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 cfs_rq *
419 is_same_group(struct sched_entity *se, struct sched_entity *pse)
421 return cfs_rq_of(se); /* always the same rq */
424 static inline struct sched_entity *parent_entity(struct sched_entity *se)
430 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
434 #endif /* CONFIG_FAIR_GROUP_SCHED */
436 static __always_inline
437 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
439 /**************************************************************
440 * Scheduling class tree data structure manipulation methods:
443 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
445 s64 delta = (s64)(vruntime - max_vruntime);
447 max_vruntime = vruntime;
452 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - min_vruntime);
456 min_vruntime = vruntime;
461 static inline int entity_before(struct sched_entity *a,
462 struct sched_entity *b)
464 return (s64)(a->vruntime - b->vruntime) < 0;
467 static void update_min_vruntime(struct cfs_rq *cfs_rq)
469 u64 vruntime = cfs_rq->min_vruntime;
472 vruntime = cfs_rq->curr->vruntime;
474 if (cfs_rq->rb_leftmost) {
475 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
480 vruntime = se->vruntime;
482 vruntime = min_vruntime(vruntime, se->vruntime);
485 /* ensure we never gain time by being placed backwards. */
486 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
489 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
494 * Enqueue an entity into the rb-tree:
496 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
498 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
499 struct rb_node *parent = NULL;
500 struct sched_entity *entry;
504 * Find the right place in the rbtree:
508 entry = rb_entry(parent, struct sched_entity, run_node);
510 * We dont care about collisions. Nodes with
511 * the same key stay together.
513 if (entity_before(se, entry)) {
514 link = &parent->rb_left;
516 link = &parent->rb_right;
522 * Maintain a cache of leftmost tree entries (it is frequently
526 cfs_rq->rb_leftmost = &se->run_node;
528 rb_link_node(&se->run_node, parent, link);
529 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
532 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
534 if (cfs_rq->rb_leftmost == &se->run_node) {
535 struct rb_node *next_node;
537 next_node = rb_next(&se->run_node);
538 cfs_rq->rb_leftmost = next_node;
541 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
544 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
546 struct rb_node *left = cfs_rq->rb_leftmost;
551 return rb_entry(left, struct sched_entity, run_node);
554 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
556 struct rb_node *next = rb_next(&se->run_node);
561 return rb_entry(next, struct sched_entity, run_node);
564 #ifdef CONFIG_SCHED_DEBUG
565 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
567 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
572 return rb_entry(last, struct sched_entity, run_node);
575 /**************************************************************
576 * Scheduling class statistics methods:
579 int sched_proc_update_handler(struct ctl_table *table, int write,
580 void __user *buffer, size_t *lenp,
583 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
584 int factor = get_update_sysctl_factor();
589 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
590 sysctl_sched_min_granularity);
592 #define WRT_SYSCTL(name) \
593 (normalized_sysctl_##name = sysctl_##name / (factor))
594 WRT_SYSCTL(sched_min_granularity);
595 WRT_SYSCTL(sched_latency);
596 WRT_SYSCTL(sched_wakeup_granularity);
606 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
608 if (unlikely(se->load.weight != NICE_0_LOAD))
609 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
615 * The idea is to set a period in which each task runs once.
617 * When there are too many tasks (sched_nr_latency) we have to stretch
618 * this period because otherwise the slices get too small.
620 * p = (nr <= nl) ? l : l*nr/nl
622 static u64 __sched_period(unsigned long nr_running)
624 u64 period = sysctl_sched_latency;
625 unsigned long nr_latency = sched_nr_latency;
627 if (unlikely(nr_running > nr_latency)) {
628 period = sysctl_sched_min_granularity;
629 period *= nr_running;
636 * We calculate the wall-time slice from the period by taking a part
637 * proportional to the weight.
641 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
643 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
645 for_each_sched_entity(se) {
646 struct load_weight *load;
647 struct load_weight lw;
649 cfs_rq = cfs_rq_of(se);
650 load = &cfs_rq->load;
652 if (unlikely(!se->on_rq)) {
655 update_load_add(&lw, se->load.weight);
658 slice = __calc_delta(slice, se->load.weight, load);
664 * We calculate the vruntime slice of a to-be-inserted task.
668 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
670 return calc_delta_fair(sched_slice(cfs_rq, se), se);
674 static unsigned long task_h_load(struct task_struct *p);
676 static inline void __update_task_entity_contrib(struct sched_entity *se);
678 /* Give new task start runnable values to heavy its load in infant time */
679 void init_task_runnable_average(struct task_struct *p)
683 p->se.avg.decay_count = 0;
684 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
685 p->se.avg.runnable_avg_sum = slice;
686 p->se.avg.runnable_avg_period = slice;
687 __update_task_entity_contrib(&p->se);
690 void init_task_runnable_average(struct task_struct *p)
696 * Update the current task's runtime statistics.
698 static void update_curr(struct cfs_rq *cfs_rq)
700 struct sched_entity *curr = cfs_rq->curr;
701 u64 now = rq_clock_task(rq_of(cfs_rq));
707 delta_exec = now - curr->exec_start;
708 if (unlikely((s64)delta_exec <= 0))
711 curr->exec_start = now;
713 schedstat_set(curr->statistics.exec_max,
714 max(delta_exec, curr->statistics.exec_max));
716 curr->sum_exec_runtime += delta_exec;
717 schedstat_add(cfs_rq, exec_clock, delta_exec);
719 curr->vruntime += calc_delta_fair(delta_exec, curr);
720 update_min_vruntime(cfs_rq);
722 if (entity_is_task(curr)) {
723 struct task_struct *curtask = task_of(curr);
725 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
726 cpuacct_charge(curtask, delta_exec);
727 account_group_exec_runtime(curtask, delta_exec);
730 account_cfs_rq_runtime(cfs_rq, delta_exec);
734 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
736 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
740 * Task is being enqueued - update stats:
742 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
745 * Are we enqueueing a waiting task? (for current tasks
746 * a dequeue/enqueue event is a NOP)
748 if (se != cfs_rq->curr)
749 update_stats_wait_start(cfs_rq, se);
753 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
755 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
756 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
757 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
758 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
759 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
760 #ifdef CONFIG_SCHEDSTATS
761 if (entity_is_task(se)) {
762 trace_sched_stat_wait(task_of(se),
763 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
766 schedstat_set(se->statistics.wait_start, 0);
770 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
773 * Mark the end of the wait period if dequeueing a
776 if (se != cfs_rq->curr)
777 update_stats_wait_end(cfs_rq, se);
781 * We are picking a new current task - update its stats:
784 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
787 * We are starting a new run period:
789 se->exec_start = rq_clock_task(rq_of(cfs_rq));
792 /**************************************************
793 * Scheduling class queueing methods:
796 #ifdef CONFIG_NUMA_BALANCING
798 * Approximate time to scan a full NUMA task in ms. The task scan period is
799 * calculated based on the tasks virtual memory size and
800 * numa_balancing_scan_size.
802 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
803 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
805 /* Portion of address space to scan in MB */
806 unsigned int sysctl_numa_balancing_scan_size = 256;
808 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
809 unsigned int sysctl_numa_balancing_scan_delay = 1000;
811 static unsigned int task_nr_scan_windows(struct task_struct *p)
813 unsigned long rss = 0;
814 unsigned long nr_scan_pages;
817 * Calculations based on RSS as non-present and empty pages are skipped
818 * by the PTE scanner and NUMA hinting faults should be trapped based
821 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
822 rss = get_mm_rss(p->mm);
826 rss = round_up(rss, nr_scan_pages);
827 return rss / nr_scan_pages;
830 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
831 #define MAX_SCAN_WINDOW 2560
833 static unsigned int task_scan_min(struct task_struct *p)
835 unsigned int scan, floor;
836 unsigned int windows = 1;
838 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
839 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
840 floor = 1000 / windows;
842 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
843 return max_t(unsigned int, floor, scan);
846 static unsigned int task_scan_max(struct task_struct *p)
848 unsigned int smin = task_scan_min(p);
851 /* Watch for min being lower than max due to floor calculations */
852 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
853 return max(smin, smax);
856 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
858 rq->nr_numa_running += (p->numa_preferred_nid != -1);
859 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
862 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
864 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
865 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
871 spinlock_t lock; /* nr_tasks, tasks */
874 struct list_head task_list;
877 nodemask_t active_nodes;
878 unsigned long total_faults;
880 * Faults_cpu is used to decide whether memory should move
881 * towards the CPU. As a consequence, these stats are weighted
882 * more by CPU use than by memory faults.
884 unsigned long *faults_cpu;
885 unsigned long faults[0];
888 /* Shared or private faults. */
889 #define NR_NUMA_HINT_FAULT_TYPES 2
891 /* Memory and CPU locality */
892 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
894 /* Averaged statistics, and temporary buffers. */
895 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
897 pid_t task_numa_group_id(struct task_struct *p)
899 return p->numa_group ? p->numa_group->gid : 0;
902 static inline int task_faults_idx(int nid, int priv)
904 return NR_NUMA_HINT_FAULT_TYPES * nid + priv;
907 static inline unsigned long task_faults(struct task_struct *p, int nid)
909 if (!p->numa_faults_memory)
912 return p->numa_faults_memory[task_faults_idx(nid, 0)] +
913 p->numa_faults_memory[task_faults_idx(nid, 1)];
916 static inline unsigned long group_faults(struct task_struct *p, int nid)
921 return p->numa_group->faults[task_faults_idx(nid, 0)] +
922 p->numa_group->faults[task_faults_idx(nid, 1)];
925 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
927 return group->faults_cpu[task_faults_idx(nid, 0)] +
928 group->faults_cpu[task_faults_idx(nid, 1)];
932 * These return the fraction of accesses done by a particular task, or
933 * task group, on a particular numa node. The group weight is given a
934 * larger multiplier, in order to group tasks together that are almost
935 * evenly spread out between numa nodes.
937 static inline unsigned long task_weight(struct task_struct *p, int nid)
939 unsigned long total_faults;
941 if (!p->numa_faults_memory)
944 total_faults = p->total_numa_faults;
949 return 1000 * task_faults(p, nid) / total_faults;
952 static inline unsigned long group_weight(struct task_struct *p, int nid)
954 if (!p->numa_group || !p->numa_group->total_faults)
957 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
960 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
961 int src_nid, int dst_cpu)
963 struct numa_group *ng = p->numa_group;
964 int dst_nid = cpu_to_node(dst_cpu);
965 int last_cpupid, this_cpupid;
967 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
970 * Multi-stage node selection is used in conjunction with a periodic
971 * migration fault to build a temporal task<->page relation. By using
972 * a two-stage filter we remove short/unlikely relations.
974 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
975 * a task's usage of a particular page (n_p) per total usage of this
976 * page (n_t) (in a given time-span) to a probability.
978 * Our periodic faults will sample this probability and getting the
979 * same result twice in a row, given these samples are fully
980 * independent, is then given by P(n)^2, provided our sample period
981 * is sufficiently short compared to the usage pattern.
983 * This quadric squishes small probabilities, making it less likely we
984 * act on an unlikely task<->page relation.
986 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
987 if (!cpupid_pid_unset(last_cpupid) &&
988 cpupid_to_nid(last_cpupid) != dst_nid)
991 /* Always allow migrate on private faults */
992 if (cpupid_match_pid(p, last_cpupid))
995 /* A shared fault, but p->numa_group has not been set up yet. */
1000 * Do not migrate if the destination is not a node that
1001 * is actively used by this numa group.
1003 if (!node_isset(dst_nid, ng->active_nodes))
1007 * Source is a node that is not actively used by this
1008 * numa group, while the destination is. Migrate.
1010 if (!node_isset(src_nid, ng->active_nodes))
1014 * Both source and destination are nodes in active
1015 * use by this numa group. Maximize memory bandwidth
1016 * by migrating from more heavily used groups, to less
1017 * heavily used ones, spreading the load around.
1018 * Use a 1/4 hysteresis to avoid spurious page movement.
1020 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1023 static unsigned long weighted_cpuload(const int cpu);
1024 static unsigned long source_load(int cpu, int type);
1025 static unsigned long target_load(int cpu, int type);
1026 static unsigned long power_of(int cpu);
1027 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1029 /* Cached statistics for all CPUs within a node */
1031 unsigned long nr_running;
1034 /* Total compute capacity of CPUs on a node */
1035 unsigned long power;
1037 /* Approximate capacity in terms of runnable tasks on a node */
1038 unsigned long capacity;
1043 * XXX borrowed from update_sg_lb_stats
1045 static void update_numa_stats(struct numa_stats *ns, int nid)
1049 memset(ns, 0, sizeof(*ns));
1050 for_each_cpu(cpu, cpumask_of_node(nid)) {
1051 struct rq *rq = cpu_rq(cpu);
1053 ns->nr_running += rq->nr_running;
1054 ns->load += weighted_cpuload(cpu);
1055 ns->power += power_of(cpu);
1061 * If we raced with hotplug and there are no CPUs left in our mask
1062 * the @ns structure is NULL'ed and task_numa_compare() will
1063 * not find this node attractive.
1065 * We'll either bail at !has_capacity, or we'll detect a huge imbalance
1071 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1072 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1073 ns->has_capacity = (ns->nr_running < ns->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;
1105 * This checks if the overall compute and NUMA accesses of the system would
1106 * be improved if the source tasks was migrated to the target dst_cpu taking
1107 * into account that it might be best if task running on the dst_cpu should
1108 * be exchanged with the source task
1110 static void task_numa_compare(struct task_numa_env *env,
1111 long taskimp, long groupimp)
1113 struct rq *src_rq = cpu_rq(env->src_cpu);
1114 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1115 struct task_struct *cur;
1116 long dst_load, src_load;
1118 long imp = (groupimp > 0) ? groupimp : taskimp;
1121 cur = ACCESS_ONCE(dst_rq->curr);
1122 if (cur->pid == 0) /* idle */
1126 * "imp" is the fault differential for the source task between the
1127 * source and destination node. Calculate the total differential for
1128 * the source task and potential destination task. The more negative
1129 * the value is, the more rmeote accesses that would be expected to
1130 * be incurred if the tasks were swapped.
1133 /* Skip this swap candidate if cannot move to the source cpu */
1134 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1138 * If dst and source tasks are in the same NUMA group, or not
1139 * in any group then look only at task weights.
1141 if (cur->numa_group == env->p->numa_group) {
1142 imp = taskimp + task_weight(cur, env->src_nid) -
1143 task_weight(cur, env->dst_nid);
1145 * Add some hysteresis to prevent swapping the
1146 * tasks within a group over tiny differences.
1148 if (cur->numa_group)
1152 * Compare the group weights. If a task is all by
1153 * itself (not part of a group), use the task weight
1156 if (env->p->numa_group)
1161 if (cur->numa_group)
1162 imp += group_weight(cur, env->src_nid) -
1163 group_weight(cur, env->dst_nid);
1165 imp += task_weight(cur, env->src_nid) -
1166 task_weight(cur, env->dst_nid);
1170 if (imp < env->best_imp)
1174 /* Is there capacity at our destination? */
1175 if (env->src_stats.has_capacity &&
1176 !env->dst_stats.has_capacity)
1182 /* Balance doesn't matter much if we're running a task per cpu */
1183 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1187 * In the overloaded case, try and keep the load balanced.
1190 dst_load = env->dst_stats.load;
1191 src_load = env->src_stats.load;
1193 /* XXX missing power terms */
1194 load = task_h_load(env->p);
1199 load = task_h_load(cur);
1204 /* make src_load the smaller */
1205 if (dst_load < src_load)
1206 swap(dst_load, src_load);
1208 if (src_load * env->imbalance_pct < dst_load * 100)
1212 task_numa_assign(env, cur, imp);
1217 static void task_numa_find_cpu(struct task_numa_env *env,
1218 long taskimp, long groupimp)
1222 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1223 /* Skip this CPU if the source task cannot migrate */
1224 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1228 task_numa_compare(env, taskimp, groupimp);
1232 static int task_numa_migrate(struct task_struct *p)
1234 struct task_numa_env env = {
1237 .src_cpu = task_cpu(p),
1238 .src_nid = task_node(p),
1240 .imbalance_pct = 112,
1246 struct sched_domain *sd;
1247 unsigned long taskweight, groupweight;
1249 long taskimp, groupimp;
1252 * Pick the lowest SD_NUMA domain, as that would have the smallest
1253 * imbalance and would be the first to start moving tasks about.
1255 * And we want to avoid any moving of tasks about, as that would create
1256 * random movement of tasks -- counter the numa conditions we're trying
1260 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1262 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1266 * Cpusets can break the scheduler domain tree into smaller
1267 * balance domains, some of which do not cross NUMA boundaries.
1268 * Tasks that are "trapped" in such domains cannot be migrated
1269 * elsewhere, so there is no point in (re)trying.
1271 if (unlikely(!sd)) {
1272 p->numa_preferred_nid = task_node(p);
1276 taskweight = task_weight(p, env.src_nid);
1277 groupweight = group_weight(p, env.src_nid);
1278 update_numa_stats(&env.src_stats, env.src_nid);
1279 env.dst_nid = p->numa_preferred_nid;
1280 taskimp = task_weight(p, env.dst_nid) - taskweight;
1281 groupimp = group_weight(p, env.dst_nid) - groupweight;
1282 update_numa_stats(&env.dst_stats, env.dst_nid);
1284 /* If the preferred nid has capacity, try to use it. */
1285 if (env.dst_stats.has_capacity)
1286 task_numa_find_cpu(&env, taskimp, groupimp);
1288 /* No space available on the preferred nid. Look elsewhere. */
1289 if (env.best_cpu == -1) {
1290 for_each_online_node(nid) {
1291 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1294 /* Only consider nodes where both task and groups benefit */
1295 taskimp = task_weight(p, nid) - taskweight;
1296 groupimp = group_weight(p, nid) - groupweight;
1297 if (taskimp < 0 && groupimp < 0)
1301 update_numa_stats(&env.dst_stats, env.dst_nid);
1302 task_numa_find_cpu(&env, taskimp, groupimp);
1306 /* No better CPU than the current one was found. */
1307 if (env.best_cpu == -1)
1310 sched_setnuma(p, env.dst_nid);
1313 * Reset the scan period if the task is being rescheduled on an
1314 * alternative node to recheck if the tasks is now properly placed.
1316 p->numa_scan_period = task_scan_min(p);
1318 if (env.best_task == NULL) {
1319 ret = migrate_task_to(p, env.best_cpu);
1321 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1325 ret = migrate_swap(p, env.best_task);
1327 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1328 put_task_struct(env.best_task);
1332 /* Attempt to migrate a task to a CPU on the preferred node. */
1333 static void numa_migrate_preferred(struct task_struct *p)
1335 /* This task has no NUMA fault statistics yet */
1336 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1339 /* Periodically retry migrating the task to the preferred node */
1340 p->numa_migrate_retry = jiffies + HZ;
1342 /* Success if task is already running on preferred CPU */
1343 if (task_node(p) == p->numa_preferred_nid)
1346 /* Otherwise, try migrate to a CPU on the preferred node */
1347 task_numa_migrate(p);
1351 * Find the nodes on which the workload is actively running. We do this by
1352 * tracking the nodes from which NUMA hinting faults are triggered. This can
1353 * be different from the set of nodes where the workload's memory is currently
1356 * The bitmask is used to make smarter decisions on when to do NUMA page
1357 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1358 * are added when they cause over 6/16 of the maximum number of faults, but
1359 * only removed when they drop below 3/16.
