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);
1198 env->best_imp = imp;
1199 env->best_cpu = env->dst_cpu;
1202 static bool load_too_imbalanced(long src_load, long dst_load,
1203 struct task_numa_env *env)
1206 long orig_src_load, orig_dst_load;
1207 long src_capacity, dst_capacity;
1210 * The load is corrected for the CPU capacity available on each node.
1213 * ------------ vs ---------
1214 * src_capacity dst_capacity
1216 src_capacity = env->src_stats.compute_capacity;
1217 dst_capacity = env->dst_stats.compute_capacity;
1219 /* We care about the slope of the imbalance, not the direction. */
1220 if (dst_load < src_load)
1221 swap(dst_load, src_load);
1223 /* Is the difference below the threshold? */
1224 imb = dst_load * src_capacity * 100 -
1225 src_load * dst_capacity * env->imbalance_pct;
1230 * The imbalance is above the allowed threshold.
1231 * Compare it with the old imbalance.
1233 orig_src_load = env->src_stats.load;
1234 orig_dst_load = env->dst_stats.load;
1236 if (orig_dst_load < orig_src_load)
1237 swap(orig_dst_load, orig_src_load);
1239 old_imb = orig_dst_load * src_capacity * 100 -
1240 orig_src_load * dst_capacity * env->imbalance_pct;
1242 /* Would this change make things worse? */
1243 return (imb > old_imb);
1247 * This checks if the overall compute and NUMA accesses of the system would
1248 * be improved if the source tasks was migrated to the target dst_cpu taking
1249 * into account that it might be best if task running on the dst_cpu should
1250 * be exchanged with the source task
1252 static void task_numa_compare(struct task_numa_env *env,
1253 long taskimp, long groupimp)
1255 struct rq *src_rq = cpu_rq(env->src_cpu);
1256 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1257 struct task_struct *cur;
1258 long src_load, dst_load;
1260 long imp = env->p->numa_group ? groupimp : taskimp;
1262 int dist = env->dist;
1266 raw_spin_lock_irq(&dst_rq->lock);
1269 * No need to move the exiting task, and this ensures that ->curr
1270 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1271 * is safe under RCU read lock.
1272 * Note that rcu_read_lock() itself can't protect from the final
1273 * put_task_struct() after the last schedule().
1275 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1277 raw_spin_unlock_irq(&dst_rq->lock);
1280 * Because we have preemption enabled we can get migrated around and
1281 * end try selecting ourselves (current == env->p) as a swap candidate.
1287 * "imp" is the fault differential for the source task between the
1288 * source and destination node. Calculate the total differential for
1289 * the source task and potential destination task. The more negative
1290 * the value is, the more rmeote accesses that would be expected to
1291 * be incurred if the tasks were swapped.
1294 /* Skip this swap candidate if cannot move to the source cpu */
1295 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1299 * If dst and source tasks are in the same NUMA group, or not
1300 * in any group then look only at task weights.
1302 if (cur->numa_group == env->p->numa_group) {
1303 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1304 task_weight(cur, env->dst_nid, dist);
1306 * Add some hysteresis to prevent swapping the
1307 * tasks within a group over tiny differences.
1309 if (cur->numa_group)
1313 * Compare the group weights. If a task is all by
1314 * itself (not part of a group), use the task weight
1317 if (cur->numa_group)
1318 imp += group_weight(cur, env->src_nid, dist) -
1319 group_weight(cur, env->dst_nid, dist);
1321 imp += task_weight(cur, env->src_nid, dist) -
1322 task_weight(cur, env->dst_nid, dist);
1326 if (imp <= env->best_imp && moveimp <= env->best_imp)
1330 /* Is there capacity at our destination? */
1331 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1332 !env->dst_stats.has_free_capacity)
1338 /* Balance doesn't matter much if we're running a task per cpu */
1339 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1340 dst_rq->nr_running == 1)
1344 * In the overloaded case, try and keep the load balanced.
1347 load = task_h_load(env->p);
1348 dst_load = env->dst_stats.load + load;
1349 src_load = env->src_stats.load - load;
1351 if (moveimp > imp && moveimp > env->best_imp) {
1353 * If the improvement from just moving env->p direction is
1354 * better than swapping tasks around, check if a move is
1355 * possible. Store a slightly smaller score than moveimp,
1356 * so an actually idle CPU will win.
1358 if (!load_too_imbalanced(src_load, dst_load, env)) {
1365 if (imp <= env->best_imp)
1369 load = task_h_load(cur);
1374 if (load_too_imbalanced(src_load, dst_load, env))
1378 * One idle CPU per node is evaluated for a task numa move.
1379 * Call select_idle_sibling to maybe find a better one.
1382 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1385 task_numa_assign(env, cur, imp);
1390 static void task_numa_find_cpu(struct task_numa_env *env,
1391 long taskimp, long groupimp)
1395 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1396 /* Skip this CPU if the source task cannot migrate */
1397 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1401 task_numa_compare(env, taskimp, groupimp);
1405 /* Only move tasks to a NUMA node less busy than the current node. */
1406 static bool numa_has_capacity(struct task_numa_env *env)
1408 struct numa_stats *src = &env->src_stats;
1409 struct numa_stats *dst = &env->dst_stats;
1411 if (src->has_free_capacity && !dst->has_free_capacity)
1415 * Only consider a task move if the source has a higher load
1416 * than the destination, corrected for CPU capacity on each node.
1418 * src->load dst->load
1419 * --------------------- vs ---------------------
1420 * src->compute_capacity dst->compute_capacity
1422 if (src->load * dst->compute_capacity * env->imbalance_pct >
1424 dst->load * src->compute_capacity * 100)
1430 static int task_numa_migrate(struct task_struct *p)
1432 struct task_numa_env env = {
1435 .src_cpu = task_cpu(p),
1436 .src_nid = task_node(p),
1438 .imbalance_pct = 112,
1444 struct sched_domain *sd;
1445 unsigned long taskweight, groupweight;
1447 long taskimp, groupimp;
1450 * Pick the lowest SD_NUMA domain, as that would have the smallest
1451 * imbalance and would be the first to start moving tasks about.
1453 * And we want to avoid any moving of tasks about, as that would create
1454 * random movement of tasks -- counter the numa conditions we're trying
1458 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1460 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1464 * Cpusets can break the scheduler domain tree into smaller
1465 * balance domains, some of which do not cross NUMA boundaries.
1466 * Tasks that are "trapped" in such domains cannot be migrated
1467 * elsewhere, so there is no point in (re)trying.
1469 if (unlikely(!sd)) {
1470 p->numa_preferred_nid = task_node(p);
1474 env.dst_nid = p->numa_preferred_nid;
1475 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1476 taskweight = task_weight(p, env.src_nid, dist);
1477 groupweight = group_weight(p, env.src_nid, dist);
1478 update_numa_stats(&env.src_stats, env.src_nid);
1479 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1480 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1481 update_numa_stats(&env.dst_stats, env.dst_nid);
1483 /* Try to find a spot on the preferred nid. */
1484 if (numa_has_capacity(&env))
1485 task_numa_find_cpu(&env, taskimp, groupimp);
1488 * Look at other nodes in these cases:
1489 * - there is no space available on the preferred_nid
1490 * - the task is part of a numa_group that is interleaved across
1491 * multiple NUMA nodes; in order to better consolidate the group,
1492 * we need to check other locations.
1494 if (env.best_cpu == -1 || (p->numa_group &&
1495 nodes_weight(p->numa_group->active_nodes) > 1)) {
1496 for_each_online_node(nid) {
1497 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1500 dist = node_distance(env.src_nid, env.dst_nid);
1501 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1503 taskweight = task_weight(p, env.src_nid, dist);
1504 groupweight = group_weight(p, env.src_nid, dist);
1507 /* Only consider nodes where both task and groups benefit */
1508 taskimp = task_weight(p, nid, dist) - taskweight;
1509 groupimp = group_weight(p, nid, dist) - groupweight;
1510 if (taskimp < 0 && groupimp < 0)
1515 update_numa_stats(&env.dst_stats, env.dst_nid);
1516 if (numa_has_capacity(&env))
1517 task_numa_find_cpu(&env, taskimp, groupimp);
1522 * If the task is part of a workload that spans multiple NUMA nodes,
1523 * and is migrating into one of the workload's active nodes, remember
1524 * this node as the task's preferred numa node, so the workload can
1526 * A task that migrated to a second choice node will be better off
1527 * trying for a better one later. Do not set the preferred node here.
1529 if (p->numa_group) {
1530 if (env.best_cpu == -1)
1535 if (node_isset(nid, p->numa_group->active_nodes))
1536 sched_setnuma(p, env.dst_nid);
1539 /* No better CPU than the current one was found. */
1540 if (env.best_cpu == -1)
1544 * Reset the scan period if the task is being rescheduled on an
1545 * alternative node to recheck if the tasks is now properly placed.
1547 p->numa_scan_period = task_scan_min(p);
1549 if (env.best_task == NULL) {
1550 ret = migrate_task_to(p, env.best_cpu);
1552 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1556 ret = migrate_swap(p, env.best_task);
1558 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1559 put_task_struct(env.best_task);
1563 /* Attempt to migrate a task to a CPU on the preferred node. */
1564 static void numa_migrate_preferred(struct task_struct *p)
1566 unsigned long interval = HZ;
1568 /* This task has no NUMA fault statistics yet */
1569 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1572 /* Periodically retry migrating the task to the preferred node */
1573 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1574 p->numa_migrate_retry = jiffies + interval;
1576 /* Success if task is already running on preferred CPU */
1577 if (task_node(p) == p->numa_preferred_nid)
1580 /* Otherwise, try migrate to a CPU on the preferred node */
1581 task_numa_migrate(p);
1585 * Find the nodes on which the workload is actively running. We do this by
1586 * tracking the nodes from which NUMA hinting faults are triggered. This can
1587 * be different from the set of nodes where the workload's memory is currently
1590 * The bitmask is used to make smarter decisions on when to do NUMA page
1591 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1592 * are added when they cause over 6/16 of the maximum number of faults, but
1593 * only removed when they drop below 3/16.
1595 static void update_numa_active_node_mask(struct numa_group *numa_group)
1597 unsigned long faults, max_faults = 0;
1600 for_each_online_node(nid) {
1601 faults = group_faults_cpu(numa_group, nid);
1602 if (faults > max_faults)
1603 max_faults = faults;
1606 for_each_online_node(nid) {
1607 faults = group_faults_cpu(numa_group, nid);
1608 if (!node_isset(nid, numa_group->active_nodes)) {
1609 if (faults > max_faults * 6 / 16)
1610 node_set(nid, numa_group->active_nodes);
1611 } else if (faults < max_faults * 3 / 16)
1612 node_clear(nid, numa_group->active_nodes);
1617 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1618 * increments. The more local the fault statistics are, the higher the scan
1619 * period will be for the next scan window. If local/(local+remote) ratio is
1620 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1621 * the scan period will decrease. Aim for 70% local accesses.
1623 #define NUMA_PERIOD_SLOTS 10
1624 #define NUMA_PERIOD_THRESHOLD 7
1627 * Increase the scan period (slow down scanning) if the majority of
1628 * our memory is already on our local node, or if the majority of
1629 * the page accesses are shared with other processes.
1630 * Otherwise, decrease the scan period.
1632 static void update_task_scan_period(struct task_struct *p,
1633 unsigned long shared, unsigned long private)
1635 unsigned int period_slot;
1639 unsigned long remote = p->numa_faults_locality[0];
1640 unsigned long local = p->numa_faults_locality[1];
1643 * If there were no record hinting faults then either the task is
1644 * completely idle or all activity is areas that are not of interest
1645 * to automatic numa balancing. Related to that, if there were failed
1646 * migration then it implies we are migrating too quickly or the local
1647 * node is overloaded. In either case, scan slower
1649 if (local + shared == 0 || p->numa_faults_locality[2]) {
1650 p->numa_scan_period = min(p->numa_scan_period_max,
1651 p->numa_scan_period << 1);
1653 p->mm->numa_next_scan = jiffies +
1654 msecs_to_jiffies(p->numa_scan_period);
1660 * Prepare to scale scan period relative to the current period.
1661 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1662 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1663 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1665 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1666 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1667 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1668 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1671 diff = slot * period_slot;
1673 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1676 * Scale scan rate increases based on sharing. There is an
1677 * inverse relationship between the degree of sharing and
1678 * the adjustment made to the scanning period. Broadly
1679 * speaking the intent is that there is little point
1680 * scanning faster if shared accesses dominate as it may
1681 * simply bounce migrations uselessly
1683 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1684 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1687 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1688 task_scan_min(p), task_scan_max(p));
1689 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1693 * Get the fraction of time the task has been running since the last
1694 * NUMA placement cycle. The scheduler keeps similar statistics, but
1695 * decays those on a 32ms period, which is orders of magnitude off
1696 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1697 * stats only if the task is so new there are no NUMA statistics yet.
