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 int select_idle_sibling(struct task_struct *p, int cpu);
669 static unsigned long task_h_load(struct task_struct *p);
671 static inline void __update_task_entity_contrib(struct sched_entity *se);
673 /* Give new task start runnable values to heavy its load in infant time */
674 void init_task_runnable_average(struct task_struct *p)
678 p->se.avg.decay_count = 0;
679 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
680 p->se.avg.runnable_avg_sum = slice;
681 p->se.avg.runnable_avg_period = slice;
682 __update_task_entity_contrib(&p->se);
685 void init_task_runnable_average(struct task_struct *p)
691 * Update the current task's runtime statistics.
693 static void update_curr(struct cfs_rq *cfs_rq)
695 struct sched_entity *curr = cfs_rq->curr;
696 u64 now = rq_clock_task(rq_of(cfs_rq));
702 delta_exec = now - curr->exec_start;
703 if (unlikely((s64)delta_exec <= 0))
706 curr->exec_start = now;
708 schedstat_set(curr->statistics.exec_max,
709 max(delta_exec, curr->statistics.exec_max));
711 curr->sum_exec_runtime += delta_exec;
712 schedstat_add(cfs_rq, exec_clock, delta_exec);
714 curr->vruntime += calc_delta_fair(delta_exec, curr);
715 update_min_vruntime(cfs_rq);
717 if (entity_is_task(curr)) {
718 struct task_struct *curtask = task_of(curr);
720 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
721 cpuacct_charge(curtask, delta_exec);
722 account_group_exec_runtime(curtask, delta_exec);
725 account_cfs_rq_runtime(cfs_rq, delta_exec);
729 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
731 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
735 * Task is being enqueued - update stats:
737 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
740 * Are we enqueueing a waiting task? (for current tasks
741 * a dequeue/enqueue event is a NOP)
743 if (se != cfs_rq->curr)
744 update_stats_wait_start(cfs_rq, se);
748 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
750 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
751 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
752 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
753 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
754 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
755 #ifdef CONFIG_SCHEDSTATS
756 if (entity_is_task(se)) {
757 trace_sched_stat_wait(task_of(se),
758 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
761 schedstat_set(se->statistics.wait_start, 0);
765 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
768 * Mark the end of the wait period if dequeueing a
771 if (se != cfs_rq->curr)
772 update_stats_wait_end(cfs_rq, se);
776 * We are picking a new current task - update its stats:
779 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
782 * We are starting a new run period:
784 se->exec_start = rq_clock_task(rq_of(cfs_rq));
787 /**************************************************
788 * Scheduling class queueing methods:
791 #ifdef CONFIG_NUMA_BALANCING
793 * Approximate time to scan a full NUMA task in ms. The task scan period is
794 * calculated based on the tasks virtual memory size and
795 * numa_balancing_scan_size.
797 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
798 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
800 /* Portion of address space to scan in MB */
801 unsigned int sysctl_numa_balancing_scan_size = 256;
803 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
804 unsigned int sysctl_numa_balancing_scan_delay = 1000;
806 static unsigned int task_nr_scan_windows(struct task_struct *p)
808 unsigned long rss = 0;
809 unsigned long nr_scan_pages;
812 * Calculations based on RSS as non-present and empty pages are skipped
813 * by the PTE scanner and NUMA hinting faults should be trapped based
816 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
817 rss = get_mm_rss(p->mm);
821 rss = round_up(rss, nr_scan_pages);
822 return rss / nr_scan_pages;
825 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
826 #define MAX_SCAN_WINDOW 2560
828 static unsigned int task_scan_min(struct task_struct *p)
830 unsigned int scan, floor;
831 unsigned int windows = 1;
833 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
834 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
835 floor = 1000 / windows;
837 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
838 return max_t(unsigned int, floor, scan);
841 static unsigned int task_scan_max(struct task_struct *p)
843 unsigned int smin = task_scan_min(p);
846 /* Watch for min being lower than max due to floor calculations */
847 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
848 return max(smin, smax);
851 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
853 rq->nr_numa_running += (p->numa_preferred_nid != -1);
854 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
857 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
859 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
860 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
866 spinlock_t lock; /* nr_tasks, tasks */
869 struct list_head task_list;
872 nodemask_t active_nodes;
873 unsigned long total_faults;
875 * Faults_cpu is used to decide whether memory should move
876 * towards the CPU. As a consequence, these stats are weighted
877 * more by CPU use than by memory faults.
879 unsigned long *faults_cpu;
880 unsigned long faults[0];
883 /* Shared or private faults. */
884 #define NR_NUMA_HINT_FAULT_TYPES 2
886 /* Memory and CPU locality */
887 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
889 /* Averaged statistics, and temporary buffers. */
890 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
892 pid_t task_numa_group_id(struct task_struct *p)
894 return p->numa_group ? p->numa_group->gid : 0;
897 static inline int task_faults_idx(int nid, int priv)
899 return NR_NUMA_HINT_FAULT_TYPES * nid + priv;
902 static inline unsigned long task_faults(struct task_struct *p, int nid)
904 if (!p->numa_faults_memory)
907 return p->numa_faults_memory[task_faults_idx(nid, 0)] +
908 p->numa_faults_memory[task_faults_idx(nid, 1)];
911 static inline unsigned long group_faults(struct task_struct *p, int nid)
916 return p->numa_group->faults[task_faults_idx(nid, 0)] +
917 p->numa_group->faults[task_faults_idx(nid, 1)];
920 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
922 return group->faults_cpu[task_faults_idx(nid, 0)] +
923 group->faults_cpu[task_faults_idx(nid, 1)];
927 * These return the fraction of accesses done by a particular task, or
928 * task group, on a particular numa node. The group weight is given a
929 * larger multiplier, in order to group tasks together that are almost
930 * evenly spread out between numa nodes.
932 static inline unsigned long task_weight(struct task_struct *p, int nid)
934 unsigned long total_faults;
936 if (!p->numa_faults_memory)
939 total_faults = p->total_numa_faults;
944 return 1000 * task_faults(p, nid) / total_faults;
947 static inline unsigned long group_weight(struct task_struct *p, int nid)
949 if (!p->numa_group || !p->numa_group->total_faults)
952 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
955 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
956 int src_nid, int dst_cpu)
958 struct numa_group *ng = p->numa_group;
959 int dst_nid = cpu_to_node(dst_cpu);
960 int last_cpupid, this_cpupid;
962 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
965 * Multi-stage node selection is used in conjunction with a periodic
966 * migration fault to build a temporal task<->page relation. By using
967 * a two-stage filter we remove short/unlikely relations.
969 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
970 * a task's usage of a particular page (n_p) per total usage of this
971 * page (n_t) (in a given time-span) to a probability.
973 * Our periodic faults will sample this probability and getting the
974 * same result twice in a row, given these samples are fully
975 * independent, is then given by P(n)^2, provided our sample period
976 * is sufficiently short compared to the usage pattern.
978 * This quadric squishes small probabilities, making it less likely we
979 * act on an unlikely task<->page relation.
981 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
982 if (!cpupid_pid_unset(last_cpupid) &&
983 cpupid_to_nid(last_cpupid) != dst_nid)
986 /* Always allow migrate on private faults */
987 if (cpupid_match_pid(p, last_cpupid))
990 /* A shared fault, but p->numa_group has not been set up yet. */
995 * Do not migrate if the destination is not a node that
996 * is actively used by this numa group.
998 if (!node_isset(dst_nid, ng->active_nodes))
1002 * Source is a node that is not actively used by this
1003 * numa group, while the destination is. Migrate.
1005 if (!node_isset(src_nid, ng->active_nodes))
1009 * Both source and destination are nodes in active
1010 * use by this numa group. Maximize memory bandwidth
1011 * by migrating from more heavily used groups, to less
1012 * heavily used ones, spreading the load around.
1013 * Use a 1/4 hysteresis to avoid spurious page movement.
1015 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1018 static unsigned long weighted_cpuload(const int cpu);
1019 static unsigned long source_load(int cpu, int type);
1020 static unsigned long target_load(int cpu, int type);
1021 static unsigned long capacity_of(int cpu);
1022 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1024 /* Cached statistics for all CPUs within a node */
1026 unsigned long nr_running;
1029 /* Total compute capacity of CPUs on a node */
1030 unsigned long compute_capacity;
1032 /* Approximate capacity in terms of runnable tasks on a node */
1033 unsigned long task_capacity;
1034 int has_free_capacity;
1038 * XXX borrowed from update_sg_lb_stats
1040 static void update_numa_stats(struct numa_stats *ns, int nid)
1042 int smt, cpu, cpus = 0;
1043 unsigned long capacity;
1045 memset(ns, 0, sizeof(*ns));
1046 for_each_cpu(cpu, cpumask_of_node(nid)) {
1047 struct rq *rq = cpu_rq(cpu);
1049 ns->nr_running += rq->nr_running;
1050 ns->load += weighted_cpuload(cpu);
1051 ns->compute_capacity += capacity_of(cpu);
1057 * If we raced with hotplug and there are no CPUs left in our mask
1058 * the @ns structure is NULL'ed and task_numa_compare() will
1059 * not find this node attractive.
1061 * We'll either bail at !has_free_capacity, or we'll detect a huge
1062 * imbalance and bail there.
1067 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1068 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1069 capacity = cpus / smt; /* cores */
1071 ns->task_capacity = min_t(unsigned, capacity,
1072 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1073 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1076 struct task_numa_env {
1077 struct task_struct *p;
1079 int src_cpu, src_nid;
1080 int dst_cpu, dst_nid;
1082 struct numa_stats src_stats, dst_stats;
1086 struct task_struct *best_task;
1091 static void task_numa_assign(struct task_numa_env *env,
1092 struct task_struct *p, long imp)
1095 put_task_struct(env->best_task);
1100 env->best_imp = imp;
1101 env->best_cpu = env->dst_cpu;
1104 static bool load_too_imbalanced(long src_load, long dst_load,
1105 struct task_numa_env *env)
1108 long orig_src_load, orig_dst_load;
1109 long src_capacity, dst_capacity;
1112 * The load is corrected for the CPU capacity available on each node.
1115 * ------------ vs ---------
1116 * src_capacity dst_capacity
1118 src_capacity = env->src_stats.compute_capacity;
1119 dst_capacity = env->dst_stats.compute_capacity;
1121 /* We care about the slope of the imbalance, not the direction. */
1122 if (dst_load < src_load)
1123 swap(dst_load, src_load);
1125 /* Is the difference below the threshold? */
1126 imb = dst_load * src_capacity * 100 -
1127 src_load * dst_capacity * env->imbalance_pct;
1132 * The imbalance is above the allowed threshold.
1133 * Compare it with the old imbalance.
1135 orig_src_load = env->src_stats.load;
1136 orig_dst_load = env->dst_stats.load;
1138 if (orig_dst_load < orig_src_load)
1139 swap(orig_dst_load, orig_src_load);
1141 old_imb = orig_dst_load * src_capacity * 100 -
1142 orig_src_load * dst_capacity * env->imbalance_pct;
1144 /* Would this change make things worse? */
1145 return (imb > old_imb);
1149 * This checks if the overall compute and NUMA accesses of the system would
1150 * be improved if the source tasks was migrated to the target dst_cpu taking
1151 * into account that it might be best if task running on the dst_cpu should
1152 * be exchanged with the source task
1154 static void task_numa_compare(struct task_numa_env *env,
1155 long taskimp, long groupimp)
1157 struct rq *src_rq = cpu_rq(env->src_cpu);
1158 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1159 struct task_struct *cur;
1160 long src_load, dst_load;
1162 long imp = env->p->numa_group ? groupimp : taskimp;
1166 cur = ACCESS_ONCE(dst_rq->curr);
1167 if (cur->pid == 0) /* idle */
1171 * "imp" is the fault differential for the source task between the
1172 * source and destination node. Calculate the total differential for
1173 * the source task and potential destination task. The more negative
1174 * the value is, the more rmeote accesses that would be expected to
1175 * be incurred if the tasks were swapped.
1178 /* Skip this swap candidate if cannot move to the source cpu */
1179 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1183 * If dst and source tasks are in the same NUMA group, or not
1184 * in any group then look only at task weights.
1186 if (cur->numa_group == env->p->numa_group) {
1187 imp = taskimp + task_weight(cur, env->src_nid) -
1188 task_weight(cur, env->dst_nid);
1190 * Add some hysteresis to prevent swapping the
1191 * tasks within a group over tiny differences.
1193 if (cur->numa_group)
1197 * Compare the group weights. If a task is all by
1198 * itself (not part of a group), use the task weight
1201 if (cur->numa_group)
1202 imp += group_weight(cur, env->src_nid) -
1203 group_weight(cur, env->dst_nid);
1205 imp += task_weight(cur, env->src_nid) -
1206 task_weight(cur, env->dst_nid);
1210 if (imp <= env->best_imp && moveimp <= env->best_imp)
1214 /* Is there capacity at our destination? */
1215 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1216 !env->dst_stats.has_free_capacity)
1222 /* Balance doesn't matter much if we're running a task per cpu */
1223 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1224 dst_rq->nr_running == 1)
1228 * In the overloaded case, try and keep the load balanced.
1231 load = task_h_load(env->p);
1232 dst_load = env->dst_stats.load + load;
1233 src_load = env->src_stats.load - load;
1235 if (moveimp > imp && moveimp > env->best_imp) {
1237 * If the improvement from just moving env->p direction is
1238 * better than swapping tasks around, check if a move is
1239 * possible. Store a slightly smaller score than moveimp,
1240 * so an actually idle CPU will win.
1242 if (!load_too_imbalanced(src_load, dst_load, env)) {
1249 if (imp <= env->best_imp)
1253 load = task_h_load(cur);
1258 if (load_too_imbalanced(src_load, dst_load, env))
1262 * One idle CPU per node is evaluated for a task numa move.
1263 * Call select_idle_sibling to maybe find a better one.
1266 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1269 task_numa_assign(env, cur, imp);
1274 static void task_numa_find_cpu(struct task_numa_env *env,
1275 long taskimp, long groupimp)
1279 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1280 /* Skip this CPU if the source task cannot migrate */
1281 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1285 task_numa_compare(env, taskimp, groupimp);
1289 static int task_numa_migrate(struct task_struct *p)
1291 struct task_numa_env env = {
1294 .src_cpu = task_cpu(p),
1295 .src_nid = task_node(p),
1297 .imbalance_pct = 112,
1303 struct sched_domain *sd;
1304 unsigned long taskweight, groupweight;
1306 long taskimp, groupimp;
1309 * Pick the lowest SD_NUMA domain, as that would have the smallest
1310 * imbalance and would be the first to start moving tasks about.
1312 * And we want to avoid any moving of tasks about, as that would create
1313 * random movement of tasks -- counter the numa conditions we're trying
1317 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1319 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1323 * Cpusets can break the scheduler domain tree into smaller
1324 * balance domains, some of which do not cross NUMA boundaries.
1325 * Tasks that are "trapped" in such domains cannot be migrated
1326 * elsewhere, so there is no point in (re)trying.
1328 if (unlikely(!sd)) {
1329 p->numa_preferred_nid = task_node(p);
1333 taskweight = task_weight(p, env.src_nid);
1334 groupweight = group_weight(p, env.src_nid);
1335 update_numa_stats(&env.src_stats, env.src_nid);
1336 env.dst_nid = p->numa_preferred_nid;
1337 taskimp = task_weight(p, env.dst_nid) - taskweight;
1338 groupimp = group_weight(p, env.dst_nid) - groupweight;
1339 update_numa_stats(&env.dst_stats, env.dst_nid);
1341 /* Try to find a spot on the preferred nid. */
1342 task_numa_find_cpu(&env, taskimp, groupimp);
1344 /* No space available on the preferred nid. Look elsewhere. */
1345 if (env.best_cpu == -1) {
1346 for_each_online_node(nid) {
1347 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1350 /* Only consider nodes where both task and groups benefit */
1351 taskimp = task_weight(p, nid) - taskweight;
1352 groupimp = group_weight(p, nid) - groupweight;
1353 if (taskimp < 0 && groupimp < 0)
1357 update_numa_stats(&env.dst_stats, env.dst_nid);
1358 task_numa_find_cpu(&env, taskimp, groupimp);
1363 * If the task is part of a workload that spans multiple NUMA nodes,
1364 * and is migrating into one of the workload's active nodes, remember
1365 * this node as the task's preferred numa node, so the workload can
1367 * A task that migrated to a second choice node will be better off
1368 * trying for a better one later. Do not set the preferred node here.
1370 if (p->numa_group) {
1371 if (env.best_cpu == -1)
1376 if (node_isset(nid, p->numa_group->active_nodes))
1377 sched_setnuma(p, env.dst_nid);
1380 /* No better CPU than the current one was found. */
1381 if (env.best_cpu == -1)
1385 * Reset the scan period if the task is being rescheduled on an
1386 * alternative node to recheck if the tasks is now properly placed.
1388 p->numa_scan_period = task_scan_min(p);
1390 if (env.best_task == NULL) {
1391 ret = migrate_task_to(p, env.best_cpu);
1393 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1397 ret = migrate_swap(p, env.best_task);
1399 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1400 put_task_struct(env.best_task);
1404 /* Attempt to migrate a task to a CPU on the preferred node. */
1405 static void numa_migrate_preferred(struct task_struct *p)
1407 unsigned long interval = HZ;
1409 /* This task has no NUMA fault statistics yet */
1410 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1413 /* Periodically retry migrating the task to the preferred node */
1414 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1415 p->numa_migrate_retry = jiffies + interval;
1417 /* Success if task is already running on preferred CPU */
1418 if (task_node(p) == p->numa_preferred_nid)
1421 /* Otherwise, try migrate to a CPU on the preferred node */
1422 task_numa_migrate(p);
1426 * Find the nodes on which the workload is actively running. We do this by
1427 * tracking the nodes from which NUMA hinting faults are triggered. This can
1428 * be different from the set of nodes where the workload's memory is currently
1431 * The bitmask is used to make smarter decisions on when to do NUMA page
1432 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1433 * are added when they cause over 6/16 of the maximum number of faults, but
1434 * only removed when they drop below 3/16.
1436 static void update_numa_active_node_mask(struct numa_group *numa_group)
1438 unsigned long faults, max_faults = 0;
1441 for_each_online_node(nid) {
1442 faults = group_faults_cpu(numa_group, nid);
1443 if (faults > max_faults)
1444 max_faults = faults;
1447 for_each_online_node(nid) {
1448 faults = group_faults_cpu(numa_group, nid);
1449 if (!node_isset(nid, numa_group->active_nodes)) {
1450 if (faults > max_faults * 6 / 16)
1451 node_set(nid, numa_group->active_nodes);
1452 } else if (faults < max_faults * 3 / 16)
1453 node_clear(nid, numa_group->active_nodes);
1458 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1459 * increments. The more local the fault statistics are, the higher the scan
1460 * period will be for the next scan window. If local/(local+remote) ratio is
1461 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1462 * the scan period will decrease. Aim for 70% local accesses.
1464 #define NUMA_PERIOD_SLOTS 10
1465 #define NUMA_PERIOD_THRESHOLD 7
1468 * Increase the scan period (slow down scanning) if the majority of
1469 * our memory is already on our local node, or if the majority of
1470 * the page accesses are shared with other processes.
1471 * Otherwise, decrease the scan period.
1473 static void update_task_scan_period(struct task_struct *p,
1474 unsigned long shared, unsigned long private)
1476 unsigned int period_slot;
1480 unsigned long remote = p->numa_faults_locality[0];
1481 unsigned long local = p->numa_faults_locality[1];
1484 * If there were no record hinting faults then either the task is
1485 * completely idle or all activity is areas that are not of interest
1486 * to automatic numa balancing. Scan slower
1488 if (local + shared == 0) {
1489 p->numa_scan_period = min(p->numa_scan_period_max,
1490 p->numa_scan_period << 1);
1492 p->mm->numa_next_scan = jiffies +
1493 msecs_to_jiffies(p->numa_scan_period);
1499 * Prepare to scale scan period relative to the current period.
1500 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1501 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1502 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1504 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1505 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1506 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1507 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1510 diff = slot * period_slot;
1512 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1515 * Scale scan rate increases based on sharing. There is an
1516 * inverse relationship between the degree of sharing and
1517 * the adjustment made to the scanning period. Broadly
1518 * speaking the intent is that there is little point
1519 * scanning faster if shared accesses dominate as it may
1520 * simply bounce migrations uselessly
1522 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1523 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1526 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1527 task_scan_min(p), task_scan_max(p));
1528 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1532 * Get the fraction of time the task has been running since the last
1533 * NUMA placement cycle. The scheduler keeps similar statistics, but
1534 * decays those on a 32ms period, which is orders of magnitude off
1535 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1536 * stats only if the task is so new there are no NUMA statistics yet.
