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
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency = 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity = 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency = 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static unsigned int get_update_sysctl_factor(void)
146 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling) {
150 case SCHED_TUNABLESCALING_NONE:
153 case SCHED_TUNABLESCALING_LINEAR:
156 case SCHED_TUNABLESCALING_LOG:
158 factor = 1 + ilog2(cpus);
165 static void update_sysctl(void)
167 unsigned int factor = get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity);
172 SET_SYSCTL(sched_latency);
173 SET_SYSCTL(sched_wakeup_granularity);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight *lw)
189 if (likely(lw->inv_weight))
192 w = scale_load_down(lw->weight);
194 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
196 else if (unlikely(!w))
197 lw->inv_weight = WMULT_CONST;
199 lw->inv_weight = WMULT_CONST / w;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
216 u64 fact = scale_load_down(weight);
217 int shift = WMULT_SHIFT;
219 __update_inv_weight(lw);
221 if (unlikely(fact >> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact = (u64)(u32)fact * lw->inv_weight;
236 return mul_u64_u32_shr(delta_exec, fact, shift);
240 const struct sched_class fair_sched_class;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct *task_of(struct sched_entity *se)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se));
262 return container_of(se, struct task_struct, se);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
310 if (cfs_rq->on_list) {
311 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq *
322 is_same_group(struct sched_entity *se, struct sched_entity *pse)
324 if (se->cfs_rq == pse->cfs_rq)
330 static inline struct sched_entity *parent_entity(struct sched_entity *se)
336 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
338 int se_depth, pse_depth;
341 * preemption test can be made between sibling entities who are in the
342 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343 * both tasks until we find their ancestors who are siblings of common
347 /* First walk up until both entities are at same depth */
348 se_depth = (*se)->depth;
349 pse_depth = (*pse)->depth;
351 while (se_depth > pse_depth) {
353 *se = parent_entity(*se);
356 while (pse_depth > se_depth) {
358 *pse = parent_entity(*pse);
361 while (!is_same_group(*se, *pse)) {
362 *se = parent_entity(*se);
363 *pse = parent_entity(*pse);
367 #else /* !CONFIG_FAIR_GROUP_SCHED */
369 static inline struct task_struct *task_of(struct sched_entity *se)
371 return container_of(se, struct task_struct, se);
374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
376 return container_of(cfs_rq, struct rq, cfs);
379 #define entity_is_task(se) 1
381 #define for_each_sched_entity(se) \
382 for (; se; se = NULL)
384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
386 return &task_rq(p)->cfs;
389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
391 struct task_struct *p = task_of(se);
392 struct rq *rq = task_rq(p);
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
414 static inline struct sched_entity *parent_entity(struct sched_entity *se)
420 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
424 #endif /* CONFIG_FAIR_GROUP_SCHED */
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
429 /**************************************************************
430 * Scheduling class tree data structure manipulation methods:
433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
435 s64 delta = (s64)(vruntime - max_vruntime);
437 max_vruntime = vruntime;
442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
444 s64 delta = (s64)(vruntime - min_vruntime);
446 min_vruntime = vruntime;
451 static inline int entity_before(struct sched_entity *a,
452 struct sched_entity *b)
454 return (s64)(a->vruntime - b->vruntime) < 0;
457 static void update_min_vruntime(struct cfs_rq *cfs_rq)
459 u64 vruntime = cfs_rq->min_vruntime;
462 vruntime = cfs_rq->curr->vruntime;
464 if (cfs_rq->rb_leftmost) {
465 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
470 vruntime = se->vruntime;
472 vruntime = min_vruntime(vruntime, se->vruntime);
475 /* ensure we never gain time by being placed backwards. */
476 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
479 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484 * Enqueue an entity into the rb-tree:
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
488 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489 struct rb_node *parent = NULL;
490 struct sched_entity *entry;
494 * Find the right place in the rbtree:
498 entry = rb_entry(parent, struct sched_entity, run_node);
500 * We dont care about collisions. Nodes with
501 * the same key stay together.
503 if (entity_before(se, entry)) {
504 link = &parent->rb_left;
506 link = &parent->rb_right;
512 * Maintain a cache of leftmost tree entries (it is frequently
516 cfs_rq->rb_leftmost = &se->run_node;
518 rb_link_node(&se->run_node, parent, link);
519 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
524 if (cfs_rq->rb_leftmost == &se->run_node) {
525 struct rb_node *next_node;
527 next_node = rb_next(&se->run_node);
528 cfs_rq->rb_leftmost = next_node;
531 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
536 struct rb_node *left = cfs_rq->rb_leftmost;
541 return rb_entry(left, struct sched_entity, run_node);
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
546 struct rb_node *next = rb_next(&se->run_node);
551 return rb_entry(next, struct sched_entity, run_node);
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
557 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
562 return rb_entry(last, struct sched_entity, run_node);
565 /**************************************************************
566 * Scheduling class statistics methods:
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570 void __user *buffer, size_t *lenp,
573 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574 unsigned int factor = get_update_sysctl_factor();
579 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580 sysctl_sched_min_granularity);
582 #define WRT_SYSCTL(name) \
583 (normalized_sysctl_##name = sysctl_##name / (factor))
584 WRT_SYSCTL(sched_min_granularity);
585 WRT_SYSCTL(sched_latency);
586 WRT_SYSCTL(sched_wakeup_granularity);
596 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
598 if (unlikely(se->load.weight != NICE_0_LOAD))
599 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
605 * The idea is to set a period in which each task runs once.
607 * When there are too many tasks (sched_nr_latency) we have to stretch
608 * this period because otherwise the slices get too small.
610 * p = (nr <= nl) ? l : l*nr/nl
612 static u64 __sched_period(unsigned long nr_running)
614 if (unlikely(nr_running > sched_nr_latency))
615 return nr_running * sysctl_sched_min_granularity;
617 return sysctl_sched_latency;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
630 for_each_sched_entity(se) {
631 struct load_weight *load;
632 struct load_weight lw;
634 cfs_rq = cfs_rq_of(se);
635 load = &cfs_rq->load;
637 if (unlikely(!se->on_rq)) {
640 update_load_add(&lw, se->load.weight);
643 slice = __calc_delta(slice, se->load.weight, load);
649 * We calculate the vruntime slice of a to-be-inserted task.
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
655 return calc_delta_fair(sched_slice(cfs_rq, se), se);
659 static int select_idle_sibling(struct task_struct *p, int cpu);
660 static unsigned long task_h_load(struct task_struct *p);
663 * We choose a half-life close to 1 scheduling period.
664 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
665 * dependent on this value.
667 #define LOAD_AVG_PERIOD 32
668 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
669 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
671 /* Give new sched_entity start runnable values to heavy its load in infant time */
672 void init_entity_runnable_average(struct sched_entity *se)
674 struct sched_avg *sa = &se->avg;
676 sa->last_update_time = 0;
678 * sched_avg's period_contrib should be strictly less then 1024, so
679 * we give it 1023 to make sure it is almost a period (1024us), and
680 * will definitely be update (after enqueue).
682 sa->period_contrib = 1023;
683 sa->load_avg = scale_load_down(se->load.weight);
684 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
685 sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
686 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
687 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
691 void init_entity_runnable_average(struct sched_entity *se)
697 * Update the current task's runtime statistics.
699 static void update_curr(struct cfs_rq *cfs_rq)
701 struct sched_entity *curr = cfs_rq->curr;
702 u64 now = rq_clock_task(rq_of(cfs_rq));
708 delta_exec = now - curr->exec_start;
709 if (unlikely((s64)delta_exec <= 0))
712 curr->exec_start = now;
714 schedstat_set(curr->statistics.exec_max,
715 max(delta_exec, curr->statistics.exec_max));
717 curr->sum_exec_runtime += delta_exec;
718 schedstat_add(cfs_rq, exec_clock, delta_exec);
720 curr->vruntime += calc_delta_fair(delta_exec, curr);
721 update_min_vruntime(cfs_rq);
723 if (entity_is_task(curr)) {
724 struct task_struct *curtask = task_of(curr);
726 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
727 cpuacct_charge(curtask, delta_exec);
728 account_group_exec_runtime(curtask, delta_exec);
731 account_cfs_rq_runtime(cfs_rq, delta_exec);
734 static void update_curr_fair(struct rq *rq)
736 update_curr(cfs_rq_of(&rq->curr->se));
740 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
742 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
746 * Task is being enqueued - update stats:
748 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
751 * Are we enqueueing a waiting task? (for current tasks
752 * a dequeue/enqueue event is a NOP)
754 if (se != cfs_rq->curr)
755 update_stats_wait_start(cfs_rq, se);
759 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
761 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
762 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
763 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
764 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
765 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
766 #ifdef CONFIG_SCHEDSTATS
767 if (entity_is_task(se)) {
768 trace_sched_stat_wait(task_of(se),
769 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
772 schedstat_set(se->statistics.wait_start, 0);
776 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
779 * Mark the end of the wait period if dequeueing a
782 if (se != cfs_rq->curr)
783 update_stats_wait_end(cfs_rq, se);
787 * We are picking a new current task - update its stats:
790 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
793 * We are starting a new run period:
795 se->exec_start = rq_clock_task(rq_of(cfs_rq));
798 /**************************************************
799 * Scheduling class queueing methods:
802 #ifdef CONFIG_NUMA_BALANCING
804 * Approximate time to scan a full NUMA task in ms. The task scan period is
805 * calculated based on the tasks virtual memory size and
806 * numa_balancing_scan_size.
808 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
809 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
811 /* Portion of address space to scan in MB */
812 unsigned int sysctl_numa_balancing_scan_size = 256;
814 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
815 unsigned int sysctl_numa_balancing_scan_delay = 1000;
817 static unsigned int task_nr_scan_windows(struct task_struct *p)
819 unsigned long rss = 0;
820 unsigned long nr_scan_pages;
823 * Calculations based on RSS as non-present and empty pages are skipped
824 * by the PTE scanner and NUMA hinting faults should be trapped based
827 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
828 rss = get_mm_rss(p->mm);
832 rss = round_up(rss, nr_scan_pages);
833 return rss / nr_scan_pages;
836 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
837 #define MAX_SCAN_WINDOW 2560
839 static unsigned int task_scan_min(struct task_struct *p)
841 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
842 unsigned int scan, floor;
843 unsigned int windows = 1;
845 if (scan_size < MAX_SCAN_WINDOW)
846 windows = MAX_SCAN_WINDOW / scan_size;
847 floor = 1000 / windows;
849 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
850 return max_t(unsigned int, floor, scan);
853 static unsigned int task_scan_max(struct task_struct *p)
855 unsigned int smin = task_scan_min(p);
858 /* Watch for min being lower than max due to floor calculations */
859 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
860 return max(smin, smax);
863 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
865 rq->nr_numa_running += (p->numa_preferred_nid != -1);
866 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
869 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
871 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
872 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
878 spinlock_t lock; /* nr_tasks, tasks */
883 nodemask_t active_nodes;
884 unsigned long total_faults;
886 * Faults_cpu is used to decide whether memory should move
887 * towards the CPU. As a consequence, these stats are weighted
888 * more by CPU use than by memory faults.
890 unsigned long *faults_cpu;
891 unsigned long faults[0];
894 /* Shared or private faults. */
895 #define NR_NUMA_HINT_FAULT_TYPES 2
897 /* Memory and CPU locality */
898 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
900 /* Averaged statistics, and temporary buffers. */
901 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
903 pid_t task_numa_group_id(struct task_struct *p)
905 return p->numa_group ? p->numa_group->gid : 0;
909 * The averaged statistics, shared & private, memory & cpu,
910 * occupy the first half of the array. The second half of the
911 * array is for current counters, which are averaged into the
912 * first set by task_numa_placement.
914 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
916 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
919 static inline unsigned long task_faults(struct task_struct *p, int nid)
924 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
925 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
928 static inline unsigned long group_faults(struct task_struct *p, int nid)
933 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
934 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
937 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
939 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
940 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
943 /* Handle placement on systems where not all nodes are directly connected. */
944 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
945 int maxdist, bool task)
947 unsigned long score = 0;
951 * All nodes are directly connected, and the same distance
952 * from each other. No need for fancy placement algorithms.
954 if (sched_numa_topology_type == NUMA_DIRECT)
958 * This code is called for each node, introducing N^2 complexity,
959 * which should be ok given the number of nodes rarely exceeds 8.
961 for_each_online_node(node) {
962 unsigned long faults;
963 int dist = node_distance(nid, node);
966 * The furthest away nodes in the system are not interesting
967 * for placement; nid was already counted.
969 if (dist == sched_max_numa_distance || node == nid)
973 * On systems with a backplane NUMA topology, compare groups
974 * of nodes, and move tasks towards the group with the most
975 * memory accesses. When comparing two nodes at distance
976 * "hoplimit", only nodes closer by than "hoplimit" are part
977 * of each group. Skip other nodes.
979 if (sched_numa_topology_type == NUMA_BACKPLANE &&
983 /* Add up the faults from nearby nodes. */
985 faults = task_faults(p, node);
987 faults = group_faults(p, node);
990 * On systems with a glueless mesh NUMA topology, there are
991 * no fixed "groups of nodes". Instead, nodes that are not
992 * directly connected bounce traffic through intermediate
993 * nodes; a numa_group can occupy any set of nodes.
994 * The further away a node is, the less the faults count.
995 * This seems to result in good task placement.
997 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
998 faults *= (sched_max_numa_distance - dist);
999 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1009 * These return the fraction of accesses done by a particular task, or
1010 * task group, on a particular numa node. The group weight is given a
1011 * larger multiplier, in order to group tasks together that are almost
1012 * evenly spread out between numa nodes.
1014 static inline unsigned long task_weight(struct task_struct *p, int nid,
1017 unsigned long faults, total_faults;
1019 if (!p->numa_faults)
1022 total_faults = p->total_numa_faults;
1027 faults = task_faults(p, nid);
1028 faults += score_nearby_nodes(p, nid, dist, true);
1030 return 1000 * faults / total_faults;
1033 static inline unsigned long group_weight(struct task_struct *p, int nid,
1036 unsigned long faults, total_faults;
1041 total_faults = p->numa_group->total_faults;
1046 faults = group_faults(p, nid);
1047 faults += score_nearby_nodes(p, nid, dist, false);
1049 return 1000 * faults / total_faults;
1052 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1053 int src_nid, int dst_cpu)
1055 struct numa_group *ng = p->numa_group;
1056 int dst_nid = cpu_to_node(dst_cpu);
1057 int last_cpupid, this_cpupid;
1059 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1062 * Multi-stage node selection is used in conjunction with a periodic
1063 * migration fault to build a temporal task<->page relation. By using
1064 * a two-stage filter we remove short/unlikely relations.
1066 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1067 * a task's usage of a particular page (n_p) per total usage of this
1068 * page (n_t) (in a given time-span) to a probability.
1070 * Our periodic faults will sample this probability and getting the
1071 * same result twice in a row, given these samples are fully
1072 * independent, is then given by P(n)^2, provided our sample period
1073 * is sufficiently short compared to the usage pattern.
1075 * This quadric squishes small probabilities, making it less likely we
1076 * act on an unlikely task<->page relation.
1078 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1079 if (!cpupid_pid_unset(last_cpupid) &&
1080 cpupid_to_nid(last_cpupid) != dst_nid)
1083 /* Always allow migrate on private faults */
1084 if (cpupid_match_pid(p, last_cpupid))
1087 /* A shared fault, but p->numa_group has not been set up yet. */
1092 * Do not migrate if the destination is not a node that
1093 * is actively used by this numa group.
1095 if (!node_isset(dst_nid, ng->active_nodes))
1099 * Source is a node that is not actively used by this
1100 * numa group, while the destination is. Migrate.
1102 if (!node_isset(src_nid, ng->active_nodes))
1106 * Both source and destination are nodes in active
1107 * use by this numa group. Maximize memory bandwidth
1108 * by migrating from more heavily used groups, to less
1109 * heavily used ones, spreading the load around.
1110 * Use a 1/4 hysteresis to avoid spurious page movement.
1112 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1115 static unsigned long weighted_cpuload(const int cpu);
1116 static unsigned long source_load(int cpu, int type);
1117 static unsigned long target_load(int cpu, int type);
1118 static unsigned long capacity_of(int cpu);
1119 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1121 /* Cached statistics for all CPUs within a node */
1123 unsigned long nr_running;
1126 /* Total compute capacity of CPUs on a node */
1127 unsigned long compute_capacity;
1129 /* Approximate capacity in terms of runnable tasks on a node */
1130 unsigned long task_capacity;
1131 int has_free_capacity;
1135 * XXX borrowed from update_sg_lb_stats
1137 static void update_numa_stats(struct numa_stats *ns, int nid)
1139 int smt, cpu, cpus = 0;
1140 unsigned long capacity;
1142 memset(ns, 0, sizeof(*ns));
1143 for_each_cpu(cpu, cpumask_of_node(nid)) {
1144 struct rq *rq = cpu_rq(cpu);
1146 ns->nr_running += rq->nr_running;
1147 ns->load += weighted_cpuload(cpu);
1148 ns->compute_capacity += capacity_of(cpu);
1154 * If we raced with hotplug and there are no CPUs left in our mask
1155 * the @ns structure is NULL'ed and task_numa_compare() will
1156 * not find this node attractive.
1158 * We'll either bail at !has_free_capacity, or we'll detect a huge
1159 * imbalance and bail there.
1164 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1165 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1166 capacity = cpus / smt; /* cores */
1168 ns->task_capacity = min_t(unsigned, capacity,
1169 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1170 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1173 struct task_numa_env {
1174 struct task_struct *p;
1176 int src_cpu, src_nid;
1177 int dst_cpu, dst_nid;
1179 struct numa_stats src_stats, dst_stats;
1184 struct task_struct *best_task;
1189 static void task_numa_assign(struct task_numa_env *env,
1190 struct task_struct *p, long imp)
1193 put_task_struct(env->best_task);
1196 env->best_imp = imp;
1197 env->best_cpu = env->dst_cpu;
1200 static bool load_too_imbalanced(long src_load, long dst_load,
1201 struct task_numa_env *env)
1204 long orig_src_load, orig_dst_load;
1205 long src_capacity, dst_capacity;
1208 * The load is corrected for the CPU capacity available on each node.
1211 * ------------ vs ---------
1212 * src_capacity dst_capacity
1214 src_capacity = env->src_stats.compute_capacity;
1215 dst_capacity = env->dst_stats.compute_capacity;
1217 /* We care about the slope of the imbalance, not the direction. */
1218 if (dst_load < src_load)
1219 swap(dst_load, src_load);
1221 /* Is the difference below the threshold? */
1222 imb = dst_load * src_capacity * 100 -
1223 src_load * dst_capacity * env->imbalance_pct;
1228 * The imbalance is above the allowed threshold.
1229 * Compare it with the old imbalance.
1231 orig_src_load = env->src_stats.load;
1232 orig_dst_load = env->dst_stats.load;
1234 if (orig_dst_load < orig_src_load)
1235 swap(orig_dst_load, orig_src_load);
1237 old_imb = orig_dst_load * src_capacity * 100 -
1238 orig_src_load * dst_capacity * env->imbalance_pct;
1240 /* Would this change make things worse? */
1241 return (imb > old_imb);
1245 * This checks if the overall compute and NUMA accesses of the system would
1246 * be improved if the source tasks was migrated to the target dst_cpu taking
1247 * into account that it might be best if task running on the dst_cpu should
1248 * be exchanged with the source task
1250 static void task_numa_compare(struct task_numa_env *env,
1251 long taskimp, long groupimp)
1253 struct rq *src_rq = cpu_rq(env->src_cpu);
1254 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1255 struct task_struct *cur;
1256 long src_load, dst_load;
1258 long imp = env->p->numa_group ? groupimp : taskimp;
1260 int dist = env->dist;
1261 bool assigned = false;
1265 raw_spin_lock_irq(&dst_rq->lock);
1268 * No need to move the exiting task or idle task.
1270 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1274 * The task_struct must be protected here to protect the
1275 * p->numa_faults access in the task_weight since the
1276 * numa_faults could already be freed in the following path:
1277 * finish_task_switch()
1278 * --> put_task_struct()
1279 * --> __put_task_struct()
1280 * --> task_numa_free()
1282 get_task_struct(cur);
1285 raw_spin_unlock_irq(&dst_rq->lock);
1288 * Because we have preemption enabled we can get migrated around and
1289 * end try selecting ourselves (current == env->p) as a swap candidate.
1295 * "imp" is the fault differential for the source task between the
1296 * source and destination node. Calculate the total differential for
1297 * the source task and potential destination task. The more negative
1298 * the value is, the more rmeote accesses that would be expected to
1299 * be incurred if the tasks were swapped.
1302 /* Skip this swap candidate if cannot move to the source cpu */
1303 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1307 * If dst and source tasks are in the same NUMA group, or not
1308 * in any group then look only at task weights.
1310 if (cur->numa_group == env->p->numa_group) {
1311 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1312 task_weight(cur, env->dst_nid, dist);
1314 * Add some hysteresis to prevent swapping the
1315 * tasks within a group over tiny differences.
1317 if (cur->numa_group)
1321 * Compare the group weights. If a task is all by
1322 * itself (not part of a group), use the task weight
1325 if (cur->numa_group)
1326 imp += group_weight(cur, env->src_nid, dist) -
1327 group_weight(cur, env->dst_nid, dist);
1329 imp += task_weight(cur, env->src_nid, dist) -
1330 task_weight(cur, env->dst_nid, dist);
1334 if (imp <= env->best_imp && moveimp <= env->best_imp)
1338 /* Is there capacity at our destination? */
1339 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1340 !env->dst_stats.has_free_capacity)
1346 /* Balance doesn't matter much if we're running a task per cpu */
1347 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1348 dst_rq->nr_running == 1)
1352 * In the overloaded case, try and keep the load balanced.
1355 load = task_h_load(env->p);
1356 dst_load = env->dst_stats.load + load;
1357 src_load = env->src_stats.load - load;
1359 if (moveimp > imp && moveimp > env->best_imp) {
1361 * If the improvement from just moving env->p direction is
1362 * better than swapping tasks around, check if a move is
1363 * possible. Store a slightly smaller score than moveimp,
1364 * so an actually idle CPU will win.
1366 if (!load_too_imbalanced(src_load, dst_load, env)) {
1368 put_task_struct(cur);
1374 if (imp <= env->best_imp)
1378 load = task_h_load(cur);
1383 if (load_too_imbalanced(src_load, dst_load, env))
1387 * One idle CPU per node is evaluated for a task numa move.
1388 * Call select_idle_sibling to maybe find a better one.