1699 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1701 u64 runtime, delta, now;
1702 /* Use the start of this time slice to avoid calculations. */
1703 now = p->se.exec_start;
1704 runtime = p->se.sum_exec_runtime;
1706 if (p->last_task_numa_placement) {
1707 delta = runtime - p->last_sum_exec_runtime;
1708 *period = now - p->last_task_numa_placement;
1710 delta = p->se.avg.load_sum / p->se.load.weight;
1711 *period = LOAD_AVG_MAX;
1714 p->last_sum_exec_runtime = runtime;
1715 p->last_task_numa_placement = now;
1721 * Determine the preferred nid for a task in a numa_group. This needs to
1722 * be done in a way that produces consistent results with group_weight,
1723 * otherwise workloads might not converge.
1725 static int preferred_group_nid(struct task_struct *p, int nid)
1730 /* Direct connections between all NUMA nodes. */
1731 if (sched_numa_topology_type == NUMA_DIRECT)
1735 * On a system with glueless mesh NUMA topology, group_weight
1736 * scores nodes according to the number of NUMA hinting faults on
1737 * both the node itself, and on nearby nodes.
1739 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1740 unsigned long score, max_score = 0;
1741 int node, max_node = nid;
1743 dist = sched_max_numa_distance;
1745 for_each_online_node(node) {
1746 score = group_weight(p, node, dist);
1747 if (score > max_score) {
1756 * Finding the preferred nid in a system with NUMA backplane
1757 * interconnect topology is more involved. The goal is to locate
1758 * tasks from numa_groups near each other in the system, and
1759 * untangle workloads from different sides of the system. This requires
1760 * searching down the hierarchy of node groups, recursively searching
1761 * inside the highest scoring group of nodes. The nodemask tricks
1762 * keep the complexity of the search down.
1764 nodes = node_online_map;
1765 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1766 unsigned long max_faults = 0;
1767 nodemask_t max_group = NODE_MASK_NONE;
1770 /* Are there nodes at this distance from each other? */
1771 if (!find_numa_distance(dist))
1774 for_each_node_mask(a, nodes) {
1775 unsigned long faults = 0;
1776 nodemask_t this_group;
1777 nodes_clear(this_group);
1779 /* Sum group's NUMA faults; includes a==b case. */
1780 for_each_node_mask(b, nodes) {
1781 if (node_distance(a, b) < dist) {
1782 faults += group_faults(p, b);
1783 node_set(b, this_group);
1784 node_clear(b, nodes);
1788 /* Remember the top group. */
1789 if (faults > max_faults) {
1790 max_faults = faults;
1791 max_group = this_group;
1793 * subtle: at the smallest distance there is
1794 * just one node left in each "group", the
1795 * winner is the preferred nid.
1800 /* Next round, evaluate the nodes within max_group. */
1808 static void task_numa_placement(struct task_struct *p)
1810 int seq, nid, max_nid = -1, max_group_nid = -1;
1811 unsigned long max_faults = 0, max_group_faults = 0;
1812 unsigned long fault_types[2] = { 0, 0 };
1813 unsigned long total_faults;
1814 u64 runtime, period;
1815 spinlock_t *group_lock = NULL;
1818 * The p->mm->numa_scan_seq field gets updated without
1819 * exclusive access. Use READ_ONCE() here to ensure
1820 * that the field is read in a single access:
1822 seq = READ_ONCE(p->mm->numa_scan_seq);
1823 if (p->numa_scan_seq == seq)
1825 p->numa_scan_seq = seq;
1826 p->numa_scan_period_max = task_scan_max(p);
1828 total_faults = p->numa_faults_locality[0] +
1829 p->numa_faults_locality[1];
1830 runtime = numa_get_avg_runtime(p, &period);
1832 /* If the task is part of a group prevent parallel updates to group stats */
1833 if (p->numa_group) {
1834 group_lock = &p->numa_group->lock;
1835 spin_lock_irq(group_lock);
1838 /* Find the node with the highest number of faults */
1839 for_each_online_node(nid) {
1840 /* Keep track of the offsets in numa_faults array */
1841 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1842 unsigned long faults = 0, group_faults = 0;
1845 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1846 long diff, f_diff, f_weight;
1848 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1849 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1850 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1851 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1853 /* Decay existing window, copy faults since last scan */
1854 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1855 fault_types[priv] += p->numa_faults[membuf_idx];
1856 p->numa_faults[membuf_idx] = 0;
1859 * Normalize the faults_from, so all tasks in a group
1860 * count according to CPU use, instead of by the raw
1861 * number of faults. Tasks with little runtime have
1862 * little over-all impact on throughput, and thus their
1863 * faults are less important.
1865 f_weight = div64_u64(runtime << 16, period + 1);
1866 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1868 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1869 p->numa_faults[cpubuf_idx] = 0;
1871 p->numa_faults[mem_idx] += diff;
1872 p->numa_faults[cpu_idx] += f_diff;
1873 faults += p->numa_faults[mem_idx];
1874 p->total_numa_faults += diff;
1875 if (p->numa_group) {
1877 * safe because we can only change our own group
1879 * mem_idx represents the offset for a given
1880 * nid and priv in a specific region because it
1881 * is at the beginning of the numa_faults array.
1883 p->numa_group->faults[mem_idx] += diff;
1884 p->numa_group->faults_cpu[mem_idx] += f_diff;
1885 p->numa_group->total_faults += diff;
1886 group_faults += p->numa_group->faults[mem_idx];
1890 if (faults > max_faults) {
1891 max_faults = faults;
1895 if (group_faults > max_group_faults) {
1896 max_group_faults = group_faults;
1897 max_group_nid = nid;
1901 update_task_scan_period(p, fault_types[0], fault_types[1]);
1903 if (p->numa_group) {
1904 update_numa_active_node_mask(p->numa_group);
1905 spin_unlock_irq(group_lock);
1906 max_nid = preferred_group_nid(p, max_group_nid);
1910 /* Set the new preferred node */
1911 if (max_nid != p->numa_preferred_nid)
1912 sched_setnuma(p, max_nid);
1914 if (task_node(p) != p->numa_preferred_nid)
1915 numa_migrate_preferred(p);
1919 static inline int get_numa_group(struct numa_group *grp)
1921 return atomic_inc_not_zero(&grp->refcount);
1924 static inline void put_numa_group(struct numa_group *grp)
1926 if (atomic_dec_and_test(&grp->refcount))
1927 kfree_rcu(grp, rcu);
1930 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1933 struct numa_group *grp, *my_grp;
1934 struct task_struct *tsk;
1936 int cpu = cpupid_to_cpu(cpupid);
1939 if (unlikely(!p->numa_group)) {
1940 unsigned int size = sizeof(struct numa_group) +
1941 4*nr_node_ids*sizeof(unsigned long);
1943 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1947 atomic_set(&grp->refcount, 1);
1948 spin_lock_init(&grp->lock);
1950 /* Second half of the array tracks nids where faults happen */
1951 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1954 node_set(task_node(current), grp->active_nodes);
1956 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1957 grp->faults[i] = p->numa_faults[i];
1959 grp->total_faults = p->total_numa_faults;
1962 rcu_assign_pointer(p->numa_group, grp);
1966 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1968 if (!cpupid_match_pid(tsk, cpupid))
1971 grp = rcu_dereference(tsk->numa_group);
1975 my_grp = p->numa_group;
1980 * Only join the other group if its bigger; if we're the bigger group,
1981 * the other task will join us.
1983 if (my_grp->nr_tasks > grp->nr_tasks)
1987 * Tie-break on the grp address.
1989 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1992 /* Always join threads in the same process. */
1993 if (tsk->mm == current->mm)
1996 /* Simple filter to avoid false positives due to PID collisions */
1997 if (flags & TNF_SHARED)
2000 /* Update priv based on whether false sharing was detected */
2003 if (join && !get_numa_group(grp))
2011 BUG_ON(irqs_disabled());
2012 double_lock_irq(&my_grp->lock, &grp->lock);
2014 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2015 my_grp->faults[i] -= p->numa_faults[i];
2016 grp->faults[i] += p->numa_faults[i];
2018 my_grp->total_faults -= p->total_numa_faults;
2019 grp->total_faults += p->total_numa_faults;
2024 spin_unlock(&my_grp->lock);
2025 spin_unlock_irq(&grp->lock);
2027 rcu_assign_pointer(p->numa_group, grp);
2029 put_numa_group(my_grp);
2037 void task_numa_free(struct task_struct *p)
2039 struct numa_group *grp = p->numa_group;
2040 void *numa_faults = p->numa_faults;
2041 unsigned long flags;
2045 spin_lock_irqsave(&grp->lock, flags);
2046 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2047 grp->faults[i] -= p->numa_faults[i];
2048 grp->total_faults -= p->total_numa_faults;
2051 spin_unlock_irqrestore(&grp->lock, flags);
2052 RCU_INIT_POINTER(p->numa_group, NULL);
2053 put_numa_group(grp);
2056 p->numa_faults = NULL;
2061 * Got a PROT_NONE fault for a page on @node.
2063 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2065 struct task_struct *p = current;
2066 bool migrated = flags & TNF_MIGRATED;
2067 int cpu_node = task_node(current);
2068 int local = !!(flags & TNF_FAULT_LOCAL);
2071 if (!static_branch_likely(&sched_numa_balancing))
2074 /* for example, ksmd faulting in a user's mm */
2078 /* Allocate buffer to track faults on a per-node basis */
2079 if (unlikely(!p->numa_faults)) {
2080 int size = sizeof(*p->numa_faults) *
2081 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2083 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2084 if (!p->numa_faults)
2087 p->total_numa_faults = 0;
2088 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2092 * First accesses are treated as private, otherwise consider accesses
2093 * to be private if the accessing pid has not changed
2095 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2098 priv = cpupid_match_pid(p, last_cpupid);
2099 if (!priv && !(flags & TNF_NO_GROUP))
2100 task_numa_group(p, last_cpupid, flags, &priv);
2104 * If a workload spans multiple NUMA nodes, a shared fault that
2105 * occurs wholly within the set of nodes that the workload is
2106 * actively using should be counted as local. This allows the
2107 * scan rate to slow down when a workload has settled down.
2109 if (!priv && !local && p->numa_group &&
2110 node_isset(cpu_node, p->numa_group->active_nodes) &&
2111 node_isset(mem_node, p->numa_group->active_nodes))
2114 task_numa_placement(p);
2117 * Retry task to preferred node migration periodically, in case it
2118 * case it previously failed, or the scheduler moved us.
2120 if (time_after(jiffies, p->numa_migrate_retry))
2121 numa_migrate_preferred(p);
2124 p->numa_pages_migrated += pages;
2125 if (flags & TNF_MIGRATE_FAIL)
2126 p->numa_faults_locality[2] += pages;
2128 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2129 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2130 p->numa_faults_locality[local] += pages;
2133 static void reset_ptenuma_scan(struct task_struct *p)
2136 * We only did a read acquisition of the mmap sem, so
2137 * p->mm->numa_scan_seq is written to without exclusive access
2138 * and the update is not guaranteed to be atomic. That's not
2139 * much of an issue though, since this is just used for
2140 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2141 * expensive, to avoid any form of compiler optimizations:
2143 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2144 p->mm->numa_scan_offset = 0;
2148 * The expensive part of numa migration is done from task_work context.
2149 * Triggered from task_tick_numa().
2151 void task_numa_work(struct callback_head *work)
2153 unsigned long migrate, next_scan, now = jiffies;
2154 struct task_struct *p = current;
2155 struct mm_struct *mm = p->mm;
2156 struct vm_area_struct *vma;
2157 unsigned long start, end;
2158 unsigned long nr_pte_updates = 0;
2159 long pages, virtpages;
2161 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2163 work->next = work; /* protect against double add */
2165 * Who cares about NUMA placement when they're dying.
2167 * NOTE: make sure not to dereference p->mm before this check,
2168 * exit_task_work() happens _after_ exit_mm() so we could be called
2169 * without p->mm even though we still had it when we enqueued this
2172 if (p->flags & PF_EXITING)
2175 if (!mm->numa_next_scan) {
2176 mm->numa_next_scan = now +
2177 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2181 * Enforce maximal scan/migration frequency..
2183 migrate = mm->numa_next_scan;
2184 if (time_before(now, migrate))
2187 if (p->numa_scan_period == 0) {
2188 p->numa_scan_period_max = task_scan_max(p);
2189 p->numa_scan_period = task_scan_min(p);
2192 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2193 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2197 * Delay this task enough that another task of this mm will likely win
2198 * the next time around.
2200 p->node_stamp += 2 * TICK_NSEC;
2202 start = mm->numa_scan_offset;
2203 pages = sysctl_numa_balancing_scan_size;
2204 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2205 virtpages = pages * 8; /* Scan up to this much virtual space */
2210 down_read(&mm->mmap_sem);
2211 vma = find_vma(mm, start);
2213 reset_ptenuma_scan(p);
2217 for (; vma; vma = vma->vm_next) {
2218 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2219 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2224 * Shared library pages mapped by multiple processes are not
2225 * migrated as it is expected they are cache replicated. Avoid
2226 * hinting faults in read-only file-backed mappings or the vdso
2227 * as migrating the pages will be of marginal benefit.
