 long orig_src_load, orig_dst_load;
1108 long src_capacity, dst_capacity;
1111 * The load is corrected for the CPU capacity available on each node.
1114 * ------------ vs ---------
1115 * src_capacity dst_capacity
1117 src_capacity = env->src_stats.compute_capacity;
1118 dst_capacity = env->dst_stats.compute_capacity;
1120 /* We care about the slope of the imbalance, not the direction. */
1121 if (dst_load < src_load)
1122 swap(dst_load, src_load);
1124 /* Is the difference below the threshold? */
1125 imb = dst_load * src_capacity * 100 -
1126 src_load * dst_capacity * env->imbalance_pct;
1131 * The imbalance is above the allowed threshold.
1132 * Compare it with the old imbalance.
1134 orig_src_load = env->src_stats.load;
1135 orig_dst_load = env->dst_stats.load;
1137 if (orig_dst_load < orig_src_load)
1138 swap(orig_dst_load, orig_src_load);
1140 old_imb = orig_dst_load * src_capacity * 100 -
1141 orig_src_load * dst_capacity * env->imbalance_pct;
1143 /* Would this change make things worse? */
1144 return (imb > old_imb);
1148 * This checks if the overall compute and NUMA accesses of the system would
1149 * be improved if the source tasks was migrated to the target dst_cpu taking
1150 * into account that it might be best if task running on the dst_cpu should
1151 * be exchanged with the source task
1153 static void task_numa_compare(struct task_numa_env *env,
1154 long taskimp, long groupimp)
1156 struct rq *src_rq = cpu_rq(env->src_cpu);
1157 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1158 struct task_struct *cur;
1159 long src_load, dst_load;
1161 long imp = env->p->numa_group ? groupimp : taskimp;
1165 cur = ACCESS_ONCE(dst_rq->curr);
1166 if (cur->pid == 0) /* idle */
1170 * "imp" is the fault differential for the source task between the
1171 * source and destination node. Calculate the total differential for
1172 * the source task and potential destination task. The more negative
1173 * the value is, the more rmeote accesses that would be expected to
1174 * be incurred if the tasks were swapped.
1177 /* Skip this swap candidate if cannot move to the source cpu */
1178 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1182 * If dst and source tasks are in the same NUMA group, or not
1183 * in any group then look only at task weights.
1185 if (cur->numa_group == env->p->numa_group) {
1186 imp = taskimp + task_weight(cur, env->src_nid) -
1187 task_weight(cur, env->dst_nid);
1189 * Add some hysteresis to prevent swapping the
1190 * tasks within a group over tiny differences.
1192 if (cur->numa_group)
1196 * Compare the group weights. If a task is all by
1197 * itself (not part of a group), use the task weight
1200 if (cur->numa_group)
1201 imp += group_weight(cur, env->src_nid) -
1202 group_weight(cur, env->dst_nid);
1204 imp += task_weight(cur, env->src_nid) -
1205 task_weight(cur, env->dst_nid);
1209 if (imp <= env->best_imp && moveimp <= env->best_imp)
1213 /* Is there capacity at our destination? */
1214 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1215 !env->dst_stats.has_free_capacity)
1221 /* Balance doesn't matter much if we're running a task per cpu */
1222 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1223 dst_rq->nr_running == 1)
1227 * In the overloaded case, try and keep the load balanced.
1230 load = task_h_load(env->p);
1231 dst_load = env->dst_stats.load + load;
1232 src_load = env->src_stats.load - load;
1234 if (moveimp > imp && moveimp > env->best_imp) {
1236 * If the improvement from just moving env->p direction is
1237 * better than swapping tasks around, check if a move is
1238 * possible. Store a slightly smaller score than moveimp,
1239 * so an actually idle CPU will win.
1241 if (!load_too_imbalanced(src_load, dst_load, env)) {
1248 if (imp <= env->best_imp)
1252 load = task_h_load(cur);
1257 if (load_too_imbalanced(src_load, dst_load, env))
1261 task_numa_assign(env, cur, imp);
1266 static void task_numa_find_cpu(struct task_numa_env *env,
1267 long taskimp, long groupimp)
1271 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1272 /* Skip this CPU if the source task cannot migrate */
1273 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1277 task_numa_compare(env, taskimp, groupimp);
1281 static int task_numa_migrate(struct task_struct *p)
1283 struct task_numa_env env = {
1286 .src_cpu = task_cpu(p),
1287 .src_nid = task_node(p),
1289 .imbalance_pct = 112,
1295 struct sched_domain *sd;
1296 unsigned long taskweight, groupweight;
1298 long taskimp, groupimp;
1301 * Pick the lowest SD_NUMA domain, as that would have the smallest
1302 * imbalance and would be the first to start moving tasks about.
1304 * And we want to avoid any moving of tasks about, as that would create
1305 * random movement of tasks -- counter the numa conditions we're trying
1309 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1311 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1315 * Cpusets can break the scheduler domain tree into smaller
1316 * balance domains, some of which do not cross NUMA boundaries.
1317 * Tasks that are "trapped" in such domains cannot be migrated
1318 * elsewhere, so there is no point in (re)trying.
1320 if (unlikely(!sd)) {
1321 p->numa_preferred_nid = task_node(p);
1325 taskweight = task_weight(p, env.src_nid);
1326 groupweight = group_weight(p, env.src_nid);
1327 update_numa_stats(&env.src_stats, env.src_nid);
1328 env.dst_nid = p->numa_preferred_nid;
1329 taskimp = task_weight(p, env.dst_nid) - taskweight;
1330 groupimp = group_weight(p, env.dst_nid) - groupweight;
1331 update_numa_stats(&env.dst_stats, env.dst_nid);
1333 /* Try to find a spot on the preferred nid. */
1334 task_numa_find_cpu(&env, taskimp, groupimp);
1336 /* No space available on the preferred nid. Look elsewhere. */
1337 if (env.best_cpu == -1) {
1338 for_each_online_node(nid) {
1339 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1342 /* Only consider nodes where both task and groups benefit */
1343 taskimp = task_weight(p, nid) - taskweight;
1344 groupimp = group_weight(p, nid) - groupweight;
1345 if (taskimp < 0 && groupimp < 0)
1349 update_numa_stats(&env.dst_stats, env.dst_nid);
1350 task_numa_find_cpu(&env, taskimp, groupimp);
1355 * If the task is part of a workload that spans multiple NUMA nodes,
1356 * and is migrating into one of the workload's active nodes, remember
1357 * this node as the task's preferred numa node, so the workload can
1359 * A task that migrated to a second choice node will be better off
1360 * trying for a better one later. Do not set the preferred node here.
1362 if (p->numa_group) {
1363 if (env.best_cpu == -1)
1368 if (node_isset(nid, p->numa_group->active_nodes))
1369 sched_setnuma(p, env.dst_nid);
1372 /* No better CPU than the current one was found. */
1373 if (env.best_cpu == -1)
1377 * Reset the scan period if the task is being rescheduled on an
1378 * alternative node to recheck if the tasks is now properly placed.
1380 p->numa_scan_period = task_scan_min(p);
1382 if (env.best_task == NULL) {
1383 ret = migrate_task_to(p, env.best_cpu);
1385 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1389 ret = migrate_swap(p, env.best_task);
1391 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1392 put_task_struct(env.best_task);
1396 /* Attempt to migrate a task to a CPU on the preferred node. */
1397 static void numa_migrate_preferred(struct task_struct *p)
1399 unsigned long interval = HZ;
1401 /* This task has no NUMA fault statistics yet */
1402 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1405 /* Periodically retry migrating the task to the preferred node */
1406 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1407 p->numa_migrate_retry = jiffies + interval;
1409 /* Success if task is already running on preferred CPU */
1410 if (task_node(p) == p->numa_preferred_nid)
1413 /* Otherwise, try migrate to a CPU on the preferred node */
1414 task_numa_migrate(p);
1418 * Find the nodes on which the workload is actively running. We do this by
1419 * tracking the nodes from which NUMA hinting faults are triggered. This can
1420 * be different from the set of nodes where the workload's memory is currently
1423 * The bitmask is used to make smarter decisions on when to do NUMA page
1424 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1425 * are added when they cause over 6/16 of the maximum number of faults, but
1426 * only removed when they drop below 3/16.
1428 static void update_numa_active_node_mask(struct numa_group *numa_group)
1430 unsigned long faults, max_faults = 0;
1433 for_each_online_node(nid) {
1434 faults = group_faults_cpu(numa_group, nid);
1435 if (faults > max_faults)
1436 max_faults = faults;
1439 for_each_online_node(nid) {
1440 faults = group_faults_cpu(numa_group, nid);
1441 if (!node_isset(nid, numa_group->active_nodes)) {
1442 if (faults > max_faults * 6 / 16)
1443 node_set(nid, numa_group->active_nodes);
1444 } else if (faults < max_faults * 3 / 16)
1445 node_clear(nid, numa_group->active_nodes);
1450 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1451 * increments. The more local the fault statistics are, the higher the scan
1452 * period will be for the next scan window. If local/(local+remote) ratio is
1453 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1454 * the scan period will decrease. Aim for 70% local accesses.
1456 #define NUMA_PERIOD_SLOTS 10
1457 #define NUMA_PERIOD_THRESHOLD 7
1460 * Increase the scan period (slow down scanning) if the majority of
1461 * our memory is already on our local node, or if the majority of
1462 * the page accesses are shared with other processes.
1463 * Otherwise, decrease the scan period.
1465 static void update_task_scan_period(struct task_struct *p,
1466 unsigned long shared, unsigned long private)
1468 unsigned int period_slot;
1472 unsigned long remote = p->numa_faults_locality[0];
1473 unsigned long local = p->numa_faults_locality[1];
1476 * If there were no record hinting faults then either the task is
1477 * completely idle or all activity is areas that are not of interest
1478 * to automatic numa balancing. Scan slower
1480 if (local + shared == 0) {
1481 p->numa_scan_period = min(p->numa_scan_period_max,
1482 p->numa_scan_period << 1);
1484 p->mm->numa_next_scan = jiffies +
1485 msecs_to_jiffies(p->numa_scan_period);
1491 * Prepare to scale scan period relative to the current period.
1492 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1493 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1494 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1496 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1497 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1498 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1499 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1502 diff = slot * period_slot;
1504 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1507 * Scale scan rate increases based on sharing. There is an
1508 * inverse relationship between the degree of sharing and
1509 * the adjustment made to the scanning period. Broadly
1510 * speaking the intent is that there is little point
1511 * scanning faster if shared accesses dominate as it may
1512 * simply bounce migrations uselessly
1514 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1515 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1518 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1519 task_scan_min(p), task_scan_max(p));
1520 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1524 * Get the fraction of time the task has been running since the last
1525 * NUMA placement cycle. The scheduler keeps similar statistics, but
1526 * decays those on a 32ms period, which is orders of magnitude off
1527 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1528 * stats only if the task is so new there are no NUMA statistics yet.
1530 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1532 u64 runtime, delta, now;
1533 /* Use the start of this time slice to avoid calculations. */
1534 now = p->se.exec_start;
1535 runtime = p->se.sum_exec_runtime;
1537 if (p->last_task_numa_placement) {
1538 delta = runtime - p->last_sum_exec_runtime;
1539 *period = now - p->last_task_numa_placement;
1541 delta = p->se.avg.runnable_avg_sum;
1542 *period = p->se.avg.runnable_avg_period;
1545 p->last_sum_exec_runtime = runtime;
1546 p->last_task_numa_placement = now;
1551 static void task_numa_placement(struct task_struct *p)
1553 int seq, nid, max_nid = -1, max_group_nid = -1;
1554 unsigned long max_faults = 0, max_group_faults = 0;
1555 unsigned long fault_types[2] = { 0, 0 };
1556 unsigned long total_faults;
1557 u64 runtime, period;
1558 spinlock_t *group_lock = NULL;
1560 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1561 if (p->numa_scan_seq == seq)
1563 p->numa_scan_seq = seq;
1564 p->numa_scan_period_max = task_scan_max(p);
1566 total_faults = p->numa_faults_locality[0] +
1567 p->numa_faults_locality[1];
1568 runtime = numa_get_avg_runtime(p, &period);
1570 /* If the task is part of a group prevent parallel updates to group stats */
1571 if (p->numa_group) {
1572 group_lock = &p->numa_group->lock;
1573 spin_lock_irq(group_lock);
1576 /* Find the node with the highest number of faults */
1577 for_each_online_node(nid) {
1578 unsigned long faults = 0, group_faults = 0;
1581 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1582 long diff, f_diff, f_weight;
1584 i = task_faults_idx(nid, priv);
1586 /* Decay existing window, copy faults since last scan */
1587 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1588 fault_types[priv] += p->numa_faults_buffer_memory[i];
1589 p->numa_faults_buffer_memory[i] = 0;
1592 * Normalize the faults_from, so all tasks in a group
1593 * count according to CPU use, instead of by the raw
1594 * number of faults. Tasks with little runtime have
1595 * little over-all impact on throughput, and thus their
1596 * faults are less important.
1598 f_weight = div64_u64(runtime << 16, period + 1);
1599 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1601 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1602 p->numa_faults_buffer_cpu[i] = 0;
1604 p->numa_faults_memory[i] += diff;
1605 p->numa_faults_cpu[i] += f_diff;
1606 faults += p->numa_faults_memory[i];
1607 p->total_numa_faults += diff;
1608 if (p->numa_group) {
1609 /* safe because we can only change our own group */
1610 p->numa_group->faults[i] += diff;
1611 p->numa_group->faults_cpu[i] += f_diff;
1612 p->numa_group->total_faults += diff;
1613 group_faults += p->numa_group->faults[i];
1617 if (faults > max_faults) {
1618 max_faults = faults;
1622 if (group_faults > max_group_faults) {
1623 max_group_faults = group_faults;
1624 max_group_nid = nid;
1628 update_task_scan_period(p, fault_types[0], fault_types[1]);
1630 if (p->numa_group) {
1631 update_numa_active_node_mask(p->numa_group);
1632 spin_unlock_irq(group_lock);
1633 max_nid = max_group_nid;
1637 /* Set the new preferred node */
1638 if (max_nid != p->numa_preferred_nid)
1639 sched_setnuma(p, max_nid);
1641 if (task_node(p) != p->numa_preferred_nid)
1642 numa_migrate_preferred(p);
1646 static inline int get_numa_group(struct numa_group *grp)
1648 return atomic_inc_not_zero(&grp->refcount);
1651 static inline void put_numa_group(struct numa_group *grp)
1653 if (atomic_dec_and_test(&grp->refcount))
1654 kfree_rcu(grp, rcu);
1657 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1660 struct numa_group *grp, *my_grp;
1661 struct task_struct *tsk;
1663 int cpu = cpupid_to_cpu(cpupid);
1666 if (unlikely(!p->numa_group)) {
1667 unsigned int size = sizeof(struct numa_group) +
1668 4*nr_node_ids*sizeof(unsigned long);
1670 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1674 atomic_set(&grp->refcount, 1);
1675 spin_lock_init(&grp->lock);
1676 INIT_LIST_HEAD(&grp->task_list);
1678 /* Second half of the array tracks nids where faults happen */
1679 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1682 node_set(task_node(current), grp->active_nodes);
1684 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1685 grp->faults[i] = p->numa_faults_memory[i];
1687 grp->total_faults = p->total_numa_faults;
1689 list_add(&p->numa_entry, &grp->task_list);
1691 rcu_assign_pointer(p->numa_group, grp);
1695 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1697 if (!cpupid_match_pid(tsk, cpupid))
1700 grp = rcu_dereference(tsk->numa_group);
1704 my_grp = p->numa_group;
1709 * Only join the other group if its bigger; if we're the bigger group,
1710 * the other task will join us.
1712 if (my_grp->nr_tasks > grp->nr_tasks)
1716 * Tie-break on the grp address.
1718 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1721 /* Always join threads in the same process. */
1722 if (tsk->mm == current->mm)
1725 /* Simple filter to avoid false positives due to PID collisions */
1726 if (flags & TNF_SHARED)
1729 /* Update priv based on whether false sharing was detected */
1732 if (join && !get_numa_group(grp))
1740 BUG_ON(irqs_disabled());
1741 double_lock_irq(&my_grp->lock, &grp->lock);
1743 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1744 my_grp->faults[i] -= p->numa_faults_memory[i];
1745 grp->faults[i] += p->numa_faults_memory[i];
1747 my_grp->total_faults -= p->total_numa_faults;
1748 grp->total_faults += p->total_numa_faults;
1750 list_move(&p->numa_entry, &grp->task_list);
1754 spin_unlock(&my_grp->lock);
1755 spin_unlock_irq(&grp->lock);
1757 rcu_assign_pointer(p->numa_group, grp);
1759 put_numa_group(my_grp);
1767 void task_numa_free(struct task_struct *p)
1769 struct numa_group *grp = p->numa_group;
1770 void *numa_faults = p->numa_faults_memory;
1771 unsigned long flags;
1775 spin_lock_irqsave(&grp->lock, flags);
1776 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1777 grp->faults[i] -= p->numa_faults_memory[i];
1778 grp->total_faults -= p->total_numa_faults;
1780 list_del(&p->numa_entry);
1782 spin_unlock_irqrestore(&grp->lock, flags);
1783 RCU_INIT_POINTER(p->numa_group, NULL);
1784 put_numa_group(grp);
1787 p->numa_faults_memory = NULL;
1788 p->numa_faults_buffer_memory = NULL;
1789 p->numa_faults_cpu= NULL;
1790 p->numa_faults_buffer_cpu = NULL;
1795 * Got a PROT_NONE fault for a page on @node.
1797 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1799 struct task_struct *p = current;
1800 bool migrated = flags & TNF_MIGRATED;
1801 int cpu_node = task_node(current);
1802 int local = !!(flags & TNF_FAULT_LOCAL);
1805 if (!numabalancing_enabled)
1808 /* for example, ksmd faulting in a user's mm */
1812 /* Do not worry about placement if exiting */
1813 if (p->state == TASK_DEAD)
1816 /* Allocate buffer to track faults on a per-node basis */
1817 if (unlikely(!p->numa_faults_memory)) {
1818 int size = sizeof(*p->numa_faults_memory) *
1819 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1821 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1822 if (!p->numa_faults_memory)
1825 BUG_ON(p->numa_faults_buffer_memory);
1827 * The averaged statistics, shared & private, memory & cpu,
1828 * occupy the first half of the array. The second half of the
1829 * array is for current counters, which are averaged into the
1830 * first set by task_numa_placement.
1832 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1833 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1834 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1835 p->total_numa_faults = 0;
1836 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1840 * First accesses are treated as private, otherwise consider accesses
1841 * to be private if the accessing pid has not changed
1843 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1846 priv = cpupid_match_pid(p, last_cpupid);
1847 if (!priv && !(flags & TNF_NO_GROUP))
1848 task_numa_group(p, last_cpupid, flags, &priv);
1852 * If a workload spans multiple NUMA nodes, a shared fault that
1853 * occurs wholly within the set of nodes that the workload is
1854 * actively using should be counted as local. This allows the
1855 * scan rate to slow down when a workload has settled down.
1857 if (!priv && !local && p->numa_group &&
1858 node_isset(cpu_node, p->numa_group->active_nodes) &&
1859 node_isset(mem_node, p->numa_group->active_nodes))
1862 task_numa_placement(p);
1865 * Retry task to preferred node migration periodically, in case it
1866 * case it previously failed, or the scheduler moved us.