1361 static void update_numa_active_node_mask(struct numa_group *numa_group)
1363 unsigned long faults, max_faults = 0;
1366 for_each_online_node(nid) {
1367 faults = group_faults_cpu(numa_group, nid);
1368 if (faults > max_faults)
1369 max_faults = faults;
1372 for_each_online_node(nid) {
1373 faults = group_faults_cpu(numa_group, nid);
1374 if (!node_isset(nid, numa_group->active_nodes)) {
1375 if (faults > max_faults * 6 / 16)
1376 node_set(nid, numa_group->active_nodes);
1377 } else if (faults < max_faults * 3 / 16)
1378 node_clear(nid, numa_group->active_nodes);
1383 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1384 * increments. The more local the fault statistics are, the higher the scan
1385 * period will be for the next scan window. If local/remote ratio is below
1386 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1387 * scan period will decrease
1389 #define NUMA_PERIOD_SLOTS 10
1390 #define NUMA_PERIOD_THRESHOLD 3
1393 * Increase the scan period (slow down scanning) if the majority of
1394 * our memory is already on our local node, or if the majority of
1395 * the page accesses are shared with other processes.
1396 * Otherwise, decrease the scan period.
1398 static void update_task_scan_period(struct task_struct *p,
1399 unsigned long shared, unsigned long private)
1401 unsigned int period_slot;
1405 unsigned long remote = p->numa_faults_locality[0];
1406 unsigned long local = p->numa_faults_locality[1];
1409 * If there were no record hinting faults then either the task is
1410 * completely idle or all activity is areas that are not of interest
1411 * to automatic numa balancing. Scan slower
1413 if (local + shared == 0) {
1414 p->numa_scan_period = min(p->numa_scan_period_max,
1415 p->numa_scan_period << 1);
1417 p->mm->numa_next_scan = jiffies +
1418 msecs_to_jiffies(p->numa_scan_period);
1424 * Prepare to scale scan period relative to the current period.
1425 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1426 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1427 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1429 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1430 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1431 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1432 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1435 diff = slot * period_slot;
1437 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1440 * Scale scan rate increases based on sharing. There is an
1441 * inverse relationship between the degree of sharing and
1442 * the adjustment made to the scanning period. Broadly
1443 * speaking the intent is that there is little point
1444 * scanning faster if shared accesses dominate as it may
1445 * simply bounce migrations uselessly
1447 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1448 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1451 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1452 task_scan_min(p), task_scan_max(p));
1453 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1457 * Get the fraction of time the task has been running since the last
1458 * NUMA placement cycle. The scheduler keeps similar statistics, but
1459 * decays those on a 32ms period, which is orders of magnitude off
1460 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1461 * stats only if the task is so new there are no NUMA statistics yet.
1463 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1465 u64 runtime, delta, now;
1466 /* Use the start of this time slice to avoid calculations. */
1467 now = p->se.exec_start;
1468 runtime = p->se.sum_exec_runtime;
1470 if (p->last_task_numa_placement) {
1471 delta = runtime - p->last_sum_exec_runtime;
1472 *period = now - p->last_task_numa_placement;
1474 delta = p->se.avg.runnable_avg_sum;
1475 *period = p->se.avg.runnable_avg_period;
1478 p->last_sum_exec_runtime = runtime;
1479 p->last_task_numa_placement = now;
1484 static void task_numa_placement(struct task_struct *p)
1486 int seq, nid, max_nid = -1, max_group_nid = -1;
1487 unsigned long max_faults = 0, max_group_faults = 0;
1488 unsigned long fault_types[2] = { 0, 0 };
1489 unsigned long total_faults;
1490 u64 runtime, period;
1491 spinlock_t *group_lock = NULL;
1493 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1494 if (p->numa_scan_seq == seq)
1496 p->numa_scan_seq = seq;
1497 p->numa_scan_period_max = task_scan_max(p);
1499 total_faults = p->numa_faults_locality[0] +
1500 p->numa_faults_locality[1];
1501 runtime = numa_get_avg_runtime(p, &period);
1503 /* If the task is part of a group prevent parallel updates to group stats */
1504 if (p->numa_group) {
1505 group_lock = &p->numa_group->lock;
1506 spin_lock(group_lock);
1509 /* Find the node with the highest number of faults */
1510 for_each_online_node(nid) {
1511 unsigned long faults = 0, group_faults = 0;
1514 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1515 long diff, f_diff, f_weight;
1517 i = task_faults_idx(nid, priv);
1519 /* Decay existing window, copy faults since last scan */
1520 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1521 fault_types[priv] += p->numa_faults_buffer_memory[i];
1522 p->numa_faults_buffer_memory[i] = 0;
1525 * Normalize the faults_from, so all tasks in a group
1526 * count according to CPU use, instead of by the raw
1527 * number of faults. Tasks with little runtime have
1528 * little over-all impact on throughput, and thus their
1529 * faults are less important.
1531 f_weight = div64_u64(runtime << 16, period + 1);
1532 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1534 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1535 p->numa_faults_buffer_cpu[i] = 0;
1537 p->numa_faults_memory[i] += diff;
1538 p->numa_faults_cpu[i] += f_diff;
1539 faults += p->numa_faults_memory[i];
1540 p->total_numa_faults += diff;
1541 if (p->numa_group) {
1542 /* safe because we can only change our own group */
1543 p->numa_group->faults[i] += diff;
1544 p->numa_group->faults_cpu[i] += f_diff;
1545 p->numa_group->total_faults += diff;
1546 group_faults += p->numa_group->faults[i];
1550 if (faults > max_faults) {
1551 max_faults = faults;
1555 if (group_faults > max_group_faults) {
1556 max_group_faults = group_faults;
1557 max_group_nid = nid;
1561 update_task_scan_period(p, fault_types[0], fault_types[1]);
1563 if (p->numa_group) {
1564 update_numa_active_node_mask(p->numa_group);
1566 * If the preferred task and group nids are different,
1567 * iterate over the nodes again to find the best place.
1569 if (max_nid != max_group_nid) {
1570 unsigned long weight, max_weight = 0;
1572 for_each_online_node(nid) {
1573 weight = task_weight(p, nid) + group_weight(p, nid);
1574 if (weight > max_weight) {
1575 max_weight = weight;
1581 spin_unlock(group_lock);
1584 /* Preferred node as the node with the most faults */
1585 if (max_faults && max_nid != p->numa_preferred_nid) {
1586 /* Update the preferred nid and migrate task if possible */
1587 sched_setnuma(p, max_nid);
1588 numa_migrate_preferred(p);
1592 static inline int get_numa_group(struct numa_group *grp)
1594 return atomic_inc_not_zero(&grp->refcount);
1597 static inline void put_numa_group(struct numa_group *grp)
1599 if (atomic_dec_and_test(&grp->refcount))
1600 kfree_rcu(grp, rcu);
1603 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1606 struct numa_group *grp, *my_grp;
1607 struct task_struct *tsk;
1609 int cpu = cpupid_to_cpu(cpupid);
1612 if (unlikely(!p->numa_group)) {
1613 unsigned int size = sizeof(struct numa_group) +
1614 4*nr_node_ids*sizeof(unsigned long);
1616 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1620 atomic_set(&grp->refcount, 1);
1621 spin_lock_init(&grp->lock);
1622 INIT_LIST_HEAD(&grp->task_list);
1624 /* Second half of the array tracks nids where faults happen */
1625 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1628 node_set(task_node(current), grp->active_nodes);
1630 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1631 grp->faults[i] = p->numa_faults_memory[i];
1633 grp->total_faults = p->total_numa_faults;
1635 list_add(&p->numa_entry, &grp->task_list);
1637 rcu_assign_pointer(p->numa_group, grp);
1641 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1643 if (!cpupid_match_pid(tsk, cpupid))
1646 grp = rcu_dereference(tsk->numa_group);
1650 my_grp = p->numa_group;
1655 * Only join the other group if its bigger; if we're the bigger group,
1656 * the other task will join us.
1658 if (my_grp->nr_tasks > grp->nr_tasks)
1662 * Tie-break on the grp address.
1664 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1667 /* Always join threads in the same process. */
1668 if (tsk->mm == current->mm)
1671 /* Simple filter to avoid false positives due to PID collisions */
1672 if (flags & TNF_SHARED)
1675 /* Update priv based on whether false sharing was detected */
1678 if (join && !get_numa_group(grp))
1686 double_lock(&my_grp->lock, &grp->lock);
1688 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1689 my_grp->faults[i] -= p->numa_faults_memory[i];
1690 grp->faults[i] += p->numa_faults_memory[i];
1692 my_grp->total_faults -= p->total_numa_faults;
1693 grp->total_faults += p->total_numa_faults;
1695 list_move(&p->numa_entry, &grp->task_list);
1699 spin_unlock(&my_grp->lock);
1700 spin_unlock(&grp->lock);
1702 rcu_assign_pointer(p->numa_group, grp);
1704 put_numa_group(my_grp);
1712 void task_numa_free(struct task_struct *p)
1714 struct numa_group *grp = p->numa_group;
1716 void *numa_faults = p->numa_faults_memory;
1719 spin_lock(&grp->lock);
1720 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1721 grp->faults[i] -= p->numa_faults_memory[i];
1722 grp->total_faults -= p->total_numa_faults;
1724 list_del(&p->numa_entry);
1726 spin_unlock(&grp->lock);
1727 rcu_assign_pointer(p->numa_group, NULL);
1728 put_numa_group(grp);
1731 p->numa_faults_memory = NULL;
1732 p->numa_faults_buffer_memory = NULL;
1733 p->numa_faults_cpu= NULL;
1734 p->numa_faults_buffer_cpu = NULL;
1739 * Got a PROT_NONE fault for a page on @node.
1741 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1743 struct task_struct *p = current;
1744 bool migrated = flags & TNF_MIGRATED;
1745 int cpu_node = task_node(current);
1748 if (!numabalancing_enabled)
1751 /* for example, ksmd faulting in a user's mm */
1755 /* Do not worry about placement if exiting */
1756 if (p->state == TASK_DEAD)
1759 /* Allocate buffer to track faults on a per-node basis */
1760 if (unlikely(!p->numa_faults_memory)) {
1761 int size = sizeof(*p->numa_faults_memory) *
1762 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1764 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1765 if (!p->numa_faults_memory)
1768 BUG_ON(p->numa_faults_buffer_memory);
1770 * The averaged statistics, shared & private, memory & cpu,
1771 * occupy the first half of the array. The second half of the
1772 * array is for current counters, which are averaged into the
1773 * first set by task_numa_placement.
1775 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1776 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1777 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1778 p->total_numa_faults = 0;
1779 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1783 * First accesses are treated as private, otherwise consider accesses
1784 * to be private if the accessing pid has not changed
1786 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1789 priv = cpupid_match_pid(p, last_cpupid);
1790 if (!priv && !(flags & TNF_NO_GROUP))
1791 task_numa_group(p, last_cpupid, flags, &priv);
1794 task_numa_placement(p);
1797 * Retry task to preferred node migration periodically, in case it
1798 * case it previously failed, or the scheduler moved us.
1800 if (time_after(jiffies, p->numa_migrate_retry))
1801 numa_migrate_preferred(p);
1804 p->numa_pages_migrated += pages;
1806 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1807 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1808 p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages;
1811 static void reset_ptenuma_scan(struct task_struct *p)
1813 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1814 p->mm->numa_scan_offset = 0;
1818 * The expensive part of numa migration is done from task_work context.
1819 * Triggered from task_tick_numa().
1821 void task_numa_work(struct callback_head *work)
1823 unsigned long migrate, next_scan, now = jiffies;
1824 struct task_struct *p = current;
1825 struct mm_struct *mm = p->mm;
1826 struct vm_area_struct *vma;
1827 unsigned long start, end;
1828 unsigned long nr_pte_updates = 0;
1831 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1833 work->next = work; /* protect against double add */
1835 * Who cares about NUMA placement when they're dying.
1837 * NOTE: make sure not to dereference p->mm before this check,
1838 * exit_task_work() happens _after_ exit_mm() so we could be called
1839 * without p->mm even though we still had it when we enqueued this
1842 if (p->flags & PF_EXITING)
1845 if (!mm->numa_next_scan) {
1846 mm->numa_next_scan = now +
1847 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1851 * Enforce maximal scan/migration frequency..
1853 migrate = mm->numa_next_scan;
1854 if (time_before(now, migrate))
1857 if (p->numa_scan_period == 0) {
1858 p->numa_scan_period_max = task_scan_max(p);
1859 p->numa_scan_period = task_scan_min(p);
1862 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1863 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1867 * Delay this task enough that another task of this mm will likely win
1868 * the next time around.
1870 p->node_stamp += 2 * TICK_NSEC;
1872 start = mm->numa_scan_offset;
1873 pages = sysctl_numa_balancing_scan_size;
1874 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1878 down_read(&mm->mmap_sem);
1879 vma = find_vma(mm, start);
1881 reset_ptenuma_scan(p);
1885 for (; vma; vma = vma->vm_next) {
1886 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1890 * Shared library pages mapped by multiple processes are not
1891 * migrated as it is expected they are cache replicated. Avoid
1892 * hinting faults in read-only file-backed mappings or the vdso
1893 * as migrating the pages will be of marginal benefit.
1896 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1900 * Skip inaccessible VMAs to avoid any confusion between
1901 * PROT_NONE and NUMA hinting ptes
1903 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1907 start = max(start, vma->vm_start);
1908 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1909 end = min(end, vma->vm_end);
1910 nr_pte_updates += change_prot_numa(vma, start, end);
1913 * Scan sysctl_numa_balancing_scan_size but ensure that
1914 * at least one PTE is updated so that unused virtual
1915 * address space is quickly skipped.
1918 pages -= (end - start) >> PAGE_SHIFT;
1923 } while (end != vma->vm_end);
1928 * It is possible to reach the end of the VMA list but the last few
1929 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1930 * would find the !migratable VMA on the next scan but not reset the
1931 * scanner to the start so check it now.
1934 mm->numa_scan_offset = start;
1936 reset_ptenuma_scan(p);
1937 up_read(&mm->mmap_sem);
1941 * Drive the periodic memory faults..
1943 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1945 struct callback_head *work = &curr->numa_work;
1949 * We don't care about NUMA placement if we don't have memory.
1951 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1955 * Using runtime rather than walltime has the dual advantage that
1956 * we (mostly) drive the selection from busy threads and that the
1957 * task needs to have done some actual work before we bother with
1960 now = curr->se.sum_exec_runtime;
1961 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1963 if (now - curr->node_stamp > period) {
1964 if (!curr->node_stamp)
1965 curr->numa_scan_period = task_scan_min(curr);
1966 curr->node_stamp += period;
1968 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1969 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1970 task_work_add(curr, work, true);
1975 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1979 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1983 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1986 #endif /* CONFIG_NUMA_BALANCING */
1989 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1991 update_load_add(&cfs_rq->load, se->load.weight);
1992 if (!parent_entity(se))
1993 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1995 if (entity_is_task(se)) {
1996 struct rq *rq = rq_of(cfs_rq);
1998 account_numa_enqueue(rq, task_of(se));
1999 list_add(&se->group_node, &rq->cfs_tasks);
2002 cfs_rq->nr_running++;
2006 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2008 update_load_sub(&cfs_rq->load, se->load.weight);
2009 if (!parent_entity(se))
2010 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2011 if (entity_is_task(se)) {
2012 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2013 list_del_init(&se->group_node);
2015 cfs_rq->nr_running--;
2018 #ifdef CONFIG_FAIR_GROUP_SCHED
2020 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2025 * Use this CPU's actual weight instead of the last load_contribution
2026 * to gain a more accurate current total weight. See
2027 * update_cfs_rq_load_contribution().
2029 tg_weight = atomic_long_read(&tg->load_avg);
2030 tg_weight -= cfs_rq->tg_load_contrib;
2031 tg_weight += cfs_rq->load.weight;
2036 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2038 long tg_weight, load, shares;
2040 tg_weight = calc_tg_weight(tg, cfs_rq);
2041 load = cfs_rq->load.weight;
2043 shares = (tg->shares * load);
2045 shares /= tg_weight;
2047 if (shares < MIN_SHARES)
2048 shares = MIN_SHARES;
2049 if (shares > tg->shares)
2050 shares = tg->shares;
2054 # else /* CONFIG_SMP */
2055 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2059 # endif /* CONFIG_SMP */
2060 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2061 unsigned long weight)
2064 /* commit outstanding execution time */
2065 if (cfs_rq->curr == se)
2066 update_curr(cfs_rq);
2067 account_entity_dequeue(cfs_rq, se);
2070 update_load_set(&se->load, weight);
2073 account_entity_enqueue(cfs_rq, se);
2076 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2078 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2080 struct task_group *tg;
2081 struct sched_entity *se;
2085 se = tg->se[cpu_of(rq_of(cfs_rq))];
2086 if (!se || throttled_hierarchy(cfs_rq))
2089 if (likely(se->load.weight == tg->shares))
2092 shares = calc_cfs_shares(cfs_rq, tg);
2094 reweight_entity(cfs_rq_of(se), se, shares);
2096 #else /* CONFIG_FAIR_GROUP_SCHED */
2097 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2100 #endif /* CONFIG_FAIR_GROUP_SCHED */
2104 * We choose a half-life close to 1 scheduling period.
2105 * Note: The tables below are dependent on this value.
2107 #define LOAD_AVG_PERIOD 32
2108 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2109 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2111 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2112 static const u32 runnable_avg_yN_inv[] = {
2113 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2114 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2115 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2116 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2117 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2118 0x85aac367, 0x82cd8698,
2122 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2123 * over-estimates when re-combining.
2125 static const u32 runnable_avg_yN_sum[] = {
2126 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2127 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2128 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2133 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2135 static __always_inline u64 decay_load(u64 val, u64 n)
2137 unsigned int local_n;
2141 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2144 /* after bounds checking we can collapse to 32-bit */
2148 * As y^PERIOD = 1/2, we can combine
2149 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2150 * With a look-up table which covers k^n (n<PERIOD)
2152 * To achieve constant time decay_load.
2154 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2155 val >>= local_n / LOAD_AVG_PERIOD;
2156 local_n %= LOAD_AVG_PERIOD;
2159 val *= runnable_avg_yN_inv[local_n];
2160 /* We don't use SRR here since we always want to round down. */
2165 * For updates fully spanning n periods, the contribution to runnable
2166 * average will be: \Sum 1024*y^n
2168 * We can compute this reasonably efficiently by combining:
2169 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2171 static u32 __compute_runnable_contrib(u64 n)
2175 if (likely(n <= LOAD_AVG_PERIOD))
2176 return runnable_avg_yN_sum[n];
2177 else if (unlikely(n >= LOAD_AVG_MAX_N))
2178 return LOAD_AVG_MAX;
2180 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2182 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2183 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2185 n -= LOAD_AVG_PERIOD;
2186 } while (n > LOAD_AVG_PERIOD);
2188 contrib = decay_load(contrib, n);
2189 return contrib + runnable_avg_yN_sum[n];
2193 * We can represent the historical contribution to runnable average as the
2194 * coefficients of a geometric series. To do this we sub-divide our runnable
2195 * history into segments of approximately 1ms (1024us); label the segment that
2196 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2198 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2200 * (now) (~1ms ago) (~2ms ago)
2202 * Let u_i denote the fraction of p_i that the entity was runnable.
2204 * We then designate the fractions u_i as our co-efficients, yielding the
2205 * following representation of historical load:
2206 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2208 * We choose y based on the with of a reasonably scheduling period, fixing:
2211 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2212 * approximately half as much as the contribution to load within the last ms
2215 * When a period "rolls over" and we have new u_0`, multiplying the previous
2216 * sum again by y is sufficient to update:
2217 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2218 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2220 static __always_inline int __update_entity_runnable_avg(u64 now,
2221 struct sched_avg *sa,
2225 u32 runnable_contrib;
2226 int delta_w, decayed = 0;
2228 delta = now - sa->last_runnable_update;
2230 * This should only happen when time goes backwards, which it
2231 * unfortunately does during sched clock init when we swap over to TSC.