1538 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1540 u64 runtime, delta, now;
1541 /* Use the start of this time slice to avoid calculations. */
1542 now = p->se.exec_start;
1543 runtime = p->se.sum_exec_runtime;
1545 if (p->last_task_numa_placement) {
1546 delta = runtime - p->last_sum_exec_runtime;
1547 *period = now - p->last_task_numa_placement;
1549 delta = p->se.avg.runnable_avg_sum;
1550 *period = p->se.avg.runnable_avg_period;
1553 p->last_sum_exec_runtime = runtime;
1554 p->last_task_numa_placement = now;
1559 static void task_numa_placement(struct task_struct *p)
1561 int seq, nid, max_nid = -1, max_group_nid = -1;
1562 unsigned long max_faults = 0, max_group_faults = 0;
1563 unsigned long fault_types[2] = { 0, 0 };
1564 unsigned long total_faults;
1565 u64 runtime, period;
1566 spinlock_t *group_lock = NULL;
1568 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1569 if (p->numa_scan_seq == seq)
1571 p->numa_scan_seq = seq;
1572 p->numa_scan_period_max = task_scan_max(p);
1574 total_faults = p->numa_faults_locality[0] +
1575 p->numa_faults_locality[1];
1576 runtime = numa_get_avg_runtime(p, &period);
1578 /* If the task is part of a group prevent parallel updates to group stats */
1579 if (p->numa_group) {
1580 group_lock = &p->numa_group->lock;
1581 spin_lock_irq(group_lock);
1584 /* Find the node with the highest number of faults */
1585 for_each_online_node(nid) {
1586 unsigned long faults = 0, group_faults = 0;
1589 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1590 long diff, f_diff, f_weight;
1592 i = task_faults_idx(nid, priv);
1594 /* Decay existing window, copy faults since last scan */
1595 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1596 fault_types[priv] += p->numa_faults_buffer_memory[i];
1597 p->numa_faults_buffer_memory[i] = 0;
1600 * Normalize the faults_from, so all tasks in a group
1601 * count according to CPU use, instead of by the raw
1602 * number of faults. Tasks with little runtime have
1603 * little over-all impact on throughput, and thus their
1604 * faults are less important.
1606 f_weight = div64_u64(runtime << 16, period + 1);
1607 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1609 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1610 p->numa_faults_buffer_cpu[i] = 0;
1612 p->numa_faults_memory[i] += diff;
1613 p->numa_faults_cpu[i] += f_diff;
1614 faults += p->numa_faults_memory[i];
1615 p->total_numa_faults += diff;
1616 if (p->numa_group) {
1617 /* safe because we can only change our own group */
1618 p->numa_group->faults[i] += diff;
1619 p->numa_group->faults_cpu[i] += f_diff;
1620 p->numa_group->total_faults += diff;
1621 group_faults += p->numa_group->faults[i];
1625 if (faults > max_faults) {
1626 max_faults = faults;
1630 if (group_faults > max_group_faults) {
1631 max_group_faults = group_faults;
1632 max_group_nid = nid;
1636 update_task_scan_period(p, fault_types[0], fault_types[1]);
1638 if (p->numa_group) {
1639 update_numa_active_node_mask(p->numa_group);
1640 spin_unlock_irq(group_lock);
1641 max_nid = max_group_nid;
1645 /* Set the new preferred node */
1646 if (max_nid != p->numa_preferred_nid)
1647 sched_setnuma(p, max_nid);
1649 if (task_node(p) != p->numa_preferred_nid)
1650 numa_migrate_preferred(p);
1654 static inline int get_numa_group(struct numa_group *grp)
1656 return atomic_inc_not_zero(&grp->refcount);
1659 static inline void put_numa_group(struct numa_group *grp)
1661 if (atomic_dec_and_test(&grp->refcount))
1662 kfree_rcu(grp, rcu);
1665 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1668 struct numa_group *grp, *my_grp;
1669 struct task_struct *tsk;
1671 int cpu = cpupid_to_cpu(cpupid);
1674 if (unlikely(!p->numa_group)) {
1675 unsigned int size = sizeof(struct numa_group) +
1676 4*nr_node_ids*sizeof(unsigned long);
1678 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1682 atomic_set(&grp->refcount, 1);
1683 spin_lock_init(&grp->lock);
1684 INIT_LIST_HEAD(&grp->task_list);
1686 /* Second half of the array tracks nids where faults happen */
1687 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1690 node_set(task_node(current), grp->active_nodes);
1692 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1693 grp->faults[i] = p->numa_faults_memory[i];
1695 grp->total_faults = p->total_numa_faults;
1697 list_add(&p->numa_entry, &grp->task_list);
1699 rcu_assign_pointer(p->numa_group, grp);
1703 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1705 if (!cpupid_match_pid(tsk, cpupid))
1708 grp = rcu_dereference(tsk->numa_group);
1712 my_grp = p->numa_group;
1717 * Only join the other group if its bigger; if we're the bigger group,
1718 * the other task will join us.
1720 if (my_grp->nr_tasks > grp->nr_tasks)
1724 * Tie-break on the grp address.
1726 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1729 /* Always join threads in the same process. */
1730 if (tsk->mm == current->mm)
1733 /* Simple filter to avoid false positives due to PID collisions */
1734 if (flags & TNF_SHARED)
1737 /* Update priv based on whether false sharing was detected */
1740 if (join && !get_numa_group(grp))
1748 BUG_ON(irqs_disabled());
1749 double_lock_irq(&my_grp->lock, &grp->lock);
1751 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1752 my_grp->faults[i] -= p->numa_faults_memory[i];
1753 grp->faults[i] += p->numa_faults_memory[i];
1755 my_grp->total_faults -= p->total_numa_faults;
1756 grp->total_faults += p->total_numa_faults;
1758 list_move(&p->numa_entry, &grp->task_list);
1762 spin_unlock(&my_grp->lock);
1763 spin_unlock_irq(&grp->lock);
1765 rcu_assign_pointer(p->numa_group, grp);
1767 put_numa_group(my_grp);
1775 void task_numa_free(struct task_struct *p)
1777 struct numa_group *grp = p->numa_group;
1778 void *numa_faults = p->numa_faults_memory;
1779 unsigned long flags;
1783 spin_lock_irqsave(&grp->lock, flags);
1784 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1785 grp->faults[i] -= p->numa_faults_memory[i];
1786 grp->total_faults -= p->total_numa_faults;
1788 list_del(&p->numa_entry);
1790 spin_unlock_irqrestore(&grp->lock, flags);
1791 RCU_INIT_POINTER(p->numa_group, NULL);
1792 put_numa_group(grp);
1795 p->numa_faults_memory = NULL;
1796 p->numa_faults_buffer_memory = NULL;
1797 p->numa_faults_cpu= NULL;
1798 p->numa_faults_buffer_cpu = NULL;
1803 * Got a PROT_NONE fault for a page on @node.
1805 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1807 struct task_struct *p = current;
1808 bool migrated = flags & TNF_MIGRATED;
1809 int cpu_node = task_node(current);
1810 int local = !!(flags & TNF_FAULT_LOCAL);
1813 if (!numabalancing_enabled)
1816 /* for example, ksmd faulting in a user's mm */
1820 /* Do not worry about placement if exiting */
1821 if (p->state == TASK_DEAD)
1824 /* Allocate buffer to track faults on a per-node basis */
1825 if (unlikely(!p->numa_faults_memory)) {
1826 int size = sizeof(*p->numa_faults_memory) *
1827 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1829 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1830 if (!p->numa_faults_memory)
1833 BUG_ON(p->numa_faults_buffer_memory);
1835 * The averaged statistics, shared & private, memory & cpu,
1836 * occupy the first half of the array. The second half of the
1837 * array is for current counters, which are averaged into the
1838 * first set by task_numa_placement.
1840 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1841 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1842 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1843 p->total_numa_faults = 0;
1844 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1848 * First accesses are treated as private, otherwise consider accesses
1849 * to be private if the accessing pid has not changed
1851 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1854 priv = cpupid_match_pid(p, last_cpupid);
1855 if (!priv && !(flags & TNF_NO_GROUP))
1856 task_numa_group(p, last_cpupid, flags, &priv);
1860 * If a workload spans multiple NUMA nodes, a shared fault that
1861 * occurs wholly within the set of nodes that the workload is
1862 * actively using should be counted as local. This allows the
1863 * scan rate to slow down when a workload has settled down.
1865 if (!priv && !local && p->numa_group &&
1866 node_isset(cpu_node, p->numa_group->active_nodes) &&
1867 node_isset(mem_node, p->numa_group->active_nodes))
1870 task_numa_placement(p);
1873 * Retry task to preferred node migration periodically, in case it
1874 * case it previously failed, or the scheduler moved us.
1876 if (time_after(jiffies, p->numa_migrate_retry))
1877 numa_migrate_preferred(p);
1880 p->numa_pages_migrated += pages;
1882 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1883 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1884 p->numa_faults_locality[local] += pages;
1887 static void reset_ptenuma_scan(struct task_struct *p)
1889 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1890 p->mm->numa_scan_offset = 0;
1894 * The expensive part of numa migration is done from task_work context.
1895 * Triggered from task_tick_numa().
1897 void task_numa_work(struct callback_head *work)
1899 unsigned long migrate, next_scan, now = jiffies;
1900 struct task_struct *p = current;
1901 struct mm_struct *mm = p->mm;
1902 struct vm_area_struct *vma;
1903 unsigned long start, end;
1904 unsigned long nr_pte_updates = 0;
1907 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1909 work->next = work; /* protect against double add */
1911 * Who cares about NUMA placement when they're dying.
1913 * NOTE: make sure not to dereference p->mm before this check,
1914 * exit_task_work() happens _after_ exit_mm() so we could be called
1915 * without p->mm even though we still had it when we enqueued this
1918 if (p->flags & PF_EXITING)
1921 if (!mm->numa_next_scan) {
1922 mm->numa_next_scan = now +
1923 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1927 * Enforce maximal scan/migration frequency..
1929 migrate = mm->numa_next_scan;
1930 if (time_before(now, migrate))
1933 if (p->numa_scan_period == 0) {
1934 p->numa_scan_period_max = task_scan_max(p);
1935 p->numa_scan_period = task_scan_min(p);
1938 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1939 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1943 * Delay this task enough that another task of this mm will likely win
1944 * the next time around.
1946 p->node_stamp += 2 * TICK_NSEC;
1948 start = mm->numa_scan_offset;
1949 pages = sysctl_numa_balancing_scan_size;
1950 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1954 down_read(&mm->mmap_sem);
1955 vma = find_vma(mm, start);
1957 reset_ptenuma_scan(p);
1961 for (; vma; vma = vma->vm_next) {
1962 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1966 * Shared library pages mapped by multiple processes are not
1967 * migrated as it is expected they are cache replicated. Avoid
1968 * hinting faults in read-only file-backed mappings or the vdso
1969 * as migrating the pages will be of marginal benefit.
1972 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1976 * Skip inaccessible VMAs to avoid any confusion between
1977 * PROT_NONE and NUMA hinting ptes
1979 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1983 start = max(start, vma->vm_start);
1984 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1985 end = min(end, vma->vm_end);
1986 nr_pte_updates += change_prot_numa(vma, start, end);
1989 * Scan sysctl_numa_balancing_scan_size but ensure that
1990 * at least one PTE is updated so that unused virtual
1991 * address space is quickly skipped.
1994 pages -= (end - start) >> PAGE_SHIFT;
2001 } while (end != vma->vm_end);
2006 * It is possible to reach the end of the VMA list but the last few
2007 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2008 * would find the !migratable VMA on the next scan but not reset the
2009 * scanner to the start so check it now.
2012 mm->numa_scan_offset = start;
2014 reset_ptenuma_scan(p);
2015 up_read(&mm->mmap_sem);
2019 * Drive the periodic memory faults..
2021 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2023 struct callback_head *work = &curr->numa_work;
2027 * We don't care about NUMA placement if we don't have memory.
2029 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2033 * Using runtime rather than walltime has the dual advantage that
2034 * we (mostly) drive the selection from busy threads and that the
2035 * task needs to have done some actual work before we bother with
2038 now = curr->se.sum_exec_runtime;
2039 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2041 if (now - curr->node_stamp > period) {
2042 if (!curr->node_stamp)
2043 curr->numa_scan_period = task_scan_min(curr);
2044 curr->node_stamp += period;
2046 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2047 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2048 task_work_add(curr, work, true);
2053 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2057 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2061 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2064 #endif /* CONFIG_NUMA_BALANCING */
2067 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2069 update_load_add(&cfs_rq->load, se->load.weight);
2070 if (!parent_entity(se))
2071 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2073 if (entity_is_task(se)) {
2074 struct rq *rq = rq_of(cfs_rq);
2076 account_numa_enqueue(rq, task_of(se));
2077 list_add(&se->group_node, &rq->cfs_tasks);
2080 cfs_rq->nr_running++;
2084 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2086 update_load_sub(&cfs_rq->load, se->load.weight);
2087 if (!parent_entity(se))
2088 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2089 if (entity_is_task(se)) {
2090 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2091 list_del_init(&se->group_node);
2093 cfs_rq->nr_running--;
2096 #ifdef CONFIG_FAIR_GROUP_SCHED
2098 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2103 * Use this CPU's actual weight instead of the last load_contribution
2104 * to gain a more accurate current total weight. See
2105 * update_cfs_rq_load_contribution().
2107 tg_weight = atomic_long_read(&tg->load_avg);
2108 tg_weight -= cfs_rq->tg_load_contrib;
2109 tg_weight += cfs_rq->load.weight;
2114 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2116 long tg_weight, load, shares;
2118 tg_weight = calc_tg_weight(tg, cfs_rq);
2119 load = cfs_rq->load.weight;
2121 shares = (tg->shares * load);
2123 shares /= tg_weight;
2125 if (shares < MIN_SHARES)
2126 shares = MIN_SHARES;
2127 if (shares > tg->shares)
2128 shares = tg->shares;
2132 # else /* CONFIG_SMP */
2133 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2137 # endif /* CONFIG_SMP */
2138 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2139 unsigned long weight)
2142 /* commit outstanding execution time */
2143 if (cfs_rq->curr == se)
2144 update_curr(cfs_rq);
2145 account_entity_dequeue(cfs_rq, se);
2148 update_load_set(&se->load, weight);
2151 account_entity_enqueue(cfs_rq, se);
2154 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2156 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2158 struct task_group *tg;
2159 struct sched_entity *se;
2163 se = tg->se[cpu_of(rq_of(cfs_rq))];
2164 if (!se || throttled_hierarchy(cfs_rq))
2167 if (likely(se->load.weight == tg->shares))
2170 shares = calc_cfs_shares(cfs_rq, tg);
2172 reweight_entity(cfs_rq_of(se), se, shares);
2174 #else /* CONFIG_FAIR_GROUP_SCHED */
2175 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2178 #endif /* CONFIG_FAIR_GROUP_SCHED */
2182 * We choose a half-life close to 1 scheduling period.
2183 * Note: The tables below are dependent on this value.
2185 #define LOAD_AVG_PERIOD 32
2186 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2187 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2189 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2190 static const u32 runnable_avg_yN_inv[] = {
2191 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2192 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2193 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2194 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2195 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2196 0x85aac367, 0x82cd8698,
2200 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2201 * over-estimates when re-combining.
2203 static const u32 runnable_avg_yN_sum[] = {
2204 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2205 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2206 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2211 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2213 static __always_inline u64 decay_load(u64 val, u64 n)
2215 unsigned int local_n;
2219 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2222 /* after bounds checking we can collapse to 32-bit */
2226 * As y^PERIOD = 1/2, we can combine
2227 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2228 * With a look-up table which covers k^n (n<PERIOD)
2230 * To achieve constant time decay_load.
2232 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2233 val >>= local_n / LOAD_AVG_PERIOD;
2234 local_n %= LOAD_AVG_PERIOD;
2237 val *= runnable_avg_yN_inv[local_n];
2238 /* We don't use SRR here since we always want to round down. */
2243 * For updates fully spanning n periods, the contribution to runnable
2244 * average will be: \Sum 1024*y^n
2246 * We can compute this reasonably efficiently by combining:
2247 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2249 static u32 __compute_runnable_contrib(u64 n)
2253 if (likely(n <= LOAD_AVG_PERIOD))
2254 return runnable_avg_yN_sum[n];
2255 else if (unlikely(n >= LOAD_AVG_MAX_N))
2256 return LOAD_AVG_MAX;
2258 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2260 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2261 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2263 n -= LOAD_AVG_PERIOD;
2264 } while (n > LOAD_AVG_PERIOD);
2266 contrib = decay_load(contrib, n);
2267 return contrib + runnable_avg_yN_sum[n];
2271 * We can represent the historical contribution to runnable average as the
2272 * coefficients of a geometric series. To do this we sub-divide our runnable
2273 * history into segments of approximately 1ms (1024us); label the segment that
2274 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2276 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2278 * (now) (~1ms ago) (~2ms ago)
2280 * Let u_i denote the fraction of p_i that the entity was runnable.
2282 * We then designate the fractions u_i as our co-efficients, yielding the
2283 * following representation of historical load:
2284 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2286 * We choose y based on the with of a reasonably scheduling period, fixing:
2289 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2290 * approximately half as much as the contribution to load within the last ms
2293 * When a period "rolls over" and we have new u_0`, multiplying the previous
2294 * sum again by y is sufficient to update:
2295 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2296 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2298 static __always_inline int __update_entity_runnable_avg(u64 now,
2299 struct sched_avg *sa,
2303 u32 runnable_contrib;
2304 int delta_w, decayed = 0;
2306 delta = now - sa->last_runnable_update;
2308 * This should only happen when time goes backwards, which it
2309 * unfortunately does during sched clock init when we swap over to TSC.
2311 if ((s64)delta < 0) {
2312 sa->last_runnable_update = now;
2317 * Use 1024ns as the unit of measurement since it's a reasonable
2318 * approximation of 1us and fast to compute.
2323 sa->last_runnable_update = now;
2325 /* delta_w is the amount already accumulated against our next period */
2326 delta_w = sa->runnable_avg_period % 1024;
2327 if (delta + delta_w >= 1024) {
2328 /* period roll-over */
2332 * Now that we know we're crossing a period boundary, figure
2333 * out how much from delta we need to complete the current
2334 * period and accrue it.
2336 delta_w = 1024 - delta_w;
2338 sa->runnable_avg_sum += delta_w;
2339 sa->runnable_avg_period += delta_w;
2343 /* Figure out how many additional periods this update spans */
2344 periods = delta / 1024;
2347 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2349 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2352 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2353 runnable_contrib = __compute_runnable_contrib(periods);
2355 sa->runnable_avg_sum += runnable_contrib;
2356 sa->runnable_avg_period += runnable_contrib;
2359 /* Remainder of delta accrued against u_0` */
2361 sa->runnable_avg_sum += delta;
2362 sa->runnable_avg_period += delta;
2367 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2368 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2370 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2371 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2373 decays -= se->avg.decay_count;
2377 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2378 se->avg.decay_count = 0;
2383 #ifdef CONFIG_FAIR_GROUP_SCHED
2384 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2387 struct task_group *tg = cfs_rq->tg;
2390 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2391 tg_contrib -= cfs_rq->tg_load_contrib;
2396 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2397 atomic_long_add(tg_contrib, &tg->load_avg);
2398 cfs_rq->tg_load_contrib += tg_contrib;
2403 * Aggregate cfs_rq runnable averages into an equivalent task_group
2404 * representation for computing load contributions.
2406 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2407 struct cfs_rq *cfs_rq)
2409 struct task_group *tg = cfs_rq->tg;
2412 /* The fraction of a cpu used by this cfs_rq */
2413 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2414 sa->runnable_avg_period + 1);
2415 contrib -= cfs_rq->tg_runnable_contrib;
2417 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2418 atomic_add(contrib, &tg->runnable_avg);
2419 cfs_rq->tg_runnable_contrib += contrib;
2423 static inline void __update_group_entity_contrib(struct sched_entity *se)
2425 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2426 struct task_group *tg = cfs_rq->tg;
2431 contrib = cfs_rq->tg_load_contrib * tg->shares;
2432 se->avg.load_avg_contrib = div_u64(contrib,
2433 atomic_long_read(&tg->load_avg) + 1);
2436 * For group entities we need to compute a correction term in the case
2437 * that they are consuming <1 cpu so that we would contribute the same
2438 * load as a task of equal weight.
2440 * Explicitly co-ordinating this measurement would be expensive, but
2441 * fortunately the sum of each cpus contribution forms a usable
2442 * lower-bound on the true value.
2444 * Consider the aggregate of 2 contributions. Either they are disjoint
2445 * (and the sum represents true value) or they are disjoint and we are
2446 * understating by the aggregate of their overlap.