1391 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1395 task_numa_assign(env, cur, imp);
1399 * The dst_rq->curr isn't assigned. The protection for task_struct is
1402 if (cur && !assigned)
1403 put_task_struct(cur);
1406 static void task_numa_find_cpu(struct task_numa_env *env,
1407 long taskimp, long groupimp)
1411 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1412 /* Skip this CPU if the source task cannot migrate */
1413 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1417 task_numa_compare(env, taskimp, groupimp);
1421 /* Only move tasks to a NUMA node less busy than the current node. */
1422 static bool numa_has_capacity(struct task_numa_env *env)
1424 struct numa_stats *src = &env->src_stats;
1425 struct numa_stats *dst = &env->dst_stats;
1427 if (src->has_free_capacity && !dst->has_free_capacity)
1431 * Only consider a task move if the source has a higher load
1432 * than the destination, corrected for CPU capacity on each node.
1434 * src->load dst->load
1435 * --------------------- vs ---------------------
1436 * src->compute_capacity dst->compute_capacity
1438 if (src->load * dst->compute_capacity * env->imbalance_pct >
1440 dst->load * src->compute_capacity * 100)
1446 static int task_numa_migrate(struct task_struct *p)
1448 struct task_numa_env env = {
1451 .src_cpu = task_cpu(p),
1452 .src_nid = task_node(p),
1454 .imbalance_pct = 112,
1460 struct sched_domain *sd;
1461 unsigned long taskweight, groupweight;
1463 long taskimp, groupimp;
1466 * Pick the lowest SD_NUMA domain, as that would have the smallest
1467 * imbalance and would be the first to start moving tasks about.
1469 * And we want to avoid any moving of tasks about, as that would create
1470 * random movement of tasks -- counter the numa conditions we're trying
1474 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1476 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1480 * Cpusets can break the scheduler domain tree into smaller
1481 * balance domains, some of which do not cross NUMA boundaries.
1482 * Tasks that are "trapped" in such domains cannot be migrated
1483 * elsewhere, so there is no point in (re)trying.
1485 if (unlikely(!sd)) {
1486 p->numa_preferred_nid = task_node(p);
1490 env.dst_nid = p->numa_preferred_nid;
1491 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1492 taskweight = task_weight(p, env.src_nid, dist);
1493 groupweight = group_weight(p, env.src_nid, dist);
1494 update_numa_stats(&env.src_stats, env.src_nid);
1495 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1496 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1497 update_numa_stats(&env.dst_stats, env.dst_nid);
1499 /* Try to find a spot on the preferred nid. */
1500 if (numa_has_capacity(&env))
1501 task_numa_find_cpu(&env, taskimp, groupimp);
1504 * Look at other nodes in these cases:
1505 * - there is no space available on the preferred_nid
1506 * - the task is part of a numa_group that is interleaved across
1507 * multiple NUMA nodes; in order to better consolidate the group,
1508 * we need to check other locations.
1510 if (env.best_cpu == -1 || (p->numa_group &&
1511 nodes_weight(p->numa_group->active_nodes) > 1)) {
1512 for_each_online_node(nid) {
1513 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1516 dist = node_distance(env.src_nid, env.dst_nid);
1517 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1519 taskweight = task_weight(p, env.src_nid, dist);
1520 groupweight = group_weight(p, env.src_nid, dist);
1523 /* Only consider nodes where both task and groups benefit */
1524 taskimp = task_weight(p, nid, dist) - taskweight;
1525 groupimp = group_weight(p, nid, dist) - groupweight;
1526 if (taskimp < 0 && groupimp < 0)
1531 update_numa_stats(&env.dst_stats, env.dst_nid);
1532 if (numa_has_capacity(&env))
1533 task_numa_find_cpu(&env, taskimp, groupimp);
1538 * If the task is part of a workload that spans multiple NUMA nodes,
1539 * and is migrating into one of the workload's active nodes, remember
1540 * this node as the task's preferred numa node, so the workload can
1542 * A task that migrated to a second choice node will be better off
1543 * trying for a better one later. Do not set the preferred node here.
1545 if (p->numa_group) {
1546 if (env.best_cpu == -1)
1551 if (node_isset(nid, p->numa_group->active_nodes))
1552 sched_setnuma(p, env.dst_nid);
1555 /* No better CPU than the current one was found. */
1556 if (env.best_cpu == -1)
1560 * Reset the scan period if the task is being rescheduled on an
1561 * alternative node to recheck if the tasks is now properly placed.
1563 p->numa_scan_period = task_scan_min(p);
1565 if (env.best_task == NULL) {
1566 ret = migrate_task_to(p, env.best_cpu);
1568 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1572 ret = migrate_swap(p, env.best_task);
1574 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1575 put_task_struct(env.best_task);
1579 /* Attempt to migrate a task to a CPU on the preferred node. */
1580 static void numa_migrate_preferred(struct task_struct *p)
1582 unsigned long interval = HZ;
1584 /* This task has no NUMA fault statistics yet */
1585 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1588 /* Periodically retry migrating the task to the preferred node */
1589 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1590 p->numa_migrate_retry = jiffies + interval;
1592 /* Success if task is already running on preferred CPU */
1593 if (task_node(p) == p->numa_preferred_nid)
1596 /* Otherwise, try migrate to a CPU on the preferred node */
1597 task_numa_migrate(p);
1601 * Find the nodes on which the workload is actively running. We do this by
1602 * tracking the nodes from which NUMA hinting faults are triggered. This can
1603 * be different from the set of nodes where the workload's memory is currently
1606 * The bitmask is used to make smarter decisions on when to do NUMA page
1607 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1608 * are added when they cause over 6/16 of the maximum number of faults, but
1609 * only removed when they drop below 3/16.
1611 static void update_numa_active_node_mask(struct numa_group *numa_group)
1613 unsigned long faults, max_faults = 0;
1616 for_each_online_node(nid) {
1617 faults = group_faults_cpu(numa_group, nid);
1618 if (faults > max_faults)
1619 max_faults = faults;
1622 for_each_online_node(nid) {
1623 faults = group_faults_cpu(numa_group, nid);
1624 if (!node_isset(nid, numa_group->active_nodes)) {
1625 if (faults > max_faults * 6 / 16)
1626 node_set(nid, numa_group->active_nodes);
1627 } else if (faults < max_faults * 3 / 16)
1628 node_clear(nid, numa_group->active_nodes);
1633 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1634 * increments. The more local the fault statistics are, the higher the scan
1635 * period will be for the next scan window. If local/(local+remote) ratio is
1636 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1637 * the scan period will decrease. Aim for 70% local accesses.
1639 #define NUMA_PERIOD_SLOTS 10
1640 #define NUMA_PERIOD_THRESHOLD 7
1643 * Increase the scan period (slow down scanning) if the majority of
1644 * our memory is already on our local node, or if the majority of
1645 * the page accesses are shared with other processes.
1646 * Otherwise, decrease the scan period.
1648 static void update_task_scan_period(struct task_struct *p,
1649 unsigned long shared, unsigned long private)
1651 unsigned int period_slot;
1655 unsigned long remote = p->numa_faults_locality[0];
1656 unsigned long local = p->numa_faults_locality[1];
1659 * If there were no record hinting faults then either the task is
1660 * completely idle or all activity is areas that are not of interest
1661 * to automatic numa balancing. Related to that, if there were failed
1662 * migration then it implies we are migrating too quickly or the local
1663 * node is overloaded. In either case, scan slower
1665 if (local + shared == 0 || p->numa_faults_locality[2]) {
1666 p->numa_scan_period = min(p->numa_scan_period_max,
1667 p->numa_scan_period << 1);
1669 p->mm->numa_next_scan = jiffies +
1670 msecs_to_jiffies(p->numa_scan_period);
1676 * Prepare to scale scan period relative to the current period.
1677 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1678 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1679 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1681 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1682 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1683 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1684 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1687 diff = slot * period_slot;
1689 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1692 * Scale scan rate increases based on sharing. There is an
1693 * inverse relationship between the degree of sharing and
1694 * the adjustment made to the scanning period. Broadly
1695 * speaking the intent is that there is little point
1696 * scanning faster if shared accesses dominate as it may
1697 * simply bounce migrations uselessly
1699 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1700 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1703 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1704 task_scan_min(p), task_scan_max(p));
1705 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1709 * Get the fraction of time the task has been running since the last
1710 * NUMA placement cycle. The scheduler keeps similar statistics, but
1711 * decays those on a 32ms period, which is orders of magnitude off
1712 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1713 * stats only if the task is so new there are no NUMA statistics yet.
1715 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1717 u64 runtime, delta, now;
1718 /* Use the start of this time slice to avoid calculations. */
1719 now = p->se.exec_start;
1720 runtime = p->se.sum_exec_runtime;
1722 if (p->last_task_numa_placement) {
1723 delta = runtime - p->last_sum_exec_runtime;
1724 *period = now - p->last_task_numa_placement;
1726 delta = p->se.avg.load_sum / p->se.load.weight;
1727 *period = LOAD_AVG_MAX;
1730 p->last_sum_exec_runtime = runtime;
1731 p->last_task_numa_placement = now;
1737 * Determine the preferred nid for a task in a numa_group. This needs to
1738 * be done in a way that produces consistent results with group_weight,
1739 * otherwise workloads might not converge.
1741 static int preferred_group_nid(struct task_struct *p, int nid)
1746 /* Direct connections between all NUMA nodes. */
1747 if (sched_numa_topology_type == NUMA_DIRECT)
1751 * On a system with glueless mesh NUMA topology, group_weight
1752 * scores nodes according to the number of NUMA hinting faults on
1753 * both the node itself, and on nearby nodes.
1755 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1756 unsigned long score, max_score = 0;
1757 int node, max_node = nid;
1759 dist = sched_max_numa_distance;
1761 for_each_online_node(node) {
1762 score = group_weight(p, node, dist);
1763 if (score > max_score) {
1772 * Finding the preferred nid in a system with NUMA backplane
1773 * interconnect topology is more involved. The goal is to locate
1774 * tasks from numa_groups near each other in the system, and
1775 * untangle workloads from different sides of the system. This requires
1776 * searching down the hierarchy of node groups, recursively searching
1777 * inside the highest scoring group of nodes. The nodemask tricks
1778 * keep the complexity of the search down.
1780 nodes = node_online_map;
1781 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1782 unsigned long max_faults = 0;
1783 nodemask_t max_group = NODE_MASK_NONE;
1786 /* Are there nodes at this distance from each other? */
1787 if (!find_numa_distance(dist))
1790 for_each_node_mask(a, nodes) {
1791 unsigned long faults = 0;
1792 nodemask_t this_group;
1793 nodes_clear(this_group);
1795 /* Sum group's NUMA faults; includes a==b case. */
1796 for_each_node_mask(b, nodes) {
1797 if (node_distance(a, b) < dist) {
1798 faults += group_faults(p, b);
1799 node_set(b, this_group);
1800 node_clear(b, nodes);
1804 /* Remember the top group. */
1805 if (faults > max_faults) {
1806 max_faults = faults;
1807 max_group = this_group;
1809 * subtle: at the smallest distance there is
1810 * just one node left in each "group", the
1811 * winner is the preferred nid.
1816 /* Next round, evaluate the nodes within max_group. */
1824 static void task_numa_placement(struct task_struct *p)
1826 int seq, nid, max_nid = -1, max_group_nid = -1;
1827 unsigned long max_faults = 0, max_group_faults = 0;
1828 unsigned long fault_types[2] = { 0, 0 };
1829 unsigned long total_faults;
1830 u64 runtime, period;
1831 spinlock_t *group_lock = NULL;
1834 * The p->mm->numa_scan_seq field gets updated without
1835 * exclusive access. Use READ_ONCE() here to ensure
1836 * that the field is read in a single access:
1838 seq = READ_ONCE(p->mm->numa_scan_seq);
1839 if (p->numa_scan_seq == seq)
1841 p->numa_scan_seq = seq;
1842 p->numa_scan_period_max = task_scan_max(p);
1844 total_faults = p->numa_faults_locality[0] +
1845 p->numa_faults_locality[1];
1846 runtime = numa_get_avg_runtime(p, &period);
1848 /* If the task is part of a group prevent parallel updates to group stats */
1849 if (p->numa_group) {
1850 group_lock = &p->numa_group->lock;
1851 spin_lock_irq(group_lock);
1854 /* Find the node with the highest number of faults */
1855 for_each_online_node(nid) {
1856 /* Keep track of the offsets in numa_faults array */
1857 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1858 unsigned long faults = 0, group_faults = 0;
1861 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1862 long diff, f_diff, f_weight;
1864 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1865 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1866 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1867 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1869 /* Decay existing window, copy faults since last scan */
1870 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1871 fault_types[priv] += p->numa_faults[membuf_idx];
1872 p->numa_faults[membuf_idx] = 0;
1875 * Normalize the faults_from, so all tasks in a group
1876 * count according to CPU use, instead of by the raw
1877 * number of faults. Tasks with little runtime have
1878 * little over-all impact on throughput, and thus their
1879 * faults are less important.
1881 f_weight = div64_u64(runtime << 16, period + 1);
1882 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1884 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1885 p->numa_faults[cpubuf_idx] = 0;
1887 p->numa_faults[mem_idx] += diff;
1888 p->numa_faults[cpu_idx] += f_diff;
1889 faults += p->numa_faults[mem_idx];
1890 p->total_numa_faults += diff;
1891 if (p->numa_group) {
1893 * safe because we can only change our own group
1895 * mem_idx represents the offset for a given
1896 * nid and priv in a specific region because it
1897 * is at the beginning of the numa_faults array.
1899 p->numa_group->faults[mem_idx] += diff;
1900 p->numa_group->faults_cpu[mem_idx] += f_diff;
1901 p->numa_group->total_faults += diff;
1902 group_faults += p->numa_group->faults[mem_idx];
1906 if (faults > max_faults) {
1907 max_faults = faults;
1911 if (group_faults > max_group_faults) {
1912 max_group_faults = group_faults;
1913 max_group_nid = nid;
1917 update_task_scan_period(p, fault_types[0], fault_types[1]);
1919 if (p->numa_group) {
1920 update_numa_active_node_mask(p->numa_group);
1921 spin_unlock_irq(group_lock);
1922 max_nid = preferred_group_nid(p, max_group_nid);
1926 /* Set the new preferred node */
1927 if (max_nid != p->numa_preferred_nid)
1928 sched_setnuma(p, max_nid);
1930 if (task_node(p) != p->numa_preferred_nid)
1931 numa_migrate_preferred(p);
1935 static inline int get_numa_group(struct numa_group *grp)
1937 return atomic_inc_not_zero(&grp->refcount);
1940 static inline void put_numa_group(struct numa_group *grp)
1942 if (atomic_dec_and_test(&grp->refcount))
1943 kfree_rcu(grp, rcu);
1946 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1949 struct numa_group *grp, *my_grp;
1950 struct task_struct *tsk;
1952 int cpu = cpupid_to_cpu(cpupid);
1955 if (unlikely(!p->numa_group)) {
1956 unsigned int size = sizeof(struct numa_group) +
1957 4*nr_node_ids*sizeof(unsigned long);
1959 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1963 atomic_set(&grp->refcount, 1);
1964 spin_lock_init(&grp->lock);
1966 /* Second half of the array tracks nids where faults happen */
1967 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1970 node_set(task_node(current), grp->active_nodes);
1972 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1973 grp->faults[i] = p->numa_faults[i];
1975 grp->total_faults = p->total_numa_faults;
1978 rcu_assign_pointer(p->numa_group, grp);
1982 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1984 if (!cpupid_match_pid(tsk, cpupid))
1987 grp = rcu_dereference(tsk->numa_group);
1991 my_grp = p->numa_group;
1996 * Only join the other group if its bigger; if we're the bigger group,
1997 * the other task will join us.
1999 if (my_grp->nr_tasks > grp->nr_tasks)
2003 * Tie-break on the grp address.
2005 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2008 /* Always join threads in the same process. */
2009 if (tsk->mm == current->mm)
2012 /* Simple filter to avoid false positives due to PID collisions */
2013 if (flags & TNF_SHARED)
2016 /* Update priv based on whether false sharing was detected */
2019 if (join && !get_numa_group(grp))
2027 BUG_ON(irqs_disabled());
2028 double_lock_irq(&my_grp->lock, &grp->lock);
2030 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2031 my_grp->faults[i] -= p->numa_faults[i];
2032 grp->faults[i] += p->numa_faults[i];
2034 my_grp->total_faults -= p->total_numa_faults;
2035 grp->total_faults += p->total_numa_faults;
2040 spin_unlock(&my_grp->lock);
2041 spin_unlock_irq(&grp->lock);
2043 rcu_assign_pointer(p->numa_group, grp);
2045 put_numa_group(my_grp);
2053 void task_numa_free(struct task_struct *p)
2055 struct numa_group *grp = p->numa_group;
2056 void *numa_faults = p->numa_faults;
2057 unsigned long flags;
2061 spin_lock_irqsave(&grp->lock, flags);
2062 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2063 grp->faults[i] -= p->numa_faults[i];
2064 grp->total_faults -= p->total_numa_faults;
2067 spin_unlock_irqrestore(&grp->lock, flags);
2068 RCU_INIT_POINTER(p->numa_group, NULL);
2069 put_numa_group(grp);
2072 p->numa_faults = NULL;
2077 * Got a PROT_NONE fault for a page on @node.
2079 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2081 struct task_struct *p = current;
2082 bool migrated = flags & TNF_MIGRATED;
2083 int cpu_node = task_node(current);
2084 int local = !!(flags & TNF_FAULT_LOCAL);
2087 if (!static_branch_likely(&sched_numa_balancing))
2090 /* for example, ksmd faulting in a user's mm */
2094 /* Allocate buffer to track faults on a per-node basis */
2095 if (unlikely(!p->numa_faults)) {
2096 int size = sizeof(*p->numa_faults) *
2097 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2099 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2100 if (!p->numa_faults)
2103 p->total_numa_faults = 0;
2104 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2108 * First accesses are treated as private, otherwise consider accesses
2109 * to be private if the accessing pid has not changed
2111 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2114 priv = cpupid_match_pid(p, last_cpupid);
2115 if (!priv && !(flags & TNF_NO_GROUP))
2116 task_numa_group(p, last_cpupid, flags, &priv);
2120 * If a workload spans multiple NUMA nodes, a shared fault that
2121 * occurs wholly within the set of nodes that the workload is
2122 * actively using should be counted as local. This allows the
2123 * scan rate to slow down when a workload has settled down.
2125 if (!priv && !local && p->numa_group &&
2126 node_isset(cpu_node, p->numa_group->active_nodes) &&
2127 node_isset(mem_node, p->numa_group->active_nodes))
2130 task_numa_placement(p);
2133 * Retry task to preferred node migration periodically, in case it
2134 * case it previously failed, or the scheduler moved us.
2136 if (time_after(jiffies, p->numa_migrate_retry))
2137 numa_migrate_preferred(p);
2140 p->numa_pages_migrated += pages;
2141 if (flags & TNF_MIGRATE_FAIL)
2142 p->numa_faults_locality[2] += pages;
2144 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2145 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2146 p->numa_faults_locality[local] += pages;
2149 static void reset_ptenuma_scan(struct task_struct *p)
2152 * We only did a read acquisition of the mmap sem, so
2153 * p->mm->numa_scan_seq is written to without exclusive access
2154 * and the update is not guaranteed to be atomic. That's not
2155 * much of an issue though, since this is just used for
2156 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2157 * expensive, to avoid any form of compiler optimizations:
2159 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2160 p->mm->numa_scan_offset = 0;
2164 * The expensive part of numa migration is done from task_work context.
2165 * Triggered from task_tick_numa().
2167 void task_numa_work(struct callback_head *work)
2169 unsigned long migrate, next_scan, now = jiffies;
2170 struct task_struct *p = current;
2171 struct mm_struct *mm = p->mm;
2172 struct vm_area_struct *vma;
2173 unsigned long start, end;
2174 unsigned long nr_pte_updates = 0;
2175 long pages, virtpages;
2177 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2179 work->next = work; /* protect against double add */
2181 * Who cares about NUMA placement when they're dying.
2183 * NOTE: make sure not to dereference p->mm before this check,
2184 * exit_task_work() happens _after_ exit_mm() so we could be called
2185 * without p->mm even though we still had it when we enqueued this
2188 if (p->flags & PF_EXITING)
2191 if (!mm->numa_next_scan) {
2192 mm->numa_next_scan = now +
2193 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2197 * Enforce maximal scan/migration frequency..
2199 migrate = mm->numa_next_scan;
2200 if (time_before(now, migrate))
2203 if (p->numa_scan_period == 0) {
2204 p->numa_scan_period_max = task_scan_max(p);
2205 p->numa_scan_period = task_scan_min(p);
2208 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2209 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2213 * Delay this task enough that another task of this mm will likely win
2214 * the next time around.
2216 p->node_stamp += 2 * TICK_NSEC;
2218 start = mm->numa_scan_offset;
2219 pages = sysctl_numa_balancing_scan_size;
2220 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2221 virtpages = pages * 8; /* Scan up to this much virtual space */
2226 down_read(&mm->mmap_sem);
2227 vma = find_vma(mm, start);
2229 reset_ptenuma_scan(p);
2233 for (; vma; vma = vma->vm_next) {
2234 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2235 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2240 * Shared library pages mapped by multiple processes are not
2241 * migrated as it is expected they are cache replicated. Avoid
2242 * hinting faults in read-only file-backed mappings or the vdso
2243 * as migrating the pages will be of marginal benefit.
2246 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2250 * Skip inaccessible VMAs to avoid any confusion between
2251 * PROT_NONE and NUMA hinting ptes
2253 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2257 start = max(start, vma->vm_start);
2258 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2259 end = min(end, vma->vm_end);
2260 nr_pte_updates = change_prot_numa(vma, start, end);
2263 * Try to scan sysctl_numa_balancing_size worth of
2264 * hpages that have at least one present PTE that
2265 * is not already pte-numa. If the VMA contains
2266 * areas that are unused or already full of prot_numa
2267 * PTEs, scan up to virtpages, to skip through those
2271 pages -= (end - start) >> PAGE_SHIFT;
2272 virtpages -= (end - start) >> PAGE_SHIFT;
2275 if (pages <= 0 || virtpages <= 0)
2279 } while (end != vma->vm_end);
2284 * It is possible to reach the end of the VMA list but the last few
2285 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2286 * would find the !migratable VMA on the next scan but not reset the
2287 * scanner to the start so check it now.
2290 mm->numa_scan_offset = start;
2292 reset_ptenuma_scan(p);
2293 up_read(&mm->mmap_sem);
2297 * Drive the periodic memory faults..
2299 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2301 struct callback_head *work = &curr->numa_work;
2305 * We don't care about NUMA placement if we don't have memory.