2233 if ((s64)delta < 0) {
2234 sa->last_runnable_update = now;
2239 * Use 1024ns as the unit of measurement since it's a reasonable
2240 * approximation of 1us and fast to compute.
2245 sa->last_runnable_update = now;
2247 /* delta_w is the amount already accumulated against our next period */
2248 delta_w = sa->runnable_avg_period % 1024;
2249 if (delta + delta_w >= 1024) {
2250 /* period roll-over */
2254 * Now that we know we're crossing a period boundary, figure
2255 * out how much from delta we need to complete the current
2256 * period and accrue it.
2258 delta_w = 1024 - delta_w;
2260 sa->runnable_avg_sum += delta_w;
2261 sa->runnable_avg_period += delta_w;
2265 /* Figure out how many additional periods this update spans */
2266 periods = delta / 1024;
2269 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2271 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2274 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2275 runnable_contrib = __compute_runnable_contrib(periods);
2277 sa->runnable_avg_sum += runnable_contrib;
2278 sa->runnable_avg_period += runnable_contrib;
2281 /* Remainder of delta accrued against u_0` */
2283 sa->runnable_avg_sum += delta;
2284 sa->runnable_avg_period += delta;
2289 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2290 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2292 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2293 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2295 decays -= se->avg.decay_count;
2299 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2300 se->avg.decay_count = 0;
2305 #ifdef CONFIG_FAIR_GROUP_SCHED
2306 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2309 struct task_group *tg = cfs_rq->tg;
2312 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2313 tg_contrib -= cfs_rq->tg_load_contrib;
2315 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2316 atomic_long_add(tg_contrib, &tg->load_avg);
2317 cfs_rq->tg_load_contrib += tg_contrib;
2322 * Aggregate cfs_rq runnable averages into an equivalent task_group
2323 * representation for computing load contributions.
2325 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2326 struct cfs_rq *cfs_rq)
2328 struct task_group *tg = cfs_rq->tg;
2331 /* The fraction of a cpu used by this cfs_rq */
2332 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2333 sa->runnable_avg_period + 1);
2334 contrib -= cfs_rq->tg_runnable_contrib;
2336 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2337 atomic_add(contrib, &tg->runnable_avg);
2338 cfs_rq->tg_runnable_contrib += contrib;
2342 static inline void __update_group_entity_contrib(struct sched_entity *se)
2344 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2345 struct task_group *tg = cfs_rq->tg;
2350 contrib = cfs_rq->tg_load_contrib * tg->shares;
2351 se->avg.load_avg_contrib = div_u64(contrib,
2352 atomic_long_read(&tg->load_avg) + 1);
2355 * For group entities we need to compute a correction term in the case
2356 * that they are consuming <1 cpu so that we would contribute the same
2357 * load as a task of equal weight.
2359 * Explicitly co-ordinating this measurement would be expensive, but
2360 * fortunately the sum of each cpus contribution forms a usable
2361 * lower-bound on the true value.
2363 * Consider the aggregate of 2 contributions. Either they are disjoint
2364 * (and the sum represents true value) or they are disjoint and we are
2365 * understating by the aggregate of their overlap.
2367 * Extending this to N cpus, for a given overlap, the maximum amount we
2368 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2369 * cpus that overlap for this interval and w_i is the interval width.
2371 * On a small machine; the first term is well-bounded which bounds the
2372 * total error since w_i is a subset of the period. Whereas on a
2373 * larger machine, while this first term can be larger, if w_i is the
2374 * of consequential size guaranteed to see n_i*w_i quickly converge to
2375 * our upper bound of 1-cpu.
2377 runnable_avg = atomic_read(&tg->runnable_avg);
2378 if (runnable_avg < NICE_0_LOAD) {
2379 se->avg.load_avg_contrib *= runnable_avg;
2380 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2384 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2385 int force_update) {}
2386 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2387 struct cfs_rq *cfs_rq) {}
2388 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2391 static inline void __update_task_entity_contrib(struct sched_entity *se)
2395 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2396 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2397 contrib /= (se->avg.runnable_avg_period + 1);
2398 se->avg.load_avg_contrib = scale_load(contrib);
2401 /* Compute the current contribution to load_avg by se, return any delta */
2402 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2404 long old_contrib = se->avg.load_avg_contrib;
2406 if (entity_is_task(se)) {
2407 __update_task_entity_contrib(se);
2409 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2410 __update_group_entity_contrib(se);
2413 return se->avg.load_avg_contrib - old_contrib;
2416 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2419 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2420 cfs_rq->blocked_load_avg -= load_contrib;
2422 cfs_rq->blocked_load_avg = 0;
2425 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2427 /* Update a sched_entity's runnable average */
2428 static inline void update_entity_load_avg(struct sched_entity *se,
2431 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2436 * For a group entity we need to use their owned cfs_rq_clock_task() in
2437 * case they are the parent of a throttled hierarchy.
2439 if (entity_is_task(se))
2440 now = cfs_rq_clock_task(cfs_rq);
2442 now = cfs_rq_clock_task(group_cfs_rq(se));
2444 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2447 contrib_delta = __update_entity_load_avg_contrib(se);
2453 cfs_rq->runnable_load_avg += contrib_delta;
2455 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2459 * Decay the load contributed by all blocked children and account this so that
2460 * their contribution may appropriately discounted when they wake up.
2462 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2464 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2467 decays = now - cfs_rq->last_decay;
2468 if (!decays && !force_update)
2471 if (atomic_long_read(&cfs_rq->removed_load)) {
2472 unsigned long removed_load;
2473 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2474 subtract_blocked_load_contrib(cfs_rq, removed_load);
2478 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2480 atomic64_add(decays, &cfs_rq->decay_counter);
2481 cfs_rq->last_decay = now;
2484 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2487 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2489 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2490 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2493 /* Add the load generated by se into cfs_rq's child load-average */
2494 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2495 struct sched_entity *se,
2499 * We track migrations using entity decay_count <= 0, on a wake-up
2500 * migration we use a negative decay count to track the remote decays
2501 * accumulated while sleeping.
2503 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2504 * are seen by enqueue_entity_load_avg() as a migration with an already
2505 * constructed load_avg_contrib.
2507 if (unlikely(se->avg.decay_count <= 0)) {
2508 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2509 if (se->avg.decay_count) {
2511 * In a wake-up migration we have to approximate the
2512 * time sleeping. This is because we can't synchronize
2513 * clock_task between the two cpus, and it is not
2514 * guaranteed to be read-safe. Instead, we can
2515 * approximate this using our carried decays, which are
2516 * explicitly atomically readable.
2518 se->avg.last_runnable_update -= (-se->avg.decay_count)
2520 update_entity_load_avg(se, 0);
2521 /* Indicate that we're now synchronized and on-rq */
2522 se->avg.decay_count = 0;
2526 __synchronize_entity_decay(se);
2529 /* migrated tasks did not contribute to our blocked load */
2531 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2532 update_entity_load_avg(se, 0);
2535 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2536 /* we force update consideration on load-balancer moves */
2537 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2541 * Remove se's load from this cfs_rq child load-average, if the entity is
2542 * transitioning to a blocked state we track its projected decay using
2545 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2546 struct sched_entity *se,
2549 update_entity_load_avg(se, 1);
2550 /* we force update consideration on load-balancer moves */
2551 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2553 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2555 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2556 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2557 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2561 * Update the rq's load with the elapsed running time before entering
2562 * idle. if the last scheduled task is not a CFS task, idle_enter will
2563 * be the only way to update the runnable statistic.
2565 void idle_enter_fair(struct rq *this_rq)
2567 update_rq_runnable_avg(this_rq, 1);
2571 * Update the rq's load with the elapsed idle time before a task is
2572 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2573 * be the only way to update the runnable statistic.
2575 void idle_exit_fair(struct rq *this_rq)
2577 update_rq_runnable_avg(this_rq, 0);
2581 static inline void update_entity_load_avg(struct sched_entity *se,
2582 int update_cfs_rq) {}
2583 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2584 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2585 struct sched_entity *se,
2587 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2588 struct sched_entity *se,
2590 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2591 int force_update) {}
2594 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2596 #ifdef CONFIG_SCHEDSTATS
2597 struct task_struct *tsk = NULL;
2599 if (entity_is_task(se))
2602 if (se->statistics.sleep_start) {
2603 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2608 if (unlikely(delta > se->statistics.sleep_max))
2609 se->statistics.sleep_max = delta;
2611 se->statistics.sleep_start = 0;
2612 se->statistics.sum_sleep_runtime += delta;
2615 account_scheduler_latency(tsk, delta >> 10, 1);
2616 trace_sched_stat_sleep(tsk, delta);
2619 if (se->statistics.block_start) {
2620 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2625 if (unlikely(delta > se->statistics.block_max))
2626 se->statistics.block_max = delta;
2628 se->statistics.block_start = 0;
2629 se->statistics.sum_sleep_runtime += delta;
2632 if (tsk->in_iowait) {
2633 se->statistics.iowait_sum += delta;
2634 se->statistics.iowait_count++;
2635 trace_sched_stat_iowait(tsk, delta);
2638 trace_sched_stat_blocked(tsk, delta);
2641 * Blocking time is in units of nanosecs, so shift by
2642 * 20 to get a milliseconds-range estimation of the
2643 * amount of time that the task spent sleeping:
2645 if (unlikely(prof_on == SLEEP_PROFILING)) {
2646 profile_hits(SLEEP_PROFILING,
2647 (void *)get_wchan(tsk),
2650 account_scheduler_latency(tsk, delta >> 10, 0);
2656 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2658 #ifdef CONFIG_SCHED_DEBUG
2659 s64 d = se->vruntime - cfs_rq->min_vruntime;
2664 if (d > 3*sysctl_sched_latency)
2665 schedstat_inc(cfs_rq, nr_spread_over);
2670 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2672 u64 vruntime = cfs_rq->min_vruntime;
2675 * The 'current' period is already promised to the current tasks,
2676 * however the extra weight of the new task will slow them down a
2677 * little, place the new task so that it fits in the slot that
2678 * stays open at the end.
2680 if (initial && sched_feat(START_DEBIT))
2681 vruntime += sched_vslice(cfs_rq, se);
2683 /* sleeps up to a single latency don't count. */
2685 unsigned long thresh = sysctl_sched_latency;
2688 * Halve their sleep time's effect, to allow
2689 * for a gentler effect of sleepers:
2691 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2697 /* ensure we never gain time by being placed backwards. */
2698 se->vruntime = max_vruntime(se->vruntime, vruntime);
2701 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2704 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2707 * Update the normalized vruntime before updating min_vruntime
2708 * through calling update_curr().
2710 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2711 se->vruntime += cfs_rq->min_vruntime;
2714 * Update run-time statistics of the 'current'.
2716 update_curr(cfs_rq);
2717 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2718 account_entity_enqueue(cfs_rq, se);
2719 update_cfs_shares(cfs_rq);
2721 if (flags & ENQUEUE_WAKEUP) {
2722 place_entity(cfs_rq, se, 0);
2723 enqueue_sleeper(cfs_rq, se);
2726 update_stats_enqueue(cfs_rq, se);
2727 check_spread(cfs_rq, se);
2728 if (se != cfs_rq->curr)
2729 __enqueue_entity(cfs_rq, se);
2732 if (cfs_rq->nr_running == 1) {
2733 list_add_leaf_cfs_rq(cfs_rq);
2734 check_enqueue_throttle(cfs_rq);
2738 static void __clear_buddies_last(struct sched_entity *se)
2740 for_each_sched_entity(se) {
2741 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2742 if (cfs_rq->last != se)
2745 cfs_rq->last = NULL;
2749 static void __clear_buddies_next(struct sched_entity *se)
2751 for_each_sched_entity(se) {
2752 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2753 if (cfs_rq->next != se)
2756 cfs_rq->next = NULL;
2760 static void __clear_buddies_skip(struct sched_entity *se)
2762 for_each_sched_entity(se) {
2763 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2764 if (cfs_rq->skip != se)
2767 cfs_rq->skip = NULL;
2771 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2773 if (cfs_rq->last == se)
2774 __clear_buddies_last(se);
2776 if (cfs_rq->next == se)
2777 __clear_buddies_next(se);
2779 if (cfs_rq->skip == se)
2780 __clear_buddies_skip(se);
2783 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2786 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2789 * Update run-time statistics of the 'current'.
2791 update_curr(cfs_rq);
2792 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2794 update_stats_dequeue(cfs_rq, se);
2795 if (flags & DEQUEUE_SLEEP) {
2796 #ifdef CONFIG_SCHEDSTATS
2797 if (entity_is_task(se)) {
2798 struct task_struct *tsk = task_of(se);
2800 if (tsk->state & TASK_INTERRUPTIBLE)
2801 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2802 if (tsk->state & TASK_UNINTERRUPTIBLE)
2803 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2808 clear_buddies(cfs_rq, se);
2810 if (se != cfs_rq->curr)
2811 __dequeue_entity(cfs_rq, se);
2813 account_entity_dequeue(cfs_rq, se);
2816 * Normalize the entity after updating the min_vruntime because the
2817 * update can refer to the ->curr item and we need to reflect this
2818 * movement in our normalized position.
2820 if (!(flags & DEQUEUE_SLEEP))
2821 se->vruntime -= cfs_rq->min_vruntime;
2823 /* return excess runtime on last dequeue */
2824 return_cfs_rq_runtime(cfs_rq);
2826 update_min_vruntime(cfs_rq);
2827 update_cfs_shares(cfs_rq);
2831 * Preempt the current task with a newly woken task if needed:
2834 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2836 unsigned long ideal_runtime, delta_exec;
2837 struct sched_entity *se;
2840 ideal_runtime = sched_slice(cfs_rq, curr);
2841 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2842 if (delta_exec > ideal_runtime) {
2843 resched_task(rq_of(cfs_rq)->curr);
2845 * The current task ran long enough, ensure it doesn't get
2846 * re-elected due to buddy favours.
2848 clear_buddies(cfs_rq, curr);
2853 * Ensure that a task that missed wakeup preemption by a
2854 * narrow margin doesn't have to wait for a full slice.
2855 * This also mitigates buddy induced latencies under load.
2857 if (delta_exec < sysctl_sched_min_granularity)
2860 se = __pick_first_entity(cfs_rq);
2861 delta = curr->vruntime - se->vruntime;
2866 if (delta > ideal_runtime)
2867 resched_task(rq_of(cfs_rq)->curr);
2871 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2873 /* 'current' is not kept within the tree. */
2876 * Any task has to be enqueued before it get to execute on
2877 * a CPU. So account for the time it spent waiting on the
2880 update_stats_wait_end(cfs_rq, se);
2881 __dequeue_entity(cfs_rq, se);
2884 update_stats_curr_start(cfs_rq, se);
2886 #ifdef CONFIG_SCHEDSTATS
2888 * Track our maximum slice length, if the CPU's load is at
2889 * least twice that of our own weight (i.e. dont track it
2890 * when there are only lesser-weight tasks around):
2892 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2893 se->statistics.slice_max = max(se->statistics.slice_max,
2894 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2897 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2901 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2904 * Pick the next process, keeping these things in mind, in this order:
2905 * 1) keep things fair between processes/task groups
2906 * 2) pick the "next" process, since someone really wants that to run
2907 * 3) pick the "last" process, for cache locality
2908 * 4) do not run the "skip" process, if something else is available
2910 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2912 struct sched_entity *se = __pick_first_entity(cfs_rq);
2913 struct sched_entity *left = se;
2916 * Avoid running the skip buddy, if running something else can
2917 * be done without getting too unfair.
2919 if (cfs_rq->skip == se) {
2920 struct sched_entity *second = __pick_next_entity(se);
2921 if (second && wakeup_preempt_entity(second, left) < 1)
2926 * Prefer last buddy, try to return the CPU to a preempted task.
2928 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2932 * Someone really wants this to run. If it's not unfair, run it.
2934 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2937 clear_buddies(cfs_rq, se);
2942 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2944 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2947 * If still on the runqueue then deactivate_task()
2948 * was not called and update_curr() has to be done:
2951 update_curr(cfs_rq);
2953 /* throttle cfs_rqs exceeding runtime */
2954 check_cfs_rq_runtime(cfs_rq);
2956 check_spread(cfs_rq, prev);
2958 update_stats_wait_start(cfs_rq, prev);
2959 /* Put 'current' back into the tree. */
2960 __enqueue_entity(cfs_rq, prev);
2961 /* in !on_rq case, update occurred at dequeue */
2962 update_entity_load_avg(prev, 1);
2964 cfs_rq->curr = NULL;
2968 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2971 * Update run-time statistics of the 'current'.
2973 update_curr(cfs_rq);
2976 * Ensure that runnable average is periodically updated.
2978 update_entity_load_avg(curr, 1);
2979 update_cfs_rq_blocked_load(cfs_rq, 1);
2980 update_cfs_shares(cfs_rq);
2982 #ifdef CONFIG_SCHED_HRTICK
2984 * queued ticks are scheduled to match the slice, so don't bother
2985 * validating it and just reschedule.
2988 resched_task(rq_of(cfs_rq)->curr);
2992 * don't let the period tick interfere with the hrtick preemption
2994 if (!sched_feat(DOUBLE_TICK) &&
2995 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2999 if (cfs_rq->nr_running > 1)
3000 check_preempt_tick(cfs_rq, curr);
3004 /**************************************************
3005 * CFS bandwidth control machinery
3008 #ifdef CONFIG_CFS_BANDWIDTH
3010 #ifdef HAVE_JUMP_LABEL
3011 static struct static_key __cfs_bandwidth_used;
3013 static inline bool cfs_bandwidth_used(void)
3015 return static_key_false(&__cfs_bandwidth_used);
3018 void cfs_bandwidth_usage_inc(void)
3020 static_key_slow_inc(&__cfs_bandwidth_used);
3023 void cfs_bandwidth_usage_dec(void)
3025 static_key_slow_dec(&__cfs_bandwidth_used);
3027 #else /* HAVE_JUMP_LABEL */
3028 static bool cfs_bandwidth_used(void)
3033 void cfs_bandwidth_usage_inc(void) {}
3034 void cfs_bandwidth_usage_dec(void) {}
3035 #endif /* HAVE_JUMP_LABEL */
3038 * default period for cfs group bandwidth.
3039 * default: 0.1s, units: nanoseconds
3041 static inline u64 default_cfs_period(void)
3043 return 100000000ULL;
3046 static inline u64 sched_cfs_bandwidth_slice(void)
3048 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3052 * Replenish runtime according to assigned quota and update expiration time.
3053 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3054 * additional synchronization around rq->lock.
3056 * requires cfs_b->lock
3058 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3062 if (cfs_b->quota == RUNTIME_INF)
3065 now = sched_clock_cpu(smp_processor_id());
3066 cfs_b->runtime = cfs_b->quota;
3067 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3070 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3072 return &tg->cfs_bandwidth;
3075 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3076 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3078 if (unlikely(cfs_rq->throttle_count))
3079 return cfs_rq->throttled_clock_task;
3081 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3084 /* returns 0 on failure to allocate runtime */
3085 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3087 struct task_group *tg = cfs_rq->tg;
3088 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3089 u64 amount = 0, min_amount, expires;
3091 /* note: this is a positive sum as runtime_remaining <= 0 */
3092 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3094 raw_spin_lock(&cfs_b->lock);
3095 if (cfs_b->quota == RUNTIME_INF)
3096 amount = min_amount;
3099 * If the bandwidth pool has become inactive, then at least one
3100 * period must have elapsed since the last consumption.