1868 if (time_after(jiffies, p->numa_migrate_retry))
1869 numa_migrate_preferred(p);
1872 p->numa_pages_migrated += pages;
1874 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1875 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1876 p->numa_faults_locality[local] += pages;
1879 static void reset_ptenuma_scan(struct task_struct *p)
1881 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1882 p->mm->numa_scan_offset = 0;
1886 * The expensive part of numa migration is done from task_work context.
1887 * Triggered from task_tick_numa().
1889 void task_numa_work(struct callback_head *work)
1891 unsigned long migrate, next_scan, now = jiffies;
1892 struct task_struct *p = current;
1893 struct mm_struct *mm = p->mm;
1894 struct vm_area_struct *vma;
1895 unsigned long start, end;
1896 unsigned long nr_pte_updates = 0;
1899 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1901 work->next = work; /* protect against double add */
1903 * Who cares about NUMA placement when they're dying.
1905 * NOTE: make sure not to dereference p->mm before this check,
1906 * exit_task_work() happens _after_ exit_mm() so we could be called
1907 * without p->mm even though we still had it when we enqueued this
1910 if (p->flags & PF_EXITING)
1913 if (!mm->numa_next_scan) {
1914 mm->numa_next_scan = now +
1915 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1919 * Enforce maximal scan/migration frequency..
1921 migrate = mm->numa_next_scan;
1922 if (time_before(now, migrate))
1925 if (p->numa_scan_period == 0) {
1926 p->numa_scan_period_max = task_scan_max(p);
1927 p->numa_scan_period = task_scan_min(p);
1930 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1931 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1935 * Delay this task enough that another task of this mm will likely win
1936 * the next time around.
1938 p->node_stamp += 2 * TICK_NSEC;
1940 start = mm->numa_scan_offset;
1941 pages = sysctl_numa_balancing_scan_size;
1942 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1946 down_read(&mm->mmap_sem);
1947 vma = find_vma(mm, start);
1949 reset_ptenuma_scan(p);
1953 for (; vma; vma = vma->vm_next) {
1954 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1958 * Shared library pages mapped by multiple processes are not
1959 * migrated as it is expected they are cache replicated. Avoid
1960 * hinting faults in read-only file-backed mappings or the vdso
1961 * as migrating the pages will be of marginal benefit.
1964 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1968 * Skip inaccessible VMAs to avoid any confusion between
1969 * PROT_NONE and NUMA hinting ptes
1971 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1975 start = max(start, vma->vm_start);
1976 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1977 end = min(end, vma->vm_end);
1978 nr_pte_updates += change_prot_numa(vma, start, end);
1981 * Scan sysctl_numa_balancing_scan_size but ensure that
1982 * at least one PTE is updated so that unused virtual
1983 * address space is quickly skipped.
1986 pages -= (end - start) >> PAGE_SHIFT;
1993 } while (end != vma->vm_end);
1998 * It is possible to reach the end of the VMA list but the last few
1999 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2000 * would find the !migratable VMA on the next scan but not reset the
2001 * scanner to the start so check it now.
2004 mm->numa_scan_offset = start;
2006 reset_ptenuma_scan(p);
2007 up_read(&mm->mmap_sem);
2011 * Drive the periodic memory faults..
2013 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2015 struct callback_head *work = &curr->numa_work;
2019 * We don't care about NUMA placement if we don't have memory.
2021 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2025 * Using runtime rather than walltime has the dual advantage that
2026 * we (mostly) drive the selection from busy threads and that the
2027 * task needs to have done some actual work before we bother with
2030 now = curr->se.sum_exec_runtime;
2031 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2033 if (now - curr->node_stamp > period) {
2034 if (!curr->node_stamp)
2035 curr->numa_scan_period = task_scan_min(curr);
2036 curr->node_stamp += period;
2038 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2039 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2040 task_work_add(curr, work, true);
2045 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2049 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2053 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2056 #endif /* CONFIG_NUMA_BALANCING */
2059 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2061 update_load_add(&cfs_rq->load, se->load.weight);
2062 if (!parent_entity(se))
2063 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2065 if (entity_is_task(se)) {
2066 struct rq *rq = rq_of(cfs_rq);
2068 account_numa_enqueue(rq, task_of(se));
2069 list_add(&se->group_node, &rq->cfs_tasks);
2072 cfs_rq->nr_running++;
2076 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2078 update_load_sub(&cfs_rq->load, se->load.weight);
2079 if (!parent_entity(se))
2080 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2081 if (entity_is_task(se)) {
2082 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2083 list_del_init(&se->group_node);
2085 cfs_rq->nr_running--;
2088 #ifdef CONFIG_FAIR_GROUP_SCHED
2090 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2095 * Use this CPU's actual weight instead of the last load_contribution
2096 * to gain a more accurate current total weight. See
2097 * update_cfs_rq_load_contribution().
2099 tg_weight = atomic_long_read(&tg->load_avg);
2100 tg_weight -= cfs_rq->tg_load_contrib;
2101 tg_weight += cfs_rq->load.weight;
2106 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2108 long tg_weight, load, shares;
2110 tg_weight = calc_tg_weight(tg, cfs_rq);
2111 load = cfs_rq->load.weight;
2113 shares = (tg->shares * load);
2115 shares /= tg_weight;
2117 if (shares < MIN_SHARES)
2118 shares = MIN_SHARES;
2119 if (shares > tg->shares)
2120 shares = tg->shares;
2124 # else /* CONFIG_SMP */
2125 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2129 # endif /* CONFIG_SMP */
2130 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2131 unsigned long weight)
2134 /* commit outstanding execution time */
2135 if (cfs_rq->curr == se)
2136 update_curr(cfs_rq);
2137 account_entity_dequeue(cfs_rq, se);
2140 update_load_set(&se->load, weight);
2143 account_entity_enqueue(cfs_rq, se);
2146 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2148 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2150 struct task_group *tg;
2151 struct sched_entity *se;
2155 se = tg->se[cpu_of(rq_of(cfs_rq))];
2156 if (!se || throttled_hierarchy(cfs_rq))
2159 if (likely(se->load.weight == tg->shares))
2162 shares = calc_cfs_shares(cfs_rq, tg);
2164 reweight_entity(cfs_rq_of(se), se, shares);
2166 #else /* CONFIG_FAIR_GROUP_SCHED */
2167 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2170 #endif /* CONFIG_FAIR_GROUP_SCHED */
2174 * We choose a half-life close to 1 scheduling period.
2175 * Note: The tables below are dependent on this value.
2177 #define LOAD_AVG_PERIOD 32
2178 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2179 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2181 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2182 static const u32 runnable_avg_yN_inv[] = {
2183 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2184 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2185 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2186 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2187 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2188 0x85aac367, 0x82cd8698,
2192 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2193 * over-estimates when re-combining.
2195 static const u32 runnable_avg_yN_sum[] = {
2196 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2197 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2198 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2203 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2205 static __always_inline u64 decay_load(u64 val, u64 n)
2207 unsigned int local_n;
2211 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2214 /* after bounds checking we can collapse to 32-bit */
2218 * As y^PERIOD = 1/2, we can combine
2219 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2220 * With a look-up table which covers k^n (n<PERIOD)
2222 * To achieve constant time decay_load.
2224 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2225 val >>= local_n / LOAD_AVG_PERIOD;
2226 local_n %= LOAD_AVG_PERIOD;
2229 val *= runnable_avg_yN_inv[local_n];
2230 /* We don't use SRR here since we always want to round down. */
2235 * For updates fully spanning n periods, the contribution to runnable
2236 * average will be: \Sum 1024*y^n
2238 * We can compute this reasonably efficiently by combining:
2239 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2241 static u32 __compute_runnable_contrib(u64 n)
2245 if (likely(n <= LOAD_AVG_PERIOD))
2246 return runnable_avg_yN_sum[n];
2247 else if (unlikely(n >= LOAD_AVG_MAX_N))
2248 return LOAD_AVG_MAX;
2250 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2252 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2253 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2255 n -= LOAD_AVG_PERIOD;
2256 } while (n > LOAD_AVG_PERIOD);
2258 contrib = decay_load(contrib, n);
2259 return contrib + runnable_avg_yN_sum[n];
2263 * We can represent the historical contribution to runnable average as the
2264 * coefficients of a geometric series. To do this we sub-divide our runnable
2265 * history into segments of approximately 1ms (1024us); label the segment that
2266 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2268 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2270 * (now) (~1ms ago) (~2ms ago)
2272 * Let u_i denote the fraction of p_i that the entity was runnable.
2274 * We then designate the fractions u_i as our co-efficients, yielding the
2275 * following representation of historical load:
2276 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2278 * We choose y based on the with of a reasonably scheduling period, fixing:
2281 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2282 * approximately half as much as the contribution to load within the last ms
2285 * When a period "rolls over" and we have new u_0`, multiplying the previous
2286 * sum again by y is sufficient to update:
2287 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2288 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2290 static __always_inline int __update_entity_runnable_avg(u64 now,
2291 struct sched_avg *sa,
2295 u32 runnable_contrib;
2296 int delta_w, decayed = 0;
2298 delta = now - sa->last_runnable_update;
2300 * This should only happen when time goes backwards, which it
2301 * unfortunately does during sched clock init when we swap over to TSC.
2303 if ((s64)delta < 0) {
2304 sa->last_runnable_update = now;
2309 * Use 1024ns as the unit of measurement since it's a reasonable
2310 * approximation of 1us and fast to compute.
2315 sa->last_runnable_update = now;
2317 /* delta_w is the amount already accumulated against our next period */
2318 delta_w = sa->runnable_avg_period % 1024;
2319 if (delta + delta_w >= 1024) {
2320 /* period roll-over */
2324 * Now that we know we're crossing a period boundary, figure
2325 * out how much from delta we need to complete the current
2326 * period and accrue it.
2328 delta_w = 1024 - delta_w;
2330 sa->runnable_avg_sum += delta_w;
2331 sa->runnable_avg_period += delta_w;
2335 /* Figure out how many additional periods this update spans */
2336 periods = delta / 1024;
2339 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2341 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2344 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2345 runnable_contrib = __compute_runnable_contrib(periods);
2347 sa->runnable_avg_sum += runnable_contrib;
2348 sa->runnable_avg_period += runnable_contrib;
2351 /* Remainder of delta accrued against u_0` */
2353 sa->runnable_avg_sum += delta;
2354 sa->runnable_avg_period += delta;
2359 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2360 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2362 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2363 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2365 decays -= se->avg.decay_count;
2369 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2370 se->avg.decay_count = 0;
2375 #ifdef CONFIG_FAIR_GROUP_SCHED
2376 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2379 struct task_group *tg = cfs_rq->tg;
2382 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2383 tg_contrib -= cfs_rq->tg_load_contrib;
2385 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2386 atomic_long_add(tg_contrib, &tg->load_avg);
2387 cfs_rq->tg_load_contrib += tg_contrib;
2392 * Aggregate cfs_rq runnable averages into an equivalent task_group
2393 * representation for computing load contributions.
2395 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2396 struct cfs_rq *cfs_rq)
2398 struct task_group *tg = cfs_rq->tg;
2401 /* The fraction of a cpu used by this cfs_rq */
2402 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2403 sa->runnable_avg_period + 1);
2404 contrib -= cfs_rq->tg_runnable_contrib;
2406 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2407 atomic_add(contrib, &tg->runnable_avg);
2408 cfs_rq->tg_runnable_contrib += contrib;
2412 static inline void __update_group_entity_contrib(struct sched_entity *se)
2414 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2415 struct task_group *tg = cfs_rq->tg;
2420 contrib = cfs_rq->tg_load_contrib * tg->shares;
2421 se->avg.load_avg_contrib = div_u64(contrib,
2422 atomic_long_read(&tg->load_avg) + 1);
2425 * For group entities we need to compute a correction term in the case
2426 * that they are consuming <1 cpu so that we would contribute the same
2427 * load as a task of equal weight.
2429 * Explicitly co-ordinating this measurement would be expensive, but
2430 * fortunately the sum of each cpus contribution forms a usable
2431 * lower-bound on the true value.
2433 * Consider the aggregate of 2 contributions. Either they are disjoint
2434 * (and the sum represents true value) or they are disjoint and we are
2435 * understating by the aggregate of their overlap.
2437 * Extending this to N cpus, for a given overlap, the maximum amount we
2438 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2439 * cpus that overlap for this interval and w_i is the interval width.
2441 * On a small machine; the first term is well-bounded which bounds the
2442 * total error since w_i is a subset of the period. Whereas on a
2443 * larger machine, while this first term can be larger, if w_i is the
2444 * of consequential size guaranteed to see n_i*w_i quickly converge to
2445 * our upper bound of 1-cpu.
2447 runnable_avg = atomic_read(&tg->runnable_avg);
2448 if (runnable_avg < NICE_0_LOAD) {
2449 se->avg.load_avg_contrib *= runnable_avg;
2450 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2454 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2456 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2457 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2459 #else /* CONFIG_FAIR_GROUP_SCHED */
2460 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2461 int force_update) {}
2462 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2463 struct cfs_rq *cfs_rq) {}
2464 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2465 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2466 #endif /* CONFIG_FAIR_GROUP_SCHED */
2468 static inline void __update_task_entity_contrib(struct sched_entity *se)
2472 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2473 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2474 contrib /= (se->avg.runnable_avg_period + 1);
2475 se->avg.load_avg_contrib = scale_load(contrib);
2478 /* Compute the current contribution to load_avg by se, return any delta */
2479 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2481 long old_contrib = se->avg.load_avg_contrib;
2483 if (entity_is_task(se)) {
2484 __update_task_entity_contrib(se);
2486 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2487 __update_group_entity_contrib(se);
2490 return se->avg.load_avg_contrib - old_contrib;
2493 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2496 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2497 cfs_rq->blocked_load_avg -= load_contrib;
2499 cfs_rq->blocked_load_avg = 0;
2502 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2504 /* Update a sched_entity's runnable average */
2505 static inline void update_entity_load_avg(struct sched_entity *se,
2508 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2513 * For a group entity we need to use their owned cfs_rq_clock_task() in
2514 * case they are the parent of a throttled hierarchy.
2516 if (entity_is_task(se))
2517 now = cfs_rq_clock_task(cfs_rq);
2519 now = cfs_rq_clock_task(group_cfs_rq(se));
2521 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2524 contrib_delta = __update_entity_load_avg_contrib(se);
2530 cfs_rq->runnable_load_avg += contrib_delta;
2532 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2536 * Decay the load contributed by all blocked children and account this so that
2537 * their contribution may appropriately discounted when they wake up.
2539 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2541 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2544 decays = now - cfs_rq->last_decay;
2545 if (!decays && !force_update)
2548 if (atomic_long_read(&cfs_rq->removed_load)) {
2549 unsigned long removed_load;
2550 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2551 subtract_blocked_load_contrib(cfs_rq, removed_load);
2555 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2557 atomic64_add(decays, &cfs_rq->decay_counter);
2558 cfs_rq->last_decay = now;
2561 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2564 /* Add the load generated by se into cfs_rq's child load-average */
2565 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2566 struct sched_entity *se,
2570 * We track migrations using entity decay_count <= 0, on a wake-up
2571 * migration we use a negative decay count to track the remote decays
2572 * accumulated while sleeping.
2574 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2575 * are seen by enqueue_entity_load_avg() as a migration with an already
2576 * constructed load_avg_contrib.
2578 if (unlikely(se->avg.decay_count <= 0)) {
2579 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2580 if (se->avg.decay_count) {
2582 * In a wake-up migration we have to approximate the
2583 * time sleeping. This is because we can't synchronize
2584 * clock_task between the two cpus, and it is not
2585 * guaranteed to be read-safe. Instead, we can
2586 * approximate this using our carried decays, which are
2587 * explicitly atomically readable.
2589 se->avg.last_runnable_update -= (-se->avg.decay_count)
2591 update_entity_load_avg(se, 0);
2592 /* Indicate that we're now synchronized and on-rq */
2593 se->avg.decay_count = 0;
2597 __synchronize_entity_decay(se);
2600 /* migrated tasks did not contribute to our blocked load */
2602 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2603 update_entity_load_avg(se, 0);
2606 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2607 /* we force update consideration on load-balancer moves */
2608 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2612 * Remove se's load from this cfs_rq child load-average, if the entity is
2613 * transitioning to a blocked state we track its projected decay using
2616 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2617 struct sched_entity *se,
2620 update_entity_load_avg(se, 1);
2621 /* we force update consideration on load-balancer moves */
2622 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2624 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2626 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2627 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2628 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2632 * Update the rq's load with the elapsed running time before entering
2633 * idle. if the last scheduled task is not a CFS task, idle_enter will
2634 * be the only way to update the runnable statistic.
2636 void idle_enter_fair(struct rq *this_rq)
2638 update_rq_runnable_avg(this_rq, 1);
2642 * Update the rq's load with the elapsed idle time before a task is
2643 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2644 * be the only way to update the runnable statistic.
2646 void idle_exit_fair(struct rq *this_rq)
2648 update_rq_runnable_avg(this_rq, 0);
2651 static int idle_balance(struct rq *this_rq);
2653 #else /* CONFIG_SMP */
2655 static inline void update_entity_load_avg(struct sched_entity *se,
2656 int update_cfs_rq) {}
2657 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2658 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2659 struct sched_entity *se,
2661 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2662 struct sched_entity *se,
2664 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2665 int force_update) {}
2667 static inline int idle_balance(struct rq *rq)
2672 #endif /* CONFIG_SMP */
2674 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2676 #ifdef CONFIG_SCHEDSTATS
2677 struct task_struct *tsk = NULL;
2679 if (entity_is_task(se))
2682 if (se->statistics.sleep_start) {
2683 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2688 if (unlikely(delta > se->statistics.sleep_max))
2689 se->statistics.sleep_max = delta;
2691 se->statistics.sleep_start = 0;
2692 se->statistics.sum_sleep_runtime += delta;
2695 account_scheduler_latency(tsk, delta >> 10, 1);
2696 trace_sched_stat_sleep(tsk, delta);
2699 if (se->statistics.block_start) {
2700 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2705 if (unlikely(delta > se->statistics.block_max))
2706 se->statistics.block_max = delta;
2708 se->statistics.block_start = 0;
2709 se->statistics.sum_sleep_runtime += delta;
2712 if (tsk->in_iowait) {
2713 se->statistics.iowait_sum += delta;
2714 se->statistics.iowait_count++;
2715 trace_sched_stat_iowait(tsk, delta);
2718 trace_sched_stat_blocked(tsk, delta);
2721 * Blocking time is in units of nanosecs, so shift by
2722 * 20 to get a milliseconds-range estimation of the
2723 * amount of time that the task spent sleeping:
2725 if (unlikely(prof_on == SLEEP_PROFILING)) {
2726 profile_hits(SLEEP_PROFILING,
2727 (void *)get_wchan(tsk),
2730 account_scheduler_latency(tsk, delta >> 10, 0);
2736 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2738 #ifdef CONFIG_SCHED_DEBUG
2739 s64 d = se->vruntime - cfs_rq->min_vruntime;
2744 if (d > 3*sysctl_sched_latency)
2745 schedstat_inc(cfs_rq, nr_spread_over);
2750 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2752 u64 vruntime = cfs_rq->min_vruntime;
2755 * The 'current' period is already promised to the current tasks,
2756 * however the extra weight of the new task will slow them down a
2757 * little, place the new task so that it fits in the slot that
2758 * stays open at the end.
