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 /* Avoid time going backwards, prevent potential divide error: */
1727 if (unlikely((s64)*period < 0))
1730 delta = p->se.avg.load_sum / p->se.load.weight;
1731 *period = LOAD_AVG_MAX;
1734 p->last_sum_exec_runtime = runtime;
1735 p->last_task_numa_placement = now;
1741 * Determine the preferred nid for a task in a numa_group. This needs to
1742 * be done in a way that produces consistent results with group_weight,
1743 * otherwise workloads might not converge.
1745 static int preferred_group_nid(struct task_struct *p, int nid)
1750 /* Direct connections between all NUMA nodes. */
1751 if (sched_numa_topology_type == NUMA_DIRECT)
1755 * On a system with glueless mesh NUMA topology, group_weight
1756 * scores nodes according to the number of NUMA hinting faults on
1757 * both the node itself, and on nearby nodes.
1759 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1760 unsigned long score, max_score = 0;
1761 int node, max_node = nid;
1763 dist = sched_max_numa_distance;
1765 for_each_online_node(node) {
1766 score = group_weight(p, node, dist);
1767 if (score > max_score) {
1776 * Finding the preferred nid in a system with NUMA backplane
1777 * interconnect topology is more involved. The goal is to locate
1778 * tasks from numa_groups near each other in the system, and
1779 * untangle workloads from different sides of the system. This requires
1780 * searching down the hierarchy of node groups, recursively searching
1781 * inside the highest scoring group of nodes. The nodemask tricks
1782 * keep the complexity of the search down.
1784 nodes = node_online_map;
1785 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1786 unsigned long max_faults = 0;
1787 nodemask_t max_group = NODE_MASK_NONE;
1790 /* Are there nodes at this distance from each other? */
1791 if (!find_numa_distance(dist))
1794 for_each_node_mask(a, nodes) {
1795 unsigned long faults = 0;
1796 nodemask_t this_group;
1797 nodes_clear(this_group);
1799 /* Sum group's NUMA faults; includes a==b case. */
1800 for_each_node_mask(b, nodes) {
1801 if (node_distance(a, b) < dist) {
1802 faults += group_faults(p, b);
1803 node_set(b, this_group);
1804 node_clear(b, nodes);
1808 /* Remember the top group. */
1809 if (faults > max_faults) {
1810 max_faults = faults;
1811 max_group = this_group;
1813 * subtle: at the smallest distance there is
1814 * just one node left in each "group", the
1815 * winner is the preferred nid.
1820 /* Next round, evaluate the nodes within max_group. */
1828 static void task_numa_placement(struct task_struct *p)
1830 int seq, nid, max_nid = -1, max_group_nid = -1;
1831 unsigned long max_faults = 0, max_group_faults = 0;
1832 unsigned long fault_types[2] = { 0, 0 };
1833 unsigned long total_faults;
1834 u64 runtime, period;
1835 spinlock_t *group_lock = NULL;
1838 * The p->mm->numa_scan_seq field gets updated without
1839 * exclusive access. Use READ_ONCE() here to ensure
1840 * that the field is read in a single access:
1842 seq = READ_ONCE(p->mm->numa_scan_seq);
1843 if (p->numa_scan_seq == seq)
1845 p->numa_scan_seq = seq;
1846 p->numa_scan_period_max = task_scan_max(p);
1848 total_faults = p->numa_faults_locality[0] +
1849 p->numa_faults_locality[1];
1850 runtime = numa_get_avg_runtime(p, &period);
1852 /* If the task is part of a group prevent parallel updates to group stats */
1853 if (p->numa_group) {
1854 group_lock = &p->numa_group->lock;
1855 spin_lock_irq(group_lock);
1858 /* Find the node with the highest number of faults */
1859 for_each_online_node(nid) {
1860 /* Keep track of the offsets in numa_faults array */
1861 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1862 unsigned long faults = 0, group_faults = 0;
1865 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1866 long diff, f_diff, f_weight;
1868 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1869 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1870 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1871 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1873 /* Decay existing window, copy faults since last scan */
1874 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1875 fault_types[priv] += p->numa_faults[membuf_idx];
1876 p->numa_faults[membuf_idx] = 0;
1879 * Normalize the faults_from, so all tasks in a group
1880 * count according to CPU use, instead of by the raw
1881 * number of faults. Tasks with little runtime have
1882 * little over-all impact on throughput, and thus their
1883 * faults are less important.
1885 f_weight = div64_u64(runtime << 16, period + 1);
1886 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1888 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1889 p->numa_faults[cpubuf_idx] = 0;
1891 p->numa_faults[mem_idx] += diff;
1892 p->numa_faults[cpu_idx] += f_diff;
1893 faults += p->numa_faults[mem_idx];
1894 p->total_numa_faults += diff;
1895 if (p->numa_group) {
1897 * safe because we can only change our own group
1899 * mem_idx represents the offset for a given
1900 * nid and priv in a specific region because it
1901 * is at the beginning of the numa_faults array.
1903 p->numa_group->faults[mem_idx] += diff;
1904 p->numa_group->faults_cpu[mem_idx] += f_diff;
1905 p->numa_group->total_faults += diff;
1906 group_faults += p->numa_group->faults[mem_idx];
1910 if (faults > max_faults) {
1911 max_faults = faults;
1915 if (group_faults > max_group_faults) {
1916 max_group_faults = group_faults;
1917 max_group_nid = nid;
1921 update_task_scan_period(p, fault_types[0], fault_types[1]);
1923 if (p->numa_group) {
1924 update_numa_active_node_mask(p->numa_group);
1925 spin_unlock_irq(group_lock);
1926 max_nid = preferred_group_nid(p, max_group_nid);
1930 /* Set the new preferred node */
1931 if (max_nid != p->numa_preferred_nid)
1932 sched_setnuma(p, max_nid);
1934 if (task_node(p) != p->numa_preferred_nid)
1935 numa_migrate_preferred(p);
1939 static inline int get_numa_group(struct numa_group *grp)
1941 return atomic_inc_not_zero(&grp->refcount);
1944 static inline void put_numa_group(struct numa_group *grp)
1946 if (atomic_dec_and_test(&grp->refcount))
1947 kfree_rcu(grp, rcu);
1950 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1953 struct numa_group *grp, *my_grp;
1954 struct task_struct *tsk;
1956 int cpu = cpupid_to_cpu(cpupid);
1959 if (unlikely(!p->numa_group)) {
1960 unsigned int size = sizeof(struct numa_group) +
1961 4*nr_node_ids*sizeof(unsigned long);
1963 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1967 atomic_set(&grp->refcount, 1);
1968 spin_lock_init(&grp->lock);
1970 /* Second half of the array tracks nids where faults happen */
1971 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1974 node_set(task_node(current), grp->active_nodes);
1976 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1977 grp->faults[i] = p->numa_faults[i];
1979 grp->total_faults = p->total_numa_faults;
1982 rcu_assign_pointer(p->numa_group, grp);
1986 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1988 if (!cpupid_match_pid(tsk, cpupid))
1991 grp = rcu_dereference(tsk->numa_group);
1995 my_grp = p->numa_group;
2000 * Only join the other group if its bigger; if we're the bigger group,
2001 * the other task will join us.
2003 if (my_grp->nr_tasks > grp->nr_tasks)
2007 * Tie-break on the grp address.
2009 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2012 /* Always join threads in the same process. */
2013 if (tsk->mm == current->mm)
2016 /* Simple filter to avoid false positives due to PID collisions */
2017 if (flags & TNF_SHARED)
2020 /* Update priv based on whether false sharing was detected */
2023 if (join && !get_numa_group(grp))
2031 BUG_ON(irqs_disabled());
2032 double_lock_irq(&my_grp->lock, &grp->lock);
2034 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2035 my_grp->faults[i] -= p->numa_faults[i];
2036 grp->faults[i] += p->numa_faults[i];
2038 my_grp->total_faults -= p->total_numa_faults;
2039 grp->total_faults += p->total_numa_faults;
2044 spin_unlock(&my_grp->lock);
2045 spin_unlock_irq(&grp->lock);
2047 rcu_assign_pointer(p->numa_group, grp);
2049 put_numa_group(my_grp);
2058 * Get rid of NUMA staticstics associated with a task (either current or dead).
2059 * If @final is set, the task is dead and has reached refcount zero, so we can
2060 * safely free all relevant data structures. Otherwise, there might be
2061 * concurrent reads from places like load balancing and procfs, and we should
2062 * reset the data back to default state without freeing ->numa_faults.
2064 void task_numa_free(struct task_struct *p, bool final)
2066 struct numa_group *grp = p->numa_group;
2067 unsigned long *numa_faults = p->numa_faults;
2068 unsigned long flags;
2075 spin_lock_irqsave(&grp->lock, flags);
2076 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2077 grp->faults[i] -= p->numa_faults[i];
2078 grp->total_faults -= p->total_numa_faults;
2081 spin_unlock_irqrestore(&grp->lock, flags);
2082 RCU_INIT_POINTER(p->numa_group, NULL);
2083 put_numa_group(grp);
2087 p->numa_faults = NULL;
2090 p->total_numa_faults = 0;
2091 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2097 * Got a PROT_NONE fault for a page on @node.
2099 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2101 struct task_struct *p = current;
2102 bool migrated = flags & TNF_MIGRATED;
2103 int cpu_node = task_node(current);
2104 int local = !!(flags & TNF_FAULT_LOCAL);
2107 if (!static_branch_likely(&sched_numa_balancing))
2110 /* for example, ksmd faulting in a user's mm */
2114 /* Allocate buffer to track faults on a per-node basis */
2115 if (unlikely(!p->numa_faults)) {
2116 int size = sizeof(*p->numa_faults) *
2117 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2119 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2120 if (!p->numa_faults)
2123 p->total_numa_faults = 0;
2124 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2128 * First accesses are treated as private, otherwise consider accesses
2129 * to be private if the accessing pid has not changed
2131 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2134 priv = cpupid_match_pid(p, last_cpupid);
2135 if (!priv && !(flags & TNF_NO_GROUP))
2136 task_numa_group(p, last_cpupid, flags, &priv);
2140 * If a workload spans multiple NUMA nodes, a shared fault that
2141 * occurs wholly within the set of nodes that the workload is
2142 * actively using should be counted as local. This allows the
2143 * scan rate to slow down when a workload has settled down.
2145 if (!priv && !local && p->numa_group &&
2146 node_isset(cpu_node, p->numa_group->active_nodes) &&
2147 node_isset(mem_node, p->numa_group->active_nodes))
2150 task_numa_placement(p);
2153 * Retry task to preferred node migration periodically, in case it
2154 * case it previously failed, or the scheduler moved us.
2156 if (time_after(jiffies, p->numa_migrate_retry))
2157 numa_migrate_preferred(p);
2160 p->numa_pages_migrated += pages;
2161 if (flags & TNF_MIGRATE_FAIL)
2162 p->numa_faults_locality[2] += pages;
2164 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2165 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2166 p->numa_faults_locality[local] += pages;
2169 static void reset_ptenuma_scan(struct task_struct *p)
2172 * We only did a read acquisition of the mmap sem, so
2173 * p->mm->numa_scan_seq is written to without exclusive access
2174 * and the update is not guaranteed to be atomic. That's not
2175 * much of an issue though, since this is just used for
2176 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2177 * expensive, to avoid any form of compiler optimizations:
2179 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2180 p->mm->numa_scan_offset = 0;
2184 * The expensive part of numa migration is done from task_work context.
2185 * Triggered from task_tick_numa().
2187 void task_numa_work(struct callback_head *work)
2189 unsigned long migrate, next_scan, now = jiffies;
2190 struct task_struct *p = current;
2191 struct mm_struct *mm = p->mm;
2192 struct vm_area_struct *vma;
2193 unsigned long start, end;
2194 unsigned long nr_pte_updates = 0;
2195 long pages, virtpages;
2197 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2199 work->next = work; /* protect against double add */
2201 * Who cares about NUMA placement when they're dying.
2203 * NOTE: make sure not to dereference p->mm before this check,
2204 * exit_task_work() happens _after_ exit_mm() so we could be called
2205 * without p->mm even though we still had it when we enqueued this
2208 if (p->flags & PF_EXITING)
2211 if (!mm->numa_next_scan) {
2212 mm->numa_next_scan = now +
2213 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2217 * Enforce maximal scan/migration frequency..
2219 migrate = mm->numa_next_scan;
2220 if (time_before(now, migrate))
2223 if (p->numa_scan_period == 0) {
2224 p->numa_scan_period_max = task_scan_max(p);
2225 p->numa_scan_period = task_scan_min(p);
2228 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2229 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2233 * Delay this task enough that another task of this mm will likely win
2234 * the next time around.
2236 p->node_stamp += 2 * TICK_NSEC;
2238 start = mm->numa_scan_offset;
2239 pages = sysctl_numa_balancing_scan_size;
2240 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2241 virtpages = pages * 8; /* Scan up to this much virtual space */
2246 if (!down_read_trylock(&mm->mmap_sem))
2248 vma = find_vma(mm, start);
2250 reset_ptenuma_scan(p);
2254 for (; vma; vma = vma->vm_next) {
2255 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2256 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2261 * Shared library pages mapped by multiple processes are not
2262 * migrated as it is expected they are cache replicated. Avoid
2263 * hinting faults in read-only file-backed mappings or the vdso
2264 * as migrating the pages will be of marginal benefit.
2267 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2271 * Skip inaccessible VMAs to avoid any confusion between
2272 * PROT_NONE and NUMA hinting ptes
2274 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2278 start = max(start, vma->vm_start);
2279 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2280 end = min(end, vma->vm_end);
2281 nr_pte_updates = change_prot_numa(vma, start, end);
2284 * Try to scan sysctl_numa_balancing_size worth of
2285 * hpages that have at least one present PTE that
2286 * is not already pte-numa. If the VMA contains
2287 * areas that are unused or already full of prot_numa
2288 * PTEs, scan up to virtpages, to skip through those
2292 pages -= (end - start) >> PAGE_SHIFT;
2293 virtpages -= (end - start) >> PAGE_SHIFT;
2296 if (pages <= 0 || virtpages <= 0)
2300 } while (end != vma->vm_end);
2305 * It is possible to reach the end of the VMA list but the last few
2306 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2307 * would find the !migratable VMA on the next scan but not reset the
2308 * scanner to the start so check it now.
2311 mm->numa_scan_offset = start;
2313 reset_ptenuma_scan(p);
2314 up_read(&mm->mmap_sem);
2318 * Drive the periodic memory faults..
2320 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2322 struct callback_head *work = &curr->numa_work;
2326 * We don't care about NUMA placement if we don't have memory.
2328 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2332 * Using runtime rather than walltime has the dual advantage that
2333 * we (mostly) drive the selection from busy threads and that the
2334 * task needs to have done some actual work before we bother with
2337 now = curr->se.sum_exec_runtime;
2338 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2340 if (now > curr->node_stamp + period) {
2341 if (!curr->node_stamp)
2342 curr->numa_scan_period = task_scan_min(curr);
2343 curr->node_stamp += period;
2345 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2346 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2347 task_work_add(curr, work, true);
2352 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2356 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2360 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2363 #endif /* CONFIG_NUMA_BALANCING */
2366 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2368 update_load_add(&cfs_rq->load, se->load.weight);
2369 if (!parent_entity(se))
2370 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2372 if (entity_is_task(se)) {
2373 struct rq *rq = rq_of(cfs_rq);
2375 account_numa_enqueue(rq, task_of(se));
2376 list_add(&se->group_node, &rq->cfs_tasks);
2379 cfs_rq->nr_running++;
2383 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2385 update_load_sub(&cfs_rq->load, se->load.weight);
2386 if (!parent_entity(se))
2387 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2388 if (entity_is_task(se)) {
2389 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2390 list_del_init(&se->group_node);
2392 cfs_rq->nr_running--;
2395 #ifdef CONFIG_FAIR_GROUP_SCHED
2397 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2399 long tg_weight, load, shares;
2402 * This really should be: cfs_rq->avg.load_avg, but instead we use
2403 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2404 * the shares for small weight interactive tasks.
2406 load = scale_load_down(cfs_rq->load.weight);
2408 tg_weight = atomic_long_read(&tg->load_avg);
2410 /* Ensure tg_weight >= load */
2411 tg_weight -= cfs_rq->tg_load_avg_contrib;
2414 shares = (tg->shares * load);
2416 shares /= tg_weight;
2418 if (shares < MIN_SHARES)
2419 shares = MIN_SHARES;
2420 if (shares > tg->shares)
2421 shares = tg->shares;
2425 # else /* CONFIG_SMP */
2426 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2430 # endif /* CONFIG_SMP */
2432 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2433 unsigned long weight)
2436 /* commit outstanding execution time */
2437 if (cfs_rq->curr == se)
2438 update_curr(cfs_rq);
2439 account_entity_dequeue(cfs_rq, se);
2442 update_load_set(&se->load, weight);
2445 account_entity_enqueue(cfs_rq, se);
2448 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2450 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2452 struct task_group *tg;
2453 struct sched_entity *se;
2457 se = tg->se[cpu_of(rq_of(cfs_rq))];
2458 if (!se || throttled_hierarchy(cfs_rq))
2461 if (likely(se->load.weight == tg->shares))
2464 shares = calc_cfs_shares(cfs_rq, tg);
2466 reweight_entity(cfs_rq_of(se), se, shares);
2468 #else /* CONFIG_FAIR_GROUP_SCHED */
2469 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2472 #endif /* CONFIG_FAIR_GROUP_SCHED */
2475 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2476 static const u32 runnable_avg_yN_inv[] = {
2477 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2478 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2479 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2480 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2481 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2482 0x85aac367, 0x82cd8698,
2486 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2487 * over-estimates when re-combining.
2489 static const u32 runnable_avg_yN_sum[] = {
2490 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2491 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2492 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2497 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2499 static __always_inline u64 decay_load(u64 val, u64 n)
2501 unsigned int local_n;
2505 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2508 /* after bounds checking we can collapse to 32-bit */
2512 * As y^PERIOD = 1/2, we can combine
2513 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2514 * With a look-up table which covers y^n (n<PERIOD)
2516 * To achieve constant time decay_load.
2518 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2519 val >>= local_n / LOAD_AVG_PERIOD;
2520 local_n %= LOAD_AVG_PERIOD;
2523 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2528 * For updates fully spanning n periods, the contribution to runnable
2529 * average will be: \Sum 1024*y^n
2531 * We can compute this reasonably efficiently by combining:
2532 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2534 static u32 __compute_runnable_contrib(u64 n)
2538 if (likely(n <= LOAD_AVG_PERIOD))
2539 return runnable_avg_yN_sum[n];
2540 else if (unlikely(n >= LOAD_AVG_MAX_N))
2541 return LOAD_AVG_MAX;
2543 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2545 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2546 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2548 n -= LOAD_AVG_PERIOD;
2549 } while (n > LOAD_AVG_PERIOD);
2551 contrib = decay_load(contrib, n);
2552 return contrib + runnable_avg_yN_sum[n];
2555 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2556 #error "load tracking assumes 2^10 as unit"
2559 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2562 * We can represent the historical contribution to runnable average as the
2563 * coefficients of a geometric series. To do this we sub-divide our runnable
2564 * history into segments of approximately 1ms (1024us); label the segment that
2565 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2567 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2569 * (now) (~1ms ago) (~2ms ago)
2571 * Let u_i denote the fraction of p_i that the entity was runnable.