2448 * Extending this to N cpus, for a given overlap, the maximum amount we
2449 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2450 * cpus that overlap for this interval and w_i is the interval width.
2452 * On a small machine; the first term is well-bounded which bounds the
2453 * total error since w_i is a subset of the period. Whereas on a
2454 * larger machine, while this first term can be larger, if w_i is the
2455 * of consequential size guaranteed to see n_i*w_i quickly converge to
2456 * our upper bound of 1-cpu.
2458 runnable_avg = atomic_read(&tg->runnable_avg);
2459 if (runnable_avg < NICE_0_LOAD) {
2460 se->avg.load_avg_contrib *= runnable_avg;
2461 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2465 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2467 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2468 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2470 #else /* CONFIG_FAIR_GROUP_SCHED */
2471 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2472 int force_update) {}
2473 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2474 struct cfs_rq *cfs_rq) {}
2475 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2476 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2477 #endif /* CONFIG_FAIR_GROUP_SCHED */
2479 static inline void __update_task_entity_contrib(struct sched_entity *se)
2483 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2484 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2485 contrib /= (se->avg.runnable_avg_period + 1);
2486 se->avg.load_avg_contrib = scale_load(contrib);
2489 /* Compute the current contribution to load_avg by se, return any delta */
2490 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2492 long old_contrib = se->avg.load_avg_contrib;
2494 if (entity_is_task(se)) {
2495 __update_task_entity_contrib(se);
2497 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2498 __update_group_entity_contrib(se);
2501 return se->avg.load_avg_contrib - old_contrib;
2504 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2507 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2508 cfs_rq->blocked_load_avg -= load_contrib;
2510 cfs_rq->blocked_load_avg = 0;
2513 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2515 /* Update a sched_entity's runnable average */
2516 static inline void update_entity_load_avg(struct sched_entity *se,
2519 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2524 * For a group entity we need to use their owned cfs_rq_clock_task() in
2525 * case they are the parent of a throttled hierarchy.
2527 if (entity_is_task(se))
2528 now = cfs_rq_clock_task(cfs_rq);
2530 now = cfs_rq_clock_task(group_cfs_rq(se));
2532 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2535 contrib_delta = __update_entity_load_avg_contrib(se);
2541 cfs_rq->runnable_load_avg += contrib_delta;
2543 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2547 * Decay the load contributed by all blocked children and account this so that
2548 * their contribution may appropriately discounted when they wake up.
2550 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2552 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2555 decays = now - cfs_rq->last_decay;
2556 if (!decays && !force_update)
2559 if (atomic_long_read(&cfs_rq->removed_load)) {
2560 unsigned long removed_load;
2561 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2562 subtract_blocked_load_contrib(cfs_rq, removed_load);
2566 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2568 atomic64_add(decays, &cfs_rq->decay_counter);
2569 cfs_rq->last_decay = now;
2572 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2575 /* Add the load generated by se into cfs_rq's child load-average */
2576 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2577 struct sched_entity *se,
2581 * We track migrations using entity decay_count <= 0, on a wake-up
2582 * migration we use a negative decay count to track the remote decays
2583 * accumulated while sleeping.
2585 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2586 * are seen by enqueue_entity_load_avg() as a migration with an already
2587 * constructed load_avg_contrib.
2589 if (unlikely(se->avg.decay_count <= 0)) {
2590 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2591 if (se->avg.decay_count) {
2593 * In a wake-up migration we have to approximate the
2594 * time sleeping. This is because we can't synchronize
2595 * clock_task between the two cpus, and it is not
2596 * guaranteed to be read-safe. Instead, we can
2597 * approximate this using our carried decays, which are
2598 * explicitly atomically readable.
2600 se->avg.last_runnable_update -= (-se->avg.decay_count)
2602 update_entity_load_avg(se, 0);
2603 /* Indicate that we're now synchronized and on-rq */
2604 se->avg.decay_count = 0;
2608 __synchronize_entity_decay(se);
2611 /* migrated tasks did not contribute to our blocked load */
2613 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2614 update_entity_load_avg(se, 0);
2617 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2618 /* we force update consideration on load-balancer moves */
2619 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2623 * Remove se's load from this cfs_rq child load-average, if the entity is
2624 * transitioning to a blocked state we track its projected decay using
2627 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2628 struct sched_entity *se,
2631 update_entity_load_avg(se, 1);
2632 /* we force update consideration on load-balancer moves */
2633 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2635 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2637 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2638 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2639 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2643 * Update the rq's load with the elapsed running time before entering
2644 * idle. if the last scheduled task is not a CFS task, idle_enter will
2645 * be the only way to update the runnable statistic.
2647 void idle_enter_fair(struct rq *this_rq)
2649 update_rq_runnable_avg(this_rq, 1);
2653 * Update the rq's load with the elapsed idle time before a task is
2654 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2655 * be the only way to update the runnable statistic.
2657 void idle_exit_fair(struct rq *this_rq)
2659 update_rq_runnable_avg(this_rq, 0);
2662 static int idle_balance(struct rq *this_rq);
2664 #else /* CONFIG_SMP */
2666 static inline void update_entity_load_avg(struct sched_entity *se,
2667 int update_cfs_rq) {}
2668 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2669 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2670 struct sched_entity *se,
2672 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2673 struct sched_entity *se,
2675 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2676 int force_update) {}
2678 static inline int idle_balance(struct rq *rq)
2683 #endif /* CONFIG_SMP */
2685 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2687 #ifdef CONFIG_SCHEDSTATS
2688 struct task_struct *tsk = NULL;
2690 if (entity_is_task(se))
2693 if (se->statistics.sleep_start) {
2694 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2699 if (unlikely(delta > se->statistics.sleep_max))
2700 se->statistics.sleep_max = delta;
2702 se->statistics.sleep_start = 0;
2703 se->statistics.sum_sleep_runtime += delta;
2706 account_scheduler_latency(tsk, delta >> 10, 1);
2707 trace_sched_stat_sleep(tsk, delta);
2710 if (se->statistics.block_start) {
2711 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2716 if (unlikely(delta > se->statistics.block_max))
2717 se->statistics.block_max = delta;
2719 se->statistics.block_start = 0;
2720 se->statistics.sum_sleep_runtime += delta;
2723 if (tsk->in_iowait) {
2724 se->statistics.iowait_sum += delta;
2725 se->statistics.iowait_count++;
2726 trace_sched_stat_iowait(tsk, delta);
2729 trace_sched_stat_blocked(tsk, delta);
2732 * Blocking time is in units of nanosecs, so shift by
2733 * 20 to get a milliseconds-range estimation of the
2734 * amount of time that the task spent sleeping:
2736 if (unlikely(prof_on == SLEEP_PROFILING)) {
2737 profile_hits(SLEEP_PROFILING,
2738 (void *)get_wchan(tsk),
2741 account_scheduler_latency(tsk, delta >> 10, 0);
2747 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2749 #ifdef CONFIG_SCHED_DEBUG
2750 s64 d = se->vruntime - cfs_rq->min_vruntime;
2755 if (d > 3*sysctl_sched_latency)
2756 schedstat_inc(cfs_rq, nr_spread_over);
2761 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2763 u64 vruntime = cfs_rq->min_vruntime;
2766 * The 'current' period is already promised to the current tasks,
2767 * however the extra weight of the new task will slow them down a
2768 * little, place the new task so that it fits in the slot that
2769 * stays open at the end.
2771 if (initial && sched_feat(START_DEBIT))
2772 vruntime += sched_vslice(cfs_rq, se);
2774 /* sleeps up to a single latency don't count. */
2776 unsigned long thresh = sysctl_sched_latency;
2779 * Halve their sleep time's effect, to allow
2780 * for a gentler effect of sleepers:
2782 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2788 /* ensure we never gain time by being placed backwards. */
2789 se->vruntime = max_vruntime(se->vruntime, vruntime);
2792 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2795 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2798 * Update the normalized vruntime before updating min_vruntime
2799 * through calling update_curr().
2801 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2802 se->vruntime += cfs_rq->min_vruntime;
2805 * Update run-time statistics of the 'current'.
2807 update_curr(cfs_rq);
2808 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2809 account_entity_enqueue(cfs_rq, se);
2810 update_cfs_shares(cfs_rq);
2812 if (flags & ENQUEUE_WAKEUP) {
2813 place_entity(cfs_rq, se, 0);
2814 enqueue_sleeper(cfs_rq, se);
2817 update_stats_enqueue(cfs_rq, se);
2818 check_spread(cfs_rq, se);
2819 if (se != cfs_rq->curr)
2820 __enqueue_entity(cfs_rq, se);
2823 if (cfs_rq->nr_running == 1) {
2824 list_add_leaf_cfs_rq(cfs_rq);
2825 check_enqueue_throttle(cfs_rq);
2829 static void __clear_buddies_last(struct sched_entity *se)
2831 for_each_sched_entity(se) {
2832 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2833 if (cfs_rq->last != se)
2836 cfs_rq->last = NULL;
2840 static void __clear_buddies_next(struct sched_entity *se)
2842 for_each_sched_entity(se) {
2843 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2844 if (cfs_rq->next != se)
2847 cfs_rq->next = NULL;
2851 static void __clear_buddies_skip(struct sched_entity *se)
2853 for_each_sched_entity(se) {
2854 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2855 if (cfs_rq->skip != se)
2858 cfs_rq->skip = NULL;
2862 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2864 if (cfs_rq->last == se)
2865 __clear_buddies_last(se);
2867 if (cfs_rq->next == se)
2868 __clear_buddies_next(se);
2870 if (cfs_rq->skip == se)
2871 __clear_buddies_skip(se);
2874 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2877 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2880 * Update run-time statistics of the 'current'.
2882 update_curr(cfs_rq);
2883 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2885 update_stats_dequeue(cfs_rq, se);
2886 if (flags & DEQUEUE_SLEEP) {
2887 #ifdef CONFIG_SCHEDSTATS
2888 if (entity_is_task(se)) {
2889 struct task_struct *tsk = task_of(se);
2891 if (tsk->state & TASK_INTERRUPTIBLE)
2892 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2893 if (tsk->state & TASK_UNINTERRUPTIBLE)
2894 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2899 clear_buddies(cfs_rq, se);
2901 if (se != cfs_rq->curr)
2902 __dequeue_entity(cfs_rq, se);
2904 account_entity_dequeue(cfs_rq, se);
2907 * Normalize the entity after updating the min_vruntime because the
2908 * update can refer to the ->curr item and we need to reflect this
2909 * movement in our normalized position.
2911 if (!(flags & DEQUEUE_SLEEP))
2912 se->vruntime -= cfs_rq->min_vruntime;
2914 /* return excess runtime on last dequeue */
2915 return_cfs_rq_runtime(cfs_rq);
2917 update_min_vruntime(cfs_rq);
2918 update_cfs_shares(cfs_rq);
2922 * Preempt the current task with a newly woken task if needed:
2925 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2927 unsigned long ideal_runtime, delta_exec;
2928 struct sched_entity *se;
2931 ideal_runtime = sched_slice(cfs_rq, curr);
2932 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2933 if (delta_exec > ideal_runtime) {
2934 resched_curr(rq_of(cfs_rq));
2936 * The current task ran long enough, ensure it doesn't get
2937 * re-elected due to buddy favours.
2939 clear_buddies(cfs_rq, curr);
2944 * Ensure that a task that missed wakeup preemption by a
2945 * narrow margin doesn't have to wait for a full slice.
2946 * This also mitigates buddy induced latencies under load.
2948 if (delta_exec < sysctl_sched_min_granularity)
2951 se = __pick_first_entity(cfs_rq);
2952 delta = curr->vruntime - se->vruntime;
2957 if (delta > ideal_runtime)
2958 resched_curr(rq_of(cfs_rq));
2962 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2964 /* 'current' is not kept within the tree. */
2967 * Any task has to be enqueued before it get to execute on
2968 * a CPU. So account for the time it spent waiting on the
2971 update_stats_wait_end(cfs_rq, se);
2972 __dequeue_entity(cfs_rq, se);
2975 update_stats_curr_start(cfs_rq, se);
2977 #ifdef CONFIG_SCHEDSTATS
2979 * Track our maximum slice length, if the CPU's load is at
2980 * least twice that of our own weight (i.e. dont track it
2981 * when there are only lesser-weight tasks around):
2983 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2984 se->statistics.slice_max = max(se->statistics.slice_max,
2985 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2988 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2992 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2995 * Pick the next process, keeping these things in mind, in this order:
2996 * 1) keep things fair between processes/task groups
2997 * 2) pick the "next" process, since someone really wants that to run
2998 * 3) pick the "last" process, for cache locality
2999 * 4) do not run the "skip" process, if something else is available
3001 static struct sched_entity *
3002 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3004 struct sched_entity *left = __pick_first_entity(cfs_rq);
3005 struct sched_entity *se;
3008 * If curr is set we have to see if its left of the leftmost entity
3009 * still in the tree, provided there was anything in the tree at all.
3011 if (!left || (curr && entity_before(curr, left)))
3014 se = left; /* ideally we run the leftmost entity */
3017 * Avoid running the skip buddy, if running something else can
3018 * be done without getting too unfair.
3020 if (cfs_rq->skip == se) {
3021 struct sched_entity *second;
3024 second = __pick_first_entity(cfs_rq);
3026 second = __pick_next_entity(se);
3027 if (!second || (curr && entity_before(curr, second)))
3031 if (second && wakeup_preempt_entity(second, left) < 1)
3036 * Prefer last buddy, try to return the CPU to a preempted task.
3038 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3042 * Someone really wants this to run. If it's not unfair, run it.
3044 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3047 clear_buddies(cfs_rq, se);
3052 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3054 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3057 * If still on the runqueue then deactivate_task()
3058 * was not called and update_curr() has to be done:
3061 update_curr(cfs_rq);
3063 /* throttle cfs_rqs exceeding runtime */
3064 check_cfs_rq_runtime(cfs_rq);
3066 check_spread(cfs_rq, prev);
3068 update_stats_wait_start(cfs_rq, prev);
3069 /* Put 'current' back into the tree. */
3070 __enqueue_entity(cfs_rq, prev);
3071 /* in !on_rq case, update occurred at dequeue */
3072 update_entity_load_avg(prev, 1);
3074 cfs_rq->curr = NULL;
3078 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3081 * Update run-time statistics of the 'current'.
3083 update_curr(cfs_rq);
3086 * Ensure that runnable average is periodically updated.
3088 update_entity_load_avg(curr, 1);
3089 update_cfs_rq_blocked_load(cfs_rq, 1);
3090 update_cfs_shares(cfs_rq);
3092 #ifdef CONFIG_SCHED_HRTICK
3094 * queued ticks are scheduled to match the slice, so don't bother
3095 * validating it and just reschedule.
3098 resched_curr(rq_of(cfs_rq));
3102 * don't let the period tick interfere with the hrtick preemption
3104 if (!sched_feat(DOUBLE_TICK) &&
3105 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3109 if (cfs_rq->nr_running > 1)
3110 check_preempt_tick(cfs_rq, curr);
3114 /**************************************************
3115 * CFS bandwidth control machinery
3118 #ifdef CONFIG_CFS_BANDWIDTH
3120 #ifdef HAVE_JUMP_LABEL
3121 static struct static_key __cfs_bandwidth_used;
3123 static inline bool cfs_bandwidth_used(void)
3125 return static_key_false(&__cfs_bandwidth_used);
3128 void cfs_bandwidth_usage_inc(void)
3130 static_key_slow_inc(&__cfs_bandwidth_used);
3133 void cfs_bandwidth_usage_dec(void)
3135 static_key_slow_dec(&__cfs_bandwidth_used);
3137 #else /* HAVE_JUMP_LABEL */
3138 static bool cfs_bandwidth_used(void)
3143 void cfs_bandwidth_usage_inc(void) {}
3144 void cfs_bandwidth_usage_dec(void) {}
3145 #endif /* HAVE_JUMP_LABEL */
3148 * default period for cfs group bandwidth.
3149 * default: 0.1s, units: nanoseconds
3151 static inline u64 default_cfs_period(void)
3153 return 100000000ULL;
3156 static inline u64 sched_cfs_bandwidth_slice(void)
3158 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3162 * Replenish runtime according to assigned quota and update expiration time.
3163 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3164 * additional synchronization around rq->lock.
3166 * requires cfs_b->lock
3168 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3172 if (cfs_b->quota == RUNTIME_INF)
3175 now = sched_clock_cpu(smp_processor_id());
3176 cfs_b->runtime = cfs_b->quota;
3177 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3180 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3182 return &tg->cfs_bandwidth;
3185 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3186 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3188 if (unlikely(cfs_rq->throttle_count))
3189 return cfs_rq->throttled_clock_task;
3191 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3194 /* returns 0 on failure to allocate runtime */
3195 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3197 struct task_group *tg = cfs_rq->tg;
3198 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3199 u64 amount = 0, min_amount, expires;
3201 /* note: this is a positive sum as runtime_remaining <= 0 */
3202 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3204 raw_spin_lock(&cfs_b->lock);
3205 if (cfs_b->quota == RUNTIME_INF)
3206 amount = min_amount;
3209 * If the bandwidth pool has become inactive, then at least one
3210 * period must have elapsed since the last consumption.
3211 * Refresh the global state and ensure bandwidth timer becomes
3214 if (!cfs_b->timer_active) {
3215 __refill_cfs_bandwidth_runtime(cfs_b);
3216 __start_cfs_bandwidth(cfs_b, false);
3219 if (cfs_b->runtime > 0) {
3220 amount = min(cfs_b->runtime, min_amount);
3221 cfs_b->runtime -= amount;
3225 expires = cfs_b->runtime_expires;
3226 raw_spin_unlock(&cfs_b->lock);
3228 cfs_rq->runtime_remaining += amount;
3230 * we may have advanced our local expiration to account for allowed
3231 * spread between our sched_clock and the one on which runtime was
3234 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3235 cfs_rq->runtime_expires = expires;
3237 return cfs_rq->runtime_remaining > 0;
3241 * Note: This depends on the synchronization provided by sched_clock and the
3242 * fact that rq->clock snapshots this value.
3244 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3246 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3248 /* if the deadline is ahead of our clock, nothing to do */
3249 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3252 if (cfs_rq->runtime_remaining < 0)
3256 * If the local deadline has passed we have to consider the
3257 * possibility that our sched_clock is 'fast' and the global deadline
3258 * has not truly expired.
3260 * Fortunately we can check determine whether this the case by checking
3261 * whether the global deadline has advanced. It is valid to compare
3262 * cfs_b->runtime_expires without any locks since we only care about
3263 * exact equality, so a partial write will still work.
3266 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3267 /* extend local deadline, drift is bounded above by 2 ticks */
3268 cfs_rq->runtime_expires += TICK_NSEC;
3270 /* global deadline is ahead, expiration has passed */
3271 cfs_rq->runtime_remaining = 0;
3275 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3277 /* dock delta_exec before expiring quota (as it could span periods) */
3278 cfs_rq->runtime_remaining -= delta_exec;
3279 expire_cfs_rq_runtime(cfs_rq);
3281 if (likely(cfs_rq->runtime_remaining > 0))
3285 * if we're unable to extend our runtime we resched so that the active
3286 * hierarchy can be throttled
3288 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3289 resched_curr(rq_of(cfs_rq));
3292 static __always_inline
3293 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3295 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3298 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3301 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3303 return cfs_bandwidth_used() && cfs_rq->throttled;
3306 /* check whether cfs_rq, or any parent, is throttled */
3307 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3309 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3313 * Ensure that neither of the group entities corresponding to src_cpu or
3314 * dest_cpu are members of a throttled hierarchy when performing group
3315 * load-balance operations.