2307 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2311 * Using runtime rather than walltime has the dual advantage that
2312 * we (mostly) drive the selection from busy threads and that the
2313 * task needs to have done some actual work before we bother with
2316 now = curr->se.sum_exec_runtime;
2317 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2319 if (now > curr->node_stamp + period) {
2320 if (!curr->node_stamp)
2321 curr->numa_scan_period = task_scan_min(curr);
2322 curr->node_stamp += period;
2324 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2325 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2326 task_work_add(curr, work, true);
2331 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2335 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2339 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2342 #endif /* CONFIG_NUMA_BALANCING */
2345 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2347 update_load_add(&cfs_rq->load, se->load.weight);
2348 if (!parent_entity(se))
2349 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2351 if (entity_is_task(se)) {
2352 struct rq *rq = rq_of(cfs_rq);
2354 account_numa_enqueue(rq, task_of(se));
2355 list_add(&se->group_node, &rq->cfs_tasks);
2358 cfs_rq->nr_running++;
2362 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2364 update_load_sub(&cfs_rq->load, se->load.weight);
2365 if (!parent_entity(se))
2366 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2367 if (entity_is_task(se)) {
2368 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2369 list_del_init(&se->group_node);
2371 cfs_rq->nr_running--;
2374 #ifdef CONFIG_FAIR_GROUP_SCHED
2376 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2381 * Use this CPU's real-time load instead of the last load contribution
2382 * as the updating of the contribution is delayed, and we will use the
2383 * the real-time load to calc the share. See update_tg_load_avg().
2385 tg_weight = atomic_long_read(&tg->load_avg);
2386 tg_weight -= cfs_rq->tg_load_avg_contrib;
2387 tg_weight += cfs_rq->load.weight;
2392 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2394 long tg_weight, load, shares;
2396 tg_weight = calc_tg_weight(tg, cfs_rq);
2397 load = cfs_rq->load.weight;
2399 shares = (tg->shares * load);
2401 shares /= tg_weight;
2403 if (shares < MIN_SHARES)
2404 shares = MIN_SHARES;
2405 if (shares > tg->shares)
2406 shares = tg->shares;
2410 # else /* CONFIG_SMP */
2411 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2415 # endif /* CONFIG_SMP */
2416 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2417 unsigned long weight)
2420 /* commit outstanding execution time */
2421 if (cfs_rq->curr == se)
2422 update_curr(cfs_rq);
2423 account_entity_dequeue(cfs_rq, se);
2426 update_load_set(&se->load, weight);
2429 account_entity_enqueue(cfs_rq, se);
2432 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2434 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2436 struct task_group *tg;
2437 struct sched_entity *se;
2441 se = tg->se[cpu_of(rq_of(cfs_rq))];
2442 if (!se || throttled_hierarchy(cfs_rq))
2445 if (likely(se->load.weight == tg->shares))
2448 shares = calc_cfs_shares(cfs_rq, tg);
2450 reweight_entity(cfs_rq_of(se), se, shares);
2452 #else /* CONFIG_FAIR_GROUP_SCHED */
2453 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2456 #endif /* CONFIG_FAIR_GROUP_SCHED */
2459 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2460 static const u32 runnable_avg_yN_inv[] = {
2461 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2462 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2463 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2464 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2465 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2466 0x85aac367, 0x82cd8698,
2470 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2471 * over-estimates when re-combining.
2473 static const u32 runnable_avg_yN_sum[] = {
2474 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2475 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2476 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2481 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2483 static __always_inline u64 decay_load(u64 val, u64 n)
2485 unsigned int local_n;
2489 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2492 /* after bounds checking we can collapse to 32-bit */
2496 * As y^PERIOD = 1/2, we can combine
2497 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2498 * With a look-up table which covers y^n (n<PERIOD)
2500 * To achieve constant time decay_load.
2502 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2503 val >>= local_n / LOAD_AVG_PERIOD;
2504 local_n %= LOAD_AVG_PERIOD;
2507 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2512 * For updates fully spanning n periods, the contribution to runnable
2513 * average will be: \Sum 1024*y^n
2515 * We can compute this reasonably efficiently by combining:
2516 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2518 static u32 __compute_runnable_contrib(u64 n)
2522 if (likely(n <= LOAD_AVG_PERIOD))
2523 return runnable_avg_yN_sum[n];
2524 else if (unlikely(n >= LOAD_AVG_MAX_N))
2525 return LOAD_AVG_MAX;
2527 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2529 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2530 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2532 n -= LOAD_AVG_PERIOD;
2533 } while (n > LOAD_AVG_PERIOD);
2535 contrib = decay_load(contrib, n);
2536 return contrib + runnable_avg_yN_sum[n];
2539 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2540 #error "load tracking assumes 2^10 as unit"
2543 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2546 * We can represent the historical contribution to runnable average as the
2547 * coefficients of a geometric series. To do this we sub-divide our runnable
2548 * history into segments of approximately 1ms (1024us); label the segment that
2549 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2551 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2553 * (now) (~1ms ago) (~2ms ago)
2555 * Let u_i denote the fraction of p_i that the entity was runnable.
2557 * We then designate the fractions u_i as our co-efficients, yielding the
2558 * following representation of historical load:
2559 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2561 * We choose y based on the with of a reasonably scheduling period, fixing:
2564 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2565 * approximately half as much as the contribution to load within the last ms
2568 * When a period "rolls over" and we have new u_0`, multiplying the previous
2569 * sum again by y is sufficient to update:
2570 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2571 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2573 static __always_inline int
2574 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2575 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2577 u64 delta, scaled_delta, periods;
2579 unsigned int delta_w, scaled_delta_w, decayed = 0;
2580 unsigned long scale_freq, scale_cpu;
2582 delta = now - sa->last_update_time;
2584 * This should only happen when time goes backwards, which it
2585 * unfortunately does during sched clock init when we swap over to TSC.
2587 if ((s64)delta < 0) {
2588 sa->last_update_time = now;
2593 * Use 1024ns as the unit of measurement since it's a reasonable
2594 * approximation of 1us and fast to compute.
2599 sa->last_update_time = now;
2601 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2602 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2604 /* delta_w is the amount already accumulated against our next period */
2605 delta_w = sa->period_contrib;
2606 if (delta + delta_w >= 1024) {
2609 /* how much left for next period will start over, we don't know yet */
2610 sa->period_contrib = 0;
2613 * Now that we know we're crossing a period boundary, figure
2614 * out how much from delta we need to complete the current
2615 * period and accrue it.
2617 delta_w = 1024 - delta_w;
2618 scaled_delta_w = cap_scale(delta_w, scale_freq);
2620 sa->load_sum += weight * scaled_delta_w;
2622 cfs_rq->runnable_load_sum +=
2623 weight * scaled_delta_w;
2627 sa->util_sum += scaled_delta_w * scale_cpu;
2631 /* Figure out how many additional periods this update spans */
2632 periods = delta / 1024;
2635 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2637 cfs_rq->runnable_load_sum =
2638 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2640 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2642 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2643 contrib = __compute_runnable_contrib(periods);
2644 contrib = cap_scale(contrib, scale_freq);
2646 sa->load_sum += weight * contrib;
2648 cfs_rq->runnable_load_sum += weight * contrib;
2651 sa->util_sum += contrib * scale_cpu;
2654 /* Remainder of delta accrued against u_0` */
2655 scaled_delta = cap_scale(delta, scale_freq);
2657 sa->load_sum += weight * scaled_delta;
2659 cfs_rq->runnable_load_sum += weight * scaled_delta;
2662 sa->util_sum += scaled_delta * scale_cpu;
2664 sa->period_contrib += delta;
2667 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2669 cfs_rq->runnable_load_avg =
2670 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2672 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2678 #ifdef CONFIG_FAIR_GROUP_SCHED
2680 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2681 * and effective_load (which is not done because it is too costly).
2683 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2685 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2687 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2688 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2689 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2693 #else /* CONFIG_FAIR_GROUP_SCHED */
2694 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2695 #endif /* CONFIG_FAIR_GROUP_SCHED */
2697 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2700 * Unsigned subtract and clamp on underflow.
2702 * Explicitly do a load-store to ensure the intermediate value never hits
2703 * memory. This allows lockless observations without ever seeing the negative
2706 #define sub_positive(_ptr, _val) do { \
2707 typeof(_ptr) ptr = (_ptr); \
2708 typeof(*ptr) val = (_val); \
2709 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2713 WRITE_ONCE(*ptr, res); \
2716 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2717 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2719 struct sched_avg *sa = &cfs_rq->avg;
2720 int decayed, removed = 0;
2722 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2723 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2724 sub_positive(&sa->load_avg, r);
2725 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2729 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2730 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2731 sub_positive(&sa->util_avg, r);
2732 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2735 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2736 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2738 #ifndef CONFIG_64BIT
2740 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2743 return decayed || removed;
2746 /* Update task and its cfs_rq load average */
2747 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2749 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2750 u64 now = cfs_rq_clock_task(cfs_rq);
2751 int cpu = cpu_of(rq_of(cfs_rq));
2754 * Track task load average for carrying it to new CPU after migrated, and
2755 * track group sched_entity load average for task_h_load calc in migration
2757 __update_load_avg(now, cpu, &se->avg,
2758 se->on_rq * scale_load_down(se->load.weight),
2759 cfs_rq->curr == se, NULL);
2761 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2762 update_tg_load_avg(cfs_rq, 0);
2765 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2767 if (!sched_feat(ATTACH_AGE_LOAD))
2771 * If we got migrated (either between CPUs or between cgroups) we'll
2772 * have aged the average right before clearing @last_update_time.
2774 if (se->avg.last_update_time) {
2775 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2776 &se->avg, 0, 0, NULL);
2779 * XXX: we could have just aged the entire load away if we've been
2780 * absent from the fair class for too long.
2785 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2786 cfs_rq->avg.load_avg += se->avg.load_avg;
2787 cfs_rq->avg.load_sum += se->avg.load_sum;
2788 cfs_rq->avg.util_avg += se->avg.util_avg;
2789 cfs_rq->avg.util_sum += se->avg.util_sum;
2792 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2794 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2795 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2796 cfs_rq->curr == se, NULL);
2798 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2799 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2800 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2801 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2804 /* Add the load generated by se into cfs_rq's load average */
2806 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2808 struct sched_avg *sa = &se->avg;
2809 u64 now = cfs_rq_clock_task(cfs_rq);
2810 int migrated, decayed;
2812 migrated = !sa->last_update_time;
2814 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2815 se->on_rq * scale_load_down(se->load.weight),
2816 cfs_rq->curr == se, NULL);
2819 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2821 cfs_rq->runnable_load_avg += sa->load_avg;
2822 cfs_rq->runnable_load_sum += sa->load_sum;
2825 attach_entity_load_avg(cfs_rq, se);
2827 if (decayed || migrated)
2828 update_tg_load_avg(cfs_rq, 0);
2831 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2833 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2835 update_load_avg(se, 1);
2837 cfs_rq->runnable_load_avg =
2838 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2839 cfs_rq->runnable_load_sum =
2840 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2844 * Task first catches up with cfs_rq, and then subtract
2845 * itself from the cfs_rq (task must be off the queue now).
2847 void remove_entity_load_avg(struct sched_entity *se)
2849 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2850 u64 last_update_time;
2852 #ifndef CONFIG_64BIT
2853 u64 last_update_time_copy;
2856 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2858 last_update_time = cfs_rq->avg.last_update_time;
2859 } while (last_update_time != last_update_time_copy);
2861 last_update_time = cfs_rq->avg.last_update_time;
2864 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2865 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2866 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2870 * Update the rq's load with the elapsed running time before entering
2871 * idle. if the last scheduled task is not a CFS task, idle_enter will
2872 * be the only way to update the runnable statistic.
2874 void idle_enter_fair(struct rq *this_rq)
2879 * Update the rq's load with the elapsed idle time before a task is
2880 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2881 * be the only way to update the runnable statistic.
2883 void idle_exit_fair(struct rq *this_rq)
2887 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2889 return cfs_rq->runnable_load_avg;
2892 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2894 return cfs_rq->avg.load_avg;
2897 static int idle_balance(struct rq *this_rq);
2899 #else /* CONFIG_SMP */
2901 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2903 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2905 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2906 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2909 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2911 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2913 static inline int idle_balance(struct rq *rq)
2918 #endif /* CONFIG_SMP */
2920 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2922 #ifdef CONFIG_SCHEDSTATS
2923 struct task_struct *tsk = NULL;
2925 if (entity_is_task(se))
2928 if (se->statistics.sleep_start) {
2929 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2934 if (unlikely(delta > se->statistics.sleep_max))
2935 se->statistics.sleep_max = delta;
2937 se->statistics.sleep_start = 0;
2938 se->statistics.sum_sleep_runtime += delta;
2941 account_scheduler_latency(tsk, delta >> 10, 1);
2942 trace_sched_stat_sleep(tsk, delta);
2945 if (se->statistics.block_start) {
2946 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2951 if (unlikely(delta > se->statistics.block_max))
2952 se->statistics.block_max = delta;
2954 se->statistics.block_start = 0;
2955 se->statistics.sum_sleep_runtime += delta;
2958 if (tsk->in_iowait) {
2959 se->statistics.iowait_sum += delta;
2960 se->statistics.iowait_count++;
2961 trace_sched_stat_iowait(tsk, delta);
2964 trace_sched_stat_blocked(tsk, delta);
2967 * Blocking time is in units of nanosecs, so shift by
2968 * 20 to get a milliseconds-range estimation of the
2969 * amount of time that the task spent sleeping:
2971 if (unlikely(prof_on == SLEEP_PROFILING)) {
2972 profile_hits(SLEEP_PROFILING,
2973 (void *)get_wchan(tsk),
2976 account_scheduler_latency(tsk, delta >> 10, 0);
2982 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2984 #ifdef CONFIG_SCHED_DEBUG
2985 s64 d = se->vruntime - cfs_rq->min_vruntime;
2990 if (d > 3*sysctl_sched_latency)
2991 schedstat_inc(cfs_rq, nr_spread_over);
2996 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2998 u64 vruntime = cfs_rq->min_vruntime;
3001 * The 'current' period is already promised to the current tasks,
3002 * however the extra weight of the new task will slow them down a
3003 * little, place the new task so that it fits in the slot that
3004 * stays open at the end.
3006 if (initial && sched_feat(START_DEBIT))
3007 vruntime += sched_vslice(cfs_rq, se);
3009 /* sleeps up to a single latency don't count. */
3011 unsigned long thresh = sysctl_sched_latency;
3014 * Halve their sleep time's effect, to allow
3015 * for a gentler effect of sleepers:
3017 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3023 /* ensure we never gain time by being placed backwards. */
3024 se->vruntime = max_vruntime(se->vruntime, vruntime);
3027 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3030 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3033 * Update the normalized vruntime before updating min_vruntime
3034 * through calling update_curr().
3036 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3037 se->vruntime += cfs_rq->min_vruntime;
3040 * Update run-time statistics of the 'current'.
3042 update_curr(cfs_rq);
3043 enqueue_entity_load_avg(cfs_rq, se);
3044 account_entity_enqueue(cfs_rq, se);
3045 update_cfs_shares(cfs_rq);
3047 if (flags & ENQUEUE_WAKEUP) {
3048 place_entity(cfs_rq, se, 0);
3049 enqueue_sleeper(cfs_rq, se);
3052 update_stats_enqueue(cfs_rq, se);
3053 check_spread(cfs_rq, se);
3054 if (se != cfs_rq->curr)
3055 __enqueue_entity(cfs_rq, se);
3058 if (cfs_rq->nr_running == 1) {
3059 list_add_leaf_cfs_rq(cfs_rq);
3060 check_enqueue_throttle(cfs_rq);
3064 static void __clear_buddies_last(struct sched_entity *se)
3066 for_each_sched_entity(se) {
3067 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3068 if (cfs_rq->last != se)
3071 cfs_rq->last = NULL;
3075 static void __clear_buddies_next(struct sched_entity *se)
3077 for_each_sched_entity(se) {
3078 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3079 if (cfs_rq->next != se)
3082 cfs_rq->next = NULL;
3086 static void __clear_buddies_skip(struct sched_entity *se)
3088 for_each_sched_entity(se) {
3089 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3090 if (cfs_rq->skip != se)
3093 cfs_rq->skip = NULL;
3097 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3099 if (cfs_rq->last == se)
3100 __clear_buddies_last(se);
3102 if (cfs_rq->next == se)
3103 __clear_buddies_next(se);
3105 if (cfs_rq->skip == se)
3106 __clear_buddies_skip(se);
3109 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3112 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3115 * Update run-time statistics of the 'current'.
3117 update_curr(cfs_rq);
3118 dequeue_entity_load_avg(cfs_rq, se);
3120 update_stats_dequeue(cfs_rq, se);
3121 if (flags & DEQUEUE_SLEEP) {
3122 #ifdef CONFIG_SCHEDSTATS
3123 if (entity_is_task(se)) {
3124 struct task_struct *tsk = task_of(se);
3126 if (tsk->state & TASK_INTERRUPTIBLE)
3127 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3128 if (tsk->state & TASK_UNINTERRUPTIBLE)
3129 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3134 clear_buddies(cfs_rq, se);
3136 if (se != cfs_rq->curr)
3137 __dequeue_entity(cfs_rq, se);
3139 account_entity_dequeue(cfs_rq, se);
3142 * Normalize the entity after updating the min_vruntime because the
3143 * update can refer to the ->curr item and we need to reflect this
3144 * movement in our normalized position.
3146 if (!(flags & DEQUEUE_SLEEP))
3147 se->vruntime -= cfs_rq->min_vruntime;
3149 /* return excess runtime on last dequeue */
3150 return_cfs_rq_runtime(cfs_rq);
3152 update_min_vruntime(cfs_rq);
3153 update_cfs_shares(cfs_rq);
3157 * Preempt the current task with a newly woken task if needed:
3160 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3162 unsigned long ideal_runtime, delta_exec;
3163 struct sched_entity *se;
3166 ideal_runtime = sched_slice(cfs_rq, curr);
3167 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3168 if (delta_exec > ideal_runtime) {
3169 resched_curr(rq_of(cfs_rq));
3171 * The current task ran long enough, ensure it doesn't get
3172 * re-elected due to buddy favours.
3174 clear_buddies(cfs_rq, curr);
3179 * Ensure that a task that missed wakeup preemption by a
3180 * narrow margin doesn't have to wait for a full slice.
3181 * This also mitigates buddy induced latencies under load.
3183 if (delta_exec < sysctl_sched_min_granularity)
3186 se = __pick_first_entity(cfs_rq);
3187 delta = curr->vruntime - se->vruntime;
3192 if (delta > ideal_runtime)
3193 resched_curr(rq_of(cfs_rq));
3197 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3199 /* 'current' is not kept within the tree. */
3202 * Any task has to be enqueued before it get to execute on
3203 * a CPU. So account for the time it spent waiting on the
3206 update_stats_wait_end(cfs_rq, se);
3207 __dequeue_entity(cfs_rq, se);
3208 update_load_avg(se, 1);
3211 update_stats_curr_start(cfs_rq, se);
3213 #ifdef CONFIG_SCHEDSTATS
3215 * Track our maximum slice length, if the CPU's load is at
3216 * least twice that of our own weight (i.e. dont track it
3217 * when there are only lesser-weight tasks around):
3219 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3220 se->statistics.slice_max = max(se->statistics.slice_max,
3221 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3224 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3228 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3231 * Pick the next process, keeping these things in mind, in this order:
3232 * 1) keep things fair between processes/task groups
3233 * 2) pick the "next" process, since someone really wants that to run
3234 * 3) pick the "last" process, for cache locality
3235 * 4) do not run the "skip" process, if something else is available
3237 static struct sched_entity *
3238 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3240 struct sched_entity *left = __pick_first_entity(cfs_rq);
3241 struct sched_entity *se;
3244 * If curr is set we have to see if its left of the leftmost entity
3245 * still in the tree, provided there was anything in the tree at all.
3247 if (!left || (curr && entity_before(curr, left)))
3250 se = left; /* ideally we run the leftmost entity */
3253 * Avoid running the skip buddy, if running something else can
3254 * be done without getting too unfair.
3256 if (cfs_rq->skip == se) {
3257 struct sched_entity *second;
3260 second = __pick_first_entity(cfs_rq);
3262 second = __pick_next_entity(se);
3263 if (!second || (curr && entity_before(curr, second)))
3267 if (second && wakeup_preempt_entity(second, left) < 1)
3272 * Prefer last buddy, try to return the CPU to a preempted task.
3274 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3278 * Someone really wants this to run. If it's not unfair, run it.
3280 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3283 clear_buddies(cfs_rq, se);
3288 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3290 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3293 * If still on the runqueue then deactivate_task()
3294 * was not called and update_curr() has to be done:
3297 update_curr(cfs_rq);
3299 /* throttle cfs_rqs exceeding runtime */
3300 check_cfs_rq_runtime(cfs_rq);
3302 check_spread(cfs_rq, prev);
3304 update_stats_wait_start(cfs_rq, prev);
3305 /* Put 'current' back into the tree. */
3306 __enqueue_entity(cfs_rq, prev);
3307 /* in !on_rq case, update occurred at dequeue */
3308 update_load_avg(prev, 0);
3310 cfs_rq->curr = NULL;
3314 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3317 * Update run-time statistics of the 'current'.
3319 update_curr(cfs_rq);
3322 * Ensure that runnable average is periodically updated.
3324 update_load_avg(curr, 1);
3325 update_cfs_shares(cfs_rq);
3327 #ifdef CONFIG_SCHED_HRTICK
3329 * queued ticks are scheduled to match the slice, so don't bother
3330 * validating it and just reschedule.
3333 resched_curr(rq_of(cfs_rq));
3337 * don't let the period tick interfere with the hrtick preemption
3339 if (!sched_feat(DOUBLE_TICK) &&
3340 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3344 if (cfs_rq->nr_running > 1)
3345 check_preempt_tick(cfs_rq, curr);
3349 /**************************************************
3350 * CFS bandwidth control machinery
3353 #ifdef CONFIG_CFS_BANDWIDTH
3355 #ifdef HAVE_JUMP_LABEL
3356 static struct static_key __cfs_bandwidth_used;
3358 static inline bool cfs_bandwidth_used(void)
3360 return static_key_false(&__cfs_bandwidth_used);
3363 void cfs_bandwidth_usage_inc(void)
3365 static_key_slow_inc(&__cfs_bandwidth_used);
3368 void cfs_bandwidth_usage_dec(void)
3370 static_key_slow_dec(&__cfs_bandwidth_used);
3372 #else /* HAVE_JUMP_LABEL */
3373 static bool cfs_bandwidth_used(void)
3378 void cfs_bandwidth_usage_inc(void) {}
3379 void cfs_bandwidth_usage_dec(void) {}
3380 #endif /* HAVE_JUMP_LABEL */
3383 * default period for cfs group bandwidth.