2573 * We then designate the fractions u_i as our co-efficients, yielding the
2574 * following representation of historical load:
2575 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2577 * We choose y based on the with of a reasonably scheduling period, fixing:
2580 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2581 * approximately half as much as the contribution to load within the last ms
2584 * When a period "rolls over" and we have new u_0`, multiplying the previous
2585 * sum again by y is sufficient to update:
2586 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2587 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2589 static __always_inline int
2590 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2591 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2593 u64 delta, scaled_delta, periods;
2595 unsigned int delta_w, scaled_delta_w, decayed = 0;
2596 unsigned long scale_freq, scale_cpu;
2598 delta = now - sa->last_update_time;
2600 * This should only happen when time goes backwards, which it
2601 * unfortunately does during sched clock init when we swap over to TSC.
2603 if ((s64)delta < 0) {
2604 sa->last_update_time = now;
2609 * Use 1024ns as the unit of measurement since it's a reasonable
2610 * approximation of 1us and fast to compute.
2615 sa->last_update_time = now;
2617 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2618 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2620 /* delta_w is the amount already accumulated against our next period */
2621 delta_w = sa->period_contrib;
2622 if (delta + delta_w >= 1024) {
2625 /* how much left for next period will start over, we don't know yet */
2626 sa->period_contrib = 0;
2629 * Now that we know we're crossing a period boundary, figure
2630 * out how much from delta we need to complete the current
2631 * period and accrue it.
2633 delta_w = 1024 - delta_w;
2634 scaled_delta_w = cap_scale(delta_w, scale_freq);
2636 sa->load_sum += weight * scaled_delta_w;
2638 cfs_rq->runnable_load_sum +=
2639 weight * scaled_delta_w;
2643 sa->util_sum += scaled_delta_w * scale_cpu;
2647 /* Figure out how many additional periods this update spans */
2648 periods = delta / 1024;
2651 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2653 cfs_rq->runnable_load_sum =
2654 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2656 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2658 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2659 contrib = __compute_runnable_contrib(periods);
2660 contrib = cap_scale(contrib, scale_freq);
2662 sa->load_sum += weight * contrib;
2664 cfs_rq->runnable_load_sum += weight * contrib;
2667 sa->util_sum += contrib * scale_cpu;
2670 /* Remainder of delta accrued against u_0` */
2671 scaled_delta = cap_scale(delta, scale_freq);
2673 sa->load_sum += weight * scaled_delta;
2675 cfs_rq->runnable_load_sum += weight * scaled_delta;
2678 sa->util_sum += scaled_delta * scale_cpu;
2680 sa->period_contrib += delta;
2683 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2685 cfs_rq->runnable_load_avg =
2686 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2688 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2694 #ifdef CONFIG_FAIR_GROUP_SCHED
2696 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2697 * and effective_load (which is not done because it is too costly).
2699 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2701 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2703 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2704 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2705 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2709 #else /* CONFIG_FAIR_GROUP_SCHED */
2710 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2711 #endif /* CONFIG_FAIR_GROUP_SCHED */
2713 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2716 * Unsigned subtract and clamp on underflow.
2718 * Explicitly do a load-store to ensure the intermediate value never hits
2719 * memory. This allows lockless observations without ever seeing the negative
2722 #define sub_positive(_ptr, _val) do { \
2723 typeof(_ptr) ptr = (_ptr); \
2724 typeof(*ptr) val = (_val); \
2725 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2729 WRITE_ONCE(*ptr, res); \
2732 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2733 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2735 struct sched_avg *sa = &cfs_rq->avg;
2736 int decayed, removed = 0;
2738 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2739 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2740 sub_positive(&sa->load_avg, r);
2741 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2745 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2746 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2747 sub_positive(&sa->util_avg, r);
2748 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2751 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2752 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2754 #ifndef CONFIG_64BIT
2756 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2759 return decayed || removed;
2762 /* Update task and its cfs_rq load average */
2763 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2765 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2766 u64 now = cfs_rq_clock_task(cfs_rq);
2767 int cpu = cpu_of(rq_of(cfs_rq));
2770 * Track task load average for carrying it to new CPU after migrated, and
2771 * track group sched_entity load average for task_h_load calc in migration
2773 __update_load_avg(now, cpu, &se->avg,
2774 se->on_rq * scale_load_down(se->load.weight),
2775 cfs_rq->curr == se, NULL);
2777 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2778 update_tg_load_avg(cfs_rq, 0);
2781 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2783 if (!sched_feat(ATTACH_AGE_LOAD))
2787 * If we got migrated (either between CPUs or between cgroups) we'll
2788 * have aged the average right before clearing @last_update_time.
2790 if (se->avg.last_update_time) {
2791 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2792 &se->avg, 0, 0, NULL);
2795 * XXX: we could have just aged the entire load away if we've been
2796 * absent from the fair class for too long.
2801 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2802 cfs_rq->avg.load_avg += se->avg.load_avg;
2803 cfs_rq->avg.load_sum += se->avg.load_sum;
2804 cfs_rq->avg.util_avg += se->avg.util_avg;
2805 cfs_rq->avg.util_sum += se->avg.util_sum;
2808 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2810 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2811 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2812 cfs_rq->curr == se, NULL);
2814 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2815 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2816 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2817 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2820 /* Add the load generated by se into cfs_rq's load average */
2822 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2824 struct sched_avg *sa = &se->avg;
2825 u64 now = cfs_rq_clock_task(cfs_rq);
2826 int migrated, decayed;
2828 migrated = !sa->last_update_time;
2830 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2831 se->on_rq * scale_load_down(se->load.weight),
2832 cfs_rq->curr == se, NULL);
2835 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2837 cfs_rq->runnable_load_avg += sa->load_avg;
2838 cfs_rq->runnable_load_sum += sa->load_sum;
2841 attach_entity_load_avg(cfs_rq, se);
2843 if (decayed || migrated)
2844 update_tg_load_avg(cfs_rq, 0);
2847 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2849 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2851 update_load_avg(se, 1);
2853 cfs_rq->runnable_load_avg =
2854 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2855 cfs_rq->runnable_load_sum =
2856 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2859 #ifndef CONFIG_64BIT
2860 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2862 u64 last_update_time_copy;
2863 u64 last_update_time;
2866 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2868 last_update_time = cfs_rq->avg.last_update_time;
2869 } while (last_update_time != last_update_time_copy);
2871 return last_update_time;
2874 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2876 return cfs_rq->avg.last_update_time;
2881 * Task first catches up with cfs_rq, and then subtract
2882 * itself from the cfs_rq (task must be off the queue now).
2884 void remove_entity_load_avg(struct sched_entity *se)
2886 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2887 u64 last_update_time;
2890 * Newly created task or never used group entity should not be removed
2891 * from its (source) cfs_rq
2893 if (se->avg.last_update_time == 0)
2896 last_update_time = cfs_rq_last_update_time(cfs_rq);
2898 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2899 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2900 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2904 * Update the rq's load with the elapsed running time before entering
2905 * idle. if the last scheduled task is not a CFS task, idle_enter will
2906 * be the only way to update the runnable statistic.
2908 void idle_enter_fair(struct rq *this_rq)
2913 * Update the rq's load with the elapsed idle time before a task is
2914 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2915 * be the only way to update the runnable statistic.
2917 void idle_exit_fair(struct rq *this_rq)
2921 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2923 return cfs_rq->runnable_load_avg;
2926 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2928 return cfs_rq->avg.load_avg;
2931 static int idle_balance(struct rq *this_rq);
2933 #else /* CONFIG_SMP */
2935 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2937 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2939 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2940 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2943 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2945 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2947 static inline int idle_balance(struct rq *rq)
2952 #endif /* CONFIG_SMP */
2954 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2956 #ifdef CONFIG_SCHEDSTATS
2957 struct task_struct *tsk = NULL;
2959 if (entity_is_task(se))
2962 if (se->statistics.sleep_start) {
2963 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2968 if (unlikely(delta > se->statistics.sleep_max))
2969 se->statistics.sleep_max = delta;
2971 se->statistics.sleep_start = 0;
2972 se->statistics.sum_sleep_runtime += delta;
2975 account_scheduler_latency(tsk, delta >> 10, 1);
2976 trace_sched_stat_sleep(tsk, delta);
2979 if (se->statistics.block_start) {
2980 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2985 if (unlikely(delta > se->statistics.block_max))
2986 se->statistics.block_max = delta;
2988 se->statistics.block_start = 0;
2989 se->statistics.sum_sleep_runtime += delta;
2992 if (tsk->in_iowait) {
2993 se->statistics.iowait_sum += delta;
2994 se->statistics.iowait_count++;
2995 trace_sched_stat_iowait(tsk, delta);
2998 trace_sched_stat_blocked(tsk, delta);
3001 * Blocking time is in units of nanosecs, so shift by
3002 * 20 to get a milliseconds-range estimation of the
3003 * amount of time that the task spent sleeping:
3005 if (unlikely(prof_on == SLEEP_PROFILING)) {
3006 profile_hits(SLEEP_PROFILING,
3007 (void *)get_wchan(tsk),
3010 account_scheduler_latency(tsk, delta >> 10, 0);
3016 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3018 #ifdef CONFIG_SCHED_DEBUG
3019 s64 d = se->vruntime - cfs_rq->min_vruntime;
3024 if (d > 3*sysctl_sched_latency)
3025 schedstat_inc(cfs_rq, nr_spread_over);
3030 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3032 u64 vruntime = cfs_rq->min_vruntime;
3035 * The 'current' period is already promised to the current tasks,
3036 * however the extra weight of the new task will slow them down a
3037 * little, place the new task so that it fits in the slot that
3038 * stays open at the end.
3040 if (initial && sched_feat(START_DEBIT))
3041 vruntime += sched_vslice(cfs_rq, se);
3043 /* sleeps up to a single latency don't count. */
3045 unsigned long thresh = sysctl_sched_latency;
3048 * Halve their sleep time's effect, to allow
3049 * for a gentler effect of sleepers:
3051 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3057 /* ensure we never gain time by being placed backwards. */
3058 se->vruntime = max_vruntime(se->vruntime, vruntime);
3061 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3064 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3067 * Update the normalized vruntime before updating min_vruntime
3068 * through calling update_curr().
3070 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3071 se->vruntime += cfs_rq->min_vruntime;
3074 * Update run-time statistics of the 'current'.
3076 update_curr(cfs_rq);
3077 enqueue_entity_load_avg(cfs_rq, se);
3078 account_entity_enqueue(cfs_rq, se);
3079 update_cfs_shares(cfs_rq);
3081 if (flags & ENQUEUE_WAKEUP) {
3082 place_entity(cfs_rq, se, 0);
3083 enqueue_sleeper(cfs_rq, se);
3086 update_stats_enqueue(cfs_rq, se);
3087 check_spread(cfs_rq, se);
3088 if (se != cfs_rq->curr)
3089 __enqueue_entity(cfs_rq, se);
3092 if (cfs_rq->nr_running == 1) {
3093 list_add_leaf_cfs_rq(cfs_rq);
3094 check_enqueue_throttle(cfs_rq);
3098 static void __clear_buddies_last(struct sched_entity *se)
3100 for_each_sched_entity(se) {
3101 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3102 if (cfs_rq->last != se)
3105 cfs_rq->last = NULL;
3109 static void __clear_buddies_next(struct sched_entity *se)
3111 for_each_sched_entity(se) {
3112 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3113 if (cfs_rq->next != se)
3116 cfs_rq->next = NULL;
3120 static void __clear_buddies_skip(struct sched_entity *se)
3122 for_each_sched_entity(se) {
3123 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3124 if (cfs_rq->skip != se)
3127 cfs_rq->skip = NULL;
3131 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3133 if (cfs_rq->last == se)
3134 __clear_buddies_last(se);
3136 if (cfs_rq->next == se)
3137 __clear_buddies_next(se);
3139 if (cfs_rq->skip == se)
3140 __clear_buddies_skip(se);
3143 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3146 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3149 * Update run-time statistics of the 'current'.
3151 update_curr(cfs_rq);
3152 dequeue_entity_load_avg(cfs_rq, se);
3154 update_stats_dequeue(cfs_rq, se);
3155 if (flags & DEQUEUE_SLEEP) {
3156 #ifdef CONFIG_SCHEDSTATS
3157 if (entity_is_task(se)) {
3158 struct task_struct *tsk = task_of(se);
3160 if (tsk->state & TASK_INTERRUPTIBLE)
3161 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3162 if (tsk->state & TASK_UNINTERRUPTIBLE)
3163 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3168 clear_buddies(cfs_rq, se);
3170 if (se != cfs_rq->curr)
3171 __dequeue_entity(cfs_rq, se);
3173 account_entity_dequeue(cfs_rq, se);
3176 * Normalize the entity after updating the min_vruntime because the
3177 * update can refer to the ->curr item and we need to reflect this
3178 * movement in our normalized position.
3180 if (!(flags & DEQUEUE_SLEEP))
3181 se->vruntime -= cfs_rq->min_vruntime;
3183 /* return excess runtime on last dequeue */
3184 return_cfs_rq_runtime(cfs_rq);
3186 update_min_vruntime(cfs_rq);
3187 update_cfs_shares(cfs_rq);
3191 * Preempt the current task with a newly woken task if needed:
3194 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3196 unsigned long ideal_runtime, delta_exec;
3197 struct sched_entity *se;
3200 ideal_runtime = sched_slice(cfs_rq, curr);
3201 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3202 if (delta_exec > ideal_runtime) {
3203 resched_curr(rq_of(cfs_rq));
3205 * The current task ran long enough, ensure it doesn't get
3206 * re-elected due to buddy favours.
3208 clear_buddies(cfs_rq, curr);
3213 * Ensure that a task that missed wakeup preemption by a
3214 * narrow margin doesn't have to wait for a full slice.
3215 * This also mitigates buddy induced latencies under load.
3217 if (delta_exec < sysctl_sched_min_granularity)
3220 se = __pick_first_entity(cfs_rq);
3221 delta = curr->vruntime - se->vruntime;
3226 if (delta > ideal_runtime)
3227 resched_curr(rq_of(cfs_rq));
3231 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3233 /* 'current' is not kept within the tree. */
3236 * Any task has to be enqueued before it get to execute on
3237 * a CPU. So account for the time it spent waiting on the
3240 update_stats_wait_end(cfs_rq, se);
3241 __dequeue_entity(cfs_rq, se);
3242 update_load_avg(se, 1);
3245 update_stats_curr_start(cfs_rq, se);
3247 #ifdef CONFIG_SCHEDSTATS
3249 * Track our maximum slice length, if the CPU's load is at
3250 * least twice that of our own weight (i.e. dont track it
3251 * when there are only lesser-weight tasks around):
3253 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3254 se->statistics.slice_max = max(se->statistics.slice_max,
3255 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3258 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3262 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3265 * Pick the next process, keeping these things in mind, in this order:
3266 * 1) keep things fair between processes/task groups
3267 * 2) pick the "next" process, since someone really wants that to run
3268 * 3) pick the "last" process, for cache locality
3269 * 4) do not run the "skip" process, if something else is available
3271 static struct sched_entity *
3272 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3274 struct sched_entity *left = __pick_first_entity(cfs_rq);
3275 struct sched_entity *se;
3278 * If curr is set we have to see if its left of the leftmost entity
3279 * still in the tree, provided there was anything in the tree at all.
3281 if (!left || (curr && entity_before(curr, left)))
3284 se = left; /* ideally we run the leftmost entity */
3287 * Avoid running the skip buddy, if running something else can
3288 * be done without getting too unfair.
3290 if (cfs_rq->skip == se) {
3291 struct sched_entity *second;
3294 second = __pick_first_entity(cfs_rq);
3296 second = __pick_next_entity(se);
3297 if (!second || (curr && entity_before(curr, second)))
3301 if (second && wakeup_preempt_entity(second, left) < 1)
3306 * Prefer last buddy, try to return the CPU to a preempted task.
3308 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3312 * Someone really wants this to run. If it's not unfair, run it.
3314 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3317 clear_buddies(cfs_rq, se);
3322 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3324 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3327 * If still on the runqueue then deactivate_task()
3328 * was not called and update_curr() has to be done:
3331 update_curr(cfs_rq);
3333 /* throttle cfs_rqs exceeding runtime */
3334 check_cfs_rq_runtime(cfs_rq);
3336 check_spread(cfs_rq, prev);
3338 update_stats_wait_start(cfs_rq, prev);
3339 /* Put 'current' back into the tree. */
3340 __enqueue_entity(cfs_rq, prev);
3341 /* in !on_rq case, update occurred at dequeue */
3342 update_load_avg(prev, 0);
3344 cfs_rq->curr = NULL;
3348 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3351 * Update run-time statistics of the 'current'.
3353 update_curr(cfs_rq);
3356 * Ensure that runnable average is periodically updated.
3358 update_load_avg(curr, 1);
3359 update_cfs_shares(cfs_rq);
3361 #ifdef CONFIG_SCHED_HRTICK
3363 * queued ticks are scheduled to match the slice, so don't bother
3364 * validating it and just reschedule.
3367 resched_curr(rq_of(cfs_rq));
3371 * don't let the period tick interfere with the hrtick preemption
3373 if (!sched_feat(DOUBLE_TICK) &&
3374 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3378 if (cfs_rq->nr_running > 1)
3379 check_preempt_tick(cfs_rq, curr);
3383 /**************************************************
3384 * CFS bandwidth control machinery
3387 #ifdef CONFIG_CFS_BANDWIDTH
3389 #ifdef HAVE_JUMP_LABEL
3390 static struct static_key __cfs_bandwidth_used;
3392 static inline bool cfs_bandwidth_used(void)
3394 return static_key_false(&__cfs_bandwidth_used);
3397 void cfs_bandwidth_usage_inc(void)
3399 static_key_slow_inc(&__cfs_bandwidth_used);
3402 void cfs_bandwidth_usage_dec(void)
3404 static_key_slow_dec(&__cfs_bandwidth_used);
3406 #else /* HAVE_JUMP_LABEL */
3407 static bool cfs_bandwidth_used(void)
3412 void cfs_bandwidth_usage_inc(void) {}
3413 void cfs_bandwidth_usage_dec(void) {}
3414 #endif /* HAVE_JUMP_LABEL */
3417 * default period for cfs group bandwidth.
3418 * default: 0.1s, units: nanoseconds
3420 static inline u64 default_cfs_period(void)
3422 return 100000000ULL;
3425 static inline u64 sched_cfs_bandwidth_slice(void)
3427 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3431 * Replenish runtime according to assigned quota and update expiration time.
3432 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3433 * additional synchronization around rq->lock.