3317 static inline int throttled_lb_pair(struct task_group *tg,
3318 int src_cpu, int dest_cpu)
3320 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3322 src_cfs_rq = tg->cfs_rq[src_cpu];
3323 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3325 return throttled_hierarchy(src_cfs_rq) ||
3326 throttled_hierarchy(dest_cfs_rq);
3329 /* updated child weight may affect parent so we have to do this bottom up */
3330 static int tg_unthrottle_up(struct task_group *tg, void *data)
3332 struct rq *rq = data;
3333 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3335 cfs_rq->throttle_count--;
3337 if (!cfs_rq->throttle_count) {
3338 /* adjust cfs_rq_clock_task() */
3339 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3340 cfs_rq->throttled_clock_task;
3347 static int tg_throttle_down(struct task_group *tg, void *data)
3349 struct rq *rq = data;
3350 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3352 /* group is entering throttled state, stop time */
3353 if (!cfs_rq->throttle_count)
3354 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3355 cfs_rq->throttle_count++;
3360 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3362 struct rq *rq = rq_of(cfs_rq);
3363 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3364 struct sched_entity *se;
3365 long task_delta, dequeue = 1;
3367 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3369 /* freeze hierarchy runnable averages while throttled */
3371 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3374 task_delta = cfs_rq->h_nr_running;
3375 for_each_sched_entity(se) {
3376 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3377 /* throttled entity or throttle-on-deactivate */
3382 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3383 qcfs_rq->h_nr_running -= task_delta;
3385 if (qcfs_rq->load.weight)
3390 sub_nr_running(rq, task_delta);
3392 cfs_rq->throttled = 1;
3393 cfs_rq->throttled_clock = rq_clock(rq);
3394 raw_spin_lock(&cfs_b->lock);
3396 * Add to the _head_ of the list, so that an already-started
3397 * distribute_cfs_runtime will not see us
3399 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3400 if (!cfs_b->timer_active)
3401 __start_cfs_bandwidth(cfs_b, false);
3402 raw_spin_unlock(&cfs_b->lock);
3405 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3407 struct rq *rq = rq_of(cfs_rq);
3408 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3409 struct sched_entity *se;
3413 se = cfs_rq->tg->se[cpu_of(rq)];
3415 cfs_rq->throttled = 0;
3417 update_rq_clock(rq);
3419 raw_spin_lock(&cfs_b->lock);
3420 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3421 list_del_rcu(&cfs_rq->throttled_list);
3422 raw_spin_unlock(&cfs_b->lock);
3424 /* update hierarchical throttle state */
3425 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3427 if (!cfs_rq->load.weight)
3430 task_delta = cfs_rq->h_nr_running;
3431 for_each_sched_entity(se) {
3435 cfs_rq = cfs_rq_of(se);
3437 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3438 cfs_rq->h_nr_running += task_delta;
3440 if (cfs_rq_throttled(cfs_rq))
3445 add_nr_running(rq, task_delta);
3447 /* determine whether we need to wake up potentially idle cpu */
3448 if (rq->curr == rq->idle && rq->cfs.nr_running)
3452 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3453 u64 remaining, u64 expires)
3455 struct cfs_rq *cfs_rq;
3457 u64 starting_runtime = remaining;
3460 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3462 struct rq *rq = rq_of(cfs_rq);
3464 raw_spin_lock(&rq->lock);
3465 if (!cfs_rq_throttled(cfs_rq))
3468 runtime = -cfs_rq->runtime_remaining + 1;
3469 if (runtime > remaining)
3470 runtime = remaining;
3471 remaining -= runtime;
3473 cfs_rq->runtime_remaining += runtime;
3474 cfs_rq->runtime_expires = expires;
3476 /* we check whether we're throttled above */
3477 if (cfs_rq->runtime_remaining > 0)
3478 unthrottle_cfs_rq(cfs_rq);
3481 raw_spin_unlock(&rq->lock);
3488 return starting_runtime - remaining;
3492 * Responsible for refilling a task_group's bandwidth and unthrottling its
3493 * cfs_rqs as appropriate. If there has been no activity within the last
3494 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3495 * used to track this state.
3497 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3499 u64 runtime, runtime_expires;
3502 /* no need to continue the timer with no bandwidth constraint */
3503 if (cfs_b->quota == RUNTIME_INF)
3504 goto out_deactivate;
3506 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3507 cfs_b->nr_periods += overrun;
3510 * idle depends on !throttled (for the case of a large deficit), and if
3511 * we're going inactive then everything else can be deferred
3513 if (cfs_b->idle && !throttled)
3514 goto out_deactivate;
3517 * if we have relooped after returning idle once, we need to update our
3518 * status as actually running, so that other cpus doing
3519 * __start_cfs_bandwidth will stop trying to cancel us.
3521 cfs_b->timer_active = 1;
3523 __refill_cfs_bandwidth_runtime(cfs_b);
3526 /* mark as potentially idle for the upcoming period */
3531 /* account preceding periods in which throttling occurred */
3532 cfs_b->nr_throttled += overrun;
3534 runtime_expires = cfs_b->runtime_expires;
3537 * This check is repeated as we are holding onto the new bandwidth while
3538 * we unthrottle. This can potentially race with an unthrottled group
3539 * trying to acquire new bandwidth from the global pool. This can result
3540 * in us over-using our runtime if it is all used during this loop, but
3541 * only by limited amounts in that extreme case.
3543 while (throttled && cfs_b->runtime > 0) {
3544 runtime = cfs_b->runtime;
3545 raw_spin_unlock(&cfs_b->lock);
3546 /* we can't nest cfs_b->lock while distributing bandwidth */
3547 runtime = distribute_cfs_runtime(cfs_b, runtime,
3549 raw_spin_lock(&cfs_b->lock);
3551 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3553 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3557 * While we are ensured activity in the period following an
3558 * unthrottle, this also covers the case in which the new bandwidth is
3559 * insufficient to cover the existing bandwidth deficit. (Forcing the
3560 * timer to remain active while there are any throttled entities.)
3567 cfs_b->timer_active = 0;
3571 /* a cfs_rq won't donate quota below this amount */
3572 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3573 /* minimum remaining period time to redistribute slack quota */
3574 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3575 /* how long we wait to gather additional slack before distributing */
3576 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3579 * Are we near the end of the current quota period?
3581 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3582 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3583 * migrate_hrtimers, base is never cleared, so we are fine.
3585 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3587 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3590 /* if the call-back is running a quota refresh is already occurring */
3591 if (hrtimer_callback_running(refresh_timer))
3594 /* is a quota refresh about to occur? */
3595 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3596 if (remaining < min_expire)
3602 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3604 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3606 /* if there's a quota refresh soon don't bother with slack */
3607 if (runtime_refresh_within(cfs_b, min_left))
3610 start_bandwidth_timer(&cfs_b->slack_timer,
3611 ns_to_ktime(cfs_bandwidth_slack_period));
3614 /* we know any runtime found here is valid as update_curr() precedes return */
3615 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3617 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3618 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3620 if (slack_runtime <= 0)
3623 raw_spin_lock(&cfs_b->lock);
3624 if (cfs_b->quota != RUNTIME_INF &&
3625 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3626 cfs_b->runtime += slack_runtime;
3628 /* we are under rq->lock, defer unthrottling using a timer */
3629 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3630 !list_empty(&cfs_b->throttled_cfs_rq))
3631 start_cfs_slack_bandwidth(cfs_b);
3633 raw_spin_unlock(&cfs_b->lock);
3635 /* even if it's not valid for return we don't want to try again */
3636 cfs_rq->runtime_remaining -= slack_runtime;
3639 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3641 if (!cfs_bandwidth_used())
3644 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3647 __return_cfs_rq_runtime(cfs_rq);
3651 * This is done with a timer (instead of inline with bandwidth return) since
3652 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3654 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3656 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3659 /* confirm we're still not at a refresh boundary */
3660 raw_spin_lock(&cfs_b->lock);
3661 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3662 raw_spin_unlock(&cfs_b->lock);
3666 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3667 runtime = cfs_b->runtime;
3669 expires = cfs_b->runtime_expires;
3670 raw_spin_unlock(&cfs_b->lock);
3675 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3677 raw_spin_lock(&cfs_b->lock);
3678 if (expires == cfs_b->runtime_expires)
3679 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3680 raw_spin_unlock(&cfs_b->lock);
3684 * When a group wakes up we want to make sure that its quota is not already
3685 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3686 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3688 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3690 if (!cfs_bandwidth_used())
3693 /* an active group must be handled by the update_curr()->put() path */
3694 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3697 /* ensure the group is not already throttled */
3698 if (cfs_rq_throttled(cfs_rq))
3701 /* update runtime allocation */
3702 account_cfs_rq_runtime(cfs_rq, 0);
3703 if (cfs_rq->runtime_remaining <= 0)
3704 throttle_cfs_rq(cfs_rq);
3707 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3708 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3710 if (!cfs_bandwidth_used())
3713 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3717 * it's possible for a throttled entity to be forced into a running
3718 * state (e.g. set_curr_task), in this case we're finished.
3720 if (cfs_rq_throttled(cfs_rq))
3723 throttle_cfs_rq(cfs_rq);
3727 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3729 struct cfs_bandwidth *cfs_b =
3730 container_of(timer, struct cfs_bandwidth, slack_timer);
3731 do_sched_cfs_slack_timer(cfs_b);
3733 return HRTIMER_NORESTART;
3736 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3738 struct cfs_bandwidth *cfs_b =
3739 container_of(timer, struct cfs_bandwidth, period_timer);
3744 raw_spin_lock(&cfs_b->lock);
3746 now = hrtimer_cb_get_time(timer);
3747 overrun = hrtimer_forward(timer, now, cfs_b->period);
3752 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3754 raw_spin_unlock(&cfs_b->lock);
3756 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3759 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3761 raw_spin_lock_init(&cfs_b->lock);
3763 cfs_b->quota = RUNTIME_INF;
3764 cfs_b->period = ns_to_ktime(default_cfs_period());
3766 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3767 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3768 cfs_b->period_timer.function = sched_cfs_period_timer;
3769 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3770 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3773 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3775 cfs_rq->runtime_enabled = 0;
3776 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3779 /* requires cfs_b->lock, may release to reprogram timer */
3780 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force)
3783 * The timer may be active because we're trying to set a new bandwidth
3784 * period or because we're racing with the tear-down path
3785 * (timer_active==0 becomes visible before the hrtimer call-back
3786 * terminates). In either case we ensure that it's re-programmed
3788 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3789 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3790 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3791 raw_spin_unlock(&cfs_b->lock);
3793 raw_spin_lock(&cfs_b->lock);
3794 /* if someone else restarted the timer then we're done */
3795 if (!force && cfs_b->timer_active)
3799 cfs_b->timer_active = 1;
3800 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3803 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3805 hrtimer_cancel(&cfs_b->period_timer);
3806 hrtimer_cancel(&cfs_b->slack_timer);
3809 static void __maybe_unused update_runtime_enabled(struct rq *rq)
3811 struct cfs_rq *cfs_rq;
3813 for_each_leaf_cfs_rq(rq, cfs_rq) {
3814 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
3816 raw_spin_lock(&cfs_b->lock);
3817 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
3818 raw_spin_unlock(&cfs_b->lock);
3822 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3824 struct cfs_rq *cfs_rq;
3826 for_each_leaf_cfs_rq(rq, cfs_rq) {
3827 if (!cfs_rq->runtime_enabled)
3831 * clock_task is not advancing so we just need to make sure
3832 * there's some valid quota amount
3834 cfs_rq->runtime_remaining = 1;
3836 * Offline rq is schedulable till cpu is completely disabled
3837 * in take_cpu_down(), so we prevent new cfs throttling here.
3839 cfs_rq->runtime_enabled = 0;
3841 if (cfs_rq_throttled(cfs_rq))
3842 unthrottle_cfs_rq(cfs_rq);
3846 #else /* CONFIG_CFS_BANDWIDTH */
3847 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3849 return rq_clock_task(rq_of(cfs_rq));
3852 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3853 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3854 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3855 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3857 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3862 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3867 static inline int throttled_lb_pair(struct task_group *tg,
3868 int src_cpu, int dest_cpu)
3873 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3875 #ifdef CONFIG_FAIR_GROUP_SCHED
3876 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3879 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3883 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3884 static inline void update_runtime_enabled(struct rq *rq) {}
3885 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3887 #endif /* CONFIG_CFS_BANDWIDTH */
3889 /**************************************************
3890 * CFS operations on tasks:
3893 #ifdef CONFIG_SCHED_HRTICK
3894 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3896 struct sched_entity *se = &p->se;
3897 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3899 WARN_ON(task_rq(p) != rq);
3901 if (cfs_rq->nr_running > 1) {
3902 u64 slice = sched_slice(cfs_rq, se);
3903 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3904 s64 delta = slice - ran;
3911 hrtick_start(rq, delta);
3916 * called from enqueue/dequeue and updates the hrtick when the
3917 * current task is from our class and nr_running is low enough
3920 static void hrtick_update(struct rq *rq)
3922 struct task_struct *curr = rq->curr;
3924 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3927 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3928 hrtick_start_fair(rq, curr);
3930 #else /* !CONFIG_SCHED_HRTICK */
3932 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3936 static inline void hrtick_update(struct rq *rq)
3942 * The enqueue_task method is called before nr_running is
3943 * increased. Here we update the fair scheduling stats and
3944 * then put the task into the rbtree:
3947 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3949 struct cfs_rq *cfs_rq;
3950 struct sched_entity *se = &p->se;
3952 for_each_sched_entity(se) {
3955 cfs_rq = cfs_rq_of(se);
3956 enqueue_entity(cfs_rq, se, flags);
3959 * end evaluation on encountering a throttled cfs_rq
3961 * note: in the case of encountering a throttled cfs_rq we will
3962 * post the final h_nr_running increment below.
3964 if (cfs_rq_throttled(cfs_rq))
3966 cfs_rq->h_nr_running++;
3968 flags = ENQUEUE_WAKEUP;
3971 for_each_sched_entity(se) {
3972 cfs_rq = cfs_rq_of(se);
3973 cfs_rq->h_nr_running++;
3975 if (cfs_rq_throttled(cfs_rq))
3978 update_cfs_shares(cfs_rq);
3979 update_entity_load_avg(se, 1);
3983 update_rq_runnable_avg(rq, rq->nr_running);
3984 add_nr_running(rq, 1);
3989 static void set_next_buddy(struct sched_entity *se);
3992 * The dequeue_task method is called before nr_running is
3993 * decreased. We remove the task from the rbtree and
3994 * update the fair scheduling stats:
3996 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3998 struct cfs_rq *cfs_rq;
3999 struct sched_entity *se = &p->se;
4000 int task_sleep = flags & DEQUEUE_SLEEP;
4002 for_each_sched_entity(se) {
4003 cfs_rq = cfs_rq_of(se);
4004 dequeue_entity(cfs_rq, se, flags);
4007 * end evaluation on encountering a throttled cfs_rq
4009 * note: in the case of encountering a throttled cfs_rq we will
4010 * post the final h_nr_running decrement below.
4012 if (cfs_rq_throttled(cfs_rq))
4014 cfs_rq->h_nr_running--;
4016 /* Don't dequeue parent if it has other entities besides us */
4017 if (cfs_rq->load.weight) {
4019 * Bias pick_next to pick a task from this cfs_rq, as
4020 * p is sleeping when it is within its sched_slice.
4022 if (task_sleep && parent_entity(se))
4023 set_next_buddy(parent_entity(se));
4025 /* avoid re-evaluating load for this entity */
4026 se = parent_entity(se);
4029 flags |= DEQUEUE_SLEEP;
4032 for_each_sched_entity(se) {
4033 cfs_rq = cfs_rq_of(se);
4034 cfs_rq->h_nr_running--;
4036 if (cfs_rq_throttled(cfs_rq))
4039 update_cfs_shares(cfs_rq);
4040 update_entity_load_avg(se, 1);
4044 sub_nr_running(rq, 1);
4045 update_rq_runnable_avg(rq, 1);
4051 /* Used instead of source_load when we know the type == 0 */
4052 static unsigned long weighted_cpuload(const int cpu)
4054 return cpu_rq(cpu)->cfs.runnable_load_avg;
4058 * Return a low guess at the load of a migration-source cpu weighted
4059 * according to the scheduling class and "nice" value.
4061 * We want to under-estimate the load of migration sources, to
4062 * balance conservatively.
4064 static unsigned long source_load(int cpu, int type)
4066 struct rq *rq = cpu_rq(cpu);
4067 unsigned long total = weighted_cpuload(cpu);
4069 if (type == 0 || !sched_feat(LB_BIAS))
4072 return min(rq->cpu_load[type-1], total);
4076 * Return a high guess at the load of a migration-target cpu weighted
4077 * according to the scheduling class and "nice" value.
4079 static unsigned long target_load(int cpu, int type)
4081 struct rq *rq = cpu_rq(cpu);
4082 unsigned long total = weighted_cpuload(cpu);
4084 if (type == 0 || !sched_feat(LB_BIAS))
4087 return max(rq->cpu_load[type-1], total);
4090 static unsigned long capacity_of(int cpu)
4092 return cpu_rq(cpu)->cpu_capacity;
4095 static unsigned long cpu_avg_load_per_task(int cpu)
4097 struct rq *rq = cpu_rq(cpu);
4098 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
4099 unsigned long load_avg = rq->cfs.runnable_load_avg;
4102 return load_avg / nr_running;
4107 static void record_wakee(struct task_struct *p)
4110 * Rough decay (wiping) for cost saving, don't worry
4111 * about the boundary, really active task won't care
4114 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4115 current->wakee_flips >>= 1;
4116 current->wakee_flip_decay_ts = jiffies;
4119 if (current->last_wakee != p) {
4120 current->last_wakee = p;
4121 current->wakee_flips++;
4125 static void task_waking_fair(struct task_struct *p)
4127 struct sched_entity *se = &p->se;
4128 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4131 #ifndef CONFIG_64BIT
4132 u64 min_vruntime_copy;
4135 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4137 min_vruntime = cfs_rq->min_vruntime;
4138 } while (min_vruntime != min_vruntime_copy);
4140 min_vruntime = cfs_rq->min_vruntime;
4143 se->vruntime -= min_vruntime;
4147 #ifdef CONFIG_FAIR_GROUP_SCHED
4149 * effective_load() calculates the load change as seen from the root_task_group
4151 * Adding load to a group doesn't make a group heavier, but can cause movement
4152 * of group shares between cpus. Assuming the shares were perfectly aligned one
4153 * can calculate the shift in shares.
4155 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4156 * on this @cpu and results in a total addition (subtraction) of @wg to the
4157 * total group weight.
4159 * Given a runqueue weight distribution (rw_i) we can compute a shares
4160 * distribution (s_i) using:
4162 * s_i = rw_i / \Sum rw_j (1)
4164 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4165 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4166 * shares distribution (s_i):
4168 * rw_i = { 2, 4, 1, 0 }
4169 * s_i = { 2/7, 4/7, 1/7, 0 }
4171 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4172 * task used to run on and the CPU the waker is running on), we need to
4173 * compute the effect of waking a task on either CPU and, in case of a sync
4174 * wakeup, compute the effect of the current task going to sleep.
4176 * So for a change of @wl to the local @cpu with an overall group weight change
4177 * of @wl we can compute the new shares distribution (s'_i) using:
4179 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4181 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4182 * differences in waking a task to CPU 0. The additional task changes the
4183 * weight and shares distributions like:
4185 * rw'_i = { 3, 4, 1, 0 }
4186 * s'_i = { 3/8, 4/8, 1/8, 0 }
4188 * We can then compute the difference in effective weight by using:
4190 * dw_i = S * (s'_i - s_i) (3)
4192 * Where 'S' is the group weight as seen by its parent.
4194 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4195 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4196 * 4/7) times the weight of the group.
4198 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4200 struct sched_entity *se = tg->se[cpu];
4202 if (!tg->parent) /* the trivial, non-cgroup case */
4205 for_each_sched_entity(se) {
4211 * W = @wg + \Sum rw_j
4213 W = wg + calc_tg_weight(tg, se->my_q);
4218 w = se->my_q->load.weight + wl;
4221 * wl = S * s'_i; see (2)
4224 wl = (w * tg->shares) / W;
4229 * Per the above, wl is the new se->load.weight value; since
4230 * those are clipped to [MIN_SHARES, ...) do so now. See
4231 * calc_cfs_shares().
4233 if (wl < MIN_SHARES)
4237 * wl = dw_i = S * (s'_i - s_i); see (3)
4239 wl -= se->load.weight;
4242 * Recursively apply this logic to all parent groups to compute
4243 * the final effective load change on the root group. Since
4244 * only the @tg group gets extra weight, all parent groups can
4245 * only redistribute existing shares. @wl is the shift in shares
4246 * resulting from this level per the above.
4255 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4262 static int wake_wide(struct task_struct *p)
4264 int factor = this_cpu_read(sd_llc_size);
4267 * Yeah, it's the switching-frequency, could means many wakee or
4268 * rapidly switch, use factor here will just help to automatically
4269 * adjust the loose-degree, so bigger node will lead to more pull.
4271 if (p->wakee_flips > factor) {
4273 * wakee is somewhat hot, it needs certain amount of cpu
4274 * resource, so if waker is far more hot, prefer to leave
4277 if (current->wakee_flips > (factor * p->wakee_flips))
4284 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4286 s64 this_load, load;
4287 int idx, this_cpu, prev_cpu;
4288 unsigned long tl_per_task;
4289 struct task_group *tg;
4290 unsigned long weight;
4294 * If we wake multiple tasks be careful to not bounce
4295 * ourselves around too much.