3384 * default: 0.1s, units: nanoseconds
3386 static inline u64 default_cfs_period(void)
3388 return 100000000ULL;
3391 static inline u64 sched_cfs_bandwidth_slice(void)
3393 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3397 * Replenish runtime according to assigned quota and update expiration time.
3398 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3399 * additional synchronization around rq->lock.
3401 * requires cfs_b->lock
3403 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3407 if (cfs_b->quota == RUNTIME_INF)
3410 now = sched_clock_cpu(smp_processor_id());
3411 cfs_b->runtime = cfs_b->quota;
3412 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3415 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3417 return &tg->cfs_bandwidth;
3420 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3421 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3423 if (unlikely(cfs_rq->throttle_count))
3424 return cfs_rq->throttled_clock_task;
3426 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3429 /* returns 0 on failure to allocate runtime */
3430 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3432 struct task_group *tg = cfs_rq->tg;
3433 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3434 u64 amount = 0, min_amount, expires;
3436 /* note: this is a positive sum as runtime_remaining <= 0 */
3437 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3439 raw_spin_lock(&cfs_b->lock);
3440 if (cfs_b->quota == RUNTIME_INF)
3441 amount = min_amount;
3443 start_cfs_bandwidth(cfs_b);
3445 if (cfs_b->runtime > 0) {
3446 amount = min(cfs_b->runtime, min_amount);
3447 cfs_b->runtime -= amount;
3451 expires = cfs_b->runtime_expires;
3452 raw_spin_unlock(&cfs_b->lock);
3454 cfs_rq->runtime_remaining += amount;
3456 * we may have advanced our local expiration to account for allowed
3457 * spread between our sched_clock and the one on which runtime was
3460 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3461 cfs_rq->runtime_expires = expires;
3463 return cfs_rq->runtime_remaining > 0;
3467 * Note: This depends on the synchronization provided by sched_clock and the
3468 * fact that rq->clock snapshots this value.
3470 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3472 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3474 /* if the deadline is ahead of our clock, nothing to do */
3475 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3478 if (cfs_rq->runtime_remaining < 0)
3482 * If the local deadline has passed we have to consider the
3483 * possibility that our sched_clock is 'fast' and the global deadline
3484 * has not truly expired.
3486 * Fortunately we can check determine whether this the case by checking
3487 * whether the global deadline has advanced. It is valid to compare
3488 * cfs_b->runtime_expires without any locks since we only care about
3489 * exact equality, so a partial write will still work.
3492 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3493 /* extend local deadline, drift is bounded above by 2 ticks */
3494 cfs_rq->runtime_expires += TICK_NSEC;
3496 /* global deadline is ahead, expiration has passed */
3497 cfs_rq->runtime_remaining = 0;
3501 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3503 /* dock delta_exec before expiring quota (as it could span periods) */
3504 cfs_rq->runtime_remaining -= delta_exec;
3505 expire_cfs_rq_runtime(cfs_rq);
3507 if (likely(cfs_rq->runtime_remaining > 0))
3511 * if we're unable to extend our runtime we resched so that the active
3512 * hierarchy can be throttled
3514 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3515 resched_curr(rq_of(cfs_rq));
3518 static __always_inline
3519 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3521 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3524 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3527 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3529 return cfs_bandwidth_used() && cfs_rq->throttled;
3532 /* check whether cfs_rq, or any parent, is throttled */
3533 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3535 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3539 * Ensure that neither of the group entities corresponding to src_cpu or
3540 * dest_cpu are members of a throttled hierarchy when performing group
3541 * load-balance operations.
3543 static inline int throttled_lb_pair(struct task_group *tg,
3544 int src_cpu, int dest_cpu)
3546 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3548 src_cfs_rq = tg->cfs_rq[src_cpu];
3549 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3551 return throttled_hierarchy(src_cfs_rq) ||
3552 throttled_hierarchy(dest_cfs_rq);
3555 /* updated child weight may affect parent so we have to do this bottom up */
3556 static int tg_unthrottle_up(struct task_group *tg, void *data)
3558 struct rq *rq = data;
3559 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3561 cfs_rq->throttle_count--;
3563 if (!cfs_rq->throttle_count) {
3564 /* adjust cfs_rq_clock_task() */
3565 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3566 cfs_rq->throttled_clock_task;
3573 static int tg_throttle_down(struct task_group *tg, void *data)
3575 struct rq *rq = data;
3576 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3578 /* group is entering throttled state, stop time */
3579 if (!cfs_rq->throttle_count)
3580 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3581 cfs_rq->throttle_count++;
3586 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3588 struct rq *rq = rq_of(cfs_rq);
3589 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3590 struct sched_entity *se;
3591 long task_delta, dequeue = 1;
3594 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3596 /* freeze hierarchy runnable averages while throttled */
3598 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3601 task_delta = cfs_rq->h_nr_running;
3602 for_each_sched_entity(se) {
3603 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3604 /* throttled entity or throttle-on-deactivate */
3609 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3610 qcfs_rq->h_nr_running -= task_delta;
3612 if (qcfs_rq->load.weight)
3617 sub_nr_running(rq, task_delta);
3619 cfs_rq->throttled = 1;
3620 cfs_rq->throttled_clock = rq_clock(rq);
3621 raw_spin_lock(&cfs_b->lock);
3622 empty = list_empty(&cfs_b->throttled_cfs_rq);
3625 * Add to the _head_ of the list, so that an already-started
3626 * distribute_cfs_runtime will not see us
3628 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3631 * If we're the first throttled task, make sure the bandwidth
3635 start_cfs_bandwidth(cfs_b);
3637 raw_spin_unlock(&cfs_b->lock);
3640 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3642 struct rq *rq = rq_of(cfs_rq);
3643 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3644 struct sched_entity *se;
3648 se = cfs_rq->tg->se[cpu_of(rq)];
3650 cfs_rq->throttled = 0;
3652 update_rq_clock(rq);
3654 raw_spin_lock(&cfs_b->lock);
3655 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3656 list_del_rcu(&cfs_rq->throttled_list);
3657 raw_spin_unlock(&cfs_b->lock);
3659 /* update hierarchical throttle state */
3660 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3662 if (!cfs_rq->load.weight)
3665 task_delta = cfs_rq->h_nr_running;
3666 for_each_sched_entity(se) {
3670 cfs_rq = cfs_rq_of(se);
3672 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3673 cfs_rq->h_nr_running += task_delta;
3675 if (cfs_rq_throttled(cfs_rq))
3680 add_nr_running(rq, task_delta);
3682 /* determine whether we need to wake up potentially idle cpu */
3683 if (rq->curr == rq->idle && rq->cfs.nr_running)
3687 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3688 u64 remaining, u64 expires)
3690 struct cfs_rq *cfs_rq;
3692 u64 starting_runtime = remaining;
3695 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3697 struct rq *rq = rq_of(cfs_rq);
3699 raw_spin_lock(&rq->lock);
3700 if (!cfs_rq_throttled(cfs_rq))
3703 runtime = -cfs_rq->runtime_remaining + 1;
3704 if (runtime > remaining)
3705 runtime = remaining;
3706 remaining -= runtime;
3708 cfs_rq->runtime_remaining += runtime;
3709 cfs_rq->runtime_expires = expires;
3711 /* we check whether we're throttled above */
3712 if (cfs_rq->runtime_remaining > 0)
3713 unthrottle_cfs_rq(cfs_rq);
3716 raw_spin_unlock(&rq->lock);
3723 return starting_runtime - remaining;
3727 * Responsible for refilling a task_group's bandwidth and unthrottling its
3728 * cfs_rqs as appropriate. If there has been no activity within the last
3729 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3730 * used to track this state.
3732 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3734 u64 runtime, runtime_expires;
3737 /* no need to continue the timer with no bandwidth constraint */
3738 if (cfs_b->quota == RUNTIME_INF)
3739 goto out_deactivate;
3741 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3742 cfs_b->nr_periods += overrun;
3745 * idle depends on !throttled (for the case of a large deficit), and if
3746 * we're going inactive then everything else can be deferred
3748 if (cfs_b->idle && !throttled)
3749 goto out_deactivate;
3751 __refill_cfs_bandwidth_runtime(cfs_b);
3754 /* mark as potentially idle for the upcoming period */
3759 /* account preceding periods in which throttling occurred */
3760 cfs_b->nr_throttled += overrun;
3762 runtime_expires = cfs_b->runtime_expires;
3765 * This check is repeated as we are holding onto the new bandwidth while
3766 * we unthrottle. This can potentially race with an unthrottled group
3767 * trying to acquire new bandwidth from the global pool. This can result
3768 * in us over-using our runtime if it is all used during this loop, but
3769 * only by limited amounts in that extreme case.
3771 while (throttled && cfs_b->runtime > 0) {
3772 runtime = cfs_b->runtime;
3773 raw_spin_unlock(&cfs_b->lock);
3774 /* we can't nest cfs_b->lock while distributing bandwidth */
3775 runtime = distribute_cfs_runtime(cfs_b, runtime,
3777 raw_spin_lock(&cfs_b->lock);
3779 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3781 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3785 * While we are ensured activity in the period following an
3786 * unthrottle, this also covers the case in which the new bandwidth is
3787 * insufficient to cover the existing bandwidth deficit. (Forcing the
3788 * timer to remain active while there are any throttled entities.)
3798 /* a cfs_rq won't donate quota below this amount */
3799 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3800 /* minimum remaining period time to redistribute slack quota */
3801 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3802 /* how long we wait to gather additional slack before distributing */
3803 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3806 * Are we near the end of the current quota period?
3808 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3809 * hrtimer base being cleared by hrtimer_start. In the case of
3810 * migrate_hrtimers, base is never cleared, so we are fine.
3812 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3814 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3817 /* if the call-back is running a quota refresh is already occurring */
3818 if (hrtimer_callback_running(refresh_timer))
3821 /* is a quota refresh about to occur? */
3822 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3823 if (remaining < min_expire)
3829 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3831 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3833 /* if there's a quota refresh soon don't bother with slack */
3834 if (runtime_refresh_within(cfs_b, min_left))
3837 hrtimer_start(&cfs_b->slack_timer,
3838 ns_to_ktime(cfs_bandwidth_slack_period),
3842 /* we know any runtime found here is valid as update_curr() precedes return */
3843 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3845 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3846 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3848 if (slack_runtime <= 0)
3851 raw_spin_lock(&cfs_b->lock);
3852 if (cfs_b->quota != RUNTIME_INF &&
3853 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3854 cfs_b->runtime += slack_runtime;
3856 /* we are under rq->lock, defer unthrottling using a timer */
3857 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3858 !list_empty(&cfs_b->throttled_cfs_rq))
3859 start_cfs_slack_bandwidth(cfs_b);
3861 raw_spin_unlock(&cfs_b->lock);
3863 /* even if it's not valid for return we don't want to try again */
3864 cfs_rq->runtime_remaining -= slack_runtime;
3867 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3869 if (!cfs_bandwidth_used())
3872 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3875 __return_cfs_rq_runtime(cfs_rq);
3879 * This is done with a timer (instead of inline with bandwidth return) since
3880 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3882 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3884 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3887 /* confirm we're still not at a refresh boundary */
3888 raw_spin_lock(&cfs_b->lock);
3889 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3890 raw_spin_unlock(&cfs_b->lock);
3894 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3895 runtime = cfs_b->runtime;
3897 expires = cfs_b->runtime_expires;
3898 raw_spin_unlock(&cfs_b->lock);
3903 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3905 raw_spin_lock(&cfs_b->lock);
3906 if (expires == cfs_b->runtime_expires)
3907 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3908 raw_spin_unlock(&cfs_b->lock);
3912 * When a group wakes up we want to make sure that its quota is not already
3913 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3914 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3916 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3918 if (!cfs_bandwidth_used())
3921 /* an active group must be handled by the update_curr()->put() path */
3922 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3925 /* ensure the group is not already throttled */
3926 if (cfs_rq_throttled(cfs_rq))
3929 /* update runtime allocation */
3930 account_cfs_rq_runtime(cfs_rq, 0);
3931 if (cfs_rq->runtime_remaining <= 0)
3932 throttle_cfs_rq(cfs_rq);
3935 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3936 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3938 if (!cfs_bandwidth_used())
3941 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3945 * it's possible for a throttled entity to be forced into a running
3946 * state (e.g. set_curr_task), in this case we're finished.
3948 if (cfs_rq_throttled(cfs_rq))
3951 throttle_cfs_rq(cfs_rq);
3955 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3957 struct cfs_bandwidth *cfs_b =
3958 container_of(timer, struct cfs_bandwidth, slack_timer);
3960 do_sched_cfs_slack_timer(cfs_b);
3962 return HRTIMER_NORESTART;
3965 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3967 struct cfs_bandwidth *cfs_b =
3968 container_of(timer, struct cfs_bandwidth, period_timer);
3972 raw_spin_lock(&cfs_b->lock);
3974 overrun = hrtimer_forward_now(timer, cfs_b->period);
3978 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3981 cfs_b->period_active = 0;
3982 raw_spin_unlock(&cfs_b->lock);
3984 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3987 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3989 raw_spin_lock_init(&cfs_b->lock);
3991 cfs_b->quota = RUNTIME_INF;
3992 cfs_b->period = ns_to_ktime(default_cfs_period());
3994 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3995 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
3996 cfs_b->period_timer.function = sched_cfs_period_timer;
3997 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3998 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4001 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4003 cfs_rq->runtime_enabled = 0;
4004 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4007 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4009 lockdep_assert_held(&cfs_b->lock);
4011 if (!cfs_b->period_active) {
4012 cfs_b->period_active = 1;
4013 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4014 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4018 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4020 /* init_cfs_bandwidth() was not called */
4021 if (!cfs_b->throttled_cfs_rq.next)
4024 hrtimer_cancel(&cfs_b->period_timer);
4025 hrtimer_cancel(&cfs_b->slack_timer);
4028 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4030 struct cfs_rq *cfs_rq;
4032 for_each_leaf_cfs_rq(rq, cfs_rq) {
4033 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4035 raw_spin_lock(&cfs_b->lock);
4036 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4037 raw_spin_unlock(&cfs_b->lock);
4041 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4043 struct cfs_rq *cfs_rq;
4045 for_each_leaf_cfs_rq(rq, cfs_rq) {
4046 if (!cfs_rq->runtime_enabled)
4050 * clock_task is not advancing so we just need to make sure
4051 * there's some valid quota amount
4053 cfs_rq->runtime_remaining = 1;
4055 * Offline rq is schedulable till cpu is completely disabled
4056 * in take_cpu_down(), so we prevent new cfs throttling here.
4058 cfs_rq->runtime_enabled = 0;
4060 if (cfs_rq_throttled(cfs_rq))
4061 unthrottle_cfs_rq(cfs_rq);
4065 #else /* CONFIG_CFS_BANDWIDTH */
4066 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4068 return rq_clock_task(rq_of(cfs_rq));
4071 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4072 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4073 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4074 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4076 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4081 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4086 static inline int throttled_lb_pair(struct task_group *tg,
4087 int src_cpu, int dest_cpu)
4092 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4094 #ifdef CONFIG_FAIR_GROUP_SCHED
4095 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4098 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4102 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4103 static inline void update_runtime_enabled(struct rq *rq) {}
4104 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4106 #endif /* CONFIG_CFS_BANDWIDTH */
4108 /**************************************************
4109 * CFS operations on tasks:
4112 #ifdef CONFIG_SCHED_HRTICK
4113 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4115 struct sched_entity *se = &p->se;
4116 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4118 WARN_ON(task_rq(p) != rq);
4120 if (cfs_rq->nr_running > 1) {
4121 u64 slice = sched_slice(cfs_rq, se);
4122 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4123 s64 delta = slice - ran;
4130 hrtick_start(rq, delta);
4135 * called from enqueue/dequeue and updates the hrtick when the
4136 * current task is from our class and nr_running is low enough
4139 static void hrtick_update(struct rq *rq)
4141 struct task_struct *curr = rq->curr;
4143 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4146 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4147 hrtick_start_fair(rq, curr);
4149 #else /* !CONFIG_SCHED_HRTICK */
4151 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4155 static inline void hrtick_update(struct rq *rq)
4161 * The enqueue_task method is called before nr_running is
4162 * increased. Here we update the fair scheduling stats and
4163 * then put the task into the rbtree:
4166 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4168 struct cfs_rq *cfs_rq;
4169 struct sched_entity *se = &p->se;
4171 for_each_sched_entity(se) {
4174 cfs_rq = cfs_rq_of(se);
4175 enqueue_entity(cfs_rq, se, flags);
4178 * end evaluation on encountering a throttled cfs_rq
4180 * note: in the case of encountering a throttled cfs_rq we will
4181 * post the final h_nr_running increment below.
4183 if (cfs_rq_throttled(cfs_rq))
4185 cfs_rq->h_nr_running++;
4187 flags = ENQUEUE_WAKEUP;
4190 for_each_sched_entity(se) {
4191 cfs_rq = cfs_rq_of(se);
4192 cfs_rq->h_nr_running++;
4194 if (cfs_rq_throttled(cfs_rq))
4197 update_load_avg(se, 1);
4198 update_cfs_shares(cfs_rq);
4202 add_nr_running(rq, 1);
4207 static void set_next_buddy(struct sched_entity *se);
4210 * The dequeue_task method is called before nr_running is
4211 * decreased. We remove the task from the rbtree and
4212 * update the fair scheduling stats:
4214 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4216 struct cfs_rq *cfs_rq;
4217 struct sched_entity *se = &p->se;
4218 int task_sleep = flags & DEQUEUE_SLEEP;
4220 for_each_sched_entity(se) {
4221 cfs_rq = cfs_rq_of(se);
4222 dequeue_entity(cfs_rq, se, flags);
4225 * end evaluation on encountering a throttled cfs_rq
4227 * note: in the case of encountering a throttled cfs_rq we will
4228 * post the final h_nr_running decrement below.
4230 if (cfs_rq_throttled(cfs_rq))
4232 cfs_rq->h_nr_running--;
4234 /* Don't dequeue parent if it has other entities besides us */
4235 if (cfs_rq->load.weight) {
4237 * Bias pick_next to pick a task from this cfs_rq, as
4238 * p is sleeping when it is within its sched_slice.
4240 if (task_sleep && parent_entity(se))
4241 set_next_buddy(parent_entity(se));
4243 /* avoid re-evaluating load for this entity */
4244 se = parent_entity(se);
4247 flags |= DEQUEUE_SLEEP;
4250 for_each_sched_entity(se) {
4251 cfs_rq = cfs_rq_of(se);
4252 cfs_rq->h_nr_running--;
4254 if (cfs_rq_throttled(cfs_rq))
4257 update_load_avg(se, 1);
4258 update_cfs_shares(cfs_rq);
4262 sub_nr_running(rq, 1);
4270 * per rq 'load' arrray crap; XXX kill this.
4274 * The exact cpuload at various idx values, calculated at every tick would be
4275 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4277 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4278 * on nth tick when cpu may be busy, then we have:
4279 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4280 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4282 * decay_load_missed() below does efficient calculation of
4283 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4284 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4286 * The calculation is approximated on a 128 point scale.
4287 * degrade_zero_ticks is the number of ticks after which load at any
4288 * particular idx is approximated to be zero.
4289 * degrade_factor is a precomputed table, a row for each load idx.
4290 * Each column corresponds to degradation factor for a power of two ticks,
4291 * based on 128 point scale.
4293 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4294 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4296 * With this power of 2 load factors, we can degrade the load n times
4297 * by looking at 1 bits in n and doing as many mult/shift instead of
4298 * n mult/shifts needed by the exact degradation.
4300 #define DEGRADE_SHIFT 7
4301 static const unsigned char
4302 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4303 static const unsigned char
4304 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4305 {0, 0, 0, 0, 0, 0, 0, 0},
4306 {64, 32, 8, 0, 0, 0, 0, 0},
4307 {96, 72, 40, 12, 1, 0, 0},
4308 {112, 98, 75, 43, 15, 1, 0},
4309 {120, 112, 98, 76, 45, 16, 2} };
4312 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4313 * would be when CPU is idle and so we just decay the old load without
4314 * adding any new load.
4316 static unsigned long
4317 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4321 if (!missed_updates)
4324 if (missed_updates >= degrade_zero_ticks[idx])
4328 return load >> missed_updates;
4330 while (missed_updates) {
4331 if (missed_updates % 2)
4332 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4334 missed_updates >>= 1;
4341 * Update rq->cpu_load[] statistics. This function is usually called every
4342 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4343 * every tick. We fix it up based on jiffies.
4345 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4346 unsigned long pending_updates)
4350 this_rq->nr_load_updates++;
4352 /* Update our load: */
4353 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4354 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4355 unsigned long old_load, new_load;
4357 /* scale is effectively 1 << i now, and >> i divides by scale */
4359 old_load = this_rq->cpu_load[i];
4360 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4361 new_load = this_load;
4363 * Round up the averaging division if load is increasing. This
4364 * prevents us from getting stuck on 9 if the load is 10, for
4367 if (new_load > old_load)
4368 new_load += scale - 1;
4370 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4373 sched_avg_update(this_rq);
4376 /* Used instead of source_load when we know the type == 0 */
4377 static unsigned long weighted_cpuload(const int cpu)
4379 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4382 #ifdef CONFIG_NO_HZ_COMMON
4384 * There is no sane way to deal with nohz on smp when using jiffies because the
4385 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4386 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4388 * Therefore we cannot use the delta approach from the regular tick since that
4389 * would seriously skew the load calculation. However we'll make do for those
4390 * updates happening while idle (nohz_idle_balance) or coming out of idle
4391 * (tick_nohz_idle_exit).
4393 * This means we might still be one tick off for nohz periods.
4397 * Called from nohz_idle_balance() to update the load ratings before doing the
4400 static void update_idle_cpu_load(struct rq *this_rq)
4402 unsigned long curr_jiffies = READ_ONCE(jiffies);
4403 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4404 unsigned long pending_updates;
4407 * bail if there's load or we're actually up-to-date.
4409 if (load || curr_jiffies == this_rq->last_load_update_tick)
4412 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4413 this_rq->last_load_update_tick = curr_jiffies;
4415 __update_cpu_load(this_rq, load, pending_updates);
4419 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4421 void update_cpu_load_nohz(void)
4423 struct rq *this_rq = this_rq();
4424 unsigned long curr_jiffies = READ_ONCE(jiffies);
4425 unsigned long pending_updates;
4427 if (curr_jiffies == this_rq->last_load_update_tick)
4430 raw_spin_lock(&this_rq->lock);
4431 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4432 if (pending_updates) {
4433 this_rq->last_load_update_tick = curr_jiffies;
4435 * We were idle, this means load 0, the current load might be
4436 * !0 due to remote wakeups and the sort.