3101 * Refresh the global state and ensure bandwidth timer becomes
3104 if (!cfs_b->timer_active) {
3105 __refill_cfs_bandwidth_runtime(cfs_b);
3106 __start_cfs_bandwidth(cfs_b);
3109 if (cfs_b->runtime > 0) {
3110 amount = min(cfs_b->runtime, min_amount);
3111 cfs_b->runtime -= amount;
3115 expires = cfs_b->runtime_expires;
3116 raw_spin_unlock(&cfs_b->lock);
3118 cfs_rq->runtime_remaining += amount;
3120 * we may have advanced our local expiration to account for allowed
3121 * spread between our sched_clock and the one on which runtime was
3124 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3125 cfs_rq->runtime_expires = expires;
3127 return cfs_rq->runtime_remaining > 0;
3131 * Note: This depends on the synchronization provided by sched_clock and the
3132 * fact that rq->clock snapshots this value.
3134 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3136 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3138 /* if the deadline is ahead of our clock, nothing to do */
3139 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3142 if (cfs_rq->runtime_remaining < 0)
3146 * If the local deadline has passed we have to consider the
3147 * possibility that our sched_clock is 'fast' and the global deadline
3148 * has not truly expired.
3150 * Fortunately we can check determine whether this the case by checking
3151 * whether the global deadline has advanced.
3154 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
3155 /* extend local deadline, drift is bounded above by 2 ticks */
3156 cfs_rq->runtime_expires += TICK_NSEC;
3158 /* global deadline is ahead, expiration has passed */
3159 cfs_rq->runtime_remaining = 0;
3163 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3165 /* dock delta_exec before expiring quota (as it could span periods) */
3166 cfs_rq->runtime_remaining -= delta_exec;
3167 expire_cfs_rq_runtime(cfs_rq);
3169 if (likely(cfs_rq->runtime_remaining > 0))
3173 * if we're unable to extend our runtime we resched so that the active
3174 * hierarchy can be throttled
3176 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3177 resched_task(rq_of(cfs_rq)->curr);
3180 static __always_inline
3181 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3183 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3186 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3189 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3191 return cfs_bandwidth_used() && cfs_rq->throttled;
3194 /* check whether cfs_rq, or any parent, is throttled */
3195 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3197 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3201 * Ensure that neither of the group entities corresponding to src_cpu or
3202 * dest_cpu are members of a throttled hierarchy when performing group
3203 * load-balance operations.
3205 static inline int throttled_lb_pair(struct task_group *tg,
3206 int src_cpu, int dest_cpu)
3208 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3210 src_cfs_rq = tg->cfs_rq[src_cpu];
3211 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3213 return throttled_hierarchy(src_cfs_rq) ||
3214 throttled_hierarchy(dest_cfs_rq);
3217 /* updated child weight may affect parent so we have to do this bottom up */
3218 static int tg_unthrottle_up(struct task_group *tg, void *data)
3220 struct rq *rq = data;
3221 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3223 cfs_rq->throttle_count--;
3225 if (!cfs_rq->throttle_count) {
3226 /* adjust cfs_rq_clock_task() */
3227 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3228 cfs_rq->throttled_clock_task;
3235 static int tg_throttle_down(struct task_group *tg, void *data)
3237 struct rq *rq = data;
3238 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3240 /* group is entering throttled state, stop time */
3241 if (!cfs_rq->throttle_count)
3242 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3243 cfs_rq->throttle_count++;
3248 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3250 struct rq *rq = rq_of(cfs_rq);
3251 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3252 struct sched_entity *se;
3253 long task_delta, dequeue = 1;
3255 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3257 /* freeze hierarchy runnable averages while throttled */
3259 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3262 task_delta = cfs_rq->h_nr_running;
3263 for_each_sched_entity(se) {
3264 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3265 /* throttled entity or throttle-on-deactivate */
3270 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3271 qcfs_rq->h_nr_running -= task_delta;
3273 if (qcfs_rq->load.weight)
3278 rq->nr_running -= task_delta;
3280 cfs_rq->throttled = 1;
3281 cfs_rq->throttled_clock = rq_clock(rq);
3282 raw_spin_lock(&cfs_b->lock);
3283 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3284 if (!cfs_b->timer_active)
3285 __start_cfs_bandwidth(cfs_b);
3286 raw_spin_unlock(&cfs_b->lock);
3289 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3291 struct rq *rq = rq_of(cfs_rq);
3292 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3293 struct sched_entity *se;
3297 se = cfs_rq->tg->se[cpu_of(rq)];
3299 cfs_rq->throttled = 0;
3301 update_rq_clock(rq);
3303 raw_spin_lock(&cfs_b->lock);
3304 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3305 list_del_rcu(&cfs_rq->throttled_list);
3306 raw_spin_unlock(&cfs_b->lock);
3308 /* update hierarchical throttle state */
3309 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3311 if (!cfs_rq->load.weight)
3314 task_delta = cfs_rq->h_nr_running;
3315 for_each_sched_entity(se) {
3319 cfs_rq = cfs_rq_of(se);
3321 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3322 cfs_rq->h_nr_running += task_delta;
3324 if (cfs_rq_throttled(cfs_rq))
3329 rq->nr_running += task_delta;
3331 /* determine whether we need to wake up potentially idle cpu */
3332 if (rq->curr == rq->idle && rq->cfs.nr_running)
3333 resched_task(rq->curr);
3336 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3337 u64 remaining, u64 expires)
3339 struct cfs_rq *cfs_rq;
3340 u64 runtime = remaining;
3343 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3345 struct rq *rq = rq_of(cfs_rq);
3347 raw_spin_lock(&rq->lock);
3348 if (!cfs_rq_throttled(cfs_rq))
3351 runtime = -cfs_rq->runtime_remaining + 1;
3352 if (runtime > remaining)
3353 runtime = remaining;
3354 remaining -= runtime;
3356 cfs_rq->runtime_remaining += runtime;
3357 cfs_rq->runtime_expires = expires;
3359 /* we check whether we're throttled above */
3360 if (cfs_rq->runtime_remaining > 0)
3361 unthrottle_cfs_rq(cfs_rq);
3364 raw_spin_unlock(&rq->lock);
3375 * Responsible for refilling a task_group's bandwidth and unthrottling its
3376 * cfs_rqs as appropriate. If there has been no activity within the last
3377 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3378 * used to track this state.
3380 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3382 u64 runtime, runtime_expires;
3383 int idle = 1, throttled;
3385 raw_spin_lock(&cfs_b->lock);
3386 /* no need to continue the timer with no bandwidth constraint */
3387 if (cfs_b->quota == RUNTIME_INF)
3390 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3391 /* idle depends on !throttled (for the case of a large deficit) */
3392 idle = cfs_b->idle && !throttled;
3393 cfs_b->nr_periods += overrun;
3395 /* if we're going inactive then everything else can be deferred */
3400 * if we have relooped after returning idle once, we need to update our
3401 * status as actually running, so that other cpus doing
3402 * __start_cfs_bandwidth will stop trying to cancel us.
3404 cfs_b->timer_active = 1;
3406 __refill_cfs_bandwidth_runtime(cfs_b);
3409 /* mark as potentially idle for the upcoming period */
3414 /* account preceding periods in which throttling occurred */
3415 cfs_b->nr_throttled += overrun;
3418 * There are throttled entities so we must first use the new bandwidth
3419 * to unthrottle them before making it generally available. This
3420 * ensures that all existing debts will be paid before a new cfs_rq is
3423 runtime = cfs_b->runtime;
3424 runtime_expires = cfs_b->runtime_expires;
3428 * This check is repeated as we are holding onto the new bandwidth
3429 * while we unthrottle. This can potentially race with an unthrottled
3430 * group trying to acquire new bandwidth from the global pool.
3432 while (throttled && runtime > 0) {
3433 raw_spin_unlock(&cfs_b->lock);
3434 /* we can't nest cfs_b->lock while distributing bandwidth */
3435 runtime = distribute_cfs_runtime(cfs_b, runtime,
3437 raw_spin_lock(&cfs_b->lock);
3439 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3442 /* return (any) remaining runtime */
3443 cfs_b->runtime = runtime;
3445 * While we are ensured activity in the period following an
3446 * unthrottle, this also covers the case in which the new bandwidth is
3447 * insufficient to cover the existing bandwidth deficit. (Forcing the
3448 * timer to remain active while there are any throttled entities.)
3453 cfs_b->timer_active = 0;
3454 raw_spin_unlock(&cfs_b->lock);
3459 /* a cfs_rq won't donate quota below this amount */
3460 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3461 /* minimum remaining period time to redistribute slack quota */
3462 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3463 /* how long we wait to gather additional slack before distributing */
3464 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3467 * Are we near the end of the current quota period?
3469 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3470 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3471 * migrate_hrtimers, base is never cleared, so we are fine.
3473 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3475 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3478 /* if the call-back is running a quota refresh is already occurring */
3479 if (hrtimer_callback_running(refresh_timer))
3482 /* is a quota refresh about to occur? */
3483 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3484 if (remaining < min_expire)
3490 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3492 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3494 /* if there's a quota refresh soon don't bother with slack */
3495 if (runtime_refresh_within(cfs_b, min_left))
3498 start_bandwidth_timer(&cfs_b->slack_timer,
3499 ns_to_ktime(cfs_bandwidth_slack_period));
3502 /* we know any runtime found here is valid as update_curr() precedes return */
3503 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3505 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3506 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3508 if (slack_runtime <= 0)
3511 raw_spin_lock(&cfs_b->lock);
3512 if (cfs_b->quota != RUNTIME_INF &&
3513 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3514 cfs_b->runtime += slack_runtime;
3516 /* we are under rq->lock, defer unthrottling using a timer */
3517 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3518 !list_empty(&cfs_b->throttled_cfs_rq))
3519 start_cfs_slack_bandwidth(cfs_b);
3521 raw_spin_unlock(&cfs_b->lock);
3523 /* even if it's not valid for return we don't want to try again */
3524 cfs_rq->runtime_remaining -= slack_runtime;
3527 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3529 if (!cfs_bandwidth_used())
3532 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3535 __return_cfs_rq_runtime(cfs_rq);
3539 * This is done with a timer (instead of inline with bandwidth return) since
3540 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3542 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3544 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3547 /* confirm we're still not at a refresh boundary */
3548 raw_spin_lock(&cfs_b->lock);
3549 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3550 raw_spin_unlock(&cfs_b->lock);
3554 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3555 runtime = cfs_b->runtime;
3558 expires = cfs_b->runtime_expires;
3559 raw_spin_unlock(&cfs_b->lock);
3564 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3566 raw_spin_lock(&cfs_b->lock);
3567 if (expires == cfs_b->runtime_expires)
3568 cfs_b->runtime = runtime;
3569 raw_spin_unlock(&cfs_b->lock);
3573 * When a group wakes up we want to make sure that its quota is not already
3574 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3575 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3577 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3579 if (!cfs_bandwidth_used())
3582 /* an active group must be handled by the update_curr()->put() path */
3583 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3586 /* ensure the group is not already throttled */
3587 if (cfs_rq_throttled(cfs_rq))
3590 /* update runtime allocation */
3591 account_cfs_rq_runtime(cfs_rq, 0);
3592 if (cfs_rq->runtime_remaining <= 0)
3593 throttle_cfs_rq(cfs_rq);
3596 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3597 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3599 if (!cfs_bandwidth_used())
3602 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3606 * it's possible for a throttled entity to be forced into a running
3607 * state (e.g. set_curr_task), in this case we're finished.
3609 if (cfs_rq_throttled(cfs_rq))
3612 throttle_cfs_rq(cfs_rq);
3615 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3617 struct cfs_bandwidth *cfs_b =
3618 container_of(timer, struct cfs_bandwidth, slack_timer);
3619 do_sched_cfs_slack_timer(cfs_b);
3621 return HRTIMER_NORESTART;
3624 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3626 struct cfs_bandwidth *cfs_b =
3627 container_of(timer, struct cfs_bandwidth, period_timer);
3633 now = hrtimer_cb_get_time(timer);
3634 overrun = hrtimer_forward(timer, now, cfs_b->period);
3639 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3642 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3645 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3647 raw_spin_lock_init(&cfs_b->lock);
3649 cfs_b->quota = RUNTIME_INF;
3650 cfs_b->period = ns_to_ktime(default_cfs_period());
3652 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3653 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3654 cfs_b->period_timer.function = sched_cfs_period_timer;
3655 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3656 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3659 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3661 cfs_rq->runtime_enabled = 0;
3662 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3665 /* requires cfs_b->lock, may release to reprogram timer */
3666 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3669 * The timer may be active because we're trying to set a new bandwidth
3670 * period or because we're racing with the tear-down path
3671 * (timer_active==0 becomes visible before the hrtimer call-back
3672 * terminates). In either case we ensure that it's re-programmed
3674 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3675 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3676 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3677 raw_spin_unlock(&cfs_b->lock);
3679 raw_spin_lock(&cfs_b->lock);
3680 /* if someone else restarted the timer then we're done */
3681 if (cfs_b->timer_active)
3685 cfs_b->timer_active = 1;
3686 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3689 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3691 hrtimer_cancel(&cfs_b->period_timer);
3692 hrtimer_cancel(&cfs_b->slack_timer);
3695 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3697 struct cfs_rq *cfs_rq;
3699 for_each_leaf_cfs_rq(rq, cfs_rq) {
3700 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3702 if (!cfs_rq->runtime_enabled)
3706 * clock_task is not advancing so we just need to make sure
3707 * there's some valid quota amount
3709 cfs_rq->runtime_remaining = cfs_b->quota;
3710 if (cfs_rq_throttled(cfs_rq))
3711 unthrottle_cfs_rq(cfs_rq);
3715 #else /* CONFIG_CFS_BANDWIDTH */
3716 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3718 return rq_clock_task(rq_of(cfs_rq));
3721 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3722 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3723 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3724 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3726 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3731 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3736 static inline int throttled_lb_pair(struct task_group *tg,
3737 int src_cpu, int dest_cpu)
3742 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3744 #ifdef CONFIG_FAIR_GROUP_SCHED
3745 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3748 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3752 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3753 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3755 #endif /* CONFIG_CFS_BANDWIDTH */
3757 /**************************************************
3758 * CFS operations on tasks:
3761 #ifdef CONFIG_SCHED_HRTICK
3762 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3764 struct sched_entity *se = &p->se;
3765 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3767 WARN_ON(task_rq(p) != rq);
3769 if (cfs_rq->nr_running > 1) {
3770 u64 slice = sched_slice(cfs_rq, se);
3771 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3772 s64 delta = slice - ran;
3781 * Don't schedule slices shorter than 10000ns, that just
3782 * doesn't make sense. Rely on vruntime for fairness.
3785 delta = max_t(s64, 10000LL, delta);
3787 hrtick_start(rq, delta);
3792 * called from enqueue/dequeue and updates the hrtick when the
3793 * current task is from our class and nr_running is low enough
3796 static void hrtick_update(struct rq *rq)
3798 struct task_struct *curr = rq->curr;
3800 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3803 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3804 hrtick_start_fair(rq, curr);
3806 #else /* !CONFIG_SCHED_HRTICK */
3808 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3812 static inline void hrtick_update(struct rq *rq)
3818 * The enqueue_task method is called before nr_running is
3819 * increased. Here we update the fair scheduling stats and
3820 * then put the task into the rbtree:
3823 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3825 struct cfs_rq *cfs_rq;
3826 struct sched_entity *se = &p->se;
3828 for_each_sched_entity(se) {
3831 cfs_rq = cfs_rq_of(se);
3832 enqueue_entity(cfs_rq, se, flags);
3835 * end evaluation on encountering a throttled cfs_rq
3837 * note: in the case of encountering a throttled cfs_rq we will
3838 * post the final h_nr_running increment below.
3840 if (cfs_rq_throttled(cfs_rq))
3842 cfs_rq->h_nr_running++;
3844 flags = ENQUEUE_WAKEUP;
3847 for_each_sched_entity(se) {
3848 cfs_rq = cfs_rq_of(se);
3849 cfs_rq->h_nr_running++;
3851 if (cfs_rq_throttled(cfs_rq))
3854 update_cfs_shares(cfs_rq);
3855 update_entity_load_avg(se, 1);
3859 update_rq_runnable_avg(rq, rq->nr_running);
3865 static void set_next_buddy(struct sched_entity *se);
3868 * The dequeue_task method is called before nr_running is
3869 * decreased. We remove the task from the rbtree and
3870 * update the fair scheduling stats:
3872 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3874 struct cfs_rq *cfs_rq;
3875 struct sched_entity *se = &p->se;
3876 int task_sleep = flags & DEQUEUE_SLEEP;
3878 for_each_sched_entity(se) {
3879 cfs_rq = cfs_rq_of(se);
3880 dequeue_entity(cfs_rq, se, flags);
3883 * end evaluation on encountering a throttled cfs_rq
3885 * note: in the case of encountering a throttled cfs_rq we will
3886 * post the final h_nr_running decrement below.
3888 if (cfs_rq_throttled(cfs_rq))
3890 cfs_rq->h_nr_running--;
3892 /* Don't dequeue parent if it has other entities besides us */
3893 if (cfs_rq->load.weight) {
3895 * Bias pick_next to pick a task from this cfs_rq, as
3896 * p is sleeping when it is within its sched_slice.
3898 if (task_sleep && parent_entity(se))
3899 set_next_buddy(parent_entity(se));
3901 /* avoid re-evaluating load for this entity */
3902 se = parent_entity(se);
3905 flags |= DEQUEUE_SLEEP;
3908 for_each_sched_entity(se) {
3909 cfs_rq = cfs_rq_of(se);
3910 cfs_rq->h_nr_running--;
3912 if (cfs_rq_throttled(cfs_rq))
3915 update_cfs_shares(cfs_rq);
3916 update_entity_load_avg(se, 1);
3921 update_rq_runnable_avg(rq, 1);
3927 /* Used instead of source_load when we know the type == 0 */
3928 static unsigned long weighted_cpuload(const int cpu)
3930 return cpu_rq(cpu)->cfs.runnable_load_avg;
3934 * Return a low guess at the load of a migration-source cpu weighted
3935 * according to the scheduling class and "nice" value.
3937 * We want to under-estimate the load of migration sources, to
3938 * balance conservatively.
3940 static unsigned long source_load(int cpu, int type)
3942 struct rq *rq = cpu_rq(cpu);
3943 unsigned long total = weighted_cpuload(cpu);
3945 if (type == 0 || !sched_feat(LB_BIAS))
3948 return min(rq->cpu_load[type-1], total);
3952 * Return a high guess at the load of a migration-target cpu weighted
3953 * according to the scheduling class and "nice" value.
3955 static unsigned long target_load(int cpu, int type)
3957 struct rq *rq = cpu_rq(cpu);
3958 unsigned long total = weighted_cpuload(cpu);
3960 if (type == 0 || !sched_feat(LB_BIAS))
3963 return max(rq->cpu_load[type-1], total);
3966 static unsigned long power_of(int cpu)
3968 return cpu_rq(cpu)->cpu_power;
3971 static unsigned long cpu_avg_load_per_task(int cpu)
3973 struct rq *rq = cpu_rq(cpu);
3974 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3975 unsigned long load_avg = rq->cfs.runnable_load_avg;
3978 return load_avg / nr_running;
3983 static void record_wakee(struct task_struct *p)
3986 * Rough decay (wiping) for cost saving, don't worry
3987 * about the boundary, really active task won't care
3990 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3991 current->wakee_flips = 0;
3992 current->wakee_flip_decay_ts = jiffies;
3995 if (current->last_wakee != p) {
3996 current->last_wakee = p;
3997 current->wakee_flips++;
4001 static void task_waking_fair(struct task_struct *p)
4003 struct sched_entity *se = &p->se;
4004 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4007 #ifndef CONFIG_64BIT
4008 u64 min_vruntime_copy;
4011 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4013 min_vruntime = cfs_rq->min_vruntime;
4014 } while (min_vruntime != min_vruntime_copy);
4016 min_vruntime = cfs_rq->min_vruntime;
4019 se->vruntime -= min_vruntime;
4023 #ifdef CONFIG_FAIR_GROUP_SCHED
4025 * effective_load() calculates the load change as seen from the root_task_group
4027 * Adding load to a group doesn't make a group heavier, but can cause movement
4028 * of group shares between cpus. Assuming the shares were perfectly aligned one
4029 * can calculate the shift in shares.