4301 this_cpu = smp_processor_id();
4302 prev_cpu = task_cpu(p);
4303 load = source_load(prev_cpu, idx);
4304 this_load = target_load(this_cpu, idx);
4307 * If sync wakeup then subtract the (maximum possible)
4308 * effect of the currently running task from the load
4309 * of the current CPU:
4312 tg = task_group(current);
4313 weight = current->se.load.weight;
4315 this_load += effective_load(tg, this_cpu, -weight, -weight);
4316 load += effective_load(tg, prev_cpu, 0, -weight);
4320 weight = p->se.load.weight;
4323 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4324 * due to the sync cause above having dropped this_load to 0, we'll
4325 * always have an imbalance, but there's really nothing you can do
4326 * about that, so that's good too.
4328 * Otherwise check if either cpus are near enough in load to allow this
4329 * task to be woken on this_cpu.
4331 if (this_load > 0) {
4332 s64 this_eff_load, prev_eff_load;
4334 this_eff_load = 100;
4335 this_eff_load *= capacity_of(prev_cpu);
4336 this_eff_load *= this_load +
4337 effective_load(tg, this_cpu, weight, weight);
4339 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4340 prev_eff_load *= capacity_of(this_cpu);
4341 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4343 balanced = this_eff_load <= prev_eff_load;
4348 * If the currently running task will sleep within
4349 * a reasonable amount of time then attract this newly
4352 if (sync && balanced)
4355 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4356 tl_per_task = cpu_avg_load_per_task(this_cpu);
4359 (this_load <= load &&
4360 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4362 * This domain has SD_WAKE_AFFINE and
4363 * p is cache cold in this domain, and
4364 * there is no bad imbalance.
4366 schedstat_inc(sd, ttwu_move_affine);
4367 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4375 * find_idlest_group finds and returns the least busy CPU group within the
4378 static struct sched_group *
4379 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4380 int this_cpu, int sd_flag)
4382 struct sched_group *idlest = NULL, *group = sd->groups;
4383 unsigned long min_load = ULONG_MAX, this_load = 0;
4384 int load_idx = sd->forkexec_idx;
4385 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4387 if (sd_flag & SD_BALANCE_WAKE)
4388 load_idx = sd->wake_idx;
4391 unsigned long load, avg_load;
4395 /* Skip over this group if it has no CPUs allowed */
4396 if (!cpumask_intersects(sched_group_cpus(group),
4397 tsk_cpus_allowed(p)))
4400 local_group = cpumask_test_cpu(this_cpu,
4401 sched_group_cpus(group));
4403 /* Tally up the load of all CPUs in the group */
4406 for_each_cpu(i, sched_group_cpus(group)) {
4407 /* Bias balancing toward cpus of our domain */
4409 load = source_load(i, load_idx);
4411 load = target_load(i, load_idx);
4416 /* Adjust by relative CPU capacity of the group */
4417 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4420 this_load = avg_load;
4421 } else if (avg_load < min_load) {
4422 min_load = avg_load;
4425 } while (group = group->next, group != sd->groups);
4427 if (!idlest || 100*this_load < imbalance*min_load)
4433 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4436 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4438 unsigned long load, min_load = ULONG_MAX;
4442 /* Traverse only the allowed CPUs */
4443 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4444 load = weighted_cpuload(i);
4446 if (load < min_load || (load == min_load && i == this_cpu)) {
4456 * Try and locate an idle CPU in the sched_domain.
4458 static int select_idle_sibling(struct task_struct *p, int target)
4460 struct sched_domain *sd;
4461 struct sched_group *sg;
4462 int i = task_cpu(p);
4464 if (idle_cpu(target))
4468 * If the prevous cpu is cache affine and idle, don't be stupid.
4470 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4474 * Otherwise, iterate the domains and find an elegible idle cpu.
4476 sd = rcu_dereference(per_cpu(sd_llc, target));
4477 for_each_lower_domain(sd) {
4480 if (!cpumask_intersects(sched_group_cpus(sg),
4481 tsk_cpus_allowed(p)))
4484 for_each_cpu(i, sched_group_cpus(sg)) {
4485 if (i == target || !idle_cpu(i))
4489 target = cpumask_first_and(sched_group_cpus(sg),
4490 tsk_cpus_allowed(p));
4494 } while (sg != sd->groups);
4501 * select_task_rq_fair: Select target runqueue for the waking task in domains
4502 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4503 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4505 * Balances load by selecting the idlest cpu in the idlest group, or under
4506 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4508 * Returns the target cpu number.
4510 * preempt must be disabled.
4513 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4515 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4516 int cpu = smp_processor_id();
4518 int want_affine = 0;
4519 int sync = wake_flags & WF_SYNC;
4521 if (p->nr_cpus_allowed == 1)
4524 if (sd_flag & SD_BALANCE_WAKE)
4525 want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4528 for_each_domain(cpu, tmp) {
4529 if (!(tmp->flags & SD_LOAD_BALANCE))
4533 * If both cpu and prev_cpu are part of this domain,
4534 * cpu is a valid SD_WAKE_AFFINE target.
4536 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4537 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4542 if (tmp->flags & sd_flag)
4546 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4549 if (sd_flag & SD_BALANCE_WAKE) {
4550 new_cpu = select_idle_sibling(p, prev_cpu);
4555 struct sched_group *group;
4558 if (!(sd->flags & sd_flag)) {
4563 group = find_idlest_group(sd, p, cpu, sd_flag);
4569 new_cpu = find_idlest_cpu(group, p, cpu);
4570 if (new_cpu == -1 || new_cpu == cpu) {
4571 /* Now try balancing at a lower domain level of cpu */
4576 /* Now try balancing at a lower domain level of new_cpu */
4578 weight = sd->span_weight;
4580 for_each_domain(cpu, tmp) {
4581 if (weight <= tmp->span_weight)
4583 if (tmp->flags & sd_flag)
4586 /* while loop will break here if sd == NULL */
4595 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4596 * cfs_rq_of(p) references at time of call are still valid and identify the
4597 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4598 * other assumptions, including the state of rq->lock, should be made.
4601 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4603 struct sched_entity *se = &p->se;
4604 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4607 * Load tracking: accumulate removed load so that it can be processed
4608 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4609 * to blocked load iff they have a positive decay-count. It can never
4610 * be negative here since on-rq tasks have decay-count == 0.
4612 if (se->avg.decay_count) {
4613 se->avg.decay_count = -__synchronize_entity_decay(se);
4614 atomic_long_add(se->avg.load_avg_contrib,
4615 &cfs_rq->removed_load);
4618 /* We have migrated, no longer consider this task hot */
4621 #endif /* CONFIG_SMP */
4623 static unsigned long
4624 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4626 unsigned long gran = sysctl_sched_wakeup_granularity;
4629 * Since its curr running now, convert the gran from real-time
4630 * to virtual-time in his units.
4632 * By using 'se' instead of 'curr' we penalize light tasks, so
4633 * they get preempted easier. That is, if 'se' < 'curr' then
4634 * the resulting gran will be larger, therefore penalizing the
4635 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4636 * be smaller, again penalizing the lighter task.
4638 * This is especially important for buddies when the leftmost
4639 * task is higher priority than the buddy.
4641 return calc_delta_fair(gran, se);
4645 * Should 'se' preempt 'curr'.
4659 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4661 s64 gran, vdiff = curr->vruntime - se->vruntime;
4666 gran = wakeup_gran(curr, se);
4673 static void set_last_buddy(struct sched_entity *se)
4675 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4678 for_each_sched_entity(se)
4679 cfs_rq_of(se)->last = se;
4682 static void set_next_buddy(struct sched_entity *se)
4684 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4687 for_each_sched_entity(se)
4688 cfs_rq_of(se)->next = se;
4691 static void set_skip_buddy(struct sched_entity *se)
4693 for_each_sched_entity(se)
4694 cfs_rq_of(se)->skip = se;
4698 * Preempt the current task with a newly woken task if needed:
4700 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4702 struct task_struct *curr = rq->curr;
4703 struct sched_entity *se = &curr->se, *pse = &p->se;
4704 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4705 int scale = cfs_rq->nr_running >= sched_nr_latency;
4706 int next_buddy_marked = 0;
4708 if (unlikely(se == pse))
4712 * This is possible from callers such as attach_tasks(), in which we
4713 * unconditionally check_prempt_curr() after an enqueue (which may have
4714 * lead to a throttle). This both saves work and prevents false
4715 * next-buddy nomination below.
4717 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4720 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4721 set_next_buddy(pse);
4722 next_buddy_marked = 1;
4726 * We can come here with TIF_NEED_RESCHED already set from new task
4729 * Note: this also catches the edge-case of curr being in a throttled
4730 * group (e.g. via set_curr_task), since update_curr() (in the
4731 * enqueue of curr) will have resulted in resched being set. This
4732 * prevents us from potentially nominating it as a false LAST_BUDDY
4735 if (test_tsk_need_resched(curr))
4738 /* Idle tasks are by definition preempted by non-idle tasks. */
4739 if (unlikely(curr->policy == SCHED_IDLE) &&
4740 likely(p->policy != SCHED_IDLE))
4744 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4745 * is driven by the tick):
4747 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4750 find_matching_se(&se, &pse);
4751 update_curr(cfs_rq_of(se));
4753 if (wakeup_preempt_entity(se, pse) == 1) {
4755 * Bias pick_next to pick the sched entity that is
4756 * triggering this preemption.
4758 if (!next_buddy_marked)
4759 set_next_buddy(pse);
4768 * Only set the backward buddy when the current task is still
4769 * on the rq. This can happen when a wakeup gets interleaved
4770 * with schedule on the ->pre_schedule() or idle_balance()
4771 * point, either of which can * drop the rq lock.
4773 * Also, during early boot the idle thread is in the fair class,
4774 * for obvious reasons its a bad idea to schedule back to it.
4776 if (unlikely(!se->on_rq || curr == rq->idle))
4779 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4783 static struct task_struct *
4784 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4786 struct cfs_rq *cfs_rq = &rq->cfs;
4787 struct sched_entity *se;
4788 struct task_struct *p;
4792 #ifdef CONFIG_FAIR_GROUP_SCHED
4793 if (!cfs_rq->nr_running)
4796 if (prev->sched_class != &fair_sched_class)
4800 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4801 * likely that a next task is from the same cgroup as the current.
4803 * Therefore attempt to avoid putting and setting the entire cgroup
4804 * hierarchy, only change the part that actually changes.
4808 struct sched_entity *curr = cfs_rq->curr;
4811 * Since we got here without doing put_prev_entity() we also
4812 * have to consider cfs_rq->curr. If it is still a runnable
4813 * entity, update_curr() will update its vruntime, otherwise
4814 * forget we've ever seen it.
4816 if (curr && curr->on_rq)
4817 update_curr(cfs_rq);
4822 * This call to check_cfs_rq_runtime() will do the throttle and
4823 * dequeue its entity in the parent(s). Therefore the 'simple'
4824 * nr_running test will indeed be correct.
4826 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4829 se = pick_next_entity(cfs_rq, curr);
4830 cfs_rq = group_cfs_rq(se);
4836 * Since we haven't yet done put_prev_entity and if the selected task
4837 * is a different task than we started out with, try and touch the
4838 * least amount of cfs_rqs.
4841 struct sched_entity *pse = &prev->se;
4843 while (!(cfs_rq = is_same_group(se, pse))) {
4844 int se_depth = se->depth;
4845 int pse_depth = pse->depth;
4847 if (se_depth <= pse_depth) {
4848 put_prev_entity(cfs_rq_of(pse), pse);
4849 pse = parent_entity(pse);
4851 if (se_depth >= pse_depth) {
4852 set_next_entity(cfs_rq_of(se), se);
4853 se = parent_entity(se);
4857 put_prev_entity(cfs_rq, pse);
4858 set_next_entity(cfs_rq, se);
4861 if (hrtick_enabled(rq))
4862 hrtick_start_fair(rq, p);
4869 if (!cfs_rq->nr_running)
4872 put_prev_task(rq, prev);
4875 se = pick_next_entity(cfs_rq, NULL);
4876 set_next_entity(cfs_rq, se);
4877 cfs_rq = group_cfs_rq(se);
4882 if (hrtick_enabled(rq))
4883 hrtick_start_fair(rq, p);
4888 new_tasks = idle_balance(rq);
4890 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4891 * possible for any higher priority task to appear. In that case we
4892 * must re-start the pick_next_entity() loop.
4904 * Account for a descheduled task:
4906 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4908 struct sched_entity *se = &prev->se;
4909 struct cfs_rq *cfs_rq;
4911 for_each_sched_entity(se) {
4912 cfs_rq = cfs_rq_of(se);
4913 put_prev_entity(cfs_rq, se);
4918 * sched_yield() is very simple
4920 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4922 static void yield_task_fair(struct rq *rq)
4924 struct task_struct *curr = rq->curr;
4925 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4926 struct sched_entity *se = &curr->se;
4929 * Are we the only task in the tree?
4931 if (unlikely(rq->nr_running == 1))
4934 clear_buddies(cfs_rq, se);
4936 if (curr->policy != SCHED_BATCH) {
4937 update_rq_clock(rq);
4939 * Update run-time statistics of the 'current'.
4941 update_curr(cfs_rq);
4943 * Tell update_rq_clock() that we've just updated,
4944 * so we don't do microscopic update in schedule()
4945 * and double the fastpath cost.
4947 rq->skip_clock_update = 1;
4953 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4955 struct sched_entity *se = &p->se;
4957 /* throttled hierarchies are not runnable */
4958 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4961 /* Tell the scheduler that we'd really like pse to run next. */
4964 yield_task_fair(rq);
4970 /**************************************************
4971 * Fair scheduling class load-balancing methods.
4975 * The purpose of load-balancing is to achieve the same basic fairness the
4976 * per-cpu scheduler provides, namely provide a proportional amount of compute
4977 * time to each task. This is expressed in the following equation:
4979 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4981 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4982 * W_i,0 is defined as:
4984 * W_i,0 = \Sum_j w_i,j (2)
4986 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4987 * is derived from the nice value as per prio_to_weight[].
4989 * The weight average is an exponential decay average of the instantaneous
4992 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4994 * C_i is the compute capacity of cpu i, typically it is the
4995 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4996 * can also include other factors [XXX].
4998 * To achieve this balance we define a measure of imbalance which follows
4999 * directly from (1):
5001 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5003 * We them move tasks around to minimize the imbalance. In the continuous
5004 * function space it is obvious this converges, in the discrete case we get
5005 * a few fun cases generally called infeasible weight scenarios.
5008 * - infeasible weights;
5009 * - local vs global optima in the discrete case. ]
5014 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5015 * for all i,j solution, we create a tree of cpus that follows the hardware
5016 * topology where each level pairs two lower groups (or better). This results
5017 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5018 * tree to only the first of the previous level and we decrease the frequency
5019 * of load-balance at each level inv. proportional to the number of cpus in
5025 * \Sum { --- * --- * 2^i } = O(n) (5)
5027 * `- size of each group
5028 * | | `- number of cpus doing load-balance
5030 * `- sum over all levels
5032 * Coupled with a limit on how many tasks we can migrate every balance pass,
5033 * this makes (5) the runtime complexity of the balancer.
5035 * An important property here is that each CPU is still (indirectly) connected
5036 * to every other cpu in at most O(log n) steps:
5038 * The adjacency matrix of the resulting graph is given by:
5041 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5044 * And you'll find that:
5046 * A^(log_2 n)_i,j != 0 for all i,j (7)
5048 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5049 * The task movement gives a factor of O(m), giving a convergence complexity
5052 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5057 * In order to avoid CPUs going idle while there's still work to do, new idle
5058 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5059 * tree itself instead of relying on other CPUs to bring it work.
5061 * This adds some complexity to both (5) and (8) but it reduces the total idle
5069 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5072 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5077 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5079 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5081 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5084 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5085 * rewrite all of this once again.]
5088 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5090 enum fbq_type { regular, remote, all };
5092 #define LBF_ALL_PINNED 0x01
5093 #define LBF_NEED_BREAK 0x02
5094 #define LBF_DST_PINNED 0x04
5095 #define LBF_SOME_PINNED 0x08
5098 struct sched_domain *sd;
5106 struct cpumask *dst_grpmask;
5108 enum cpu_idle_type idle;
5110 /* The set of CPUs under consideration for load-balancing */
5111 struct cpumask *cpus;
5116 unsigned int loop_break;
5117 unsigned int loop_max;
5119 enum fbq_type fbq_type;
5120 struct list_head tasks;
5124 * Is this task likely cache-hot:
5126 static int task_hot(struct task_struct *p, struct lb_env *env)
5130 lockdep_assert_held(&env->src_rq->lock);
5132 if (p->sched_class != &fair_sched_class)
5135 if (unlikely(p->policy == SCHED_IDLE))
5139 * Buddy candidates are cache hot:
5141 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5142 (&p->se == cfs_rq_of(&p->se)->next ||
5143 &p->se == cfs_rq_of(&p->se)->last))
5146 if (sysctl_sched_migration_cost == -1)
5148 if (sysctl_sched_migration_cost == 0)
5151 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5153 return delta < (s64)sysctl_sched_migration_cost;
5156 #ifdef CONFIG_NUMA_BALANCING
5157 /* Returns true if the destination node has incurred more faults */
5158 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5160 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5161 int src_nid, dst_nid;
5163 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5164 !(env->sd->flags & SD_NUMA)) {
5168 src_nid = cpu_to_node(env->src_cpu);
5169 dst_nid = cpu_to_node(env->dst_cpu);
5171 if (src_nid == dst_nid)
5175 /* Task is already in the group's interleave set. */
5176 if (node_isset(src_nid, numa_group->active_nodes))
5179 /* Task is moving into the group's interleave set. */
5180 if (node_isset(dst_nid, numa_group->active_nodes))
5183 return group_faults(p, dst_nid) > group_faults(p, src_nid);
5186 /* Encourage migration to the preferred node. */
5187 if (dst_nid == p->numa_preferred_nid)
5190 return task_faults(p, dst_nid) > task_faults(p, src_nid);
5194 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5196 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5197 int src_nid, dst_nid;
5199 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5202 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5205 src_nid = cpu_to_node(env->src_cpu);
5206 dst_nid = cpu_to_node(env->dst_cpu);
5208 if (src_nid == dst_nid)
5212 /* Task is moving within/into the group's interleave set. */
5213 if (node_isset(dst_nid, numa_group->active_nodes))
5216 /* Task is moving out of the group's interleave set. */
5217 if (node_isset(src_nid, numa_group->active_nodes))
5220 return group_faults(p, dst_nid) < group_faults(p, src_nid);
5223 /* Migrating away from the preferred node is always bad. */
5224 if (src_nid == p->numa_preferred_nid)
5227 return task_faults(p, dst_nid) < task_faults(p, src_nid);
5231 static inline bool migrate_improves_locality(struct task_struct *p,
5237 static inline bool migrate_degrades_locality(struct task_struct *p,
5245 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5248 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5250 int tsk_cache_hot = 0;
5252 lockdep_assert_held(&env->src_rq->lock);
5255 * We do not migrate tasks that are:
5256 * 1) throttled_lb_pair, or
5257 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5258 * 3) running (obviously), or
5259 * 4) are cache-hot on their current CPU.
5261 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5264 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5267 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5269 env->flags |= LBF_SOME_PINNED;
5272 * Remember if this task can be migrated to any other cpu in
5273 * our sched_group. We may want to revisit it if we couldn't
5274 * meet load balance goals by pulling other tasks on src_cpu.
5276 * Also avoid computing new_dst_cpu if we have already computed
5277 * one in current iteration.
5279 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5282 /* Prevent to re-select dst_cpu via env's cpus */
5283 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5284 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5285 env->flags |= LBF_DST_PINNED;
5286 env->new_dst_cpu = cpu;
5294 /* Record that we found atleast one task that could run on dst_cpu */
5295 env->flags &= ~LBF_ALL_PINNED;
5297 if (task_running(env->src_rq, p)) {
5298 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5303 * Aggressive migration if:
5304 * 1) destination numa is preferred
5305 * 2) task is cache cold, or
5306 * 3) too many balance attempts have failed.
5308 tsk_cache_hot = task_hot(p, env);
5310 tsk_cache_hot = migrate_degrades_locality(p, env);
5312 if (migrate_improves_locality(p, env)) {
5313 #ifdef CONFIG_SCHEDSTATS
5314 if (tsk_cache_hot) {
5315 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5316 schedstat_inc(p, se.statistics.nr_forced_migrations);
5322 if (!tsk_cache_hot ||
5323 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5325 if (tsk_cache_hot) {
5326 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5327 schedstat_inc(p, se.statistics.nr_forced_migrations);
5333 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5338 * detach_task() -- detach the task for the migration specified in env
5340 static void detach_task(struct task_struct *p, struct lb_env *env)
5342 lockdep_assert_held(&env->src_rq->lock);
5344 deactivate_task(env->src_rq, p, 0);
5345 p->on_rq = TASK_ON_RQ_MIGRATING;
5346 set_task_cpu(p, env->dst_cpu);
5350 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5351 * part of active balancing operations within "domain".