2760 if (initial && sched_feat(START_DEBIT))
2761 vruntime += sched_vslice(cfs_rq, se);
2763 /* sleeps up to a single latency don't count. */
2765 unsigned long thresh = sysctl_sched_latency;
2768 * Halve their sleep time's effect, to allow
2769 * for a gentler effect of sleepers:
2771 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2777 /* ensure we never gain time by being placed backwards. */
2778 se->vruntime = max_vruntime(se->vruntime, vruntime);
2781 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2784 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2787 * Update the normalized vruntime before updating min_vruntime
2788 * through calling update_curr().
2790 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2791 se->vruntime += cfs_rq->min_vruntime;
2794 * Update run-time statistics of the 'current'.
2796 update_curr(cfs_rq);
2797 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2798 account_entity_enqueue(cfs_rq, se);
2799 update_cfs_shares(cfs_rq);
2801 if (flags & ENQUEUE_WAKEUP) {
2802 place_entity(cfs_rq, se, 0);
2803 enqueue_sleeper(cfs_rq, se);
2806 update_stats_enqueue(cfs_rq, se);
2807 check_spread(cfs_rq, se);
2808 if (se != cfs_rq->curr)
2809 __enqueue_entity(cfs_rq, se);
2812 if (cfs_rq->nr_running == 1) {
2813 list_add_leaf_cfs_rq(cfs_rq);
2814 check_enqueue_throttle(cfs_rq);
2818 static void __clear_buddies_last(struct sched_entity *se)
2820 for_each_sched_entity(se) {
2821 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2822 if (cfs_rq->last != se)
2825 cfs_rq->last = NULL;
2829 static void __clear_buddies_next(struct sched_entity *se)
2831 for_each_sched_entity(se) {
2832 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2833 if (cfs_rq->next != se)
2836 cfs_rq->next = NULL;
2840 static void __clear_buddies_skip(struct sched_entity *se)
2842 for_each_sched_entity(se) {
2843 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2844 if (cfs_rq->skip != se)
2847 cfs_rq->skip = NULL;
2851 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2853 if (cfs_rq->last == se)
2854 __clear_buddies_last(se);
2856 if (cfs_rq->next == se)
2857 __clear_buddies_next(se);
2859 if (cfs_rq->skip == se)
2860 __clear_buddies_skip(se);
2863 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2866 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2869 * Update run-time statistics of the 'current'.
2871 update_curr(cfs_rq);
2872 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2874 update_stats_dequeue(cfs_rq, se);
2875 if (flags & DEQUEUE_SLEEP) {
2876 #ifdef CONFIG_SCHEDSTATS
2877 if (entity_is_task(se)) {
2878 struct task_struct *tsk = task_of(se);
2880 if (tsk->state & TASK_INTERRUPTIBLE)
2881 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2882 if (tsk->state & TASK_UNINTERRUPTIBLE)
2883 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2888 clear_buddies(cfs_rq, se);
2890 if (se != cfs_rq->curr)
2891 __dequeue_entity(cfs_rq, se);
2893 account_entity_dequeue(cfs_rq, se);
2896 * Normalize the entity after updating the min_vruntime because the
2897 * update can refer to the ->curr item and we need to reflect this
2898 * movement in our normalized position.
2900 if (!(flags & DEQUEUE_SLEEP))
2901 se->vruntime -= cfs_rq->min_vruntime;
2903 /* return excess runtime on last dequeue */
2904 return_cfs_rq_runtime(cfs_rq);
2906 update_min_vruntime(cfs_rq);
2907 update_cfs_shares(cfs_rq);
2911 * Preempt the current task with a newly woken task if needed:
2914 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2916 unsigned long ideal_runtime, delta_exec;
2917 struct sched_entity *se;
2920 ideal_runtime = sched_slice(cfs_rq, curr);
2921 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2922 if (delta_exec > ideal_runtime) {
2923 resched_curr(rq_of(cfs_rq));
2925 * The current task ran long enough, ensure it doesn't get
2926 * re-elected due to buddy favours.
2928 clear_buddies(cfs_rq, curr);
2933 * Ensure that a task that missed wakeup preemption by a
2934 * narrow margin doesn't have to wait for a full slice.
2935 * This also mitigates buddy induced latencies under load.
2937 if (delta_exec < sysctl_sched_min_granularity)
2940 se = __pick_first_entity(cfs_rq);
2941 delta = curr->vruntime - se->vruntime;
2946 if (delta > ideal_runtime)
2947 resched_curr(rq_of(cfs_rq));
2951 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2953 /* 'current' is not kept within the tree. */
2956 * Any task has to be enqueued before it get to execute on
2957 * a CPU. So account for the time it spent waiting on the
2960 update_stats_wait_end(cfs_rq, se);
2961 __dequeue_entity(cfs_rq, se);
2964 update_stats_curr_start(cfs_rq, se);
2966 #ifdef CONFIG_SCHEDSTATS
2968 * Track our maximum slice length, if the CPU's load is at
2969 * least twice that of our own weight (i.e. dont track it
2970 * when there are only lesser-weight tasks around):
2972 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2973 se->statistics.slice_max = max(se->statistics.slice_max,
2974 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2977 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2981 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2984 * Pick the next process, keeping these things in mind, in this order:
2985 * 1) keep things fair between processes/task groups
2986 * 2) pick the "next" process, since someone really wants that to run
2987 * 3) pick the "last" process, for cache locality
2988 * 4) do not run the "skip" process, if something else is available
2990 static struct sched_entity *
2991 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2993 struct sched_entity *left = __pick_first_entity(cfs_rq);
2994 struct sched_entity *se;
2997 * If curr is set we have to see if its left of the leftmost entity
2998 * still in the tree, provided there was anything in the tree at all.
3000 if (!left || (curr && entity_before(curr, left)))
3003 se = left; /* ideally we run the leftmost entity */
3006 * Avoid running the skip buddy, if running something else can
3007 * be done without getting too unfair.
3009 if (cfs_rq->skip == se) {
3010 struct sched_entity *second;
3013 second = __pick_first_entity(cfs_rq);
3015 second = __pick_next_entity(se);
3016 if (!second || (curr && entity_before(curr, second)))
3020 if (second && wakeup_preempt_entity(second, left) < 1)
3025 * Prefer last buddy, try to return the CPU to a preempted task.
3027 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3031 * Someone really wants this to run. If it's not unfair, run it.
3033 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3036 clear_buddies(cfs_rq, se);
3041 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3043 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3046 * If still on the runqueue then deactivate_task()
3047 * was not called and update_curr() has to be done:
3050 update_curr(cfs_rq);
3052 /* throttle cfs_rqs exceeding runtime */
3053 check_cfs_rq_runtime(cfs_rq);
3055 check_spread(cfs_rq, prev);
3057 update_stats_wait_start(cfs_rq, prev);
3058 /* Put 'current' back into the tree. */
3059 __enqueue_entity(cfs_rq, prev);
3060 /* in !on_rq case, update occurred at dequeue */
3061 update_entity_load_avg(prev, 1);
3063 cfs_rq->curr = NULL;
3067 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3070 * Update run-time statistics of the 'current'.
3072 update_curr(cfs_rq);
3075 * Ensure that runnable average is periodically updated.
3077 update_entity_load_avg(curr, 1);
3078 update_cfs_rq_blocked_load(cfs_rq, 1);
3079 update_cfs_shares(cfs_rq);
3081 #ifdef CONFIG_SCHED_HRTICK
3083 * queued ticks are scheduled to match the slice, so don't bother
3084 * validating it and just reschedule.
3087 resched_curr(rq_of(cfs_rq));
3091 * don't let the period tick interfere with the hrtick preemption
3093 if (!sched_feat(DOUBLE_TICK) &&
3094 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3098 if (cfs_rq->nr_running > 1)
3099 check_preempt_tick(cfs_rq, curr);
3103 /**************************************************
3104 * CFS bandwidth control machinery
3107 #ifdef CONFIG_CFS_BANDWIDTH
3109 #ifdef HAVE_JUMP_LABEL
3110 static struct static_key __cfs_bandwidth_used;
3112 static inline bool cfs_bandwidth_used(void)
3114 return static_key_false(&__cfs_bandwidth_used);
3117 void cfs_bandwidth_usage_inc(void)
3119 static_key_slow_inc(&__cfs_bandwidth_used);
3122 void cfs_bandwidth_usage_dec(void)
3124 static_key_slow_dec(&__cfs_bandwidth_used);
3126 #else /* HAVE_JUMP_LABEL */
3127 static bool cfs_bandwidth_used(void)
3132 void cfs_bandwidth_usage_inc(void) {}
3133 void cfs_bandwidth_usage_dec(void) {}
3134 #endif /* HAVE_JUMP_LABEL */
3137 * default period for cfs group bandwidth.
3138 * default: 0.1s, units: nanoseconds
3140 static inline u64 default_cfs_period(void)
3142 return 100000000ULL;
3145 static inline u64 sched_cfs_bandwidth_slice(void)
3147 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3151 * Replenish runtime according to assigned quota and update expiration time.
3152 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3153 * additional synchronization around rq->lock.
3155 * requires cfs_b->lock
3157 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3161 if (cfs_b->quota == RUNTIME_INF)
3164 now = sched_clock_cpu(smp_processor_id());
3165 cfs_b->runtime = cfs_b->quota;
3166 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3169 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3171 return &tg->cfs_bandwidth;
3174 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3175 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3177 if (unlikely(cfs_rq->throttle_count))
3178 return cfs_rq->throttled_clock_task;
3180 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3183 /* returns 0 on failure to allocate runtime */
3184 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3186 struct task_group *tg = cfs_rq->tg;
3187 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3188 u64 amount = 0, min_amount, expires;
3190 /* note: this is a positive sum as runtime_remaining <= 0 */
3191 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3193 raw_spin_lock(&cfs_b->lock);
3194 if (cfs_b->quota == RUNTIME_INF)
3195 amount = min_amount;
3198 * If the bandwidth pool has become inactive, then at least one
3199 * period must have elapsed since the last consumption.
3200 * Refresh the global state and ensure bandwidth timer becomes
3203 if (!cfs_b->timer_active) {
3204 __refill_cfs_bandwidth_runtime(cfs_b);
3205 __start_cfs_bandwidth(cfs_b, false);
3208 if (cfs_b->runtime > 0) {
3209 amount = min(cfs_b->runtime, min_amount);
3210 cfs_b->runtime -= amount;
3214 expires = cfs_b->runtime_expires;
3215 raw_spin_unlock(&cfs_b->lock);
3217 cfs_rq->runtime_remaining += amount;
3219 * we may have advanced our local expiration to account for allowed
3220 * spread between our sched_clock and the one on which runtime was
3223 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3224 cfs_rq->runtime_expires = expires;
3226 return cfs_rq->runtime_remaining > 0;
3230 * Note: This depends on the synchronization provided by sched_clock and the
3231 * fact that rq->clock snapshots this value.
3233 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3235 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3237 /* if the deadline is ahead of our clock, nothing to do */
3238 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3241 if (cfs_rq->runtime_remaining < 0)
3245 * If the local deadline has passed we have to consider the
3246 * possibility that our sched_clock is 'fast' and the global deadline
3247 * has not truly expired.
3249 * Fortunately we can check determine whether this the case by checking
3250 * whether the global deadline has advanced. It is valid to compare
3251 * cfs_b->runtime_expires without any locks since we only care about
3252 * exact equality, so a partial write will still work.
3255 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3256 /* extend local deadline, drift is bounded above by 2 ticks */
3257 cfs_rq->runtime_expires += TICK_NSEC;
3259 /* global deadline is ahead, expiration has passed */
3260 cfs_rq->runtime_remaining = 0;
3264 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3266 /* dock delta_exec before expiring quota (as it could span periods) */
3267 cfs_rq->runtime_remaining -= delta_exec;
3268 expire_cfs_rq_runtime(cfs_rq);
3270 if (likely(cfs_rq->runtime_remaining > 0))
3274 * if we're unable to extend our runtime we resched so that the active
3275 * hierarchy can be throttled
3277 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3278 resched_curr(rq_of(cfs_rq));
3281 static __always_inline
3282 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3284 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3287 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3290 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3292 return cfs_bandwidth_used() && cfs_rq->throttled;
3295 /* check whether cfs_rq, or any parent, is throttled */
3296 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3298 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3302 * Ensure that neither of the group entities corresponding to src_cpu or
3303 * dest_cpu are members of a throttled hierarchy when performing group
3304 * load-balance operations.
3306 static inline int throttled_lb_pair(struct task_group *tg,
3307 int src_cpu, int dest_cpu)
3309 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3311 src_cfs_rq = tg->cfs_rq[src_cpu];
3312 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3314 return throttled_hierarchy(src_cfs_rq) ||
3315 throttled_hierarchy(dest_cfs_rq);
3318 /* updated child weight may affect parent so we have to do this bottom up */
3319 static int tg_unthrottle_up(struct task_group *tg, void *data)
3321 struct rq *rq = data;
3322 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3324 cfs_rq->throttle_count--;
3326 if (!cfs_rq->throttle_count) {
3327 /* adjust cfs_rq_clock_task() */
3328 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3329 cfs_rq->throttled_clock_task;
3336 static int tg_throttle_down(struct task_group *tg, void *data)
3338 struct rq *rq = data;
3339 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3341 /* group is entering throttled state, stop time */
3342 if (!cfs_rq->throttle_count)
3343 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3344 cfs_rq->throttle_count++;
3349 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3351 struct rq *rq = rq_of(cfs_rq);
3352 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3353 struct sched_entity *se;
3354 long task_delta, dequeue = 1;
3356 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3358 /* freeze hierarchy runnable averages while throttled */
3360 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3363 task_delta = cfs_rq->h_nr_running;
3364 for_each_sched_entity(se) {
3365 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3366 /* throttled entity or throttle-on-deactivate */
3371 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3372 qcfs_rq->h_nr_running -= task_delta;
3374 if (qcfs_rq->load.weight)
3379 sub_nr_running(rq, task_delta);
3381 cfs_rq->throttled = 1;
3382 cfs_rq->throttled_clock = rq_clock(rq);
3383 raw_spin_lock(&cfs_b->lock);
3385 * Add to the _head_ of the list, so that an already-started
3386 * distribute_cfs_runtime will not see us
3388 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3389 if (!cfs_b->timer_active)
3390 __start_cfs_bandwidth(cfs_b, false);
3391 raw_spin_unlock(&cfs_b->lock);
3394 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3396 struct rq *rq = rq_of(cfs_rq);
3397 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3398 struct sched_entity *se;
3402 se = cfs_rq->tg->se[cpu_of(rq)];
3404 cfs_rq->throttled = 0;
3406 update_rq_clock(rq);
3408 raw_spin_lock(&cfs_b->lock);
3409 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3410 list_del_rcu(&cfs_rq->throttled_list);
3411 raw_spin_unlock(&cfs_b->lock);
3413 /* update hierarchical throttle state */
3414 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3416 if (!cfs_rq->load.weight)
3419 task_delta = cfs_rq->h_nr_running;
3420 for_each_sched_entity(se) {
3424 cfs_rq = cfs_rq_of(se);
3426 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3427 cfs_rq->h_nr_running += task_delta;
3429 if (cfs_rq_throttled(cfs_rq))
3434 add_nr_running(rq, task_delta);
3436 /* determine whether we need to wake up potentially idle cpu */
3437 if (rq->curr == rq->idle && rq->cfs.nr_running)
3441 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3442 u64 remaining, u64 expires)
3444 struct cfs_rq *cfs_rq;
3446 u64 starting_runtime = remaining;
3449 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3451 struct rq *rq = rq_of(cfs_rq);
3453 raw_spin_lock(&rq->lock);
3454 if (!cfs_rq_throttled(cfs_rq))
3457 runtime = -cfs_rq->runtime_remaining + 1;
3458 if (runtime > remaining)
3459 runtime = remaining;
3460 remaining -= runtime;
3462 cfs_rq->runtime_remaining += runtime;
3463 cfs_rq->runtime_expires = expires;
3465 /* we check whether we're throttled above */
3466 if (cfs_rq->runtime_remaining > 0)
3467 unthrottle_cfs_rq(cfs_rq);
3470 raw_spin_unlock(&rq->lock);
3477 return starting_runtime - remaining;
3481 * Responsible for refilling a task_group's bandwidth and unthrottling its
3482 * cfs_rqs as appropriate. If there has been no activity within the last
3483 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3484 * used to track this state.
3486 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3488 u64 runtime, runtime_expires;
3491 /* no need to continue the timer with no bandwidth constraint */
3492 if (cfs_b->quota == RUNTIME_INF)
3493 goto out_deactivate;
3495 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3496 cfs_b->nr_periods += overrun;
3499 * idle depends on !throttled (for the case of a large deficit), and if
3500 * we're going inactive then everything else can be deferred
3502 if (cfs_b->idle && !throttled)
3503 goto out_deactivate;
3506 * if we have relooped after returning idle once, we need to update our
3507 * status as actually running, so that other cpus doing
3508 * __start_cfs_bandwidth will stop trying to cancel us.
3510 cfs_b->timer_active = 1;
3512 __refill_cfs_bandwidth_runtime(cfs_b);
3515 /* mark as potentially idle for the upcoming period */
3520 /* account preceding periods in which throttling occurred */
3521 cfs_b->nr_throttled += overrun;
3523 runtime_expires = cfs_b->runtime_expires;
3526 * This check is repeated as we are holding onto the new bandwidth while
3527 * we unthrottle. This can potentially race with an unthrottled group
3528 * trying to acquire new bandwidth from the global pool. This can result
3529 * in us over-using our runtime if it is all used during this loop, but
3530 * only by limited amounts in that extreme case.
3532 while (throttled && cfs_b->runtime > 0) {
3533 runtime = cfs_b->runtime;
3534 raw_spin_unlock(&cfs_b->lock);
3535 /* we can't nest cfs_b->lock while distributing bandwidth */
3536 runtime = distribute_cfs_runtime(cfs_b, runtime,
3538 raw_spin_lock(&cfs_b->lock);
3540 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3542 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3546 * While we are ensured activity in the period following an
3547 * unthrottle, this also covers the case in which the new bandwidth is
3548 * insufficient to cover the existing bandwidth deficit. (Forcing the
3549 * timer to remain active while there are any throttled entities.)
3556 cfs_b->timer_active = 0;
3560 /* a cfs_rq won't donate quota below this amount */
3561 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3562 /* minimum remaining period time to redistribute slack quota */
3563 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3564 /* how long we wait to gather additional slack before distributing */
3565 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3568 * Are we near the end of the current quota period?
3570 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3571 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3572 * migrate_hrtimers, base is never cleared, so we are fine.