2230 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2234 * Skip inaccessible VMAs to avoid any confusion between
2235 * PROT_NONE and NUMA hinting ptes
2237 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2241 start = max(start, vma->vm_start);
2242 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2243 end = min(end, vma->vm_end);
2244 nr_pte_updates = change_prot_numa(vma, start, end);
2247 * Try to scan sysctl_numa_balancing_size worth of
2248 * hpages that have at least one present PTE that
2249 * is not already pte-numa. If the VMA contains
2250 * areas that are unused or already full of prot_numa
2251 * PTEs, scan up to virtpages, to skip through those
2255 pages -= (end - start) >> PAGE_SHIFT;
2256 virtpages -= (end - start) >> PAGE_SHIFT;
2259 if (pages <= 0 || virtpages <= 0)
2263 } while (end != vma->vm_end);
2268 * It is possible to reach the end of the VMA list but the last few
2269 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2270 * would find the !migratable VMA on the next scan but not reset the
2271 * scanner to the start so check it now.
2274 mm->numa_scan_offset = start;
2276 reset_ptenuma_scan(p);
2277 up_read(&mm->mmap_sem);
2281 * Drive the periodic memory faults..
2283 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2285 struct callback_head *work = &curr->numa_work;
2289 * We don't care about NUMA placement if we don't have memory.
2291 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2295 * Using runtime rather than walltime has the dual advantage that
2296 * we (mostly) drive the selection from busy threads and that the
2297 * task needs to have done some actual work before we bother with
2300 now = curr->se.sum_exec_runtime;
2301 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2303 if (now > curr->node_stamp + period) {
2304 if (!curr->node_stamp)
2305 curr->numa_scan_period = task_scan_min(curr);
2306 curr->node_stamp += period;
2308 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2309 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2310 task_work_add(curr, work, true);
2315 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2319 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2323 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2326 #endif /* CONFIG_NUMA_BALANCING */
2329 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2331 update_load_add(&cfs_rq->load, se->load.weight);
2332 if (!parent_entity(se))
2333 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2335 if (entity_is_task(se)) {
2336 struct rq *rq = rq_of(cfs_rq);
2338 account_numa_enqueue(rq, task_of(se));
2339 list_add(&se->group_node, &rq->cfs_tasks);
2342 cfs_rq->nr_running++;
2346 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2348 update_load_sub(&cfs_rq->load, se->load.weight);
2349 if (!parent_entity(se))
2350 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2351 if (entity_is_task(se)) {
2352 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2353 list_del_init(&se->group_node);
2355 cfs_rq->nr_running--;
2358 #ifdef CONFIG_FAIR_GROUP_SCHED
2360 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2365 * Use this CPU's real-time load instead of the last load contribution
2366 * as the updating of the contribution is delayed, and we will use the
2367 * the real-time load to calc the share. See update_tg_load_avg().
2369 tg_weight = atomic_long_read(&tg->load_avg);
2370 tg_weight -= cfs_rq->tg_load_avg_contrib;
2371 tg_weight += cfs_rq->load.weight;
2376 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2378 long tg_weight, load, shares;
2380 tg_weight = calc_tg_weight(tg, cfs_rq);
2381 load = cfs_rq->load.weight;
2383 shares = (tg->shares * load);
2385 shares /= tg_weight;
2387 if (shares < MIN_SHARES)
2388 shares = MIN_SHARES;
2389 if (shares > tg->shares)
2390 shares = tg->shares;
2394 # else /* CONFIG_SMP */
2395 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2399 # endif /* CONFIG_SMP */
2400 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2401 unsigned long weight)
2404 /* commit outstanding execution time */
2405 if (cfs_rq->curr == se)
2406 update_curr(cfs_rq);
2407 account_entity_dequeue(cfs_rq, se);
2410 update_load_set(&se->load, weight);
2413 account_entity_enqueue(cfs_rq, se);
2416 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2418 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2420 struct task_group *tg;
2421 struct sched_entity *se;
2425 se = tg->se[cpu_of(rq_of(cfs_rq))];
2426 if (!se || throttled_hierarchy(cfs_rq))
2429 if (likely(se->load.weight == tg->shares))
2432 shares = calc_cfs_shares(cfs_rq, tg);
2434 reweight_entity(cfs_rq_of(se), se, shares);
2436 #else /* CONFIG_FAIR_GROUP_SCHED */
2437 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2440 #endif /* CONFIG_FAIR_GROUP_SCHED */
2443 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2444 static const u32 runnable_avg_yN_inv[] = {
2445 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2446 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2447 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2448 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2449 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2450 0x85aac367, 0x82cd8698,
2454 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2455 * over-estimates when re-combining.
2457 static const u32 runnable_avg_yN_sum[] = {
2458 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2459 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2460 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2465 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2467 static __always_inline u64 decay_load(u64 val, u64 n)
2469 unsigned int local_n;
2473 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2476 /* after bounds checking we can collapse to 32-bit */
2480 * As y^PERIOD = 1/2, we can combine
2481 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2482 * With a look-up table which covers y^n (n<PERIOD)
2484 * To achieve constant time decay_load.
2486 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2487 val >>= local_n / LOAD_AVG_PERIOD;
2488 local_n %= LOAD_AVG_PERIOD;
2491 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2496 * For updates fully spanning n periods, the contribution to runnable
2497 * average will be: \Sum 1024*y^n
2499 * We can compute this reasonably efficiently by combining:
2500 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2502 static u32 __compute_runnable_contrib(u64 n)
2506 if (likely(n <= LOAD_AVG_PERIOD))
2507 return runnable_avg_yN_sum[n];
2508 else if (unlikely(n >= LOAD_AVG_MAX_N))
2509 return LOAD_AVG_MAX;
2511 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2513 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2514 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2516 n -= LOAD_AVG_PERIOD;
2517 } while (n > LOAD_AVG_PERIOD);
2519 contrib = decay_load(contrib, n);
2520 return contrib + runnable_avg_yN_sum[n];
2523 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2524 #error "load tracking assumes 2^10 as unit"
2527 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2530 * We can represent the historical contribution to runnable average as the
2531 * coefficients of a geometric series. To do this we sub-divide our runnable
2532 * history into segments of approximately 1ms (1024us); label the segment that
2533 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2535 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2537 * (now) (~1ms ago) (~2ms ago)
2539 * Let u_i denote the fraction of p_i that the entity was runnable.
2541 * We then designate the fractions u_i as our co-efficients, yielding the
2542 * following representation of historical load:
2543 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2545 * We choose y based on the with of a reasonably scheduling period, fixing:
2548 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2549 * approximately half as much as the contribution to load within the last ms
2552 * When a period "rolls over" and we have new u_0`, multiplying the previous
2553 * sum again by y is sufficient to update:
2554 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2555 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2557 static __always_inline int
2558 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2559 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2561 u64 delta, scaled_delta, periods;
2563 unsigned int delta_w, scaled_delta_w, decayed = 0;
2564 unsigned long scale_freq, scale_cpu;
2566 delta = now - sa->last_update_time;
2568 * This should only happen when time goes backwards, which it
2569 * unfortunately does during sched clock init when we swap over to TSC.
2571 if ((s64)delta < 0) {
2572 sa->last_update_time = now;
2577 * Use 1024ns as the unit of measurement since it's a reasonable
2578 * approximation of 1us and fast to compute.
2583 sa->last_update_time = now;
2585 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2586 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2588 /* delta_w is the amount already accumulated against our next period */
2589 delta_w = sa->period_contrib;
2590 if (delta + delta_w >= 1024) {
2593 /* how much left for next period will start over, we don't know yet */
2594 sa->period_contrib = 0;
2597 * Now that we know we're crossing a period boundary, figure
2598 * out how much from delta we need to complete the current
2599 * period and accrue it.
2601 delta_w = 1024 - delta_w;
2602 scaled_delta_w = cap_scale(delta_w, scale_freq);
2604 sa->load_sum += weight * scaled_delta_w;
2606 cfs_rq->runnable_load_sum +=
2607 weight * scaled_delta_w;
2611 sa->util_sum += scaled_delta_w * scale_cpu;
2615 /* Figure out how many additional periods this update spans */
2616 periods = delta / 1024;
2619 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2621 cfs_rq->runnable_load_sum =
2622 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2624 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2626 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2627 contrib = __compute_runnable_contrib(periods);
2628 contrib = cap_scale(contrib, scale_freq);
2630 sa->load_sum += weight * contrib;
2632 cfs_rq->runnable_load_sum += weight * contrib;
2635 sa->util_sum += contrib * scale_cpu;
2638 /* Remainder of delta accrued against u_0` */
2639 scaled_delta = cap_scale(delta, scale_freq);
2641 sa->load_sum += weight * scaled_delta;
2643 cfs_rq->runnable_load_sum += weight * scaled_delta;
2646 sa->util_sum += scaled_delta * scale_cpu;
2648 sa->period_contrib += delta;
2651 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2653 cfs_rq->runnable_load_avg =
2654 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2656 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2662 #ifdef CONFIG_FAIR_GROUP_SCHED
2664 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2665 * and effective_load (which is not done because it is too costly).
2667 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2669 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2671 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2672 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2673 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2677 #else /* CONFIG_FAIR_GROUP_SCHED */
2678 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2679 #endif /* CONFIG_FAIR_GROUP_SCHED */
2681 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2684 * Unsigned subtract and clamp on underflow.
2686 * Explicitly do a load-store to ensure the intermediate value never hits
2687 * memory. This allows lockless observations without ever seeing the negative
2690 #define sub_positive(_ptr, _val) do { \
2691 typeof(_ptr) ptr = (_ptr); \
2692 typeof(*ptr) val = (_val); \
2693 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2697 WRITE_ONCE(*ptr, res); \
2700 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2701 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2703 struct sched_avg *sa = &cfs_rq->avg;
2704 int decayed, removed = 0;
2706 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2707 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2708 sub_positive(&sa->load_avg, r);
2709 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2713 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2714 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2715 sub_positive(&sa->util_avg, r);
2716 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2719 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2720 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2722 #ifndef CONFIG_64BIT
2724 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2727 return decayed || removed;
2730 /* Update task and its cfs_rq load average */
2731 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2733 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2734 u64 now = cfs_rq_clock_task(cfs_rq);
2735 int cpu = cpu_of(rq_of(cfs_rq));
2738 * Track task load average for carrying it to new CPU after migrated, and
2739 * track group sched_entity load average for task_h_load calc in migration
2741 __update_load_avg(now, cpu, &se->avg,
2742 se->on_rq * scale_load_down(se->load.weight),
2743 cfs_rq->curr == se, NULL);
2745 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2746 update_tg_load_avg(cfs_rq, 0);
2749 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2751 if (!sched_feat(ATTACH_AGE_LOAD))
2755 * If we got migrated (either between CPUs or between cgroups) we'll
2756 * have aged the average right before clearing @last_update_time.
2758 if (se->avg.last_update_time) {
2759 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2760 &se->avg, 0, 0, NULL);
2763 * XXX: we could have just aged the entire load away if we've been
2764 * absent from the fair class for too long.
2769 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2770 cfs_rq->avg.load_avg += se->avg.load_avg;
2771 cfs_rq->avg.load_sum += se->avg.load_sum;
2772 cfs_rq->avg.util_avg += se->avg.util_avg;
2773 cfs_rq->avg.util_sum += se->avg.util_sum;
2776 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2778 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2779 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2780 cfs_rq->curr == se, NULL);
2782 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2783 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2784 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2785 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2788 /* Add the load generated by se into cfs_rq's load average */
2790 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2792 struct sched_avg *sa = &se->avg;
2793 u64 now = cfs_rq_clock_task(cfs_rq);
2794 int migrated, decayed;
2796 migrated = !sa->last_update_time;
2798 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2799 se->on_rq * scale_load_down(se->load.weight),
2800 cfs_rq->curr == se, NULL);
2803 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2805 cfs_rq->runnable_load_avg += sa->load_avg;
2806 cfs_rq->runnable_load_sum += sa->load_sum;
2809 attach_entity_load_avg(cfs_rq, se);
2811 if (decayed || migrated)
2812 update_tg_load_avg(cfs_rq, 0);
2815 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2817 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2819 update_load_avg(se, 1);
2821 cfs_rq->runnable_load_avg =
2822 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2823 cfs_rq->runnable_load_sum =
2824 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2828 * Task first catches up with cfs_rq, and then subtract
2829 * itself from the cfs_rq (task must be off the queue now).
2831 void remove_entity_load_avg(struct sched_entity *se)
2833 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2834 u64 last_update_time;
2836 #ifndef CONFIG_64BIT
2837 u64 last_update_time_copy;
2840 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2842 last_update_time = cfs_rq->avg.last_update_time;
2843 } while (last_update_time != last_update_time_copy);
2845 last_update_time = cfs_rq->avg.last_update_time;
2848 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2849 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2850 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2854 * Update the rq's load with the elapsed running time before entering
2855 * idle. if the last scheduled task is not a CFS task, idle_enter will
2856 * be the only way to update the runnable statistic.