3435 * requires cfs_b->lock
3437 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3441 if (cfs_b->quota == RUNTIME_INF)
3444 now = sched_clock_cpu(smp_processor_id());
3445 cfs_b->runtime = cfs_b->quota;
3446 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3449 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3451 return &tg->cfs_bandwidth;
3454 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3455 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3457 if (unlikely(cfs_rq->throttle_count))
3458 return cfs_rq->throttled_clock_task;
3460 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3463 /* returns 0 on failure to allocate runtime */
3464 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3466 struct task_group *tg = cfs_rq->tg;
3467 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3468 u64 amount = 0, min_amount, expires;
3470 /* note: this is a positive sum as runtime_remaining <= 0 */
3471 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3473 raw_spin_lock(&cfs_b->lock);
3474 if (cfs_b->quota == RUNTIME_INF)
3475 amount = min_amount;
3477 start_cfs_bandwidth(cfs_b);
3479 if (cfs_b->runtime > 0) {
3480 amount = min(cfs_b->runtime, min_amount);
3481 cfs_b->runtime -= amount;
3485 expires = cfs_b->runtime_expires;
3486 raw_spin_unlock(&cfs_b->lock);
3488 cfs_rq->runtime_remaining += amount;
3490 * we may have advanced our local expiration to account for allowed
3491 * spread between our sched_clock and the one on which runtime was
3494 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3495 cfs_rq->runtime_expires = expires;
3497 return cfs_rq->runtime_remaining > 0;
3501 * Note: This depends on the synchronization provided by sched_clock and the
3502 * fact that rq->clock snapshots this value.
3504 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3506 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3508 /* if the deadline is ahead of our clock, nothing to do */
3509 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3512 if (cfs_rq->runtime_remaining < 0)
3516 * If the local deadline has passed we have to consider the
3517 * possibility that our sched_clock is 'fast' and the global deadline
3518 * has not truly expired.
3520 * Fortunately we can check determine whether this the case by checking
3521 * whether the global deadline has advanced. It is valid to compare
3522 * cfs_b->runtime_expires without any locks since we only care about
3523 * exact equality, so a partial write will still work.
3526 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3527 /* extend local deadline, drift is bounded above by 2 ticks */
3528 cfs_rq->runtime_expires += TICK_NSEC;
3530 /* global deadline is ahead, expiration has passed */
3531 cfs_rq->runtime_remaining = 0;
3535 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3537 /* dock delta_exec before expiring quota (as it could span periods) */
3538 cfs_rq->runtime_remaining -= delta_exec;
3539 expire_cfs_rq_runtime(cfs_rq);
3541 if (likely(cfs_rq->runtime_remaining > 0))
3545 * if we're unable to extend our runtime we resched so that the active
3546 * hierarchy can be throttled
3548 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3549 resched_curr(rq_of(cfs_rq));
3552 static __always_inline
3553 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3555 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3558 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3561 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3563 return cfs_bandwidth_used() && cfs_rq->throttled;
3566 /* check whether cfs_rq, or any parent, is throttled */
3567 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3569 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3573 * Ensure that neither of the group entities corresponding to src_cpu or
3574 * dest_cpu are members of a throttled hierarchy when performing group
3575 * load-balance operations.
3577 static inline int throttled_lb_pair(struct task_group *tg,
3578 int src_cpu, int dest_cpu)
3580 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3582 src_cfs_rq = tg->cfs_rq[src_cpu];
3583 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3585 return throttled_hierarchy(src_cfs_rq) ||
3586 throttled_hierarchy(dest_cfs_rq);
3589 /* updated child weight may affect parent so we have to do this bottom up */
3590 static int tg_unthrottle_up(struct task_group *tg, void *data)
3592 struct rq *rq = data;
3593 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3595 cfs_rq->throttle_count--;
3597 if (!cfs_rq->throttle_count) {
3598 /* adjust cfs_rq_clock_task() */
3599 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3600 cfs_rq->throttled_clock_task;
3607 static int tg_throttle_down(struct task_group *tg, void *data)
3609 struct rq *rq = data;
3610 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3612 /* group is entering throttled state, stop time */
3613 if (!cfs_rq->throttle_count)
3614 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3615 cfs_rq->throttle_count++;
3620 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3622 struct rq *rq = rq_of(cfs_rq);
3623 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3624 struct sched_entity *se;
3625 long task_delta, dequeue = 1;
3628 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3630 /* freeze hierarchy runnable averages while throttled */
3632 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3635 task_delta = cfs_rq->h_nr_running;
3636 for_each_sched_entity(se) {
3637 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3638 /* throttled entity or throttle-on-deactivate */
3643 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3644 qcfs_rq->h_nr_running -= task_delta;
3646 if (qcfs_rq->load.weight)
3651 sub_nr_running(rq, task_delta);
3653 cfs_rq->throttled = 1;
3654 cfs_rq->throttled_clock = rq_clock(rq);
3655 raw_spin_lock(&cfs_b->lock);
3656 empty = list_empty(&cfs_b->throttled_cfs_rq);
3659 * Add to the _head_ of the list, so that an already-started
3660 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
3661 * not running add to the tail so that later runqueues don't get starved.
3663 if (cfs_b->distribute_running)
3664 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3666 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3669 * If we're the first throttled task, make sure the bandwidth
3673 start_cfs_bandwidth(cfs_b);
3675 raw_spin_unlock(&cfs_b->lock);
3678 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3680 struct rq *rq = rq_of(cfs_rq);
3681 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3682 struct sched_entity *se;
3686 se = cfs_rq->tg->se[cpu_of(rq)];
3688 cfs_rq->throttled = 0;
3690 update_rq_clock(rq);
3692 raw_spin_lock(&cfs_b->lock);
3693 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3694 list_del_rcu(&cfs_rq->throttled_list);
3695 raw_spin_unlock(&cfs_b->lock);
3697 /* update hierarchical throttle state */
3698 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3700 if (!cfs_rq->load.weight)
3703 task_delta = cfs_rq->h_nr_running;
3704 for_each_sched_entity(se) {
3708 cfs_rq = cfs_rq_of(se);
3710 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3711 cfs_rq->h_nr_running += task_delta;
3713 if (cfs_rq_throttled(cfs_rq))
3718 add_nr_running(rq, task_delta);
3720 /* determine whether we need to wake up potentially idle cpu */
3721 if (rq->curr == rq->idle && rq->cfs.nr_running)
3725 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3726 u64 remaining, u64 expires)
3728 struct cfs_rq *cfs_rq;
3730 u64 starting_runtime = remaining;
3733 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3735 struct rq *rq = rq_of(cfs_rq);
3737 raw_spin_lock(&rq->lock);
3738 if (!cfs_rq_throttled(cfs_rq))
3741 runtime = -cfs_rq->runtime_remaining + 1;
3742 if (runtime > remaining)
3743 runtime = remaining;
3744 remaining -= runtime;
3746 cfs_rq->runtime_remaining += runtime;
3747 cfs_rq->runtime_expires = expires;
3749 /* we check whether we're throttled above */
3750 if (cfs_rq->runtime_remaining > 0)
3751 unthrottle_cfs_rq(cfs_rq);
3754 raw_spin_unlock(&rq->lock);
3761 return starting_runtime - remaining;
3765 * Responsible for refilling a task_group's bandwidth and unthrottling its
3766 * cfs_rqs as appropriate. If there has been no activity within the last
3767 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3768 * used to track this state.
3770 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3772 u64 runtime, runtime_expires;
3775 /* no need to continue the timer with no bandwidth constraint */
3776 if (cfs_b->quota == RUNTIME_INF)
3777 goto out_deactivate;
3779 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3780 cfs_b->nr_periods += overrun;
3783 * idle depends on !throttled (for the case of a large deficit), and if
3784 * we're going inactive then everything else can be deferred
3786 if (cfs_b->idle && !throttled)
3787 goto out_deactivate;
3789 __refill_cfs_bandwidth_runtime(cfs_b);
3792 /* mark as potentially idle for the upcoming period */
3797 /* account preceding periods in which throttling occurred */
3798 cfs_b->nr_throttled += overrun;
3800 runtime_expires = cfs_b->runtime_expires;
3803 * This check is repeated as we are holding onto the new bandwidth while
3804 * we unthrottle. This can potentially race with an unthrottled group
3805 * trying to acquire new bandwidth from the global pool. This can result
3806 * in us over-using our runtime if it is all used during this loop, but
3807 * only by limited amounts in that extreme case.
3809 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
3810 runtime = cfs_b->runtime;
3811 cfs_b->distribute_running = 1;
3812 raw_spin_unlock(&cfs_b->lock);
3813 /* we can't nest cfs_b->lock while distributing bandwidth */
3814 runtime = distribute_cfs_runtime(cfs_b, runtime,
3816 raw_spin_lock(&cfs_b->lock);
3818 cfs_b->distribute_running = 0;
3819 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3821 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3825 * While we are ensured activity in the period following an
3826 * unthrottle, this also covers the case in which the new bandwidth is
3827 * insufficient to cover the existing bandwidth deficit. (Forcing the
3828 * timer to remain active while there are any throttled entities.)
3838 /* a cfs_rq won't donate quota below this amount */
3839 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3840 /* minimum remaining period time to redistribute slack quota */
3841 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3842 /* how long we wait to gather additional slack before distributing */
3843 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3846 * Are we near the end of the current quota period?
3848 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3849 * hrtimer base being cleared by hrtimer_start. In the case of
3850 * migrate_hrtimers, base is never cleared, so we are fine.
3852 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3854 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3857 /* if the call-back is running a quota refresh is already occurring */
3858 if (hrtimer_callback_running(refresh_timer))
3861 /* is a quota refresh about to occur? */
3862 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3863 if (remaining < (s64)min_expire)
3869 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3871 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3873 /* if there's a quota refresh soon don't bother with slack */
3874 if (runtime_refresh_within(cfs_b, min_left))
3877 hrtimer_start(&cfs_b->slack_timer,
3878 ns_to_ktime(cfs_bandwidth_slack_period),
3882 /* we know any runtime found here is valid as update_curr() precedes return */
3883 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3885 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3886 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3888 if (slack_runtime <= 0)
3891 raw_spin_lock(&cfs_b->lock);
3892 if (cfs_b->quota != RUNTIME_INF &&
3893 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3894 cfs_b->runtime += slack_runtime;
3896 /* we are under rq->lock, defer unthrottling using a timer */
3897 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3898 !list_empty(&cfs_b->throttled_cfs_rq))
3899 start_cfs_slack_bandwidth(cfs_b);
3901 raw_spin_unlock(&cfs_b->lock);
3903 /* even if it's not valid for return we don't want to try again */
3904 cfs_rq->runtime_remaining -= slack_runtime;
3907 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3909 if (!cfs_bandwidth_used())
3912 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3915 __return_cfs_rq_runtime(cfs_rq);
3919 * This is done with a timer (instead of inline with bandwidth return) since
3920 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3922 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3924 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3927 /* confirm we're still not at a refresh boundary */
3928 raw_spin_lock(&cfs_b->lock);
3929 if (cfs_b->distribute_running) {
3930 raw_spin_unlock(&cfs_b->lock);
3934 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3935 raw_spin_unlock(&cfs_b->lock);
3939 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3940 runtime = cfs_b->runtime;
3942 expires = cfs_b->runtime_expires;
3944 cfs_b->distribute_running = 1;
3946 raw_spin_unlock(&cfs_b->lock);
3951 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3953 raw_spin_lock(&cfs_b->lock);
3954 if (expires == cfs_b->runtime_expires)
3955 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3956 cfs_b->distribute_running = 0;
3957 raw_spin_unlock(&cfs_b->lock);
3961 * When a group wakes up we want to make sure that its quota is not already
3962 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3963 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3965 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3967 if (!cfs_bandwidth_used())
3970 /* Synchronize hierarchical throttle counter: */
3971 if (unlikely(!cfs_rq->throttle_uptodate)) {
3972 struct rq *rq = rq_of(cfs_rq);
3973 struct cfs_rq *pcfs_rq;
3974 struct task_group *tg;
3976 cfs_rq->throttle_uptodate = 1;
3978 /* Get closest up-to-date node, because leaves go first: */
3979 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
3980 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
3981 if (pcfs_rq->throttle_uptodate)
3985 cfs_rq->throttle_count = pcfs_rq->throttle_count;
3986 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3990 /* an active group must be handled by the update_curr()->put() path */
3991 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3994 /* ensure the group is not already throttled */
3995 if (cfs_rq_throttled(cfs_rq))
3998 /* update runtime allocation */
3999 account_cfs_rq_runtime(cfs_rq, 0);
4000 if (cfs_rq->runtime_remaining <= 0)
4001 throttle_cfs_rq(cfs_rq);
4004 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4005 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4007 if (!cfs_bandwidth_used())
4010 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4014 * it's possible for a throttled entity to be forced into a running
4015 * state (e.g. set_curr_task), in this case we're finished.
4017 if (cfs_rq_throttled(cfs_rq))
4020 throttle_cfs_rq(cfs_rq);
4024 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4026 struct cfs_bandwidth *cfs_b =
4027 container_of(timer, struct cfs_bandwidth, slack_timer);
4029 do_sched_cfs_slack_timer(cfs_b);
4031 return HRTIMER_NORESTART;
4034 extern const u64 max_cfs_quota_period;
4036 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4038 struct cfs_bandwidth *cfs_b =
4039 container_of(timer, struct cfs_bandwidth, period_timer);
4044 raw_spin_lock(&cfs_b->lock);
4046 overrun = hrtimer_forward_now(timer, cfs_b->period);
4051 u64 new, old = ktime_to_ns(cfs_b->period);
4054 * Grow period by a factor of 2 to avoid losing precision.
4055 * Precision loss in the quota/period ratio can cause __cfs_schedulable
4059 if (new < max_cfs_quota_period) {
4060 cfs_b->period = ns_to_ktime(new);
4063 pr_warn_ratelimited(
4064 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4066 div_u64(new, NSEC_PER_USEC),
4067 div_u64(cfs_b->quota, NSEC_PER_USEC));
4069 pr_warn_ratelimited(
4070 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4072 div_u64(old, NSEC_PER_USEC),
4073 div_u64(cfs_b->quota, NSEC_PER_USEC));
4076 /* reset count so we don't come right back in here */
4080 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4083 cfs_b->period_active = 0;
4084 raw_spin_unlock(&cfs_b->lock);
4086 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4089 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4091 raw_spin_lock_init(&cfs_b->lock);
4093 cfs_b->quota = RUNTIME_INF;
4094 cfs_b->period = ns_to_ktime(default_cfs_period());
4096 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4097 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4098 cfs_b->period_timer.function = sched_cfs_period_timer;
4099 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4100 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4101 cfs_b->distribute_running = 0;
4104 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4106 cfs_rq->runtime_enabled = 0;
4107 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4110 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4112 lockdep_assert_held(&cfs_b->lock);
4114 if (!cfs_b->period_active) {
4115 cfs_b->period_active = 1;
4116 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4117 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4121 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4123 /* init_cfs_bandwidth() was not called */
4124 if (!cfs_b->throttled_cfs_rq.next)
4127 hrtimer_cancel(&cfs_b->period_timer);
4128 hrtimer_cancel(&cfs_b->slack_timer);
4131 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4133 struct cfs_rq *cfs_rq;
4135 for_each_leaf_cfs_rq(rq, cfs_rq) {
4136 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4138 raw_spin_lock(&cfs_b->lock);
4139 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4140 raw_spin_unlock(&cfs_b->lock);
4144 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4146 struct cfs_rq *cfs_rq;
4148 for_each_leaf_cfs_rq(rq, cfs_rq) {
4149 if (!cfs_rq->runtime_enabled)
4153 * clock_task is not advancing so we just need to make sure
4154 * there's some valid quota amount
4156 cfs_rq->runtime_remaining = 1;
4158 * Offline rq is schedulable till cpu is completely disabled
4159 * in take_cpu_down(), so we prevent new cfs throttling here.
4161 cfs_rq->runtime_enabled = 0;
4163 if (cfs_rq_throttled(cfs_rq))
4164 unthrottle_cfs_rq(cfs_rq);
4168 #else /* CONFIG_CFS_BANDWIDTH */
4169 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4171 return rq_clock_task(rq_of(cfs_rq));
4174 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4175 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4176 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4177 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4179 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4184 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4189 static inline int throttled_lb_pair(struct task_group *tg,
4190 int src_cpu, int dest_cpu)
4195 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4197 #ifdef CONFIG_FAIR_GROUP_SCHED
4198 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4201 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4205 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4206 static inline void update_runtime_enabled(struct rq *rq) {}
4207 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4209 #endif /* CONFIG_CFS_BANDWIDTH */
4211 /**************************************************
4212 * CFS operations on tasks:
4215 #ifdef CONFIG_SCHED_HRTICK
4216 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4218 struct sched_entity *se = &p->se;
4219 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4221 WARN_ON(task_rq(p) != rq);
4223 if (cfs_rq->nr_running > 1) {
4224 u64 slice = sched_slice(cfs_rq, se);
4225 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4226 s64 delta = slice - ran;
4233 hrtick_start(rq, delta);
4238 * called from enqueue/dequeue and updates the hrtick when the
4239 * current task is from our class and nr_running is low enough
4242 static void hrtick_update(struct rq *rq)
4244 struct task_struct *curr = rq->curr;
4246 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4249 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4250 hrtick_start_fair(rq, curr);
4252 #else /* !CONFIG_SCHED_HRTICK */
4254 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4258 static inline void hrtick_update(struct rq *rq)
4264 * The enqueue_task method is called before nr_running is
4265 * increased. Here we update the fair scheduling stats and
4266 * then put the task into the rbtree:
4269 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4271 struct cfs_rq *cfs_rq;
4272 struct sched_entity *se = &p->se;
4274 for_each_sched_entity(se) {
4277 cfs_rq = cfs_rq_of(se);
4278 enqueue_entity(cfs_rq, se, flags);
4281 * end evaluation on encountering a throttled cfs_rq
4283 * note: in the case of encountering a throttled cfs_rq we will
4284 * post the final h_nr_running increment below.
4286 if (cfs_rq_throttled(cfs_rq))
4288 cfs_rq->h_nr_running++;
4290 flags = ENQUEUE_WAKEUP;
4293 for_each_sched_entity(se) {
4294 cfs_rq = cfs_rq_of(se);
4295 cfs_rq->h_nr_running++;
4297 if (cfs_rq_throttled(cfs_rq))
4300 update_load_avg(se, 1);
4301 update_cfs_shares(cfs_rq);
4305 add_nr_running(rq, 1);
4310 static void set_next_buddy(struct sched_entity *se);
4313 * The dequeue_task method is called before nr_running is
4314 * decreased. We remove the task from the rbtree and
4315 * update the fair scheduling stats:
4317 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4319 struct cfs_rq *cfs_rq;
4320 struct sched_entity *se = &p->se;
4321 int task_sleep = flags & DEQUEUE_SLEEP;
4323 for_each_sched_entity(se) {
4324 cfs_rq = cfs_rq_of(se);
4325 dequeue_entity(cfs_rq, se, flags);
4328 * end evaluation on encountering a throttled cfs_rq
4330 * note: in the case of encountering a throttled cfs_rq we will
4331 * post the final h_nr_running decrement below.
4333 if (cfs_rq_throttled(cfs_rq))
4335 cfs_rq->h_nr_running--;
4337 /* Don't dequeue parent if it has other entities besides us */
4338 if (cfs_rq->load.weight) {
4339 /* Avoid re-evaluating load for this entity: */
4340 se = parent_entity(se);
4342 * Bias pick_next to pick a task from this cfs_rq, as
4343 * p is sleeping when it is within its sched_slice.