5353 * Returns a task if successful and NULL otherwise.
5355 static struct task_struct *detach_one_task(struct lb_env *env)
5357 struct task_struct *p, *n;
5359 lockdep_assert_held(&env->src_rq->lock);
5361 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5362 if (!can_migrate_task(p, env))
5365 detach_task(p, env);
5368 * Right now, this is only the second place where
5369 * lb_gained[env->idle] is updated (other is detach_tasks)
5370 * so we can safely collect stats here rather than
5371 * inside detach_tasks().
5373 schedstat_inc(env->sd, lb_gained[env->idle]);
5379 static const unsigned int sched_nr_migrate_break = 32;
5382 * detach_tasks() -- tries to detach up to imbalance weighted load from
5383 * busiest_rq, as part of a balancing operation within domain "sd".
5385 * Returns number of detached tasks if successful and 0 otherwise.
5387 static int detach_tasks(struct lb_env *env)
5389 struct list_head *tasks = &env->src_rq->cfs_tasks;
5390 struct task_struct *p;
5394 lockdep_assert_held(&env->src_rq->lock);
5396 if (env->imbalance <= 0)
5399 while (!list_empty(tasks)) {
5400 p = list_first_entry(tasks, struct task_struct, se.group_node);
5403 /* We've more or less seen every task there is, call it quits */
5404 if (env->loop > env->loop_max)
5407 /* take a breather every nr_migrate tasks */
5408 if (env->loop > env->loop_break) {
5409 env->loop_break += sched_nr_migrate_break;
5410 env->flags |= LBF_NEED_BREAK;
5414 if (!can_migrate_task(p, env))
5417 load = task_h_load(p);
5419 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5422 if ((load / 2) > env->imbalance)
5425 detach_task(p, env);
5426 list_add(&p->se.group_node, &env->tasks);
5429 env->imbalance -= load;
5431 #ifdef CONFIG_PREEMPT
5433 * NEWIDLE balancing is a source of latency, so preemptible
5434 * kernels will stop after the first task is detached to minimize
5435 * the critical section.
5437 if (env->idle == CPU_NEWLY_IDLE)
5442 * We only want to steal up to the prescribed amount of
5445 if (env->imbalance <= 0)
5450 list_move_tail(&p->se.group_node, tasks);
5454 * Right now, this is one of only two places we collect this stat
5455 * so we can safely collect detach_one_task() stats here rather
5456 * than inside detach_one_task().
5458 schedstat_add(env->sd, lb_gained[env->idle], detached);
5464 * attach_task() -- attach the task detached by detach_task() to its new rq.
5466 static void attach_task(struct rq *rq, struct task_struct *p)
5468 lockdep_assert_held(&rq->lock);
5470 BUG_ON(task_rq(p) != rq);
5471 p->on_rq = TASK_ON_RQ_QUEUED;
5472 activate_task(rq, p, 0);
5473 check_preempt_curr(rq, p, 0);
5477 * attach_one_task() -- attaches the task returned from detach_one_task() to
5480 static void attach_one_task(struct rq *rq, struct task_struct *p)
5482 raw_spin_lock(&rq->lock);
5484 raw_spin_unlock(&rq->lock);
5488 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5491 static void attach_tasks(struct lb_env *env)
5493 struct list_head *tasks = &env->tasks;
5494 struct task_struct *p;
5496 raw_spin_lock(&env->dst_rq->lock);
5498 while (!list_empty(tasks)) {
5499 p = list_first_entry(tasks, struct task_struct, se.group_node);
5500 list_del_init(&p->se.group_node);
5502 attach_task(env->dst_rq, p);
5505 raw_spin_unlock(&env->dst_rq->lock);
5508 #ifdef CONFIG_FAIR_GROUP_SCHED
5510 * update tg->load_weight by folding this cpu's load_avg
5512 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5514 struct sched_entity *se = tg->se[cpu];
5515 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5517 /* throttled entities do not contribute to load */
5518 if (throttled_hierarchy(cfs_rq))
5521 update_cfs_rq_blocked_load(cfs_rq, 1);
5524 update_entity_load_avg(se, 1);
5526 * We pivot on our runnable average having decayed to zero for
5527 * list removal. This generally implies that all our children
5528 * have also been removed (modulo rounding error or bandwidth
5529 * control); however, such cases are rare and we can fix these
5532 * TODO: fix up out-of-order children on enqueue.
5534 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5535 list_del_leaf_cfs_rq(cfs_rq);
5537 struct rq *rq = rq_of(cfs_rq);
5538 update_rq_runnable_avg(rq, rq->nr_running);
5542 static void update_blocked_averages(int cpu)
5544 struct rq *rq = cpu_rq(cpu);
5545 struct cfs_rq *cfs_rq;
5546 unsigned long flags;
5548 raw_spin_lock_irqsave(&rq->lock, flags);
5549 update_rq_clock(rq);
5551 * Iterates the task_group tree in a bottom up fashion, see
5552 * list_add_leaf_cfs_rq() for details.
5554 for_each_leaf_cfs_rq(rq, cfs_rq) {
5556 * Note: We may want to consider periodically releasing
5557 * rq->lock about these updates so that creating many task
5558 * groups does not result in continually extending hold time.
5560 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5563 raw_spin_unlock_irqrestore(&rq->lock, flags);
5567 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5568 * This needs to be done in a top-down fashion because the load of a child
5569 * group is a fraction of its parents load.
5571 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5573 struct rq *rq = rq_of(cfs_rq);
5574 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5575 unsigned long now = jiffies;
5578 if (cfs_rq->last_h_load_update == now)
5581 cfs_rq->h_load_next = NULL;
5582 for_each_sched_entity(se) {
5583 cfs_rq = cfs_rq_of(se);
5584 cfs_rq->h_load_next = se;
5585 if (cfs_rq->last_h_load_update == now)
5590 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5591 cfs_rq->last_h_load_update = now;
5594 while ((se = cfs_rq->h_load_next) != NULL) {
5595 load = cfs_rq->h_load;
5596 load = div64_ul(load * se->avg.load_avg_contrib,
5597 cfs_rq->runnable_load_avg + 1);
5598 cfs_rq = group_cfs_rq(se);
5599 cfs_rq->h_load = load;
5600 cfs_rq->last_h_load_update = now;
5604 static unsigned long task_h_load(struct task_struct *p)
5606 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5608 update_cfs_rq_h_load(cfs_rq);
5609 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5610 cfs_rq->runnable_load_avg + 1);
5613 static inline void update_blocked_averages(int cpu)
5617 static unsigned long task_h_load(struct task_struct *p)
5619 return p->se.avg.load_avg_contrib;
5623 /********** Helpers for find_busiest_group ************************/
5632 * sg_lb_stats - stats of a sched_group required for load_balancing
5634 struct sg_lb_stats {
5635 unsigned long avg_load; /*Avg load across the CPUs of the group */
5636 unsigned long group_load; /* Total load over the CPUs of the group */
5637 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5638 unsigned long load_per_task;
5639 unsigned long group_capacity;
5640 unsigned int sum_nr_running; /* Nr tasks running in the group */
5641 unsigned int group_capacity_factor;
5642 unsigned int idle_cpus;
5643 unsigned int group_weight;
5644 enum group_type group_type;
5645 int group_has_free_capacity;
5646 #ifdef CONFIG_NUMA_BALANCING
5647 unsigned int nr_numa_running;
5648 unsigned int nr_preferred_running;
5653 * sd_lb_stats - Structure to store the statistics of a sched_domain
5654 * during load balancing.
5656 struct sd_lb_stats {
5657 struct sched_group *busiest; /* Busiest group in this sd */
5658 struct sched_group *local; /* Local group in this sd */
5659 unsigned long total_load; /* Total load of all groups in sd */
5660 unsigned long total_capacity; /* Total capacity of all groups in sd */
5661 unsigned long avg_load; /* Average load across all groups in sd */
5663 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5664 struct sg_lb_stats local_stat; /* Statistics of the local group */
5667 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5670 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5671 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5672 * We must however clear busiest_stat::avg_load because
5673 * update_sd_pick_busiest() reads this before assignment.
5675 *sds = (struct sd_lb_stats){
5679 .total_capacity = 0UL,
5682 .sum_nr_running = 0,
5683 .group_type = group_other,
5689 * get_sd_load_idx - Obtain the load index for a given sched domain.
5690 * @sd: The sched_domain whose load_idx is to be obtained.
5691 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5693 * Return: The load index.
5695 static inline int get_sd_load_idx(struct sched_domain *sd,
5696 enum cpu_idle_type idle)
5702 load_idx = sd->busy_idx;
5705 case CPU_NEWLY_IDLE:
5706 load_idx = sd->newidle_idx;
5709 load_idx = sd->idle_idx;
5716 static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu)
5718 return SCHED_CAPACITY_SCALE;
5721 unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu)
5723 return default_scale_capacity(sd, cpu);
5726 static unsigned long default_scale_smt_capacity(struct sched_domain *sd, int cpu)
5728 unsigned long weight = sd->span_weight;
5729 unsigned long smt_gain = sd->smt_gain;
5736 unsigned long __weak arch_scale_smt_capacity(struct sched_domain *sd, int cpu)
5738 return default_scale_smt_capacity(sd, cpu);
5741 static unsigned long scale_rt_capacity(int cpu)
5743 struct rq *rq = cpu_rq(cpu);
5744 u64 total, available, age_stamp, avg;
5748 * Since we're reading these variables without serialization make sure
5749 * we read them once before doing sanity checks on them.
5751 age_stamp = ACCESS_ONCE(rq->age_stamp);
5752 avg = ACCESS_ONCE(rq->rt_avg);
5754 delta = rq_clock(rq) - age_stamp;
5755 if (unlikely(delta < 0))
5758 total = sched_avg_period() + delta;
5760 if (unlikely(total < avg)) {
5761 /* Ensures that capacity won't end up being negative */
5764 available = total - avg;
5767 if (unlikely((s64)total < SCHED_CAPACITY_SCALE))
5768 total = SCHED_CAPACITY_SCALE;
5770 total >>= SCHED_CAPACITY_SHIFT;
5772 return div_u64(available, total);
5775 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
5777 unsigned long weight = sd->span_weight;
5778 unsigned long capacity = SCHED_CAPACITY_SCALE;
5779 struct sched_group *sdg = sd->groups;
5781 if ((sd->flags & SD_SHARE_CPUCAPACITY) && weight > 1) {
5782 if (sched_feat(ARCH_CAPACITY))
5783 capacity *= arch_scale_smt_capacity(sd, cpu);
5785 capacity *= default_scale_smt_capacity(sd, cpu);
5787 capacity >>= SCHED_CAPACITY_SHIFT;
5790 sdg->sgc->capacity_orig = capacity;
5792 if (sched_feat(ARCH_CAPACITY))
5793 capacity *= arch_scale_freq_capacity(sd, cpu);
5795 capacity *= default_scale_capacity(sd, cpu);
5797 capacity >>= SCHED_CAPACITY_SHIFT;
5799 capacity *= scale_rt_capacity(cpu);
5800 capacity >>= SCHED_CAPACITY_SHIFT;
5805 cpu_rq(cpu)->cpu_capacity = capacity;
5806 sdg->sgc->capacity = capacity;
5809 void update_group_capacity(struct sched_domain *sd, int cpu)
5811 struct sched_domain *child = sd->child;
5812 struct sched_group *group, *sdg = sd->groups;
5813 unsigned long capacity, capacity_orig;
5814 unsigned long interval;
5816 interval = msecs_to_jiffies(sd->balance_interval);
5817 interval = clamp(interval, 1UL, max_load_balance_interval);
5818 sdg->sgc->next_update = jiffies + interval;
5821 update_cpu_capacity(sd, cpu);
5825 capacity_orig = capacity = 0;
5827 if (child->flags & SD_OVERLAP) {
5829 * SD_OVERLAP domains cannot assume that child groups
5830 * span the current group.
5833 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5834 struct sched_group_capacity *sgc;
5835 struct rq *rq = cpu_rq(cpu);
5838 * build_sched_domains() -> init_sched_groups_capacity()
5839 * gets here before we've attached the domains to the
5842 * Use capacity_of(), which is set irrespective of domains
5843 * in update_cpu_capacity().
5845 * This avoids capacity/capacity_orig from being 0 and
5846 * causing divide-by-zero issues on boot.
5848 * Runtime updates will correct capacity_orig.
5850 if (unlikely(!rq->sd)) {
5851 capacity_orig += capacity_of(cpu);
5852 capacity += capacity_of(cpu);
5856 sgc = rq->sd->groups->sgc;
5857 capacity_orig += sgc->capacity_orig;
5858 capacity += sgc->capacity;
5862 * !SD_OVERLAP domains can assume that child groups
5863 * span the current group.
5866 group = child->groups;
5868 capacity_orig += group->sgc->capacity_orig;
5869 capacity += group->sgc->capacity;
5870 group = group->next;
5871 } while (group != child->groups);
5874 sdg->sgc->capacity_orig = capacity_orig;
5875 sdg->sgc->capacity = capacity;
5879 * Try and fix up capacity for tiny siblings, this is needed when
5880 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5881 * which on its own isn't powerful enough.
5883 * See update_sd_pick_busiest() and check_asym_packing().
5886 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5889 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5891 if (!(sd->flags & SD_SHARE_CPUCAPACITY))
5895 * If ~90% of the cpu_capacity is still there, we're good.
5897 if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29)
5904 * Group imbalance indicates (and tries to solve) the problem where balancing
5905 * groups is inadequate due to tsk_cpus_allowed() constraints.
5907 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5908 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5911 * { 0 1 2 3 } { 4 5 6 7 }
5914 * If we were to balance group-wise we'd place two tasks in the first group and
5915 * two tasks in the second group. Clearly this is undesired as it will overload
5916 * cpu 3 and leave one of the cpus in the second group unused.
5918 * The current solution to this issue is detecting the skew in the first group
5919 * by noticing the lower domain failed to reach balance and had difficulty
5920 * moving tasks due to affinity constraints.
5922 * When this is so detected; this group becomes a candidate for busiest; see
5923 * update_sd_pick_busiest(). And calculate_imbalance() and
5924 * find_busiest_group() avoid some of the usual balance conditions to allow it
5925 * to create an effective group imbalance.
5927 * This is a somewhat tricky proposition since the next run might not find the
5928 * group imbalance and decide the groups need to be balanced again. A most
5929 * subtle and fragile situation.
5932 static inline int sg_imbalanced(struct sched_group *group)
5934 return group->sgc->imbalance;
5938 * Compute the group capacity factor.
5940 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5941 * first dividing out the smt factor and computing the actual number of cores
5942 * and limit unit capacity with that.
5944 static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group)
5946 unsigned int capacity_factor, smt, cpus;
5947 unsigned int capacity, capacity_orig;
5949 capacity = group->sgc->capacity;
5950 capacity_orig = group->sgc->capacity_orig;
5951 cpus = group->group_weight;
5953 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5954 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig);
5955 capacity_factor = cpus / smt; /* cores */
5957 capacity_factor = min_t(unsigned,
5958 capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE));
5959 if (!capacity_factor)
5960 capacity_factor = fix_small_capacity(env->sd, group);
5962 return capacity_factor;
5965 static enum group_type
5966 group_classify(struct sched_group *group, struct sg_lb_stats *sgs)
5968 if (sgs->sum_nr_running > sgs->group_capacity_factor)
5969 return group_overloaded;
5971 if (sg_imbalanced(group))
5972 return group_imbalanced;
5978 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5979 * @env: The load balancing environment.
5980 * @group: sched_group whose statistics are to be updated.
5981 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5982 * @local_group: Does group contain this_cpu.
5983 * @sgs: variable to hold the statistics for this group.
5984 * @overload: Indicate more than one runnable task for any CPU.
5986 static inline void update_sg_lb_stats(struct lb_env *env,
5987 struct sched_group *group, int load_idx,
5988 int local_group, struct sg_lb_stats *sgs,
5994 memset(sgs, 0, sizeof(*sgs));
5996 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5997 struct rq *rq = cpu_rq(i);
5999 /* Bias balancing toward cpus of our domain */
6001 load = target_load(i, load_idx);
6003 load = source_load(i, load_idx);
6005 sgs->group_load += load;
6006 sgs->sum_nr_running += rq->nr_running;
6008 if (rq->nr_running > 1)
6011 #ifdef CONFIG_NUMA_BALANCING
6012 sgs->nr_numa_running += rq->nr_numa_running;
6013 sgs->nr_preferred_running += rq->nr_preferred_running;
6015 sgs->sum_weighted_load += weighted_cpuload(i);
6020 /* Adjust by relative CPU capacity of the group */
6021 sgs->group_capacity = group->sgc->capacity;
6022 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6024 if (sgs->sum_nr_running)
6025 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6027 sgs->group_weight = group->group_weight;
6028 sgs->group_capacity_factor = sg_capacity_factor(env, group);
6029 sgs->group_type = group_classify(group, sgs);
6031 if (sgs->group_capacity_factor > sgs->sum_nr_running)
6032 sgs->group_has_free_capacity = 1;
6036 * update_sd_pick_busiest - return 1 on busiest group
6037 * @env: The load balancing environment.
6038 * @sds: sched_domain statistics
6039 * @sg: sched_group candidate to be checked for being the busiest
6040 * @sgs: sched_group statistics
6042 * Determine if @sg is a busier group than the previously selected
6045 * Return: %true if @sg is a busier group than the previously selected
6046 * busiest group. %false otherwise.
6048 static bool update_sd_pick_busiest(struct lb_env *env,
6049 struct sd_lb_stats *sds,
6050 struct sched_group *sg,
6051 struct sg_lb_stats *sgs)
6053 struct sg_lb_stats *busiest = &sds->busiest_stat;
6055 if (sgs->group_type > busiest->group_type)
6058 if (sgs->group_type < busiest->group_type)
6061 if (sgs->avg_load <= busiest->avg_load)
6064 /* This is the busiest node in its class. */
6065 if (!(env->sd->flags & SD_ASYM_PACKING))
6069 * ASYM_PACKING needs to move all the work to the lowest
6070 * numbered CPUs in the group, therefore mark all groups
6071 * higher than ourself as busy.
6073 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6077 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6084 #ifdef CONFIG_NUMA_BALANCING
6085 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6087 if (sgs->sum_nr_running > sgs->nr_numa_running)
6089 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6094 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6096 if (rq->nr_running > rq->nr_numa_running)
6098 if (rq->nr_running > rq->nr_preferred_running)
6103 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6108 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6112 #endif /* CONFIG_NUMA_BALANCING */
6115 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6116 * @env: The load balancing environment.
6117 * @sds: variable to hold the statistics for this sched_domain.
6119 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6121 struct sched_domain *child = env->sd->child;
6122 struct sched_group *sg = env->sd->groups;
6123 struct sg_lb_stats tmp_sgs;
6124 int load_idx, prefer_sibling = 0;
6125 bool overload = false;
6127 if (child && child->flags & SD_PREFER_SIBLING)
6130 load_idx = get_sd_load_idx(env->sd, env->idle);
6133 struct sg_lb_stats *sgs = &tmp_sgs;
6136 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6139 sgs = &sds->local_stat;
6141 if (env->idle != CPU_NEWLY_IDLE ||
6142 time_after_eq(jiffies, sg->sgc->next_update))
6143 update_group_capacity(env->sd, env->dst_cpu);
6146 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6153 * In case the child domain prefers tasks go to siblings
6154 * first, lower the sg capacity factor to one so that we'll try
6155 * and move all the excess tasks away. We lower the capacity
6156 * of a group only if the local group has the capacity to fit
6157 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6158 * extra check prevents the case where you always pull from the
6159 * heaviest group when it is already under-utilized (possible
6160 * with a large weight task outweighs the tasks on the system).
6162 if (prefer_sibling && sds->local &&
6163 sds->local_stat.group_has_free_capacity)
6164 sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U);
6166 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6168 sds->busiest_stat = *sgs;
6172 /* Now, start updating sd_lb_stats */
6173 sds->total_load += sgs->group_load;
6174 sds->total_capacity += sgs->group_capacity;
6177 } while (sg != env->sd->groups);
6179 if (env->sd->flags & SD_NUMA)
6180 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6182 if (!env->sd->parent) {
6183 /* update overload indicator if we are at root domain */
6184 if (env->dst_rq->rd->overload != overload)
6185 env->dst_rq->rd->overload = overload;
6191 * check_asym_packing - Check to see if the group is packed into the
6194 * This is primarily intended to used at the sibling level. Some
6195 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6196 * case of POWER7, it can move to lower SMT modes only when higher
6197 * threads are idle. When in lower SMT modes, the threads will
6198 * perform better since they share less core resources. Hence when we
6199 * have idle threads, we want them to be the higher ones.