4438 __update_cpu_load(this_rq, 0, pending_updates);
4440 raw_spin_unlock(&this_rq->lock);
4442 #endif /* CONFIG_NO_HZ */
4445 * Called from scheduler_tick()
4447 void update_cpu_load_active(struct rq *this_rq)
4449 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4451 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4453 this_rq->last_load_update_tick = jiffies;
4454 __update_cpu_load(this_rq, load, 1);
4458 * Return a low guess at the load of a migration-source cpu weighted
4459 * according to the scheduling class and "nice" value.
4461 * We want to under-estimate the load of migration sources, to
4462 * balance conservatively.
4464 static unsigned long source_load(int cpu, int type)
4466 struct rq *rq = cpu_rq(cpu);
4467 unsigned long total = weighted_cpuload(cpu);
4469 if (type == 0 || !sched_feat(LB_BIAS))
4472 return min(rq->cpu_load[type-1], total);
4476 * Return a high guess at the load of a migration-target cpu weighted
4477 * according to the scheduling class and "nice" value.
4479 static unsigned long target_load(int cpu, int type)
4481 struct rq *rq = cpu_rq(cpu);
4482 unsigned long total = weighted_cpuload(cpu);
4484 if (type == 0 || !sched_feat(LB_BIAS))
4487 return max(rq->cpu_load[type-1], total);
4490 static unsigned long capacity_of(int cpu)
4492 return cpu_rq(cpu)->cpu_capacity;
4495 static unsigned long capacity_orig_of(int cpu)
4497 return cpu_rq(cpu)->cpu_capacity_orig;
4500 static unsigned long cpu_avg_load_per_task(int cpu)
4502 struct rq *rq = cpu_rq(cpu);
4503 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4504 unsigned long load_avg = weighted_cpuload(cpu);
4507 return load_avg / nr_running;
4512 static void record_wakee(struct task_struct *p)
4515 * Rough decay (wiping) for cost saving, don't worry
4516 * about the boundary, really active task won't care
4519 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4520 current->wakee_flips >>= 1;
4521 current->wakee_flip_decay_ts = jiffies;
4524 if (current->last_wakee != p) {
4525 current->last_wakee = p;
4526 current->wakee_flips++;
4530 static void task_waking_fair(struct task_struct *p)
4532 struct sched_entity *se = &p->se;
4533 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4536 #ifndef CONFIG_64BIT
4537 u64 min_vruntime_copy;
4540 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4542 min_vruntime = cfs_rq->min_vruntime;
4543 } while (min_vruntime != min_vruntime_copy);
4545 min_vruntime = cfs_rq->min_vruntime;
4548 se->vruntime -= min_vruntime;
4552 #ifdef CONFIG_FAIR_GROUP_SCHED
4554 * effective_load() calculates the load change as seen from the root_task_group
4556 * Adding load to a group doesn't make a group heavier, but can cause movement
4557 * of group shares between cpus. Assuming the shares were perfectly aligned one
4558 * can calculate the shift in shares.
4560 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4561 * on this @cpu and results in a total addition (subtraction) of @wg to the
4562 * total group weight.
4564 * Given a runqueue weight distribution (rw_i) we can compute a shares
4565 * distribution (s_i) using:
4567 * s_i = rw_i / \Sum rw_j (1)
4569 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4570 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4571 * shares distribution (s_i):
4573 * rw_i = { 2, 4, 1, 0 }
4574 * s_i = { 2/7, 4/7, 1/7, 0 }
4576 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4577 * task used to run on and the CPU the waker is running on), we need to
4578 * compute the effect of waking a task on either CPU and, in case of a sync
4579 * wakeup, compute the effect of the current task going to sleep.
4581 * So for a change of @wl to the local @cpu with an overall group weight change
4582 * of @wl we can compute the new shares distribution (s'_i) using:
4584 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4586 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4587 * differences in waking a task to CPU 0. The additional task changes the
4588 * weight and shares distributions like:
4590 * rw'_i = { 3, 4, 1, 0 }
4591 * s'_i = { 3/8, 4/8, 1/8, 0 }
4593 * We can then compute the difference in effective weight by using:
4595 * dw_i = S * (s'_i - s_i) (3)
4597 * Where 'S' is the group weight as seen by its parent.
4599 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4600 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4601 * 4/7) times the weight of the group.
4603 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4605 struct sched_entity *se = tg->se[cpu];
4607 if (!tg->parent) /* the trivial, non-cgroup case */
4610 for_each_sched_entity(se) {
4611 struct cfs_rq *cfs_rq = se->my_q;
4612 long W, w = cfs_rq_load_avg(cfs_rq);
4617 * W = @wg + \Sum rw_j
4619 W = wg + atomic_long_read(&tg->load_avg);
4621 /* Ensure \Sum rw_j >= rw_i */
4622 W -= cfs_rq->tg_load_avg_contrib;
4631 * wl = S * s'_i; see (2)
4634 wl = (w * (long)tg->shares) / W;
4639 * Per the above, wl is the new se->load.weight value; since
4640 * those are clipped to [MIN_SHARES, ...) do so now. See
4641 * calc_cfs_shares().
4643 if (wl < MIN_SHARES)
4647 * wl = dw_i = S * (s'_i - s_i); see (3)
4649 wl -= se->avg.load_avg;
4652 * Recursively apply this logic to all parent groups to compute
4653 * the final effective load change on the root group. Since
4654 * only the @tg group gets extra weight, all parent groups can
4655 * only redistribute existing shares. @wl is the shift in shares
4656 * resulting from this level per the above.
4665 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4673 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4674 * A waker of many should wake a different task than the one last awakened
4675 * at a frequency roughly N times higher than one of its wakees. In order
4676 * to determine whether we should let the load spread vs consolodating to
4677 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4678 * partner, and a factor of lls_size higher frequency in the other. With
4679 * both conditions met, we can be relatively sure that the relationship is
4680 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4681 * being client/server, worker/dispatcher, interrupt source or whatever is
4682 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4684 static int wake_wide(struct task_struct *p)
4686 unsigned int master = current->wakee_flips;
4687 unsigned int slave = p->wakee_flips;
4688 int factor = this_cpu_read(sd_llc_size);
4691 swap(master, slave);
4692 if (slave < factor || master < slave * factor)
4697 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4699 s64 this_load, load;
4700 s64 this_eff_load, prev_eff_load;
4701 int idx, this_cpu, prev_cpu;
4702 struct task_group *tg;
4703 unsigned long weight;
4707 this_cpu = smp_processor_id();
4708 prev_cpu = task_cpu(p);
4709 load = source_load(prev_cpu, idx);
4710 this_load = target_load(this_cpu, idx);
4713 * If sync wakeup then subtract the (maximum possible)
4714 * effect of the currently running task from the load
4715 * of the current CPU:
4718 tg = task_group(current);
4719 weight = current->se.avg.load_avg;
4721 this_load += effective_load(tg, this_cpu, -weight, -weight);
4722 load += effective_load(tg, prev_cpu, 0, -weight);
4726 weight = p->se.avg.load_avg;
4729 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4730 * due to the sync cause above having dropped this_load to 0, we'll
4731 * always have an imbalance, but there's really nothing you can do
4732 * about that, so that's good too.
4734 * Otherwise check if either cpus are near enough in load to allow this
4735 * task to be woken on this_cpu.
4737 this_eff_load = 100;
4738 this_eff_load *= capacity_of(prev_cpu);
4740 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4741 prev_eff_load *= capacity_of(this_cpu);
4743 if (this_load > 0) {
4744 this_eff_load *= this_load +
4745 effective_load(tg, this_cpu, weight, weight);
4747 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4750 balanced = this_eff_load <= prev_eff_load;
4752 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4757 schedstat_inc(sd, ttwu_move_affine);
4758 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4764 * find_idlest_group finds and returns the least busy CPU group within the
4767 static struct sched_group *
4768 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4769 int this_cpu, int sd_flag)
4771 struct sched_group *idlest = NULL, *group = sd->groups;
4772 unsigned long min_load = ULONG_MAX, this_load = 0;
4773 int load_idx = sd->forkexec_idx;
4774 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4776 if (sd_flag & SD_BALANCE_WAKE)
4777 load_idx = sd->wake_idx;
4780 unsigned long load, avg_load;
4784 /* Skip over this group if it has no CPUs allowed */
4785 if (!cpumask_intersects(sched_group_cpus(group),
4786 tsk_cpus_allowed(p)))
4789 local_group = cpumask_test_cpu(this_cpu,
4790 sched_group_cpus(group));
4792 /* Tally up the load of all CPUs in the group */
4795 for_each_cpu(i, sched_group_cpus(group)) {
4796 /* Bias balancing toward cpus of our domain */
4798 load = source_load(i, load_idx);
4800 load = target_load(i, load_idx);
4805 /* Adjust by relative CPU capacity of the group */
4806 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4809 this_load = avg_load;
4810 } else if (avg_load < min_load) {
4811 min_load = avg_load;
4814 } while (group = group->next, group != sd->groups);
4816 if (!idlest || 100*this_load < imbalance*min_load)
4822 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4825 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4827 unsigned long load, min_load = ULONG_MAX;
4828 unsigned int min_exit_latency = UINT_MAX;
4829 u64 latest_idle_timestamp = 0;
4830 int least_loaded_cpu = this_cpu;
4831 int shallowest_idle_cpu = -1;
4834 /* Traverse only the allowed CPUs */
4835 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4837 struct rq *rq = cpu_rq(i);
4838 struct cpuidle_state *idle = idle_get_state(rq);
4839 if (idle && idle->exit_latency < min_exit_latency) {
4841 * We give priority to a CPU whose idle state
4842 * has the smallest exit latency irrespective
4843 * of any idle timestamp.
4845 min_exit_latency = idle->exit_latency;
4846 latest_idle_timestamp = rq->idle_stamp;
4847 shallowest_idle_cpu = i;
4848 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4849 rq->idle_stamp > latest_idle_timestamp) {
4851 * If equal or no active idle state, then
4852 * the most recently idled CPU might have
4855 latest_idle_timestamp = rq->idle_stamp;
4856 shallowest_idle_cpu = i;
4858 } else if (shallowest_idle_cpu == -1) {
4859 load = weighted_cpuload(i);
4860 if (load < min_load || (load == min_load && i == this_cpu)) {
4862 least_loaded_cpu = i;
4867 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4871 * Try and locate an idle CPU in the sched_domain.
4873 static int select_idle_sibling(struct task_struct *p, int target)
4875 struct sched_domain *sd;
4876 struct sched_group *sg;
4877 int i = task_cpu(p);
4879 if (idle_cpu(target))
4883 * If the prevous cpu is cache affine and idle, don't be stupid.
4885 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4889 * Otherwise, iterate the domains and find an elegible idle cpu.
4891 sd = rcu_dereference(per_cpu(sd_llc, target));
4892 for_each_lower_domain(sd) {
4895 if (!cpumask_intersects(sched_group_cpus(sg),
4896 tsk_cpus_allowed(p)))
4899 for_each_cpu(i, sched_group_cpus(sg)) {
4900 if (i == target || !idle_cpu(i))
4904 target = cpumask_first_and(sched_group_cpus(sg),
4905 tsk_cpus_allowed(p));
4909 } while (sg != sd->groups);
4916 * cpu_util returns the amount of capacity of a CPU that is used by CFS
4917 * tasks. The unit of the return value must be the one of capacity so we can
4918 * compare the utilization with the capacity of the CPU that is available for
4919 * CFS task (ie cpu_capacity).
4921 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
4922 * recent utilization of currently non-runnable tasks on a CPU. It represents
4923 * the amount of utilization of a CPU in the range [0..capacity_orig] where
4924 * capacity_orig is the cpu_capacity available at the highest frequency
4925 * (arch_scale_freq_capacity()).
4926 * The utilization of a CPU converges towards a sum equal to or less than the
4927 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
4928 * the running time on this CPU scaled by capacity_curr.
4930 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
4931 * higher than capacity_orig because of unfortunate rounding in
4932 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
4933 * the average stabilizes with the new running time. We need to check that the
4934 * utilization stays within the range of [0..capacity_orig] and cap it if
4935 * necessary. Without utilization capping, a group could be seen as overloaded
4936 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
4937 * available capacity. We allow utilization to overshoot capacity_curr (but not
4938 * capacity_orig) as it useful for predicting the capacity required after task
4939 * migrations (scheduler-driven DVFS).
4941 static int cpu_util(int cpu)
4943 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
4944 unsigned long capacity = capacity_orig_of(cpu);
4946 return (util >= capacity) ? capacity : util;
4950 * select_task_rq_fair: Select target runqueue for the waking task in domains
4951 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4952 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4954 * Balances load by selecting the idlest cpu in the idlest group, or under
4955 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4957 * Returns the target cpu number.
4959 * preempt must be disabled.
4962 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4964 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4965 int cpu = smp_processor_id();
4966 int new_cpu = prev_cpu;
4967 int want_affine = 0;
4968 int sync = wake_flags & WF_SYNC;
4970 if (sd_flag & SD_BALANCE_WAKE)
4971 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4974 for_each_domain(cpu, tmp) {
4975 if (!(tmp->flags & SD_LOAD_BALANCE))
4979 * If both cpu and prev_cpu are part of this domain,
4980 * cpu is a valid SD_WAKE_AFFINE target.
4982 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4983 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4988 if (tmp->flags & sd_flag)
4990 else if (!want_affine)
4995 sd = NULL; /* Prefer wake_affine over balance flags */
4996 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5001 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5002 new_cpu = select_idle_sibling(p, new_cpu);
5005 struct sched_group *group;
5008 if (!(sd->flags & sd_flag)) {
5013 group = find_idlest_group(sd, p, cpu, sd_flag);
5019 new_cpu = find_idlest_cpu(group, p, cpu);
5020 if (new_cpu == -1 || new_cpu == cpu) {
5021 /* Now try balancing at a lower domain level of cpu */
5026 /* Now try balancing at a lower domain level of new_cpu */
5028 weight = sd->span_weight;
5030 for_each_domain(cpu, tmp) {
5031 if (weight <= tmp->span_weight)
5033 if (tmp->flags & sd_flag)
5036 /* while loop will break here if sd == NULL */
5044 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5045 * cfs_rq_of(p) references at time of call are still valid and identify the
5046 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5047 * other assumptions, including the state of rq->lock, should be made.
5049 static void migrate_task_rq_fair(struct task_struct *p)
5052 * We are supposed to update the task to "current" time, then its up to date
5053 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5054 * what current time is, so simply throw away the out-of-date time. This
5055 * will result in the wakee task is less decayed, but giving the wakee more
5056 * load sounds not bad.
5058 remove_entity_load_avg(&p->se);
5060 /* Tell new CPU we are migrated */
5061 p->se.avg.last_update_time = 0;
5063 /* We have migrated, no longer consider this task hot */
5064 p->se.exec_start = 0;
5067 static void task_dead_fair(struct task_struct *p)
5069 remove_entity_load_avg(&p->se);
5071 #endif /* CONFIG_SMP */
5073 static unsigned long
5074 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5076 unsigned long gran = sysctl_sched_wakeup_granularity;
5079 * Since its curr running now, convert the gran from real-time
5080 * to virtual-time in his units.
5082 * By using 'se' instead of 'curr' we penalize light tasks, so
5083 * they get preempted easier. That is, if 'se' < 'curr' then
5084 * the resulting gran will be larger, therefore penalizing the
5085 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5086 * be smaller, again penalizing the lighter task.
5088 * This is especially important for buddies when the leftmost
5089 * task is higher priority than the buddy.
5091 return calc_delta_fair(gran, se);
5095 * Should 'se' preempt 'curr'.
5109 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5111 s64 gran, vdiff = curr->vruntime - se->vruntime;
5116 gran = wakeup_gran(curr, se);
5123 static void set_last_buddy(struct sched_entity *se)
5125 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5128 for_each_sched_entity(se)
5129 cfs_rq_of(se)->last = se;
5132 static void set_next_buddy(struct sched_entity *se)
5134 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5137 for_each_sched_entity(se)
5138 cfs_rq_of(se)->next = se;
5141 static void set_skip_buddy(struct sched_entity *se)
5143 for_each_sched_entity(se)
5144 cfs_rq_of(se)->skip = se;
5148 * Preempt the current task with a newly woken task if needed:
5150 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5152 struct task_struct *curr = rq->curr;
5153 struct sched_entity *se = &curr->se, *pse = &p->se;
5154 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5155 int scale = cfs_rq->nr_running >= sched_nr_latency;
5156 int next_buddy_marked = 0;
5158 if (unlikely(se == pse))
5162 * This is possible from callers such as attach_tasks(), in which we
5163 * unconditionally check_prempt_curr() after an enqueue (which may have
5164 * lead to a throttle). This both saves work and prevents false
5165 * next-buddy nomination below.
5167 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5170 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5171 set_next_buddy(pse);
5172 next_buddy_marked = 1;
5176 * We can come here with TIF_NEED_RESCHED already set from new task
5179 * Note: this also catches the edge-case of curr being in a throttled
5180 * group (e.g. via set_curr_task), since update_curr() (in the
5181 * enqueue of curr) will have resulted in resched being set. This
5182 * prevents us from potentially nominating it as a false LAST_BUDDY
5185 if (test_tsk_need_resched(curr))
5188 /* Idle tasks are by definition preempted by non-idle tasks. */
5189 if (unlikely(curr->policy == SCHED_IDLE) &&
5190 likely(p->policy != SCHED_IDLE))
5194 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5195 * is driven by the tick):
5197 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5200 find_matching_se(&se, &pse);
5201 update_curr(cfs_rq_of(se));
5203 if (wakeup_preempt_entity(se, pse) == 1) {
5205 * Bias pick_next to pick the sched entity that is
5206 * triggering this preemption.
5208 if (!next_buddy_marked)
5209 set_next_buddy(pse);
5218 * Only set the backward buddy when the current task is still
5219 * on the rq. This can happen when a wakeup gets interleaved
5220 * with schedule on the ->pre_schedule() or idle_balance()
5221 * point, either of which can * drop the rq lock.
5223 * Also, during early boot the idle thread is in the fair class,
5224 * for obvious reasons its a bad idea to schedule back to it.
5226 if (unlikely(!se->on_rq || curr == rq->idle))
5229 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5233 static struct task_struct *
5234 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5236 struct cfs_rq *cfs_rq = &rq->cfs;
5237 struct sched_entity *se;
5238 struct task_struct *p;
5242 #ifdef CONFIG_FAIR_GROUP_SCHED
5243 if (!cfs_rq->nr_running)
5246 if (prev->sched_class != &fair_sched_class)
5250 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5251 * likely that a next task is from the same cgroup as the current.
5253 * Therefore attempt to avoid putting and setting the entire cgroup
5254 * hierarchy, only change the part that actually changes.
5258 struct sched_entity *curr = cfs_rq->curr;
5261 * Since we got here without doing put_prev_entity() we also
5262 * have to consider cfs_rq->curr. If it is still a runnable
5263 * entity, update_curr() will update its vruntime, otherwise
5264 * forget we've ever seen it.
5268 update_curr(cfs_rq);
5273 * This call to check_cfs_rq_runtime() will do the
5274 * throttle and dequeue its entity in the parent(s).
5275 * Therefore the 'simple' nr_running test will indeed
5278 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5282 se = pick_next_entity(cfs_rq, curr);
5283 cfs_rq = group_cfs_rq(se);
5289 * Since we haven't yet done put_prev_entity and if the selected task
5290 * is a different task than we started out with, try and touch the
5291 * least amount of cfs_rqs.
5294 struct sched_entity *pse = &prev->se;
5296 while (!(cfs_rq = is_same_group(se, pse))) {
5297 int se_depth = se->depth;
5298 int pse_depth = pse->depth;
5300 if (se_depth <= pse_depth) {
5301 put_prev_entity(cfs_rq_of(pse), pse);
5302 pse = parent_entity(pse);
5304 if (se_depth >= pse_depth) {
5305 set_next_entity(cfs_rq_of(se), se);
5306 se = parent_entity(se);
5310 put_prev_entity(cfs_rq, pse);
5311 set_next_entity(cfs_rq, se);
5314 if (hrtick_enabled(rq))
5315 hrtick_start_fair(rq, p);
5322 if (!cfs_rq->nr_running)
5325 put_prev_task(rq, prev);
5328 se = pick_next_entity(cfs_rq, NULL);
5329 set_next_entity(cfs_rq, se);
5330 cfs_rq = group_cfs_rq(se);
5335 if (hrtick_enabled(rq))
5336 hrtick_start_fair(rq, p);
5342 * This is OK, because current is on_cpu, which avoids it being picked
5343 * for load-balance and preemption/IRQs are still disabled avoiding
5344 * further scheduler activity on it and we're being very careful to
5345 * re-start the picking loop.
5347 lockdep_unpin_lock(&rq->lock);
5348 new_tasks = idle_balance(rq);
5349 lockdep_pin_lock(&rq->lock);
5351 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5352 * possible for any higher priority task to appear. In that case we
5353 * must re-start the pick_next_entity() loop.
5365 * Account for a descheduled task:
5367 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5369 struct sched_entity *se = &prev->se;
5370 struct cfs_rq *cfs_rq;
5372 for_each_sched_entity(se) {
5373 cfs_rq = cfs_rq_of(se);
5374 put_prev_entity(cfs_rq, se);
5379 * sched_yield() is very simple
5381 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5383 static void yield_task_fair(struct rq *rq)
5385 struct task_struct *curr = rq->curr;
5386 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5387 struct sched_entity *se = &curr->se;
5390 * Are we the only task in the tree?
5392 if (unlikely(rq->nr_running == 1))
5395 clear_buddies(cfs_rq, se);
5397 if (curr->policy != SCHED_BATCH) {
5398 update_rq_clock(rq);
5400 * Update run-time statistics of the 'current'.
5402 update_curr(cfs_rq);
5404 * Tell update_rq_clock() that we've just updated,
5405 * so we don't do microscopic update in schedule()
5406 * and double the fastpath cost.
5408 rq_clock_skip_update(rq, true);
5414 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5416 struct sched_entity *se = &p->se;
5418 /* throttled hierarchies are not runnable */
5419 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5422 /* Tell the scheduler that we'd really like pse to run next. */
5425 yield_task_fair(rq);
5431 /**************************************************
5432 * Fair scheduling class load-balancing methods.