4345 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4349 flags |= DEQUEUE_SLEEP;
4352 for_each_sched_entity(se) {
4353 cfs_rq = cfs_rq_of(se);
4354 cfs_rq->h_nr_running--;
4356 if (cfs_rq_throttled(cfs_rq))
4359 update_load_avg(se, 1);
4360 update_cfs_shares(cfs_rq);
4364 sub_nr_running(rq, 1);
4372 * per rq 'load' arrray crap; XXX kill this.
4376 * The exact cpuload at various idx values, calculated at every tick would be
4377 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4379 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4380 * on nth tick when cpu may be busy, then we have:
4381 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4382 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4384 * decay_load_missed() below does efficient calculation of
4385 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4386 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4388 * The calculation is approximated on a 128 point scale.
4389 * degrade_zero_ticks is the number of ticks after which load at any
4390 * particular idx is approximated to be zero.
4391 * degrade_factor is a precomputed table, a row for each load idx.
4392 * Each column corresponds to degradation factor for a power of two ticks,
4393 * based on 128 point scale.
4395 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4396 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4398 * With this power of 2 load factors, we can degrade the load n times
4399 * by looking at 1 bits in n and doing as many mult/shift instead of
4400 * n mult/shifts needed by the exact degradation.
4402 #define DEGRADE_SHIFT 7
4403 static const unsigned char
4404 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4405 static const unsigned char
4406 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4407 {0, 0, 0, 0, 0, 0, 0, 0},
4408 {64, 32, 8, 0, 0, 0, 0, 0},
4409 {96, 72, 40, 12, 1, 0, 0},
4410 {112, 98, 75, 43, 15, 1, 0},
4411 {120, 112, 98, 76, 45, 16, 2} };
4414 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4415 * would be when CPU is idle and so we just decay the old load without
4416 * adding any new load.
4418 static unsigned long
4419 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4423 if (!missed_updates)
4426 if (missed_updates >= degrade_zero_ticks[idx])
4430 return load >> missed_updates;
4432 while (missed_updates) {
4433 if (missed_updates % 2)
4434 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4436 missed_updates >>= 1;
4443 * Update rq->cpu_load[] statistics. This function is usually called every
4444 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4445 * every tick. We fix it up based on jiffies.
4447 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4448 unsigned long pending_updates)
4452 this_rq->nr_load_updates++;
4454 /* Update our load: */
4455 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4456 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4457 unsigned long old_load, new_load;
4459 /* scale is effectively 1 << i now, and >> i divides by scale */
4461 old_load = this_rq->cpu_load[i];
4462 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4463 new_load = this_load;
4465 * Round up the averaging division if load is increasing. This
4466 * prevents us from getting stuck on 9 if the load is 10, for
4469 if (new_load > old_load)
4470 new_load += scale - 1;
4472 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4475 sched_avg_update(this_rq);
4478 /* Used instead of source_load when we know the type == 0 */
4479 static unsigned long weighted_cpuload(const int cpu)
4481 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4484 #ifdef CONFIG_NO_HZ_COMMON
4486 * There is no sane way to deal with nohz on smp when using jiffies because the
4487 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4488 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4490 * Therefore we cannot use the delta approach from the regular tick since that
4491 * would seriously skew the load calculation. However we'll make do for those
4492 * updates happening while idle (nohz_idle_balance) or coming out of idle
4493 * (tick_nohz_idle_exit).
4495 * This means we might still be one tick off for nohz periods.
4499 * Called from nohz_idle_balance() to update the load ratings before doing the
4502 static void update_idle_cpu_load(struct rq *this_rq)
4504 unsigned long curr_jiffies = READ_ONCE(jiffies);
4505 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4506 unsigned long pending_updates;
4509 * bail if there's load or we're actually up-to-date.
4511 if (load || curr_jiffies == this_rq->last_load_update_tick)
4514 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4515 this_rq->last_load_update_tick = curr_jiffies;
4517 __update_cpu_load(this_rq, load, pending_updates);
4521 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4523 void update_cpu_load_nohz(void)
4525 struct rq *this_rq = this_rq();
4526 unsigned long curr_jiffies = READ_ONCE(jiffies);
4527 unsigned long pending_updates;
4529 if (curr_jiffies == this_rq->last_load_update_tick)
4532 raw_spin_lock(&this_rq->lock);
4533 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4534 if (pending_updates) {
4535 this_rq->last_load_update_tick = curr_jiffies;
4537 * We were idle, this means load 0, the current load might be
4538 * !0 due to remote wakeups and the sort.
4540 __update_cpu_load(this_rq, 0, pending_updates);
4542 raw_spin_unlock(&this_rq->lock);
4544 #endif /* CONFIG_NO_HZ */
4547 * Called from scheduler_tick()
4549 void update_cpu_load_active(struct rq *this_rq)
4551 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4553 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4555 this_rq->last_load_update_tick = jiffies;
4556 __update_cpu_load(this_rq, load, 1);
4560 * Return a low guess at the load of a migration-source cpu weighted
4561 * according to the scheduling class and "nice" value.
4563 * We want to under-estimate the load of migration sources, to
4564 * balance conservatively.
4566 static unsigned long source_load(int cpu, int type)
4568 struct rq *rq = cpu_rq(cpu);
4569 unsigned long total = weighted_cpuload(cpu);
4571 if (type == 0 || !sched_feat(LB_BIAS))
4574 return min(rq->cpu_load[type-1], total);
4578 * Return a high guess at the load of a migration-target cpu weighted
4579 * according to the scheduling class and "nice" value.
4581 static unsigned long target_load(int cpu, int type)
4583 struct rq *rq = cpu_rq(cpu);
4584 unsigned long total = weighted_cpuload(cpu);
4586 if (type == 0 || !sched_feat(LB_BIAS))
4589 return max(rq->cpu_load[type-1], total);
4592 static unsigned long capacity_of(int cpu)
4594 return cpu_rq(cpu)->cpu_capacity;
4597 static unsigned long capacity_orig_of(int cpu)
4599 return cpu_rq(cpu)->cpu_capacity_orig;
4602 static unsigned long cpu_avg_load_per_task(int cpu)
4604 struct rq *rq = cpu_rq(cpu);
4605 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4606 unsigned long load_avg = weighted_cpuload(cpu);
4609 return load_avg / nr_running;
4614 static void record_wakee(struct task_struct *p)
4617 * Rough decay (wiping) for cost saving, don't worry
4618 * about the boundary, really active task won't care
4621 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4622 current->wakee_flips >>= 1;
4623 current->wakee_flip_decay_ts = jiffies;
4626 if (current->last_wakee != p) {
4627 current->last_wakee = p;
4628 current->wakee_flips++;
4632 static void task_waking_fair(struct task_struct *p)
4634 struct sched_entity *se = &p->se;
4635 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4638 #ifndef CONFIG_64BIT
4639 u64 min_vruntime_copy;
4642 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4644 min_vruntime = cfs_rq->min_vruntime;
4645 } while (min_vruntime != min_vruntime_copy);
4647 min_vruntime = cfs_rq->min_vruntime;
4650 se->vruntime -= min_vruntime;
4654 #ifdef CONFIG_FAIR_GROUP_SCHED
4656 * effective_load() calculates the load change as seen from the root_task_group
4658 * Adding load to a group doesn't make a group heavier, but can cause movement
4659 * of group shares between cpus. Assuming the shares were perfectly aligned one
4660 * can calculate the shift in shares.
4662 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4663 * on this @cpu and results in a total addition (subtraction) of @wg to the
4664 * total group weight.
4666 * Given a runqueue weight distribution (rw_i) we can compute a shares
4667 * distribution (s_i) using:
4669 * s_i = rw_i / \Sum rw_j (1)
4671 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4672 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4673 * shares distribution (s_i):
4675 * rw_i = { 2, 4, 1, 0 }
4676 * s_i = { 2/7, 4/7, 1/7, 0 }
4678 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4679 * task used to run on and the CPU the waker is running on), we need to
4680 * compute the effect of waking a task on either CPU and, in case of a sync
4681 * wakeup, compute the effect of the current task going to sleep.
4683 * So for a change of @wl to the local @cpu with an overall group weight change
4684 * of @wl we can compute the new shares distribution (s'_i) using:
4686 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4688 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4689 * differences in waking a task to CPU 0. The additional task changes the
4690 * weight and shares distributions like:
4692 * rw'_i = { 3, 4, 1, 0 }
4693 * s'_i = { 3/8, 4/8, 1/8, 0 }
4695 * We can then compute the difference in effective weight by using:
4697 * dw_i = S * (s'_i - s_i) (3)
4699 * Where 'S' is the group weight as seen by its parent.
4701 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4702 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4703 * 4/7) times the weight of the group.
4705 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4707 struct sched_entity *se = tg->se[cpu];
4709 if (!tg->parent) /* the trivial, non-cgroup case */
4712 for_each_sched_entity(se) {
4713 struct cfs_rq *cfs_rq = se->my_q;
4714 long W, w = cfs_rq_load_avg(cfs_rq);
4719 * W = @wg + \Sum rw_j
4721 W = wg + atomic_long_read(&tg->load_avg);
4723 /* Ensure \Sum rw_j >= rw_i */
4724 W -= cfs_rq->tg_load_avg_contrib;
4733 * wl = S * s'_i; see (2)
4736 wl = (w * (long)tg->shares) / W;
4741 * Per the above, wl is the new se->load.weight value; since
4742 * those are clipped to [MIN_SHARES, ...) do so now. See
4743 * calc_cfs_shares().
4745 if (wl < MIN_SHARES)
4749 * wl = dw_i = S * (s'_i - s_i); see (3)
4751 wl -= se->avg.load_avg;
4754 * Recursively apply this logic to all parent groups to compute
4755 * the final effective load change on the root group. Since
4756 * only the @tg group gets extra weight, all parent groups can
4757 * only redistribute existing shares. @wl is the shift in shares
4758 * resulting from this level per the above.
4767 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4775 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4776 * A waker of many should wake a different task than the one last awakened
4777 * at a frequency roughly N times higher than one of its wakees. In order
4778 * to determine whether we should let the load spread vs consolodating to
4779 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4780 * partner, and a factor of lls_size higher frequency in the other. With
4781 * both conditions met, we can be relatively sure that the relationship is
4782 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4783 * being client/server, worker/dispatcher, interrupt source or whatever is
4784 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4786 static int wake_wide(struct task_struct *p)
4788 unsigned int master = current->wakee_flips;
4789 unsigned int slave = p->wakee_flips;
4790 int factor = this_cpu_read(sd_llc_size);
4793 swap(master, slave);
4794 if (slave < factor || master < slave * factor)
4799 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4801 s64 this_load, load;
4802 s64 this_eff_load, prev_eff_load;
4803 int idx, this_cpu, prev_cpu;
4804 struct task_group *tg;
4805 unsigned long weight;
4809 this_cpu = smp_processor_id();
4810 prev_cpu = task_cpu(p);
4811 load = source_load(prev_cpu, idx);
4812 this_load = target_load(this_cpu, idx);
4815 * If sync wakeup then subtract the (maximum possible)
4816 * effect of the currently running task from the load
4817 * of the current CPU:
4820 tg = task_group(current);
4821 weight = current->se.avg.load_avg;
4823 this_load += effective_load(tg, this_cpu, -weight, -weight);
4824 load += effective_load(tg, prev_cpu, 0, -weight);
4828 weight = p->se.avg.load_avg;
4831 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4832 * due to the sync cause above having dropped this_load to 0, we'll
4833 * always have an imbalance, but there's really nothing you can do
4834 * about that, so that's good too.
4836 * Otherwise check if either cpus are near enough in load to allow this
4837 * task to be woken on this_cpu.
4839 this_eff_load = 100;
4840 this_eff_load *= capacity_of(prev_cpu);
4842 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4843 prev_eff_load *= capacity_of(this_cpu);
4845 if (this_load > 0) {
4846 this_eff_load *= this_load +
4847 effective_load(tg, this_cpu, weight, weight);
4849 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4852 balanced = this_eff_load <= prev_eff_load;
4854 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4859 schedstat_inc(sd, ttwu_move_affine);
4860 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4866 * find_idlest_group finds and returns the least busy CPU group within the
4869 static struct sched_group *
4870 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4871 int this_cpu, int sd_flag)
4873 struct sched_group *idlest = NULL, *group = sd->groups;
4874 unsigned long min_load = ULONG_MAX, this_load = 0;
4875 int load_idx = sd->forkexec_idx;
4876 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4878 if (sd_flag & SD_BALANCE_WAKE)
4879 load_idx = sd->wake_idx;
4882 unsigned long load, avg_load;
4886 /* Skip over this group if it has no CPUs allowed */
4887 if (!cpumask_intersects(sched_group_cpus(group),
4888 tsk_cpus_allowed(p)))
4891 local_group = cpumask_test_cpu(this_cpu,
4892 sched_group_cpus(group));
4894 /* Tally up the load of all CPUs in the group */
4897 for_each_cpu(i, sched_group_cpus(group)) {
4898 /* Bias balancing toward cpus of our domain */
4900 load = source_load(i, load_idx);
4902 load = target_load(i, load_idx);
4907 /* Adjust by relative CPU capacity of the group */
4908 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4911 this_load = avg_load;
4912 } else if (avg_load < min_load) {
4913 min_load = avg_load;
4916 } while (group = group->next, group != sd->groups);
4918 if (!idlest || 100*this_load < imbalance*min_load)
4924 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4927 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4929 unsigned long load, min_load = ULONG_MAX;
4930 unsigned int min_exit_latency = UINT_MAX;
4931 u64 latest_idle_timestamp = 0;
4932 int least_loaded_cpu = this_cpu;
4933 int shallowest_idle_cpu = -1;
4936 /* Traverse only the allowed CPUs */
4937 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4939 struct rq *rq = cpu_rq(i);
4940 struct cpuidle_state *idle = idle_get_state(rq);
4941 if (idle && idle->exit_latency < min_exit_latency) {
4943 * We give priority to a CPU whose idle state
4944 * has the smallest exit latency irrespective
4945 * of any idle timestamp.
4947 min_exit_latency = idle->exit_latency;
4948 latest_idle_timestamp = rq->idle_stamp;
4949 shallowest_idle_cpu = i;
4950 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4951 rq->idle_stamp > latest_idle_timestamp) {
4953 * If equal or no active idle state, then
4954 * the most recently idled CPU might have
4957 latest_idle_timestamp = rq->idle_stamp;
4958 shallowest_idle_cpu = i;
4960 } else if (shallowest_idle_cpu == -1) {
4961 load = weighted_cpuload(i);
4962 if (load < min_load || (load == min_load && i == this_cpu)) {
4964 least_loaded_cpu = i;
4969 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4973 * Try and locate an idle CPU in the sched_domain.
4975 static int select_idle_sibling(struct task_struct *p, int target)
4977 struct sched_domain *sd;
4978 struct sched_group *sg;
4979 int i = task_cpu(p);
4981 if (idle_cpu(target))
4985 * If the prevous cpu is cache affine and idle, don't be stupid.
4987 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4991 * Otherwise, iterate the domains and find an elegible idle cpu.
4993 sd = rcu_dereference(per_cpu(sd_llc, target));
4994 for_each_lower_domain(sd) {
4997 if (!cpumask_intersects(sched_group_cpus(sg),
4998 tsk_cpus_allowed(p)))
5001 for_each_cpu(i, sched_group_cpus(sg)) {
5002 if (i == target || !idle_cpu(i))
5006 target = cpumask_first_and(sched_group_cpus(sg),
5007 tsk_cpus_allowed(p));
5011 } while (sg != sd->groups);
5018 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5019 * tasks. The unit of the return value must be the one of capacity so we can
5020 * compare the utilization with the capacity of the CPU that is available for
5021 * CFS task (ie cpu_capacity).
5023 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5024 * recent utilization of currently non-runnable tasks on a CPU. It represents
5025 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5026 * capacity_orig is the cpu_capacity available at the highest frequency
5027 * (arch_scale_freq_capacity()).
5028 * The utilization of a CPU converges towards a sum equal to or less than the
5029 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5030 * the running time on this CPU scaled by capacity_curr.
5032 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5033 * higher than capacity_orig because of unfortunate rounding in
5034 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5035 * the average stabilizes with the new running time. We need to check that the
5036 * utilization stays within the range of [0..capacity_orig] and cap it if
5037 * necessary. Without utilization capping, a group could be seen as overloaded
5038 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5039 * available capacity. We allow utilization to overshoot capacity_curr (but not
5040 * capacity_orig) as it useful for predicting the capacity required after task
5041 * migrations (scheduler-driven DVFS).
5043 static int cpu_util(int cpu)
5045 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5046 unsigned long capacity = capacity_orig_of(cpu);
5048 return (util >= capacity) ? capacity : util;
5052 * select_task_rq_fair: Select target runqueue for the waking task in domains
5053 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5054 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5056 * Balances load by selecting the idlest cpu in the idlest group, or under
5057 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5059 * Returns the target cpu number.
5061 * preempt must be disabled.
5064 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5066 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5067 int cpu = smp_processor_id();
5068 int new_cpu = prev_cpu;
5069 int want_affine = 0;
5070 int sync = wake_flags & WF_SYNC;
5072 if (sd_flag & SD_BALANCE_WAKE)
5073 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5076 for_each_domain(cpu, tmp) {
5077 if (!(tmp->flags & SD_LOAD_BALANCE))
5081 * If both cpu and prev_cpu are part of this domain,
5082 * cpu is a valid SD_WAKE_AFFINE target.
5084 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5085 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5090 if (tmp->flags & sd_flag)
5092 else if (!want_affine)
5097 sd = NULL; /* Prefer wake_affine over balance flags */
5098 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5103 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5104 new_cpu = select_idle_sibling(p, new_cpu);
5107 struct sched_group *group;
5110 if (!(sd->flags & sd_flag)) {
5115 group = find_idlest_group(sd, p, cpu, sd_flag);
5121 new_cpu = find_idlest_cpu(group, p, cpu);
5122 if (new_cpu == -1 || new_cpu == cpu) {
5123 /* Now try balancing at a lower domain level of cpu */
5128 /* Now try balancing at a lower domain level of new_cpu */
5130 weight = sd->span_weight;
5132 for_each_domain(cpu, tmp) {
5133 if (weight <= tmp->span_weight)
5135 if (tmp->flags & sd_flag)
5138 /* while loop will break here if sd == NULL */
5146 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5147 * cfs_rq_of(p) references at time of call are still valid and identify the
5148 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5149 * other assumptions, including the state of rq->lock, should be made.
5151 static void migrate_task_rq_fair(struct task_struct *p)
5154 * We are supposed to update the task to "current" time, then its up to date
5155 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5156 * what current time is, so simply throw away the out-of-date time. This
5157 * will result in the wakee task is less decayed, but giving the wakee more
5158 * load sounds not bad.