4031 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4032 * on this @cpu and results in a total addition (subtraction) of @wg to the
4033 * total group weight.
4035 * Given a runqueue weight distribution (rw_i) we can compute a shares
4036 * distribution (s_i) using:
4038 * s_i = rw_i / \Sum rw_j (1)
4040 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4041 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4042 * shares distribution (s_i):
4044 * rw_i = { 2, 4, 1, 0 }
4045 * s_i = { 2/7, 4/7, 1/7, 0 }
4047 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4048 * task used to run on and the CPU the waker is running on), we need to
4049 * compute the effect of waking a task on either CPU and, in case of a sync
4050 * wakeup, compute the effect of the current task going to sleep.
4052 * So for a change of @wl to the local @cpu with an overall group weight change
4053 * of @wl we can compute the new shares distribution (s'_i) using:
4055 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4057 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4058 * differences in waking a task to CPU 0. The additional task changes the
4059 * weight and shares distributions like:
4061 * rw'_i = { 3, 4, 1, 0 }
4062 * s'_i = { 3/8, 4/8, 1/8, 0 }
4064 * We can then compute the difference in effective weight by using:
4066 * dw_i = S * (s'_i - s_i) (3)
4068 * Where 'S' is the group weight as seen by its parent.
4070 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4071 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4072 * 4/7) times the weight of the group.
4074 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4076 struct sched_entity *se = tg->se[cpu];
4078 if (!tg->parent) /* the trivial, non-cgroup case */
4081 for_each_sched_entity(se) {
4087 * W = @wg + \Sum rw_j
4089 W = wg + calc_tg_weight(tg, se->my_q);
4094 w = se->my_q->load.weight + wl;
4097 * wl = S * s'_i; see (2)
4100 wl = (w * tg->shares) / W;
4105 * Per the above, wl is the new se->load.weight value; since
4106 * those are clipped to [MIN_SHARES, ...) do so now. See
4107 * calc_cfs_shares().
4109 if (wl < MIN_SHARES)
4113 * wl = dw_i = S * (s'_i - s_i); see (3)
4115 wl -= se->load.weight;
4118 * Recursively apply this logic to all parent groups to compute
4119 * the final effective load change on the root group. Since
4120 * only the @tg group gets extra weight, all parent groups can
4121 * only redistribute existing shares. @wl is the shift in shares
4122 * resulting from this level per the above.
4131 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4138 static int wake_wide(struct task_struct *p)
4140 int factor = this_cpu_read(sd_llc_size);
4143 * Yeah, it's the switching-frequency, could means many wakee or
4144 * rapidly switch, use factor here will just help to automatically
4145 * adjust the loose-degree, so bigger node will lead to more pull.
4147 if (p->wakee_flips > factor) {
4149 * wakee is somewhat hot, it needs certain amount of cpu
4150 * resource, so if waker is far more hot, prefer to leave
4153 if (current->wakee_flips > (factor * p->wakee_flips))
4160 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4162 s64 this_load, load;
4163 int idx, this_cpu, prev_cpu;
4164 unsigned long tl_per_task;
4165 struct task_group *tg;
4166 unsigned long weight;
4170 * If we wake multiple tasks be careful to not bounce
4171 * ourselves around too much.
4177 this_cpu = smp_processor_id();
4178 prev_cpu = task_cpu(p);
4179 load = source_load(prev_cpu, idx);
4180 this_load = target_load(this_cpu, idx);
4183 * If sync wakeup then subtract the (maximum possible)
4184 * effect of the currently running task from the load
4185 * of the current CPU:
4188 tg = task_group(current);
4189 weight = current->se.load.weight;
4191 this_load += effective_load(tg, this_cpu, -weight, -weight);
4192 load += effective_load(tg, prev_cpu, 0, -weight);
4196 weight = p->se.load.weight;
4199 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4200 * due to the sync cause above having dropped this_load to 0, we'll
4201 * always have an imbalance, but there's really nothing you can do
4202 * about that, so that's good too.
4204 * Otherwise check if either cpus are near enough in load to allow this
4205 * task to be woken on this_cpu.
4207 if (this_load > 0) {
4208 s64 this_eff_load, prev_eff_load;
4210 this_eff_load = 100;
4211 this_eff_load *= power_of(prev_cpu);
4212 this_eff_load *= this_load +
4213 effective_load(tg, this_cpu, weight, weight);
4215 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4216 prev_eff_load *= power_of(this_cpu);
4217 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4219 balanced = this_eff_load <= prev_eff_load;
4224 * If the currently running task will sleep within
4225 * a reasonable amount of time then attract this newly
4228 if (sync && balanced)
4231 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4232 tl_per_task = cpu_avg_load_per_task(this_cpu);
4235 (this_load <= load &&
4236 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4238 * This domain has SD_WAKE_AFFINE and
4239 * p is cache cold in this domain, and
4240 * there is no bad imbalance.
4242 schedstat_inc(sd, ttwu_move_affine);
4243 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4251 * find_idlest_group finds and returns the least busy CPU group within the
4254 static struct sched_group *
4255 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4256 int this_cpu, int sd_flag)
4258 struct sched_group *idlest = NULL, *group = sd->groups;
4259 unsigned long min_load = ULONG_MAX, this_load = 0;
4260 int load_idx = sd->forkexec_idx;
4261 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4263 if (sd_flag & SD_BALANCE_WAKE)
4264 load_idx = sd->wake_idx;
4267 unsigned long load, avg_load;
4271 /* Skip over this group if it has no CPUs allowed */
4272 if (!cpumask_intersects(sched_group_cpus(group),
4273 tsk_cpus_allowed(p)))
4276 local_group = cpumask_test_cpu(this_cpu,
4277 sched_group_cpus(group));
4279 /* Tally up the load of all CPUs in the group */
4282 for_each_cpu(i, sched_group_cpus(group)) {
4283 /* Bias balancing toward cpus of our domain */
4285 load = source_load(i, load_idx);
4287 load = target_load(i, load_idx);
4292 /* Adjust by relative CPU power of the group */
4293 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4296 this_load = avg_load;
4297 } else if (avg_load < min_load) {
4298 min_load = avg_load;
4301 } while (group = group->next, group != sd->groups);
4303 if (!idlest || 100*this_load < imbalance*min_load)
4309 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4312 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4314 unsigned long load, min_load = ULONG_MAX;
4318 /* Traverse only the allowed CPUs */
4319 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4320 load = weighted_cpuload(i);
4322 if (load < min_load || (load == min_load && i == this_cpu)) {
4332 * Try and locate an idle CPU in the sched_domain.
4334 static int select_idle_sibling(struct task_struct *p, int target)
4336 struct sched_domain *sd;
4337 struct sched_group *sg;
4338 int i = task_cpu(p);
4340 if (idle_cpu(target))
4344 * If the prevous cpu is cache affine and idle, don't be stupid.
4346 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4350 * Otherwise, iterate the domains and find an elegible idle cpu.
4352 sd = rcu_dereference(per_cpu(sd_llc, target));
4353 for_each_lower_domain(sd) {
4356 if (!cpumask_intersects(sched_group_cpus(sg),
4357 tsk_cpus_allowed(p)))
4360 for_each_cpu(i, sched_group_cpus(sg)) {
4361 if (i == target || !idle_cpu(i))
4365 target = cpumask_first_and(sched_group_cpus(sg),
4366 tsk_cpus_allowed(p));
4370 } while (sg != sd->groups);
4377 * sched_balance_self: balance the current task (running on cpu) in domains
4378 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4381 * Balance, ie. select the least loaded group.
4383 * Returns the target CPU number, or the same CPU if no balancing is needed.
4385 * preempt must be disabled.
4388 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4390 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4391 int cpu = smp_processor_id();
4393 int want_affine = 0;
4394 int sync = wake_flags & WF_SYNC;
4396 if (p->nr_cpus_allowed == 1)
4399 if (sd_flag & SD_BALANCE_WAKE) {
4400 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4406 for_each_domain(cpu, tmp) {
4407 if (!(tmp->flags & SD_LOAD_BALANCE))
4411 * If both cpu and prev_cpu are part of this domain,
4412 * cpu is a valid SD_WAKE_AFFINE target.
4414 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4415 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4420 if (tmp->flags & sd_flag)
4425 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4428 new_cpu = select_idle_sibling(p, prev_cpu);
4433 struct sched_group *group;
4436 if (!(sd->flags & sd_flag)) {
4441 group = find_idlest_group(sd, p, cpu, sd_flag);
4447 new_cpu = find_idlest_cpu(group, p, cpu);
4448 if (new_cpu == -1 || new_cpu == cpu) {
4449 /* Now try balancing at a lower domain level of cpu */
4454 /* Now try balancing at a lower domain level of new_cpu */
4456 weight = sd->span_weight;
4458 for_each_domain(cpu, tmp) {
4459 if (weight <= tmp->span_weight)
4461 if (tmp->flags & sd_flag)
4464 /* while loop will break here if sd == NULL */
4473 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4474 * cfs_rq_of(p) references at time of call are still valid and identify the
4475 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4476 * other assumptions, including the state of rq->lock, should be made.
4479 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4481 struct sched_entity *se = &p->se;
4482 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4485 * Load tracking: accumulate removed load so that it can be processed
4486 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4487 * to blocked load iff they have a positive decay-count. It can never
4488 * be negative here since on-rq tasks have decay-count == 0.
4490 if (se->avg.decay_count) {
4491 se->avg.decay_count = -__synchronize_entity_decay(se);
4492 atomic_long_add(se->avg.load_avg_contrib,
4493 &cfs_rq->removed_load);
4496 #endif /* CONFIG_SMP */
4498 static unsigned long
4499 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4501 unsigned long gran = sysctl_sched_wakeup_granularity;
4504 * Since its curr running now, convert the gran from real-time
4505 * to virtual-time in his units.
4507 * By using 'se' instead of 'curr' we penalize light tasks, so
4508 * they get preempted easier. That is, if 'se' < 'curr' then
4509 * the resulting gran will be larger, therefore penalizing the
4510 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4511 * be smaller, again penalizing the lighter task.
4513 * This is especially important for buddies when the leftmost
4514 * task is higher priority than the buddy.
4516 return calc_delta_fair(gran, se);
4520 * Should 'se' preempt 'curr'.
4534 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4536 s64 gran, vdiff = curr->vruntime - se->vruntime;
4541 gran = wakeup_gran(curr, se);
4548 static void set_last_buddy(struct sched_entity *se)
4550 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4553 for_each_sched_entity(se)
4554 cfs_rq_of(se)->last = se;
4557 static void set_next_buddy(struct sched_entity *se)
4559 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4562 for_each_sched_entity(se)
4563 cfs_rq_of(se)->next = se;
4566 static void set_skip_buddy(struct sched_entity *se)
4568 for_each_sched_entity(se)
4569 cfs_rq_of(se)->skip = se;
4573 * Preempt the current task with a newly woken task if needed:
4575 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4577 struct task_struct *curr = rq->curr;
4578 struct sched_entity *se = &curr->se, *pse = &p->se;
4579 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4580 int scale = cfs_rq->nr_running >= sched_nr_latency;
4581 int next_buddy_marked = 0;
4583 if (unlikely(se == pse))
4587 * This is possible from callers such as move_task(), in which we
4588 * unconditionally check_prempt_curr() after an enqueue (which may have
4589 * lead to a throttle). This both saves work and prevents false
4590 * next-buddy nomination below.
4592 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4595 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4596 set_next_buddy(pse);
4597 next_buddy_marked = 1;
4601 * We can come here with TIF_NEED_RESCHED already set from new task
4604 * Note: this also catches the edge-case of curr being in a throttled
4605 * group (e.g. via set_curr_task), since update_curr() (in the
4606 * enqueue of curr) will have resulted in resched being set. This
4607 * prevents us from potentially nominating it as a false LAST_BUDDY
4610 if (test_tsk_need_resched(curr))
4613 /* Idle tasks are by definition preempted by non-idle tasks. */
4614 if (unlikely(curr->policy == SCHED_IDLE) &&
4615 likely(p->policy != SCHED_IDLE))
4619 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4620 * is driven by the tick):
4622 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4625 find_matching_se(&se, &pse);
4626 update_curr(cfs_rq_of(se));
4628 if (wakeup_preempt_entity(se, pse) == 1) {
4630 * Bias pick_next to pick the sched entity that is
4631 * triggering this preemption.
4633 if (!next_buddy_marked)
4634 set_next_buddy(pse);
4643 * Only set the backward buddy when the current task is still
4644 * on the rq. This can happen when a wakeup gets interleaved
4645 * with schedule on the ->pre_schedule() or idle_balance()
4646 * point, either of which can * drop the rq lock.
4648 * Also, during early boot the idle thread is in the fair class,
4649 * for obvious reasons its a bad idea to schedule back to it.
4651 if (unlikely(!se->on_rq || curr == rq->idle))
4654 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4658 static struct task_struct *
4659 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4661 struct task_struct *p;
4662 struct cfs_rq *cfs_rq = &rq->cfs;
4663 struct sched_entity *se;
4665 if (!cfs_rq->nr_running)
4669 prev->sched_class->put_prev_task(rq, prev);
4672 se = pick_next_entity(cfs_rq);
4673 set_next_entity(cfs_rq, se);
4674 cfs_rq = group_cfs_rq(se);
4678 if (hrtick_enabled(rq))
4679 hrtick_start_fair(rq, p);
4685 * Account for a descheduled task:
4687 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4689 struct sched_entity *se = &prev->se;
4690 struct cfs_rq *cfs_rq;
4692 for_each_sched_entity(se) {
4693 cfs_rq = cfs_rq_of(se);
4694 put_prev_entity(cfs_rq, se);
4699 * sched_yield() is very simple
4701 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4703 static void yield_task_fair(struct rq *rq)
4705 struct task_struct *curr = rq->curr;
4706 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4707 struct sched_entity *se = &curr->se;
4710 * Are we the only task in the tree?
4712 if (unlikely(rq->nr_running == 1))
4715 clear_buddies(cfs_rq, se);
4717 if (curr->policy != SCHED_BATCH) {
4718 update_rq_clock(rq);
4720 * Update run-time statistics of the 'current'.
4722 update_curr(cfs_rq);
4724 * Tell update_rq_clock() that we've just updated,
4725 * so we don't do microscopic update in schedule()
4726 * and double the fastpath cost.
4728 rq->skip_clock_update = 1;
4734 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4736 struct sched_entity *se = &p->se;
4738 /* throttled hierarchies are not runnable */
4739 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4742 /* Tell the scheduler that we'd really like pse to run next. */
4745 yield_task_fair(rq);
4751 /**************************************************
4752 * Fair scheduling class load-balancing methods.
4756 * The purpose of load-balancing is to achieve the same basic fairness the
4757 * per-cpu scheduler provides, namely provide a proportional amount of compute
4758 * time to each task. This is expressed in the following equation:
4760 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4762 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4763 * W_i,0 is defined as:
4765 * W_i,0 = \Sum_j w_i,j (2)
4767 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4768 * is derived from the nice value as per prio_to_weight[].
4770 * The weight average is an exponential decay average of the instantaneous
4773 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4775 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4776 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4777 * can also include other factors [XXX].
4779 * To achieve this balance we define a measure of imbalance which follows
4780 * directly from (1):
4782 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4784 * We them move tasks around to minimize the imbalance. In the continuous
4785 * function space it is obvious this converges, in the discrete case we get
4786 * a few fun cases generally called infeasible weight scenarios.
4789 * - infeasible weights;
4790 * - local vs global optima in the discrete case. ]
4795 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4796 * for all i,j solution, we create a tree of cpus that follows the hardware
4797 * topology where each level pairs two lower groups (or better). This results
4798 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4799 * tree to only the first of the previous level and we decrease the frequency
4800 * of load-balance at each level inv. proportional to the number of cpus in
4806 * \Sum { --- * --- * 2^i } = O(n) (5)
4808 * `- size of each group
4809 * | | `- number of cpus doing load-balance
4811 * `- sum over all levels
4813 * Coupled with a limit on how many tasks we can migrate every balance pass,
4814 * this makes (5) the runtime complexity of the balancer.
4816 * An important property here is that each CPU is still (indirectly) connected
4817 * to every other cpu in at most O(log n) steps:
4819 * The adjacency matrix of the resulting graph is given by:
4822 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4825 * And you'll find that:
4827 * A^(log_2 n)_i,j != 0 for all i,j (7)
4829 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4830 * The task movement gives a factor of O(m), giving a convergence complexity
4833 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4838 * In order to avoid CPUs going idle while there's still work to do, new idle
4839 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4840 * tree itself instead of relying on other CPUs to bring it work.
4842 * This adds some complexity to both (5) and (8) but it reduces the total idle
4850 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4853 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4858 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4860 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4862 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4865 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4866 * rewrite all of this once again.]
4869 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4871 enum fbq_type { regular, remote, all };
4873 #define LBF_ALL_PINNED 0x01
4874 #define LBF_NEED_BREAK 0x02
4875 #define LBF_DST_PINNED 0x04
4876 #define LBF_SOME_PINNED 0x08
4879 struct sched_domain *sd;
4887 struct cpumask *dst_grpmask;
4889 enum cpu_idle_type idle;
4891 /* The set of CPUs under consideration for load-balancing */
4892 struct cpumask *cpus;
4897 unsigned int loop_break;
4898 unsigned int loop_max;
4900 enum fbq_type fbq_type;
4904 * move_task - move a task from one runqueue to another runqueue.
4905 * Both runqueues must be locked.
4907 static void move_task(struct task_struct *p, struct lb_env *env)
4909 deactivate_task(env->src_rq, p, 0);
4910 set_task_cpu(p, env->dst_cpu);
4911 activate_task(env->dst_rq, p, 0);
4912 check_preempt_curr(env->dst_rq, p, 0);
4916 * Is this task likely cache-hot:
4919 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4923 if (p->sched_class != &fair_sched_class)
4926 if (unlikely(p->policy == SCHED_IDLE))
4930 * Buddy candidates are cache hot:
4932 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4933 (&p->se == cfs_rq_of(&p->se)->next ||
4934 &p->se == cfs_rq_of(&p->se)->last))
4937 if (sysctl_sched_migration_cost == -1)
4939 if (sysctl_sched_migration_cost == 0)
4942 delta = now - p->se.exec_start;
4944 return delta < (s64)sysctl_sched_migration_cost;
4947 #ifdef CONFIG_NUMA_BALANCING
4948 /* Returns true if the destination node has incurred more faults */
4949 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4951 int src_nid, dst_nid;
4953 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
4954 !(env->sd->flags & SD_NUMA)) {
4958 src_nid = cpu_to_node(env->src_cpu);
4959 dst_nid = cpu_to_node(env->dst_cpu);
4961 if (src_nid == dst_nid)
4964 /* Always encourage migration to the preferred node. */
4965 if (dst_nid == p->numa_preferred_nid)
4968 /* If both task and group weight improve, this move is a winner. */
4969 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4970 group_weight(p, dst_nid) > group_weight(p, src_nid))
4977 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4979 int src_nid, dst_nid;
4981 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4984 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
4987 src_nid = cpu_to_node(env->src_cpu);
4988 dst_nid = cpu_to_node(env->dst_cpu);
4990 if (src_nid == dst_nid)
4993 /* Migrating away from the preferred node is always bad. */
4994 if (src_nid == p->numa_preferred_nid)
4997 /* If either task or group weight get worse, don't do it. */
4998 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
4999 group_weight(p, dst_nid) < group_weight(p, src_nid))
5006 static inline bool migrate_improves_locality(struct task_struct *p,
5012 static inline bool migrate_degrades_locality(struct task_struct *p,
5020 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5023 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5025 int tsk_cache_hot = 0;
5027 * We do not migrate tasks that are:
5028 * 1) throttled_lb_pair, or
5029 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5030 * 3) running (obviously), or
5031 * 4) are cache-hot on their current CPU.