2858 void idle_enter_fair(struct rq *this_rq)
2863 * Update the rq's load with the elapsed idle time before a task is
2864 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2865 * be the only way to update the runnable statistic.
2867 void idle_exit_fair(struct rq *this_rq)
2871 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2873 return cfs_rq->runnable_load_avg;
2876 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2878 return cfs_rq->avg.load_avg;
2881 static int idle_balance(struct rq *this_rq);
2883 #else /* CONFIG_SMP */
2885 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2887 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2889 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2890 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2893 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2895 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2897 static inline int idle_balance(struct rq *rq)
2902 #endif /* CONFIG_SMP */
2904 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2906 #ifdef CONFIG_SCHEDSTATS
2907 struct task_struct *tsk = NULL;
2909 if (entity_is_task(se))
2912 if (se->statistics.sleep_start) {
2913 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2918 if (unlikely(delta > se->statistics.sleep_max))
2919 se->statistics.sleep_max = delta;
2921 se->statistics.sleep_start = 0;
2922 se->statistics.sum_sleep_runtime += delta;
2925 account_scheduler_latency(tsk, delta >> 10, 1);
2926 trace_sched_stat_sleep(tsk, delta);
2929 if (se->statistics.block_start) {
2930 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2935 if (unlikely(delta > se->statistics.block_max))
2936 se->statistics.block_max = delta;
2938 se->statistics.block_start = 0;
2939 se->statistics.sum_sleep_runtime += delta;
2942 if (tsk->in_iowait) {
2943 se->statistics.iowait_sum += delta;
2944 se->statistics.iowait_count++;
2945 trace_sched_stat_iowait(tsk, delta);
2948 trace_sched_stat_blocked(tsk, delta);
2949 trace_sched_blocked_reason(tsk);
2952 * Blocking time is in units of nanosecs, so shift by
2953 * 20 to get a milliseconds-range estimation of the
2954 * amount of time that the task spent sleeping:
2956 if (unlikely(prof_on == SLEEP_PROFILING)) {
2957 profile_hits(SLEEP_PROFILING,
2958 (void *)get_wchan(tsk),
2961 account_scheduler_latency(tsk, delta >> 10, 0);
2967 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2969 #ifdef CONFIG_SCHED_DEBUG
2970 s64 d = se->vruntime - cfs_rq->min_vruntime;
2975 if (d > 3*sysctl_sched_latency)
2976 schedstat_inc(cfs_rq, nr_spread_over);
2981 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2983 u64 vruntime = cfs_rq->min_vruntime;
2986 * The 'current' period is already promised to the current tasks,
2987 * however the extra weight of the new task will slow them down a
2988 * little, place the new task so that it fits in the slot that
2989 * stays open at the end.
2991 if (initial && sched_feat(START_DEBIT))
2992 vruntime += sched_vslice(cfs_rq, se);
2994 /* sleeps up to a single latency don't count. */
2996 unsigned long thresh = sysctl_sched_latency;
2999 * Halve their sleep time's effect, to allow
3000 * for a gentler effect of sleepers:
3002 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3008 /* ensure we never gain time by being placed backwards. */
3009 se->vruntime = max_vruntime(se->vruntime, vruntime);
3012 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3015 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3018 * Update the normalized vruntime before updating min_vruntime
3019 * through calling update_curr().
3021 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3022 se->vruntime += cfs_rq->min_vruntime;
3025 * Update run-time statistics of the 'current'.
3027 update_curr(cfs_rq);
3028 enqueue_entity_load_avg(cfs_rq, se);
3029 account_entity_enqueue(cfs_rq, se);
3030 update_cfs_shares(cfs_rq);
3032 if (flags & ENQUEUE_WAKEUP) {
3033 place_entity(cfs_rq, se, 0);
3034 enqueue_sleeper(cfs_rq, se);
3037 update_stats_enqueue(cfs_rq, se);
3038 check_spread(cfs_rq, se);
3039 if (se != cfs_rq->curr)
3040 __enqueue_entity(cfs_rq, se);
3043 if (cfs_rq->nr_running == 1) {
3044 list_add_leaf_cfs_rq(cfs_rq);
3045 check_enqueue_throttle(cfs_rq);
3049 static void __clear_buddies_last(struct sched_entity *se)
3051 for_each_sched_entity(se) {
3052 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3053 if (cfs_rq->last != se)
3056 cfs_rq->last = NULL;
3060 static void __clear_buddies_next(struct sched_entity *se)
3062 for_each_sched_entity(se) {
3063 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3064 if (cfs_rq->next != se)
3067 cfs_rq->next = NULL;
3071 static void __clear_buddies_skip(struct sched_entity *se)
3073 for_each_sched_entity(se) {
3074 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3075 if (cfs_rq->skip != se)
3078 cfs_rq->skip = NULL;
3082 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3084 if (cfs_rq->last == se)
3085 __clear_buddies_last(se);
3087 if (cfs_rq->next == se)
3088 __clear_buddies_next(se);
3090 if (cfs_rq->skip == se)
3091 __clear_buddies_skip(se);
3094 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3097 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3100 * Update run-time statistics of the 'current'.
3102 update_curr(cfs_rq);
3103 dequeue_entity_load_avg(cfs_rq, se);
3105 update_stats_dequeue(cfs_rq, se);
3106 if (flags & DEQUEUE_SLEEP) {
3107 #ifdef CONFIG_SCHEDSTATS
3108 if (entity_is_task(se)) {
3109 struct task_struct *tsk = task_of(se);
3111 if (tsk->state & TASK_INTERRUPTIBLE)
3112 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3113 if (tsk->state & TASK_UNINTERRUPTIBLE)
3114 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3119 clear_buddies(cfs_rq, se);
3121 if (se != cfs_rq->curr)
3122 __dequeue_entity(cfs_rq, se);
3124 account_entity_dequeue(cfs_rq, se);
3127 * Normalize the entity after updating the min_vruntime because the
3128 * update can refer to the ->curr item and we need to reflect this
3129 * movement in our normalized position.
3131 if (!(flags & DEQUEUE_SLEEP))
3132 se->vruntime -= cfs_rq->min_vruntime;
3134 /* return excess runtime on last dequeue */
3135 return_cfs_rq_runtime(cfs_rq);
3137 update_min_vruntime(cfs_rq);
3138 update_cfs_shares(cfs_rq);
3142 * Preempt the current task with a newly woken task if needed:
3145 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3147 unsigned long ideal_runtime, delta_exec;
3148 struct sched_entity *se;
3151 ideal_runtime = sched_slice(cfs_rq, curr);
3152 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3153 if (delta_exec > ideal_runtime) {
3154 resched_curr(rq_of(cfs_rq));
3156 * The current task ran long enough, ensure it doesn't get
3157 * re-elected due to buddy favours.
3159 clear_buddies(cfs_rq, curr);
3164 * Ensure that a task that missed wakeup preemption by a
3165 * narrow margin doesn't have to wait for a full slice.
3166 * This also mitigates buddy induced latencies under load.
3168 if (delta_exec < sysctl_sched_min_granularity)
3171 se = __pick_first_entity(cfs_rq);
3172 delta = curr->vruntime - se->vruntime;
3177 if (delta > ideal_runtime)
3178 resched_curr(rq_of(cfs_rq));
3182 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3184 /* 'current' is not kept within the tree. */
3187 * Any task has to be enqueued before it get to execute on
3188 * a CPU. So account for the time it spent waiting on the
3191 update_stats_wait_end(cfs_rq, se);
3192 __dequeue_entity(cfs_rq, se);
3193 update_load_avg(se, 1);
3196 update_stats_curr_start(cfs_rq, se);
3198 #ifdef CONFIG_SCHEDSTATS
3200 * Track our maximum slice length, if the CPU's load is at
3201 * least twice that of our own weight (i.e. dont track it
3202 * when there are only lesser-weight tasks around):
3204 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3205 se->statistics.slice_max = max(se->statistics.slice_max,
3206 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3209 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3213 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3216 * Pick the next process, keeping these things in mind, in this order:
3217 * 1) keep things fair between processes/task groups
3218 * 2) pick the "next" process, since someone really wants that to run
3219 * 3) pick the "last" process, for cache locality
3220 * 4) do not run the "skip" process, if something else is available
3222 static struct sched_entity *
3223 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3225 struct sched_entity *left = __pick_first_entity(cfs_rq);
3226 struct sched_entity *se;
3229 * If curr is set we have to see if its left of the leftmost entity
3230 * still in the tree, provided there was anything in the tree at all.
3232 if (!left || (curr && entity_before(curr, left)))
3235 se = left; /* ideally we run the leftmost entity */
3238 * Avoid running the skip buddy, if running something else can
3239 * be done without getting too unfair.
3241 if (cfs_rq->skip == se) {
3242 struct sched_entity *second;
3245 second = __pick_first_entity(cfs_rq);
3247 second = __pick_next_entity(se);
3248 if (!second || (curr && entity_before(curr, second)))
3252 if (second && wakeup_preempt_entity(second, left) < 1)
3257 * Prefer last buddy, try to return the CPU to a preempted task.
3259 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3263 * Someone really wants this to run. If it's not unfair, run it.
3265 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3268 clear_buddies(cfs_rq, se);
3273 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3275 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3278 * If still on the runqueue then deactivate_task()
3279 * was not called and update_curr() has to be done:
3282 update_curr(cfs_rq);
3284 /* throttle cfs_rqs exceeding runtime */
3285 check_cfs_rq_runtime(cfs_rq);
3287 check_spread(cfs_rq, prev);
3289 update_stats_wait_start(cfs_rq, prev);
3290 /* Put 'current' back into the tree. */
3291 __enqueue_entity(cfs_rq, prev);
3292 /* in !on_rq case, update occurred at dequeue */
3293 update_load_avg(prev, 0);
3295 cfs_rq->curr = NULL;
3299 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3302 * Update run-time statistics of the 'current'.
3304 update_curr(cfs_rq);
3307 * Ensure that runnable average is periodically updated.
3309 update_load_avg(curr, 1);
3310 update_cfs_shares(cfs_rq);
3312 #ifdef CONFIG_SCHED_HRTICK
3314 * queued ticks are scheduled to match the slice, so don't bother
3315 * validating it and just reschedule.
3318 resched_curr(rq_of(cfs_rq));
3322 * don't let the period tick interfere with the hrtick preemption
3324 if (!sched_feat(DOUBLE_TICK) &&
3325 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3329 if (cfs_rq->nr_running > 1)
3330 check_preempt_tick(cfs_rq, curr);
3334 /**************************************************
3335 * CFS bandwidth control machinery
3338 #ifdef CONFIG_CFS_BANDWIDTH
3340 #ifdef HAVE_JUMP_LABEL
3341 static struct static_key __cfs_bandwidth_used;
3343 static inline bool cfs_bandwidth_used(void)
3345 return static_key_false(&__cfs_bandwidth_used);
3348 void cfs_bandwidth_usage_inc(void)
3350 static_key_slow_inc(&__cfs_bandwidth_used);
3353 void cfs_bandwidth_usage_dec(void)
3355 static_key_slow_dec(&__cfs_bandwidth_used);
3357 #else /* HAVE_JUMP_LABEL */
3358 static bool cfs_bandwidth_used(void)
3363 void cfs_bandwidth_usage_inc(void) {}
3364 void cfs_bandwidth_usage_dec(void) {}
3365 #endif /* HAVE_JUMP_LABEL */
3368 * default period for cfs group bandwidth.
3369 * default: 0.1s, units: nanoseconds
3371 static inline u64 default_cfs_period(void)
3373 return 100000000ULL;
3376 static inline u64 sched_cfs_bandwidth_slice(void)
3378 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3382 * Replenish runtime according to assigned quota and update expiration time.
3383 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3384 * additional synchronization around rq->lock.
3386 * requires cfs_b->lock
3388 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3392 if (cfs_b->quota == RUNTIME_INF)
3395 now = sched_clock_cpu(smp_processor_id());
3396 cfs_b->runtime = cfs_b->quota;
3397 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3400 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3402 return &tg->cfs_bandwidth;
3405 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3406 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3408 if (unlikely(cfs_rq->throttle_count))
3409 return cfs_rq->throttled_clock_task;
3411 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3414 /* returns 0 on failure to allocate runtime */
3415 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3417 struct task_group *tg = cfs_rq->tg;
3418 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3419 u64 amount = 0, min_amount, expires;
3421 /* note: this is a positive sum as runtime_remaining <= 0 */
3422 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3424 raw_spin_lock(&cfs_b->lock);
3425 if (cfs_b->quota == RUNTIME_INF)
3426 amount = min_amount;
3428 start_cfs_bandwidth(cfs_b);
3430 if (cfs_b->runtime > 0) {
3431 amount = min(cfs_b->runtime, min_amount);
3432 cfs_b->runtime -= amount;
3436 expires = cfs_b->runtime_expires;
3437 raw_spin_unlock(&cfs_b->lock);
3439 cfs_rq->runtime_remaining += amount;
3441 * we may have advanced our local expiration to account for allowed
3442 * spread between our sched_clock and the one on which runtime was
3445 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3446 cfs_rq->runtime_expires = expires;
3448 return cfs_rq->runtime_remaining > 0;
3452 * Note: This depends on the synchronization provided by sched_clock and the
3453 * fact that rq->clock snapshots this value.