5436 * The purpose of load-balancing is to achieve the same basic fairness the
5437 * per-cpu scheduler provides, namely provide a proportional amount of compute
5438 * time to each task. This is expressed in the following equation:
5440 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5442 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5443 * W_i,0 is defined as:
5445 * W_i,0 = \Sum_j w_i,j (2)
5447 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5448 * is derived from the nice value as per prio_to_weight[].
5450 * The weight average is an exponential decay average of the instantaneous
5453 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5455 * C_i is the compute capacity of cpu i, typically it is the
5456 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5457 * can also include other factors [XXX].
5459 * To achieve this balance we define a measure of imbalance which follows
5460 * directly from (1):
5462 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5464 * We them move tasks around to minimize the imbalance. In the continuous
5465 * function space it is obvious this converges, in the discrete case we get
5466 * a few fun cases generally called infeasible weight scenarios.
5469 * - infeasible weights;
5470 * - local vs global optima in the discrete case. ]
5475 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5476 * for all i,j solution, we create a tree of cpus that follows the hardware
5477 * topology where each level pairs two lower groups (or better). This results
5478 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5479 * tree to only the first of the previous level and we decrease the frequency
5480 * of load-balance at each level inv. proportional to the number of cpus in
5486 * \Sum { --- * --- * 2^i } = O(n) (5)
5488 * `- size of each group
5489 * | | `- number of cpus doing load-balance
5491 * `- sum over all levels
5493 * Coupled with a limit on how many tasks we can migrate every balance pass,
5494 * this makes (5) the runtime complexity of the balancer.
5496 * An important property here is that each CPU is still (indirectly) connected
5497 * to every other cpu in at most O(log n) steps:
5499 * The adjacency matrix of the resulting graph is given by:
5502 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5505 * And you'll find that:
5507 * A^(log_2 n)_i,j != 0 for all i,j (7)
5509 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5510 * The task movement gives a factor of O(m), giving a convergence complexity
5513 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5518 * In order to avoid CPUs going idle while there's still work to do, new idle
5519 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5520 * tree itself instead of relying on other CPUs to bring it work.
5522 * This adds some complexity to both (5) and (8) but it reduces the total idle
5530 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5533 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5538 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5540 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5542 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5545 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5546 * rewrite all of this once again.]
5549 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5551 enum fbq_type { regular, remote, all };
5553 #define LBF_ALL_PINNED 0x01
5554 #define LBF_NEED_BREAK 0x02
5555 #define LBF_DST_PINNED 0x04
5556 #define LBF_SOME_PINNED 0x08
5559 struct sched_domain *sd;
5567 struct cpumask *dst_grpmask;
5569 enum cpu_idle_type idle;
5571 /* The set of CPUs under consideration for load-balancing */
5572 struct cpumask *cpus;
5577 unsigned int loop_break;
5578 unsigned int loop_max;
5580 enum fbq_type fbq_type;
5581 struct list_head tasks;
5585 * Is this task likely cache-hot:
5587 static int task_hot(struct task_struct *p, struct lb_env *env)
5591 lockdep_assert_held(&env->src_rq->lock);
5593 if (p->sched_class != &fair_sched_class)
5596 if (unlikely(p->policy == SCHED_IDLE))
5600 * Buddy candidates are cache hot:
5602 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5603 (&p->se == cfs_rq_of(&p->se)->next ||
5604 &p->se == cfs_rq_of(&p->se)->last))
5607 if (sysctl_sched_migration_cost == -1)
5609 if (sysctl_sched_migration_cost == 0)
5612 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5614 return delta < (s64)sysctl_sched_migration_cost;
5617 #ifdef CONFIG_NUMA_BALANCING
5619 * Returns 1, if task migration degrades locality
5620 * Returns 0, if task migration improves locality i.e migration preferred.
5621 * Returns -1, if task migration is not affected by locality.
5623 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5625 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5626 unsigned long src_faults, dst_faults;
5627 int src_nid, dst_nid;
5629 if (!static_branch_likely(&sched_numa_balancing))
5632 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5635 src_nid = cpu_to_node(env->src_cpu);
5636 dst_nid = cpu_to_node(env->dst_cpu);
5638 if (src_nid == dst_nid)
5641 /* Migrating away from the preferred node is always bad. */
5642 if (src_nid == p->numa_preferred_nid) {
5643 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5649 /* Encourage migration to the preferred node. */
5650 if (dst_nid == p->numa_preferred_nid)
5654 src_faults = group_faults(p, src_nid);
5655 dst_faults = group_faults(p, dst_nid);
5657 src_faults = task_faults(p, src_nid);
5658 dst_faults = task_faults(p, dst_nid);
5661 return dst_faults < src_faults;
5665 static inline int migrate_degrades_locality(struct task_struct *p,
5673 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5676 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5680 lockdep_assert_held(&env->src_rq->lock);
5683 * We do not migrate tasks that are:
5684 * 1) throttled_lb_pair, or
5685 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5686 * 3) running (obviously), or
5687 * 4) are cache-hot on their current CPU.
5689 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5692 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5695 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5697 env->flags |= LBF_SOME_PINNED;
5700 * Remember if this task can be migrated to any other cpu in
5701 * our sched_group. We may want to revisit it if we couldn't
5702 * meet load balance goals by pulling other tasks on src_cpu.
5704 * Also avoid computing new_dst_cpu if we have already computed
5705 * one in current iteration.
5707 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5710 /* Prevent to re-select dst_cpu via env's cpus */
5711 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5712 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5713 env->flags |= LBF_DST_PINNED;
5714 env->new_dst_cpu = cpu;
5722 /* Record that we found atleast one task that could run on dst_cpu */
5723 env->flags &= ~LBF_ALL_PINNED;
5725 if (task_running(env->src_rq, p)) {
5726 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5731 * Aggressive migration if:
5732 * 1) destination numa is preferred
5733 * 2) task is cache cold, or
5734 * 3) too many balance attempts have failed.
5736 tsk_cache_hot = migrate_degrades_locality(p, env);
5737 if (tsk_cache_hot == -1)
5738 tsk_cache_hot = task_hot(p, env);
5740 if (tsk_cache_hot <= 0 ||
5741 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5742 if (tsk_cache_hot == 1) {
5743 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5744 schedstat_inc(p, se.statistics.nr_forced_migrations);
5749 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5754 * detach_task() -- detach the task for the migration specified in env
5756 static void detach_task(struct task_struct *p, struct lb_env *env)
5758 lockdep_assert_held(&env->src_rq->lock);
5760 deactivate_task(env->src_rq, p, 0);
5761 p->on_rq = TASK_ON_RQ_MIGRATING;
5762 set_task_cpu(p, env->dst_cpu);
5766 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5767 * part of active balancing operations within "domain".
5769 * Returns a task if successful and NULL otherwise.
5771 static struct task_struct *detach_one_task(struct lb_env *env)
5773 struct task_struct *p, *n;
5775 lockdep_assert_held(&env->src_rq->lock);
5777 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5778 if (!can_migrate_task(p, env))
5781 detach_task(p, env);
5784 * Right now, this is only the second place where
5785 * lb_gained[env->idle] is updated (other is detach_tasks)
5786 * so we can safely collect stats here rather than
5787 * inside detach_tasks().
5789 schedstat_inc(env->sd, lb_gained[env->idle]);
5795 static const unsigned int sched_nr_migrate_break = 32;
5798 * detach_tasks() -- tries to detach up to imbalance weighted load from
5799 * busiest_rq, as part of a balancing operation within domain "sd".
5801 * Returns number of detached tasks if successful and 0 otherwise.
5803 static int detach_tasks(struct lb_env *env)
5805 struct list_head *tasks = &env->src_rq->cfs_tasks;
5806 struct task_struct *p;
5810 lockdep_assert_held(&env->src_rq->lock);
5812 if (env->imbalance <= 0)
5815 while (!list_empty(tasks)) {
5817 * We don't want to steal all, otherwise we may be treated likewise,
5818 * which could at worst lead to a livelock crash.
5820 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5823 p = list_first_entry(tasks, struct task_struct, se.group_node);
5826 /* We've more or less seen every task there is, call it quits */
5827 if (env->loop > env->loop_max)
5830 /* take a breather every nr_migrate tasks */
5831 if (env->loop > env->loop_break) {
5832 env->loop_break += sched_nr_migrate_break;
5833 env->flags |= LBF_NEED_BREAK;
5837 if (!can_migrate_task(p, env))
5840 load = task_h_load(p);
5842 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5845 if ((load / 2) > env->imbalance)
5848 detach_task(p, env);
5849 list_add(&p->se.group_node, &env->tasks);
5852 env->imbalance -= load;
5854 #ifdef CONFIG_PREEMPT
5856 * NEWIDLE balancing is a source of latency, so preemptible
5857 * kernels will stop after the first task is detached to minimize
5858 * the critical section.
5860 if (env->idle == CPU_NEWLY_IDLE)
5865 * We only want to steal up to the prescribed amount of
5868 if (env->imbalance <= 0)
5873 list_move_tail(&p->se.group_node, tasks);
5877 * Right now, this is one of only two places we collect this stat
5878 * so we can safely collect detach_one_task() stats here rather
5879 * than inside detach_one_task().
5881 schedstat_add(env->sd, lb_gained[env->idle], detached);
5887 * attach_task() -- attach the task detached by detach_task() to its new rq.
5889 static void attach_task(struct rq *rq, struct task_struct *p)
5891 lockdep_assert_held(&rq->lock);
5893 BUG_ON(task_rq(p) != rq);
5894 p->on_rq = TASK_ON_RQ_QUEUED;
5895 activate_task(rq, p, 0);
5896 check_preempt_curr(rq, p, 0);
5900 * attach_one_task() -- attaches the task returned from detach_one_task() to
5903 static void attach_one_task(struct rq *rq, struct task_struct *p)
5905 raw_spin_lock(&rq->lock);
5907 raw_spin_unlock(&rq->lock);
5911 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5914 static void attach_tasks(struct lb_env *env)
5916 struct list_head *tasks = &env->tasks;
5917 struct task_struct *p;
5919 raw_spin_lock(&env->dst_rq->lock);
5921 while (!list_empty(tasks)) {
5922 p = list_first_entry(tasks, struct task_struct, se.group_node);
5923 list_del_init(&p->se.group_node);
5925 attach_task(env->dst_rq, p);
5928 raw_spin_unlock(&env->dst_rq->lock);
5931 #ifdef CONFIG_FAIR_GROUP_SCHED
5932 static void update_blocked_averages(int cpu)
5934 struct rq *rq = cpu_rq(cpu);
5935 struct cfs_rq *cfs_rq;
5936 unsigned long flags;
5938 raw_spin_lock_irqsave(&rq->lock, flags);
5939 update_rq_clock(rq);
5942 * Iterates the task_group tree in a bottom up fashion, see
5943 * list_add_leaf_cfs_rq() for details.
5945 for_each_leaf_cfs_rq(rq, cfs_rq) {
5946 /* throttled entities do not contribute to load */
5947 if (throttled_hierarchy(cfs_rq))
5950 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
5951 update_tg_load_avg(cfs_rq, 0);
5953 raw_spin_unlock_irqrestore(&rq->lock, flags);
5957 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5958 * This needs to be done in a top-down fashion because the load of a child
5959 * group is a fraction of its parents load.
5961 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5963 struct rq *rq = rq_of(cfs_rq);
5964 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5965 unsigned long now = jiffies;
5968 if (cfs_rq->last_h_load_update == now)
5971 cfs_rq->h_load_next = NULL;
5972 for_each_sched_entity(se) {
5973 cfs_rq = cfs_rq_of(se);
5974 cfs_rq->h_load_next = se;
5975 if (cfs_rq->last_h_load_update == now)
5980 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
5981 cfs_rq->last_h_load_update = now;
5984 while ((se = cfs_rq->h_load_next) != NULL) {
5985 load = cfs_rq->h_load;
5986 load = div64_ul(load * se->avg.load_avg,
5987 cfs_rq_load_avg(cfs_rq) + 1);
5988 cfs_rq = group_cfs_rq(se);
5989 cfs_rq->h_load = load;
5990 cfs_rq->last_h_load_update = now;
5994 static unsigned long task_h_load(struct task_struct *p)
5996 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5998 update_cfs_rq_h_load(cfs_rq);
5999 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6000 cfs_rq_load_avg(cfs_rq) + 1);
6003 static inline void update_blocked_averages(int cpu)
6005 struct rq *rq = cpu_rq(cpu);
6006 struct cfs_rq *cfs_rq = &rq->cfs;
6007 unsigned long flags;
6009 raw_spin_lock_irqsave(&rq->lock, flags);
6010 update_rq_clock(rq);
6011 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6012 raw_spin_unlock_irqrestore(&rq->lock, flags);
6015 static unsigned long task_h_load(struct task_struct *p)
6017 return p->se.avg.load_avg;
6021 /********** Helpers for find_busiest_group ************************/
6030 * sg_lb_stats - stats of a sched_group required for load_balancing
6032 struct sg_lb_stats {
6033 unsigned long avg_load; /*Avg load across the CPUs of the group */
6034 unsigned long group_load; /* Total load over the CPUs of the group */
6035 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6036 unsigned long load_per_task;
6037 unsigned long group_capacity;
6038 unsigned long group_util; /* Total utilization of the group */
6039 unsigned int sum_nr_running; /* Nr tasks running in the group */
6040 unsigned int idle_cpus;
6041 unsigned int group_weight;
6042 enum group_type group_type;
6043 int group_no_capacity;
6044 #ifdef CONFIG_NUMA_BALANCING
6045 unsigned int nr_numa_running;
6046 unsigned int nr_preferred_running;
6051 * sd_lb_stats - Structure to store the statistics of a sched_domain
6052 * during load balancing.
6054 struct sd_lb_stats {
6055 struct sched_group *busiest; /* Busiest group in this sd */
6056 struct sched_group *local; /* Local group in this sd */
6057 unsigned long total_load; /* Total load of all groups in sd */
6058 unsigned long total_capacity; /* Total capacity of all groups in sd */
6059 unsigned long avg_load; /* Average load across all groups in sd */
6061 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6062 struct sg_lb_stats local_stat; /* Statistics of the local group */
6065 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6068 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6069 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6070 * We must however clear busiest_stat::avg_load because
6071 * update_sd_pick_busiest() reads this before assignment.
6073 *sds = (struct sd_lb_stats){
6077 .total_capacity = 0UL,
6080 .sum_nr_running = 0,
6081 .group_type = group_other,
6087 * get_sd_load_idx - Obtain the load index for a given sched domain.
6088 * @sd: The sched_domain whose load_idx is to be obtained.
6089 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6091 * Return: The load index.
6093 static inline int get_sd_load_idx(struct sched_domain *sd,
6094 enum cpu_idle_type idle)
6100 load_idx = sd->busy_idx;
6103 case CPU_NEWLY_IDLE:
6104 load_idx = sd->newidle_idx;
6107 load_idx = sd->idle_idx;
6114 static unsigned long scale_rt_capacity(int cpu)
6116 struct rq *rq = cpu_rq(cpu);
6117 u64 total, used, age_stamp, avg;
6121 * Since we're reading these variables without serialization make sure
6122 * we read them once before doing sanity checks on them.
6124 age_stamp = READ_ONCE(rq->age_stamp);
6125 avg = READ_ONCE(rq->rt_avg);
6126 delta = __rq_clock_broken(rq) - age_stamp;
6128 if (unlikely(delta < 0))
6131 total = sched_avg_period() + delta;
6133 used = div_u64(avg, total);
6135 if (likely(used < SCHED_CAPACITY_SCALE))
6136 return SCHED_CAPACITY_SCALE - used;
6141 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6143 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6144 struct sched_group *sdg = sd->groups;
6146 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6148 capacity *= scale_rt_capacity(cpu);
6149 capacity >>= SCHED_CAPACITY_SHIFT;
6154 cpu_rq(cpu)->cpu_capacity = capacity;
6155 sdg->sgc->capacity = capacity;
6158 void update_group_capacity(struct sched_domain *sd, int cpu)
6160 struct sched_domain *child = sd->child;
6161 struct sched_group *group, *sdg = sd->groups;
6162 unsigned long capacity;
6163 unsigned long interval;
6165 interval = msecs_to_jiffies(sd->balance_interval);
6166 interval = clamp(interval, 1UL, max_load_balance_interval);
6167 sdg->sgc->next_update = jiffies + interval;
6170 update_cpu_capacity(sd, cpu);
6176 if (child->flags & SD_OVERLAP) {
6178 * SD_OVERLAP domains cannot assume that child groups
6179 * span the current group.
6182 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6183 struct sched_group_capacity *sgc;
6184 struct rq *rq = cpu_rq(cpu);
6187 * build_sched_domains() -> init_sched_groups_capacity()
6188 * gets here before we've attached the domains to the
6191 * Use capacity_of(), which is set irrespective of domains
6192 * in update_cpu_capacity().
6194 * This avoids capacity from being 0 and
6195 * causing divide-by-zero issues on boot.
6197 if (unlikely(!rq->sd)) {
6198 capacity += capacity_of(cpu);
6202 sgc = rq->sd->groups->sgc;
6203 capacity += sgc->capacity;
6207 * !SD_OVERLAP domains can assume that child groups
6208 * span the current group.
6211 group = child->groups;
6213 capacity += group->sgc->capacity;
6214 group = group->next;
6215 } while (group != child->groups);
6218 sdg->sgc->capacity = capacity;
6222 * Check whether the capacity of the rq has been noticeably reduced by side
6223 * activity. The imbalance_pct is used for the threshold.
6224 * Return true is the capacity is reduced
6227 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6229 return ((rq->cpu_capacity * sd->imbalance_pct) <
6230 (rq->cpu_capacity_orig * 100));
6234 * Group imbalance indicates (and tries to solve) the problem where balancing
6235 * groups is inadequate due to tsk_cpus_allowed() constraints.
6237 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6238 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6241 * { 0 1 2 3 } { 4 5 6 7 }
6244 * If we were to balance group-wise we'd place two tasks in the first group and
6245 * two tasks in the second group. Clearly this is undesired as it will overload
6246 * cpu 3 and leave one of the cpus in the second group unused.
6248 * The current solution to this issue is detecting the skew in the first group
6249 * by noticing the lower domain failed to reach balance and had difficulty
6250 * moving tasks due to affinity constraints.
6252 * When this is so detected; this group becomes a candidate for busiest; see
6253 * update_sd_pick_busiest(). And calculate_imbalance() and
6254 * find_busiest_group() avoid some of the usual balance conditions to allow it
6255 * to create an effective group imbalance.
6257 * This is a somewhat tricky proposition since the next run might not find the
6258 * group imbalance and decide the groups need to be balanced again. A most
6259 * subtle and fragile situation.
6262 static inline int sg_imbalanced(struct sched_group *group)
6264 return group->sgc->imbalance;
6268 * group_has_capacity returns true if the group has spare capacity that could
6269 * be used by some tasks.
6270 * We consider that a group has spare capacity if the * number of task is
6271 * smaller than the number of CPUs or if the utilization is lower than the
6272 * available capacity for CFS tasks.
6273 * For the latter, we use a threshold to stabilize the state, to take into
6274 * account the variance of the tasks' load and to return true if the available
6275 * capacity in meaningful for the load balancer.
6276 * As an example, an available capacity of 1% can appear but it doesn't make
6277 * any benefit for the load balance.
6280 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6282 if (sgs->sum_nr_running < sgs->group_weight)
6285 if ((sgs->group_capacity * 100) >
6286 (sgs->group_util * env->sd->imbalance_pct))
6293 * group_is_overloaded returns true if the group has more tasks than it can
6295 * group_is_overloaded is not equals to !group_has_capacity because a group
6296 * with the exact right number of tasks, has no more spare capacity but is not
6297 * overloaded so both group_has_capacity and group_is_overloaded return
6301 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6303 if (sgs->sum_nr_running <= sgs->group_weight)
6306 if ((sgs->group_capacity * 100) <
6307 (sgs->group_util * env->sd->imbalance_pct))
6314 group_type group_classify(struct sched_group *group,
6315 struct sg_lb_stats *sgs)
6317 if (sgs->group_no_capacity)
6318 return group_overloaded;
6320 if (sg_imbalanced(group))
6321 return group_imbalanced;
6327 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6328 * @env: The load balancing environment.
6329 * @group: sched_group whose statistics are to be updated.
6330 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6331 * @local_group: Does group contain this_cpu.
6332 * @sgs: variable to hold the statistics for this group.
6333 * @overload: Indicate more than one runnable task for any CPU.
6335 static inline void update_sg_lb_stats(struct lb_env *env,
6336 struct sched_group *group, int load_idx,
6337 int local_group, struct sg_lb_stats *sgs,
6343 memset(sgs, 0, sizeof(*sgs));
6345 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6346 struct rq *rq = cpu_rq(i);
6348 /* Bias balancing toward cpus of our domain */
6350 load = target_load(i, load_idx);
6352 load = source_load(i, load_idx);
6354 sgs->group_load += load;
6355 sgs->group_util += cpu_util(i);
6356 sgs->sum_nr_running += rq->cfs.h_nr_running;
6358 if (rq->nr_running > 1)
6361 #ifdef CONFIG_NUMA_BALANCING
6362 sgs->nr_numa_running += rq->nr_numa_running;
6363 sgs->nr_preferred_running += rq->nr_preferred_running;
6365 sgs->sum_weighted_load += weighted_cpuload(i);
6370 /* Adjust by relative CPU capacity of the group */
6371 sgs->group_capacity = group->sgc->capacity;
6372 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6374 if (sgs->sum_nr_running)
6375 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6377 sgs->group_weight = group->group_weight;
6379 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6380 sgs->group_type = group_classify(group, sgs);
6384 * update_sd_pick_busiest - return 1 on busiest group
6385 * @env: The load balancing environment.
6386 * @sds: sched_domain statistics
6387 * @sg: sched_group candidate to be checked for being the busiest
6388 * @sgs: sched_group statistics
6390 * Determine if @sg is a busier group than the previously selected
6393 * Return: %true if @sg is a busier group than the previously selected
6394 * busiest group. %false otherwise.
6396 static bool update_sd_pick_busiest(struct lb_env *env,
6397 struct sd_lb_stats *sds,
6398 struct sched_group *sg,
6399 struct sg_lb_stats *sgs)
6401 struct sg_lb_stats *busiest = &sds->busiest_stat;
6403 if (sgs->group_type > busiest->group_type)
6406 if (sgs->group_type < busiest->group_type)
6409 if (sgs->avg_load <= busiest->avg_load)
6412 /* This is the busiest node in its class. */
6413 if (!(env->sd->flags & SD_ASYM_PACKING))
6417 * ASYM_PACKING needs to move all the work to the lowest
6418 * numbered CPUs in the group, therefore mark all groups
6419 * higher than ourself as busy.