5033 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5036 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5039 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5041 env->flags |= LBF_SOME_PINNED;
5044 * Remember if this task can be migrated to any other cpu in
5045 * our sched_group. We may want to revisit it if we couldn't
5046 * meet load balance goals by pulling other tasks on src_cpu.
5048 * Also avoid computing new_dst_cpu if we have already computed
5049 * one in current iteration.
5051 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5054 /* Prevent to re-select dst_cpu via env's cpus */
5055 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5056 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5057 env->flags |= LBF_DST_PINNED;
5058 env->new_dst_cpu = cpu;
5066 /* Record that we found atleast one task that could run on dst_cpu */
5067 env->flags &= ~LBF_ALL_PINNED;
5069 if (task_running(env->src_rq, p)) {
5070 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5075 * Aggressive migration if:
5076 * 1) destination numa is preferred
5077 * 2) task is cache cold, or
5078 * 3) too many balance attempts have failed.
5080 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
5082 tsk_cache_hot = migrate_degrades_locality(p, env);
5084 if (migrate_improves_locality(p, env)) {
5085 #ifdef CONFIG_SCHEDSTATS
5086 if (tsk_cache_hot) {
5087 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5088 schedstat_inc(p, se.statistics.nr_forced_migrations);
5094 if (!tsk_cache_hot ||
5095 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5097 if (tsk_cache_hot) {
5098 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5099 schedstat_inc(p, se.statistics.nr_forced_migrations);
5105 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5110 * move_one_task tries to move exactly one task from busiest to this_rq, as
5111 * part of active balancing operations within "domain".
5112 * Returns 1 if successful and 0 otherwise.
5114 * Called with both runqueues locked.
5116 static int move_one_task(struct lb_env *env)
5118 struct task_struct *p, *n;
5120 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5121 if (!can_migrate_task(p, env))
5126 * Right now, this is only the second place move_task()
5127 * is called, so we can safely collect move_task()
5128 * stats here rather than inside move_task().
5130 schedstat_inc(env->sd, lb_gained[env->idle]);
5136 static const unsigned int sched_nr_migrate_break = 32;
5139 * move_tasks tries to move up to imbalance weighted load from busiest to
5140 * this_rq, as part of a balancing operation within domain "sd".
5141 * Returns 1 if successful and 0 otherwise.
5143 * Called with both runqueues locked.
5145 static int move_tasks(struct lb_env *env)
5147 struct list_head *tasks = &env->src_rq->cfs_tasks;
5148 struct task_struct *p;
5152 if (env->imbalance <= 0)
5155 while (!list_empty(tasks)) {
5156 p = list_first_entry(tasks, struct task_struct, se.group_node);
5159 /* We've more or less seen every task there is, call it quits */
5160 if (env->loop > env->loop_max)
5163 /* take a breather every nr_migrate tasks */
5164 if (env->loop > env->loop_break) {
5165 env->loop_break += sched_nr_migrate_break;
5166 env->flags |= LBF_NEED_BREAK;
5170 if (!can_migrate_task(p, env))
5173 load = task_h_load(p);
5175 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5178 if ((load / 2) > env->imbalance)
5183 env->imbalance -= load;
5185 #ifdef CONFIG_PREEMPT
5187 * NEWIDLE balancing is a source of latency, so preemptible
5188 * kernels will stop after the first task is pulled to minimize
5189 * the critical section.
5191 if (env->idle == CPU_NEWLY_IDLE)
5196 * We only want to steal up to the prescribed amount of
5199 if (env->imbalance <= 0)
5204 list_move_tail(&p->se.group_node, tasks);
5208 * Right now, this is one of only two places move_task() is called,
5209 * so we can safely collect move_task() stats here rather than
5210 * inside move_task().
5212 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5217 #ifdef CONFIG_FAIR_GROUP_SCHED
5219 * update tg->load_weight by folding this cpu's load_avg
5221 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5223 struct sched_entity *se = tg->se[cpu];
5224 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5226 /* throttled entities do not contribute to load */
5227 if (throttled_hierarchy(cfs_rq))
5230 update_cfs_rq_blocked_load(cfs_rq, 1);
5233 update_entity_load_avg(se, 1);
5235 * We pivot on our runnable average having decayed to zero for
5236 * list removal. This generally implies that all our children
5237 * have also been removed (modulo rounding error or bandwidth
5238 * control); however, such cases are rare and we can fix these
5241 * TODO: fix up out-of-order children on enqueue.
5243 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5244 list_del_leaf_cfs_rq(cfs_rq);
5246 struct rq *rq = rq_of(cfs_rq);
5247 update_rq_runnable_avg(rq, rq->nr_running);
5251 static void update_blocked_averages(int cpu)
5253 struct rq *rq = cpu_rq(cpu);
5254 struct cfs_rq *cfs_rq;
5255 unsigned long flags;
5257 raw_spin_lock_irqsave(&rq->lock, flags);
5258 update_rq_clock(rq);
5260 * Iterates the task_group tree in a bottom up fashion, see
5261 * list_add_leaf_cfs_rq() for details.
5263 for_each_leaf_cfs_rq(rq, cfs_rq) {
5265 * Note: We may want to consider periodically releasing
5266 * rq->lock about these updates so that creating many task
5267 * groups does not result in continually extending hold time.
5269 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5272 raw_spin_unlock_irqrestore(&rq->lock, flags);
5276 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5277 * This needs to be done in a top-down fashion because the load of a child
5278 * group is a fraction of its parents load.
5280 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5282 struct rq *rq = rq_of(cfs_rq);
5283 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5284 unsigned long now = jiffies;
5287 if (cfs_rq->last_h_load_update == now)
5290 cfs_rq->h_load_next = NULL;
5291 for_each_sched_entity(se) {
5292 cfs_rq = cfs_rq_of(se);
5293 cfs_rq->h_load_next = se;
5294 if (cfs_rq->last_h_load_update == now)
5299 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5300 cfs_rq->last_h_load_update = now;
5303 while ((se = cfs_rq->h_load_next) != NULL) {
5304 load = cfs_rq->h_load;
5305 load = div64_ul(load * se->avg.load_avg_contrib,
5306 cfs_rq->runnable_load_avg + 1);
5307 cfs_rq = group_cfs_rq(se);
5308 cfs_rq->h_load = load;
5309 cfs_rq->last_h_load_update = now;
5313 static unsigned long task_h_load(struct task_struct *p)
5315 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5317 update_cfs_rq_h_load(cfs_rq);
5318 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5319 cfs_rq->runnable_load_avg + 1);
5322 static inline void update_blocked_averages(int cpu)
5326 static unsigned long task_h_load(struct task_struct *p)
5328 return p->se.avg.load_avg_contrib;
5332 /********** Helpers for find_busiest_group ************************/
5334 * sg_lb_stats - stats of a sched_group required for load_balancing
5336 struct sg_lb_stats {
5337 unsigned long avg_load; /*Avg load across the CPUs of the group */
5338 unsigned long group_load; /* Total load over the CPUs of the group */
5339 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5340 unsigned long load_per_task;
5341 unsigned long group_power;
5342 unsigned int sum_nr_running; /* Nr tasks running in the group */
5343 unsigned int group_capacity;
5344 unsigned int idle_cpus;
5345 unsigned int group_weight;
5346 int group_imb; /* Is there an imbalance in the group ? */
5347 int group_has_capacity; /* Is there extra capacity in the group? */
5348 #ifdef CONFIG_NUMA_BALANCING
5349 unsigned int nr_numa_running;
5350 unsigned int nr_preferred_running;
5355 * sd_lb_stats - Structure to store the statistics of a sched_domain
5356 * during load balancing.
5358 struct sd_lb_stats {
5359 struct sched_group *busiest; /* Busiest group in this sd */
5360 struct sched_group *local; /* Local group in this sd */
5361 unsigned long total_load; /* Total load of all groups in sd */
5362 unsigned long total_pwr; /* Total power of all groups in sd */
5363 unsigned long avg_load; /* Average load across all groups in sd */
5365 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5366 struct sg_lb_stats local_stat; /* Statistics of the local group */
5369 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5372 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5373 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5374 * We must however clear busiest_stat::avg_load because
5375 * update_sd_pick_busiest() reads this before assignment.
5377 *sds = (struct sd_lb_stats){
5389 * get_sd_load_idx - Obtain the load index for a given sched domain.
5390 * @sd: The sched_domain whose load_idx is to be obtained.
5391 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5393 * Return: The load index.
5395 static inline int get_sd_load_idx(struct sched_domain *sd,
5396 enum cpu_idle_type idle)
5402 load_idx = sd->busy_idx;
5405 case CPU_NEWLY_IDLE:
5406 load_idx = sd->newidle_idx;
5409 load_idx = sd->idle_idx;
5416 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5418 return SCHED_POWER_SCALE;
5421 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5423 return default_scale_freq_power(sd, cpu);
5426 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5428 unsigned long weight = sd->span_weight;
5429 unsigned long smt_gain = sd->smt_gain;
5436 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5438 return default_scale_smt_power(sd, cpu);
5441 static unsigned long scale_rt_power(int cpu)
5443 struct rq *rq = cpu_rq(cpu);
5444 u64 total, available, age_stamp, avg;
5447 * Since we're reading these variables without serialization make sure
5448 * we read them once before doing sanity checks on them.
5450 age_stamp = ACCESS_ONCE(rq->age_stamp);
5451 avg = ACCESS_ONCE(rq->rt_avg);
5453 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5455 if (unlikely(total < avg)) {
5456 /* Ensures that power won't end up being negative */
5459 available = total - avg;
5462 if (unlikely((s64)total < SCHED_POWER_SCALE))
5463 total = SCHED_POWER_SCALE;
5465 total >>= SCHED_POWER_SHIFT;
5467 return div_u64(available, total);
5470 static void update_cpu_power(struct sched_domain *sd, int cpu)
5472 unsigned long weight = sd->span_weight;
5473 unsigned long power = SCHED_POWER_SCALE;
5474 struct sched_group *sdg = sd->groups;
5476 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5477 if (sched_feat(ARCH_POWER))
5478 power *= arch_scale_smt_power(sd, cpu);
5480 power *= default_scale_smt_power(sd, cpu);
5482 power >>= SCHED_POWER_SHIFT;
5485 sdg->sgp->power_orig = power;
5487 if (sched_feat(ARCH_POWER))
5488 power *= arch_scale_freq_power(sd, cpu);
5490 power *= default_scale_freq_power(sd, cpu);
5492 power >>= SCHED_POWER_SHIFT;
5494 power *= scale_rt_power(cpu);
5495 power >>= SCHED_POWER_SHIFT;
5500 cpu_rq(cpu)->cpu_power = power;
5501 sdg->sgp->power = power;
5504 void update_group_power(struct sched_domain *sd, int cpu)
5506 struct sched_domain *child = sd->child;
5507 struct sched_group *group, *sdg = sd->groups;
5508 unsigned long power, power_orig;
5509 unsigned long interval;
5511 interval = msecs_to_jiffies(sd->balance_interval);
5512 interval = clamp(interval, 1UL, max_load_balance_interval);
5513 sdg->sgp->next_update = jiffies + interval;
5516 update_cpu_power(sd, cpu);
5520 power_orig = power = 0;
5522 if (child->flags & SD_OVERLAP) {
5524 * SD_OVERLAP domains cannot assume that child groups
5525 * span the current group.
5528 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5529 struct sched_group_power *sgp;
5530 struct rq *rq = cpu_rq(cpu);
5533 * build_sched_domains() -> init_sched_groups_power()
5534 * gets here before we've attached the domains to the
5537 * Use power_of(), which is set irrespective of domains
5538 * in update_cpu_power().
5540 * This avoids power/power_orig from being 0 and
5541 * causing divide-by-zero issues on boot.
5543 * Runtime updates will correct power_orig.
5545 if (unlikely(!rq->sd)) {
5546 power_orig += power_of(cpu);
5547 power += power_of(cpu);
5551 sgp = rq->sd->groups->sgp;
5552 power_orig += sgp->power_orig;
5553 power += sgp->power;
5557 * !SD_OVERLAP domains can assume that child groups
5558 * span the current group.
5561 group = child->groups;
5563 power_orig += group->sgp->power_orig;
5564 power += group->sgp->power;
5565 group = group->next;
5566 } while (group != child->groups);
5569 sdg->sgp->power_orig = power_orig;
5570 sdg->sgp->power = power;
5574 * Try and fix up capacity for tiny siblings, this is needed when
5575 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5576 * which on its own isn't powerful enough.
5578 * See update_sd_pick_busiest() and check_asym_packing().
5581 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5584 * Only siblings can have significantly less than SCHED_POWER_SCALE
5586 if (!(sd->flags & SD_SHARE_CPUPOWER))
5590 * If ~90% of the cpu_power is still there, we're good.
5592 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5599 * Group imbalance indicates (and tries to solve) the problem where balancing
5600 * groups is inadequate due to tsk_cpus_allowed() constraints.
5602 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5603 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5606 * { 0 1 2 3 } { 4 5 6 7 }
5609 * If we were to balance group-wise we'd place two tasks in the first group and
5610 * two tasks in the second group. Clearly this is undesired as it will overload
5611 * cpu 3 and leave one of the cpus in the second group unused.
5613 * The current solution to this issue is detecting the skew in the first group
5614 * by noticing the lower domain failed to reach balance and had difficulty
5615 * moving tasks due to affinity constraints.
5617 * When this is so detected; this group becomes a candidate for busiest; see
5618 * update_sd_pick_busiest(). And calculate_imbalance() and
5619 * find_busiest_group() avoid some of the usual balance conditions to allow it
5620 * to create an effective group imbalance.
5622 * This is a somewhat tricky proposition since the next run might not find the
5623 * group imbalance and decide the groups need to be balanced again. A most
5624 * subtle and fragile situation.
5627 static inline int sg_imbalanced(struct sched_group *group)
5629 return group->sgp->imbalance;
5633 * Compute the group capacity.
5635 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5636 * first dividing out the smt factor and computing the actual number of cores
5637 * and limit power unit capacity with that.
5639 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5641 unsigned int capacity, smt, cpus;
5642 unsigned int power, power_orig;
5644 power = group->sgp->power;
5645 power_orig = group->sgp->power_orig;
5646 cpus = group->group_weight;
5648 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5649 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5650 capacity = cpus / smt; /* cores */
5652 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5654 capacity = fix_small_capacity(env->sd, group);
5660 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5661 * @env: The load balancing environment.
5662 * @group: sched_group whose statistics are to be updated.
5663 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5664 * @local_group: Does group contain this_cpu.
5665 * @sgs: variable to hold the statistics for this group.
5667 static inline void update_sg_lb_stats(struct lb_env *env,
5668 struct sched_group *group, int load_idx,
5669 int local_group, struct sg_lb_stats *sgs)
5674 memset(sgs, 0, sizeof(*sgs));
5676 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5677 struct rq *rq = cpu_rq(i);
5679 /* Bias balancing toward cpus of our domain */
5681 load = target_load(i, load_idx);
5683 load = source_load(i, load_idx);
5685 sgs->group_load += load;
5686 sgs->sum_nr_running += rq->nr_running;
5687 #ifdef CONFIG_NUMA_BALANCING
5688 sgs->nr_numa_running += rq->nr_numa_running;
5689 sgs->nr_preferred_running += rq->nr_preferred_running;
5691 sgs->sum_weighted_load += weighted_cpuload(i);
5696 /* Adjust by relative CPU power of the group */
5697 sgs->group_power = group->sgp->power;
5698 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5700 if (sgs->sum_nr_running)
5701 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5703 sgs->group_weight = group->group_weight;
5705 sgs->group_imb = sg_imbalanced(group);
5706 sgs->group_capacity = sg_capacity(env, group);
5708 if (sgs->group_capacity > sgs->sum_nr_running)
5709 sgs->group_has_capacity = 1;
5713 * update_sd_pick_busiest - return 1 on busiest group
5714 * @env: The load balancing environment.
5715 * @sds: sched_domain statistics
5716 * @sg: sched_group candidate to be checked for being the busiest
5717 * @sgs: sched_group statistics
5719 * Determine if @sg is a busier group than the previously selected
5722 * Return: %true if @sg is a busier group than the previously selected
5723 * busiest group. %false otherwise.
5725 static bool update_sd_pick_busiest(struct lb_env *env,
5726 struct sd_lb_stats *sds,
5727 struct sched_group *sg,
5728 struct sg_lb_stats *sgs)
5730 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5733 if (sgs->sum_nr_running > sgs->group_capacity)
5740 * ASYM_PACKING needs to move all the work to the lowest
5741 * numbered CPUs in the group, therefore mark all groups
5742 * higher than ourself as busy.
5744 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5745 env->dst_cpu < group_first_cpu(sg)) {
5749 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5756 #ifdef CONFIG_NUMA_BALANCING
5757 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5759 if (sgs->sum_nr_running > sgs->nr_numa_running)
5761 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5766 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5768 if (rq->nr_running > rq->nr_numa_running)
5770 if (rq->nr_running > rq->nr_preferred_running)
5775 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5780 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5784 #endif /* CONFIG_NUMA_BALANCING */
5787 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5788 * @env: The load balancing environment.
5789 * @sds: variable to hold the statistics for this sched_domain.
5791 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5793 struct sched_domain *child = env->sd->child;
5794 struct sched_group *sg = env->sd->groups;
5795 struct sg_lb_stats tmp_sgs;
5796 int load_idx, prefer_sibling = 0;
5798 if (child && child->flags & SD_PREFER_SIBLING)
5801 load_idx = get_sd_load_idx(env->sd, env->idle);
5804 struct sg_lb_stats *sgs = &tmp_sgs;
5807 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5810 sgs = &sds->local_stat;
5812 if (env->idle != CPU_NEWLY_IDLE ||
5813 time_after_eq(jiffies, sg->sgp->next_update))
5814 update_group_power(env->sd, env->dst_cpu);
5817 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5823 * In case the child domain prefers tasks go to siblings
5824 * first, lower the sg capacity to one so that we'll try
5825 * and move all the excess tasks away. We lower the capacity
5826 * of a group only if the local group has the capacity to fit
5827 * these excess tasks, i.e. nr_running < group_capacity. The
5828 * extra check prevents the case where you always pull from the
5829 * heaviest group when it is already under-utilized (possible
5830 * with a large weight task outweighs the tasks on the system).
5832 if (prefer_sibling && sds->local &&
5833 sds->local_stat.group_has_capacity)
5834 sgs->group_capacity = min(sgs->group_capacity, 1U);
5836 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5838 sds->busiest_stat = *sgs;
5842 /* Now, start updating sd_lb_stats */
5843 sds->total_load += sgs->group_load;
5844 sds->total_pwr += sgs->group_power;
5847 } while (sg != env->sd->groups);
5849 if (env->sd->flags & SD_NUMA)
5850 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5854 * check_asym_packing - Check to see if the group is packed into the
5857 * This is primarily intended to used at the sibling level. Some
5858 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5859 * case of POWER7, it can move to lower SMT modes only when higher
5860 * threads are idle. When in lower SMT modes, the threads will
5861 * perform better since they share less core resources. Hence when we
5862 * have idle threads, we want them to be the higher ones.