5160 remove_entity_load_avg(&p->se);
5162 /* Tell new CPU we are migrated */
5163 p->se.avg.last_update_time = 0;
5165 /* We have migrated, no longer consider this task hot */
5166 p->se.exec_start = 0;
5169 static void task_dead_fair(struct task_struct *p)
5171 remove_entity_load_avg(&p->se);
5173 #endif /* CONFIG_SMP */
5175 static unsigned long
5176 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5178 unsigned long gran = sysctl_sched_wakeup_granularity;
5181 * Since its curr running now, convert the gran from real-time
5182 * to virtual-time in his units.
5184 * By using 'se' instead of 'curr' we penalize light tasks, so
5185 * they get preempted easier. That is, if 'se' < 'curr' then
5186 * the resulting gran will be larger, therefore penalizing the
5187 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5188 * be smaller, again penalizing the lighter task.
5190 * This is especially important for buddies when the leftmost
5191 * task is higher priority than the buddy.
5193 return calc_delta_fair(gran, se);
5197 * Should 'se' preempt 'curr'.
5211 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5213 s64 gran, vdiff = curr->vruntime - se->vruntime;
5218 gran = wakeup_gran(curr, se);
5225 static void set_last_buddy(struct sched_entity *se)
5227 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5230 for_each_sched_entity(se)
5231 cfs_rq_of(se)->last = se;
5234 static void set_next_buddy(struct sched_entity *se)
5236 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5239 for_each_sched_entity(se)
5240 cfs_rq_of(se)->next = se;
5243 static void set_skip_buddy(struct sched_entity *se)
5245 for_each_sched_entity(se)
5246 cfs_rq_of(se)->skip = se;
5250 * Preempt the current task with a newly woken task if needed:
5252 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5254 struct task_struct *curr = rq->curr;
5255 struct sched_entity *se = &curr->se, *pse = &p->se;
5256 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5257 int scale = cfs_rq->nr_running >= sched_nr_latency;
5258 int next_buddy_marked = 0;
5260 if (unlikely(se == pse))
5264 * This is possible from callers such as attach_tasks(), in which we
5265 * unconditionally check_prempt_curr() after an enqueue (which may have
5266 * lead to a throttle). This both saves work and prevents false
5267 * next-buddy nomination below.
5269 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5272 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5273 set_next_buddy(pse);
5274 next_buddy_marked = 1;
5278 * We can come here with TIF_NEED_RESCHED already set from new task
5281 * Note: this also catches the edge-case of curr being in a throttled
5282 * group (e.g. via set_curr_task), since update_curr() (in the
5283 * enqueue of curr) will have resulted in resched being set. This
5284 * prevents us from potentially nominating it as a false LAST_BUDDY
5287 if (test_tsk_need_resched(curr))
5290 /* Idle tasks are by definition preempted by non-idle tasks. */
5291 if (unlikely(curr->policy == SCHED_IDLE) &&
5292 likely(p->policy != SCHED_IDLE))
5296 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5297 * is driven by the tick):
5299 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5302 find_matching_se(&se, &pse);
5303 update_curr(cfs_rq_of(se));
5305 if (wakeup_preempt_entity(se, pse) == 1) {
5307 * Bias pick_next to pick the sched entity that is
5308 * triggering this preemption.
5310 if (!next_buddy_marked)
5311 set_next_buddy(pse);
5320 * Only set the backward buddy when the current task is still
5321 * on the rq. This can happen when a wakeup gets interleaved
5322 * with schedule on the ->pre_schedule() or idle_balance()
5323 * point, either of which can * drop the rq lock.
5325 * Also, during early boot the idle thread is in the fair class,
5326 * for obvious reasons its a bad idea to schedule back to it.
5328 if (unlikely(!se->on_rq || curr == rq->idle))
5331 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5335 static struct task_struct *
5336 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5338 struct cfs_rq *cfs_rq = &rq->cfs;
5339 struct sched_entity *se;
5340 struct task_struct *p;
5344 #ifdef CONFIG_FAIR_GROUP_SCHED
5345 if (!cfs_rq->nr_running)
5348 if (prev->sched_class != &fair_sched_class)
5352 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5353 * likely that a next task is from the same cgroup as the current.
5355 * Therefore attempt to avoid putting and setting the entire cgroup
5356 * hierarchy, only change the part that actually changes.
5360 struct sched_entity *curr = cfs_rq->curr;
5363 * Since we got here without doing put_prev_entity() we also
5364 * have to consider cfs_rq->curr. If it is still a runnable
5365 * entity, update_curr() will update its vruntime, otherwise
5366 * forget we've ever seen it.
5370 update_curr(cfs_rq);
5375 * This call to check_cfs_rq_runtime() will do the
5376 * throttle and dequeue its entity in the parent(s).
5377 * Therefore the 'simple' nr_running test will indeed
5380 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5384 se = pick_next_entity(cfs_rq, curr);
5385 cfs_rq = group_cfs_rq(se);
5391 * Since we haven't yet done put_prev_entity and if the selected task
5392 * is a different task than we started out with, try and touch the
5393 * least amount of cfs_rqs.
5396 struct sched_entity *pse = &prev->se;
5398 while (!(cfs_rq = is_same_group(se, pse))) {
5399 int se_depth = se->depth;
5400 int pse_depth = pse->depth;
5402 if (se_depth <= pse_depth) {
5403 put_prev_entity(cfs_rq_of(pse), pse);
5404 pse = parent_entity(pse);
5406 if (se_depth >= pse_depth) {
5407 set_next_entity(cfs_rq_of(se), se);
5408 se = parent_entity(se);
5412 put_prev_entity(cfs_rq, pse);
5413 set_next_entity(cfs_rq, se);
5416 if (hrtick_enabled(rq))
5417 hrtick_start_fair(rq, p);
5424 if (!cfs_rq->nr_running)
5427 put_prev_task(rq, prev);
5430 se = pick_next_entity(cfs_rq, NULL);
5431 set_next_entity(cfs_rq, se);
5432 cfs_rq = group_cfs_rq(se);
5437 if (hrtick_enabled(rq))
5438 hrtick_start_fair(rq, p);
5444 * This is OK, because current is on_cpu, which avoids it being picked
5445 * for load-balance and preemption/IRQs are still disabled avoiding
5446 * further scheduler activity on it and we're being very careful to
5447 * re-start the picking loop.
5449 lockdep_unpin_lock(&rq->lock);
5450 new_tasks = idle_balance(rq);
5451 lockdep_pin_lock(&rq->lock);
5453 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5454 * possible for any higher priority task to appear. In that case we
5455 * must re-start the pick_next_entity() loop.
5467 * Account for a descheduled task:
5469 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5471 struct sched_entity *se = &prev->se;
5472 struct cfs_rq *cfs_rq;
5474 for_each_sched_entity(se) {
5475 cfs_rq = cfs_rq_of(se);
5476 put_prev_entity(cfs_rq, se);
5481 * sched_yield() is very simple
5483 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5485 static void yield_task_fair(struct rq *rq)
5487 struct task_struct *curr = rq->curr;
5488 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5489 struct sched_entity *se = &curr->se;
5492 * Are we the only task in the tree?
5494 if (unlikely(rq->nr_running == 1))
5497 clear_buddies(cfs_rq, se);
5499 if (curr->policy != SCHED_BATCH) {
5500 update_rq_clock(rq);
5502 * Update run-time statistics of the 'current'.
5504 update_curr(cfs_rq);
5506 * Tell update_rq_clock() that we've just updated,
5507 * so we don't do microscopic update in schedule()
5508 * and double the fastpath cost.
5510 rq_clock_skip_update(rq, true);
5516 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5518 struct sched_entity *se = &p->se;
5520 /* throttled hierarchies are not runnable */
5521 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5524 /* Tell the scheduler that we'd really like pse to run next. */
5527 yield_task_fair(rq);
5533 /**************************************************
5534 * Fair scheduling class load-balancing methods.
5538 * The purpose of load-balancing is to achieve the same basic fairness the
5539 * per-cpu scheduler provides, namely provide a proportional amount of compute
5540 * time to each task. This is expressed in the following equation:
5542 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5544 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5545 * W_i,0 is defined as:
5547 * W_i,0 = \Sum_j w_i,j (2)
5549 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5550 * is derived from the nice value as per prio_to_weight[].
5552 * The weight average is an exponential decay average of the instantaneous
5555 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5557 * C_i is the compute capacity of cpu i, typically it is the
5558 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5559 * can also include other factors [XXX].
5561 * To achieve this balance we define a measure of imbalance which follows
5562 * directly from (1):
5564 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5566 * We them move tasks around to minimize the imbalance. In the continuous
5567 * function space it is obvious this converges, in the discrete case we get
5568 * a few fun cases generally called infeasible weight scenarios.
5571 * - infeasible weights;
5572 * - local vs global optima in the discrete case. ]
5577 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5578 * for all i,j solution, we create a tree of cpus that follows the hardware
5579 * topology where each level pairs two lower groups (or better). This results
5580 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5581 * tree to only the first of the previous level and we decrease the frequency
5582 * of load-balance at each level inv. proportional to the number of cpus in
5588 * \Sum { --- * --- * 2^i } = O(n) (5)
5590 * `- size of each group
5591 * | | `- number of cpus doing load-balance
5593 * `- sum over all levels
5595 * Coupled with a limit on how many tasks we can migrate every balance pass,
5596 * this makes (5) the runtime complexity of the balancer.
5598 * An important property here is that each CPU is still (indirectly) connected
5599 * to every other cpu in at most O(log n) steps:
5601 * The adjacency matrix of the resulting graph is given by:
5604 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5607 * And you'll find that:
5609 * A^(log_2 n)_i,j != 0 for all i,j (7)
5611 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5612 * The task movement gives a factor of O(m), giving a convergence complexity
5615 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5620 * In order to avoid CPUs going idle while there's still work to do, new idle
5621 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5622 * tree itself instead of relying on other CPUs to bring it work.
5624 * This adds some complexity to both (5) and (8) but it reduces the total idle
5632 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5635 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5640 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5642 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5644 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5647 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5648 * rewrite all of this once again.]
5651 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5653 enum fbq_type { regular, remote, all };
5655 #define LBF_ALL_PINNED 0x01
5656 #define LBF_NEED_BREAK 0x02
5657 #define LBF_DST_PINNED 0x04
5658 #define LBF_SOME_PINNED 0x08
5661 struct sched_domain *sd;
5669 struct cpumask *dst_grpmask;
5671 enum cpu_idle_type idle;
5673 /* The set of CPUs under consideration for load-balancing */
5674 struct cpumask *cpus;
5679 unsigned int loop_break;
5680 unsigned int loop_max;
5682 enum fbq_type fbq_type;
5683 struct list_head tasks;
5687 * Is this task likely cache-hot:
5689 static int task_hot(struct task_struct *p, struct lb_env *env)
5693 lockdep_assert_held(&env->src_rq->lock);
5695 if (p->sched_class != &fair_sched_class)
5698 if (unlikely(p->policy == SCHED_IDLE))
5702 * Buddy candidates are cache hot:
5704 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5705 (&p->se == cfs_rq_of(&p->se)->next ||
5706 &p->se == cfs_rq_of(&p->se)->last))
5709 if (sysctl_sched_migration_cost == -1)
5711 if (sysctl_sched_migration_cost == 0)
5714 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5716 return delta < (s64)sysctl_sched_migration_cost;
5719 #ifdef CONFIG_NUMA_BALANCING
5721 * Returns 1, if task migration degrades locality
5722 * Returns 0, if task migration improves locality i.e migration preferred.
5723 * Returns -1, if task migration is not affected by locality.
5725 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5727 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5728 unsigned long src_faults, dst_faults;
5729 int src_nid, dst_nid;
5731 if (!static_branch_likely(&sched_numa_balancing))
5734 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5737 src_nid = cpu_to_node(env->src_cpu);
5738 dst_nid = cpu_to_node(env->dst_cpu);
5740 if (src_nid == dst_nid)
5743 /* Migrating away from the preferred node is always bad. */
5744 if (src_nid == p->numa_preferred_nid) {
5745 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5751 /* Encourage migration to the preferred node. */
5752 if (dst_nid == p->numa_preferred_nid)
5756 src_faults = group_faults(p, src_nid);
5757 dst_faults = group_faults(p, dst_nid);
5759 src_faults = task_faults(p, src_nid);
5760 dst_faults = task_faults(p, dst_nid);
5763 return dst_faults < src_faults;
5767 static inline int migrate_degrades_locality(struct task_struct *p,
5775 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5778 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5782 lockdep_assert_held(&env->src_rq->lock);
5785 * We do not migrate tasks that are:
5786 * 1) throttled_lb_pair, or
5787 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5788 * 3) running (obviously), or
5789 * 4) are cache-hot on their current CPU.
5791 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5794 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5797 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5799 env->flags |= LBF_SOME_PINNED;
5802 * Remember if this task can be migrated to any other cpu in
5803 * our sched_group. We may want to revisit it if we couldn't
5804 * meet load balance goals by pulling other tasks on src_cpu.
5806 * Also avoid computing new_dst_cpu if we have already computed
5807 * one in current iteration.
5809 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5812 /* Prevent to re-select dst_cpu via env's cpus */
5813 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5814 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5815 env->flags |= LBF_DST_PINNED;
5816 env->new_dst_cpu = cpu;
5824 /* Record that we found atleast one task that could run on dst_cpu */
5825 env->flags &= ~LBF_ALL_PINNED;
5827 if (task_running(env->src_rq, p)) {
5828 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5833 * Aggressive migration if:
5834 * 1) destination numa is preferred
5835 * 2) task is cache cold, or
5836 * 3) too many balance attempts have failed.
5838 tsk_cache_hot = migrate_degrades_locality(p, env);
5839 if (tsk_cache_hot == -1)
5840 tsk_cache_hot = task_hot(p, env);
5842 if (tsk_cache_hot <= 0 ||
5843 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5844 if (tsk_cache_hot == 1) {
5845 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5846 schedstat_inc(p, se.statistics.nr_forced_migrations);
5851 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5856 * detach_task() -- detach the task for the migration specified in env
5858 static void detach_task(struct task_struct *p, struct lb_env *env)
5860 lockdep_assert_held(&env->src_rq->lock);
5862 deactivate_task(env->src_rq, p, 0);
5863 p->on_rq = TASK_ON_RQ_MIGRATING;
5864 set_task_cpu(p, env->dst_cpu);
5868 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5869 * part of active balancing operations within "domain".
5871 * Returns a task if successful and NULL otherwise.
5873 static struct task_struct *detach_one_task(struct lb_env *env)
5875 struct task_struct *p, *n;
5877 lockdep_assert_held(&env->src_rq->lock);
5879 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5880 if (!can_migrate_task(p, env))
5883 detach_task(p, env);
5886 * Right now, this is only the second place where
5887 * lb_gained[env->idle] is updated (other is detach_tasks)
5888 * so we can safely collect stats here rather than
5889 * inside detach_tasks().
5891 schedstat_inc(env->sd, lb_gained[env->idle]);
5897 static const unsigned int sched_nr_migrate_break = 32;
5900 * detach_tasks() -- tries to detach up to imbalance weighted load from
5901 * busiest_rq, as part of a balancing operation within domain "sd".
5903 * Returns number of detached tasks if successful and 0 otherwise.
5905 static int detach_tasks(struct lb_env *env)
5907 struct list_head *tasks = &env->src_rq->cfs_tasks;
5908 struct task_struct *p;
5912 lockdep_assert_held(&env->src_rq->lock);
5914 if (env->imbalance <= 0)
5917 while (!list_empty(tasks)) {
5919 * We don't want to steal all, otherwise we may be treated likewise,
5920 * which could at worst lead to a livelock crash.
5922 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5925 p = list_first_entry(tasks, struct task_struct, se.group_node);
5928 /* We've more or less seen every task there is, call it quits */
5929 if (env->loop > env->loop_max)
5932 /* take a breather every nr_migrate tasks */
5933 if (env->loop > env->loop_break) {
5934 env->loop_break += sched_nr_migrate_break;
5935 env->flags |= LBF_NEED_BREAK;
5939 if (!can_migrate_task(p, env))
5943 * Depending of the number of CPUs and tasks and the
5944 * cgroup hierarchy, task_h_load() can return a null
5945 * value. Make sure that env->imbalance decreases
5946 * otherwise detach_tasks() will stop only after
5947 * detaching up to loop_max tasks.
5949 load = max_t(unsigned long, task_h_load(p), 1);
5952 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5955 if ((load / 2) > env->imbalance)
5958 detach_task(p, env);
5959 list_add(&p->se.group_node, &env->tasks);
5962 env->imbalance -= load;
5964 #ifdef CONFIG_PREEMPT
5966 * NEWIDLE balancing is a source of latency, so preemptible
5967 * kernels will stop after the first task is detached to minimize
5968 * the critical section.
5970 if (env->idle == CPU_NEWLY_IDLE)
5975 * We only want to steal up to the prescribed amount of
5978 if (env->imbalance <= 0)
5983 list_move_tail(&p->se.group_node, tasks);
5987 * Right now, this is one of only two places we collect this stat
5988 * so we can safely collect detach_one_task() stats here rather
5989 * than inside detach_one_task().
5991 schedstat_add(env->sd, lb_gained[env->idle], detached);
5997 * attach_task() -- attach the task detached by detach_task() to its new rq.
5999 static void attach_task(struct rq *rq, struct task_struct *p)
6001 lockdep_assert_held(&rq->lock);
6003 BUG_ON(task_rq(p) != rq);
6004 p->on_rq = TASK_ON_RQ_QUEUED;
6005 activate_task(rq, p, 0);
6006 check_preempt_curr(rq, p, 0);
6010 * attach_one_task() -- attaches the task returned from detach_one_task() to
6013 static void attach_one_task(struct rq *rq, struct task_struct *p)
6015 raw_spin_lock(&rq->lock);
6017 raw_spin_unlock(&rq->lock);
6021 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6024 static void attach_tasks(struct lb_env *env)
6026 struct list_head *tasks = &env->tasks;
6027 struct task_struct *p;
6029 raw_spin_lock(&env->dst_rq->lock);
6031 while (!list_empty(tasks)) {
6032 p = list_first_entry(tasks, struct task_struct, se.group_node);
6033 list_del_init(&p->se.group_node);
6035 attach_task(env->dst_rq, p);
6038 raw_spin_unlock(&env->dst_rq->lock);
6041 #ifdef CONFIG_FAIR_GROUP_SCHED
6042 static void update_blocked_averages(int cpu)
6044 struct rq *rq = cpu_rq(cpu);
6045 struct cfs_rq *cfs_rq;
6046 unsigned long flags;
6048 raw_spin_lock_irqsave(&rq->lock, flags);
6049 update_rq_clock(rq);
6052 * Iterates the task_group tree in a bottom up fashion, see
6053 * list_add_leaf_cfs_rq() for details.
6055 for_each_leaf_cfs_rq(rq, cfs_rq) {
6056 /* throttled entities do not contribute to load */
6057 if (throttled_hierarchy(cfs_rq))
6060 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6061 update_tg_load_avg(cfs_rq, 0);
6063 raw_spin_unlock_irqrestore(&rq->lock, flags);
6067 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6068 * This needs to be done in a top-down fashion because the load of a child
6069 * group is a fraction of its parents load.