3455 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3457 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3459 /* if the deadline is ahead of our clock, nothing to do */
3460 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3463 if (cfs_rq->runtime_remaining < 0)
3467 * If the local deadline has passed we have to consider the
3468 * possibility that our sched_clock is 'fast' and the global deadline
3469 * has not truly expired.
3471 * Fortunately we can check determine whether this the case by checking
3472 * whether the global deadline has advanced. It is valid to compare
3473 * cfs_b->runtime_expires without any locks since we only care about
3474 * exact equality, so a partial write will still work.
3477 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3478 /* extend local deadline, drift is bounded above by 2 ticks */
3479 cfs_rq->runtime_expires += TICK_NSEC;
3481 /* global deadline is ahead, expiration has passed */
3482 cfs_rq->runtime_remaining = 0;
3486 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3488 /* dock delta_exec before expiring quota (as it could span periods) */
3489 cfs_rq->runtime_remaining -= delta_exec;
3490 expire_cfs_rq_runtime(cfs_rq);
3492 if (likely(cfs_rq->runtime_remaining > 0))
3496 * if we're unable to extend our runtime we resched so that the active
3497 * hierarchy can be throttled
3499 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3500 resched_curr(rq_of(cfs_rq));
3503 static __always_inline
3504 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3506 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3509 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3512 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3514 return cfs_bandwidth_used() && cfs_rq->throttled;
3517 /* check whether cfs_rq, or any parent, is throttled */
3518 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3520 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3524 * Ensure that neither of the group entities corresponding to src_cpu or
3525 * dest_cpu are members of a throttled hierarchy when performing group
3526 * load-balance operations.
3528 static inline int throttled_lb_pair(struct task_group *tg,
3529 int src_cpu, int dest_cpu)
3531 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3533 src_cfs_rq = tg->cfs_rq[src_cpu];
3534 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3536 return throttled_hierarchy(src_cfs_rq) ||
3537 throttled_hierarchy(dest_cfs_rq);
3540 /* updated child weight may affect parent so we have to do this bottom up */
3541 static int tg_unthrottle_up(struct task_group *tg, void *data)
3543 struct rq *rq = data;
3544 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3546 cfs_rq->throttle_count--;
3548 if (!cfs_rq->throttle_count) {
3549 /* adjust cfs_rq_clock_task() */
3550 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3551 cfs_rq->throttled_clock_task;
3558 static int tg_throttle_down(struct task_group *tg, void *data)
3560 struct rq *rq = data;
3561 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3563 /* group is entering throttled state, stop time */
3564 if (!cfs_rq->throttle_count)
3565 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3566 cfs_rq->throttle_count++;
3571 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3573 struct rq *rq = rq_of(cfs_rq);
3574 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3575 struct sched_entity *se;
3576 long task_delta, dequeue = 1;
3579 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3581 /* freeze hierarchy runnable averages while throttled */
3583 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3586 task_delta = cfs_rq->h_nr_running;
3587 for_each_sched_entity(se) {
3588 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3589 /* throttled entity or throttle-on-deactivate */
3594 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3595 qcfs_rq->h_nr_running -= task_delta;
3597 if (qcfs_rq->load.weight)
3602 sub_nr_running(rq, task_delta);
3604 cfs_rq->throttled = 1;
3605 cfs_rq->throttled_clock = rq_clock(rq);
3606 raw_spin_lock(&cfs_b->lock);
3607 empty = list_empty(&cfs_b->throttled_cfs_rq);
3610 * Add to the _head_ of the list, so that an already-started
3611 * distribute_cfs_runtime will not see us
3613 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3616 * If we're the first throttled task, make sure the bandwidth
3620 start_cfs_bandwidth(cfs_b);
3622 raw_spin_unlock(&cfs_b->lock);
3625 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3627 struct rq *rq = rq_of(cfs_rq);
3628 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3629 struct sched_entity *se;
3633 se = cfs_rq->tg->se[cpu_of(rq)];
3635 cfs_rq->throttled = 0;
3637 update_rq_clock(rq);
3639 raw_spin_lock(&cfs_b->lock);
3640 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3641 list_del_rcu(&cfs_rq->throttled_list);
3642 raw_spin_unlock(&cfs_b->lock);
3644 /* update hierarchical throttle state */
3645 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3647 if (!cfs_rq->load.weight)
3650 task_delta = cfs_rq->h_nr_running;
3651 for_each_sched_entity(se) {
3655 cfs_rq = cfs_rq_of(se);
3657 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3658 cfs_rq->h_nr_running += task_delta;
3660 if (cfs_rq_throttled(cfs_rq))
3665 add_nr_running(rq, task_delta);
3667 /* determine whether we need to wake up potentially idle cpu */
3668 if (rq->curr == rq->idle && rq->cfs.nr_running)
3672 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3673 u64 remaining, u64 expires)
3675 struct cfs_rq *cfs_rq;
3677 u64 starting_runtime = remaining;
3680 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3682 struct rq *rq = rq_of(cfs_rq);
3684 raw_spin_lock(&rq->lock);
3685 if (!cfs_rq_throttled(cfs_rq))
3688 runtime = -cfs_rq->runtime_remaining + 1;
3689 if (runtime > remaining)
3690 runtime = remaining;
3691 remaining -= runtime;
3693 cfs_rq->runtime_remaining += runtime;
3694 cfs_rq->runtime_expires = expires;
3696 /* we check whether we're throttled above */
3697 if (cfs_rq->runtime_remaining > 0)
3698 unthrottle_cfs_rq(cfs_rq);
3701 raw_spin_unlock(&rq->lock);
3708 return starting_runtime - remaining;
3712 * Responsible for refilling a task_group's bandwidth and unthrottling its
3713 * cfs_rqs as appropriate. If there has been no activity within the last
3714 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3715 * used to track this state.
3717 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3719 u64 runtime, runtime_expires;
3722 /* no need to continue the timer with no bandwidth constraint */
3723 if (cfs_b->quota == RUNTIME_INF)
3724 goto out_deactivate;
3726 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3727 cfs_b->nr_periods += overrun;
3730 * idle depends on !throttled (for the case of a large deficit), and if
3731 * we're going inactive then everything else can be deferred
3733 if (cfs_b->idle && !throttled)
3734 goto out_deactivate;
3736 __refill_cfs_bandwidth_runtime(cfs_b);
3739 /* mark as potentially idle for the upcoming period */
3744 /* account preceding periods in which throttling occurred */
3745 cfs_b->nr_throttled += overrun;
3747 runtime_expires = cfs_b->runtime_expires;
3750 * This check is repeated as we are holding onto the new bandwidth while
3751 * we unthrottle. This can potentially race with an unthrottled group
3752 * trying to acquire new bandwidth from the global pool. This can result
3753 * in us over-using our runtime if it is all used during this loop, but
3754 * only by limited amounts in that extreme case.
3756 while (throttled && cfs_b->runtime > 0) {
3757 runtime = cfs_b->runtime;
3758 raw_spin_unlock(&cfs_b->lock);
3759 /* we can't nest cfs_b->lock while distributing bandwidth */
3760 runtime = distribute_cfs_runtime(cfs_b, runtime,
3762 raw_spin_lock(&cfs_b->lock);
3764 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3766 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3770 * While we are ensured activity in the period following an
3771 * unthrottle, this also covers the case in which the new bandwidth is
3772 * insufficient to cover the existing bandwidth deficit. (Forcing the
3773 * timer to remain active while there are any throttled entities.)
3783 /* a cfs_rq won't donate quota below this amount */
3784 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3785 /* minimum remaining period time to redistribute slack quota */
3786 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3787 /* how long we wait to gather additional slack before distributing */
3788 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3791 * Are we near the end of the current quota period?
3793 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3794 * hrtimer base being cleared by hrtimer_start. In the case of
3795 * migrate_hrtimers, base is never cleared, so we are fine.
3797 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3799 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3802 /* if the call-back is running a quota refresh is already occurring */
3803 if (hrtimer_callback_running(refresh_timer))
3806 /* is a quota refresh about to occur? */
3807 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3808 if (remaining < min_expire)
3814 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3816 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3818 /* if there's a quota refresh soon don't bother with slack */
3819 if (runtime_refresh_within(cfs_b, min_left))
3822 hrtimer_start(&cfs_b->slack_timer,
3823 ns_to_ktime(cfs_bandwidth_slack_period),
3827 /* we know any runtime found here is valid as update_curr() precedes return */
3828 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3830 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3831 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3833 if (slack_runtime <= 0)
3836 raw_spin_lock(&cfs_b->lock);
3837 if (cfs_b->quota != RUNTIME_INF &&
3838 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3839 cfs_b->runtime += slack_runtime;
3841 /* we are under rq->lock, defer unthrottling using a timer */
3842 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3843 !list_empty(&cfs_b->throttled_cfs_rq))
3844 start_cfs_slack_bandwidth(cfs_b);
3846 raw_spin_unlock(&cfs_b->lock);
3848 /* even if it's not valid for return we don't want to try again */
3849 cfs_rq->runtime_remaining -= slack_runtime;
3852 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3854 if (!cfs_bandwidth_used())
3857 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3860 __return_cfs_rq_runtime(cfs_rq);
3864 * This is done with a timer (instead of inline with bandwidth return) since
3865 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3867 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3869 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3872 /* confirm we're still not at a refresh boundary */
3873 raw_spin_lock(&cfs_b->lock);
3874 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3875 raw_spin_unlock(&cfs_b->lock);
3879 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3880 runtime = cfs_b->runtime;
3882 expires = cfs_b->runtime_expires;
3883 raw_spin_unlock(&cfs_b->lock);
3888 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3890 raw_spin_lock(&cfs_b->lock);
3891 if (expires == cfs_b->runtime_expires)
3892 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3893 raw_spin_unlock(&cfs_b->lock);
3897 * When a group wakes up we want to make sure that its quota is not already
3898 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3899 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3901 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3903 if (!cfs_bandwidth_used())
3906 /* an active group must be handled by the update_curr()->put() path */
3907 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3910 /* ensure the group is not already throttled */
3911 if (cfs_rq_throttled(cfs_rq))
3914 /* update runtime allocation */
3915 account_cfs_rq_runtime(cfs_rq, 0);
3916 if (cfs_rq->runtime_remaining <= 0)
3917 throttle_cfs_rq(cfs_rq);
3920 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3921 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3923 if (!cfs_bandwidth_used())
3926 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3930 * it's possible for a throttled entity to be forced into a running
3931 * state (e.g. set_curr_task), in this case we're finished.
3933 if (cfs_rq_throttled(cfs_rq))
3936 throttle_cfs_rq(cfs_rq);
3940 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3942 struct cfs_bandwidth *cfs_b =
3943 container_of(timer, struct cfs_bandwidth, slack_timer);
3945 do_sched_cfs_slack_timer(cfs_b);
3947 return HRTIMER_NORESTART;
3950 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3952 struct cfs_bandwidth *cfs_b =
3953 container_of(timer, struct cfs_bandwidth, period_timer);
3957 raw_spin_lock(&cfs_b->lock);
3959 overrun = hrtimer_forward_now(timer, cfs_b->period);
3963 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3966 cfs_b->period_active = 0;
3967 raw_spin_unlock(&cfs_b->lock);
3969 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3972 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3974 raw_spin_lock_init(&cfs_b->lock);
3976 cfs_b->quota = RUNTIME_INF;
3977 cfs_b->period = ns_to_ktime(default_cfs_period());
3979 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3980 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
3981 cfs_b->period_timer.function = sched_cfs_period_timer;
3982 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3983 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3986 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3988 cfs_rq->runtime_enabled = 0;
3989 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3992 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3994 lockdep_assert_held(&cfs_b->lock);
3996 if (!cfs_b->period_active) {
3997 cfs_b->period_active = 1;
3998 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
3999 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4003 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4005 /* init_cfs_bandwidth() was not called */
4006 if (!cfs_b->throttled_cfs_rq.next)
4009 hrtimer_cancel(&cfs_b->period_timer);
4010 hrtimer_cancel(&cfs_b->slack_timer);
4013 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4015 struct cfs_rq *cfs_rq;
4017 for_each_leaf_cfs_rq(rq, cfs_rq) {
4018 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4020 raw_spin_lock(&cfs_b->lock);
4021 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4022 raw_spin_unlock(&cfs_b->lock);
4026 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4028 struct cfs_rq *cfs_rq;
4030 for_each_leaf_cfs_rq(rq, cfs_rq) {
4031 if (!cfs_rq->runtime_enabled)
4035 * clock_task is not advancing so we just need to make sure
4036 * there's some valid quota amount
4038 cfs_rq->runtime_remaining = 1;
4040 * Offline rq is schedulable till cpu is completely disabled
4041 * in take_cpu_down(), so we prevent new cfs throttling here.