5864 * This packing function is run on idle threads. It checks to see if
5865 * the busiest CPU in this domain (core in the P7 case) has a higher
5866 * CPU number than the packing function is being run on. Here we are
5867 * assuming lower CPU number will be equivalent to lower a SMT thread
5870 * Return: 1 when packing is required and a task should be moved to
5871 * this CPU. The amount of the imbalance is returned in *imbalance.
5873 * @env: The load balancing environment.
5874 * @sds: Statistics of the sched_domain which is to be packed
5876 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5880 if (!(env->sd->flags & SD_ASYM_PACKING))
5886 busiest_cpu = group_first_cpu(sds->busiest);
5887 if (env->dst_cpu > busiest_cpu)
5890 env->imbalance = DIV_ROUND_CLOSEST(
5891 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5898 * fix_small_imbalance - Calculate the minor imbalance that exists
5899 * amongst the groups of a sched_domain, during
5901 * @env: The load balancing environment.
5902 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5905 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5907 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5908 unsigned int imbn = 2;
5909 unsigned long scaled_busy_load_per_task;
5910 struct sg_lb_stats *local, *busiest;
5912 local = &sds->local_stat;
5913 busiest = &sds->busiest_stat;
5915 if (!local->sum_nr_running)
5916 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5917 else if (busiest->load_per_task > local->load_per_task)
5920 scaled_busy_load_per_task =
5921 (busiest->load_per_task * SCHED_POWER_SCALE) /
5922 busiest->group_power;
5924 if (busiest->avg_load + scaled_busy_load_per_task >=
5925 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5926 env->imbalance = busiest->load_per_task;
5931 * OK, we don't have enough imbalance to justify moving tasks,
5932 * however we may be able to increase total CPU power used by
5936 pwr_now += busiest->group_power *
5937 min(busiest->load_per_task, busiest->avg_load);
5938 pwr_now += local->group_power *
5939 min(local->load_per_task, local->avg_load);
5940 pwr_now /= SCHED_POWER_SCALE;
5942 /* Amount of load we'd subtract */
5943 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5944 busiest->group_power;
5945 if (busiest->avg_load > tmp) {
5946 pwr_move += busiest->group_power *
5947 min(busiest->load_per_task,
5948 busiest->avg_load - tmp);
5951 /* Amount of load we'd add */
5952 if (busiest->avg_load * busiest->group_power <
5953 busiest->load_per_task * SCHED_POWER_SCALE) {
5954 tmp = (busiest->avg_load * busiest->group_power) /
5957 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5960 pwr_move += local->group_power *
5961 min(local->load_per_task, local->avg_load + tmp);
5962 pwr_move /= SCHED_POWER_SCALE;
5964 /* Move if we gain throughput */
5965 if (pwr_move > pwr_now)
5966 env->imbalance = busiest->load_per_task;
5970 * calculate_imbalance - Calculate the amount of imbalance present within the
5971 * groups of a given sched_domain during load balance.
5972 * @env: load balance environment
5973 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5975 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5977 unsigned long max_pull, load_above_capacity = ~0UL;
5978 struct sg_lb_stats *local, *busiest;
5980 local = &sds->local_stat;
5981 busiest = &sds->busiest_stat;
5983 if (busiest->group_imb) {
5985 * In the group_imb case we cannot rely on group-wide averages
5986 * to ensure cpu-load equilibrium, look at wider averages. XXX
5988 busiest->load_per_task =
5989 min(busiest->load_per_task, sds->avg_load);
5993 * In the presence of smp nice balancing, certain scenarios can have
5994 * max load less than avg load(as we skip the groups at or below
5995 * its cpu_power, while calculating max_load..)
5997 if (busiest->avg_load <= sds->avg_load ||
5998 local->avg_load >= sds->avg_load) {
6000 return fix_small_imbalance(env, sds);
6003 if (!busiest->group_imb) {
6005 * Don't want to pull so many tasks that a group would go idle.
6006 * Except of course for the group_imb case, since then we might
6007 * have to drop below capacity to reach cpu-load equilibrium.
6009 load_above_capacity =
6010 (busiest->sum_nr_running - busiest->group_capacity);
6012 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
6013 load_above_capacity /= busiest->group_power;
6017 * We're trying to get all the cpus to the average_load, so we don't
6018 * want to push ourselves above the average load, nor do we wish to
6019 * reduce the max loaded cpu below the average load. At the same time,
6020 * we also don't want to reduce the group load below the group capacity
6021 * (so that we can implement power-savings policies etc). Thus we look
6022 * for the minimum possible imbalance.
6024 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6026 /* How much load to actually move to equalise the imbalance */
6027 env->imbalance = min(
6028 max_pull * busiest->group_power,
6029 (sds->avg_load - local->avg_load) * local->group_power
6030 ) / SCHED_POWER_SCALE;
6033 * if *imbalance is less than the average load per runnable task
6034 * there is no guarantee that any tasks will be moved so we'll have
6035 * a think about bumping its value to force at least one task to be
6038 if (env->imbalance < busiest->load_per_task)
6039 return fix_small_imbalance(env, sds);
6042 /******* find_busiest_group() helpers end here *********************/
6045 * find_busiest_group - Returns the busiest group within the sched_domain
6046 * if there is an imbalance. If there isn't an imbalance, and
6047 * the user has opted for power-savings, it returns a group whose
6048 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6049 * such a group exists.
6051 * Also calculates the amount of weighted load which should be moved
6052 * to restore balance.
6054 * @env: The load balancing environment.
6056 * Return: - The busiest group if imbalance exists.
6057 * - If no imbalance and user has opted for power-savings balance,
6058 * return the least loaded group whose CPUs can be
6059 * put to idle by rebalancing its tasks onto our group.
6061 static struct sched_group *find_busiest_group(struct lb_env *env)
6063 struct sg_lb_stats *local, *busiest;
6064 struct sd_lb_stats sds;
6066 init_sd_lb_stats(&sds);
6069 * Compute the various statistics relavent for load balancing at
6072 update_sd_lb_stats(env, &sds);
6073 local = &sds.local_stat;
6074 busiest = &sds.busiest_stat;
6076 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6077 check_asym_packing(env, &sds))
6080 /* There is no busy sibling group to pull tasks from */
6081 if (!sds.busiest || busiest->sum_nr_running == 0)
6084 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
6087 * If the busiest group is imbalanced the below checks don't
6088 * work because they assume all things are equal, which typically
6089 * isn't true due to cpus_allowed constraints and the like.
6091 if (busiest->group_imb)
6094 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6095 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
6096 !busiest->group_has_capacity)
6100 * If the local group is more busy than the selected busiest group
6101 * don't try and pull any tasks.
6103 if (local->avg_load >= busiest->avg_load)
6107 * Don't pull any tasks if this group is already above the domain
6110 if (local->avg_load >= sds.avg_load)
6113 if (env->idle == CPU_IDLE) {
6115 * This cpu is idle. If the busiest group load doesn't
6116 * have more tasks than the number of available cpu's and
6117 * there is no imbalance between this and busiest group
6118 * wrt to idle cpu's, it is balanced.
6120 if ((local->idle_cpus < busiest->idle_cpus) &&
6121 busiest->sum_nr_running <= busiest->group_weight)
6125 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6126 * imbalance_pct to be conservative.
6128 if (100 * busiest->avg_load <=
6129 env->sd->imbalance_pct * local->avg_load)
6134 /* Looks like there is an imbalance. Compute it */
6135 calculate_imbalance(env, &sds);
6144 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6146 static struct rq *find_busiest_queue(struct lb_env *env,
6147 struct sched_group *group)
6149 struct rq *busiest = NULL, *rq;
6150 unsigned long busiest_load = 0, busiest_power = 1;
6153 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6154 unsigned long power, capacity, wl;
6158 rt = fbq_classify_rq(rq);
6161 * We classify groups/runqueues into three groups:
6162 * - regular: there are !numa tasks
6163 * - remote: there are numa tasks that run on the 'wrong' node
6164 * - all: there is no distinction
6166 * In order to avoid migrating ideally placed numa tasks,
6167 * ignore those when there's better options.
6169 * If we ignore the actual busiest queue to migrate another
6170 * task, the next balance pass can still reduce the busiest
6171 * queue by moving tasks around inside the node.
6173 * If we cannot move enough load due to this classification
6174 * the next pass will adjust the group classification and
6175 * allow migration of more tasks.
6177 * Both cases only affect the total convergence complexity.
6179 if (rt > env->fbq_type)
6182 power = power_of(i);
6183 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
6185 capacity = fix_small_capacity(env->sd, group);
6187 wl = weighted_cpuload(i);
6190 * When comparing with imbalance, use weighted_cpuload()
6191 * which is not scaled with the cpu power.
6193 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6197 * For the load comparisons with the other cpu's, consider
6198 * the weighted_cpuload() scaled with the cpu power, so that
6199 * the load can be moved away from the cpu that is potentially
6200 * running at a lower capacity.
6202 * Thus we're looking for max(wl_i / power_i), crosswise
6203 * multiplication to rid ourselves of the division works out
6204 * to: wl_i * power_j > wl_j * power_i; where j is our
6207 if (wl * busiest_power > busiest_load * power) {
6209 busiest_power = power;
6218 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6219 * so long as it is large enough.
6221 #define MAX_PINNED_INTERVAL 512
6223 /* Working cpumask for load_balance and load_balance_newidle. */
6224 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6226 static int need_active_balance(struct lb_env *env)
6228 struct sched_domain *sd = env->sd;
6230 if (env->idle == CPU_NEWLY_IDLE) {
6233 * ASYM_PACKING needs to force migrate tasks from busy but
6234 * higher numbered CPUs in order to pack all tasks in the
6235 * lowest numbered CPUs.
6237 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6241 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6244 static int active_load_balance_cpu_stop(void *data);
6246 static int should_we_balance(struct lb_env *env)
6248 struct sched_group *sg = env->sd->groups;
6249 struct cpumask *sg_cpus, *sg_mask;
6250 int cpu, balance_cpu = -1;
6253 * In the newly idle case, we will allow all the cpu's
6254 * to do the newly idle load balance.
6256 if (env->idle == CPU_NEWLY_IDLE)
6259 sg_cpus = sched_group_cpus(sg);
6260 sg_mask = sched_group_mask(sg);
6261 /* Try to find first idle cpu */
6262 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6263 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6270 if (balance_cpu == -1)
6271 balance_cpu = group_balance_cpu(sg);
6274 * First idle cpu or the first cpu(busiest) in this sched group
6275 * is eligible for doing load balancing at this and above domains.
6277 return balance_cpu == env->dst_cpu;
6281 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6282 * tasks if there is an imbalance.
6284 static int load_balance(int this_cpu, struct rq *this_rq,
6285 struct sched_domain *sd, enum cpu_idle_type idle,
6286 int *continue_balancing)
6288 int ld_moved, cur_ld_moved, active_balance = 0;
6289 struct sched_domain *sd_parent = sd->parent;
6290 struct sched_group *group;
6292 unsigned long flags;
6293 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6295 struct lb_env env = {
6297 .dst_cpu = this_cpu,
6299 .dst_grpmask = sched_group_cpus(sd->groups),
6301 .loop_break = sched_nr_migrate_break,
6307 * For NEWLY_IDLE load_balancing, we don't need to consider
6308 * other cpus in our group
6310 if (idle == CPU_NEWLY_IDLE)
6311 env.dst_grpmask = NULL;
6313 cpumask_copy(cpus, cpu_active_mask);
6315 schedstat_inc(sd, lb_count[idle]);
6318 if (!should_we_balance(&env)) {
6319 *continue_balancing = 0;
6323 group = find_busiest_group(&env);
6325 schedstat_inc(sd, lb_nobusyg[idle]);
6329 busiest = find_busiest_queue(&env, group);
6331 schedstat_inc(sd, lb_nobusyq[idle]);
6335 BUG_ON(busiest == env.dst_rq);
6337 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6340 if (busiest->nr_running > 1) {
6342 * Attempt to move tasks. If find_busiest_group has found
6343 * an imbalance but busiest->nr_running <= 1, the group is
6344 * still unbalanced. ld_moved simply stays zero, so it is
6345 * correctly treated as an imbalance.
6347 env.flags |= LBF_ALL_PINNED;
6348 env.src_cpu = busiest->cpu;
6349 env.src_rq = busiest;
6350 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6353 local_irq_save(flags);
6354 double_rq_lock(env.dst_rq, busiest);
6357 * cur_ld_moved - load moved in current iteration
6358 * ld_moved - cumulative load moved across iterations
6360 cur_ld_moved = move_tasks(&env);
6361 ld_moved += cur_ld_moved;
6362 double_rq_unlock(env.dst_rq, busiest);
6363 local_irq_restore(flags);
6366 * some other cpu did the load balance for us.
6368 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6369 resched_cpu(env.dst_cpu);
6371 if (env.flags & LBF_NEED_BREAK) {
6372 env.flags &= ~LBF_NEED_BREAK;
6377 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6378 * us and move them to an alternate dst_cpu in our sched_group
6379 * where they can run. The upper limit on how many times we
6380 * iterate on same src_cpu is dependent on number of cpus in our
6383 * This changes load balance semantics a bit on who can move
6384 * load to a given_cpu. In addition to the given_cpu itself
6385 * (or a ilb_cpu acting on its behalf where given_cpu is
6386 * nohz-idle), we now have balance_cpu in a position to move
6387 * load to given_cpu. In rare situations, this may cause
6388 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6389 * _independently_ and at _same_ time to move some load to
6390 * given_cpu) causing exceess load to be moved to given_cpu.
6391 * This however should not happen so much in practice and
6392 * moreover subsequent load balance cycles should correct the
6393 * excess load moved.
6395 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6397 /* Prevent to re-select dst_cpu via env's cpus */
6398 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6400 env.dst_rq = cpu_rq(env.new_dst_cpu);
6401 env.dst_cpu = env.new_dst_cpu;
6402 env.flags &= ~LBF_DST_PINNED;
6404 env.loop_break = sched_nr_migrate_break;
6407 * Go back to "more_balance" rather than "redo" since we
6408 * need to continue with same src_cpu.
6414 * We failed to reach balance because of affinity.
6417 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6419 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6420 *group_imbalance = 1;
6421 } else if (*group_imbalance)
6422 *group_imbalance = 0;
6425 /* All tasks on this runqueue were pinned by CPU affinity */
6426 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6427 cpumask_clear_cpu(cpu_of(busiest), cpus);
6428 if (!cpumask_empty(cpus)) {
6430 env.loop_break = sched_nr_migrate_break;
6438 schedstat_inc(sd, lb_failed[idle]);
6440 * Increment the failure counter only on periodic balance.
6441 * We do not want newidle balance, which can be very
6442 * frequent, pollute the failure counter causing
6443 * excessive cache_hot migrations and active balances.
6445 if (idle != CPU_NEWLY_IDLE)
6446 sd->nr_balance_failed++;
6448 if (need_active_balance(&env)) {
6449 raw_spin_lock_irqsave(&busiest->lock, flags);
6451 /* don't kick the active_load_balance_cpu_stop,
6452 * if the curr task on busiest cpu can't be
6455 if (!cpumask_test_cpu(this_cpu,
6456 tsk_cpus_allowed(busiest->curr))) {
6457 raw_spin_unlock_irqrestore(&busiest->lock,
6459 env.flags |= LBF_ALL_PINNED;
6460 goto out_one_pinned;
6464 * ->active_balance synchronizes accesses to
6465 * ->active_balance_work. Once set, it's cleared
6466 * only after active load balance is finished.
6468 if (!busiest->active_balance) {
6469 busiest->active_balance = 1;
6470 busiest->push_cpu = this_cpu;
6473 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6475 if (active_balance) {
6476 stop_one_cpu_nowait(cpu_of(busiest),
6477 active_load_balance_cpu_stop, busiest,
6478 &busiest->active_balance_work);
6482 * We've kicked active balancing, reset the failure
6485 sd->nr_balance_failed = sd->cache_nice_tries+1;
6488 sd->nr_balance_failed = 0;
6490 if (likely(!active_balance)) {
6491 /* We were unbalanced, so reset the balancing interval */
6492 sd->balance_interval = sd->min_interval;
6495 * If we've begun active balancing, start to back off. This
6496 * case may not be covered by the all_pinned logic if there
6497 * is only 1 task on the busy runqueue (because we don't call
6500 if (sd->balance_interval < sd->max_interval)
6501 sd->balance_interval *= 2;
6507 schedstat_inc(sd, lb_balanced[idle]);
6509 sd->nr_balance_failed = 0;
6512 /* tune up the balancing interval */
6513 if (((env.flags & LBF_ALL_PINNED) &&
6514 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6515 (sd->balance_interval < sd->max_interval))
6516 sd->balance_interval *= 2;
6524 * idle_balance is called by schedule() if this_cpu is about to become
6525 * idle. Attempts to pull tasks from other CPUs.
6527 int idle_balance(struct rq *this_rq)
6529 struct sched_domain *sd;
6530 int pulled_task = 0;
6531 unsigned long next_balance = jiffies + HZ;
6533 int this_cpu = this_rq->cpu;
6535 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6539 * Drop the rq->lock, but keep IRQ/preempt disabled.
6541 raw_spin_unlock(&this_rq->lock);
6543 update_blocked_averages(this_cpu);
6545 for_each_domain(this_cpu, sd) {
6546 unsigned long interval;
6547 int continue_balancing = 1;
6548 u64 t0, domain_cost;
6550 if (!(sd->flags & SD_LOAD_BALANCE))
6553 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6556 if (sd->flags & SD_BALANCE_NEWIDLE) {
6557 t0 = sched_clock_cpu(this_cpu);
6559 /* If we've pulled tasks over stop searching: */
6560 pulled_task = load_balance(this_cpu, this_rq,
6562 &continue_balancing);
6564 domain_cost = sched_clock_cpu(this_cpu) - t0;
6565 if (domain_cost > sd->max_newidle_lb_cost)
6566 sd->max_newidle_lb_cost = domain_cost;
6568 curr_cost += domain_cost;
6571 interval = msecs_to_jiffies(sd->balance_interval);
6572 if (time_after(next_balance, sd->last_balance + interval))
6573 next_balance = sd->last_balance + interval;
6579 raw_spin_lock(&this_rq->lock);
6582 * While browsing the domains, we released the rq lock.
6583 * A task could have be enqueued in the meantime
6585 if (this_rq->nr_running && !pulled_task)
6588 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6590 * We are going idle. next_balance may be set based on
6591 * a busy processor. So reset next_balance.
6593 this_rq->next_balance = next_balance;
6596 if (curr_cost > this_rq->max_idle_balance_cost)
6597 this_rq->max_idle_balance_cost = curr_cost;
6603 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6604 * running tasks off the busiest CPU onto idle CPUs. It requires at
6605 * least 1 task to be running on each physical CPU where possible, and
6606 * avoids physical / logical imbalances.
6608 static int active_load_balance_cpu_stop(void *data)
6610 struct rq *busiest_rq = data;
6611 int busiest_cpu = cpu_of(busiest_rq);
6612 int target_cpu = busiest_rq->push_cpu;
6613 struct rq *target_rq = cpu_rq(target_cpu);
6614 struct sched_domain *sd;
6616 raw_spin_lock_irq(&busiest_rq->lock);
6618 /* make sure the requested cpu hasn't gone down in the meantime */
6619 if (unlikely(busiest_cpu != smp_processor_id() ||
6620 !busiest_rq->active_balance))
6623 /* Is there any task to move? */
6624 if (busiest_rq->nr_running <= 1)
6628 * This condition is "impossible", if it occurs
6629 * we need to fix it. Originally reported by
6630 * Bjorn Helgaas on a 128-cpu setup.