6421 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6425 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6432 #ifdef CONFIG_NUMA_BALANCING
6433 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6435 if (sgs->sum_nr_running > sgs->nr_numa_running)
6437 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6442 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6444 if (rq->nr_running > rq->nr_numa_running)
6446 if (rq->nr_running > rq->nr_preferred_running)
6451 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6456 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6460 #endif /* CONFIG_NUMA_BALANCING */
6463 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6464 * @env: The load balancing environment.
6465 * @sds: variable to hold the statistics for this sched_domain.
6467 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6469 struct sched_domain *child = env->sd->child;
6470 struct sched_group *sg = env->sd->groups;
6471 struct sg_lb_stats tmp_sgs;
6472 int load_idx, prefer_sibling = 0;
6473 bool overload = false;
6475 if (child && child->flags & SD_PREFER_SIBLING)
6478 load_idx = get_sd_load_idx(env->sd, env->idle);
6481 struct sg_lb_stats *sgs = &tmp_sgs;
6484 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6487 sgs = &sds->local_stat;
6489 if (env->idle != CPU_NEWLY_IDLE ||
6490 time_after_eq(jiffies, sg->sgc->next_update))
6491 update_group_capacity(env->sd, env->dst_cpu);
6494 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6501 * In case the child domain prefers tasks go to siblings
6502 * first, lower the sg capacity so that we'll try
6503 * and move all the excess tasks away. We lower the capacity
6504 * of a group only if the local group has the capacity to fit
6505 * these excess tasks. The extra check prevents the case where
6506 * you always pull from the heaviest group when it is already
6507 * under-utilized (possible with a large weight task outweighs
6508 * the tasks on the system).
6510 if (prefer_sibling && sds->local &&
6511 group_has_capacity(env, &sds->local_stat) &&
6512 (sgs->sum_nr_running > 1)) {
6513 sgs->group_no_capacity = 1;
6514 sgs->group_type = group_classify(sg, sgs);
6517 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6519 sds->busiest_stat = *sgs;
6523 /* Now, start updating sd_lb_stats */
6524 sds->total_load += sgs->group_load;
6525 sds->total_capacity += sgs->group_capacity;
6528 } while (sg != env->sd->groups);
6530 if (env->sd->flags & SD_NUMA)
6531 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6533 if (!env->sd->parent) {
6534 /* update overload indicator if we are at root domain */
6535 if (env->dst_rq->rd->overload != overload)
6536 env->dst_rq->rd->overload = overload;
6542 * check_asym_packing - Check to see if the group is packed into the
6545 * This is primarily intended to used at the sibling level. Some
6546 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6547 * case of POWER7, it can move to lower SMT modes only when higher
6548 * threads are idle. When in lower SMT modes, the threads will
6549 * perform better since they share less core resources. Hence when we
6550 * have idle threads, we want them to be the higher ones.
6552 * This packing function is run on idle threads. It checks to see if
6553 * the busiest CPU in this domain (core in the P7 case) has a higher
6554 * CPU number than the packing function is being run on. Here we are
6555 * assuming lower CPU number will be equivalent to lower a SMT thread
6558 * Return: 1 when packing is required and a task should be moved to
6559 * this CPU. The amount of the imbalance is returned in *imbalance.
6561 * @env: The load balancing environment.
6562 * @sds: Statistics of the sched_domain which is to be packed
6564 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6568 if (!(env->sd->flags & SD_ASYM_PACKING))
6574 busiest_cpu = group_first_cpu(sds->busiest);
6575 if (env->dst_cpu > busiest_cpu)
6578 env->imbalance = DIV_ROUND_CLOSEST(
6579 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6580 SCHED_CAPACITY_SCALE);
6586 * fix_small_imbalance - Calculate the minor imbalance that exists
6587 * amongst the groups of a sched_domain, during
6589 * @env: The load balancing environment.
6590 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6593 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6595 unsigned long tmp, capa_now = 0, capa_move = 0;
6596 unsigned int imbn = 2;
6597 unsigned long scaled_busy_load_per_task;
6598 struct sg_lb_stats *local, *busiest;
6600 local = &sds->local_stat;
6601 busiest = &sds->busiest_stat;
6603 if (!local->sum_nr_running)
6604 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6605 else if (busiest->load_per_task > local->load_per_task)
6608 scaled_busy_load_per_task =
6609 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6610 busiest->group_capacity;
6612 if (busiest->avg_load + scaled_busy_load_per_task >=
6613 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6614 env->imbalance = busiest->load_per_task;
6619 * OK, we don't have enough imbalance to justify moving tasks,
6620 * however we may be able to increase total CPU capacity used by
6624 capa_now += busiest->group_capacity *
6625 min(busiest->load_per_task, busiest->avg_load);
6626 capa_now += local->group_capacity *
6627 min(local->load_per_task, local->avg_load);
6628 capa_now /= SCHED_CAPACITY_SCALE;
6630 /* Amount of load we'd subtract */
6631 if (busiest->avg_load > scaled_busy_load_per_task) {
6632 capa_move += busiest->group_capacity *
6633 min(busiest->load_per_task,
6634 busiest->avg_load - scaled_busy_load_per_task);
6637 /* Amount of load we'd add */
6638 if (busiest->avg_load * busiest->group_capacity <
6639 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6640 tmp = (busiest->avg_load * busiest->group_capacity) /
6641 local->group_capacity;
6643 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6644 local->group_capacity;
6646 capa_move += local->group_capacity *
6647 min(local->load_per_task, local->avg_load + tmp);
6648 capa_move /= SCHED_CAPACITY_SCALE;
6650 /* Move if we gain throughput */
6651 if (capa_move > capa_now)
6652 env->imbalance = busiest->load_per_task;
6656 * calculate_imbalance - Calculate the amount of imbalance present within the
6657 * groups of a given sched_domain during load balance.
6658 * @env: load balance environment
6659 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6661 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6663 unsigned long max_pull, load_above_capacity = ~0UL;
6664 struct sg_lb_stats *local, *busiest;
6666 local = &sds->local_stat;
6667 busiest = &sds->busiest_stat;
6669 if (busiest->group_type == group_imbalanced) {
6671 * In the group_imb case we cannot rely on group-wide averages
6672 * to ensure cpu-load equilibrium, look at wider averages. XXX
6674 busiest->load_per_task =
6675 min(busiest->load_per_task, sds->avg_load);
6679 * In the presence of smp nice balancing, certain scenarios can have
6680 * max load less than avg load(as we skip the groups at or below
6681 * its cpu_capacity, while calculating max_load..)
6683 if (busiest->avg_load <= sds->avg_load ||
6684 local->avg_load >= sds->avg_load) {
6686 return fix_small_imbalance(env, sds);
6690 * If there aren't any idle cpus, avoid creating some.
6692 if (busiest->group_type == group_overloaded &&
6693 local->group_type == group_overloaded) {
6694 load_above_capacity = busiest->sum_nr_running *
6696 if (load_above_capacity > busiest->group_capacity)
6697 load_above_capacity -= busiest->group_capacity;
6699 load_above_capacity = ~0UL;
6703 * We're trying to get all the cpus to the average_load, so we don't
6704 * want to push ourselves above the average load, nor do we wish to
6705 * reduce the max loaded cpu below the average load. At the same time,
6706 * we also don't want to reduce the group load below the group capacity
6707 * (so that we can implement power-savings policies etc). Thus we look
6708 * for the minimum possible imbalance.
6710 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6712 /* How much load to actually move to equalise the imbalance */
6713 env->imbalance = min(
6714 max_pull * busiest->group_capacity,
6715 (sds->avg_load - local->avg_load) * local->group_capacity
6716 ) / SCHED_CAPACITY_SCALE;
6719 * if *imbalance is less than the average load per runnable task
6720 * there is no guarantee that any tasks will be moved so we'll have
6721 * a think about bumping its value to force at least one task to be
6724 if (env->imbalance < busiest->load_per_task)
6725 return fix_small_imbalance(env, sds);
6728 /******* find_busiest_group() helpers end here *********************/
6731 * find_busiest_group - Returns the busiest group within the sched_domain
6732 * if there is an imbalance. If there isn't an imbalance, and
6733 * the user has opted for power-savings, it returns a group whose
6734 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6735 * such a group exists.
6737 * Also calculates the amount of weighted load which should be moved
6738 * to restore balance.
6740 * @env: The load balancing environment.
6742 * Return: - The busiest group if imbalance exists.
6743 * - If no imbalance and user has opted for power-savings balance,
6744 * return the least loaded group whose CPUs can be
6745 * put to idle by rebalancing its tasks onto our group.
6747 static struct sched_group *find_busiest_group(struct lb_env *env)
6749 struct sg_lb_stats *local, *busiest;
6750 struct sd_lb_stats sds;
6752 init_sd_lb_stats(&sds);
6755 * Compute the various statistics relavent for load balancing at
6758 update_sd_lb_stats(env, &sds);
6759 local = &sds.local_stat;
6760 busiest = &sds.busiest_stat;
6762 /* ASYM feature bypasses nice load balance check */
6763 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6764 check_asym_packing(env, &sds))
6767 /* There is no busy sibling group to pull tasks from */
6768 if (!sds.busiest || busiest->sum_nr_running == 0)
6771 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6772 / sds.total_capacity;
6775 * If the busiest group is imbalanced the below checks don't
6776 * work because they assume all things are equal, which typically
6777 * isn't true due to cpus_allowed constraints and the like.
6779 if (busiest->group_type == group_imbalanced)
6782 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6783 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6784 busiest->group_no_capacity)
6788 * If the local group is busier than the selected busiest group
6789 * don't try and pull any tasks.
6791 if (local->avg_load >= busiest->avg_load)
6795 * Don't pull any tasks if this group is already above the domain
6798 if (local->avg_load >= sds.avg_load)
6801 if (env->idle == CPU_IDLE) {
6803 * This cpu is idle. If the busiest group is not overloaded
6804 * and there is no imbalance between this and busiest group
6805 * wrt idle cpus, it is balanced. The imbalance becomes
6806 * significant if the diff is greater than 1 otherwise we
6807 * might end up to just move the imbalance on another group
6809 if ((busiest->group_type != group_overloaded) &&
6810 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6814 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6815 * imbalance_pct to be conservative.
6817 if (100 * busiest->avg_load <=
6818 env->sd->imbalance_pct * local->avg_load)
6823 /* Looks like there is an imbalance. Compute it */
6824 calculate_imbalance(env, &sds);
6833 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6835 static struct rq *find_busiest_queue(struct lb_env *env,
6836 struct sched_group *group)
6838 struct rq *busiest = NULL, *rq;
6839 unsigned long busiest_load = 0, busiest_capacity = 1;
6842 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6843 unsigned long capacity, wl;
6847 rt = fbq_classify_rq(rq);
6850 * We classify groups/runqueues into three groups:
6851 * - regular: there are !numa tasks
6852 * - remote: there are numa tasks that run on the 'wrong' node
6853 * - all: there is no distinction
6855 * In order to avoid migrating ideally placed numa tasks,
6856 * ignore those when there's better options.
6858 * If we ignore the actual busiest queue to migrate another
6859 * task, the next balance pass can still reduce the busiest
6860 * queue by moving tasks around inside the node.
6862 * If we cannot move enough load due to this classification
6863 * the next pass will adjust the group classification and
6864 * allow migration of more tasks.
6866 * Both cases only affect the total convergence complexity.
6868 if (rt > env->fbq_type)
6871 capacity = capacity_of(i);
6873 wl = weighted_cpuload(i);
6876 * When comparing with imbalance, use weighted_cpuload()
6877 * which is not scaled with the cpu capacity.
6880 if (rq->nr_running == 1 && wl > env->imbalance &&
6881 !check_cpu_capacity(rq, env->sd))
6885 * For the load comparisons with the other cpu's, consider
6886 * the weighted_cpuload() scaled with the cpu capacity, so
6887 * that the load can be moved away from the cpu that is
6888 * potentially running at a lower capacity.
6890 * Thus we're looking for max(wl_i / capacity_i), crosswise
6891 * multiplication to rid ourselves of the division works out
6892 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6893 * our previous maximum.
6895 if (wl * busiest_capacity > busiest_load * capacity) {
6897 busiest_capacity = capacity;
6906 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6907 * so long as it is large enough.
6909 #define MAX_PINNED_INTERVAL 512
6911 /* Working cpumask for load_balance and load_balance_newidle. */
6912 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6914 static int need_active_balance(struct lb_env *env)
6916 struct sched_domain *sd = env->sd;
6918 if (env->idle == CPU_NEWLY_IDLE) {
6921 * ASYM_PACKING needs to force migrate tasks from busy but
6922 * higher numbered CPUs in order to pack all tasks in the
6923 * lowest numbered CPUs.
6925 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6930 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
6931 * It's worth migrating the task if the src_cpu's capacity is reduced
6932 * because of other sched_class or IRQs if more capacity stays
6933 * available on dst_cpu.
6935 if ((env->idle != CPU_NOT_IDLE) &&
6936 (env->src_rq->cfs.h_nr_running == 1)) {
6937 if ((check_cpu_capacity(env->src_rq, sd)) &&
6938 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
6942 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6945 static int active_load_balance_cpu_stop(void *data);
6947 static int should_we_balance(struct lb_env *env)
6949 struct sched_group *sg = env->sd->groups;
6950 struct cpumask *sg_cpus, *sg_mask;
6951 int cpu, balance_cpu = -1;
6954 * In the newly idle case, we will allow all the cpu's
6955 * to do the newly idle load balance.
6957 if (env->idle == CPU_NEWLY_IDLE)
6960 sg_cpus = sched_group_cpus(sg);
6961 sg_mask = sched_group_mask(sg);
6962 /* Try to find first idle cpu */
6963 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6964 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6971 if (balance_cpu == -1)
6972 balance_cpu = group_balance_cpu(sg);
6975 * First idle cpu or the first cpu(busiest) in this sched group
6976 * is eligible for doing load balancing at this and above domains.
6978 return balance_cpu == env->dst_cpu;
6982 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6983 * tasks if there is an imbalance.
6985 static int load_balance(int this_cpu, struct rq *this_rq,
6986 struct sched_domain *sd, enum cpu_idle_type idle,
6987 int *continue_balancing)
6989 int ld_moved, cur_ld_moved, active_balance = 0;
6990 struct sched_domain *sd_parent = sd->parent;
6991 struct sched_group *group;
6993 unsigned long flags;
6994 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
6996 struct lb_env env = {
6998 .dst_cpu = this_cpu,
7000 .dst_grpmask = sched_group_cpus(sd->groups),
7002 .loop_break = sched_nr_migrate_break,
7005 .tasks = LIST_HEAD_INIT(env.tasks),
7009 * For NEWLY_IDLE load_balancing, we don't need to consider
7010 * other cpus in our group
7012 if (idle == CPU_NEWLY_IDLE)
7013 env.dst_grpmask = NULL;
7015 cpumask_copy(cpus, cpu_active_mask);
7017 schedstat_inc(sd, lb_count[idle]);
7020 if (!should_we_balance(&env)) {
7021 *continue_balancing = 0;
7025 group = find_busiest_group(&env);
7027 schedstat_inc(sd, lb_nobusyg[idle]);
7031 busiest = find_busiest_queue(&env, group);
7033 schedstat_inc(sd, lb_nobusyq[idle]);
7037 BUG_ON(busiest == env.dst_rq);
7039 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7041 env.src_cpu = busiest->cpu;
7042 env.src_rq = busiest;
7045 if (busiest->nr_running > 1) {
7047 * Attempt to move tasks. If find_busiest_group has found
7048 * an imbalance but busiest->nr_running <= 1, the group is
7049 * still unbalanced. ld_moved simply stays zero, so it is
7050 * correctly treated as an imbalance.
7052 env.flags |= LBF_ALL_PINNED;
7053 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7056 raw_spin_lock_irqsave(&busiest->lock, flags);
7059 * cur_ld_moved - load moved in current iteration
7060 * ld_moved - cumulative load moved across iterations
7062 cur_ld_moved = detach_tasks(&env);
7065 * We've detached some tasks from busiest_rq. Every
7066 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7067 * unlock busiest->lock, and we are able to be sure
7068 * that nobody can manipulate the tasks in parallel.
7069 * See task_rq_lock() family for the details.
7072 raw_spin_unlock(&busiest->lock);
7076 ld_moved += cur_ld_moved;
7079 local_irq_restore(flags);
7081 if (env.flags & LBF_NEED_BREAK) {
7082 env.flags &= ~LBF_NEED_BREAK;
7087 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7088 * us and move them to an alternate dst_cpu in our sched_group
7089 * where they can run. The upper limit on how many times we
7090 * iterate on same src_cpu is dependent on number of cpus in our
7093 * This changes load balance semantics a bit on who can move
7094 * load to a given_cpu. In addition to the given_cpu itself
7095 * (or a ilb_cpu acting on its behalf where given_cpu is
7096 * nohz-idle), we now have balance_cpu in a position to move
7097 * load to given_cpu. In rare situations, this may cause
7098 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7099 * _independently_ and at _same_ time to move some load to
7100 * given_cpu) causing exceess load to be moved to given_cpu.
7101 * This however should not happen so much in practice and
7102 * moreover subsequent load balance cycles should correct the
7103 * excess load moved.
7105 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7107 /* Prevent to re-select dst_cpu via env's cpus */
7108 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7110 env.dst_rq = cpu_rq(env.new_dst_cpu);
7111 env.dst_cpu = env.new_dst_cpu;
7112 env.flags &= ~LBF_DST_PINNED;
7114 env.loop_break = sched_nr_migrate_break;
7117 * Go back to "more_balance" rather than "redo" since we
7118 * need to continue with same src_cpu.
7124 * We failed to reach balance because of affinity.
7127 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7129 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7130 *group_imbalance = 1;
7133 /* All tasks on this runqueue were pinned by CPU affinity */
7134 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7135 cpumask_clear_cpu(cpu_of(busiest), cpus);
7136 if (!cpumask_empty(cpus)) {
7138 env.loop_break = sched_nr_migrate_break;
7141 goto out_all_pinned;
7146 schedstat_inc(sd, lb_failed[idle]);
7148 * Increment the failure counter only on periodic balance.
7149 * We do not want newidle balance, which can be very
7150 * frequent, pollute the failure counter causing
7151 * excessive cache_hot migrations and active balances.
7153 if (idle != CPU_NEWLY_IDLE)
7154 sd->nr_balance_failed++;
7156 if (need_active_balance(&env)) {
7157 raw_spin_lock_irqsave(&busiest->lock, flags);
7159 /* don't kick the active_load_balance_cpu_stop,
7160 * if the curr task on busiest cpu can't be
7163 if (!cpumask_test_cpu(this_cpu,
7164 tsk_cpus_allowed(busiest->curr))) {
7165 raw_spin_unlock_irqrestore(&busiest->lock,
7167 env.flags |= LBF_ALL_PINNED;
7168 goto out_one_pinned;
7172 * ->active_balance synchronizes accesses to
7173 * ->active_balance_work. Once set, it's cleared
7174 * only after active load balance is finished.
7176 if (!busiest->active_balance) {
7177 busiest->active_balance = 1;
7178 busiest->push_cpu = this_cpu;
7181 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7183 if (active_balance) {
7184 stop_one_cpu_nowait(cpu_of(busiest),
7185 active_load_balance_cpu_stop, busiest,
7186 &busiest->active_balance_work);
7190 * We've kicked active balancing, reset the failure
7193 sd->nr_balance_failed = sd->cache_nice_tries+1;
7196 sd->nr_balance_failed = 0;
7198 if (likely(!active_balance)) {
7199 /* We were unbalanced, so reset the balancing interval */
7200 sd->balance_interval = sd->min_interval;
7203 * If we've begun active balancing, start to back off. This
7204 * case may not be covered by the all_pinned logic if there
7205 * is only 1 task on the busy runqueue (because we don't call
7208 if (sd->balance_interval < sd->max_interval)
7209 sd->balance_interval *= 2;
7216 * We reach balance although we may have faced some affinity
7217 * constraints. Clear the imbalance flag if it was set.
7220 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7222 if (*group_imbalance)
7223 *group_imbalance = 0;
7228 * We reach balance because all tasks are pinned at this level so
7229 * we can't migrate them. Let the imbalance flag set so parent level
7230 * can try to migrate them.
7232 schedstat_inc(sd, lb_balanced[idle]);
7234 sd->nr_balance_failed = 0;
7237 /* tune up the balancing interval */
7238 if (((env.flags & LBF_ALL_PINNED) &&
7239 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7240 (sd->balance_interval < sd->max_interval))
7241 sd->balance_interval *= 2;
7248 static inline unsigned long
7249 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7251 unsigned long interval = sd->balance_interval;
7254 interval *= sd->busy_factor;
7256 /* scale ms to jiffies */
7257 interval = msecs_to_jiffies(interval);
7258 interval = clamp(interval, 1UL, max_load_balance_interval);
7264 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7266 unsigned long interval, next;
7268 interval = get_sd_balance_interval(sd, cpu_busy);
7269 next = sd->last_balance + interval;
7271 if (time_after(*next_balance, next))
7272 *next_balance = next;
7276 * idle_balance is called by schedule() if this_cpu is about to become
7277 * idle. Attempts to pull tasks from other CPUs.
7279 static int idle_balance(struct rq *this_rq)
7281 unsigned long next_balance = jiffies + HZ;
7282 int this_cpu = this_rq->cpu;
7283 struct sched_domain *sd;
7284 int pulled_task = 0;
7287 idle_enter_fair(this_rq);
7290 * We must set idle_stamp _before_ calling idle_balance(), such that we
7291 * measure the duration of idle_balance() as idle time.
7293 this_rq->idle_stamp = rq_clock(this_rq);
7295 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7296 !this_rq->rd->overload) {
7298 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7300 update_next_balance(sd, 0, &next_balance);
7306 raw_spin_unlock(&this_rq->lock);
7308 update_blocked_averages(this_cpu);
7310 for_each_domain(this_cpu, sd) {
7311 int continue_balancing = 1;
7312 u64 t0, domain_cost;
7314 if (!(sd->flags & SD_LOAD_BALANCE))
7317 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7318 update_next_balance(sd, 0, &next_balance);
7322 if (sd->flags & SD_BALANCE_NEWIDLE) {
7323 t0 = sched_clock_cpu(this_cpu);
7325 pulled_task = load_balance(this_cpu, this_rq,
7327 &continue_balancing);
7329 domain_cost = sched_clock_cpu(this_cpu) - t0;
7330 if (domain_cost > sd->max_newidle_lb_cost)
7331 sd->max_newidle_lb_cost = domain_cost;
7333 curr_cost += domain_cost;
7336 update_next_balance(sd, 0, &next_balance);
7339 * Stop searching for tasks to pull if there are
7340 * now runnable tasks on this rq.