3574 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3576 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3579 /* if the call-back is running a quota refresh is already occurring */
3580 if (hrtimer_callback_running(refresh_timer))
3583 /* is a quota refresh about to occur? */
3584 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3585 if (remaining < min_expire)
3591 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3593 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3595 /* if there's a quota refresh soon don't bother with slack */
3596 if (runtime_refresh_within(cfs_b, min_left))
3599 start_bandwidth_timer(&cfs_b->slack_timer,
3600 ns_to_ktime(cfs_bandwidth_slack_period));
3603 /* we know any runtime found here is valid as update_curr() precedes return */
3604 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3606 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3607 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3609 if (slack_runtime <= 0)
3612 raw_spin_lock(&cfs_b->lock);
3613 if (cfs_b->quota != RUNTIME_INF &&
3614 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3615 cfs_b->runtime += slack_runtime;
3617 /* we are under rq->lock, defer unthrottling using a timer */
3618 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3619 !list_empty(&cfs_b->throttled_cfs_rq))
3620 start_cfs_slack_bandwidth(cfs_b);
3622 raw_spin_unlock(&cfs_b->lock);
3624 /* even if it's not valid for return we don't want to try again */
3625 cfs_rq->runtime_remaining -= slack_runtime;
3628 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3630 if (!cfs_bandwidth_used())
3633 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3636 __return_cfs_rq_runtime(cfs_rq);
3640 * This is done with a timer (instead of inline with bandwidth return) since
3641 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3643 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3645 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3648 /* confirm we're still not at a refresh boundary */
3649 raw_spin_lock(&cfs_b->lock);
3650 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3651 raw_spin_unlock(&cfs_b->lock);
3655 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3656 runtime = cfs_b->runtime;
3658 expires = cfs_b->runtime_expires;
3659 raw_spin_unlock(&cfs_b->lock);
3664 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3666 raw_spin_lock(&cfs_b->lock);
3667 if (expires == cfs_b->runtime_expires)
3668 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3669 raw_spin_unlock(&cfs_b->lock);
3673 * When a group wakes up we want to make sure that its quota is not already
3674 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3675 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3677 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3679 if (!cfs_bandwidth_used())
3682 /* an active group must be handled by the update_curr()->put() path */
3683 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3686 /* ensure the group is not already throttled */
3687 if (cfs_rq_throttled(cfs_rq))
3690 /* update runtime allocation */
3691 account_cfs_rq_runtime(cfs_rq, 0);
3692 if (cfs_rq->runtime_remaining <= 0)
3693 throttle_cfs_rq(cfs_rq);
3696 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3697 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3699 if (!cfs_bandwidth_used())
3702 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3706 * it's possible for a throttled entity to be forced into a running
3707 * state (e.g. set_curr_task), in this case we're finished.
3709 if (cfs_rq_throttled(cfs_rq))
3712 throttle_cfs_rq(cfs_rq);
3716 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3718 struct cfs_bandwidth *cfs_b =
3719 container_of(timer, struct cfs_bandwidth, slack_timer);
3720 do_sched_cfs_slack_timer(cfs_b);
3722 return HRTIMER_NORESTART;
3725 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3727 struct cfs_bandwidth *cfs_b =
3728 container_of(timer, struct cfs_bandwidth, period_timer);
3733 raw_spin_lock(&cfs_b->lock);
3735 now = hrtimer_cb_get_time(timer);
3736 overrun = hrtimer_forward(timer, now, cfs_b->period);
3741 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3743 raw_spin_unlock(&cfs_b->lock);
3745 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3748 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3750 raw_spin_lock_init(&cfs_b->lock);
3752 cfs_b->quota = RUNTIME_INF;
3753 cfs_b->period = ns_to_ktime(default_cfs_period());
3755 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3756 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3757 cfs_b->period_timer.function = sched_cfs_period_timer;
3758 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3759 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3762 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3764 cfs_rq->runtime_enabled = 0;
3765 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3768 /* requires cfs_b->lock, may release to reprogram timer */
3769 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force)
3772 * The timer may be active because we're trying to set a new bandwidth
3773 * period or because we're racing with the tear-down path
3774 * (timer_active==0 becomes visible before the hrtimer call-back
3775 * terminates). In either case we ensure that it's re-programmed
3777 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3778 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3779 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3780 raw_spin_unlock(&cfs_b->lock);
3782 raw_spin_lock(&cfs_b->lock);
3783 /* if someone else restarted the timer then we're done */
3784 if (!force && cfs_b->timer_active)
3788 cfs_b->timer_active = 1;
3789 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3792 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3794 hrtimer_cancel(&cfs_b->period_timer);
3795 hrtimer_cancel(&cfs_b->slack_timer);
3798 static void __maybe_unused update_runtime_enabled(struct rq *rq)
3800 struct cfs_rq *cfs_rq;
3802 for_each_leaf_cfs_rq(rq, cfs_rq) {
3803 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
3805 raw_spin_lock(&cfs_b->lock);
3806 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
3807 raw_spin_unlock(&cfs_b->lock);
3811 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3813 struct cfs_rq *cfs_rq;
3815 for_each_leaf_cfs_rq(rq, cfs_rq) {
3816 if (!cfs_rq->runtime_enabled)
3820 * clock_task is not advancing so we just need to make sure
3821 * there's some valid quota amount
3823 cfs_rq->runtime_remaining = 1;
3825 * Offline rq is schedulable till cpu is completely disabled
3826 * in take_cpu_down(), so we prevent new cfs throttling here.
3828 cfs_rq->runtime_enabled = 0;
3830 if (cfs_rq_throttled(cfs_rq))
3831 unthrottle_cfs_rq(cfs_rq);
3835 #else /* CONFIG_CFS_BANDWIDTH */
3836 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3838 return rq_clock_task(rq_of(cfs_rq));
3841 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3842 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3843 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3844 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3846 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3851 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3856 static inline int throttled_lb_pair(struct task_group *tg,
3857 int src_cpu, int dest_cpu)
3862 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3864 #ifdef CONFIG_FAIR_GROUP_SCHED
3865 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3868 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3872 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3873 static inline void update_runtime_enabled(struct rq *rq) {}
3874 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3876 #endif /* CONFIG_CFS_BANDWIDTH */
3878 /**************************************************
3879 * CFS operations on tasks:
3882 #ifdef CONFIG_SCHED_HRTICK
3883 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3885 struct sched_entity *se = &p->se;
3886 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3888 WARN_ON(task_rq(p) != rq);
3890 if (cfs_rq->nr_running > 1) {
3891 u64 slice = sched_slice(cfs_rq, se);
3892 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3893 s64 delta = slice - ran;
3900 hrtick_start(rq, delta);
3905 * called from enqueue/dequeue and updates the hrtick when the
3906 * current task is from our class and nr_running is low enough
3909 static void hrtick_update(struct rq *rq)
3911 struct task_struct *curr = rq->curr;
3913 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3916 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3917 hrtick_start_fair(rq, curr);
3919 #else /* !CONFIG_SCHED_HRTICK */
3921 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3925 static inline void hrtick_update(struct rq *rq)
3931 * The enqueue_task method is called before nr_running is
3932 * increased. Here we update the fair scheduling stats and
3933 * then put the task into the rbtree:
3936 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3938 struct cfs_rq *cfs_rq;
3939 struct sched_entity *se = &p->se;
3941 for_each_sched_entity(se) {
3944 cfs_rq = cfs_rq_of(se);
3945 enqueue_entity(cfs_rq, se, flags);
3948 * end evaluation on encountering a throttled cfs_rq
3950 * note: in the case of encountering a throttled cfs_rq we will
3951 * post the final h_nr_running increment below.
3953 if (cfs_rq_throttled(cfs_rq))
3955 cfs_rq->h_nr_running++;
3957 flags = ENQUEUE_WAKEUP;
3960 for_each_sched_entity(se) {
3961 cfs_rq = cfs_rq_of(se);
3962 cfs_rq->h_nr_running++;
3964 if (cfs_rq_throttled(cfs_rq))
3967 update_cfs_shares(cfs_rq);
3968 update_entity_load_avg(se, 1);
3972 update_rq_runnable_avg(rq, rq->nr_running);
3973 add_nr_running(rq, 1);
3978 static void set_next_buddy(struct sched_entity *se);
3981 * The dequeue_task method is called before nr_running is
3982 * decreased. We remove the task from the rbtree and
3983 * update the fair scheduling stats:
3985 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3987 struct cfs_rq *cfs_rq;
3988 struct sched_entity *se = &p->se;
3989 int task_sleep = flags & DEQUEUE_SLEEP;
3991 for_each_sched_entity(se) {
3992 cfs_rq = cfs_rq_of(se);
3993 dequeue_entity(cfs_rq, se, flags);
3996 * end evaluation on encountering a throttled cfs_rq
3998 * note: in the case of encountering a throttled cfs_rq we will
3999 * post the final h_nr_running decrement below.
4001 if (cfs_rq_throttled(cfs_rq))
4003 cfs_rq->h_nr_running--;
4005 /* Don't dequeue parent if it has other entities besides us */
4006 if (cfs_rq->load.weight) {
4008 * Bias pick_next to pick a task from this cfs_rq, as
4009 * p is sleeping when it is within its sched_slice.
4011 if (task_sleep && parent_entity(se))
4012 set_next_buddy(parent_entity(se));
4014 /* avoid re-evaluating load for this entity */
4015 se = parent_entity(se);
4018 flags |= DEQUEUE_SLEEP;
4021 for_each_sched_entity(se) {
4022 cfs_rq = cfs_rq_of(se);
4023 cfs_rq->h_nr_running--;
4025 if (cfs_rq_throttled(cfs_rq))
4028 update_cfs_shares(cfs_rq);
4029 update_entity_load_avg(se, 1);
4033 sub_nr_running(rq, 1);
4034 update_rq_runnable_avg(rq, 1);
4040 /* Used instead of source_load when we know the type == 0 */
4041 static unsigned long weighted_cpuload(const int cpu)
4043 return cpu_rq(cpu)->cfs.runnable_load_avg;
4047 * Return a low guess at the load of a migration-source cpu weighted
4048 * according to the scheduling class and "nice" value.
4050 * We want to under-estimate the load of migration sources, to
4051 * balance conservatively.
4053 static unsigned long source_load(int cpu, int type)
4055 struct rq *rq = cpu_rq(cpu);
4056 unsigned long total = weighted_cpuload(cpu);
4058 if (type == 0 || !sched_feat(LB_BIAS))
4061 return min(rq->cpu_load[type-1], total);
4065 * Return a high guess at the load of a migration-target cpu weighted
4066 * according to the scheduling class and "nice" value.
4068 static unsigned long target_load(int cpu, int type)
4070 struct rq *rq = cpu_rq(cpu);
4071 unsigned long total = weighted_cpuload(cpu);
4073 if (type == 0 || !sched_feat(LB_BIAS))
4076 return max(rq->cpu_load[type-1], total);
4079 static unsigned long capacity_of(int cpu)
4081 return cpu_rq(cpu)->cpu_capacity;
4084 static unsigned long cpu_avg_load_per_task(int cpu)
4086 struct rq *rq = cpu_rq(cpu);
4087 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
4088 unsigned long load_avg = rq->cfs.runnable_load_avg;
4091 return load_avg / nr_running;
4096 static void record_wakee(struct task_struct *p)
4099 * Rough decay (wiping) for cost saving, don't worry
4100 * about the boundary, really active task won't care
4103 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4104 current->wakee_flips >>= 1;
4105 current->wakee_flip_decay_ts = jiffies;
4108 if (current->last_wakee != p) {
4109 current->last_wakee = p;
4110 current->wakee_flips++;
4114 static void task_waking_fair(struct task_struct *p)
4116 struct sched_entity *se = &p->se;
4117 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4120 #ifndef CONFIG_64BIT
4121 u64 min_vruntime_copy;
4124 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4126 min_vruntime = cfs_rq->min_vruntime;
4127 } while (min_vruntime != min_vruntime_copy);
4129 min_vruntime = cfs_rq->min_vruntime;
4132 se->vruntime -= min_vruntime;
4136 #ifdef CONFIG_FAIR_GROUP_SCHED
4138 * effective_load() calculates the load change as seen from the root_task_group
4140 * Adding load to a group doesn't make a group heavier, but can cause movement
4141 * of group shares between cpus. Assuming the shares were perfectly aligned one
4142 * can calculate the shift in shares.
4144 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4145 * on this @cpu and results in a total addition (subtraction) of @wg to the
4146 * total group weight.
4148 * Given a runqueue weight distribution (rw_i) we can compute a shares
4149 * distribution (s_i) using:
4151 * s_i = rw_i / \Sum rw_j (1)
4153 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4154 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4155 * shares distribution (s_i):
4157 * rw_i = { 2, 4, 1, 0 }
4158 * s_i = { 2/7, 4/7, 1/7, 0 }
4160 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4161 * task used to run on and the CPU the waker is running on), we need to
4162 * compute the effect of waking a task on either CPU and, in case of a sync
4163 * wakeup, compute the effect of the current task going to sleep.
4165 * So for a change of @wl to the local @cpu with an overall group weight change
4166 * of @wl we can compute the new shares distribution (s'_i) using:
4168 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4170 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4171 * differences in waking a task to CPU 0. The additional task changes the
4172 * weight and shares distributions like:
4174 * rw'_i = { 3, 4, 1, 0 }
4175 * s'_i = { 3/8, 4/8, 1/8, 0 }
4177 * We can then compute the difference in effective weight by using:
4179 * dw_i = S * (s'_i - s_i) (3)
4181 * Where 'S' is the group weight as seen by its parent.
4183 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4184 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4185 * 4/7) times the weight of the group.
4187 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4189 struct sched_entity *se = tg->se[cpu];
4191 if (!tg->parent) /* the trivial, non-cgroup case */
4194 for_each_sched_entity(se) {
4200 * W = @wg + \Sum rw_j
4202 W = wg + calc_tg_weight(tg, se->my_q);
4207 w = se->my_q->load.weight + wl;
4210 * wl = S * s'_i; see (2)
4213 wl = (w * tg->shares) / W;
4218 * Per the above, wl is the new se->load.weight value; since
4219 * those are clipped to [MIN_SHARES, ...) do so now. See
4220 * calc_cfs_shares().
4222 if (wl < MIN_SHARES)
4226 * wl = dw_i = S * (s'_i - s_i); see (3)
4228 wl -= se->load.weight;
4231 * Recursively apply this logic to all parent groups to compute
4232 * the final effective load change on the root group. Since
4233 * only the @tg group gets extra weight, all parent groups can
4234 * only redistribute existing shares. @wl is the shift in shares
4235 * resulting from this level per the above.
4244 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4251 static int wake_wide(struct task_struct *p)
4253 int factor = this_cpu_read(sd_llc_size);
4256 * Yeah, it's the switching-frequency, could means many wakee or
4257 * rapidly switch, use factor here will just help to automatically
4258 * adjust the loose-degree, so bigger node will lead to more pull.
4260 if (p->wakee_flips > factor) {
4262 * wakee is somewhat hot, it needs certain amount of cpu
4263 * resource, so if waker is far more hot, prefer to leave
4266 if (current->wakee_flips > (factor * p->wakee_flips))
4273 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4275 s64 this_load, load;
4276 int idx, this_cpu, prev_cpu;
4277 unsigned long tl_per_task;
4278 struct task_group *tg;
4279 unsigned long weight;
4283 * If we wake multiple tasks be careful to not bounce
4284 * ourselves around too much.
4290 this_cpu = smp_processor_id();
4291 prev_cpu = task_cpu(p);
4292 load = source_load(prev_cpu, idx);
4293 this_load = target_load(this_cpu, idx);
4296 * If sync wakeup then subtract the (maximum possible)
4297 * effect of the currently running task from the load
4298 * of the current CPU:
4301 tg = task_group(current);
4302 weight = current->se.load.weight;
4304 this_load += effective_load(tg, this_cpu, -weight, -weight);
4305 load += effective_load(tg, prev_cpu, 0, -weight);
4309 weight = p->se.load.weight;
4312 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4313 * due to the sync cause above having dropped this_load to 0, we'll
4314 * always have an imbalance, but there's really nothing you can do
4315 * about that, so that's good too.
4317 * Otherwise check if either cpus are near enough in load to allow this
4318 * task to be woken on this_cpu.
4320 if (this_load > 0) {
4321 s64 this_eff_load, prev_eff_load;
4323 this_eff_load = 100;
4324 this_eff_load *= capacity_of(prev_cpu);
4325 this_eff_load *= this_load +
4326 effective_load(tg, this_cpu, weight, weight);
4328 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4329 prev_eff_load *= capacity_of(this_cpu);
4330 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4332 balanced = this_eff_load <= prev_eff_load;
4337 * If the currently running task will sleep within
4338 * a reasonable amount of time then attract this newly
4341 if (sync && balanced)
4344 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4345 tl_per_task = cpu_avg_load_per_task(this_cpu);
4348 (this_load <= load &&
4349 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4351 * This domain has SD_WAKE_AFFINE and
4352 * p is cache cold in this domain, and
4353 * there is no bad imbalance.
4355 schedstat_inc(sd, ttwu_move_affine);
4356 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4364 * find_idlest_group finds and returns the least busy CPU group within the
4367 static struct sched_group *
4368 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4369 int this_cpu, int sd_flag)
4371 struct sched_group *idlest = NULL, *group = sd->groups;
4372 unsigned long min_load = ULONG_MAX, this_load = 0;
4373 int load_idx = sd->forkexec_idx;
4374 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4376 if (sd_flag & SD_BALANCE_WAKE)
4377 load_idx = sd->wake_idx;
4380 unsigned long load, avg_load;
4384 /* Skip over this group if it has no CPUs allowed */
4385 if (!cpumask_intersects(sched_group_cpus(group),
4386 tsk_cpus_allowed(p)))
4389 local_group = cpumask_test_cpu(this_cpu,
4390 sched_group_cpus(group));
4392 /* Tally up the load of all CPUs in the group */
4395 for_each_cpu(i, sched_group_cpus(group)) {
4396 /* Bias balancing toward cpus of our domain */
4398 load = source_load(i, load_idx);
4400 load = target_load(i, load_idx);
4405 /* Adjust by relative CPU capacity of the group */
4406 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4409 this_load = avg_load;
4410 } else if (avg_load < min_load) {
4411 min_load = avg_load;
4414 } while (group = group->next, group != sd->groups);
4416 if (!idlest || 100*this_load < imbalance*min_load)
4422 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4425 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4427 unsigned long load, min_load = ULONG_MAX;
4431 /* Traverse only the allowed CPUs */
4432 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4433 load = weighted_cpuload(i);
4435 if (load < min_load || (load == min_load && i == this_cpu)) {
4445 * Try and locate an idle CPU in the sched_domain.
4447 static int select_idle_sibling(struct task_struct *p, int target)
4449 struct sched_domain *sd;
4450 struct sched_group *sg;
4451 int i = task_cpu(p);
4453 if (idle_cpu(target))
4457 * If the prevous cpu is cache affine and idle, don't be stupid.
4459 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4463 * Otherwise, iterate the domains and find an elegible idle cpu.
4465 sd = rcu_dereference(per_cpu(sd_llc, target));
4466 for_each_lower_domain(sd) {
4469 if (!cpumask_intersects(sched_group_cpus(sg),
4470 tsk_cpus_allowed(p)))
4473 for_each_cpu(i, sched_group_cpus(sg)) {
4474 if (i == target || !idle_cpu(i))
4478 target = cpumask_first_and(sched_group_cpus(sg),
4479 tsk_cpus_allowed(p));
4483 } while (sg != sd->groups);
4490 * select_task_rq_fair: Select target runqueue for the waking task in domains
4491 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4492 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4494 * Balances load by selecting the idlest cpu in the idlest group, or under
4495 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4497 * Returns the target cpu number.
4499 * preempt must be disabled.
4502 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4504 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4505 int cpu = smp_processor_id();
4507 int want_affine = 0;
4508 int sync = wake_flags & WF_SYNC;
4510 if (p->nr_cpus_allowed == 1)
4513 if (sd_flag & SD_BALANCE_WAKE) {
4514 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4520 for_each_domain(cpu, tmp) {
4521 if (!(tmp->flags & SD_LOAD_BALANCE))
4525 * If both cpu and prev_cpu are part of this domain,
4526 * cpu is a valid SD_WAKE_AFFINE target.