6632 BUG_ON(busiest_rq == target_rq);
6634 /* move a task from busiest_rq to target_rq */
6635 double_lock_balance(busiest_rq, target_rq);
6637 /* Search for an sd spanning us and the target CPU. */
6639 for_each_domain(target_cpu, sd) {
6640 if ((sd->flags & SD_LOAD_BALANCE) &&
6641 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6646 struct lb_env env = {
6648 .dst_cpu = target_cpu,
6649 .dst_rq = target_rq,
6650 .src_cpu = busiest_rq->cpu,
6651 .src_rq = busiest_rq,
6655 schedstat_inc(sd, alb_count);
6657 if (move_one_task(&env))
6658 schedstat_inc(sd, alb_pushed);
6660 schedstat_inc(sd, alb_failed);
6663 double_unlock_balance(busiest_rq, target_rq);
6665 busiest_rq->active_balance = 0;
6666 raw_spin_unlock_irq(&busiest_rq->lock);
6670 #ifdef CONFIG_NO_HZ_COMMON
6672 * idle load balancing details
6673 * - When one of the busy CPUs notice that there may be an idle rebalancing
6674 * needed, they will kick the idle load balancer, which then does idle
6675 * load balancing for all the idle CPUs.
6678 cpumask_var_t idle_cpus_mask;
6680 unsigned long next_balance; /* in jiffy units */
6681 } nohz ____cacheline_aligned;
6683 static inline int find_new_ilb(void)
6685 int ilb = cpumask_first(nohz.idle_cpus_mask);
6687 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6694 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6695 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6696 * CPU (if there is one).
6698 static void nohz_balancer_kick(void)
6702 nohz.next_balance++;
6704 ilb_cpu = find_new_ilb();
6706 if (ilb_cpu >= nr_cpu_ids)
6709 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6712 * Use smp_send_reschedule() instead of resched_cpu().
6713 * This way we generate a sched IPI on the target cpu which
6714 * is idle. And the softirq performing nohz idle load balance
6715 * will be run before returning from the IPI.
6717 smp_send_reschedule(ilb_cpu);
6721 static inline void nohz_balance_exit_idle(int cpu)
6723 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6724 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6725 atomic_dec(&nohz.nr_cpus);
6726 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6730 static inline void set_cpu_sd_state_busy(void)
6732 struct sched_domain *sd;
6733 int cpu = smp_processor_id();
6736 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6738 if (!sd || !sd->nohz_idle)
6742 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6747 void set_cpu_sd_state_idle(void)
6749 struct sched_domain *sd;
6750 int cpu = smp_processor_id();
6753 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6755 if (!sd || sd->nohz_idle)
6759 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6765 * This routine will record that the cpu is going idle with tick stopped.
6766 * This info will be used in performing idle load balancing in the future.
6768 void nohz_balance_enter_idle(int cpu)
6771 * If this cpu is going down, then nothing needs to be done.
6773 if (!cpu_active(cpu))
6776 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6779 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6780 atomic_inc(&nohz.nr_cpus);
6781 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6784 static int sched_ilb_notifier(struct notifier_block *nfb,
6785 unsigned long action, void *hcpu)
6787 switch (action & ~CPU_TASKS_FROZEN) {
6789 nohz_balance_exit_idle(smp_processor_id());
6797 static DEFINE_SPINLOCK(balancing);
6800 * Scale the max load_balance interval with the number of CPUs in the system.
6801 * This trades load-balance latency on larger machines for less cross talk.
6803 void update_max_interval(void)
6805 max_load_balance_interval = HZ*num_online_cpus()/10;
6809 * It checks each scheduling domain to see if it is due to be balanced,
6810 * and initiates a balancing operation if so.
6812 * Balancing parameters are set up in init_sched_domains.
6814 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
6816 int continue_balancing = 1;
6818 unsigned long interval;
6819 struct sched_domain *sd;
6820 /* Earliest time when we have to do rebalance again */
6821 unsigned long next_balance = jiffies + 60*HZ;
6822 int update_next_balance = 0;
6823 int need_serialize, need_decay = 0;
6826 update_blocked_averages(cpu);
6829 for_each_domain(cpu, sd) {
6831 * Decay the newidle max times here because this is a regular
6832 * visit to all the domains. Decay ~1% per second.
6834 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6835 sd->max_newidle_lb_cost =
6836 (sd->max_newidle_lb_cost * 253) / 256;
6837 sd->next_decay_max_lb_cost = jiffies + HZ;
6840 max_cost += sd->max_newidle_lb_cost;
6842 if (!(sd->flags & SD_LOAD_BALANCE))
6846 * Stop the load balance at this level. There is another
6847 * CPU in our sched group which is doing load balancing more
6850 if (!continue_balancing) {
6856 interval = sd->balance_interval;
6857 if (idle != CPU_IDLE)
6858 interval *= sd->busy_factor;
6860 /* scale ms to jiffies */
6861 interval = msecs_to_jiffies(interval);
6862 interval = clamp(interval, 1UL, max_load_balance_interval);
6864 need_serialize = sd->flags & SD_SERIALIZE;
6866 if (need_serialize) {
6867 if (!spin_trylock(&balancing))
6871 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6872 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6874 * The LBF_DST_PINNED logic could have changed
6875 * env->dst_cpu, so we can't know our idle
6876 * state even if we migrated tasks. Update it.
6878 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6880 sd->last_balance = jiffies;
6883 spin_unlock(&balancing);
6885 if (time_after(next_balance, sd->last_balance + interval)) {
6886 next_balance = sd->last_balance + interval;
6887 update_next_balance = 1;
6892 * Ensure the rq-wide value also decays but keep it at a
6893 * reasonable floor to avoid funnies with rq->avg_idle.
6895 rq->max_idle_balance_cost =
6896 max((u64)sysctl_sched_migration_cost, max_cost);
6901 * next_balance will be updated only when there is a need.
6902 * When the cpu is attached to null domain for ex, it will not be
6905 if (likely(update_next_balance))
6906 rq->next_balance = next_balance;
6909 #ifdef CONFIG_NO_HZ_COMMON
6911 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6912 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6914 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
6916 int this_cpu = this_rq->cpu;
6920 if (idle != CPU_IDLE ||
6921 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6924 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6925 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6929 * If this cpu gets work to do, stop the load balancing
6930 * work being done for other cpus. Next load
6931 * balancing owner will pick it up.
6936 rq = cpu_rq(balance_cpu);
6938 raw_spin_lock_irq(&rq->lock);
6939 update_rq_clock(rq);
6940 update_idle_cpu_load(rq);
6941 raw_spin_unlock_irq(&rq->lock);
6943 rebalance_domains(rq, CPU_IDLE);
6945 if (time_after(this_rq->next_balance, rq->next_balance))
6946 this_rq->next_balance = rq->next_balance;
6948 nohz.next_balance = this_rq->next_balance;
6950 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6954 * Current heuristic for kicking the idle load balancer in the presence
6955 * of an idle cpu is the system.
6956 * - This rq has more than one task.
6957 * - At any scheduler domain level, this cpu's scheduler group has multiple
6958 * busy cpu's exceeding the group's power.
6959 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6960 * domain span are idle.
6962 static inline int nohz_kick_needed(struct rq *rq)
6964 unsigned long now = jiffies;
6965 struct sched_domain *sd;
6966 struct sched_group_power *sgp;
6967 int nr_busy, cpu = rq->cpu;
6969 if (unlikely(rq->idle_balance))
6973 * We may be recently in ticked or tickless idle mode. At the first
6974 * busy tick after returning from idle, we will update the busy stats.
6976 set_cpu_sd_state_busy();
6977 nohz_balance_exit_idle(cpu);
6980 * None are in tickless mode and hence no need for NOHZ idle load
6983 if (likely(!atomic_read(&nohz.nr_cpus)))
6986 if (time_before(now, nohz.next_balance))
6989 if (rq->nr_running >= 2)
6993 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6996 sgp = sd->groups->sgp;
6997 nr_busy = atomic_read(&sgp->nr_busy_cpus);
7000 goto need_kick_unlock;
7003 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7005 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7006 sched_domain_span(sd)) < cpu))
7007 goto need_kick_unlock;
7018 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7022 * run_rebalance_domains is triggered when needed from the scheduler tick.
7023 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7025 static void run_rebalance_domains(struct softirq_action *h)
7027 struct rq *this_rq = this_rq();
7028 enum cpu_idle_type idle = this_rq->idle_balance ?
7029 CPU_IDLE : CPU_NOT_IDLE;
7031 rebalance_domains(this_rq, idle);
7034 * If this cpu has a pending nohz_balance_kick, then do the
7035 * balancing on behalf of the other idle cpus whose ticks are
7038 nohz_idle_balance(this_rq, idle);
7041 static inline int on_null_domain(struct rq *rq)
7043 return !rcu_dereference_sched(rq->sd);
7047 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7049 void trigger_load_balance(struct rq *rq)
7051 /* Don't need to rebalance while attached to NULL domain */
7052 if (unlikely(on_null_domain(rq)))
7055 if (time_after_eq(jiffies, rq->next_balance))
7056 raise_softirq(SCHED_SOFTIRQ);
7057 #ifdef CONFIG_NO_HZ_COMMON
7058 if (nohz_kick_needed(rq))
7059 nohz_balancer_kick();
7063 static void rq_online_fair(struct rq *rq)
7068 static void rq_offline_fair(struct rq *rq)
7072 /* Ensure any throttled groups are reachable by pick_next_task */
7073 unthrottle_offline_cfs_rqs(rq);
7076 #endif /* CONFIG_SMP */
7079 * scheduler tick hitting a task of our scheduling class:
7081 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7083 struct cfs_rq *cfs_rq;
7084 struct sched_entity *se = &curr->se;
7086 for_each_sched_entity(se) {
7087 cfs_rq = cfs_rq_of(se);
7088 entity_tick(cfs_rq, se, queued);
7091 if (numabalancing_enabled)
7092 task_tick_numa(rq, curr);
7094 update_rq_runnable_avg(rq, 1);
7098 * called on fork with the child task as argument from the parent's context
7099 * - child not yet on the tasklist
7100 * - preemption disabled
7102 static void task_fork_fair(struct task_struct *p)
7104 struct cfs_rq *cfs_rq;
7105 struct sched_entity *se = &p->se, *curr;
7106 int this_cpu = smp_processor_id();
7107 struct rq *rq = this_rq();
7108 unsigned long flags;
7110 raw_spin_lock_irqsave(&rq->lock, flags);
7112 update_rq_clock(rq);
7114 cfs_rq = task_cfs_rq(current);
7115 curr = cfs_rq->curr;
7118 * Not only the cpu but also the task_group of the parent might have
7119 * been changed after parent->se.parent,cfs_rq were copied to
7120 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7121 * of child point to valid ones.
7124 __set_task_cpu(p, this_cpu);
7127 update_curr(cfs_rq);
7130 se->vruntime = curr->vruntime;
7131 place_entity(cfs_rq, se, 1);
7133 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7135 * Upon rescheduling, sched_class::put_prev_task() will place
7136 * 'current' within the tree based on its new key value.
7138 swap(curr->vruntime, se->vruntime);
7139 resched_task(rq->curr);
7142 se->vruntime -= cfs_rq->min_vruntime;
7144 raw_spin_unlock_irqrestore(&rq->lock, flags);
7148 * Priority of the task has changed. Check to see if we preempt
7152 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7158 * Reschedule if we are currently running on this runqueue and
7159 * our priority decreased, or if we are not currently running on
7160 * this runqueue and our priority is higher than the current's
7162 if (rq->curr == p) {
7163 if (p->prio > oldprio)
7164 resched_task(rq->curr);
7166 check_preempt_curr(rq, p, 0);
7169 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7171 struct sched_entity *se = &p->se;
7172 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7175 * Ensure the task's vruntime is normalized, so that when its
7176 * switched back to the fair class the enqueue_entity(.flags=0) will
7177 * do the right thing.
7179 * If it was on_rq, then the dequeue_entity(.flags=0) will already
7180 * have normalized the vruntime, if it was !on_rq, then only when
7181 * the task is sleeping will it still have non-normalized vruntime.
7183 if (!se->on_rq && p->state != TASK_RUNNING) {
7185 * Fix up our vruntime so that the current sleep doesn't
7186 * cause 'unlimited' sleep bonus.
7188 place_entity(cfs_rq, se, 0);
7189 se->vruntime -= cfs_rq->min_vruntime;
7194 * Remove our load from contribution when we leave sched_fair
7195 * and ensure we don't carry in an old decay_count if we
7198 if (se->avg.decay_count) {
7199 __synchronize_entity_decay(se);
7200 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7206 * We switched to the sched_fair class.
7208 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7214 * We were most likely switched from sched_rt, so
7215 * kick off the schedule if running, otherwise just see
7216 * if we can still preempt the current task.
7219 resched_task(rq->curr);
7221 check_preempt_curr(rq, p, 0);
7224 /* Account for a task changing its policy or group.
7226 * This routine is mostly called to set cfs_rq->curr field when a task
7227 * migrates between groups/classes.
7229 static void set_curr_task_fair(struct rq *rq)
7231 struct sched_entity *se = &rq->curr->se;
7233 for_each_sched_entity(se) {
7234 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7236 set_next_entity(cfs_rq, se);
7237 /* ensure bandwidth has been allocated on our new cfs_rq */
7238 account_cfs_rq_runtime(cfs_rq, 0);
7242 void init_cfs_rq(struct cfs_rq *cfs_rq)
7244 cfs_rq->tasks_timeline = RB_ROOT;
7245 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7246 #ifndef CONFIG_64BIT
7247 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7250 atomic64_set(&cfs_rq->decay_counter, 1);
7251 atomic_long_set(&cfs_rq->removed_load, 0);
7255 #ifdef CONFIG_FAIR_GROUP_SCHED
7256 static void task_move_group_fair(struct task_struct *p, int on_rq)
7258 struct sched_entity *se = &p->se;
7259 struct cfs_rq *cfs_rq;
7262 * If the task was not on the rq at the time of this cgroup movement
7263 * it must have been asleep, sleeping tasks keep their ->vruntime
7264 * absolute on their old rq until wakeup (needed for the fair sleeper
7265 * bonus in place_entity()).
7267 * If it was on the rq, we've just 'preempted' it, which does convert
7268 * ->vruntime to a relative base.
7270 * Make sure both cases convert their relative position when migrating
7271 * to another cgroup's rq. This does somewhat interfere with the
7272 * fair sleeper stuff for the first placement, but who cares.
7275 * When !on_rq, vruntime of the task has usually NOT been normalized.
7276 * But there are some cases where it has already been normalized:
7278 * - Moving a forked child which is waiting for being woken up by
7279 * wake_up_new_task().
7280 * - Moving a task which has been woken up by try_to_wake_up() and
7281 * waiting for actually being woken up by sched_ttwu_pending().
7283 * To prevent boost or penalty in the new cfs_rq caused by delta
7284 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7286 if (!on_rq && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7290 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7291 set_task_rq(p, task_cpu(p));
7292 se->depth = se->parent ? se->parent->depth + 1 : 0;
7294 cfs_rq = cfs_rq_of(se);
7295 se->vruntime += cfs_rq->min_vruntime;
7298 * migrate_task_rq_fair() will have removed our previous
7299 * contribution, but we must synchronize for ongoing future
7302 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7303 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7308 void free_fair_sched_group(struct task_group *tg)
7312 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7314 for_each_possible_cpu(i) {
7316 kfree(tg->cfs_rq[i]);
7325 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7327 struct cfs_rq *cfs_rq;
7328 struct sched_entity *se;
7331 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7334 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7338 tg->shares = NICE_0_LOAD;
7340 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7342 for_each_possible_cpu(i) {
7343 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7344 GFP_KERNEL, cpu_to_node(i));
7348 se = kzalloc_node(sizeof(struct sched_entity),
7349 GFP_KERNEL, cpu_to_node(i));
7353 init_cfs_rq(cfs_rq);
7354 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7365 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7367 struct rq *rq = cpu_rq(cpu);
7368 unsigned long flags;
7371 * Only empty task groups can be destroyed; so we can speculatively
7372 * check on_list without danger of it being re-added.
7374 if (!tg->cfs_rq[cpu]->on_list)
7377 raw_spin_lock_irqsave(&rq->lock, flags);
7378 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7379 raw_spin_unlock_irqrestore(&rq->lock, flags);
7382 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7383 struct sched_entity *se, int cpu,
7384 struct sched_entity *parent)
7386 struct rq *rq = cpu_rq(cpu);
7390 init_cfs_rq_runtime(cfs_rq);
7392 tg->cfs_rq[cpu] = cfs_rq;
7395 /* se could be NULL for root_task_group */
7400 se->cfs_rq = &rq->cfs;
7403 se->cfs_rq = parent->my_q;
7404 se->depth = parent->depth + 1;
7408 /* guarantee group entities always have weight */
7409 update_load_set(&se->load, NICE_0_LOAD);
7410 se->parent = parent;
7413 static DEFINE_MUTEX(shares_mutex);
7415 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7418 unsigned long flags;
7421 * We can't change the weight of the root cgroup.
7426 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7428 mutex_lock(&shares_mutex);
7429 if (tg->shares == shares)
7432 tg->shares = shares;
7433 for_each_possible_cpu(i) {
7434 struct rq *rq = cpu_rq(i);
7435 struct sched_entity *se;
7438 /* Propagate contribution to hierarchy */
7439 raw_spin_lock_irqsave(&rq->lock, flags);
7441 /* Possible calls to update_curr() need rq clock */
7442 update_rq_clock(rq);
7443 for_each_sched_entity(se)
7444 update_cfs_shares(group_cfs_rq(se));
7445 raw_spin_unlock_irqrestore(&rq->lock, flags);
7449 mutex_unlock(&shares_mutex);
7452 #else /* CONFIG_FAIR_GROUP_SCHED */
7454 void free_fair_sched_group(struct task_group *tg) { }
7456 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7461 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7463 #endif /* CONFIG_FAIR_GROUP_SCHED */
7466 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7468 struct sched_entity *se = &task->se;
7469 unsigned int rr_interval = 0;
7472 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7475 if (rq->cfs.load.weight)
7476 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7482 * All the scheduling class methods:
7484 const struct sched_class fair_sched_class = {
7485 .next = &idle_sched_class,
7486 .enqueue_task = enqueue_task_fair,
7487 .dequeue_task = dequeue_task_fair,
7488 .yield_task = yield_task_fair,
7489 .yield_to_task = yield_to_task_fair,
7491 .check_preempt_curr = check_preempt_wakeup,
7493 .pick_next_task = pick_next_task_fair,
7494 .put_prev_task = put_prev_task_fair,
7497 .select_task_rq = select_task_rq_fair,
7498 .migrate_task_rq = migrate_task_rq_fair,
7500 .rq_online = rq_online_fair,
7501 .rq_offline = rq_offline_fair,
7503 .task_waking = task_waking_fair,
7506 .set_curr_task = set_curr_task_fair,
7507 .task_tick = task_tick_fair,
7508 .task_fork = task_fork_fair,
7510 .prio_changed = prio_changed_fair,
7511 .switched_from = switched_from_fair,
7512 .switched_to = switched_to_fair,
7514 .get_rr_interval = get_rr_interval_fair,
7516 #ifdef CONFIG_FAIR_GROUP_SCHED
7517 .task_move_group = task_move_group_fair,
7521 #ifdef CONFIG_SCHED_DEBUG
7522 void print_cfs_stats(struct seq_file *m, int cpu)
7524 struct cfs_rq *cfs_rq;
7527 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7528 print_cfs_rq(m, cpu, cfs_rq);
7533 __init void init_sched_fair_class(void)
7536 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7538 #ifdef CONFIG_NO_HZ_COMMON
7539 nohz.next_balance = jiffies;
7540 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7541 cpu_notifier(sched_ilb_notifier, 0);