6071 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6073 struct rq *rq = rq_of(cfs_rq);
6074 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6075 unsigned long now = jiffies;
6078 if (cfs_rq->last_h_load_update == now)
6081 WRITE_ONCE(cfs_rq->h_load_next, NULL);
6082 for_each_sched_entity(se) {
6083 cfs_rq = cfs_rq_of(se);
6084 WRITE_ONCE(cfs_rq->h_load_next, se);
6085 if (cfs_rq->last_h_load_update == now)
6090 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6091 cfs_rq->last_h_load_update = now;
6094 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
6095 load = cfs_rq->h_load;
6096 load = div64_ul(load * se->avg.load_avg,
6097 cfs_rq_load_avg(cfs_rq) + 1);
6098 cfs_rq = group_cfs_rq(se);
6099 cfs_rq->h_load = load;
6100 cfs_rq->last_h_load_update = now;
6104 static unsigned long task_h_load(struct task_struct *p)
6106 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6108 update_cfs_rq_h_load(cfs_rq);
6109 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6110 cfs_rq_load_avg(cfs_rq) + 1);
6113 static inline void update_blocked_averages(int cpu)
6115 struct rq *rq = cpu_rq(cpu);
6116 struct cfs_rq *cfs_rq = &rq->cfs;
6117 unsigned long flags;
6119 raw_spin_lock_irqsave(&rq->lock, flags);
6120 update_rq_clock(rq);
6121 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6122 raw_spin_unlock_irqrestore(&rq->lock, flags);
6125 static unsigned long task_h_load(struct task_struct *p)
6127 return p->se.avg.load_avg;
6131 /********** Helpers for find_busiest_group ************************/
6140 * sg_lb_stats - stats of a sched_group required for load_balancing
6142 struct sg_lb_stats {
6143 unsigned long avg_load; /*Avg load across the CPUs of the group */
6144 unsigned long group_load; /* Total load over the CPUs of the group */
6145 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6146 unsigned long load_per_task;
6147 unsigned long group_capacity;
6148 unsigned long group_util; /* Total utilization of the group */
6149 unsigned int sum_nr_running; /* Nr tasks running in the group */
6150 unsigned int idle_cpus;
6151 unsigned int group_weight;
6152 enum group_type group_type;
6153 int group_no_capacity;
6154 #ifdef CONFIG_NUMA_BALANCING
6155 unsigned int nr_numa_running;
6156 unsigned int nr_preferred_running;
6161 * sd_lb_stats - Structure to store the statistics of a sched_domain
6162 * during load balancing.
6164 struct sd_lb_stats {
6165 struct sched_group *busiest; /* Busiest group in this sd */
6166 struct sched_group *local; /* Local group in this sd */
6167 unsigned long total_load; /* Total load of all groups in sd */
6168 unsigned long total_capacity; /* Total capacity of all groups in sd */
6169 unsigned long avg_load; /* Average load across all groups in sd */
6171 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6172 struct sg_lb_stats local_stat; /* Statistics of the local group */
6175 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6178 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6179 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6180 * We must however clear busiest_stat::avg_load because
6181 * update_sd_pick_busiest() reads this before assignment.
6183 *sds = (struct sd_lb_stats){
6187 .total_capacity = 0UL,
6190 .sum_nr_running = 0,
6191 .group_type = group_other,
6197 * get_sd_load_idx - Obtain the load index for a given sched domain.
6198 * @sd: The sched_domain whose load_idx is to be obtained.
6199 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6201 * Return: The load index.
6203 static inline int get_sd_load_idx(struct sched_domain *sd,
6204 enum cpu_idle_type idle)
6210 load_idx = sd->busy_idx;
6213 case CPU_NEWLY_IDLE:
6214 load_idx = sd->newidle_idx;
6217 load_idx = sd->idle_idx;
6224 static unsigned long scale_rt_capacity(int cpu)
6226 struct rq *rq = cpu_rq(cpu);
6227 u64 total, used, age_stamp, avg;
6231 * Since we're reading these variables without serialization make sure
6232 * we read them once before doing sanity checks on them.
6234 age_stamp = READ_ONCE(rq->age_stamp);
6235 avg = READ_ONCE(rq->rt_avg);
6236 delta = __rq_clock_broken(rq) - age_stamp;
6238 if (unlikely(delta < 0))
6241 total = sched_avg_period() + delta;
6243 used = div_u64(avg, total);
6245 if (likely(used < SCHED_CAPACITY_SCALE))
6246 return SCHED_CAPACITY_SCALE - used;
6251 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6253 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6254 struct sched_group *sdg = sd->groups;
6256 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6258 capacity *= scale_rt_capacity(cpu);
6259 capacity >>= SCHED_CAPACITY_SHIFT;
6264 cpu_rq(cpu)->cpu_capacity = capacity;
6265 sdg->sgc->capacity = capacity;
6268 void update_group_capacity(struct sched_domain *sd, int cpu)
6270 struct sched_domain *child = sd->child;
6271 struct sched_group *group, *sdg = sd->groups;
6272 unsigned long capacity;
6273 unsigned long interval;
6275 interval = msecs_to_jiffies(sd->balance_interval);
6276 interval = clamp(interval, 1UL, max_load_balance_interval);
6277 sdg->sgc->next_update = jiffies + interval;
6280 update_cpu_capacity(sd, cpu);
6286 if (child->flags & SD_OVERLAP) {
6288 * SD_OVERLAP domains cannot assume that child groups
6289 * span the current group.
6292 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6293 struct sched_group_capacity *sgc;
6294 struct rq *rq = cpu_rq(cpu);
6297 * build_sched_domains() -> init_sched_groups_capacity()
6298 * gets here before we've attached the domains to the
6301 * Use capacity_of(), which is set irrespective of domains
6302 * in update_cpu_capacity().
6304 * This avoids capacity from being 0 and
6305 * causing divide-by-zero issues on boot.
6307 if (unlikely(!rq->sd)) {
6308 capacity += capacity_of(cpu);
6312 sgc = rq->sd->groups->sgc;
6313 capacity += sgc->capacity;
6317 * !SD_OVERLAP domains can assume that child groups
6318 * span the current group.
6321 group = child->groups;
6323 capacity += group->sgc->capacity;
6324 group = group->next;
6325 } while (group != child->groups);
6328 sdg->sgc->capacity = capacity;
6332 * Check whether the capacity of the rq has been noticeably reduced by side
6333 * activity. The imbalance_pct is used for the threshold.
6334 * Return true is the capacity is reduced
6337 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6339 return ((rq->cpu_capacity * sd->imbalance_pct) <
6340 (rq->cpu_capacity_orig * 100));
6344 * Group imbalance indicates (and tries to solve) the problem where balancing
6345 * groups is inadequate due to tsk_cpus_allowed() constraints.
6347 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6348 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6351 * { 0 1 2 3 } { 4 5 6 7 }
6354 * If we were to balance group-wise we'd place two tasks in the first group and
6355 * two tasks in the second group. Clearly this is undesired as it will overload
6356 * cpu 3 and leave one of the cpus in the second group unused.
6358 * The current solution to this issue is detecting the skew in the first group
6359 * by noticing the lower domain failed to reach balance and had difficulty
6360 * moving tasks due to affinity constraints.
6362 * When this is so detected; this group becomes a candidate for busiest; see
6363 * update_sd_pick_busiest(). And calculate_imbalance() and
6364 * find_busiest_group() avoid some of the usual balance conditions to allow it
6365 * to create an effective group imbalance.
6367 * This is a somewhat tricky proposition since the next run might not find the
6368 * group imbalance and decide the groups need to be balanced again. A most
6369 * subtle and fragile situation.
6372 static inline int sg_imbalanced(struct sched_group *group)
6374 return group->sgc->imbalance;
6378 * group_has_capacity returns true if the group has spare capacity that could
6379 * be used by some tasks.
6380 * We consider that a group has spare capacity if the * number of task is
6381 * smaller than the number of CPUs or if the utilization is lower than the
6382 * available capacity for CFS tasks.
6383 * For the latter, we use a threshold to stabilize the state, to take into
6384 * account the variance of the tasks' load and to return true if the available
6385 * capacity in meaningful for the load balancer.
6386 * As an example, an available capacity of 1% can appear but it doesn't make
6387 * any benefit for the load balance.
6390 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6392 if (sgs->sum_nr_running < sgs->group_weight)
6395 if ((sgs->group_capacity * 100) >
6396 (sgs->group_util * env->sd->imbalance_pct))
6403 * group_is_overloaded returns true if the group has more tasks than it can
6405 * group_is_overloaded is not equals to !group_has_capacity because a group
6406 * with the exact right number of tasks, has no more spare capacity but is not
6407 * overloaded so both group_has_capacity and group_is_overloaded return
6411 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6413 if (sgs->sum_nr_running <= sgs->group_weight)
6416 if ((sgs->group_capacity * 100) <
6417 (sgs->group_util * env->sd->imbalance_pct))
6424 group_type group_classify(struct sched_group *group,
6425 struct sg_lb_stats *sgs)
6427 if (sgs->group_no_capacity)
6428 return group_overloaded;
6430 if (sg_imbalanced(group))
6431 return group_imbalanced;
6437 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6438 * @env: The load balancing environment.
6439 * @group: sched_group whose statistics are to be updated.
6440 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6441 * @local_group: Does group contain this_cpu.
6442 * @sgs: variable to hold the statistics for this group.
6443 * @overload: Indicate more than one runnable task for any CPU.
6445 static inline void update_sg_lb_stats(struct lb_env *env,
6446 struct sched_group *group, int load_idx,
6447 int local_group, struct sg_lb_stats *sgs,
6453 memset(sgs, 0, sizeof(*sgs));
6455 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6456 struct rq *rq = cpu_rq(i);
6458 /* Bias balancing toward cpus of our domain */
6460 load = target_load(i, load_idx);
6462 load = source_load(i, load_idx);
6464 sgs->group_load += load;
6465 sgs->group_util += cpu_util(i);
6466 sgs->sum_nr_running += rq->cfs.h_nr_running;
6468 if (rq->nr_running > 1)
6471 #ifdef CONFIG_NUMA_BALANCING
6472 sgs->nr_numa_running += rq->nr_numa_running;
6473 sgs->nr_preferred_running += rq->nr_preferred_running;
6475 sgs->sum_weighted_load += weighted_cpuload(i);
6480 /* Adjust by relative CPU capacity of the group */
6481 sgs->group_capacity = group->sgc->capacity;
6482 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6484 if (sgs->sum_nr_running)
6485 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6487 sgs->group_weight = group->group_weight;
6489 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6490 sgs->group_type = group_classify(group, sgs);
6494 * update_sd_pick_busiest - return 1 on busiest group
6495 * @env: The load balancing environment.
6496 * @sds: sched_domain statistics
6497 * @sg: sched_group candidate to be checked for being the busiest
6498 * @sgs: sched_group statistics
6500 * Determine if @sg is a busier group than the previously selected
6503 * Return: %true if @sg is a busier group than the previously selected
6504 * busiest group. %false otherwise.
6506 static bool update_sd_pick_busiest(struct lb_env *env,
6507 struct sd_lb_stats *sds,
6508 struct sched_group *sg,
6509 struct sg_lb_stats *sgs)
6511 struct sg_lb_stats *busiest = &sds->busiest_stat;
6513 if (sgs->group_type > busiest->group_type)
6516 if (sgs->group_type < busiest->group_type)
6519 if (sgs->avg_load <= busiest->avg_load)
6522 /* This is the busiest node in its class. */
6523 if (!(env->sd->flags & SD_ASYM_PACKING))
6527 * ASYM_PACKING needs to move all the work to the lowest
6528 * numbered CPUs in the group, therefore mark all groups
6529 * higher than ourself as busy.
6531 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6535 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6542 #ifdef CONFIG_NUMA_BALANCING
6543 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6545 if (sgs->sum_nr_running > sgs->nr_numa_running)
6547 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6552 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6554 if (rq->nr_running > rq->nr_numa_running)
6556 if (rq->nr_running > rq->nr_preferred_running)
6561 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6566 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6570 #endif /* CONFIG_NUMA_BALANCING */
6573 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6574 * @env: The load balancing environment.
6575 * @sds: variable to hold the statistics for this sched_domain.
6577 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6579 struct sched_domain *child = env->sd->child;
6580 struct sched_group *sg = env->sd->groups;
6581 struct sg_lb_stats tmp_sgs;
6582 int load_idx, prefer_sibling = 0;
6583 bool overload = false;
6585 if (child && child->flags & SD_PREFER_SIBLING)
6588 load_idx = get_sd_load_idx(env->sd, env->idle);
6591 struct sg_lb_stats *sgs = &tmp_sgs;
6594 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6597 sgs = &sds->local_stat;
6599 if (env->idle != CPU_NEWLY_IDLE ||
6600 time_after_eq(jiffies, sg->sgc->next_update))
6601 update_group_capacity(env->sd, env->dst_cpu);
6604 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6611 * In case the child domain prefers tasks go to siblings
6612 * first, lower the sg capacity so that we'll try
6613 * and move all the excess tasks away. We lower the capacity
6614 * of a group only if the local group has the capacity to fit
6615 * these excess tasks. The extra check prevents the case where
6616 * you always pull from the heaviest group when it is already
6617 * under-utilized (possible with a large weight task outweighs
6618 * the tasks on the system).
6620 if (prefer_sibling && sds->local &&
6621 group_has_capacity(env, &sds->local_stat) &&
6622 (sgs->sum_nr_running > 1)) {
6623 sgs->group_no_capacity = 1;
6624 sgs->group_type = group_classify(sg, sgs);
6627 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6629 sds->busiest_stat = *sgs;
6633 /* Now, start updating sd_lb_stats */
6634 sds->total_load += sgs->group_load;
6635 sds->total_capacity += sgs->group_capacity;
6638 } while (sg != env->sd->groups);
6640 if (env->sd->flags & SD_NUMA)
6641 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6643 if (!env->sd->parent) {
6644 /* update overload indicator if we are at root domain */
6645 if (env->dst_rq->rd->overload != overload)
6646 env->dst_rq->rd->overload = overload;
6652 * check_asym_packing - Check to see if the group is packed into the
6655 * This is primarily intended to used at the sibling level. Some
6656 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6657 * case of POWER7, it can move to lower SMT modes only when higher
6658 * threads are idle. When in lower SMT modes, the threads will
6659 * perform better since they share less core resources. Hence when we
6660 * have idle threads, we want them to be the higher ones.
6662 * This packing function is run on idle threads. It checks to see if
6663 * the busiest CPU in this domain (core in the P7 case) has a higher
6664 * CPU number than the packing function is being run on. Here we are
6665 * assuming lower CPU number will be equivalent to lower a SMT thread
6668 * Return: 1 when packing is required and a task should be moved to
6669 * this CPU. The amount of the imbalance is returned in *imbalance.
6671 * @env: The load balancing environment.
6672 * @sds: Statistics of the sched_domain which is to be packed
6674 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6678 if (!(env->sd->flags & SD_ASYM_PACKING))
6684 busiest_cpu = group_first_cpu(sds->busiest);
6685 if (env->dst_cpu > busiest_cpu)
6688 env->imbalance = DIV_ROUND_CLOSEST(
6689 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6690 SCHED_CAPACITY_SCALE);
6696 * fix_small_imbalance - Calculate the minor imbalance that exists
6697 * amongst the groups of a sched_domain, during
6699 * @env: The load balancing environment.
6700 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6703 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6705 unsigned long tmp, capa_now = 0, capa_move = 0;
6706 unsigned int imbn = 2;
6707 unsigned long scaled_busy_load_per_task;
6708 struct sg_lb_stats *local, *busiest;
6710 local = &sds->local_stat;
6711 busiest = &sds->busiest_stat;
6713 if (!local->sum_nr_running)
6714 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6715 else if (busiest->load_per_task > local->load_per_task)
6718 scaled_busy_load_per_task =
6719 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6720 busiest->group_capacity;
6722 if (busiest->avg_load + scaled_busy_load_per_task >=
6723 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6724 env->imbalance = busiest->load_per_task;
6729 * OK, we don't have enough imbalance to justify moving tasks,
6730 * however we may be able to increase total CPU capacity used by
6734 capa_now += busiest->group_capacity *
6735 min(busiest->load_per_task, busiest->avg_load);
6736 capa_now += local->group_capacity *
6737 min(local->load_per_task, local->avg_load);
6738 capa_now /= SCHED_CAPACITY_SCALE;
6740 /* Amount of load we'd subtract */
6741 if (busiest->avg_load > scaled_busy_load_per_task) {
6742 capa_move += busiest->group_capacity *
6743 min(busiest->load_per_task,
6744 busiest->avg_load - scaled_busy_load_per_task);
6747 /* Amount of load we'd add */
6748 if (busiest->avg_load * busiest->group_capacity <
6749 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6750 tmp = (busiest->avg_load * busiest->group_capacity) /
6751 local->group_capacity;
6753 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6754 local->group_capacity;
6756 capa_move += local->group_capacity *
6757 min(local->load_per_task, local->avg_load + tmp);
6758 capa_move /= SCHED_CAPACITY_SCALE;
6760 /* Move if we gain throughput */
6761 if (capa_move > capa_now)
6762 env->imbalance = busiest->load_per_task;
6766 * calculate_imbalance - Calculate the amount of imbalance present within the
6767 * groups of a given sched_domain during load balance.
6768 * @env: load balance environment
6769 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6771 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6773 unsigned long max_pull, load_above_capacity = ~0UL;
6774 struct sg_lb_stats *local, *busiest;
6776 local = &sds->local_stat;
6777 busiest = &sds->busiest_stat;
6779 if (busiest->group_type == group_imbalanced) {
6781 * In the group_imb case we cannot rely on group-wide averages
6782 * to ensure cpu-load equilibrium, look at wider averages. XXX
6784 busiest->load_per_task =
6785 min(busiest->load_per_task, sds->avg_load);
6789 * In the presence of smp nice balancing, certain scenarios can have
6790 * max load less than avg load(as we skip the groups at or below
6791 * its cpu_capacity, while calculating max_load..)
6793 if (busiest->avg_load <= sds->avg_load ||
6794 local->avg_load >= sds->avg_load) {
6796 return fix_small_imbalance(env, sds);
6800 * If there aren't any idle cpus, avoid creating some.
6802 if (busiest->group_type == group_overloaded &&
6803 local->group_type == group_overloaded) {
6804 load_above_capacity = busiest->sum_nr_running *
6806 if (load_above_capacity > busiest->group_capacity)
6807 load_above_capacity -= busiest->group_capacity;
6809 load_above_capacity = ~0UL;
6813 * We're trying to get all the cpus to the average_load, so we don't
6814 * want to push ourselves above the average load, nor do we wish to
6815 * reduce the max loaded cpu below the average load. At the same time,
6816 * we also don't want to reduce the group load below the group capacity
6817 * (so that we can implement power-savings policies etc). Thus we look
6818 * for the minimum possible imbalance.