4528 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4529 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4534 if (tmp->flags & sd_flag)
4538 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4541 if (sd_flag & SD_BALANCE_WAKE) {
4542 new_cpu = select_idle_sibling(p, prev_cpu);
4547 struct sched_group *group;
4550 if (!(sd->flags & sd_flag)) {
4555 group = find_idlest_group(sd, p, cpu, sd_flag);
4561 new_cpu = find_idlest_cpu(group, p, cpu);
4562 if (new_cpu == -1 || new_cpu == cpu) {
4563 /* Now try balancing at a lower domain level of cpu */
4568 /* Now try balancing at a lower domain level of new_cpu */
4570 weight = sd->span_weight;
4572 for_each_domain(cpu, tmp) {
4573 if (weight <= tmp->span_weight)
4575 if (tmp->flags & sd_flag)
4578 /* while loop will break here if sd == NULL */
4587 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4588 * cfs_rq_of(p) references at time of call are still valid and identify the
4589 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4590 * other assumptions, including the state of rq->lock, should be made.
4593 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4595 struct sched_entity *se = &p->se;
4596 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4599 * Load tracking: accumulate removed load so that it can be processed
4600 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4601 * to blocked load iff they have a positive decay-count. It can never
4602 * be negative here since on-rq tasks have decay-count == 0.
4604 if (se->avg.decay_count) {
4605 se->avg.decay_count = -__synchronize_entity_decay(se);
4606 atomic_long_add(se->avg.load_avg_contrib,
4607 &cfs_rq->removed_load);
4610 /* We have migrated, no longer consider this task hot */
4613 #endif /* CONFIG_SMP */
4615 static unsigned long
4616 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4618 unsigned long gran = sysctl_sched_wakeup_granularity;
4621 * Since its curr running now, convert the gran from real-time
4622 * to virtual-time in his units.
4624 * By using 'se' instead of 'curr' we penalize light tasks, so
4625 * they get preempted easier. That is, if 'se' < 'curr' then
4626 * the resulting gran will be larger, therefore penalizing the
4627 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4628 * be smaller, again penalizing the lighter task.
4630 * This is especially important for buddies when the leftmost
4631 * task is higher priority than the buddy.
4633 return calc_delta_fair(gran, se);
4637 * Should 'se' preempt 'curr'.
4651 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4653 s64 gran, vdiff = curr->vruntime - se->vruntime;
4658 gran = wakeup_gran(curr, se);
4665 static void set_last_buddy(struct sched_entity *se)
4667 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4670 for_each_sched_entity(se)
4671 cfs_rq_of(se)->last = se;
4674 static void set_next_buddy(struct sched_entity *se)
4676 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4679 for_each_sched_entity(se)
4680 cfs_rq_of(se)->next = se;
4683 static void set_skip_buddy(struct sched_entity *se)
4685 for_each_sched_entity(se)
4686 cfs_rq_of(se)->skip = se;
4690 * Preempt the current task with a newly woken task if needed:
4692 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4694 struct task_struct *curr = rq->curr;
4695 struct sched_entity *se = &curr->se, *pse = &p->se;
4696 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4697 int scale = cfs_rq->nr_running >= sched_nr_latency;
4698 int next_buddy_marked = 0;
4700 if (unlikely(se == pse))
4704 * This is possible from callers such as attach_tasks(), in which we
4705 * unconditionally check_prempt_curr() after an enqueue (which may have
4706 * lead to a throttle). This both saves work and prevents false
4707 * next-buddy nomination below.
4709 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4712 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4713 set_next_buddy(pse);
4714 next_buddy_marked = 1;
4718 * We can come here with TIF_NEED_RESCHED already set from new task
4721 * Note: this also catches the edge-case of curr being in a throttled
4722 * group (e.g. via set_curr_task), since update_curr() (in the
4723 * enqueue of curr) will have resulted in resched being set. This
4724 * prevents us from potentially nominating it as a false LAST_BUDDY
4727 if (test_tsk_need_resched(curr))
4730 /* Idle tasks are by definition preempted by non-idle tasks. */
4731 if (unlikely(curr->policy == SCHED_IDLE) &&
4732 likely(p->policy != SCHED_IDLE))
4736 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4737 * is driven by the tick):
4739 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4742 find_matching_se(&se, &pse);
4743 update_curr(cfs_rq_of(se));
4745 if (wakeup_preempt_entity(se, pse) == 1) {
4747 * Bias pick_next to pick the sched entity that is
4748 * triggering this preemption.
4750 if (!next_buddy_marked)
4751 set_next_buddy(pse);
4760 * Only set the backward buddy when the current task is still
4761 * on the rq. This can happen when a wakeup gets interleaved
4762 * with schedule on the ->pre_schedule() or idle_balance()
4763 * point, either of which can * drop the rq lock.
4765 * Also, during early boot the idle thread is in the fair class,
4766 * for obvious reasons its a bad idea to schedule back to it.
4768 if (unlikely(!se->on_rq || curr == rq->idle))
4771 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4775 static struct task_struct *
4776 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4778 struct cfs_rq *cfs_rq = &rq->cfs;
4779 struct sched_entity *se;
4780 struct task_struct *p;
4784 #ifdef CONFIG_FAIR_GROUP_SCHED
4785 if (!cfs_rq->nr_running)
4788 if (prev->sched_class != &fair_sched_class)
4792 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4793 * likely that a next task is from the same cgroup as the current.
4795 * Therefore attempt to avoid putting and setting the entire cgroup
4796 * hierarchy, only change the part that actually changes.
4800 struct sched_entity *curr = cfs_rq->curr;
4803 * Since we got here without doing put_prev_entity() we also
4804 * have to consider cfs_rq->curr. If it is still a runnable
4805 * entity, update_curr() will update its vruntime, otherwise
4806 * forget we've ever seen it.
4808 if (curr && curr->on_rq)
4809 update_curr(cfs_rq);
4814 * This call to check_cfs_rq_runtime() will do the throttle and
4815 * dequeue its entity in the parent(s). Therefore the 'simple'
4816 * nr_running test will indeed be correct.
4818 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4821 se = pick_next_entity(cfs_rq, curr);
4822 cfs_rq = group_cfs_rq(se);
4828 * Since we haven't yet done put_prev_entity and if the selected task
4829 * is a different task than we started out with, try and touch the
4830 * least amount of cfs_rqs.
4833 struct sched_entity *pse = &prev->se;
4835 while (!(cfs_rq = is_same_group(se, pse))) {
4836 int se_depth = se->depth;
4837 int pse_depth = pse->depth;
4839 if (se_depth <= pse_depth) {
4840 put_prev_entity(cfs_rq_of(pse), pse);
4841 pse = parent_entity(pse);
4843 if (se_depth >= pse_depth) {
4844 set_next_entity(cfs_rq_of(se), se);
4845 se = parent_entity(se);
4849 put_prev_entity(cfs_rq, pse);
4850 set_next_entity(cfs_rq, se);
4853 if (hrtick_enabled(rq))
4854 hrtick_start_fair(rq, p);
4861 if (!cfs_rq->nr_running)
4864 put_prev_task(rq, prev);
4867 se = pick_next_entity(cfs_rq, NULL);
4868 set_next_entity(cfs_rq, se);
4869 cfs_rq = group_cfs_rq(se);
4874 if (hrtick_enabled(rq))
4875 hrtick_start_fair(rq, p);
4880 new_tasks = idle_balance(rq);
4882 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4883 * possible for any higher priority task to appear. In that case we
4884 * must re-start the pick_next_entity() loop.
4896 * Account for a descheduled task:
4898 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4900 struct sched_entity *se = &prev->se;
4901 struct cfs_rq *cfs_rq;
4903 for_each_sched_entity(se) {
4904 cfs_rq = cfs_rq_of(se);
4905 put_prev_entity(cfs_rq, se);
4910 * sched_yield() is very simple
4912 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4914 static void yield_task_fair(struct rq *rq)
4916 struct task_struct *curr = rq->curr;
4917 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4918 struct sched_entity *se = &curr->se;
4921 * Are we the only task in the tree?
4923 if (unlikely(rq->nr_running == 1))
4926 clear_buddies(cfs_rq, se);
4928 if (curr->policy != SCHED_BATCH) {
4929 update_rq_clock(rq);
4931 * Update run-time statistics of the 'current'.
4933 update_curr(cfs_rq);
4935 * Tell update_rq_clock() that we've just updated,
4936 * so we don't do microscopic update in schedule()
4937 * and double the fastpath cost.
4939 rq->skip_clock_update = 1;
4945 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4947 struct sched_entity *se = &p->se;
4949 /* throttled hierarchies are not runnable */
4950 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4953 /* Tell the scheduler that we'd really like pse to run next. */
4956 yield_task_fair(rq);
4962 /**************************************************
4963 * Fair scheduling class load-balancing methods.
4967 * The purpose of load-balancing is to achieve the same basic fairness the
4968 * per-cpu scheduler provides, namely provide a proportional amount of compute
4969 * time to each task. This is expressed in the following equation:
4971 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4973 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4974 * W_i,0 is defined as:
4976 * W_i,0 = \Sum_j w_i,j (2)
4978 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4979 * is derived from the nice value as per prio_to_weight[].
4981 * The weight average is an exponential decay average of the instantaneous
4984 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4986 * C_i is the compute capacity of cpu i, typically it is the
4987 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4988 * can also include other factors [XXX].
4990 * To achieve this balance we define a measure of imbalance which follows
4991 * directly from (1):
4993 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
4995 * We them move tasks around to minimize the imbalance. In the continuous
4996 * function space it is obvious this converges, in the discrete case we get
4997 * a few fun cases generally called infeasible weight scenarios.
5000 * - infeasible weights;
5001 * - local vs global optima in the discrete case. ]
5006 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5007 * for all i,j solution, we create a tree of cpus that follows the hardware
5008 * topology where each level pairs two lower groups (or better). This results
5009 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5010 * tree to only the first of the previous level and we decrease the frequency
5011 * of load-balance at each level inv. proportional to the number of cpus in
5017 * \Sum { --- * --- * 2^i } = O(n) (5)
5019 * `- size of each group
5020 * | | `- number of cpus doing load-balance
5022 * `- sum over all levels
5024 * Coupled with a limit on how many tasks we can migrate every balance pass,
5025 * this makes (5) the runtime complexity of the balancer.
5027 * An important property here is that each CPU is still (indirectly) connected
5028 * to every other cpu in at most O(log n) steps:
5030 * The adjacency matrix of the resulting graph is given by:
5033 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5036 * And you'll find that:
5038 * A^(log_2 n)_i,j != 0 for all i,j (7)
5040 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5041 * The task movement gives a factor of O(m), giving a convergence complexity
5044 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5049 * In order to avoid CPUs going idle while there's still work to do, new idle
5050 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5051 * tree itself instead of relying on other CPUs to bring it work.
5053 * This adds some complexity to both (5) and (8) but it reduces the total idle
5061 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5064 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5069 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5071 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5073 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5076 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5077 * rewrite all of this once again.]
5080 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5082 enum fbq_type { regular, remote, all };
5084 #define LBF_ALL_PINNED 0x01
5085 #define LBF_NEED_BREAK 0x02
5086 #define LBF_DST_PINNED 0x04
5087 #define LBF_SOME_PINNED 0x08
5090 struct sched_domain *sd;
5098 struct cpumask *dst_grpmask;
5100 enum cpu_idle_type idle;
5102 /* The set of CPUs under consideration for load-balancing */
5103 struct cpumask *cpus;
5108 unsigned int loop_break;
5109 unsigned int loop_max;
5111 enum fbq_type fbq_type;
5112 struct list_head tasks;
5116 * Is this task likely cache-hot:
5118 static int task_hot(struct task_struct *p, struct lb_env *env)
5122 lockdep_assert_held(&env->src_rq->lock);
5124 if (p->sched_class != &fair_sched_class)
5127 if (unlikely(p->policy == SCHED_IDLE))
5131 * Buddy candidates are cache hot:
5133 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5134 (&p->se == cfs_rq_of(&p->se)->next ||
5135 &p->se == cfs_rq_of(&p->se)->last))
5138 if (sysctl_sched_migration_cost == -1)
5140 if (sysctl_sched_migration_cost == 0)
5143 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5145 return delta < (s64)sysctl_sched_migration_cost;
5148 #ifdef CONFIG_NUMA_BALANCING
5149 /* Returns true if the destination node has incurred more faults */
5150 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5152 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5153 int src_nid, dst_nid;
5155 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5156 !(env->sd->flags & SD_NUMA)) {
5160 src_nid = cpu_to_node(env->src_cpu);
5161 dst_nid = cpu_to_node(env->dst_cpu);
5163 if (src_nid == dst_nid)
5167 /* Task is already in the group's interleave set. */
5168 if (node_isset(src_nid, numa_group->active_nodes))
5171 /* Task is moving into the group's interleave set. */
5172 if (node_isset(dst_nid, numa_group->active_nodes))
5175 return group_faults(p, dst_nid) > group_faults(p, src_nid);
5178 /* Encourage migration to the preferred node. */
5179 if (dst_nid == p->numa_preferred_nid)
5182 return task_faults(p, dst_nid) > task_faults(p, src_nid);
5186 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5188 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5189 int src_nid, dst_nid;
5191 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5194 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5197 src_nid = cpu_to_node(env->src_cpu);
5198 dst_nid = cpu_to_node(env->dst_cpu);
5200 if (src_nid == dst_nid)
5204 /* Task is moving within/into the group's interleave set. */
5205 if (node_isset(dst_nid, numa_group->active_nodes))
5208 /* Task is moving out of the group's interleave set. */
5209 if (node_isset(src_nid, numa_group->active_nodes))
5212 return group_faults(p, dst_nid) < group_faults(p, src_nid);
5215 /* Migrating away from the preferred node is always bad. */
5216 if (src_nid == p->numa_preferred_nid)
5219 return task_faults(p, dst_nid) < task_faults(p, src_nid);
5223 static inline bool migrate_improves_locality(struct task_struct *p,
5229 static inline bool migrate_degrades_locality(struct task_struct *p,
5237 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5240 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5242 int tsk_cache_hot = 0;
5244 lockdep_assert_held(&env->src_rq->lock);
5247 * We do not migrate tasks that are:
5248 * 1) throttled_lb_pair, or
5249 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5250 * 3) running (obviously), or
5251 * 4) are cache-hot on their current CPU.
5253 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5256 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5259 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5261 env->flags |= LBF_SOME_PINNED;
5264 * Remember if this task can be migrated to any other cpu in
5265 * our sched_group. We may want to revisit it if we couldn't
5266 * meet load balance goals by pulling other tasks on src_cpu.
5268 * Also avoid computing new_dst_cpu if we have already computed
5269 * one in current iteration.
5271 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5274 /* Prevent to re-select dst_cpu via env's cpus */
5275 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5276 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5277 env->flags |= LBF_DST_PINNED;
5278 env->new_dst_cpu = cpu;
5286 /* Record that we found atleast one task that could run on dst_cpu */
5287 env->flags &= ~LBF_ALL_PINNED;
5289 if (task_running(env->src_rq, p)) {
5290 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5295 * Aggressive migration if:
5296 * 1) destination numa is preferred
5297 * 2) task is cache cold, or
5298 * 3) too many balance attempts have failed.
5300 tsk_cache_hot = task_hot(p, env);
5302 tsk_cache_hot = migrate_degrades_locality(p, env);
5304 if (migrate_improves_locality(p, env)) {
5305 #ifdef CONFIG_SCHEDSTATS
5306 if (tsk_cache_hot) {
5307 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5308 schedstat_inc(p, se.statistics.nr_forced_migrations);
5314 if (!tsk_cache_hot ||
5315 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5317 if (tsk_cache_hot) {
5318 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5319 schedstat_inc(p, se.statistics.nr_forced_migrations);
5325 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5330 * detach_task() -- detach the task for the migration specified in env
5332 static void detach_task(struct task_struct *p, struct lb_env *env)
5334 lockdep_assert_held(&env->src_rq->lock);
5336 deactivate_task(env->src_rq, p, 0);
5337 p->on_rq = TASK_ON_RQ_MIGRATING;
5338 set_task_cpu(p, env->dst_cpu);
5342 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5343 * part of active balancing operations within "domain".
5345 * Returns a task if successful and NULL otherwise.
5347 static struct task_struct *detach_one_task(struct lb_env *env)
5349 struct task_struct *p, *n;
5351 lockdep_assert_held(&env->src_rq->lock);
5353 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5354 if (!can_migrate_task(p, env))
5357 detach_task(p, env);
5360 * Right now, this is only the second place where
5361 * lb_gained[env->idle] is updated (other is detach_tasks)
5362 * so we can safely collect stats here rather than
5363 * inside detach_tasks().
5365 schedstat_inc(env->sd, lb_gained[env->idle]);
5371 static const unsigned int sched_nr_migrate_break = 32;
5374 * detach_tasks() -- tries to detach up to imbalance weighted load from
5375 * busiest_rq, as part of a balancing operation within domain "sd".
5377 * Returns number of detached tasks if successful and 0 otherwise.
5379 static int detach_tasks(struct lb_env *env)
5381 struct list_head *tasks = &env->src_rq->cfs_tasks;
5382 struct task_struct *p;
5386 lockdep_assert_held(&env->src_rq->lock);
5388 if (env->imbalance <= 0)
5391 while (!list_empty(tasks)) {
5392 p = list_first_entry(tasks, struct task_struct, se.group_node);
5395 /* We've more or less seen every task there is, call it quits */
5396 if (env->loop > env->loop_max)
5399 /* take a breather every nr_migrate tasks */
5400 if (env->loop > env->loop_break) {
5401 env->loop_break += sched_nr_migrate_break;
5402 env->flags |= LBF_NEED_BREAK;
5406 if (!can_migrate_task(p, env))
5409 load = task_h_load(p);
5411 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5414 if ((load / 2) > env->imbalance)
5417 detach_task(p, env);
5418 list_add(&p->se.group_node, &env->tasks);
5421 env->imbalance -= load;
5423 #ifdef CONFIG_PREEMPT
5425 * NEWIDLE balancing is a source of latency, so preemptible
5426 * kernels will stop after the first task is detached to minimize
5427 * the critical section.
5429 if (env->idle == CPU_NEWLY_IDLE)
5434 * We only want to steal up to the prescribed amount of
5437 if (env->imbalance <= 0)
5442 list_move_tail(&p->se.group_node, tasks);
5446 * Right now, this is one of only two places we collect this stat
5447 * so we can safely collect detach_one_task() stats here rather
5448 * than inside detach_one_task().
5450 schedstat_add(env->sd, lb_gained[env->idle], detached);
5456 * attach_task() -- attach the task detached by detach_task() to its new rq.
5458 static void attach_task(struct rq *rq, struct task_struct *p)
5460 lockdep_assert_held(&rq->lock);
5462 BUG_ON(task_rq(p) != rq);
5463 p->on_rq = TASK_ON_RQ_QUEUED;
5464 activate_task(rq, p, 0);
5465 check_preempt_curr(rq, p, 0);
5469 * attach_one_task() -- attaches the task returned from detach_one_task() to
5472 static void attach_one_task(struct rq *rq, struct task_struct *p)
5474 raw_spin_lock(&rq->lock);
5476 raw_spin_unlock(&rq->lock);
5480 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5483 static void attach_tasks(struct lb_env *env)
5485 struct list_head *tasks = &env->tasks;
5486 struct task_struct *p;
5488 raw_spin_lock(&env->dst_rq->lock);
5490 while (!list_empty(tasks)) {
5491 p = list_first_entry(tasks, struct task_struct, se.group_node);
5492 list_del_init(&p->se.group_node);
5494 attach_task(env->dst_rq, p);
5497 raw_spin_unlock(&env->dst_rq->lock);
5500 #ifdef CONFIG_FAIR_GROUP_SCHED
5502 * update tg->load_weight by folding this cpu's load_avg
5504 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5506 struct sched_entity *se = tg->se[cpu];
5507 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5509 /* throttled entities do not contribute to load */
5510 if (throttled_hierarchy(cfs_rq))
5513 update_cfs_rq_blocked_load(cfs_rq, 1);
5516 update_entity_load_avg(se, 1);
5518 * We pivot on our runnable average having decayed to zero for
5519 * list removal. This generally implies that all our children
5520 * have also been removed (modulo rounding error or bandwidth
5521 * control); however, such cases are rare and we can fix these
5524 * TODO: fix up out-of-order children on enqueue.