7342 if (pulled_task || this_rq->nr_running > 0)
7347 raw_spin_lock(&this_rq->lock);
7349 if (curr_cost > this_rq->max_idle_balance_cost)
7350 this_rq->max_idle_balance_cost = curr_cost;
7353 * While browsing the domains, we released the rq lock, a task could
7354 * have been enqueued in the meantime. Since we're not going idle,
7355 * pretend we pulled a task.
7357 if (this_rq->cfs.h_nr_running && !pulled_task)
7361 /* Move the next balance forward */
7362 if (time_after(this_rq->next_balance, next_balance))
7363 this_rq->next_balance = next_balance;
7365 /* Is there a task of a high priority class? */
7366 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7370 idle_exit_fair(this_rq);
7371 this_rq->idle_stamp = 0;
7378 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7379 * running tasks off the busiest CPU onto idle CPUs. It requires at
7380 * least 1 task to be running on each physical CPU where possible, and
7381 * avoids physical / logical imbalances.
7383 static int active_load_balance_cpu_stop(void *data)
7385 struct rq *busiest_rq = data;
7386 int busiest_cpu = cpu_of(busiest_rq);
7387 int target_cpu = busiest_rq->push_cpu;
7388 struct rq *target_rq = cpu_rq(target_cpu);
7389 struct sched_domain *sd;
7390 struct task_struct *p = NULL;
7392 raw_spin_lock_irq(&busiest_rq->lock);
7394 /* make sure the requested cpu hasn't gone down in the meantime */
7395 if (unlikely(busiest_cpu != smp_processor_id() ||
7396 !busiest_rq->active_balance))
7399 /* Is there any task to move? */
7400 if (busiest_rq->nr_running <= 1)
7404 * This condition is "impossible", if it occurs
7405 * we need to fix it. Originally reported by
7406 * Bjorn Helgaas on a 128-cpu setup.
7408 BUG_ON(busiest_rq == target_rq);
7410 /* Search for an sd spanning us and the target CPU. */
7412 for_each_domain(target_cpu, sd) {
7413 if ((sd->flags & SD_LOAD_BALANCE) &&
7414 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7419 struct lb_env env = {
7421 .dst_cpu = target_cpu,
7422 .dst_rq = target_rq,
7423 .src_cpu = busiest_rq->cpu,
7424 .src_rq = busiest_rq,
7428 schedstat_inc(sd, alb_count);
7430 p = detach_one_task(&env);
7432 schedstat_inc(sd, alb_pushed);
7434 schedstat_inc(sd, alb_failed);
7438 busiest_rq->active_balance = 0;
7439 raw_spin_unlock(&busiest_rq->lock);
7442 attach_one_task(target_rq, p);
7449 static inline int on_null_domain(struct rq *rq)
7451 return unlikely(!rcu_dereference_sched(rq->sd));
7454 #ifdef CONFIG_NO_HZ_COMMON
7456 * idle load balancing details
7457 * - When one of the busy CPUs notice that there may be an idle rebalancing
7458 * needed, they will kick the idle load balancer, which then does idle
7459 * load balancing for all the idle CPUs.
7462 cpumask_var_t idle_cpus_mask;
7464 unsigned long next_balance; /* in jiffy units */
7465 } nohz ____cacheline_aligned;
7467 static inline int find_new_ilb(void)
7469 int ilb = cpumask_first(nohz.idle_cpus_mask);
7471 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7478 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7479 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7480 * CPU (if there is one).
7482 static void nohz_balancer_kick(void)
7486 nohz.next_balance++;
7488 ilb_cpu = find_new_ilb();
7490 if (ilb_cpu >= nr_cpu_ids)
7493 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7496 * Use smp_send_reschedule() instead of resched_cpu().
7497 * This way we generate a sched IPI on the target cpu which
7498 * is idle. And the softirq performing nohz idle load balance
7499 * will be run before returning from the IPI.
7501 smp_send_reschedule(ilb_cpu);
7505 static inline void nohz_balance_exit_idle(int cpu)
7507 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7509 * Completely isolated CPUs don't ever set, so we must test.
7511 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7512 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7513 atomic_dec(&nohz.nr_cpus);
7515 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7519 static inline void set_cpu_sd_state_busy(void)
7521 struct sched_domain *sd;
7522 int cpu = smp_processor_id();
7525 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7527 if (!sd || !sd->nohz_idle)
7531 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7536 void set_cpu_sd_state_idle(void)
7538 struct sched_domain *sd;
7539 int cpu = smp_processor_id();
7542 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7544 if (!sd || sd->nohz_idle)
7548 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7554 * This routine will record that the cpu is going idle with tick stopped.
7555 * This info will be used in performing idle load balancing in the future.
7557 void nohz_balance_enter_idle(int cpu)
7560 * If this cpu is going down, then nothing needs to be done.
7562 if (!cpu_active(cpu))
7565 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7569 * If we're a completely isolated CPU, we don't play.
7571 if (on_null_domain(cpu_rq(cpu)))
7574 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7575 atomic_inc(&nohz.nr_cpus);
7576 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7579 static int sched_ilb_notifier(struct notifier_block *nfb,
7580 unsigned long action, void *hcpu)
7582 switch (action & ~CPU_TASKS_FROZEN) {
7584 nohz_balance_exit_idle(smp_processor_id());
7592 static DEFINE_SPINLOCK(balancing);
7595 * Scale the max load_balance interval with the number of CPUs in the system.
7596 * This trades load-balance latency on larger machines for less cross talk.
7598 void update_max_interval(void)
7600 max_load_balance_interval = HZ*num_online_cpus()/10;
7604 * It checks each scheduling domain to see if it is due to be balanced,
7605 * and initiates a balancing operation if so.
7607 * Balancing parameters are set up in init_sched_domains.
7609 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7611 int continue_balancing = 1;
7613 unsigned long interval;
7614 struct sched_domain *sd;
7615 /* Earliest time when we have to do rebalance again */
7616 unsigned long next_balance = jiffies + 60*HZ;
7617 int update_next_balance = 0;
7618 int need_serialize, need_decay = 0;
7621 update_blocked_averages(cpu);
7624 for_each_domain(cpu, sd) {
7626 * Decay the newidle max times here because this is a regular
7627 * visit to all the domains. Decay ~1% per second.
7629 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7630 sd->max_newidle_lb_cost =
7631 (sd->max_newidle_lb_cost * 253) / 256;
7632 sd->next_decay_max_lb_cost = jiffies + HZ;
7635 max_cost += sd->max_newidle_lb_cost;
7637 if (!(sd->flags & SD_LOAD_BALANCE))
7641 * Stop the load balance at this level. There is another
7642 * CPU in our sched group which is doing load balancing more
7645 if (!continue_balancing) {
7651 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7653 need_serialize = sd->flags & SD_SERIALIZE;
7654 if (need_serialize) {
7655 if (!spin_trylock(&balancing))
7659 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7660 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7662 * The LBF_DST_PINNED logic could have changed
7663 * env->dst_cpu, so we can't know our idle
7664 * state even if we migrated tasks. Update it.
7666 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7668 sd->last_balance = jiffies;
7669 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7672 spin_unlock(&balancing);
7674 if (time_after(next_balance, sd->last_balance + interval)) {
7675 next_balance = sd->last_balance + interval;
7676 update_next_balance = 1;
7681 * Ensure the rq-wide value also decays but keep it at a
7682 * reasonable floor to avoid funnies with rq->avg_idle.
7684 rq->max_idle_balance_cost =
7685 max((u64)sysctl_sched_migration_cost, max_cost);
7690 * next_balance will be updated only when there is a need.
7691 * When the cpu is attached to null domain for ex, it will not be
7694 if (likely(update_next_balance)) {
7695 rq->next_balance = next_balance;
7697 #ifdef CONFIG_NO_HZ_COMMON
7699 * If this CPU has been elected to perform the nohz idle
7700 * balance. Other idle CPUs have already rebalanced with
7701 * nohz_idle_balance() and nohz.next_balance has been
7702 * updated accordingly. This CPU is now running the idle load
7703 * balance for itself and we need to update the
7704 * nohz.next_balance accordingly.
7706 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
7707 nohz.next_balance = rq->next_balance;
7712 #ifdef CONFIG_NO_HZ_COMMON
7714 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7715 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7717 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7719 int this_cpu = this_rq->cpu;
7722 /* Earliest time when we have to do rebalance again */
7723 unsigned long next_balance = jiffies + 60*HZ;
7724 int update_next_balance = 0;
7726 if (idle != CPU_IDLE ||
7727 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7730 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7731 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7735 * If this cpu gets work to do, stop the load balancing
7736 * work being done for other cpus. Next load
7737 * balancing owner will pick it up.
7742 rq = cpu_rq(balance_cpu);
7745 * If time for next balance is due,
7748 if (time_after_eq(jiffies, rq->next_balance)) {
7749 raw_spin_lock_irq(&rq->lock);
7750 update_rq_clock(rq);
7751 update_idle_cpu_load(rq);
7752 raw_spin_unlock_irq(&rq->lock);
7753 rebalance_domains(rq, CPU_IDLE);
7756 if (time_after(next_balance, rq->next_balance)) {
7757 next_balance = rq->next_balance;
7758 update_next_balance = 1;
7763 * next_balance will be updated only when there is a need.
7764 * When the CPU is attached to null domain for ex, it will not be
7767 if (likely(update_next_balance))
7768 nohz.next_balance = next_balance;
7770 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7774 * Current heuristic for kicking the idle load balancer in the presence
7775 * of an idle cpu in the system.
7776 * - This rq has more than one task.
7777 * - This rq has at least one CFS task and the capacity of the CPU is
7778 * significantly reduced because of RT tasks or IRQs.
7779 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7780 * multiple busy cpu.
7781 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7782 * domain span are idle.
7784 static inline bool nohz_kick_needed(struct rq *rq)
7786 unsigned long now = jiffies;
7787 struct sched_domain *sd;
7788 struct sched_group_capacity *sgc;
7789 int nr_busy, cpu = rq->cpu;
7792 if (unlikely(rq->idle_balance))
7796 * We may be recently in ticked or tickless idle mode. At the first
7797 * busy tick after returning from idle, we will update the busy stats.
7799 set_cpu_sd_state_busy();
7800 nohz_balance_exit_idle(cpu);
7803 * None are in tickless mode and hence no need for NOHZ idle load
7806 if (likely(!atomic_read(&nohz.nr_cpus)))
7809 if (time_before(now, nohz.next_balance))
7812 if (rq->nr_running >= 2)
7816 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7818 sgc = sd->groups->sgc;
7819 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7828 sd = rcu_dereference(rq->sd);
7830 if ((rq->cfs.h_nr_running >= 1) &&
7831 check_cpu_capacity(rq, sd)) {
7837 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7838 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7839 sched_domain_span(sd)) < cpu)) {
7849 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7853 * run_rebalance_domains is triggered when needed from the scheduler tick.
7854 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7856 static void run_rebalance_domains(struct softirq_action *h)
7858 struct rq *this_rq = this_rq();
7859 enum cpu_idle_type idle = this_rq->idle_balance ?
7860 CPU_IDLE : CPU_NOT_IDLE;
7863 * If this cpu has a pending nohz_balance_kick, then do the
7864 * balancing on behalf of the other idle cpus whose ticks are
7865 * stopped. Do nohz_idle_balance *before* rebalance_domains to
7866 * give the idle cpus a chance to load balance. Else we may
7867 * load balance only within the local sched_domain hierarchy
7868 * and abort nohz_idle_balance altogether if we pull some load.
7870 nohz_idle_balance(this_rq, idle);
7871 rebalance_domains(this_rq, idle);
7875 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7877 void trigger_load_balance(struct rq *rq)
7879 /* Don't need to rebalance while attached to NULL domain */
7880 if (unlikely(on_null_domain(rq)))
7883 if (time_after_eq(jiffies, rq->next_balance))
7884 raise_softirq(SCHED_SOFTIRQ);
7885 #ifdef CONFIG_NO_HZ_COMMON
7886 if (nohz_kick_needed(rq))
7887 nohz_balancer_kick();
7891 static void rq_online_fair(struct rq *rq)
7895 update_runtime_enabled(rq);
7898 static void rq_offline_fair(struct rq *rq)
7902 /* Ensure any throttled groups are reachable by pick_next_task */
7903 unthrottle_offline_cfs_rqs(rq);
7906 #endif /* CONFIG_SMP */
7909 * scheduler tick hitting a task of our scheduling class:
7911 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7913 struct cfs_rq *cfs_rq;
7914 struct sched_entity *se = &curr->se;
7916 for_each_sched_entity(se) {
7917 cfs_rq = cfs_rq_of(se);
7918 entity_tick(cfs_rq, se, queued);
7921 if (static_branch_unlikely(&sched_numa_balancing))
7922 task_tick_numa(rq, curr);
7926 * called on fork with the child task as argument from the parent's context
7927 * - child not yet on the tasklist
7928 * - preemption disabled
7930 static void task_fork_fair(struct task_struct *p)
7932 struct cfs_rq *cfs_rq;
7933 struct sched_entity *se = &p->se, *curr;
7934 int this_cpu = smp_processor_id();
7935 struct rq *rq = this_rq();
7936 unsigned long flags;
7938 raw_spin_lock_irqsave(&rq->lock, flags);
7940 update_rq_clock(rq);
7942 cfs_rq = task_cfs_rq(current);
7943 curr = cfs_rq->curr;
7946 * Not only the cpu but also the task_group of the parent might have
7947 * been changed after parent->se.parent,cfs_rq were copied to
7948 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7949 * of child point to valid ones.
7952 __set_task_cpu(p, this_cpu);
7955 update_curr(cfs_rq);
7958 se->vruntime = curr->vruntime;
7959 place_entity(cfs_rq, se, 1);
7961 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7963 * Upon rescheduling, sched_class::put_prev_task() will place
7964 * 'current' within the tree based on its new key value.
7966 swap(curr->vruntime, se->vruntime);
7970 se->vruntime -= cfs_rq->min_vruntime;
7972 raw_spin_unlock_irqrestore(&rq->lock, flags);
7976 * Priority of the task has changed. Check to see if we preempt
7980 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7982 if (!task_on_rq_queued(p))
7986 * Reschedule if we are currently running on this runqueue and
7987 * our priority decreased, or if we are not currently running on
7988 * this runqueue and our priority is higher than the current's
7990 if (rq->curr == p) {
7991 if (p->prio > oldprio)
7994 check_preempt_curr(rq, p, 0);
7997 static inline bool vruntime_normalized(struct task_struct *p)
7999 struct sched_entity *se = &p->se;
8002 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8003 * the dequeue_entity(.flags=0) will already have normalized the
8010 * When !on_rq, vruntime of the task has usually NOT been normalized.
8011 * But there are some cases where it has already been normalized:
8013 * - A forked child which is waiting for being woken up by
8014 * wake_up_new_task().
8015 * - A task which has been woken up by try_to_wake_up() and
8016 * waiting for actually being woken up by sched_ttwu_pending().
8018 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8024 static void detach_task_cfs_rq(struct task_struct *p)
8026 struct sched_entity *se = &p->se;
8027 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8029 if (!vruntime_normalized(p)) {
8031 * Fix up our vruntime so that the current sleep doesn't
8032 * cause 'unlimited' sleep bonus.
8034 place_entity(cfs_rq, se, 0);
8035 se->vruntime -= cfs_rq->min_vruntime;
8038 /* Catch up with the cfs_rq and remove our load when we leave */
8039 detach_entity_load_avg(cfs_rq, se);
8042 static void attach_task_cfs_rq(struct task_struct *p)
8044 struct sched_entity *se = &p->se;
8045 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8047 #ifdef CONFIG_FAIR_GROUP_SCHED
8049 * Since the real-depth could have been changed (only FAIR
8050 * class maintain depth value), reset depth properly.
8052 se->depth = se->parent ? se->parent->depth + 1 : 0;
8055 /* Synchronize task with its cfs_rq */
8056 attach_entity_load_avg(cfs_rq, se);
8058 if (!vruntime_normalized(p))
8059 se->vruntime += cfs_rq->min_vruntime;
8062 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8064 detach_task_cfs_rq(p);
8067 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8069 attach_task_cfs_rq(p);
8071 if (task_on_rq_queued(p)) {
8073 * We were most likely switched from sched_rt, so
8074 * kick off the schedule if running, otherwise just see
8075 * if we can still preempt the current task.
8080 check_preempt_curr(rq, p, 0);
8084 /* Account for a task changing its policy or group.
8086 * This routine is mostly called to set cfs_rq->curr field when a task
8087 * migrates between groups/classes.
8089 static void set_curr_task_fair(struct rq *rq)
8091 struct sched_entity *se = &rq->curr->se;
8093 for_each_sched_entity(se) {
8094 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8096 set_next_entity(cfs_rq, se);
8097 /* ensure bandwidth has been allocated on our new cfs_rq */
8098 account_cfs_rq_runtime(cfs_rq, 0);
8102 void init_cfs_rq(struct cfs_rq *cfs_rq)
8104 cfs_rq->tasks_timeline = RB_ROOT;
8105 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8106 #ifndef CONFIG_64BIT
8107 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8110 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8111 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8115 #ifdef CONFIG_FAIR_GROUP_SCHED
8116 static void task_move_group_fair(struct task_struct *p)
8118 detach_task_cfs_rq(p);
8119 set_task_rq(p, task_cpu(p));
8122 /* Tell se's cfs_rq has been changed -- migrated */
8123 p->se.avg.last_update_time = 0;
8125 attach_task_cfs_rq(p);
8128 void free_fair_sched_group(struct task_group *tg)
8132 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8134 for_each_possible_cpu(i) {
8136 kfree(tg->cfs_rq[i]);
8139 remove_entity_load_avg(tg->se[i]);
8148 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8150 struct cfs_rq *cfs_rq;
8151 struct sched_entity *se;
8154 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8157 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8161 tg->shares = NICE_0_LOAD;
8163 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8165 for_each_possible_cpu(i) {
8166 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8167 GFP_KERNEL, cpu_to_node(i));
8171 se = kzalloc_node(sizeof(struct sched_entity),
8172 GFP_KERNEL, cpu_to_node(i));
8176 init_cfs_rq(cfs_rq);
8177 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8178 init_entity_runnable_average(se);
8189 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8191 struct rq *rq = cpu_rq(cpu);
8192 unsigned long flags;
8195 * Only empty task groups can be destroyed; so we can speculatively
8196 * check on_list without danger of it being re-added.
8198 if (!tg->cfs_rq[cpu]->on_list)
8201 raw_spin_lock_irqsave(&rq->lock, flags);
8202 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8203 raw_spin_unlock_irqrestore(&rq->lock, flags);
8206 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8207 struct sched_entity *se, int cpu,
8208 struct sched_entity *parent)
8210 struct rq *rq = cpu_rq(cpu);
8214 init_cfs_rq_runtime(cfs_rq);
8216 tg->cfs_rq[cpu] = cfs_rq;
8219 /* se could be NULL for root_task_group */
8224 se->cfs_rq = &rq->cfs;
8227 se->cfs_rq = parent->my_q;
8228 se->depth = parent->depth + 1;
8232 /* guarantee group entities always have weight */
8233 update_load_set(&se->load, NICE_0_LOAD);
8234 se->parent = parent;
8237 static DEFINE_MUTEX(shares_mutex);
8239 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8242 unsigned long flags;
8245 * We can't change the weight of the root cgroup.
8250 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8252 mutex_lock(&shares_mutex);
8253 if (tg->shares == shares)
8256 tg->shares = shares;
8257 for_each_possible_cpu(i) {
8258 struct rq *rq = cpu_rq(i);
8259 struct sched_entity *se;
8262 /* Propagate contribution to hierarchy */
8263 raw_spin_lock_irqsave(&rq->lock, flags);
8265 /* Possible calls to update_curr() need rq clock */
8266 update_rq_clock(rq);
8267 for_each_sched_entity(se)
8268 update_cfs_shares(group_cfs_rq(se));
8269 raw_spin_unlock_irqrestore(&rq->lock, flags);
8273 mutex_unlock(&shares_mutex);
8276 #else /* CONFIG_FAIR_GROUP_SCHED */
8278 void free_fair_sched_group(struct task_group *tg) { }
8280 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8285 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8287 #endif /* CONFIG_FAIR_GROUP_SCHED */
8290 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8292 struct sched_entity *se = &task->se;
8293 unsigned int rr_interval = 0;
8296 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8299 if (rq->cfs.load.weight)
8300 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8306 * All the scheduling class methods:
8308 const struct sched_class fair_sched_class = {
8309 .next = &idle_sched_class,
8310 .enqueue_task = enqueue_task_fair,
8311 .dequeue_task = dequeue_task_fair,
8312 .yield_task = yield_task_fair,
8313 .yield_to_task = yield_to_task_fair,
8315 .check_preempt_curr = check_preempt_wakeup,
8317 .pick_next_task = pick_next_task_fair,
8318 .put_prev_task = put_prev_task_fair,
8321 .select_task_rq = select_task_rq_fair,
8322 .migrate_task_rq = migrate_task_rq_fair,
8324 .rq_online = rq_online_fair,
8325 .rq_offline = rq_offline_fair,
8327 .task_waking = task_waking_fair,
8328 .task_dead = task_dead_fair,
8329 .set_cpus_allowed = set_cpus_allowed_common,
8332 .set_curr_task = set_curr_task_fair,
8333 .task_tick = task_tick_fair,
8334 .task_fork = task_fork_fair,
8336 .prio_changed = prio_changed_fair,
8337 .switched_from = switched_from_fair,
8338 .switched_to = switched_to_fair,
8340 .get_rr_interval = get_rr_interval_fair,
8342 .update_curr = update_curr_fair,
8344 #ifdef CONFIG_FAIR_GROUP_SCHED
8345 .task_move_group = task_move_group_fair,
8349 #ifdef CONFIG_SCHED_DEBUG
8350 void print_cfs_stats(struct seq_file *m, int cpu)
8352 struct cfs_rq *cfs_rq;
8355 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8356 print_cfs_rq(m, cpu, cfs_rq);
8360 #ifdef CONFIG_NUMA_BALANCING
8361 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8364 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8366 for_each_online_node(node) {
8367 if (p->numa_faults) {
8368 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8369 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8371 if (p->numa_group) {
8372 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8373 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8375 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8378 #endif /* CONFIG_NUMA_BALANCING */
8379 #endif /* CONFIG_SCHED_DEBUG */
8381 __init void init_sched_fair_class(void)
8384 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8386 #ifdef CONFIG_NO_HZ_COMMON
8387 nohz.next_balance = jiffies;
8388 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8389 cpu_notifier(sched_ilb_notifier, 0);