4043 cfs_rq->runtime_enabled = 0;
4045 if (cfs_rq_throttled(cfs_rq))
4046 unthrottle_cfs_rq(cfs_rq);
4050 #else /* CONFIG_CFS_BANDWIDTH */
4051 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4053 return rq_clock_task(rq_of(cfs_rq));
4056 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4057 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4058 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4059 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4061 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4066 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4071 static inline int throttled_lb_pair(struct task_group *tg,
4072 int src_cpu, int dest_cpu)
4077 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4079 #ifdef CONFIG_FAIR_GROUP_SCHED
4080 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4083 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4087 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4088 static inline void update_runtime_enabled(struct rq *rq) {}
4089 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4091 #endif /* CONFIG_CFS_BANDWIDTH */
4093 /**************************************************
4094 * CFS operations on tasks:
4097 #ifdef CONFIG_SCHED_HRTICK
4098 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4100 struct sched_entity *se = &p->se;
4101 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4103 WARN_ON(task_rq(p) != rq);
4105 if (cfs_rq->nr_running > 1) {
4106 u64 slice = sched_slice(cfs_rq, se);
4107 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4108 s64 delta = slice - ran;
4115 hrtick_start(rq, delta);
4120 * called from enqueue/dequeue and updates the hrtick when the
4121 * current task is from our class and nr_running is low enough
4124 static void hrtick_update(struct rq *rq)
4126 struct task_struct *curr = rq->curr;
4128 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4131 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4132 hrtick_start_fair(rq, curr);
4134 #else /* !CONFIG_SCHED_HRTICK */
4136 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4140 static inline void hrtick_update(struct rq *rq)
4146 * The enqueue_task method is called before nr_running is
4147 * increased. Here we update the fair scheduling stats and
4148 * then put the task into the rbtree:
4151 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4153 struct cfs_rq *cfs_rq;
4154 struct sched_entity *se = &p->se;
4156 for_each_sched_entity(se) {
4159 cfs_rq = cfs_rq_of(se);
4160 enqueue_entity(cfs_rq, se, flags);
4163 * end evaluation on encountering a throttled cfs_rq
4165 * note: in the case of encountering a throttled cfs_rq we will
4166 * post the final h_nr_running increment below.
4168 if (cfs_rq_throttled(cfs_rq))
4170 cfs_rq->h_nr_running++;
4172 flags = ENQUEUE_WAKEUP;
4175 for_each_sched_entity(se) {
4176 cfs_rq = cfs_rq_of(se);
4177 cfs_rq->h_nr_running++;
4179 if (cfs_rq_throttled(cfs_rq))
4182 update_load_avg(se, 1);
4183 update_cfs_shares(cfs_rq);
4187 add_nr_running(rq, 1);
4192 static void set_next_buddy(struct sched_entity *se);
4195 * The dequeue_task method is called before nr_running is
4196 * decreased. We remove the task from the rbtree and
4197 * update the fair scheduling stats:
4199 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4201 struct cfs_rq *cfs_rq;
4202 struct sched_entity *se = &p->se;
4203 int task_sleep = flags & DEQUEUE_SLEEP;
4205 for_each_sched_entity(se) {
4206 cfs_rq = cfs_rq_of(se);
4207 dequeue_entity(cfs_rq, se, flags);
4210 * end evaluation on encountering a throttled cfs_rq
4212 * note: in the case of encountering a throttled cfs_rq we will
4213 * post the final h_nr_running decrement below.
4215 if (cfs_rq_throttled(cfs_rq))
4217 cfs_rq->h_nr_running--;
4219 /* Don't dequeue parent if it has other entities besides us */
4220 if (cfs_rq->load.weight) {
4222 * Bias pick_next to pick a task from this cfs_rq, as
4223 * p is sleeping when it is within its sched_slice.
4225 if (task_sleep && parent_entity(se))
4226 set_next_buddy(parent_entity(se));
4228 /* avoid re-evaluating load for this entity */
4229 se = parent_entity(se);
4232 flags |= DEQUEUE_SLEEP;
4235 for_each_sched_entity(se) {
4236 cfs_rq = cfs_rq_of(se);
4237 cfs_rq->h_nr_running--;
4239 if (cfs_rq_throttled(cfs_rq))
4242 update_load_avg(se, 1);
4243 update_cfs_shares(cfs_rq);
4247 sub_nr_running(rq, 1);
4255 * per rq 'load' arrray crap; XXX kill this.
4259 * The exact cpuload at various idx values, calculated at every tick would be
4260 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4262 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4263 * on nth tick when cpu may be busy, then we have:
4264 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4265 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4267 * decay_load_missed() below does efficient calculation of
4268 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4269 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4271 * The calculation is approximated on a 128 point scale.
4272 * degrade_zero_ticks is the number of ticks after which load at any
4273 * particular idx is approximated to be zero.
4274 * degrade_factor is a precomputed table, a row for each load idx.
4275 * Each column corresponds to degradation factor for a power of two ticks,
4276 * based on 128 point scale.
4278 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4279 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4281 * With this power of 2 load factors, we can degrade the load n times
4282 * by looking at 1 bits in n and doing as many mult/shift instead of
4283 * n mult/shifts needed by the exact degradation.
4285 #define DEGRADE_SHIFT 7
4286 static const unsigned char
4287 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4288 static const unsigned char
4289 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4290 {0, 0, 0, 0, 0, 0, 0, 0},
4291 {64, 32, 8, 0, 0, 0, 0, 0},
4292 {96, 72, 40, 12, 1, 0, 0},
4293 {112, 98, 75, 43, 15, 1, 0},
4294 {120, 112, 98, 76, 45, 16, 2} };
4297 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4298 * would be when CPU is idle and so we just decay the old load without
4299 * adding any new load.
4301 static unsigned long
4302 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4306 if (!missed_updates)
4309 if (missed_updates >= degrade_zero_ticks[idx])
4313 return load >> missed_updates;
4315 while (missed_updates) {
4316 if (missed_updates % 2)
4317 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4319 missed_updates >>= 1;
4326 * Update rq->cpu_load[] statistics. This function is usually called every
4327 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4328 * every tick. We fix it up based on jiffies.
4330 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4331 unsigned long pending_updates)
4335 this_rq->nr_load_updates++;
4337 /* Update our load: */
4338 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4339 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4340 unsigned long old_load, new_load;
4342 /* scale is effectively 1 << i now, and >> i divides by scale */
4344 old_load = this_rq->cpu_load[i];
4345 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4346 new_load = this_load;
4348 * Round up the averaging division if load is increasing. This
4349 * prevents us from getting stuck on 9 if the load is 10, for
4352 if (new_load > old_load)
4353 new_load += scale - 1;
4355 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4358 sched_avg_update(this_rq);
4361 /* Used instead of source_load when we know the type == 0 */
4362 static unsigned long weighted_cpuload(const int cpu)
4364 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4367 #ifdef CONFIG_NO_HZ_COMMON
4369 * There is no sane way to deal with nohz on smp when using jiffies because the
4370 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4371 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4373 * Therefore we cannot use the delta approach from the regular tick since that
4374 * would seriously skew the load calculation. However we'll make do for those
4375 * updates happening while idle (nohz_idle_balance) or coming out of idle
4376 * (tick_nohz_idle_exit).
4378 * This means we might still be one tick off for nohz periods.
4382 * Called from nohz_idle_balance() to update the load ratings before doing the
4385 static void update_idle_cpu_load(struct rq *this_rq)
4387 unsigned long curr_jiffies = READ_ONCE(jiffies);
4388 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4389 unsigned long pending_updates;
4392 * bail if there's load or we're actually up-to-date.
4394 if (load || curr_jiffies == this_rq->last_load_update_tick)
4397 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4398 this_rq->last_load_update_tick = curr_jiffies;
4400 __update_cpu_load(this_rq, load, pending_updates);
4404 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4406 void update_cpu_load_nohz(void)
4408 struct rq *this_rq = this_rq();
4409 unsigned long curr_jiffies = READ_ONCE(jiffies);
4410 unsigned long pending_updates;
4412 if (curr_jiffies == this_rq->last_load_update_tick)
4415 raw_spin_lock(&this_rq->lock);
4416 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4417 if (pending_updates) {
4418 this_rq->last_load_update_tick = curr_jiffies;
4420 * We were idle, this means load 0, the current load might be
4421 * !0 due to remote wakeups and the sort.
4423 __update_cpu_load(this_rq, 0, pending_updates);
4425 raw_spin_unlock(&this_rq->lock);
4427 #endif /* CONFIG_NO_HZ */
4430 * Called from scheduler_tick()
4432 void update_cpu_load_active(struct rq *this_rq)
4434 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4436 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4438 this_rq->last_load_update_tick = jiffies;
4439 __update_cpu_load(this_rq, load, 1);
4443 * Return a low guess at the load of a migration-source cpu weighted
4444 * according to the scheduling class and "nice" value.
4446 * We want to under-estimate the load of migration sources, to
4447 * balance conservatively.
4449 static unsigned long source_load(int cpu, int type)
4451 struct rq *rq = cpu_rq(cpu);
4452 unsigned long total = weighted_cpuload(cpu);
4454 if (type == 0 || !sched_feat(LB_BIAS))
4457 return min(rq->cpu_load[type-1], total);
4461 * Return a high guess at the load of a migration-target cpu weighted
4462 * according to the scheduling class and "nice" value.
4464 static unsigned long target_load(int cpu, int type)
4466 struct rq *rq = cpu_rq(cpu);
4467 unsigned long total = weighted_cpuload(cpu);
4469 if (type == 0 || !sched_feat(LB_BIAS))
4472 return max(rq->cpu_load[type-1], total);
4475 static unsigned long capacity_of(int cpu)
4477 return cpu_rq(cpu)->cpu_capacity;
4480 static unsigned long capacity_orig_of(int cpu)
4482 return cpu_rq(cpu)->cpu_capacity_orig;
4485 static unsigned long cpu_avg_load_per_task(int cpu)
4487 struct rq *rq = cpu_rq(cpu);
4488 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4489 unsigned long load_avg = weighted_cpuload(cpu);
4492 return load_avg / nr_running;
4497 static void record_wakee(struct task_struct *p)
4500 * Rough decay (wiping) for cost saving, don't worry
4501 * about the boundary, really active task won't care
4504 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4505 current->wakee_flips >>= 1;
4506 current->wakee_flip_decay_ts = jiffies;
4509 if (current->last_wakee != p) {
4510 current->last_wakee = p;
4511 current->wakee_flips++;
4515 static void task_waking_fair(struct task_struct *p)
4517 struct sched_entity *se = &p->se;
4518 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4521 #ifndef CONFIG_64BIT
4522 u64 min_vruntime_copy;
4525 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4527 min_vruntime = cfs_rq->min_vruntime;
4528 } while (min_vruntime != min_vruntime_copy);
4530 min_vruntime = cfs_rq->min_vruntime;
4533 se->vruntime -= min_vruntime;
4537 #ifdef CONFIG_FAIR_GROUP_SCHED
4539 * effective_load() calculates the load change as seen from the root_task_group
4541 * Adding load to a group doesn't make a group heavier, but can cause movement
4542 * of group shares between cpus. Assuming the shares were perfectly aligned one
4543 * can calculate the shift in shares.
4545 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4546 * on this @cpu and results in a total addition (subtraction) of @wg to the
4547 * total group weight.
4549 * Given a runqueue weight distribution (rw_i) we can compute a shares
4550 * distribution (s_i) using:
4552 * s_i = rw_i / \Sum rw_j (1)
4554 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4555 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4556 * shares distribution (s_i):
4558 * rw_i = { 2, 4, 1, 0 }
4559 * s_i = { 2/7, 4/7, 1/7, 0 }
4561 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4562 * task used to run on and the CPU the waker is running on), we need to
4563 * compute the effect of waking a task on either CPU and, in case of a sync
4564 * wakeup, compute the effect of the current task going to sleep.
4566 * So for a change of @wl to the local @cpu with an overall group weight change
4567 * of @wl we can compute the new shares distribution (s'_i) using:
4569 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4571 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4572 * differences in waking a task to CPU 0. The additional task changes the
4573 * weight and shares distributions like:
4575 * rw'_i = { 3, 4, 1, 0 }
4576 * s'_i = { 3/8, 4/8, 1/8, 0 }
4578 * We can then compute the difference in effective weight by using:
4580 * dw_i = S * (s'_i - s_i) (3)
4582 * Where 'S' is the group weight as seen by its parent.
4584 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4585 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4586 * 4/7) times the weight of the group.
4588 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4590 struct sched_entity *se = tg->se[cpu];
4592 if (!tg->parent) /* the trivial, non-cgroup case */
4595 for_each_sched_entity(se) {
4596 struct cfs_rq *cfs_rq = se->my_q;
4597 long W, w = cfs_rq_load_avg(cfs_rq);
4602 * W = @wg + \Sum rw_j
4604 W = wg + atomic_long_read(&tg->load_avg);
4606 /* Ensure \Sum rw_j >= rw_i */
4607 W -= cfs_rq->tg_load_avg_contrib;
4616 * wl = S * s'_i; see (2)
4619 wl = (w * (long)tg->shares) / W;
4624 * Per the above, wl is the new se->load.weight value; since
4625 * those are clipped to [MIN_SHARES, ...) do so now. See
4626 * calc_cfs_shares().