5526 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5527 list_del_leaf_cfs_rq(cfs_rq);
5529 struct rq *rq = rq_of(cfs_rq);
5530 update_rq_runnable_avg(rq, rq->nr_running);
5534 static void update_blocked_averages(int cpu)
5536 struct rq *rq = cpu_rq(cpu);
5537 struct cfs_rq *cfs_rq;
5538 unsigned long flags;
5540 raw_spin_lock_irqsave(&rq->lock, flags);
5541 update_rq_clock(rq);
5543 * Iterates the task_group tree in a bottom up fashion, see
5544 * list_add_leaf_cfs_rq() for details.
5546 for_each_leaf_cfs_rq(rq, cfs_rq) {
5548 * Note: We may want to consider periodically releasing
5549 * rq->lock about these updates so that creating many task
5550 * groups does not result in continually extending hold time.
5552 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5555 raw_spin_unlock_irqrestore(&rq->lock, flags);
5559 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5560 * This needs to be done in a top-down fashion because the load of a child
5561 * group is a fraction of its parents load.
5563 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5565 struct rq *rq = rq_of(cfs_rq);
5566 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5567 unsigned long now = jiffies;
5570 if (cfs_rq->last_h_load_update == now)
5573 cfs_rq->h_load_next = NULL;
5574 for_each_sched_entity(se) {
5575 cfs_rq = cfs_rq_of(se);
5576 cfs_rq->h_load_next = se;
5577 if (cfs_rq->last_h_load_update == now)
5582 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5583 cfs_rq->last_h_load_update = now;
5586 while ((se = cfs_rq->h_load_next) != NULL) {
5587 load = cfs_rq->h_load;
5588 load = div64_ul(load * se->avg.load_avg_contrib,
5589 cfs_rq->runnable_load_avg + 1);
5590 cfs_rq = group_cfs_rq(se);
5591 cfs_rq->h_load = load;
5592 cfs_rq->last_h_load_update = now;
5596 static unsigned long task_h_load(struct task_struct *p)
5598 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5600 update_cfs_rq_h_load(cfs_rq);
5601 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5602 cfs_rq->runnable_load_avg + 1);
5605 static inline void update_blocked_averages(int cpu)
5609 static unsigned long task_h_load(struct task_struct *p)
5611 return p->se.avg.load_avg_contrib;
5615 /********** Helpers for find_busiest_group ************************/
5624 * sg_lb_stats - stats of a sched_group required for load_balancing
5626 struct sg_lb_stats {
5627 unsigned long avg_load; /*Avg load across the CPUs of the group */
5628 unsigned long group_load; /* Total load over the CPUs of the group */
5629 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5630 unsigned long load_per_task;
5631 unsigned long group_capacity;
5632 unsigned int sum_nr_running; /* Nr tasks running in the group */
5633 unsigned int group_capacity_factor;
5634 unsigned int idle_cpus;
5635 unsigned int group_weight;
5636 enum group_type group_type;
5637 int group_has_free_capacity;
5638 #ifdef CONFIG_NUMA_BALANCING
5639 unsigned int nr_numa_running;
5640 unsigned int nr_preferred_running;
5645 * sd_lb_stats - Structure to store the statistics of a sched_domain
5646 * during load balancing.
5648 struct sd_lb_stats {
5649 struct sched_group *busiest; /* Busiest group in this sd */
5650 struct sched_group *local; /* Local group in this sd */
5651 unsigned long total_load; /* Total load of all groups in sd */
5652 unsigned long total_capacity; /* Total capacity of all groups in sd */
5653 unsigned long avg_load; /* Average load across all groups in sd */
5655 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5656 struct sg_lb_stats local_stat; /* Statistics of the local group */
5659 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5662 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5663 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5664 * We must however clear busiest_stat::avg_load because
5665 * update_sd_pick_busiest() reads this before assignment.
5667 *sds = (struct sd_lb_stats){
5671 .total_capacity = 0UL,
5674 .sum_nr_running = 0,
5675 .group_type = group_other,
5681 * get_sd_load_idx - Obtain the load index for a given sched domain.
5682 * @sd: The sched_domain whose load_idx is to be obtained.
5683 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5685 * Return: The load index.
5687 static inline int get_sd_load_idx(struct sched_domain *sd,
5688 enum cpu_idle_type idle)
5694 load_idx = sd->busy_idx;
5697 case CPU_NEWLY_IDLE:
5698 load_idx = sd->newidle_idx;
5701 load_idx = sd->idle_idx;
5708 static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu)
5710 return SCHED_CAPACITY_SCALE;
5713 unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu)
5715 return default_scale_capacity(sd, cpu);
5718 static unsigned long default_scale_smt_capacity(struct sched_domain *sd, int cpu)
5720 unsigned long weight = sd->span_weight;
5721 unsigned long smt_gain = sd->smt_gain;
5728 unsigned long __weak arch_scale_smt_capacity(struct sched_domain *sd, int cpu)
5730 return default_scale_smt_capacity(sd, cpu);
5733 static unsigned long scale_rt_capacity(int cpu)
5735 struct rq *rq = cpu_rq(cpu);
5736 u64 total, available, age_stamp, avg;
5740 * Since we're reading these variables without serialization make sure
5741 * we read them once before doing sanity checks on them.
5743 age_stamp = ACCESS_ONCE(rq->age_stamp);
5744 avg = ACCESS_ONCE(rq->rt_avg);
5746 delta = rq_clock(rq) - age_stamp;
5747 if (unlikely(delta < 0))
5750 total = sched_avg_period() + delta;
5752 if (unlikely(total < avg)) {
5753 /* Ensures that capacity won't end up being negative */
5756 available = total - avg;
5759 if (unlikely((s64)total < SCHED_CAPACITY_SCALE))
5760 total = SCHED_CAPACITY_SCALE;
5762 total >>= SCHED_CAPACITY_SHIFT;
5764 return div_u64(available, total);
5767 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
5769 unsigned long weight = sd->span_weight;
5770 unsigned long capacity = SCHED_CAPACITY_SCALE;
5771 struct sched_group *sdg = sd->groups;
5773 if ((sd->flags & SD_SHARE_CPUCAPACITY) && weight > 1) {
5774 if (sched_feat(ARCH_CAPACITY))
5775 capacity *= arch_scale_smt_capacity(sd, cpu);
5777 capacity *= default_scale_smt_capacity(sd, cpu);
5779 capacity >>= SCHED_CAPACITY_SHIFT;
5782 sdg->sgc->capacity_orig = capacity;
5784 if (sched_feat(ARCH_CAPACITY))
5785 capacity *= arch_scale_freq_capacity(sd, cpu);
5787 capacity *= default_scale_capacity(sd, cpu);
5789 capacity >>= SCHED_CAPACITY_SHIFT;
5791 capacity *= scale_rt_capacity(cpu);
5792 capacity >>= SCHED_CAPACITY_SHIFT;
5797 cpu_rq(cpu)->cpu_capacity = capacity;
5798 sdg->sgc->capacity = capacity;
5801 void update_group_capacity(struct sched_domain *sd, int cpu)
5803 struct sched_domain *child = sd->child;
5804 struct sched_group *group, *sdg = sd->groups;
5805 unsigned long capacity, capacity_orig;
5806 unsigned long interval;
5808 interval = msecs_to_jiffies(sd->balance_interval);
5809 interval = clamp(interval, 1UL, max_load_balance_interval);
5810 sdg->sgc->next_update = jiffies + interval;
5813 update_cpu_capacity(sd, cpu);
5817 capacity_orig = capacity = 0;
5819 if (child->flags & SD_OVERLAP) {
5821 * SD_OVERLAP domains cannot assume that child groups
5822 * span the current group.
5825 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5826 struct sched_group_capacity *sgc;
5827 struct rq *rq = cpu_rq(cpu);
5830 * build_sched_domains() -> init_sched_groups_capacity()
5831 * gets here before we've attached the domains to the
5834 * Use capacity_of(), which is set irrespective of domains
5835 * in update_cpu_capacity().
5837 * This avoids capacity/capacity_orig from being 0 and
5838 * causing divide-by-zero issues on boot.
5840 * Runtime updates will correct capacity_orig.
5842 if (unlikely(!rq->sd)) {
5843 capacity_orig += capacity_of(cpu);
5844 capacity += capacity_of(cpu);
5848 sgc = rq->sd->groups->sgc;
5849 capacity_orig += sgc->capacity_orig;
5850 capacity += sgc->capacity;
5854 * !SD_OVERLAP domains can assume that child groups
5855 * span the current group.
5858 group = child->groups;
5860 capacity_orig += group->sgc->capacity_orig;
5861 capacity += group->sgc->capacity;
5862 group = group->next;
5863 } while (group != child->groups);
5866 sdg->sgc->capacity_orig = capacity_orig;
5867 sdg->sgc->capacity = capacity;
5871 * Try and fix up capacity for tiny siblings, this is needed when
5872 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5873 * which on its own isn't powerful enough.
5875 * See update_sd_pick_busiest() and check_asym_packing().
5878 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5881 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5883 if (!(sd->flags & SD_SHARE_CPUCAPACITY))
5887 * If ~90% of the cpu_capacity is still there, we're good.
5889 if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29)
5896 * Group imbalance indicates (and tries to solve) the problem where balancing
5897 * groups is inadequate due to tsk_cpus_allowed() constraints.
5899 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5900 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5903 * { 0 1 2 3 } { 4 5 6 7 }
5906 * If we were to balance group-wise we'd place two tasks in the first group and
5907 * two tasks in the second group. Clearly this is undesired as it will overload
5908 * cpu 3 and leave one of the cpus in the second group unused.
5910 * The current solution to this issue is detecting the skew in the first group
5911 * by noticing the lower domain failed to reach balance and had difficulty
5912 * moving tasks due to affinity constraints.
5914 * When this is so detected; this group becomes a candidate for busiest; see
5915 * update_sd_pick_busiest(). And calculate_imbalance() and
5916 * find_busiest_group() avoid some of the usual balance conditions to allow it
5917 * to create an effective group imbalance.
5919 * This is a somewhat tricky proposition since the next run might not find the
5920 * group imbalance and decide the groups need to be balanced again. A most
5921 * subtle and fragile situation.
5924 static inline int sg_imbalanced(struct sched_group *group)
5926 return group->sgc->imbalance;
5930 * Compute the group capacity factor.
5932 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5933 * first dividing out the smt factor and computing the actual number of cores
5934 * and limit unit capacity with that.
5936 static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group)
5938 unsigned int capacity_factor, smt, cpus;
5939 unsigned int capacity, capacity_orig;
5941 capacity = group->sgc->capacity;
5942 capacity_orig = group->sgc->capacity_orig;
5943 cpus = group->group_weight;
5945 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5946 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig);
5947 capacity_factor = cpus / smt; /* cores */
5949 capacity_factor = min_t(unsigned,
5950 capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE));
5951 if (!capacity_factor)
5952 capacity_factor = fix_small_capacity(env->sd, group);
5954 return capacity_factor;
5957 static enum group_type
5958 group_classify(struct sched_group *group, struct sg_lb_stats *sgs)
5960 if (sgs->sum_nr_running > sgs->group_capacity_factor)
5961 return group_overloaded;
5963 if (sg_imbalanced(group))
5964 return group_imbalanced;
5970 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5971 * @env: The load balancing environment.
5972 * @group: sched_group whose statistics are to be updated.
5973 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5974 * @local_group: Does group contain this_cpu.
5975 * @sgs: variable to hold the statistics for this group.
5976 * @overload: Indicate more than one runnable task for any CPU.
5978 static inline void update_sg_lb_stats(struct lb_env *env,
5979 struct sched_group *group, int load_idx,
5980 int local_group, struct sg_lb_stats *sgs,
5986 memset(sgs, 0, sizeof(*sgs));
5988 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5989 struct rq *rq = cpu_rq(i);
5991 /* Bias balancing toward cpus of our domain */
5993 load = target_load(i, load_idx);
5995 load = source_load(i, load_idx);
5997 sgs->group_load += load;
5998 sgs->sum_nr_running += rq->nr_running;
6000 if (rq->nr_running > 1)
6003 #ifdef CONFIG_NUMA_BALANCING
6004 sgs->nr_numa_running += rq->nr_numa_running;
6005 sgs->nr_preferred_running += rq->nr_preferred_running;
6007 sgs->sum_weighted_load += weighted_cpuload(i);
6012 /* Adjust by relative CPU capacity of the group */
6013 sgs->group_capacity = group->sgc->capacity;
6014 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6016 if (sgs->sum_nr_running)
6017 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6019 sgs->group_weight = group->group_weight;
6020 sgs->group_capacity_factor = sg_capacity_factor(env, group);
6021 sgs->group_type = group_classify(group, sgs);
6023 if (sgs->group_capacity_factor > sgs->sum_nr_running)
6024 sgs->group_has_free_capacity = 1;
6028 * update_sd_pick_busiest - return 1 on busiest group
6029 * @env: The load balancing environment.
6030 * @sds: sched_domain statistics
6031 * @sg: sched_group candidate to be checked for being the busiest
6032 * @sgs: sched_group statistics
6034 * Determine if @sg is a busier group than the previously selected
6037 * Return: %true if @sg is a busier group than the previously selected
6038 * busiest group. %false otherwise.
6040 static bool update_sd_pick_busiest(struct lb_env *env,
6041 struct sd_lb_stats *sds,
6042 struct sched_group *sg,
6043 struct sg_lb_stats *sgs)
6045 struct sg_lb_stats *busiest = &sds->busiest_stat;
6047 if (sgs->group_type > busiest->group_type)
6050 if (sgs->group_type < busiest->group_type)
6053 if (sgs->avg_load <= busiest->avg_load)
6056 /* This is the busiest node in its class. */
6057 if (!(env->sd->flags & SD_ASYM_PACKING))
6061 * ASYM_PACKING needs to move all the work to the lowest
6062 * numbered CPUs in the group, therefore mark all groups
6063 * higher than ourself as busy.
6065 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6069 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6076 #ifdef CONFIG_NUMA_BALANCING
6077 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6079 if (sgs->sum_nr_running > sgs->nr_numa_running)
6081 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6086 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6088 if (rq->nr_running > rq->nr_numa_running)
6090 if (rq->nr_running > rq->nr_preferred_running)
6095 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6100 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6104 #endif /* CONFIG_NUMA_BALANCING */
6107 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6108 * @env: The load balancing environment.
6109 * @sds: variable to hold the statistics for this sched_domain.
6111 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6113 struct sched_domain *child = env->sd->child;
6114 struct sched_group *sg = env->sd->groups;
6115 struct sg_lb_stats tmp_sgs;
6116 int load_idx, prefer_sibling = 0;
6117 bool overload = false;
6119 if (child && child->flags & SD_PREFER_SIBLING)
6122 load_idx = get_sd_load_idx(env->sd, env->idle);
6125 struct sg_lb_stats *sgs = &tmp_sgs;
6128 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6131 sgs = &sds->local_stat;
6133 if (env->idle != CPU_NEWLY_IDLE ||
6134 time_after_eq(jiffies, sg->sgc->next_update))
6135 update_group_capacity(env->sd, env->dst_cpu);
6138 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6145 * In case the child domain prefers tasks go to siblings
6146 * first, lower the sg capacity factor to one so that we'll try
6147 * and move all the excess tasks away. We lower the capacity
6148 * of a group only if the local group has the capacity to fit
6149 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6150 * extra check prevents the case where you always pull from the
6151 * heaviest group when it is already under-utilized (possible
6152 * with a large weight task outweighs the tasks on the system).
6154 if (prefer_sibling && sds->local &&
6155 sds->local_stat.group_has_free_capacity)
6156 sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U);
6158 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6160 sds->busiest_stat = *sgs;
6164 /* Now, start updating sd_lb_stats */
6165 sds->total_load += sgs->group_load;
6166 sds->total_capacity += sgs->group_capacity;
6169 } while (sg != env->sd->groups);
6171 if (env->sd->flags & SD_NUMA)
6172 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6174 if (!env->sd->parent) {
6175 /* update overload indicator if we are at root domain */
6176 if (env->dst_rq->rd->overload != overload)
6177 env->dst_rq->rd->overload = overload;
6183 * check_asym_packing - Check to see if the group is packed into the
6186 * This is primarily intended to used at the sibling level. Some
6187 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6188 * case of POWER7, it can move to lower SMT modes only when higher
6189 * threads are idle. When in lower SMT modes, the threads will
6190 * perform better since they share less core resources. Hence when we
6191 * have idle threads, we want them to be the higher ones.
6193 * This packing function is run on idle threads. It checks to see if
6194 * the busiest CPU in this domain (core in the P7 case) has a higher
6195 * CPU number than the packing function is being run on. Here we are
6196 * assuming lower CPU number will be equivalent to lower a SMT thread
6199 * Return: 1 when packing is required and a task should be moved to
6200 * this CPU. The amount of the imbalance is returned in *imbalance.
6202 * @env: The load balancing environment.
6203 * @sds: Statistics of the sched_domain which is to be packed
6205 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6209 if (!(env->sd->flags & SD_ASYM_PACKING))
6215 busiest_cpu = group_first_cpu(sds->busiest);
6216 if (env->dst_cpu > busiest_cpu)
6219 env->imbalance = DIV_ROUND_CLOSEST(
6220 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6221 SCHED_CAPACITY_SCALE);
6227 * fix_small_imbalance - Calculate the minor imbalance that exists
6228 * amongst the groups of a sched_domain, during
6230 * @env: The load balancing environment.
6231 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6234 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6236 unsigned long tmp, capa_now = 0, capa_move = 0;
6237 unsigned int imbn = 2;
6238 unsigned long scaled_busy_load_per_task;
6239 struct sg_lb_stats *local, *busiest;
6241 local = &sds->local_stat;
6242 busiest = &sds->busiest_stat;
6244 if (!local->sum_nr_running)
6245 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6246 else if (busiest->load_per_task > local->load_per_task)
6249 scaled_busy_load_per_task =
6250 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6251 busiest->group_capacity;
6253 if (busiest->avg_load + scaled_busy_load_per_task >=
6254 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6255 env->imbalance = busiest->load_per_task;
6260 * OK, we don't have enough imbalance to justify moving tasks,
6261 * however we may be able to increase total CPU capacity used by
6265 capa_now += busiest->group_capacity *
6266 min(busiest->load_per_task, busiest->avg_load);
6267 capa_now += local->group_capacity *
6268 min(local->load_per_task, local->avg_load);
6269 capa_now /= SCHED_CAPACITY_SCALE;
6271 /* Amount of load we'd subtract */
6272 if (busiest->avg_load > scaled_busy_load_per_task) {
6273 capa_move += busiest->group_capacity *
6274 min(busiest->load_per_task,
6275 busiest->avg_load - scaled_busy_load_per_task);
6278 /* Amount of load we'd add */
6279 if (busiest->avg_load * busiest->group_capacity <
6280 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6281 tmp = (busiest->avg_load * busiest->group_capacity) /
6282 local->group_capacity;
6284 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6285 local->group_capacity;
6287 capa_move += local->group_capacity *
6288 min(local->load_per_task, local->avg_load + tmp);
6289 capa_move /= SCHED_CAPACITY_SCALE;
6291 /* Move if we gain throughput */
6292 if (capa_move > capa_now)
6293 env->imbalance = busiest->load_per_task;
6297 * calculate_imbalance - Calculate the amount of imbalance present within the
6298 * groups of a given sched_domain during load balance.