6201 * This packing function is run on idle threads. It checks to see if
6202 * the busiest CPU in this domain (core in the P7 case) has a higher
6203 * CPU number than the packing function is being run on. Here we are
6204 * assuming lower CPU number will be equivalent to lower a SMT thread
6207 * Return: 1 when packing is required and a task should be moved to
6208 * this CPU. The amount of the imbalance is returned in *imbalance.
6210 * @env: The load balancing environment.
6211 * @sds: Statistics of the sched_domain which is to be packed
6213 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6217 if (!(env->sd->flags & SD_ASYM_PACKING))
6223 busiest_cpu = group_first_cpu(sds->busiest);
6224 if (env->dst_cpu > busiest_cpu)
6227 env->imbalance = DIV_ROUND_CLOSEST(
6228 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6229 SCHED_CAPACITY_SCALE);
6235 * fix_small_imbalance - Calculate the minor imbalance that exists
6236 * amongst the groups of a sched_domain, during
6238 * @env: The load balancing environment.
6239 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6242 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6244 unsigned long tmp, capa_now = 0, capa_move = 0;
6245 unsigned int imbn = 2;
6246 unsigned long scaled_busy_load_per_task;
6247 struct sg_lb_stats *local, *busiest;
6249 local = &sds->local_stat;
6250 busiest = &sds->busiest_stat;
6252 if (!local->sum_nr_running)
6253 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6254 else if (busiest->load_per_task > local->load_per_task)
6257 scaled_busy_load_per_task =
6258 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6259 busiest->group_capacity;
6261 if (busiest->avg_load + scaled_busy_load_per_task >=
6262 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6263 env->imbalance = busiest->load_per_task;
6268 * OK, we don't have enough imbalance to justify moving tasks,
6269 * however we may be able to increase total CPU capacity used by
6273 capa_now += busiest->group_capacity *
6274 min(busiest->load_per_task, busiest->avg_load);
6275 capa_now += local->group_capacity *
6276 min(local->load_per_task, local->avg_load);
6277 capa_now /= SCHED_CAPACITY_SCALE;
6279 /* Amount of load we'd subtract */
6280 if (busiest->avg_load > scaled_busy_load_per_task) {
6281 capa_move += busiest->group_capacity *
6282 min(busiest->load_per_task,
6283 busiest->avg_load - scaled_busy_load_per_task);
6286 /* Amount of load we'd add */
6287 if (busiest->avg_load * busiest->group_capacity <
6288 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6289 tmp = (busiest->avg_load * busiest->group_capacity) /
6290 local->group_capacity;
6292 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6293 local->group_capacity;
6295 capa_move += local->group_capacity *
6296 min(local->load_per_task, local->avg_load + tmp);
6297 capa_move /= SCHED_CAPACITY_SCALE;
6299 /* Move if we gain throughput */
6300 if (capa_move > capa_now)
6301 env->imbalance = busiest->load_per_task;
6305 * calculate_imbalance - Calculate the amount of imbalance present within the
6306 * groups of a given sched_domain during load balance.
6307 * @env: load balance environment
6308 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6310 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6312 unsigned long max_pull, load_above_capacity = ~0UL;
6313 struct sg_lb_stats *local, *busiest;
6315 local = &sds->local_stat;
6316 busiest = &sds->busiest_stat;
6318 if (busiest->group_type == group_imbalanced) {
6320 * In the group_imb case we cannot rely on group-wide averages
6321 * to ensure cpu-load equilibrium, look at wider averages. XXX
6323 busiest->load_per_task =
6324 min(busiest->load_per_task, sds->avg_load);
6328 * In the presence of smp nice balancing, certain scenarios can have
6329 * max load less than avg load(as we skip the groups at or below
6330 * its cpu_capacity, while calculating max_load..)
6332 if (busiest->avg_load <= sds->avg_load ||
6333 local->avg_load >= sds->avg_load) {
6335 return fix_small_imbalance(env, sds);
6339 * If there aren't any idle cpus, avoid creating some.
6341 if (busiest->group_type == group_overloaded &&
6342 local->group_type == group_overloaded) {
6343 load_above_capacity =
6344 (busiest->sum_nr_running - busiest->group_capacity_factor);
6346 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE);
6347 load_above_capacity /= busiest->group_capacity;
6351 * We're trying to get all the cpus to the average_load, so we don't
6352 * want to push ourselves above the average load, nor do we wish to
6353 * reduce the max loaded cpu below the average load. At the same time,
6354 * we also don't want to reduce the group load below the group capacity
6355 * (so that we can implement power-savings policies etc). Thus we look
6356 * for the minimum possible imbalance.
6358 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6360 /* How much load to actually move to equalise the imbalance */
6361 env->imbalance = min(
6362 max_pull * busiest->group_capacity,
6363 (sds->avg_load - local->avg_load) * local->group_capacity
6364 ) / SCHED_CAPACITY_SCALE;
6367 * if *imbalance is less than the average load per runnable task
6368 * there is no guarantee that any tasks will be moved so we'll have
6369 * a think about bumping its value to force at least one task to be
6372 if (env->imbalance < busiest->load_per_task)
6373 return fix_small_imbalance(env, sds);
6376 /******* find_busiest_group() helpers end here *********************/
6379 * find_busiest_group - Returns the busiest group within the sched_domain
6380 * if there is an imbalance. If there isn't an imbalance, and
6381 * the user has opted for power-savings, it returns a group whose
6382 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6383 * such a group exists.
6385 * Also calculates the amount of weighted load which should be moved
6386 * to restore balance.
6388 * @env: The load balancing environment.
6390 * Return: - The busiest group if imbalance exists.
6391 * - If no imbalance and user has opted for power-savings balance,
6392 * return the least loaded group whose CPUs can be
6393 * put to idle by rebalancing its tasks onto our group.
6395 static struct sched_group *find_busiest_group(struct lb_env *env)
6397 struct sg_lb_stats *local, *busiest;
6398 struct sd_lb_stats sds;
6400 init_sd_lb_stats(&sds);
6403 * Compute the various statistics relavent for load balancing at
6406 update_sd_lb_stats(env, &sds);
6407 local = &sds.local_stat;
6408 busiest = &sds.busiest_stat;
6410 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6411 check_asym_packing(env, &sds))
6414 /* There is no busy sibling group to pull tasks from */
6415 if (!sds.busiest || busiest->sum_nr_running == 0)
6418 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6419 / sds.total_capacity;
6422 * If the busiest group is imbalanced the below checks don't
6423 * work because they assume all things are equal, which typically
6424 * isn't true due to cpus_allowed constraints and the like.
6426 if (busiest->group_type == group_imbalanced)
6429 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6430 if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity &&
6431 !busiest->group_has_free_capacity)
6435 * If the local group is more busy than the selected busiest group
6436 * don't try and pull any tasks.
6438 if (local->avg_load >= busiest->avg_load)
6442 * Don't pull any tasks if this group is already above the domain
6445 if (local->avg_load >= sds.avg_load)
6448 if (env->idle == CPU_IDLE) {
6450 * This cpu is idle. If the busiest group load doesn't
6451 * have more tasks than the number of available cpu's and
6452 * there is no imbalance between this and busiest group
6453 * wrt to idle cpu's, it is balanced.
6455 if ((local->idle_cpus < busiest->idle_cpus) &&
6456 busiest->sum_nr_running <= busiest->group_weight)
6460 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6461 * imbalance_pct to be conservative.
6463 if (100 * busiest->avg_load <=
6464 env->sd->imbalance_pct * local->avg_load)
6469 /* Looks like there is an imbalance. Compute it */
6470 calculate_imbalance(env, &sds);
6479 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6481 static struct rq *find_busiest_queue(struct lb_env *env,
6482 struct sched_group *group)
6484 struct rq *busiest = NULL, *rq;
6485 unsigned long busiest_load = 0, busiest_capacity = 1;
6488 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6489 unsigned long capacity, capacity_factor, wl;
6493 rt = fbq_classify_rq(rq);
6496 * We classify groups/runqueues into three groups:
6497 * - regular: there are !numa tasks
6498 * - remote: there are numa tasks that run on the 'wrong' node
6499 * - all: there is no distinction
6501 * In order to avoid migrating ideally placed numa tasks,
6502 * ignore those when there's better options.
6504 * If we ignore the actual busiest queue to migrate another
6505 * task, the next balance pass can still reduce the busiest
6506 * queue by moving tasks around inside the node.
6508 * If we cannot move enough load due to this classification
6509 * the next pass will adjust the group classification and
6510 * allow migration of more tasks.
6512 * Both cases only affect the total convergence complexity.
6514 if (rt > env->fbq_type)
6517 capacity = capacity_of(i);
6518 capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE);
6519 if (!capacity_factor)
6520 capacity_factor = fix_small_capacity(env->sd, group);
6522 wl = weighted_cpuload(i);
6525 * When comparing with imbalance, use weighted_cpuload()
6526 * which is not scaled with the cpu capacity.
6528 if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance)
6532 * For the load comparisons with the other cpu's, consider
6533 * the weighted_cpuload() scaled with the cpu capacity, so
6534 * that the load can be moved away from the cpu that is
6535 * potentially running at a lower capacity.
6537 * Thus we're looking for max(wl_i / capacity_i), crosswise
6538 * multiplication to rid ourselves of the division works out
6539 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6540 * our previous maximum.
6542 if (wl * busiest_capacity > busiest_load * capacity) {
6544 busiest_capacity = capacity;
6553 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6554 * so long as it is large enough.
6556 #define MAX_PINNED_INTERVAL 512
6558 /* Working cpumask for load_balance and load_balance_newidle. */
6559 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6561 static int need_active_balance(struct lb_env *env)
6563 struct sched_domain *sd = env->sd;
6565 if (env->idle == CPU_NEWLY_IDLE) {
6568 * ASYM_PACKING needs to force migrate tasks from busy but
6569 * higher numbered CPUs in order to pack all tasks in the
6570 * lowest numbered CPUs.
6572 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6576 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6579 static int active_load_balance_cpu_stop(void *data);
6581 static int should_we_balance(struct lb_env *env)
6583 struct sched_group *sg = env->sd->groups;
6584 struct cpumask *sg_cpus, *sg_mask;
6585 int cpu, balance_cpu = -1;
6588 * In the newly idle case, we will allow all the cpu's
6589 * to do the newly idle load balance.
6591 if (env->idle == CPU_NEWLY_IDLE)
6594 sg_cpus = sched_group_cpus(sg);
6595 sg_mask = sched_group_mask(sg);
6596 /* Try to find first idle cpu */
6597 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6598 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6605 if (balance_cpu == -1)
6606 balance_cpu = group_balance_cpu(sg);
6609 * First idle cpu or the first cpu(busiest) in this sched group
6610 * is eligible for doing load balancing at this and above domains.
6612 return balance_cpu == env->dst_cpu;
6616 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6617 * tasks if there is an imbalance.
6619 static int load_balance(int this_cpu, struct rq *this_rq,
6620 struct sched_domain *sd, enum cpu_idle_type idle,
6621 int *continue_balancing)
6623 int ld_moved, cur_ld_moved, active_balance = 0;
6624 struct sched_domain *sd_parent = sd->parent;
6625 struct sched_group *group;
6627 unsigned long flags;
6628 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6630 struct lb_env env = {
6632 .dst_cpu = this_cpu,
6634 .dst_grpmask = sched_group_cpus(sd->groups),
6636 .loop_break = sched_nr_migrate_break,
6639 .tasks = LIST_HEAD_INIT(env.tasks),
6643 * For NEWLY_IDLE load_balancing, we don't need to consider
6644 * other cpus in our group
6646 if (idle == CPU_NEWLY_IDLE)
6647 env.dst_grpmask = NULL;
6649 cpumask_copy(cpus, cpu_active_mask);
6651 schedstat_inc(sd, lb_count[idle]);
6654 if (!should_we_balance(&env)) {
6655 *continue_balancing = 0;
6659 group = find_busiest_group(&env);
6661 schedstat_inc(sd, lb_nobusyg[idle]);
6665 busiest = find_busiest_queue(&env, group);
6667 schedstat_inc(sd, lb_nobusyq[idle]);
6671 BUG_ON(busiest == env.dst_rq);
6673 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6676 if (busiest->nr_running > 1) {
6678 * Attempt to move tasks. If find_busiest_group has found
6679 * an imbalance but busiest->nr_running <= 1, the group is
6680 * still unbalanced. ld_moved simply stays zero, so it is
6681 * correctly treated as an imbalance.
6683 env.flags |= LBF_ALL_PINNED;
6684 env.src_cpu = busiest->cpu;
6685 env.src_rq = busiest;
6686 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6689 raw_spin_lock_irqsave(&busiest->lock, flags);
6692 * cur_ld_moved - load moved in current iteration
6693 * ld_moved - cumulative load moved across iterations
6695 cur_ld_moved = detach_tasks(&env);
6698 * We've detached some tasks from busiest_rq. Every
6699 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6700 * unlock busiest->lock, and we are able to be sure
6701 * that nobody can manipulate the tasks in parallel.
6702 * See task_rq_lock() family for the details.
6705 raw_spin_unlock(&busiest->lock);
6709 ld_moved += cur_ld_moved;
6712 local_irq_restore(flags);
6715 * some other cpu did the load balance for us.
6717 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6718 resched_cpu(env.dst_cpu);
6720 if (env.flags & LBF_NEED_BREAK) {
6721 env.flags &= ~LBF_NEED_BREAK;
6726 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6727 * us and move them to an alternate dst_cpu in our sched_group
6728 * where they can run. The upper limit on how many times we
6729 * iterate on same src_cpu is dependent on number of cpus in our
6732 * This changes load balance semantics a bit on who can move
6733 * load to a given_cpu. In addition to the given_cpu itself
6734 * (or a ilb_cpu acting on its behalf where given_cpu is
6735 * nohz-idle), we now have balance_cpu in a position to move
6736 * load to given_cpu. In rare situations, this may cause
6737 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6738 * _independently_ and at _same_ time to move some load to
6739 * given_cpu) causing exceess load to be moved to given_cpu.
6740 * This however should not happen so much in practice and
6741 * moreover subsequent load balance cycles should correct the
6742 * excess load moved.
6744 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6746 /* Prevent to re-select dst_cpu via env's cpus */
6747 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6749 env.dst_rq = cpu_rq(env.new_dst_cpu);
6750 env.dst_cpu = env.new_dst_cpu;
6751 env.flags &= ~LBF_DST_PINNED;
6753 env.loop_break = sched_nr_migrate_break;
6756 * Go back to "more_balance" rather than "redo" since we
6757 * need to continue with same src_cpu.
6763 * We failed to reach balance because of affinity.
6766 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
6768 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
6769 *group_imbalance = 1;
6772 /* All tasks on this runqueue were pinned by CPU affinity */
6773 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6774 cpumask_clear_cpu(cpu_of(busiest), cpus);
6775 if (!cpumask_empty(cpus)) {
6777 env.loop_break = sched_nr_migrate_break;
6780 goto out_all_pinned;
6785 schedstat_inc(sd, lb_failed[idle]);
6787 * Increment the failure counter only on periodic balance.
6788 * We do not want newidle balance, which can be very
6789 * frequent, pollute the failure counter causing
6790 * excessive cache_hot migrations and active balances.
6792 if (idle != CPU_NEWLY_IDLE)
6793 sd->nr_balance_failed++;
6795 if (need_active_balance(&env)) {
6796 raw_spin_lock_irqsave(&busiest->lock, flags);
6798 /* don't kick the active_load_balance_cpu_stop,
6799 * if the curr task on busiest cpu can't be
6802 if (!cpumask_test_cpu(this_cpu,
6803 tsk_cpus_allowed(busiest->curr))) {
6804 raw_spin_unlock_irqrestore(&busiest->lock,
6806 env.flags |= LBF_ALL_PINNED;
6807 goto out_one_pinned;
6811 * ->active_balance synchronizes accesses to
6812 * ->active_balance_work. Once set, it's cleared
6813 * only after active load balance is finished.
6815 if (!busiest->active_balance) {
6816 busiest->active_balance = 1;
6817 busiest->push_cpu = this_cpu;
6820 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6822 if (active_balance) {
6823 stop_one_cpu_nowait(cpu_of(busiest),
6824 active_load_balance_cpu_stop, busiest,
6825 &busiest->active_balance_work);
6829 * We've kicked active balancing, reset the failure
6832 sd->nr_balance_failed = sd->cache_nice_tries+1;
6835 sd->nr_balance_failed = 0;
6837 if (likely(!active_balance)) {
6838 /* We were unbalanced, so reset the balancing interval */
6839 sd->balance_interval = sd->min_interval;
6842 * If we've begun active balancing, start to back off. This
6843 * case may not be covered by the all_pinned logic if there
6844 * is only 1 task on the busy runqueue (because we don't call
6847 if (sd->balance_interval < sd->max_interval)
6848 sd->balance_interval *= 2;
6855 * We reach balance although we may have faced some affinity
6856 * constraints. Clear the imbalance flag if it was set.
6859 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
6861 if (*group_imbalance)
6862 *group_imbalance = 0;
6867 * We reach balance because all tasks are pinned at this level so
6868 * we can't migrate them. Let the imbalance flag set so parent level
6869 * can try to migrate them.
6871 schedstat_inc(sd, lb_balanced[idle]);
6873 sd->nr_balance_failed = 0;
6876 /* tune up the balancing interval */
6877 if (((env.flags & LBF_ALL_PINNED) &&
6878 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6879 (sd->balance_interval < sd->max_interval))
6880 sd->balance_interval *= 2;
6887 static inline unsigned long
6888 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
6890 unsigned long interval = sd->balance_interval;
6893 interval *= sd->busy_factor;
6895 /* scale ms to jiffies */
6896 interval = msecs_to_jiffies(interval);
6897 interval = clamp(interval, 1UL, max_load_balance_interval);
6903 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
6905 unsigned long interval, next;
6907 interval = get_sd_balance_interval(sd, cpu_busy);
6908 next = sd->last_balance + interval;
6910 if (time_after(*next_balance, next))
6911 *next_balance = next;
6915 * idle_balance is called by schedule() if this_cpu is about to become
6916 * idle. Attempts to pull tasks from other CPUs.
6918 static int idle_balance(struct rq *this_rq)
6920 unsigned long next_balance = jiffies + HZ;
6921 int this_cpu = this_rq->cpu;
6922 struct sched_domain *sd;
6923 int pulled_task = 0;
6926 idle_enter_fair(this_rq);
6929 * We must set idle_stamp _before_ calling idle_balance(), such that we
6930 * measure the duration of idle_balance() as idle time.
6932 this_rq->idle_stamp = rq_clock(this_rq);
6934 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
6935 !this_rq->rd->overload) {
6937 sd = rcu_dereference_check_sched_domain(this_rq->sd);
6939 update_next_balance(sd, 0, &next_balance);
6946 * Drop the rq->lock, but keep IRQ/preempt disabled.
6948 raw_spin_unlock(&this_rq->lock);
6950 update_blocked_averages(this_cpu);
6952 for_each_domain(this_cpu, sd) {
6953 int continue_balancing = 1;
6954 u64 t0, domain_cost;
6956 if (!(sd->flags & SD_LOAD_BALANCE))
6959 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
6960 update_next_balance(sd, 0, &next_balance);
6964 if (sd->flags & SD_BALANCE_NEWIDLE) {
6965 t0 = sched_clock_cpu(this_cpu);
6967 pulled_task = load_balance(this_cpu, this_rq,
6969 &continue_balancing);
6971 domain_cost = sched_clock_cpu(this_cpu) - t0;
6972 if (domain_cost > sd->max_newidle_lb_cost)
6973 sd->max_newidle_lb_cost = domain_cost;
6975 curr_cost += domain_cost;
6978 update_next_balance(sd, 0, &next_balance);
6981 * Stop searching for tasks to pull if there are
6982 * now runnable tasks on this rq.
6984 if (pulled_task || this_rq->nr_running > 0)
6989 raw_spin_lock(&this_rq->lock);
6991 if (curr_cost > this_rq->max_idle_balance_cost)
6992 this_rq->max_idle_balance_cost = curr_cost;
6995 * While browsing the domains, we released the rq lock, a task could
6996 * have been enqueued in the meantime. Since we're not going idle,
6997 * pretend we pulled a task.
6999 if (this_rq->cfs.h_nr_running && !pulled_task)
7003 /* Move the next balance forward */
7004 if (time_after(this_rq->next_balance, next_balance))
7005 this_rq->next_balance = next_balance;
7007 /* Is there a task of a high priority class? */
7008 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7012 idle_exit_fair(this_rq);
7013 this_rq->idle_stamp = 0;
7020 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7021 * running tasks off the busiest CPU onto idle CPUs. It requires at
7022 * least 1 task to be running on each physical CPU where possible, and
7023 * avoids physical / logical imbalances.