6820 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6822 /* How much load to actually move to equalise the imbalance */
6823 env->imbalance = min(
6824 max_pull * busiest->group_capacity,
6825 (sds->avg_load - local->avg_load) * local->group_capacity
6826 ) / SCHED_CAPACITY_SCALE;
6829 * if *imbalance is less than the average load per runnable task
6830 * there is no guarantee that any tasks will be moved so we'll have
6831 * a think about bumping its value to force at least one task to be
6834 if (env->imbalance < busiest->load_per_task)
6835 return fix_small_imbalance(env, sds);
6838 /******* find_busiest_group() helpers end here *********************/
6841 * find_busiest_group - Returns the busiest group within the sched_domain
6842 * if there is an imbalance. If there isn't an imbalance, and
6843 * the user has opted for power-savings, it returns a group whose
6844 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6845 * such a group exists.
6847 * Also calculates the amount of weighted load which should be moved
6848 * to restore balance.
6850 * @env: The load balancing environment.
6852 * Return: - The busiest group if imbalance exists.
6853 * - If no imbalance and user has opted for power-savings balance,
6854 * return the least loaded group whose CPUs can be
6855 * put to idle by rebalancing its tasks onto our group.
6857 static struct sched_group *find_busiest_group(struct lb_env *env)
6859 struct sg_lb_stats *local, *busiest;
6860 struct sd_lb_stats sds;
6862 init_sd_lb_stats(&sds);
6865 * Compute the various statistics relavent for load balancing at
6868 update_sd_lb_stats(env, &sds);
6869 local = &sds.local_stat;
6870 busiest = &sds.busiest_stat;
6872 /* ASYM feature bypasses nice load balance check */
6873 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6874 check_asym_packing(env, &sds))
6877 /* There is no busy sibling group to pull tasks from */
6878 if (!sds.busiest || busiest->sum_nr_running == 0)
6881 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6882 / sds.total_capacity;
6885 * If the busiest group is imbalanced the below checks don't
6886 * work because they assume all things are equal, which typically
6887 * isn't true due to cpus_allowed constraints and the like.
6889 if (busiest->group_type == group_imbalanced)
6892 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6893 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6894 busiest->group_no_capacity)
6898 * If the local group is busier than the selected busiest group
6899 * don't try and pull any tasks.
6901 if (local->avg_load >= busiest->avg_load)
6905 * Don't pull any tasks if this group is already above the domain
6908 if (local->avg_load >= sds.avg_load)
6911 if (env->idle == CPU_IDLE) {
6913 * This cpu is idle. If the busiest group is not overloaded
6914 * and there is no imbalance between this and busiest group
6915 * wrt idle cpus, it is balanced. The imbalance becomes
6916 * significant if the diff is greater than 1 otherwise we
6917 * might end up to just move the imbalance on another group
6919 if ((busiest->group_type != group_overloaded) &&
6920 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6924 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6925 * imbalance_pct to be conservative.
6927 if (100 * busiest->avg_load <=
6928 env->sd->imbalance_pct * local->avg_load)
6933 /* Looks like there is an imbalance. Compute it */
6934 calculate_imbalance(env, &sds);
6943 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6945 static struct rq *find_busiest_queue(struct lb_env *env,
6946 struct sched_group *group)
6948 struct rq *busiest = NULL, *rq;
6949 unsigned long busiest_load = 0, busiest_capacity = 1;
6952 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6953 unsigned long capacity, wl;
6957 rt = fbq_classify_rq(rq);
6960 * We classify groups/runqueues into three groups:
6961 * - regular: there are !numa tasks
6962 * - remote: there are numa tasks that run on the 'wrong' node
6963 * - all: there is no distinction
6965 * In order to avoid migrating ideally placed numa tasks,
6966 * ignore those when there's better options.
6968 * If we ignore the actual busiest queue to migrate another
6969 * task, the next balance pass can still reduce the busiest
6970 * queue by moving tasks around inside the node.
6972 * If we cannot move enough load due to this classification
6973 * the next pass will adjust the group classification and
6974 * allow migration of more tasks.
6976 * Both cases only affect the total convergence complexity.
6978 if (rt > env->fbq_type)
6981 capacity = capacity_of(i);
6983 wl = weighted_cpuload(i);
6986 * When comparing with imbalance, use weighted_cpuload()
6987 * which is not scaled with the cpu capacity.
6990 if (rq->nr_running == 1 && wl > env->imbalance &&
6991 !check_cpu_capacity(rq, env->sd))
6995 * For the load comparisons with the other cpu's, consider
6996 * the weighted_cpuload() scaled with the cpu capacity, so
6997 * that the load can be moved away from the cpu that is
6998 * potentially running at a lower capacity.
7000 * Thus we're looking for max(wl_i / capacity_i), crosswise
7001 * multiplication to rid ourselves of the division works out
7002 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7003 * our previous maximum.
7005 if (wl * busiest_capacity > busiest_load * capacity) {
7007 busiest_capacity = capacity;
7016 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7017 * so long as it is large enough.
7019 #define MAX_PINNED_INTERVAL 512
7021 /* Working cpumask for load_balance and load_balance_newidle. */
7022 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7024 static int need_active_balance(struct lb_env *env)
7026 struct sched_domain *sd = env->sd;
7028 if (env->idle == CPU_NEWLY_IDLE) {
7031 * ASYM_PACKING needs to force migrate tasks from busy but
7032 * higher numbered CPUs in order to pack all tasks in the
7033 * lowest numbered CPUs.
7035 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7040 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7041 * It's worth migrating the task if the src_cpu's capacity is reduced
7042 * because of other sched_class or IRQs if more capacity stays
7043 * available on dst_cpu.
7045 if ((env->idle != CPU_NOT_IDLE) &&
7046 (env->src_rq->cfs.h_nr_running == 1)) {
7047 if ((check_cpu_capacity(env->src_rq, sd)) &&
7048 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7052 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7055 static int active_load_balance_cpu_stop(void *data);
7057 static int should_we_balance(struct lb_env *env)
7059 struct sched_group *sg = env->sd->groups;
7060 struct cpumask *sg_cpus, *sg_mask;
7061 int cpu, balance_cpu = -1;
7064 * In the newly idle case, we will allow all the cpu's
7065 * to do the newly idle load balance.
7067 if (env->idle == CPU_NEWLY_IDLE)
7070 sg_cpus = sched_group_cpus(sg);
7071 sg_mask = sched_group_mask(sg);
7072 /* Try to find first idle cpu */
7073 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7074 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7081 if (balance_cpu == -1)
7082 balance_cpu = group_balance_cpu(sg);
7085 * First idle cpu or the first cpu(busiest) in this sched group
7086 * is eligible for doing load balancing at this and above domains.
7088 return balance_cpu == env->dst_cpu;
7092 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7093 * tasks if there is an imbalance.
7095 static int load_balance(int this_cpu, struct rq *this_rq,
7096 struct sched_domain *sd, enum cpu_idle_type idle,
7097 int *continue_balancing)
7099 int ld_moved, cur_ld_moved, active_balance = 0;
7100 struct sched_domain *sd_parent = sd->parent;
7101 struct sched_group *group;
7103 unsigned long flags;
7104 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7106 struct lb_env env = {
7108 .dst_cpu = this_cpu,
7110 .dst_grpmask = sched_group_cpus(sd->groups),
7112 .loop_break = sched_nr_migrate_break,
7115 .tasks = LIST_HEAD_INIT(env.tasks),
7119 * For NEWLY_IDLE load_balancing, we don't need to consider
7120 * other cpus in our group
7122 if (idle == CPU_NEWLY_IDLE)
7123 env.dst_grpmask = NULL;
7125 cpumask_copy(cpus, cpu_active_mask);
7127 schedstat_inc(sd, lb_count[idle]);
7130 if (!should_we_balance(&env)) {
7131 *continue_balancing = 0;
7135 group = find_busiest_group(&env);
7137 schedstat_inc(sd, lb_nobusyg[idle]);
7141 busiest = find_busiest_queue(&env, group);
7143 schedstat_inc(sd, lb_nobusyq[idle]);
7147 BUG_ON(busiest == env.dst_rq);
7149 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7151 env.src_cpu = busiest->cpu;
7152 env.src_rq = busiest;
7155 if (busiest->nr_running > 1) {
7157 * Attempt to move tasks. If find_busiest_group has found
7158 * an imbalance but busiest->nr_running <= 1, the group is
7159 * still unbalanced. ld_moved simply stays zero, so it is
7160 * correctly treated as an imbalance.
7162 env.flags |= LBF_ALL_PINNED;
7163 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7166 raw_spin_lock_irqsave(&busiest->lock, flags);
7169 * cur_ld_moved - load moved in current iteration
7170 * ld_moved - cumulative load moved across iterations
7172 cur_ld_moved = detach_tasks(&env);
7175 * We've detached some tasks from busiest_rq. Every
7176 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7177 * unlock busiest->lock, and we are able to be sure
7178 * that nobody can manipulate the tasks in parallel.
7179 * See task_rq_lock() family for the details.
7182 raw_spin_unlock(&busiest->lock);
7186 ld_moved += cur_ld_moved;
7189 local_irq_restore(flags);
7191 if (env.flags & LBF_NEED_BREAK) {
7192 env.flags &= ~LBF_NEED_BREAK;
7197 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7198 * us and move them to an alternate dst_cpu in our sched_group
7199 * where they can run. The upper limit on how many times we
7200 * iterate on same src_cpu is dependent on number of cpus in our
7203 * This changes load balance semantics a bit on who can move
7204 * load to a given_cpu. In addition to the given_cpu itself
7205 * (or a ilb_cpu acting on its behalf where given_cpu is
7206 * nohz-idle), we now have balance_cpu in a position to move
7207 * load to given_cpu. In rare situations, this may cause
7208 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7209 * _independently_ and at _same_ time to move some load to
7210 * given_cpu) causing exceess load to be moved to given_cpu.
7211 * This however should not happen so much in practice and
7212 * moreover subsequent load balance cycles should correct the
7213 * excess load moved.
7215 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7217 /* Prevent to re-select dst_cpu via env's cpus */
7218 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7220 env.dst_rq = cpu_rq(env.new_dst_cpu);
7221 env.dst_cpu = env.new_dst_cpu;
7222 env.flags &= ~LBF_DST_PINNED;
7224 env.loop_break = sched_nr_migrate_break;
7227 * Go back to "more_balance" rather than "redo" since we
7228 * need to continue with same src_cpu.
7234 * We failed to reach balance because of affinity.
7237 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7239 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7240 *group_imbalance = 1;
7243 /* All tasks on this runqueue were pinned by CPU affinity */
7244 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7245 cpumask_clear_cpu(cpu_of(busiest), cpus);
7246 if (!cpumask_empty(cpus)) {
7248 env.loop_break = sched_nr_migrate_break;
7251 goto out_all_pinned;
7256 schedstat_inc(sd, lb_failed[idle]);
7258 * Increment the failure counter only on periodic balance.
7259 * We do not want newidle balance, which can be very
7260 * frequent, pollute the failure counter causing
7261 * excessive cache_hot migrations and active balances.
7263 if (idle != CPU_NEWLY_IDLE)
7264 sd->nr_balance_failed++;
7266 if (need_active_balance(&env)) {
7267 raw_spin_lock_irqsave(&busiest->lock, flags);
7269 /* don't kick the active_load_balance_cpu_stop,
7270 * if the curr task on busiest cpu can't be
7273 if (!cpumask_test_cpu(this_cpu,
7274 tsk_cpus_allowed(busiest->curr))) {
7275 raw_spin_unlock_irqrestore(&busiest->lock,
7277 env.flags |= LBF_ALL_PINNED;
7278 goto out_one_pinned;
7282 * ->active_balance synchronizes accesses to
7283 * ->active_balance_work. Once set, it's cleared
7284 * only after active load balance is finished.
7286 if (!busiest->active_balance) {
7287 busiest->active_balance = 1;
7288 busiest->push_cpu = this_cpu;
7291 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7293 if (active_balance) {
7294 stop_one_cpu_nowait(cpu_of(busiest),
7295 active_load_balance_cpu_stop, busiest,
7296 &busiest->active_balance_work);
7300 * We've kicked active balancing, reset the failure
7303 sd->nr_balance_failed = sd->cache_nice_tries+1;
7306 sd->nr_balance_failed = 0;
7308 if (likely(!active_balance)) {
7309 /* We were unbalanced, so reset the balancing interval */
7310 sd->balance_interval = sd->min_interval;
7313 * If we've begun active balancing, start to back off. This
7314 * case may not be covered by the all_pinned logic if there
7315 * is only 1 task on the busy runqueue (because we don't call
7318 if (sd->balance_interval < sd->max_interval)
7319 sd->balance_interval *= 2;
7326 * We reach balance although we may have faced some affinity
7327 * constraints. Clear the imbalance flag only if other tasks got
7328 * a chance to move and fix the imbalance.
7330 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
7331 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7333 if (*group_imbalance)
7334 *group_imbalance = 0;
7339 * We reach balance because all tasks are pinned at this level so
7340 * we can't migrate them. Let the imbalance flag set so parent level
7341 * can try to migrate them.
7343 schedstat_inc(sd, lb_balanced[idle]);
7345 sd->nr_balance_failed = 0;
7351 * idle_balance() disregards balance intervals, so we could repeatedly
7352 * reach this code, which would lead to balance_interval skyrocketting
7353 * in a short amount of time. Skip the balance_interval increase logic
7356 if (env.idle == CPU_NEWLY_IDLE)
7359 /* tune up the balancing interval */
7360 if (((env.flags & LBF_ALL_PINNED) &&
7361 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7362 (sd->balance_interval < sd->max_interval))
7363 sd->balance_interval *= 2;
7368 static inline unsigned long
7369 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7371 unsigned long interval = sd->balance_interval;
7374 interval *= sd->busy_factor;
7376 /* scale ms to jiffies */
7377 interval = msecs_to_jiffies(interval);
7378 interval = clamp(interval, 1UL, max_load_balance_interval);
7384 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7386 unsigned long interval, next;
7388 interval = get_sd_balance_interval(sd, cpu_busy);
7389 next = sd->last_balance + interval;
7391 if (time_after(*next_balance, next))
7392 *next_balance = next;
7396 * idle_balance is called by schedule() if this_cpu is about to become
7397 * idle. Attempts to pull tasks from other CPUs.
7399 static int idle_balance(struct rq *this_rq)
7401 unsigned long next_balance = jiffies + HZ;
7402 int this_cpu = this_rq->cpu;
7403 struct sched_domain *sd;
7404 int pulled_task = 0;
7407 idle_enter_fair(this_rq);
7410 * We must set idle_stamp _before_ calling idle_balance(), such that we
7411 * measure the duration of idle_balance() as idle time.
7413 this_rq->idle_stamp = rq_clock(this_rq);
7415 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7416 !this_rq->rd->overload) {
7418 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7420 update_next_balance(sd, 0, &next_balance);
7426 raw_spin_unlock(&this_rq->lock);
7428 update_blocked_averages(this_cpu);
7430 for_each_domain(this_cpu, sd) {
7431 int continue_balancing = 1;
7432 u64 t0, domain_cost;
7434 if (!(sd->flags & SD_LOAD_BALANCE))
7437 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7438 update_next_balance(sd, 0, &next_balance);
7442 if (sd->flags & SD_BALANCE_NEWIDLE) {
7443 t0 = sched_clock_cpu(this_cpu);
7445 pulled_task = load_balance(this_cpu, this_rq,
7447 &continue_balancing);
7449 domain_cost = sched_clock_cpu(this_cpu) - t0;
7450 if (domain_cost > sd->max_newidle_lb_cost)
7451 sd->max_newidle_lb_cost = domain_cost;
7453 curr_cost += domain_cost;
7456 update_next_balance(sd, 0, &next_balance);
7459 * Stop searching for tasks to pull if there are
7460 * now runnable tasks on this rq.
7462 if (pulled_task || this_rq->nr_running > 0)
7467 raw_spin_lock(&this_rq->lock);
7469 if (curr_cost > this_rq->max_idle_balance_cost)
7470 this_rq->max_idle_balance_cost = curr_cost;
7473 * While browsing the domains, we released the rq lock, a task could
7474 * have been enqueued in the meantime. Since we're not going idle,
7475 * pretend we pulled a task.
7477 if (this_rq->cfs.h_nr_running && !pulled_task)
7481 /* Move the next balance forward */
7482 if (time_after(this_rq->next_balance, next_balance))
7483 this_rq->next_balance = next_balance;
7485 /* Is there a task of a high priority class? */
7486 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7490 idle_exit_fair(this_rq);
7491 this_rq->idle_stamp = 0;
7498 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7499 * running tasks off the busiest CPU onto idle CPUs. It requires at
7500 * least 1 task to be running on each physical CPU where possible, and
7501 * avoids physical / logical imbalances.
7503 static int active_load_balance_cpu_stop(void *data)
7505 struct rq *busiest_rq = data;
7506 int busiest_cpu = cpu_of(busiest_rq);
7507 int target_cpu = busiest_rq->push_cpu;
7508 struct rq *target_rq = cpu_rq(target_cpu);
7509 struct sched_domain *sd;
7510 struct task_struct *p = NULL;
7512 raw_spin_lock_irq(&busiest_rq->lock);
7514 /* make sure the requested cpu hasn't gone down in the meantime */
7515 if (unlikely(busiest_cpu != smp_processor_id() ||
7516 !busiest_rq->active_balance))
7519 /* Is there any task to move? */
7520 if (busiest_rq->nr_running <= 1)
7524 * This condition is "impossible", if it occurs
7525 * we need to fix it. Originally reported by
7526 * Bjorn Helgaas on a 128-cpu setup.
7528 BUG_ON(busiest_rq == target_rq);
7530 /* Search for an sd spanning us and the target CPU. */
7532 for_each_domain(target_cpu, sd) {
7533 if ((sd->flags & SD_LOAD_BALANCE) &&
7534 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7539 struct lb_env env = {
7541 .dst_cpu = target_cpu,
7542 .dst_rq = target_rq,
7543 .src_cpu = busiest_rq->cpu,
7544 .src_rq = busiest_rq,
7548 schedstat_inc(sd, alb_count);
7550 p = detach_one_task(&env);
7552 schedstat_inc(sd, alb_pushed);
7554 schedstat_inc(sd, alb_failed);
7558 busiest_rq->active_balance = 0;
7559 raw_spin_unlock(&busiest_rq->lock);
7562 attach_one_task(target_rq, p);
7569 static inline int on_null_domain(struct rq *rq)
7571 return unlikely(!rcu_dereference_sched(rq->sd));
7574 #ifdef CONFIG_NO_HZ_COMMON
7576 * idle load balancing details
7577 * - When one of the busy CPUs notice that there may be an idle rebalancing
7578 * needed, they will kick the idle load balancer, which then does idle
7579 * load balancing for all the idle CPUs.
7582 cpumask_var_t idle_cpus_mask;
7584 unsigned long next_balance; /* in jiffy units */
7585 } nohz ____cacheline_aligned;
7587 static inline int find_new_ilb(void)
7589 int ilb = cpumask_first(nohz.idle_cpus_mask);
7591 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7598 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7599 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7600 * CPU (if there is one).
7602 static void nohz_balancer_kick(void)
7606 nohz.next_balance++;
7608 ilb_cpu = find_new_ilb();
7610 if (ilb_cpu >= nr_cpu_ids)
7613 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7616 * Use smp_send_reschedule() instead of resched_cpu().