6299 * @env: load balance environment
6300 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6302 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6304 unsigned long max_pull, load_above_capacity = ~0UL;
6305 struct sg_lb_stats *local, *busiest;
6307 local = &sds->local_stat;
6308 busiest = &sds->busiest_stat;
6310 if (busiest->group_type == group_imbalanced) {
6312 * In the group_imb case we cannot rely on group-wide averages
6313 * to ensure cpu-load equilibrium, look at wider averages. XXX
6315 busiest->load_per_task =
6316 min(busiest->load_per_task, sds->avg_load);
6320 * In the presence of smp nice balancing, certain scenarios can have
6321 * max load less than avg load(as we skip the groups at or below
6322 * its cpu_capacity, while calculating max_load..)
6324 if (busiest->avg_load <= sds->avg_load ||
6325 local->avg_load >= sds->avg_load) {
6327 return fix_small_imbalance(env, sds);
6331 * If there aren't any idle cpus, avoid creating some.
6333 if (busiest->group_type == group_overloaded &&
6334 local->group_type == group_overloaded) {
6335 load_above_capacity =
6336 (busiest->sum_nr_running - busiest->group_capacity_factor);
6338 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE);
6339 load_above_capacity /= busiest->group_capacity;
6343 * We're trying to get all the cpus to the average_load, so we don't
6344 * want to push ourselves above the average load, nor do we wish to
6345 * reduce the max loaded cpu below the average load. At the same time,
6346 * we also don't want to reduce the group load below the group capacity
6347 * (so that we can implement power-savings policies etc). Thus we look
6348 * for the minimum possible imbalance.
6350 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6352 /* How much load to actually move to equalise the imbalance */
6353 env->imbalance = min(
6354 max_pull * busiest->group_capacity,
6355 (sds->avg_load - local->avg_load) * local->group_capacity
6356 ) / SCHED_CAPACITY_SCALE;
6359 * if *imbalance is less than the average load per runnable task
6360 * there is no guarantee that any tasks will be moved so we'll have
6361 * a think about bumping its value to force at least one task to be
6364 if (env->imbalance < busiest->load_per_task)
6365 return fix_small_imbalance(env, sds);
6368 /******* find_busiest_group() helpers end here *********************/
6371 * find_busiest_group - Returns the busiest group within the sched_domain
6372 * if there is an imbalance. If there isn't an imbalance, and
6373 * the user has opted for power-savings, it returns a group whose
6374 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6375 * such a group exists.
6377 * Also calculates the amount of weighted load which should be moved
6378 * to restore balance.
6380 * @env: The load balancing environment.
6382 * Return: - The busiest group if imbalance exists.
6383 * - If no imbalance and user has opted for power-savings balance,
6384 * return the least loaded group whose CPUs can be
6385 * put to idle by rebalancing its tasks onto our group.
6387 static struct sched_group *find_busiest_group(struct lb_env *env)
6389 struct sg_lb_stats *local, *busiest;
6390 struct sd_lb_stats sds;
6392 init_sd_lb_stats(&sds);
6395 * Compute the various statistics relavent for load balancing at
6398 update_sd_lb_stats(env, &sds);
6399 local = &sds.local_stat;
6400 busiest = &sds.busiest_stat;
6402 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6403 check_asym_packing(env, &sds))
6406 /* There is no busy sibling group to pull tasks from */
6407 if (!sds.busiest || busiest->sum_nr_running == 0)
6410 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6411 / sds.total_capacity;
6414 * If the busiest group is imbalanced the below checks don't
6415 * work because they assume all things are equal, which typically
6416 * isn't true due to cpus_allowed constraints and the like.
6418 if (busiest->group_type == group_imbalanced)
6421 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6422 if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity &&
6423 !busiest->group_has_free_capacity)
6427 * If the local group is more busy than the selected busiest group
6428 * don't try and pull any tasks.
6430 if (local->avg_load >= busiest->avg_load)
6434 * Don't pull any tasks if this group is already above the domain
6437 if (local->avg_load >= sds.avg_load)
6440 if (env->idle == CPU_IDLE) {
6442 * This cpu is idle. If the busiest group load doesn't
6443 * have more tasks than the number of available cpu's and
6444 * there is no imbalance between this and busiest group
6445 * wrt to idle cpu's, it is balanced.
6447 if ((local->idle_cpus < busiest->idle_cpus) &&
6448 busiest->sum_nr_running <= busiest->group_weight)
6452 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6453 * imbalance_pct to be conservative.
6455 if (100 * busiest->avg_load <=
6456 env->sd->imbalance_pct * local->avg_load)
6461 /* Looks like there is an imbalance. Compute it */
6462 calculate_imbalance(env, &sds);
6471 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6473 static struct rq *find_busiest_queue(struct lb_env *env,
6474 struct sched_group *group)
6476 struct rq *busiest = NULL, *rq;
6477 unsigned long busiest_load = 0, busiest_capacity = 1;
6480 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6481 unsigned long capacity, capacity_factor, wl;
6485 rt = fbq_classify_rq(rq);
6488 * We classify groups/runqueues into three groups:
6489 * - regular: there are !numa tasks
6490 * - remote: there are numa tasks that run on the 'wrong' node
6491 * - all: there is no distinction
6493 * In order to avoid migrating ideally placed numa tasks,
6494 * ignore those when there's better options.
6496 * If we ignore the actual busiest queue to migrate another
6497 * task, the next balance pass can still reduce the busiest
6498 * queue by moving tasks around inside the node.
6500 * If we cannot move enough load due to this classification
6501 * the next pass will adjust the group classification and
6502 * allow migration of more tasks.
6504 * Both cases only affect the total convergence complexity.
6506 if (rt > env->fbq_type)
6509 capacity = capacity_of(i);
6510 capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE);
6511 if (!capacity_factor)
6512 capacity_factor = fix_small_capacity(env->sd, group);
6514 wl = weighted_cpuload(i);
6517 * When comparing with imbalance, use weighted_cpuload()
6518 * which is not scaled with the cpu capacity.
6520 if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance)
6524 * For the load comparisons with the other cpu's, consider
6525 * the weighted_cpuload() scaled with the cpu capacity, so
6526 * that the load can be moved away from the cpu that is
6527 * potentially running at a lower capacity.
6529 * Thus we're looking for max(wl_i / capacity_i), crosswise
6530 * multiplication to rid ourselves of the division works out
6531 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6532 * our previous maximum.
6534 if (wl * busiest_capacity > busiest_load * capacity) {
6536 busiest_capacity = capacity;
6545 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6546 * so long as it is large enough.
6548 #define MAX_PINNED_INTERVAL 512
6550 /* Working cpumask for load_balance and load_balance_newidle. */
6551 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6553 static int need_active_balance(struct lb_env *env)
6555 struct sched_domain *sd = env->sd;
6557 if (env->idle == CPU_NEWLY_IDLE) {
6560 * ASYM_PACKING needs to force migrate tasks from busy but
6561 * higher numbered CPUs in order to pack all tasks in the
6562 * lowest numbered CPUs.
6564 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6568 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6571 static int active_load_balance_cpu_stop(void *data);
6573 static int should_we_balance(struct lb_env *env)
6575 struct sched_group *sg = env->sd->groups;
6576 struct cpumask *sg_cpus, *sg_mask;
6577 int cpu, balance_cpu = -1;
6580 * In the newly idle case, we will allow all the cpu's
6581 * to do the newly idle load balance.
6583 if (env->idle == CPU_NEWLY_IDLE)
6586 sg_cpus = sched_group_cpus(sg);
6587 sg_mask = sched_group_mask(sg);
6588 /* Try to find first idle cpu */
6589 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6590 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6597 if (balance_cpu == -1)
6598 balance_cpu = group_balance_cpu(sg);
6601 * First idle cpu or the first cpu(busiest) in this sched group
6602 * is eligible for doing load balancing at this and above domains.
6604 return balance_cpu == env->dst_cpu;
6608 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6609 * tasks if there is an imbalance.
6611 static int load_balance(int this_cpu, struct rq *this_rq,
6612 struct sched_domain *sd, enum cpu_idle_type idle,
6613 int *continue_balancing)
6615 int ld_moved, cur_ld_moved, active_balance = 0;
6616 struct sched_domain *sd_parent = sd->parent;
6617 struct sched_group *group;
6619 unsigned long flags;
6620 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6622 struct lb_env env = {
6624 .dst_cpu = this_cpu,
6626 .dst_grpmask = sched_group_cpus(sd->groups),
6628 .loop_break = sched_nr_migrate_break,
6631 .tasks = LIST_HEAD_INIT(env.tasks),
6635 * For NEWLY_IDLE load_balancing, we don't need to consider
6636 * other cpus in our group
6638 if (idle == CPU_NEWLY_IDLE)
6639 env.dst_grpmask = NULL;
6641 cpumask_copy(cpus, cpu_active_mask);
6643 schedstat_inc(sd, lb_count[idle]);
6646 if (!should_we_balance(&env)) {
6647 *continue_balancing = 0;
6651 group = find_busiest_group(&env);
6653 schedstat_inc(sd, lb_nobusyg[idle]);
6657 busiest = find_busiest_queue(&env, group);
6659 schedstat_inc(sd, lb_nobusyq[idle]);
6663 BUG_ON(busiest == env.dst_rq);
6665 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6668 if (busiest->nr_running > 1) {
6670 * Attempt to move tasks. If find_busiest_group has found
6671 * an imbalance but busiest->nr_running <= 1, the group is
6672 * still unbalanced. ld_moved simply stays zero, so it is
6673 * correctly treated as an imbalance.
6675 env.flags |= LBF_ALL_PINNED;
6676 env.src_cpu = busiest->cpu;
6677 env.src_rq = busiest;
6678 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6681 raw_spin_lock_irqsave(&busiest->lock, flags);
6684 * cur_ld_moved - load moved in current iteration
6685 * ld_moved - cumulative load moved across iterations
6687 cur_ld_moved = detach_tasks(&env);
6690 * We've detached some tasks from busiest_rq. Every
6691 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6692 * unlock busiest->lock, and we are able to be sure
6693 * that nobody can manipulate the tasks in parallel.
6694 * See task_rq_lock() family for the details.
6697 raw_spin_unlock(&busiest->lock);
6701 ld_moved += cur_ld_moved;
6704 local_irq_restore(flags);
6707 * some other cpu did the load balance for us.
6709 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6710 resched_cpu(env.dst_cpu);
6712 if (env.flags & LBF_NEED_BREAK) {
6713 env.flags &= ~LBF_NEED_BREAK;
6718 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6719 * us and move them to an alternate dst_cpu in our sched_group
6720 * where they can run. The upper limit on how many times we
6721 * iterate on same src_cpu is dependent on number of cpus in our
6724 * This changes load balance semantics a bit on who can move
6725 * load to a given_cpu. In addition to the given_cpu itself
6726 * (or a ilb_cpu acting on its behalf where given_cpu is
6727 * nohz-idle), we now have balance_cpu in a position to move
6728 * load to given_cpu. In rare situations, this may cause
6729 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6730 * _independently_ and at _same_ time to move some load to
6731 * given_cpu) causing exceess load to be moved to given_cpu.
6732 * This however should not happen so much in practice and
6733 * moreover subsequent load balance cycles should correct the
6734 * excess load moved.
6736 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6738 /* Prevent to re-select dst_cpu via env's cpus */
6739 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6741 env.dst_rq = cpu_rq(env.new_dst_cpu);
6742 env.dst_cpu = env.new_dst_cpu;
6743 env.flags &= ~LBF_DST_PINNED;
6745 env.loop_break = sched_nr_migrate_break;
6748 * Go back to "more_balance" rather than "redo" since we
6749 * need to continue with same src_cpu.
6755 * We failed to reach balance because of affinity.
6758 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
6760 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6761 *group_imbalance = 1;
6762 } else if (*group_imbalance)
6763 *group_imbalance = 0;
6766 /* All tasks on this runqueue were pinned by CPU affinity */
6767 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6768 cpumask_clear_cpu(cpu_of(busiest), cpus);
6769 if (!cpumask_empty(cpus)) {
6771 env.loop_break = sched_nr_migrate_break;
6779 schedstat_inc(sd, lb_failed[idle]);
6781 * Increment the failure counter only on periodic balance.
6782 * We do not want newidle balance, which can be very
6783 * frequent, pollute the failure counter causing
6784 * excessive cache_hot migrations and active balances.
6786 if (idle != CPU_NEWLY_IDLE)
6787 sd->nr_balance_failed++;
6789 if (need_active_balance(&env)) {
6790 raw_spin_lock_irqsave(&busiest->lock, flags);
6792 /* don't kick the active_load_balance_cpu_stop,
6793 * if the curr task on busiest cpu can't be
6796 if (!cpumask_test_cpu(this_cpu,
6797 tsk_cpus_allowed(busiest->curr))) {
6798 raw_spin_unlock_irqrestore(&busiest->lock,
6800 env.flags |= LBF_ALL_PINNED;
6801 goto out_one_pinned;
6805 * ->active_balance synchronizes accesses to
6806 * ->active_balance_work. Once set, it's cleared
6807 * only after active load balance is finished.
6809 if (!busiest->active_balance) {
6810 busiest->active_balance = 1;
6811 busiest->push_cpu = this_cpu;
6814 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6816 if (active_balance) {
6817 stop_one_cpu_nowait(cpu_of(busiest),
6818 active_load_balance_cpu_stop, busiest,
6819 &busiest->active_balance_work);
6823 * We've kicked active balancing, reset the failure
6826 sd->nr_balance_failed = sd->cache_nice_tries+1;
6829 sd->nr_balance_failed = 0;
6831 if (likely(!active_balance)) {
6832 /* We were unbalanced, so reset the balancing interval */
6833 sd->balance_interval = sd->min_interval;
6836 * If we've begun active balancing, start to back off. This
6837 * case may not be covered by the all_pinned logic if there
6838 * is only 1 task on the busy runqueue (because we don't call
6841 if (sd->balance_interval < sd->max_interval)
6842 sd->balance_interval *= 2;
6848 schedstat_inc(sd, lb_balanced[idle]);
6850 sd->nr_balance_failed = 0;
6853 /* tune up the balancing interval */
6854 if (((env.flags & LBF_ALL_PINNED) &&
6855 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6856 (sd->balance_interval < sd->max_interval))
6857 sd->balance_interval *= 2;
6864 static inline unsigned long
6865 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
6867 unsigned long interval = sd->balance_interval;
6870 interval *= sd->busy_factor;
6872 /* scale ms to jiffies */
6873 interval = msecs_to_jiffies(interval);
6874 interval = clamp(interval, 1UL, max_load_balance_interval);
6880 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
6882 unsigned long interval, next;
6884 interval = get_sd_balance_interval(sd, cpu_busy);
6885 next = sd->last_balance + interval;
6887 if (time_after(*next_balance, next))
6888 *next_balance = next;
6892 * idle_balance is called by schedule() if this_cpu is about to become
6893 * idle. Attempts to pull tasks from other CPUs.
6895 static int idle_balance(struct rq *this_rq)
6897 unsigned long next_balance = jiffies + HZ;
6898 int this_cpu = this_rq->cpu;
6899 struct sched_domain *sd;
6900 int pulled_task = 0;
6903 idle_enter_fair(this_rq);
6906 * We must set idle_stamp _before_ calling idle_balance(), such that we
6907 * measure the duration of idle_balance() as idle time.
6909 this_rq->idle_stamp = rq_clock(this_rq);
6911 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
6912 !this_rq->rd->overload) {
6914 sd = rcu_dereference_check_sched_domain(this_rq->sd);
6916 update_next_balance(sd, 0, &next_balance);
6923 * Drop the rq->lock, but keep IRQ/preempt disabled.
6925 raw_spin_unlock(&this_rq->lock);
6927 update_blocked_averages(this_cpu);
6929 for_each_domain(this_cpu, sd) {
6930 int continue_balancing = 1;
6931 u64 t0, domain_cost;
6933 if (!(sd->flags & SD_LOAD_BALANCE))
6936 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
6937 update_next_balance(sd, 0, &next_balance);
6941 if (sd->flags & SD_BALANCE_NEWIDLE) {
6942 t0 = sched_clock_cpu(this_cpu);
6944 pulled_task = load_balance(this_cpu, this_rq,
6946 &continue_balancing);
6948 domain_cost = sched_clock_cpu(this_cpu) - t0;
6949 if (domain_cost > sd->max_newidle_lb_cost)
6950 sd->max_newidle_lb_cost = domain_cost;
6952 curr_cost += domain_cost;
6955 update_next_balance(sd, 0, &next_balance);
6958 * Stop searching for tasks to pull if there are
6959 * now runnable tasks on this rq.
6961 if (pulled_task || this_rq->nr_running > 0)
6966 raw_spin_lock(&this_rq->lock);
6968 if (curr_cost > this_rq->max_idle_balance_cost)
6969 this_rq->max_idle_balance_cost = curr_cost;
6972 * While browsing the domains, we released the rq lock, a task could
6973 * have been enqueued in the meantime. Since we're not going idle,
6974 * pretend we pulled a task.
6976 if (this_rq->cfs.h_nr_running && !pulled_task)
6980 /* Move the next balance forward */
6981 if (time_after(this_rq->next_balance, next_balance))
6982 this_rq->next_balance = next_balance;
6984 /* Is there a task of a high priority class? */
6985 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
6989 idle_exit_fair(this_rq);
6990 this_rq->idle_stamp = 0;
6997 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6998 * running tasks off the busiest CPU onto idle CPUs. It requires at
6999 * least 1 task to be running on each physical CPU where possible, and
7000 * avoids physical / logical imbalances.
7002 static int active_load_balance_cpu_stop(void *data)
7004 struct rq *busiest_rq = data;
7005 int busiest_cpu = cpu_of(busiest_rq);
7006 int target_cpu = busiest_rq->push_cpu;
7007 struct rq *target_rq = cpu_rq(target_cpu);
7008 struct sched_domain *sd;
7009 struct task_struct *p = NULL;
7011 raw_spin_lock_irq(&busiest_rq->lock);
7013 /* make sure the requested cpu hasn't gone down in the meantime */
7014 if (unlikely(busiest_cpu != smp_processor_id() ||
7015 !busiest_rq->active_balance))
7018 /* Is there any task to move? */
7019 if (busiest_rq->nr_running <= 1)
7023 * This condition is "impossible", if it occurs
7024 * we need to fix it. Originally reported by
7025 * Bjorn Helgaas on a 128-cpu setup.