4628 if (wl < MIN_SHARES)
4632 * wl = dw_i = S * (s'_i - s_i); see (3)
4634 wl -= se->avg.load_avg;
4637 * Recursively apply this logic to all parent groups to compute
4638 * the final effective load change on the root group. Since
4639 * only the @tg group gets extra weight, all parent groups can
4640 * only redistribute existing shares. @wl is the shift in shares
4641 * resulting from this level per the above.
4650 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4658 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4659 * A waker of many should wake a different task than the one last awakened
4660 * at a frequency roughly N times higher than one of its wakees. In order
4661 * to determine whether we should let the load spread vs consolodating to
4662 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4663 * partner, and a factor of lls_size higher frequency in the other. With
4664 * both conditions met, we can be relatively sure that the relationship is
4665 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4666 * being client/server, worker/dispatcher, interrupt source or whatever is
4667 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4669 static int wake_wide(struct task_struct *p)
4671 unsigned int master = current->wakee_flips;
4672 unsigned int slave = p->wakee_flips;
4673 int factor = this_cpu_read(sd_llc_size);
4676 swap(master, slave);
4677 if (slave < factor || master < slave * factor)
4682 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4684 s64 this_load, load;
4685 s64 this_eff_load, prev_eff_load;
4686 int idx, this_cpu, prev_cpu;
4687 struct task_group *tg;
4688 unsigned long weight;
4692 this_cpu = smp_processor_id();
4693 prev_cpu = task_cpu(p);
4694 load = source_load(prev_cpu, idx);
4695 this_load = target_load(this_cpu, idx);
4698 * If sync wakeup then subtract the (maximum possible)
4699 * effect of the currently running task from the load
4700 * of the current CPU:
4703 tg = task_group(current);
4704 weight = current->se.avg.load_avg;
4706 this_load += effective_load(tg, this_cpu, -weight, -weight);
4707 load += effective_load(tg, prev_cpu, 0, -weight);
4711 weight = p->se.avg.load_avg;
4714 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4715 * due to the sync cause above having dropped this_load to 0, we'll
4716 * always have an imbalance, but there's really nothing you can do
4717 * about that, so that's good too.
4719 * Otherwise check if either cpus are near enough in load to allow this
4720 * task to be woken on this_cpu.
4722 this_eff_load = 100;
4723 this_eff_load *= capacity_of(prev_cpu);
4725 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4726 prev_eff_load *= capacity_of(this_cpu);
4728 if (this_load > 0) {
4729 this_eff_load *= this_load +
4730 effective_load(tg, this_cpu, weight, weight);
4732 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4735 balanced = this_eff_load <= prev_eff_load;
4737 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4742 schedstat_inc(sd, ttwu_move_affine);
4743 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4749 * find_idlest_group finds and returns the least busy CPU group within the
4752 static struct sched_group *
4753 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4754 int this_cpu, int sd_flag)
4756 struct sched_group *idlest = NULL, *group = sd->groups;
4757 unsigned long min_load = ULONG_MAX, this_load = 0;
4758 int load_idx = sd->forkexec_idx;
4759 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4761 if (sd_flag & SD_BALANCE_WAKE)
4762 load_idx = sd->wake_idx;
4765 unsigned long load, avg_load;
4769 /* Skip over this group if it has no CPUs allowed */
4770 if (!cpumask_intersects(sched_group_cpus(group),
4771 tsk_cpus_allowed(p)))
4774 local_group = cpumask_test_cpu(this_cpu,
4775 sched_group_cpus(group));
4777 /* Tally up the load of all CPUs in the group */
4780 for_each_cpu(i, sched_group_cpus(group)) {
4781 /* Bias balancing toward cpus of our domain */
4783 load = source_load(i, load_idx);
4785 load = target_load(i, load_idx);
4790 /* Adjust by relative CPU capacity of the group */
4791 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4794 this_load = avg_load;
4795 } else if (avg_load < min_load) {
4796 min_load = avg_load;
4799 } while (group = group->next, group != sd->groups);
4801 if (!idlest || 100*this_load < imbalance*min_load)
4807 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4810 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4812 unsigned long load, min_load = ULONG_MAX;
4813 unsigned int min_exit_latency = UINT_MAX;
4814 u64 latest_idle_timestamp = 0;
4815 int least_loaded_cpu = this_cpu;
4816 int shallowest_idle_cpu = -1;
4819 /* Traverse only the allowed CPUs */
4820 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4822 struct rq *rq = cpu_rq(i);
4823 struct cpuidle_state *idle = idle_get_state(rq);
4824 if (idle && idle->exit_latency < min_exit_latency) {
4826 * We give priority to a CPU whose idle state
4827 * has the smallest exit latency irrespective
4828 * of any idle timestamp.
4830 min_exit_latency = idle->exit_latency;
4831 latest_idle_timestamp = rq->idle_stamp;
4832 shallowest_idle_cpu = i;
4833 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4834 rq->idle_stamp > latest_idle_timestamp) {
4836 * If equal or no active idle state, then
4837 * the most recently idled CPU might have
4840 latest_idle_timestamp = rq->idle_stamp;
4841 shallowest_idle_cpu = i;
4843 } else if (shallowest_idle_cpu == -1) {
4844 load = weighted_cpuload(i);
4845 if (load < min_load || (load == min_load && i == this_cpu)) {
4847 least_loaded_cpu = i;
4852 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4856 * Try and locate an idle CPU in the sched_domain.
4858 static int select_idle_sibling(struct task_struct *p, int target)
4860 struct sched_domain *sd;
4861 struct sched_group *sg;
4862 int i = task_cpu(p);
4864 if (idle_cpu(target))
4868 * If the prevous cpu is cache affine and idle, don't be stupid.
4870 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4874 * Otherwise, iterate the domains and find an elegible idle cpu.
4876 sd = rcu_dereference(per_cpu(sd_llc, target));
4877 for_each_lower_domain(sd) {
4880 if (!cpumask_intersects(sched_group_cpus(sg),
4881 tsk_cpus_allowed(p)))
4884 for_each_cpu(i, sched_group_cpus(sg)) {
4885 if (i == target || !idle_cpu(i))
4889 target = cpumask_first_and(sched_group_cpus(sg),
4890 tsk_cpus_allowed(p));
4894 } while (sg != sd->groups);
4901 * cpu_util returns the amount of capacity of a CPU that is used by CFS
4902 * tasks. The unit of the return value must be the one of capacity so we can
4903 * compare the utilization with the capacity of the CPU that is available for
4904 * CFS task (ie cpu_capacity).
4906 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
4907 * recent utilization of currently non-runnable tasks on a CPU. It represents
4908 * the amount of utilization of a CPU in the range [0..capacity_orig] where
4909 * capacity_orig is the cpu_capacity available at the highest frequency
4910 * (arch_scale_freq_capacity()).
4911 * The utilization of a CPU converges towards a sum equal to or less than the
4912 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
4913 * the running time on this CPU scaled by capacity_curr.
4915 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
4916 * higher than capacity_orig because of unfortunate rounding in
4917 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
4918 * the average stabilizes with the new running time. We need to check that the
4919 * utilization stays within the range of [0..capacity_orig] and cap it if
4920 * necessary. Without utilization capping, a group could be seen as overloaded
4921 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
4922 * available capacity. We allow utilization to overshoot capacity_curr (but not
4923 * capacity_orig) as it useful for predicting the capacity required after task
4924 * migrations (scheduler-driven DVFS).
4926 static int cpu_util(int cpu)
4928 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
4929 unsigned long capacity = capacity_orig_of(cpu);
4931 return (util >= capacity) ? capacity : util;
4935 * select_task_rq_fair: Select target runqueue for the waking task in domains
4936 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4937 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4939 * Balances load by selecting the idlest cpu in the idlest group, or under
4940 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4942 * Returns the target cpu number.
4944 * preempt must be disabled.
4947 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4949 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4950 int cpu = smp_processor_id();
4951 int new_cpu = prev_cpu;
4952 int want_affine = 0;
4953 int sync = wake_flags & WF_SYNC;
4955 if (sd_flag & SD_BALANCE_WAKE)
4956 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4959 for_each_domain(cpu, tmp) {
4960 if (!(tmp->flags & SD_LOAD_BALANCE))
4964 * If both cpu and prev_cpu are part of this domain,
4965 * cpu is a valid SD_WAKE_AFFINE target.
4967 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4968 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4973 if (tmp->flags & sd_flag)
4975 else if (!want_affine)
4980 sd = NULL; /* Prefer wake_affine over balance flags */
4981 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4986 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
4987 new_cpu = select_idle_sibling(p, new_cpu);
4990 struct sched_group *group;
4993 if (!(sd->flags & sd_flag)) {
4998 group = find_idlest_group(sd, p, cpu, sd_flag);
5004 new_cpu = find_idlest_cpu(group, p, cpu);
5005 if (new_cpu == -1 || new_cpu == cpu) {
5006 /* Now try balancing at a lower domain level of cpu */
5011 /* Now try balancing at a lower domain level of new_cpu */
5013 weight = sd->span_weight;
5015 for_each_domain(cpu, tmp) {
5016 if (weight <= tmp->span_weight)
5018 if (tmp->flags & sd_flag)
5021 /* while loop will break here if sd == NULL */
5029 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5030 * cfs_rq_of(p) references at time of call are still valid and identify the
5031 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5032 * other assumptions, including the state of rq->lock, should be made.
5034 static void migrate_task_rq_fair(struct task_struct *p)
5037 * We are supposed to update the task to "current" time, then its up to date
5038 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5039 * what current time is, so simply throw away the out-of-date time. This
5040 * will result in the wakee task is less decayed, but giving the wakee more
5041 * load sounds not bad.
5043 remove_entity_load_avg(&p->se);
5045 /* Tell new CPU we are migrated */
5046 p->se.avg.last_update_time = 0;
5048 /* We have migrated, no longer consider this task hot */
5049 p->se.exec_start = 0;
5052 static void task_dead_fair(struct task_struct *p)
5054 remove_entity_load_avg(&p->se);
5056 #endif /* CONFIG_SMP */
5058 static unsigned long
5059 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5061 unsigned long gran = sysctl_sched_wakeup_granularity;
5064 * Since its curr running now, convert the gran from real-time
5065 * to virtual-time in his units.
5067 * By using 'se' instead of 'curr' we penalize light tasks, so
5068 * they get preempted easier. That is, if 'se' < 'curr' then
5069 * the resulting gran will be larger, therefore penalizing the
5070 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5071 * be smaller, again penalizing the lighter task.
5073 * This is especially important for buddies when the leftmost
5074 * task is higher priority than the buddy.
5076 return calc_delta_fair(gran, se);
5080 * Should 'se' preempt 'curr'.
5094 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5096 s64 gran, vdiff = curr->vruntime - se->vruntime;
5101 gran = wakeup_gran(curr, se);
5108 static void set_last_buddy(struct sched_entity *se)
5110 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5113 for_each_sched_entity(se)
5114 cfs_rq_of(se)->last = se;
5117 static void set_next_buddy(struct sched_entity *se)
5119 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5122 for_each_sched_entity(se)
5123 cfs_rq_of(se)->next = se;
5126 static void set_skip_buddy(struct sched_entity *se)
5128 for_each_sched_entity(se)
5129 cfs_rq_of(se)->skip = se;
5133 * Preempt the current task with a newly woken task if needed:
5135 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5137 struct task_struct *curr = rq->curr;
5138 struct sched_entity *se = &curr->se, *pse = &p->se;
5139 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5140 int scale = cfs_rq->nr_running >= sched_nr_latency;
5141 int next_buddy_marked = 0;
5143 if (unlikely(se == pse))
5147 * This is possible from callers such as attach_tasks(), in which we
5148 * unconditionally check_prempt_curr() after an enqueue (which may have
5149 * lead to a throttle). This both saves work and prevents false
5150 * next-buddy nomination below.
5152 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5155 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5156 set_next_buddy(pse);
5157 next_buddy_marked = 1;
5161 * We can come here with TIF_NEED_RESCHED already set from new task
5164 * Note: this also catches the edge-case of curr being in a throttled
5165 * group (e.g. via set_curr_task), since update_curr() (in the
5166 * enqueue of curr) will have resulted in resched being set. This
5167 * prevents us from potentially nominating it as a false LAST_BUDDY
5170 if (test_tsk_need_resched(curr))
5173 /* Idle tasks are by definition preempted by non-idle tasks. */
5174 if (unlikely(curr->policy == SCHED_IDLE) &&
5175 likely(p->policy != SCHED_IDLE))
5179 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5180 * is driven by the tick):
5182 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5185 find_matching_se(&se, &pse);
5186 update_curr(cfs_rq_of(se));
5188 if (wakeup_preempt_entity(se, pse) == 1) {
5190 * Bias pick_next to pick the sched entity that is
5191 * triggering this preemption.