7025 static int active_load_balance_cpu_stop(void *data)
7027 struct rq *busiest_rq = data;
7028 int busiest_cpu = cpu_of(busiest_rq);
7029 int target_cpu = busiest_rq->push_cpu;
7030 struct rq *target_rq = cpu_rq(target_cpu);
7031 struct sched_domain *sd;
7032 struct task_struct *p = NULL;
7034 raw_spin_lock_irq(&busiest_rq->lock);
7036 /* make sure the requested cpu hasn't gone down in the meantime */
7037 if (unlikely(busiest_cpu != smp_processor_id() ||
7038 !busiest_rq->active_balance))
7041 /* Is there any task to move? */
7042 if (busiest_rq->nr_running <= 1)
7046 * This condition is "impossible", if it occurs
7047 * we need to fix it. Originally reported by
7048 * Bjorn Helgaas on a 128-cpu setup.
7050 BUG_ON(busiest_rq == target_rq);
7052 /* Search for an sd spanning us and the target CPU. */
7054 for_each_domain(target_cpu, sd) {
7055 if ((sd->flags & SD_LOAD_BALANCE) &&
7056 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7061 struct lb_env env = {
7063 .dst_cpu = target_cpu,
7064 .dst_rq = target_rq,
7065 .src_cpu = busiest_rq->cpu,
7066 .src_rq = busiest_rq,
7070 schedstat_inc(sd, alb_count);
7072 p = detach_one_task(&env);
7074 schedstat_inc(sd, alb_pushed);
7076 schedstat_inc(sd, alb_failed);
7080 busiest_rq->active_balance = 0;
7081 raw_spin_unlock(&busiest_rq->lock);
7084 attach_one_task(target_rq, p);
7091 static inline int on_null_domain(struct rq *rq)
7093 return unlikely(!rcu_dereference_sched(rq->sd));
7096 #ifdef CONFIG_NO_HZ_COMMON
7098 * idle load balancing details
7099 * - When one of the busy CPUs notice that there may be an idle rebalancing
7100 * needed, they will kick the idle load balancer, which then does idle
7101 * load balancing for all the idle CPUs.
7104 cpumask_var_t idle_cpus_mask;
7106 unsigned long next_balance; /* in jiffy units */
7107 } nohz ____cacheline_aligned;
7109 static inline int find_new_ilb(void)
7111 int ilb = cpumask_first(nohz.idle_cpus_mask);
7113 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7120 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7121 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7122 * CPU (if there is one).
7124 static void nohz_balancer_kick(void)
7128 nohz.next_balance++;
7130 ilb_cpu = find_new_ilb();
7132 if (ilb_cpu >= nr_cpu_ids)
7135 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7138 * Use smp_send_reschedule() instead of resched_cpu().
7139 * This way we generate a sched IPI on the target cpu which
7140 * is idle. And the softirq performing nohz idle load balance
7141 * will be run before returning from the IPI.
7143 smp_send_reschedule(ilb_cpu);
7147 static inline void nohz_balance_exit_idle(int cpu)
7149 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7151 * Completely isolated CPUs don't ever set, so we must test.
7153 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7154 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7155 atomic_dec(&nohz.nr_cpus);
7157 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7161 static inline void set_cpu_sd_state_busy(void)
7163 struct sched_domain *sd;
7164 int cpu = smp_processor_id();
7167 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7169 if (!sd || !sd->nohz_idle)
7173 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7178 void set_cpu_sd_state_idle(void)
7180 struct sched_domain *sd;
7181 int cpu = smp_processor_id();
7184 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7186 if (!sd || sd->nohz_idle)
7190 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7196 * This routine will record that the cpu is going idle with tick stopped.
7197 * This info will be used in performing idle load balancing in the future.
7199 void nohz_balance_enter_idle(int cpu)
7202 * If this cpu is going down, then nothing needs to be done.
7204 if (!cpu_active(cpu))
7207 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7211 * If we're a completely isolated CPU, we don't play.
7213 if (on_null_domain(cpu_rq(cpu)))
7216 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7217 atomic_inc(&nohz.nr_cpus);
7218 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7221 static int sched_ilb_notifier(struct notifier_block *nfb,
7222 unsigned long action, void *hcpu)
7224 switch (action & ~CPU_TASKS_FROZEN) {
7226 nohz_balance_exit_idle(smp_processor_id());
7234 static DEFINE_SPINLOCK(balancing);
7237 * Scale the max load_balance interval with the number of CPUs in the system.
7238 * This trades load-balance latency on larger machines for less cross talk.
7240 void update_max_interval(void)
7242 max_load_balance_interval = HZ*num_online_cpus()/10;
7246 * It checks each scheduling domain to see if it is due to be balanced,
7247 * and initiates a balancing operation if so.
7249 * Balancing parameters are set up in init_sched_domains.
7251 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7253 int continue_balancing = 1;
7255 unsigned long interval;
7256 struct sched_domain *sd;
7257 /* Earliest time when we have to do rebalance again */
7258 unsigned long next_balance = jiffies + 60*HZ;
7259 int update_next_balance = 0;
7260 int need_serialize, need_decay = 0;
7263 update_blocked_averages(cpu);
7266 for_each_domain(cpu, sd) {
7268 * Decay the newidle max times here because this is a regular
7269 * visit to all the domains. Decay ~1% per second.
7271 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7272 sd->max_newidle_lb_cost =
7273 (sd->max_newidle_lb_cost * 253) / 256;
7274 sd->next_decay_max_lb_cost = jiffies + HZ;
7277 max_cost += sd->max_newidle_lb_cost;
7279 if (!(sd->flags & SD_LOAD_BALANCE))
7283 * Stop the load balance at this level. There is another
7284 * CPU in our sched group which is doing load balancing more
7287 if (!continue_balancing) {
7293 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7295 need_serialize = sd->flags & SD_SERIALIZE;
7296 if (need_serialize) {
7297 if (!spin_trylock(&balancing))
7301 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7302 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7304 * The LBF_DST_PINNED logic could have changed
7305 * env->dst_cpu, so we can't know our idle
7306 * state even if we migrated tasks. Update it.
7308 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7310 sd->last_balance = jiffies;
7311 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7314 spin_unlock(&balancing);
7316 if (time_after(next_balance, sd->last_balance + interval)) {
7317 next_balance = sd->last_balance + interval;
7318 update_next_balance = 1;
7323 * Ensure the rq-wide value also decays but keep it at a
7324 * reasonable floor to avoid funnies with rq->avg_idle.
7326 rq->max_idle_balance_cost =
7327 max((u64)sysctl_sched_migration_cost, max_cost);
7332 * next_balance will be updated only when there is a need.
7333 * When the cpu is attached to null domain for ex, it will not be
7336 if (likely(update_next_balance))
7337 rq->next_balance = next_balance;
7340 #ifdef CONFIG_NO_HZ_COMMON
7342 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7343 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7345 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7347 int this_cpu = this_rq->cpu;
7351 if (idle != CPU_IDLE ||
7352 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7355 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7356 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7360 * If this cpu gets work to do, stop the load balancing
7361 * work being done for other cpus. Next load
7362 * balancing owner will pick it up.
7367 rq = cpu_rq(balance_cpu);
7370 * If time for next balance is due,
7373 if (time_after_eq(jiffies, rq->next_balance)) {
7374 raw_spin_lock_irq(&rq->lock);
7375 update_rq_clock(rq);
7376 update_idle_cpu_load(rq);
7377 raw_spin_unlock_irq(&rq->lock);
7378 rebalance_domains(rq, CPU_IDLE);
7381 if (time_after(this_rq->next_balance, rq->next_balance))
7382 this_rq->next_balance = rq->next_balance;
7384 nohz.next_balance = this_rq->next_balance;
7386 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7390 * Current heuristic for kicking the idle load balancer in the presence
7391 * of an idle cpu is the system.
7392 * - This rq has more than one task.
7393 * - At any scheduler domain level, this cpu's scheduler group has multiple
7394 * busy cpu's exceeding the group's capacity.
7395 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7396 * domain span are idle.
7398 static inline int nohz_kick_needed(struct rq *rq)
7400 unsigned long now = jiffies;
7401 struct sched_domain *sd;
7402 struct sched_group_capacity *sgc;
7403 int nr_busy, cpu = rq->cpu;
7405 if (unlikely(rq->idle_balance))
7409 * We may be recently in ticked or tickless idle mode. At the first
7410 * busy tick after returning from idle, we will update the busy stats.
7412 set_cpu_sd_state_busy();
7413 nohz_balance_exit_idle(cpu);
7416 * None are in tickless mode and hence no need for NOHZ idle load
7419 if (likely(!atomic_read(&nohz.nr_cpus)))
7422 if (time_before(now, nohz.next_balance))
7425 if (rq->nr_running >= 2)
7429 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7432 sgc = sd->groups->sgc;
7433 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7436 goto need_kick_unlock;
7439 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7441 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7442 sched_domain_span(sd)) < cpu))
7443 goto need_kick_unlock;
7454 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7458 * run_rebalance_domains is triggered when needed from the scheduler tick.
7459 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7461 static void run_rebalance_domains(struct softirq_action *h)
7463 struct rq *this_rq = this_rq();
7464 enum cpu_idle_type idle = this_rq->idle_balance ?
7465 CPU_IDLE : CPU_NOT_IDLE;
7467 rebalance_domains(this_rq, idle);
7470 * If this cpu has a pending nohz_balance_kick, then do the
7471 * balancing on behalf of the other idle cpus whose ticks are
7474 nohz_idle_balance(this_rq, idle);
7478 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7480 void trigger_load_balance(struct rq *rq)
7482 /* Don't need to rebalance while attached to NULL domain */
7483 if (unlikely(on_null_domain(rq)))
7486 if (time_after_eq(jiffies, rq->next_balance))
7487 raise_softirq(SCHED_SOFTIRQ);
7488 #ifdef CONFIG_NO_HZ_COMMON
7489 if (nohz_kick_needed(rq))
7490 nohz_balancer_kick();
7494 static void rq_online_fair(struct rq *rq)
7498 update_runtime_enabled(rq);
7501 static void rq_offline_fair(struct rq *rq)
7505 /* Ensure any throttled groups are reachable by pick_next_task */
7506 unthrottle_offline_cfs_rqs(rq);
7509 #endif /* CONFIG_SMP */
7512 * scheduler tick hitting a task of our scheduling class:
7514 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7516 struct cfs_rq *cfs_rq;
7517 struct sched_entity *se = &curr->se;
7519 for_each_sched_entity(se) {
7520 cfs_rq = cfs_rq_of(se);
7521 entity_tick(cfs_rq, se, queued);
7524 if (numabalancing_enabled)
7525 task_tick_numa(rq, curr);
7527 update_rq_runnable_avg(rq, 1);
7531 * called on fork with the child task as argument from the parent's context
7532 * - child not yet on the tasklist
7533 * - preemption disabled
7535 static void task_fork_fair(struct task_struct *p)
7537 struct cfs_rq *cfs_rq;
7538 struct sched_entity *se = &p->se, *curr;
7539 int this_cpu = smp_processor_id();
7540 struct rq *rq = this_rq();
7541 unsigned long flags;
7543 raw_spin_lock_irqsave(&rq->lock, flags);
7545 update_rq_clock(rq);
7547 cfs_rq = task_cfs_rq(current);
7548 curr = cfs_rq->curr;
7551 * Not only the cpu but also the task_group of the parent might have
7552 * been changed after parent->se.parent,cfs_rq were copied to
7553 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7554 * of child point to valid ones.
7557 __set_task_cpu(p, this_cpu);
7560 update_curr(cfs_rq);
7563 se->vruntime = curr->vruntime;
7564 place_entity(cfs_rq, se, 1);
7566 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7568 * Upon rescheduling, sched_class::put_prev_task() will place
7569 * 'current' within the tree based on its new key value.
7571 swap(curr->vruntime, se->vruntime);
7575 se->vruntime -= cfs_rq->min_vruntime;
7577 raw_spin_unlock_irqrestore(&rq->lock, flags);
7581 * Priority of the task has changed. Check to see if we preempt
7585 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7587 if (!task_on_rq_queued(p))
7591 * Reschedule if we are currently running on this runqueue and
7592 * our priority decreased, or if we are not currently running on
7593 * this runqueue and our priority is higher than the current's
7595 if (rq->curr == p) {
7596 if (p->prio > oldprio)
7599 check_preempt_curr(rq, p, 0);
7602 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7604 struct sched_entity *se = &p->se;
7605 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7608 * Ensure the task's vruntime is normalized, so that when it's
7609 * switched back to the fair class the enqueue_entity(.flags=0) will
7610 * do the right thing.
7612 * If it's queued, then the dequeue_entity(.flags=0) will already
7613 * have normalized the vruntime, if it's !queued, then only when
7614 * the task is sleeping will it still have non-normalized vruntime.
7616 if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) {
7618 * Fix up our vruntime so that the current sleep doesn't
7619 * cause 'unlimited' sleep bonus.
7621 place_entity(cfs_rq, se, 0);
7622 se->vruntime -= cfs_rq->min_vruntime;
7627 * Remove our load from contribution when we leave sched_fair
7628 * and ensure we don't carry in an old decay_count if we
7631 if (se->avg.decay_count) {
7632 __synchronize_entity_decay(se);
7633 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7639 * We switched to the sched_fair class.
7641 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7643 #ifdef CONFIG_FAIR_GROUP_SCHED
7644 struct sched_entity *se = &p->se;
7646 * Since the real-depth could have been changed (only FAIR
7647 * class maintain depth value), reset depth properly.
7649 se->depth = se->parent ? se->parent->depth + 1 : 0;
7651 if (!task_on_rq_queued(p))
7655 * We were most likely switched from sched_rt, so
7656 * kick off the schedule if running, otherwise just see
7657 * if we can still preempt the current task.
7662 check_preempt_curr(rq, p, 0);
7665 /* Account for a task changing its policy or group.
7667 * This routine is mostly called to set cfs_rq->curr field when a task
7668 * migrates between groups/classes.
7670 static void set_curr_task_fair(struct rq *rq)
7672 struct sched_entity *se = &rq->curr->se;
7674 for_each_sched_entity(se) {
7675 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7677 set_next_entity(cfs_rq, se);
7678 /* ensure bandwidth has been allocated on our new cfs_rq */
7679 account_cfs_rq_runtime(cfs_rq, 0);
7683 void init_cfs_rq(struct cfs_rq *cfs_rq)
7685 cfs_rq->tasks_timeline = RB_ROOT;
7686 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7687 #ifndef CONFIG_64BIT
7688 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7691 atomic64_set(&cfs_rq->decay_counter, 1);
7692 atomic_long_set(&cfs_rq->removed_load, 0);
7696 #ifdef CONFIG_FAIR_GROUP_SCHED
7697 static void task_move_group_fair(struct task_struct *p, int queued)
7699 struct sched_entity *se = &p->se;
7700 struct cfs_rq *cfs_rq;
7703 * If the task was not on the rq at the time of this cgroup movement
7704 * it must have been asleep, sleeping tasks keep their ->vruntime
7705 * absolute on their old rq until wakeup (needed for the fair sleeper
7706 * bonus in place_entity()).
7708 * If it was on the rq, we've just 'preempted' it, which does convert
7709 * ->vruntime to a relative base.
7711 * Make sure both cases convert their relative position when migrating
7712 * to another cgroup's rq. This does somewhat interfere with the
7713 * fair sleeper stuff for the first placement, but who cares.
7716 * When !queued, vruntime of the task has usually NOT been normalized.
7717 * But there are some cases where it has already been normalized:
7719 * - Moving a forked child which is waiting for being woken up by
7720 * wake_up_new_task().
7721 * - Moving a task which has been woken up by try_to_wake_up() and
7722 * waiting for actually being woken up by sched_ttwu_pending().
7724 * To prevent boost or penalty in the new cfs_rq caused by delta
7725 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7727 if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7731 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7732 set_task_rq(p, task_cpu(p));
7733 se->depth = se->parent ? se->parent->depth + 1 : 0;
7735 cfs_rq = cfs_rq_of(se);
7736 se->vruntime += cfs_rq->min_vruntime;
7739 * migrate_task_rq_fair() will have removed our previous
7740 * contribution, but we must synchronize for ongoing future
7743 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7744 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7749 void free_fair_sched_group(struct task_group *tg)
7753 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7755 for_each_possible_cpu(i) {
7757 kfree(tg->cfs_rq[i]);
7766 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7768 struct cfs_rq *cfs_rq;
7769 struct sched_entity *se;
7772 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7775 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7779 tg->shares = NICE_0_LOAD;
7781 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7783 for_each_possible_cpu(i) {
7784 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7785 GFP_KERNEL, cpu_to_node(i));
7789 se = kzalloc_node(sizeof(struct sched_entity),
7790 GFP_KERNEL, cpu_to_node(i));
7794 init_cfs_rq(cfs_rq);
7795 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7806 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7808 struct rq *rq = cpu_rq(cpu);
7809 unsigned long flags;
7812 * Only empty task groups can be destroyed; so we can speculatively
7813 * check on_list without danger of it being re-added.
7815 if (!tg->cfs_rq[cpu]->on_list)
7818 raw_spin_lock_irqsave(&rq->lock, flags);
7819 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7820 raw_spin_unlock_irqrestore(&rq->lock, flags);
7823 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7824 struct sched_entity *se, int cpu,
7825 struct sched_entity *parent)
7827 struct rq *rq = cpu_rq(cpu);
7831 init_cfs_rq_runtime(cfs_rq);
7833 tg->cfs_rq[cpu] = cfs_rq;
7836 /* se could be NULL for root_task_group */
7841 se->cfs_rq = &rq->cfs;
7844 se->cfs_rq = parent->my_q;
7845 se->depth = parent->depth + 1;
7849 /* guarantee group entities always have weight */
7850 update_load_set(&se->load, NICE_0_LOAD);
7851 se->parent = parent;
7854 static DEFINE_MUTEX(shares_mutex);
7856 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7859 unsigned long flags;
7862 * We can't change the weight of the root cgroup.
7867 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7869 mutex_lock(&shares_mutex);
7870 if (tg->shares == shares)
7873 tg->shares = shares;
7874 for_each_possible_cpu(i) {
7875 struct rq *rq = cpu_rq(i);
7876 struct sched_entity *se;
7879 /* Propagate contribution to hierarchy */
7880 raw_spin_lock_irqsave(&rq->lock, flags);
7882 /* Possible calls to update_curr() need rq clock */
7883 update_rq_clock(rq);
7884 for_each_sched_entity(se)
7885 update_cfs_shares(group_cfs_rq(se));
7886 raw_spin_unlock_irqrestore(&rq->lock, flags);
7890 mutex_unlock(&shares_mutex);
7893 #else /* CONFIG_FAIR_GROUP_SCHED */
7895 void free_fair_sched_group(struct task_group *tg) { }
7897 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7902 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7904 #endif /* CONFIG_FAIR_GROUP_SCHED */
7907 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7909 struct sched_entity *se = &task->se;
7910 unsigned int rr_interval = 0;
7913 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7916 if (rq->cfs.load.weight)
7917 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7923 * All the scheduling class methods:
7925 const struct sched_class fair_sched_class = {
7926 .next = &idle_sched_class,
7927 .enqueue_task = enqueue_task_fair,
7928 .dequeue_task = dequeue_task_fair,
7929 .yield_task = yield_task_fair,
7930 .yield_to_task = yield_to_task_fair,
7932 .check_preempt_curr = check_preempt_wakeup,
7934 .pick_next_task = pick_next_task_fair,
7935 .put_prev_task = put_prev_task_fair,
7938 .select_task_rq = select_task_rq_fair,
7939 .migrate_task_rq = migrate_task_rq_fair,
7941 .rq_online = rq_online_fair,
7942 .rq_offline = rq_offline_fair,
7944 .task_waking = task_waking_fair,
7947 .set_curr_task = set_curr_task_fair,
7948 .task_tick = task_tick_fair,
7949 .task_fork = task_fork_fair,
7951 .prio_changed = prio_changed_fair,
7952 .switched_from = switched_from_fair,
7953 .switched_to = switched_to_fair,
7955 .get_rr_interval = get_rr_interval_fair,
7957 #ifdef CONFIG_FAIR_GROUP_SCHED
7958 .task_move_group = task_move_group_fair,
7962 #ifdef CONFIG_SCHED_DEBUG
7963 void print_cfs_stats(struct seq_file *m, int cpu)
7965 struct cfs_rq *cfs_rq;
7968 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7969 print_cfs_rq(m, cpu, cfs_rq);
7974 __init void init_sched_fair_class(void)
7977 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7979 #ifdef CONFIG_NO_HZ_COMMON
7980 nohz.next_balance = jiffies;
7981 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7982 cpu_notifier(sched_ilb_notifier, 0);