7027 BUG_ON(busiest_rq == target_rq);
7029 /* Search for an sd spanning us and the target CPU. */
7031 for_each_domain(target_cpu, sd) {
7032 if ((sd->flags & SD_LOAD_BALANCE) &&
7033 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7038 struct lb_env env = {
7040 .dst_cpu = target_cpu,
7041 .dst_rq = target_rq,
7042 .src_cpu = busiest_rq->cpu,
7043 .src_rq = busiest_rq,
7047 schedstat_inc(sd, alb_count);
7049 p = detach_one_task(&env);
7051 schedstat_inc(sd, alb_pushed);
7053 schedstat_inc(sd, alb_failed);
7057 busiest_rq->active_balance = 0;
7058 raw_spin_unlock(&busiest_rq->lock);
7061 attach_one_task(target_rq, p);
7068 static inline int on_null_domain(struct rq *rq)
7070 return unlikely(!rcu_dereference_sched(rq->sd));
7073 #ifdef CONFIG_NO_HZ_COMMON
7075 * idle load balancing details
7076 * - When one of the busy CPUs notice that there may be an idle rebalancing
7077 * needed, they will kick the idle load balancer, which then does idle
7078 * load balancing for all the idle CPUs.
7081 cpumask_var_t idle_cpus_mask;
7083 unsigned long next_balance; /* in jiffy units */
7084 } nohz ____cacheline_aligned;
7086 static inline int find_new_ilb(void)
7088 int ilb = cpumask_first(nohz.idle_cpus_mask);
7090 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7097 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7098 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7099 * CPU (if there is one).
7101 static void nohz_balancer_kick(void)
7105 nohz.next_balance++;
7107 ilb_cpu = find_new_ilb();
7109 if (ilb_cpu >= nr_cpu_ids)
7112 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7115 * Use smp_send_reschedule() instead of resched_cpu().
7116 * This way we generate a sched IPI on the target cpu which
7117 * is idle. And the softirq performing nohz idle load balance
7118 * will be run before returning from the IPI.
7120 smp_send_reschedule(ilb_cpu);
7124 static inline void nohz_balance_exit_idle(int cpu)
7126 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7128 * Completely isolated CPUs don't ever set, so we must test.
7130 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7131 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7132 atomic_dec(&nohz.nr_cpus);
7134 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7138 static inline void set_cpu_sd_state_busy(void)
7140 struct sched_domain *sd;
7141 int cpu = smp_processor_id();
7144 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7146 if (!sd || !sd->nohz_idle)
7150 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7155 void set_cpu_sd_state_idle(void)
7157 struct sched_domain *sd;
7158 int cpu = smp_processor_id();
7161 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7163 if (!sd || sd->nohz_idle)
7167 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7173 * This routine will record that the cpu is going idle with tick stopped.
7174 * This info will be used in performing idle load balancing in the future.
7176 void nohz_balance_enter_idle(int cpu)
7179 * If this cpu is going down, then nothing needs to be done.
7181 if (!cpu_active(cpu))
7184 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7188 * If we're a completely isolated CPU, we don't play.
7190 if (on_null_domain(cpu_rq(cpu)))
7193 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7194 atomic_inc(&nohz.nr_cpus);
7195 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7198 static int sched_ilb_notifier(struct notifier_block *nfb,
7199 unsigned long action, void *hcpu)
7201 switch (action & ~CPU_TASKS_FROZEN) {
7203 nohz_balance_exit_idle(smp_processor_id());
7211 static DEFINE_SPINLOCK(balancing);
7214 * Scale the max load_balance interval with the number of CPUs in the system.
7215 * This trades load-balance latency on larger machines for less cross talk.
7217 void update_max_interval(void)
7219 max_load_balance_interval = HZ*num_online_cpus()/10;
7223 * It checks each scheduling domain to see if it is due to be balanced,
7224 * and initiates a balancing operation if so.
7226 * Balancing parameters are set up in init_sched_domains.
7228 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7230 int continue_balancing = 1;
7232 unsigned long interval;
7233 struct sched_domain *sd;
7234 /* Earliest time when we have to do rebalance again */
7235 unsigned long next_balance = jiffies + 60*HZ;
7236 int update_next_balance = 0;
7237 int need_serialize, need_decay = 0;
7240 update_blocked_averages(cpu);
7243 for_each_domain(cpu, sd) {
7245 * Decay the newidle max times here because this is a regular
7246 * visit to all the domains. Decay ~1% per second.
7248 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7249 sd->max_newidle_lb_cost =
7250 (sd->max_newidle_lb_cost * 253) / 256;
7251 sd->next_decay_max_lb_cost = jiffies + HZ;
7254 max_cost += sd->max_newidle_lb_cost;
7256 if (!(sd->flags & SD_LOAD_BALANCE))
7260 * Stop the load balance at this level. There is another
7261 * CPU in our sched group which is doing load balancing more
7264 if (!continue_balancing) {
7270 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7272 need_serialize = sd->flags & SD_SERIALIZE;
7273 if (need_serialize) {
7274 if (!spin_trylock(&balancing))
7278 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7279 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7281 * The LBF_DST_PINNED logic could have changed
7282 * env->dst_cpu, so we can't know our idle
7283 * state even if we migrated tasks. Update it.
7285 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7287 sd->last_balance = jiffies;
7288 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7291 spin_unlock(&balancing);
7293 if (time_after(next_balance, sd->last_balance + interval)) {
7294 next_balance = sd->last_balance + interval;
7295 update_next_balance = 1;
7300 * Ensure the rq-wide value also decays but keep it at a
7301 * reasonable floor to avoid funnies with rq->avg_idle.
7303 rq->max_idle_balance_cost =
7304 max((u64)sysctl_sched_migration_cost, max_cost);
7309 * next_balance will be updated only when there is a need.
7310 * When the cpu is attached to null domain for ex, it will not be
7313 if (likely(update_next_balance))
7314 rq->next_balance = next_balance;
7317 #ifdef CONFIG_NO_HZ_COMMON
7319 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7320 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7322 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7324 int this_cpu = this_rq->cpu;
7328 if (idle != CPU_IDLE ||
7329 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7332 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7333 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7337 * If this cpu gets work to do, stop the load balancing
7338 * work being done for other cpus. Next load
7339 * balancing owner will pick it up.
7344 rq = cpu_rq(balance_cpu);
7347 * If time for next balance is due,
7350 if (time_after_eq(jiffies, rq->next_balance)) {
7351 raw_spin_lock_irq(&rq->lock);
7352 update_rq_clock(rq);
7353 update_idle_cpu_load(rq);
7354 raw_spin_unlock_irq(&rq->lock);
7355 rebalance_domains(rq, CPU_IDLE);
7358 if (time_after(this_rq->next_balance, rq->next_balance))
7359 this_rq->next_balance = rq->next_balance;
7361 nohz.next_balance = this_rq->next_balance;
7363 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7367 * Current heuristic for kicking the idle load balancer in the presence
7368 * of an idle cpu is the system.
7369 * - This rq has more than one task.
7370 * - At any scheduler domain level, this cpu's scheduler group has multiple
7371 * busy cpu's exceeding the group's capacity.
7372 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7373 * domain span are idle.
7375 static inline int nohz_kick_needed(struct rq *rq)
7377 unsigned long now = jiffies;
7378 struct sched_domain *sd;
7379 struct sched_group_capacity *sgc;
7380 int nr_busy, cpu = rq->cpu;
7382 if (unlikely(rq->idle_balance))
7386 * We may be recently in ticked or tickless idle mode. At the first
7387 * busy tick after returning from idle, we will update the busy stats.
7389 set_cpu_sd_state_busy();
7390 nohz_balance_exit_idle(cpu);
7393 * None are in tickless mode and hence no need for NOHZ idle load
7396 if (likely(!atomic_read(&nohz.nr_cpus)))
7399 if (time_before(now, nohz.next_balance))
7402 if (rq->nr_running >= 2)
7406 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7409 sgc = sd->groups->sgc;
7410 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7413 goto need_kick_unlock;
7416 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7418 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7419 sched_domain_span(sd)) < cpu))
7420 goto need_kick_unlock;
7431 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7435 * run_rebalance_domains is triggered when needed from the scheduler tick.
7436 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7438 static void run_rebalance_domains(struct softirq_action *h)
7440 struct rq *this_rq = this_rq();
7441 enum cpu_idle_type idle = this_rq->idle_balance ?
7442 CPU_IDLE : CPU_NOT_IDLE;
7444 rebalance_domains(this_rq, idle);
7447 * If this cpu has a pending nohz_balance_kick, then do the
7448 * balancing on behalf of the other idle cpus whose ticks are
7451 nohz_idle_balance(this_rq, idle);
7455 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7457 void trigger_load_balance(struct rq *rq)
7459 /* Don't need to rebalance while attached to NULL domain */
7460 if (unlikely(on_null_domain(rq)))
7463 if (time_after_eq(jiffies, rq->next_balance))
7464 raise_softirq(SCHED_SOFTIRQ);
7465 #ifdef CONFIG_NO_HZ_COMMON
7466 if (nohz_kick_needed(rq))
7467 nohz_balancer_kick();
7471 static void rq_online_fair(struct rq *rq)
7475 update_runtime_enabled(rq);
7478 static void rq_offline_fair(struct rq *rq)
7482 /* Ensure any throttled groups are reachable by pick_next_task */
7483 unthrottle_offline_cfs_rqs(rq);
7486 #endif /* CONFIG_SMP */
7489 * scheduler tick hitting a task of our scheduling class:
7491 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7493 struct cfs_rq *cfs_rq;
7494 struct sched_entity *se = &curr->se;
7496 for_each_sched_entity(se) {
7497 cfs_rq = cfs_rq_of(se);
7498 entity_tick(cfs_rq, se, queued);
7501 if (numabalancing_enabled)
7502 task_tick_numa(rq, curr);
7504 update_rq_runnable_avg(rq, 1);
7508 * called on fork with the child task as argument from the parent's context
7509 * - child not yet on the tasklist
7510 * - preemption disabled
7512 static void task_fork_fair(struct task_struct *p)
7514 struct cfs_rq *cfs_rq;
7515 struct sched_entity *se = &p->se, *curr;
7516 int this_cpu = smp_processor_id();
7517 struct rq *rq = this_rq();
7518 unsigned long flags;
7520 raw_spin_lock_irqsave(&rq->lock, flags);
7522 update_rq_clock(rq);
7524 cfs_rq = task_cfs_rq(current);
7525 curr = cfs_rq->curr;
7528 * Not only the cpu but also the task_group of the parent might have
7529 * been changed after parent->se.parent,cfs_rq were copied to
7530 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7531 * of child point to valid ones.
7534 __set_task_cpu(p, this_cpu);
7537 update_curr(cfs_rq);
7540 se->vruntime = curr->vruntime;
7541 place_entity(cfs_rq, se, 1);
7543 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7545 * Upon rescheduling, sched_class::put_prev_task() will place
7546 * 'current' within the tree based on its new key value.
7548 swap(curr->vruntime, se->vruntime);
7552 se->vruntime -= cfs_rq->min_vruntime;
7554 raw_spin_unlock_irqrestore(&rq->lock, flags);
7558 * Priority of the task has changed. Check to see if we preempt
7562 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7564 if (!task_on_rq_queued(p))
7568 * Reschedule if we are currently running on this runqueue and
7569 * our priority decreased, or if we are not currently running on
7570 * this runqueue and our priority is higher than the current's
7572 if (rq->curr == p) {
7573 if (p->prio > oldprio)
7576 check_preempt_curr(rq, p, 0);
7579 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7581 struct sched_entity *se = &p->se;
7582 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7585 * Ensure the task's vruntime is normalized, so that when it's
7586 * switched back to the fair class the enqueue_entity(.flags=0) will
7587 * do the right thing.
7589 * If it's queued, then the dequeue_entity(.flags=0) will already
7590 * have normalized the vruntime, if it's !queued, then only when
7591 * the task is sleeping will it still have non-normalized vruntime.
7593 if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) {
7595 * Fix up our vruntime so that the current sleep doesn't
7596 * cause 'unlimited' sleep bonus.
7598 place_entity(cfs_rq, se, 0);
7599 se->vruntime -= cfs_rq->min_vruntime;
7604 * Remove our load from contribution when we leave sched_fair
7605 * and ensure we don't carry in an old decay_count if we
7608 if (se->avg.decay_count) {
7609 __synchronize_entity_decay(se);
7610 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7616 * We switched to the sched_fair class.
7618 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7620 #ifdef CONFIG_FAIR_GROUP_SCHED
7621 struct sched_entity *se = &p->se;
7623 * Since the real-depth could have been changed (only FAIR
7624 * class maintain depth value), reset depth properly.
7626 se->depth = se->parent ? se->parent->depth + 1 : 0;
7628 if (!task_on_rq_queued(p))
7632 * We were most likely switched from sched_rt, so
7633 * kick off the schedule if running, otherwise just see
7634 * if we can still preempt the current task.
7639 check_preempt_curr(rq, p, 0);
7642 /* Account for a task changing its policy or group.
7644 * This routine is mostly called to set cfs_rq->curr field when a task
7645 * migrates between groups/classes.
7647 static void set_curr_task_fair(struct rq *rq)
7649 struct sched_entity *se = &rq->curr->se;
7651 for_each_sched_entity(se) {
7652 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7654 set_next_entity(cfs_rq, se);
7655 /* ensure bandwidth has been allocated on our new cfs_rq */
7656 account_cfs_rq_runtime(cfs_rq, 0);
7660 void init_cfs_rq(struct cfs_rq *cfs_rq)
7662 cfs_rq->tasks_timeline = RB_ROOT;
7663 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7664 #ifndef CONFIG_64BIT
7665 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7668 atomic64_set(&cfs_rq->decay_counter, 1);
7669 atomic_long_set(&cfs_rq->removed_load, 0);
7673 #ifdef CONFIG_FAIR_GROUP_SCHED
7674 static void task_move_group_fair(struct task_struct *p, int queued)
7676 struct sched_entity *se = &p->se;
7677 struct cfs_rq *cfs_rq;
7680 * If the task was not on the rq at the time of this cgroup movement
7681 * it must have been asleep, sleeping tasks keep their ->vruntime
7682 * absolute on their old rq until wakeup (needed for the fair sleeper
7683 * bonus in place_entity()).
7685 * If it was on the rq, we've just 'preempted' it, which does convert
7686 * ->vruntime to a relative base.
7688 * Make sure both cases convert their relative position when migrating
7689 * to another cgroup's rq. This does somewhat interfere with the
7690 * fair sleeper stuff for the first placement, but who cares.
7693 * When !queued, vruntime of the task has usually NOT been normalized.
7694 * But there are some cases where it has already been normalized:
7696 * - Moving a forked child which is waiting for being woken up by
7697 * wake_up_new_task().
7698 * - Moving a task which has been woken up by try_to_wake_up() and
7699 * waiting for actually being woken up by sched_ttwu_pending().
7701 * To prevent boost or penalty in the new cfs_rq caused by delta
7702 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7704 if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7708 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7709 set_task_rq(p, task_cpu(p));
7710 se->depth = se->parent ? se->parent->depth + 1 : 0;
7712 cfs_rq = cfs_rq_of(se);
7713 se->vruntime += cfs_rq->min_vruntime;
7716 * migrate_task_rq_fair() will have removed our previous
7717 * contribution, but we must synchronize for ongoing future
7720 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7721 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7726 void free_fair_sched_group(struct task_group *tg)
7730 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7732 for_each_possible_cpu(i) {
7734 kfree(tg->cfs_rq[i]);
7743 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7745 struct cfs_rq *cfs_rq;
7746 struct sched_entity *se;
7749 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7752 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7756 tg->shares = NICE_0_LOAD;
7758 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7760 for_each_possible_cpu(i) {
7761 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7762 GFP_KERNEL, cpu_to_node(i));
7766 se = kzalloc_node(sizeof(struct sched_entity),
7767 GFP_KERNEL, cpu_to_node(i));
7771 init_cfs_rq(cfs_rq);
7772 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7783 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7785 struct rq *rq = cpu_rq(cpu);
7786 unsigned long flags;
7789 * Only empty task groups can be destroyed; so we can speculatively
7790 * check on_list without danger of it being re-added.
7792 if (!tg->cfs_rq[cpu]->on_list)
7795 raw_spin_lock_irqsave(&rq->lock, flags);
7796 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7797 raw_spin_unlock_irqrestore(&rq->lock, flags);
7800 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7801 struct sched_entity *se, int cpu,
7802 struct sched_entity *parent)
7804 struct rq *rq = cpu_rq(cpu);
7808 init_cfs_rq_runtime(cfs_rq);
7810 tg->cfs_rq[cpu] = cfs_rq;
7813 /* se could be NULL for root_task_group */
7818 se->cfs_rq = &rq->cfs;
7821 se->cfs_rq = parent->my_q;
7822 se->depth = parent->depth + 1;
7826 /* guarantee group entities always have weight */
7827 update_load_set(&se->load, NICE_0_LOAD);
7828 se->parent = parent;
7831 static DEFINE_MUTEX(shares_mutex);
7833 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7836 unsigned long flags;
7839 * We can't change the weight of the root cgroup.
7844 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7846 mutex_lock(&shares_mutex);
7847 if (tg->shares == shares)
7850 tg->shares = shares;
7851 for_each_possible_cpu(i) {
7852 struct rq *rq = cpu_rq(i);
7853 struct sched_entity *se;
7856 /* Propagate contribution to hierarchy */
7857 raw_spin_lock_irqsave(&rq->lock, flags);
7859 /* Possible calls to update_curr() need rq clock */
7860 update_rq_clock(rq);
7861 for_each_sched_entity(se)
7862 update_cfs_shares(group_cfs_rq(se));
7863 raw_spin_unlock_irqrestore(&rq->lock, flags);
7867 mutex_unlock(&shares_mutex);
7870 #else /* CONFIG_FAIR_GROUP_SCHED */
7872 void free_fair_sched_group(struct task_group *tg) { }
7874 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7879 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7881 #endif /* CONFIG_FAIR_GROUP_SCHED */
7884 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7886 struct sched_entity *se = &task->se;
7887 unsigned int rr_interval = 0;
7890 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7893 if (rq->cfs.load.weight)
7894 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7900 * All the scheduling class methods:
7902 const struct sched_class fair_sched_class = {
7903 .next = &idle_sched_class,
7904 .enqueue_task = enqueue_task_fair,
7905 .dequeue_task = dequeue_task_fair,
7906 .yield_task = yield_task_fair,
7907 .yield_to_task = yield_to_task_fair,
7909 .check_preempt_curr = check_preempt_wakeup,
7911 .pick_next_task = pick_next_task_fair,
7912 .put_prev_task = put_prev_task_fair,
7915 .select_task_rq = select_task_rq_fair,
7916 .migrate_task_rq = migrate_task_rq_fair,
7918 .rq_online = rq_online_fair,
7919 .rq_offline = rq_offline_fair,
7921 .task_waking = task_waking_fair,
7924 .set_curr_task = set_curr_task_fair,
7925 .task_tick = task_tick_fair,
7926 .task_fork = task_fork_fair,
7928 .prio_changed = prio_changed_fair,
7929 .switched_from = switched_from_fair,
7930 .switched_to = switched_to_fair,
7932 .get_rr_interval = get_rr_interval_fair,
7934 #ifdef CONFIG_FAIR_GROUP_SCHED
7935 .task_move_group = task_move_group_fair,
7939 #ifdef CONFIG_SCHED_DEBUG
7940 void print_cfs_stats(struct seq_file *m, int cpu)
7942 struct cfs_rq *cfs_rq;
7945 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7946 print_cfs_rq(m, cpu, cfs_rq);
7951 __init void init_sched_fair_class(void)
7954 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7956 #ifdef CONFIG_NO_HZ_COMMON
7957 nohz.next_balance = jiffies;
7958 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7959 cpu_notifier(sched_ilb_notifier, 0);