7617 * This way we generate a sched IPI on the target cpu which
7618 * is idle. And the softirq performing nohz idle load balance
7619 * will be run before returning from the IPI.
7621 smp_send_reschedule(ilb_cpu);
7625 static inline void nohz_balance_exit_idle(int cpu)
7627 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7629 * Completely isolated CPUs don't ever set, so we must test.
7631 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7632 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7633 atomic_dec(&nohz.nr_cpus);
7635 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7639 static inline void set_cpu_sd_state_busy(void)
7641 struct sched_domain *sd;
7642 int cpu = smp_processor_id();
7645 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7647 if (!sd || !sd->nohz_idle)
7651 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7656 void set_cpu_sd_state_idle(void)
7658 struct sched_domain *sd;
7659 int cpu = smp_processor_id();
7662 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7664 if (!sd || sd->nohz_idle)
7668 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7674 * This routine will record that the cpu is going idle with tick stopped.
7675 * This info will be used in performing idle load balancing in the future.
7677 void nohz_balance_enter_idle(int cpu)
7680 * If this cpu is going down, then nothing needs to be done.
7682 if (!cpu_active(cpu))
7685 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7689 * If we're a completely isolated CPU, we don't play.
7691 if (on_null_domain(cpu_rq(cpu)))
7694 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7695 atomic_inc(&nohz.nr_cpus);
7696 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7699 static int sched_ilb_notifier(struct notifier_block *nfb,
7700 unsigned long action, void *hcpu)
7702 switch (action & ~CPU_TASKS_FROZEN) {
7704 nohz_balance_exit_idle(smp_processor_id());
7712 static DEFINE_SPINLOCK(balancing);
7715 * Scale the max load_balance interval with the number of CPUs in the system.
7716 * This trades load-balance latency on larger machines for less cross talk.
7718 void update_max_interval(void)
7720 max_load_balance_interval = HZ*num_online_cpus()/10;
7724 * It checks each scheduling domain to see if it is due to be balanced,
7725 * and initiates a balancing operation if so.
7727 * Balancing parameters are set up in init_sched_domains.
7729 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7731 int continue_balancing = 1;
7733 unsigned long interval;
7734 struct sched_domain *sd;
7735 /* Earliest time when we have to do rebalance again */
7736 unsigned long next_balance = jiffies + 60*HZ;
7737 int update_next_balance = 0;
7738 int need_serialize, need_decay = 0;
7741 update_blocked_averages(cpu);
7744 for_each_domain(cpu, sd) {
7746 * Decay the newidle max times here because this is a regular
7747 * visit to all the domains. Decay ~1% per second.
7749 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7750 sd->max_newidle_lb_cost =
7751 (sd->max_newidle_lb_cost * 253) / 256;
7752 sd->next_decay_max_lb_cost = jiffies + HZ;
7755 max_cost += sd->max_newidle_lb_cost;
7757 if (!(sd->flags & SD_LOAD_BALANCE))
7761 * Stop the load balance at this level. There is another
7762 * CPU in our sched group which is doing load balancing more
7765 if (!continue_balancing) {
7771 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7773 need_serialize = sd->flags & SD_SERIALIZE;
7774 if (need_serialize) {
7775 if (!spin_trylock(&balancing))
7779 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7780 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7782 * The LBF_DST_PINNED logic could have changed
7783 * env->dst_cpu, so we can't know our idle
7784 * state even if we migrated tasks. Update it.
7786 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7788 sd->last_balance = jiffies;
7789 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7792 spin_unlock(&balancing);
7794 if (time_after(next_balance, sd->last_balance + interval)) {
7795 next_balance = sd->last_balance + interval;
7796 update_next_balance = 1;
7801 * Ensure the rq-wide value also decays but keep it at a
7802 * reasonable floor to avoid funnies with rq->avg_idle.
7804 rq->max_idle_balance_cost =
7805 max((u64)sysctl_sched_migration_cost, max_cost);
7810 * next_balance will be updated only when there is a need.
7811 * When the cpu is attached to null domain for ex, it will not be
7814 if (likely(update_next_balance)) {
7815 rq->next_balance = next_balance;
7817 #ifdef CONFIG_NO_HZ_COMMON
7819 * If this CPU has been elected to perform the nohz idle
7820 * balance. Other idle CPUs have already rebalanced with
7821 * nohz_idle_balance() and nohz.next_balance has been
7822 * updated accordingly. This CPU is now running the idle load
7823 * balance for itself and we need to update the
7824 * nohz.next_balance accordingly.
7826 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
7827 nohz.next_balance = rq->next_balance;
7832 #ifdef CONFIG_NO_HZ_COMMON
7834 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7835 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7837 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7839 int this_cpu = this_rq->cpu;
7842 /* Earliest time when we have to do rebalance again */
7843 unsigned long next_balance = jiffies + 60*HZ;
7844 int update_next_balance = 0;
7846 if (idle != CPU_IDLE ||
7847 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7850 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7851 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7855 * If this cpu gets work to do, stop the load balancing
7856 * work being done for other cpus. Next load
7857 * balancing owner will pick it up.
7862 rq = cpu_rq(balance_cpu);
7865 * If time for next balance is due,
7868 if (time_after_eq(jiffies, rq->next_balance)) {
7869 raw_spin_lock_irq(&rq->lock);
7870 update_rq_clock(rq);
7871 update_idle_cpu_load(rq);
7872 raw_spin_unlock_irq(&rq->lock);
7873 rebalance_domains(rq, CPU_IDLE);
7876 if (time_after(next_balance, rq->next_balance)) {
7877 next_balance = rq->next_balance;
7878 update_next_balance = 1;
7883 * next_balance will be updated only when there is a need.
7884 * When the CPU is attached to null domain for ex, it will not be
7887 if (likely(update_next_balance))
7888 nohz.next_balance = next_balance;
7890 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7894 * Current heuristic for kicking the idle load balancer in the presence
7895 * of an idle cpu in the system.
7896 * - This rq has more than one task.
7897 * - This rq has at least one CFS task and the capacity of the CPU is
7898 * significantly reduced because of RT tasks or IRQs.
7899 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7900 * multiple busy cpu.
7901 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7902 * domain span are idle.
7904 static inline bool nohz_kick_needed(struct rq *rq)
7906 unsigned long now = jiffies;
7907 struct sched_domain *sd;
7908 struct sched_group_capacity *sgc;
7909 int nr_busy, cpu = rq->cpu;
7912 if (unlikely(rq->idle_balance))
7916 * We may be recently in ticked or tickless idle mode. At the first
7917 * busy tick after returning from idle, we will update the busy stats.
7919 set_cpu_sd_state_busy();
7920 nohz_balance_exit_idle(cpu);
7923 * None are in tickless mode and hence no need for NOHZ idle load
7926 if (likely(!atomic_read(&nohz.nr_cpus)))
7929 if (time_before(now, nohz.next_balance))
7932 if (rq->nr_running >= 2)
7936 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7938 sgc = sd->groups->sgc;
7939 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7948 sd = rcu_dereference(rq->sd);
7950 if ((rq->cfs.h_nr_running >= 1) &&
7951 check_cpu_capacity(rq, sd)) {
7957 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7958 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7959 sched_domain_span(sd)) < cpu)) {
7969 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7973 * run_rebalance_domains is triggered when needed from the scheduler tick.
7974 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7976 static void run_rebalance_domains(struct softirq_action *h)
7978 struct rq *this_rq = this_rq();
7979 enum cpu_idle_type idle = this_rq->idle_balance ?
7980 CPU_IDLE : CPU_NOT_IDLE;
7983 * If this cpu has a pending nohz_balance_kick, then do the
7984 * balancing on behalf of the other idle cpus whose ticks are
7985 * stopped. Do nohz_idle_balance *before* rebalance_domains to
7986 * give the idle cpus a chance to load balance. Else we may
7987 * load balance only within the local sched_domain hierarchy
7988 * and abort nohz_idle_balance altogether if we pull some load.
7990 nohz_idle_balance(this_rq, idle);
7991 rebalance_domains(this_rq, idle);
7995 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7997 void trigger_load_balance(struct rq *rq)
7999 /* Don't need to rebalance while attached to NULL domain */
8000 if (unlikely(on_null_domain(rq)))
8003 if (time_after_eq(jiffies, rq->next_balance))
8004 raise_softirq(SCHED_SOFTIRQ);
8005 #ifdef CONFIG_NO_HZ_COMMON
8006 if (nohz_kick_needed(rq))
8007 nohz_balancer_kick();
8011 static void rq_online_fair(struct rq *rq)
8015 update_runtime_enabled(rq);
8018 static void rq_offline_fair(struct rq *rq)
8022 /* Ensure any throttled groups are reachable by pick_next_task */
8023 unthrottle_offline_cfs_rqs(rq);
8026 #endif /* CONFIG_SMP */
8029 * scheduler tick hitting a task of our scheduling class:
8031 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8033 struct cfs_rq *cfs_rq;
8034 struct sched_entity *se = &curr->se;
8036 for_each_sched_entity(se) {
8037 cfs_rq = cfs_rq_of(se);
8038 entity_tick(cfs_rq, se, queued);
8041 if (static_branch_unlikely(&sched_numa_balancing))
8042 task_tick_numa(rq, curr);
8046 * called on fork with the child task as argument from the parent's context
8047 * - child not yet on the tasklist
8048 * - preemption disabled
8050 static void task_fork_fair(struct task_struct *p)
8052 struct cfs_rq *cfs_rq;
8053 struct sched_entity *se = &p->se, *curr;
8054 int this_cpu = smp_processor_id();
8055 struct rq *rq = this_rq();
8056 unsigned long flags;
8058 raw_spin_lock_irqsave(&rq->lock, flags);
8060 update_rq_clock(rq);
8062 cfs_rq = task_cfs_rq(current);
8063 curr = cfs_rq->curr;
8066 * Not only the cpu but also the task_group of the parent might have
8067 * been changed after parent->se.parent,cfs_rq were copied to
8068 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8069 * of child point to valid ones.
8072 __set_task_cpu(p, this_cpu);
8075 update_curr(cfs_rq);
8078 se->vruntime = curr->vruntime;
8079 place_entity(cfs_rq, se, 1);
8081 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8083 * Upon rescheduling, sched_class::put_prev_task() will place
8084 * 'current' within the tree based on its new key value.
8086 swap(curr->vruntime, se->vruntime);
8090 se->vruntime -= cfs_rq->min_vruntime;
8092 raw_spin_unlock_irqrestore(&rq->lock, flags);
8096 * Priority of the task has changed. Check to see if we preempt
8100 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8102 if (!task_on_rq_queued(p))
8106 * Reschedule if we are currently running on this runqueue and
8107 * our priority decreased, or if we are not currently running on
8108 * this runqueue and our priority is higher than the current's
8110 if (rq->curr == p) {
8111 if (p->prio > oldprio)
8114 check_preempt_curr(rq, p, 0);
8117 static inline bool vruntime_normalized(struct task_struct *p)
8119 struct sched_entity *se = &p->se;
8122 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8123 * the dequeue_entity(.flags=0) will already have normalized the
8130 * When !on_rq, vruntime of the task has usually NOT been normalized.
8131 * But there are some cases where it has already been normalized:
8133 * - A forked child which is waiting for being woken up by
8134 * wake_up_new_task().
8135 * - A task which has been woken up by try_to_wake_up() and
8136 * waiting for actually being woken up by sched_ttwu_pending().
8138 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8144 static void detach_task_cfs_rq(struct task_struct *p)
8146 struct sched_entity *se = &p->se;
8147 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8149 if (!vruntime_normalized(p)) {
8151 * Fix up our vruntime so that the current sleep doesn't
8152 * cause 'unlimited' sleep bonus.
8154 place_entity(cfs_rq, se, 0);
8155 se->vruntime -= cfs_rq->min_vruntime;
8158 /* Catch up with the cfs_rq and remove our load when we leave */
8159 detach_entity_load_avg(cfs_rq, se);
8162 static void attach_task_cfs_rq(struct task_struct *p)
8164 struct sched_entity *se = &p->se;
8165 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8167 #ifdef CONFIG_FAIR_GROUP_SCHED
8169 * Since the real-depth could have been changed (only FAIR
8170 * class maintain depth value), reset depth properly.
8172 se->depth = se->parent ? se->parent->depth + 1 : 0;
8175 /* Synchronize task with its cfs_rq */
8176 attach_entity_load_avg(cfs_rq, se);
8178 if (!vruntime_normalized(p))
8179 se->vruntime += cfs_rq->min_vruntime;
8182 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8184 detach_task_cfs_rq(p);
8187 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8189 attach_task_cfs_rq(p);
8191 if (task_on_rq_queued(p)) {
8193 * We were most likely switched from sched_rt, so
8194 * kick off the schedule if running, otherwise just see
8195 * if we can still preempt the current task.
8200 check_preempt_curr(rq, p, 0);
8204 /* Account for a task changing its policy or group.
8206 * This routine is mostly called to set cfs_rq->curr field when a task
8207 * migrates between groups/classes.
8209 static void set_curr_task_fair(struct rq *rq)
8211 struct sched_entity *se = &rq->curr->se;
8213 for_each_sched_entity(se) {
8214 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8216 set_next_entity(cfs_rq, se);
8217 /* ensure bandwidth has been allocated on our new cfs_rq */
8218 account_cfs_rq_runtime(cfs_rq, 0);
8222 void init_cfs_rq(struct cfs_rq *cfs_rq)
8224 cfs_rq->tasks_timeline = RB_ROOT;
8225 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8226 #ifndef CONFIG_64BIT
8227 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8230 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8231 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8235 #ifdef CONFIG_FAIR_GROUP_SCHED
8236 static void task_move_group_fair(struct task_struct *p)
8238 detach_task_cfs_rq(p);
8239 set_task_rq(p, task_cpu(p));
8242 /* Tell se's cfs_rq has been changed -- migrated */
8243 p->se.avg.last_update_time = 0;
8245 attach_task_cfs_rq(p);
8248 void free_fair_sched_group(struct task_group *tg)
8252 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8254 for_each_possible_cpu(i) {
8256 kfree(tg->cfs_rq[i]);
8265 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8267 struct cfs_rq *cfs_rq;
8268 struct sched_entity *se;
8271 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8274 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8278 tg->shares = NICE_0_LOAD;
8280 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8282 for_each_possible_cpu(i) {
8283 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8284 GFP_KERNEL, cpu_to_node(i));
8288 se = kzalloc_node(sizeof(struct sched_entity),
8289 GFP_KERNEL, cpu_to_node(i));
8293 init_cfs_rq(cfs_rq);
8294 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8295 init_entity_runnable_average(se);
8306 void unregister_fair_sched_group(struct task_group *tg)
8308 unsigned long flags;
8312 for_each_possible_cpu(cpu) {
8314 remove_entity_load_avg(tg->se[cpu]);
8317 * Only empty task groups can be destroyed; so we can speculatively
8318 * check on_list without danger of it being re-added.
8320 if (!tg->cfs_rq[cpu]->on_list)
8325 raw_spin_lock_irqsave(&rq->lock, flags);
8326 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8327 raw_spin_unlock_irqrestore(&rq->lock, flags);
8331 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8332 struct sched_entity *se, int cpu,
8333 struct sched_entity *parent)
8335 struct rq *rq = cpu_rq(cpu);
8339 init_cfs_rq_runtime(cfs_rq);
8341 tg->cfs_rq[cpu] = cfs_rq;
8344 /* se could be NULL for root_task_group */
8349 se->cfs_rq = &rq->cfs;
8352 se->cfs_rq = parent->my_q;
8353 se->depth = parent->depth + 1;
8357 /* guarantee group entities always have weight */
8358 update_load_set(&se->load, NICE_0_LOAD);
8359 se->parent = parent;
8362 static DEFINE_MUTEX(shares_mutex);
8364 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8367 unsigned long flags;
8370 * We can't change the weight of the root cgroup.
8375 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8377 mutex_lock(&shares_mutex);
8378 if (tg->shares == shares)
8381 tg->shares = shares;
8382 for_each_possible_cpu(i) {
8383 struct rq *rq = cpu_rq(i);
8384 struct sched_entity *se;
8387 /* Propagate contribution to hierarchy */
8388 raw_spin_lock_irqsave(&rq->lock, flags);
8390 /* Possible calls to update_curr() need rq clock */
8391 update_rq_clock(rq);
8392 for_each_sched_entity(se)
8393 update_cfs_shares(group_cfs_rq(se));
8394 raw_spin_unlock_irqrestore(&rq->lock, flags);
8398 mutex_unlock(&shares_mutex);
8401 #else /* CONFIG_FAIR_GROUP_SCHED */
8403 void free_fair_sched_group(struct task_group *tg) { }
8405 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8410 void unregister_fair_sched_group(struct task_group *tg) { }
8412 #endif /* CONFIG_FAIR_GROUP_SCHED */
8415 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8417 struct sched_entity *se = &task->se;
8418 unsigned int rr_interval = 0;
8421 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8424 if (rq->cfs.load.weight)
8425 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8431 * All the scheduling class methods:
8433 const struct sched_class fair_sched_class = {
8434 .next = &idle_sched_class,
8435 .enqueue_task = enqueue_task_fair,
8436 .dequeue_task = dequeue_task_fair,
8437 .yield_task = yield_task_fair,
8438 .yield_to_task = yield_to_task_fair,
8440 .check_preempt_curr = check_preempt_wakeup,
8442 .pick_next_task = pick_next_task_fair,
8443 .put_prev_task = put_prev_task_fair,
8446 .select_task_rq = select_task_rq_fair,
8447 .migrate_task_rq = migrate_task_rq_fair,
8449 .rq_online = rq_online_fair,
8450 .rq_offline = rq_offline_fair,
8452 .task_waking = task_waking_fair,
8453 .task_dead = task_dead_fair,
8454 .set_cpus_allowed = set_cpus_allowed_common,
8457 .set_curr_task = set_curr_task_fair,
8458 .task_tick = task_tick_fair,
8459 .task_fork = task_fork_fair,
8461 .prio_changed = prio_changed_fair,
8462 .switched_from = switched_from_fair,
8463 .switched_to = switched_to_fair,
8465 .get_rr_interval = get_rr_interval_fair,
8467 .update_curr = update_curr_fair,
8469 #ifdef CONFIG_FAIR_GROUP_SCHED
8470 .task_move_group = task_move_group_fair,
8474 #ifdef CONFIG_SCHED_DEBUG
8475 void print_cfs_stats(struct seq_file *m, int cpu)
8477 struct cfs_rq *cfs_rq;
8480 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8481 print_cfs_rq(m, cpu, cfs_rq);
8485 #ifdef CONFIG_NUMA_BALANCING
8486 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8489 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8491 for_each_online_node(node) {
8492 if (p->numa_faults) {
8493 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8494 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8496 if (p->numa_group) {
8497 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8498 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8500 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8503 #endif /* CONFIG_NUMA_BALANCING */
8504 #endif /* CONFIG_SCHED_DEBUG */
8506 __init void init_sched_fair_class(void)
8509 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8511 #ifdef CONFIG_NO_HZ_COMMON
8512 nohz.next_balance = jiffies;
8513 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8514 cpu_notifier(sched_ilb_notifier, 0);