5193 if (!next_buddy_marked)
5194 set_next_buddy(pse);
5203 * Only set the backward buddy when the current task is still
5204 * on the rq. This can happen when a wakeup gets interleaved
5205 * with schedule on the ->pre_schedule() or idle_balance()
5206 * point, either of which can * drop the rq lock.
5208 * Also, during early boot the idle thread is in the fair class,
5209 * for obvious reasons its a bad idea to schedule back to it.
5211 if (unlikely(!se->on_rq || curr == rq->idle))
5214 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5218 static struct task_struct *
5219 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5221 struct cfs_rq *cfs_rq = &rq->cfs;
5222 struct sched_entity *se;
5223 struct task_struct *p;
5227 #ifdef CONFIG_FAIR_GROUP_SCHED
5228 if (!cfs_rq->nr_running)
5231 if (prev->sched_class != &fair_sched_class)
5235 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5236 * likely that a next task is from the same cgroup as the current.
5238 * Therefore attempt to avoid putting and setting the entire cgroup
5239 * hierarchy, only change the part that actually changes.
5243 struct sched_entity *curr = cfs_rq->curr;
5246 * Since we got here without doing put_prev_entity() we also
5247 * have to consider cfs_rq->curr. If it is still a runnable
5248 * entity, update_curr() will update its vruntime, otherwise
5249 * forget we've ever seen it.
5253 update_curr(cfs_rq);
5258 * This call to check_cfs_rq_runtime() will do the
5259 * throttle and dequeue its entity in the parent(s).
5260 * Therefore the 'simple' nr_running test will indeed
5263 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5267 se = pick_next_entity(cfs_rq, curr);
5268 cfs_rq = group_cfs_rq(se);
5274 * Since we haven't yet done put_prev_entity and if the selected task
5275 * is a different task than we started out with, try and touch the
5276 * least amount of cfs_rqs.
5279 struct sched_entity *pse = &prev->se;
5281 while (!(cfs_rq = is_same_group(se, pse))) {
5282 int se_depth = se->depth;
5283 int pse_depth = pse->depth;
5285 if (se_depth <= pse_depth) {
5286 put_prev_entity(cfs_rq_of(pse), pse);
5287 pse = parent_entity(pse);
5289 if (se_depth >= pse_depth) {
5290 set_next_entity(cfs_rq_of(se), se);
5291 se = parent_entity(se);
5295 put_prev_entity(cfs_rq, pse);
5296 set_next_entity(cfs_rq, se);
5299 if (hrtick_enabled(rq))
5300 hrtick_start_fair(rq, p);
5307 if (!cfs_rq->nr_running)
5310 put_prev_task(rq, prev);
5313 se = pick_next_entity(cfs_rq, NULL);
5314 set_next_entity(cfs_rq, se);
5315 cfs_rq = group_cfs_rq(se);
5320 if (hrtick_enabled(rq))
5321 hrtick_start_fair(rq, p);
5327 * This is OK, because current is on_cpu, which avoids it being picked
5328 * for load-balance and preemption/IRQs are still disabled avoiding
5329 * further scheduler activity on it and we're being very careful to
5330 * re-start the picking loop.
5332 lockdep_unpin_lock(&rq->lock);
5333 new_tasks = idle_balance(rq);
5334 lockdep_pin_lock(&rq->lock);
5336 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5337 * possible for any higher priority task to appear. In that case we
5338 * must re-start the pick_next_entity() loop.
5350 * Account for a descheduled task:
5352 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5354 struct sched_entity *se = &prev->se;
5355 struct cfs_rq *cfs_rq;
5357 for_each_sched_entity(se) {
5358 cfs_rq = cfs_rq_of(se);
5359 put_prev_entity(cfs_rq, se);
5364 * sched_yield() is very simple
5366 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5368 static void yield_task_fair(struct rq *rq)
5370 struct task_struct *curr = rq->curr;
5371 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5372 struct sched_entity *se = &curr->se;
5375 * Are we the only task in the tree?
5377 if (unlikely(rq->nr_running == 1))
5380 clear_buddies(cfs_rq, se);
5382 if (curr->policy != SCHED_BATCH) {
5383 update_rq_clock(rq);
5385 * Update run-time statistics of the 'current'.
5387 update_curr(cfs_rq);
5389 * Tell update_rq_clock() that we've just updated,
5390 * so we don't do microscopic update in schedule()
5391 * and double the fastpath cost.
5393 rq_clock_skip_update(rq, true);
5399 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5401 struct sched_entity *se = &p->se;
5403 /* throttled hierarchies are not runnable */
5404 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5407 /* Tell the scheduler that we'd really like pse to run next. */
5410 yield_task_fair(rq);
5416 /**************************************************
5417 * Fair scheduling class load-balancing methods.
5421 * The purpose of load-balancing is to achieve the same basic fairness the
5422 * per-cpu scheduler provides, namely provide a proportional amount of compute
5423 * time to each task. This is expressed in the following equation:
5425 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5427 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5428 * W_i,0 is defined as:
5430 * W_i,0 = \Sum_j w_i,j (2)
5432 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5433 * is derived from the nice value as per prio_to_weight[].
5435 * The weight average is an exponential decay average of the instantaneous
5438 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5440 * C_i is the compute capacity of cpu i, typically it is the
5441 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5442 * can also include other factors [XXX].
5444 * To achieve this balance we define a measure of imbalance which follows
5445 * directly from (1):
5447 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5449 * We them move tasks around to minimize the imbalance. In the continuous
5450 * function space it is obvious this converges, in the discrete case we get
5451 * a few fun cases generally called infeasible weight scenarios.
5454 * - infeasible weights;
5455 * - local vs global optima in the discrete case. ]
5460 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5461 * for all i,j solution, we create a tree of cpus that follows the hardware
5462 * topology where each level pairs two lower groups (or better). This results
5463 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5464 * tree to only the first of the previous level and we decrease the frequency
5465 * of load-balance at each level inv. proportional to the number of cpus in
5471 * \Sum { --- * --- * 2^i } = O(n) (5)
5473 * `- size of each group
5474 * | | `- number of cpus doing load-balance
5476 * `- sum over all levels
5478 * Coupled with a limit on how many tasks we can migrate every balance pass,
5479 * this makes (5) the runtime complexity of the balancer.
5481 * An important property here is that each CPU is still (indirectly) connected
5482 * to every other cpu in at most O(log n) steps:
5484 * The adjacency matrix of the resulting graph is given by:
5487 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5490 * And you'll find that:
5492 * A^(log_2 n)_i,j != 0 for all i,j (7)
5494 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5495 * The task movement gives a factor of O(m), giving a convergence complexity
5498 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5503 * In order to avoid CPUs going idle while there's still work to do, new idle
5504 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5505 * tree itself instead of relying on other CPUs to bring it work.
5507 * This adds some complexity to both (5) and (8) but it reduces the total idle
5515 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5518 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5523 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5525 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5527 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5530 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5531 * rewrite all of this once again.]
5534 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5536 enum fbq_type { regular, remote, all };
5538 #define LBF_ALL_PINNED 0x01
5539 #define LBF_NEED_BREAK 0x02
5540 #define LBF_DST_PINNED 0x04
5541 #define LBF_SOME_PINNED 0x08
5544 struct sched_domain *sd;
5552 struct cpumask *dst_grpmask;
5554 enum cpu_idle_type idle;
5556 /* The set of CPUs under consideration for load-balancing */
5557 struct cpumask *cpus;
5562 unsigned int loop_break;
5563 unsigned int loop_max;
5565 enum fbq_type fbq_type;
5566 struct list_head tasks;
5570 * Is this task likely cache-hot:
5572 static int task_hot(struct task_struct *p, struct lb_env *env)
5576 lockdep_assert_held(&env->src_rq->lock);
5578 if (p->sched_class != &fair_sched_class)
5581 if (unlikely(p->policy == SCHED_IDLE))
5585 * Buddy candidates are cache hot:
5587 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5588 (&p->se == cfs_rq_of(&p->se)->next ||
5589 &p->se == cfs_rq_of(&p->se)->last))
5592 if (sysctl_sched_migration_cost == -1)
5594 if (sysctl_sched_migration_cost == 0)
5597 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5599 return delta < (s64)sysctl_sched_migration_cost;
5602 #ifdef CONFIG_NUMA_BALANCING
5604 * Returns 1, if task migration degrades locality
5605 * Returns 0, if task migration improves locality i.e migration preferred.
5606 * Returns -1, if task migration is not affected by locality.
5608 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5610 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5611 unsigned long src_faults, dst_faults;
5612 int src_nid, dst_nid;
5614 if (!static_branch_likely(&sched_numa_balancing))
5617 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5620 src_nid = cpu_to_node(env->src_cpu);
5621 dst_nid = cpu_to_node(env->dst_cpu);
5623 if (src_nid == dst_nid)
5626 /* Migrating away from the preferred node is always bad. */
5627 if (src_nid == p->numa_preferred_nid) {
5628 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5634 /* Encourage migration to the preferred node. */
5635 if (dst_nid == p->numa_preferred_nid)
5639 src_faults = group_faults(p, src_nid);
5640 dst_faults = group_faults(p, dst_nid);
5642 src_faults = task_faults(p, src_nid);
5643 dst_faults = task_faults(p, dst_nid);
5646 return dst_faults < src_faults;
5650 static inline int migrate_degrades_locality(struct task_struct *p,
5658 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5661 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5665 lockdep_assert_held(&env->src_rq->lock);
5668 * We do not migrate tasks that are:
5669 * 1) throttled_lb_pair, or
5670 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5671 * 3) running (obviously), or
5672 * 4) are cache-hot on their current CPU.
5674 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5677 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5680 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5682 env->flags |= LBF_SOME_PINNED;
5685 * Remember if this task can be migrated to any other cpu in
5686 * our sched_group. We may want to revisit it if we couldn't
5687 * meet load balance goals by pulling other tasks on src_cpu.
5689 * Also avoid computing new_dst_cpu if we have already computed
5690 * one in current iteration.
5692 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5695 /* Prevent to re-select dst_cpu via env's cpus */
5696 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5697 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5698 env->flags |= LBF_DST_PINNED;
5699 env->new_dst_cpu = cpu;
5707 /* Record that we found atleast one task that could run on dst_cpu */
5708 env->flags &= ~LBF_ALL_PINNED;
5710 if (task_running(env->src_rq, p)) {
5711 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5716 * Aggressive migration if:
5717 * 1) destination numa is preferred
5718 * 2) task is cache cold, or
5719 * 3) too many balance attempts have failed.
5721 tsk_cache_hot = migrate_degrades_locality(p, env);
5722 if (tsk_cache_hot == -1)
5723 tsk_cache_hot = task_hot(p, env);
5725 if (tsk_cache_hot <= 0 ||
5726 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5727 if (tsk_cache_hot == 1) {
5728 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5729 schedstat_inc(p, se.statistics.nr_forced_migrations);
5734 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5739 * detach_task() -- detach the task for the migration specified in env
5741 static void detach_task(struct task_struct *p, struct lb_env *env)
5743 lockdep_assert_held(&env->src_rq->lock);
5745 deactivate_task(env->src_rq, p, 0);
5746 p->on_rq = TASK_ON_RQ_MIGRATING;
5747 set_task_cpu(p, env->dst_cpu);
5751 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5752 * part of active balancing operations within "domain".
5754 * Returns a task if successful and NULL otherwise.
5756 static struct task_struct *detach_one_task(struct lb_env *env)
5758 struct task_struct *p, *n;
5760 lockdep_assert_held(&env->src_rq->lock);
5762 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5763 if (!can_migrate_task(p, env))
5766 detach_task(p, env);
5769 * Right now, this is only the second place where
5770 * lb_gained[env->idle] is updated (other is detach_tasks)
5771 * so we can safely collect stats here rather than
5772 * inside detach_tasks().
5774 schedstat_inc(env->sd, lb_gained[env->idle]);
5780 static const unsigned int sched_nr_migrate_break = 32;
5783 * detach_tasks() -- tries to detach up to imbalance weighted load from
5784 * busiest_rq, as part of a balancing operation within domain "sd".
5786 * Returns number of detached tasks if successful and 0 otherwise.
5788 static int detach_tasks(struct lb_env *env)
5790 struct list_head *tasks = &env->src_rq->cfs_tasks;
5791 struct task_struct *p;
5795 lockdep_assert_held(&env->src_rq->lock);
5797 if (env->imbalance <= 0)
5800 while (!list_empty(tasks)) {
5802 * We don't want to steal all, otherwise we may be treated likewise,
5803 * which could at worst lead to a livelock crash.
5805 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5808 p = list_first_entry(tasks, struct task_struct, se.group_node);
5811 /* We've more or less seen every task there is, call it quits */
5812 if (env->loop > env->loop_max)
5815 /* take a breather every nr_migrate tasks */
5816 if (env->loop > env->loop_break) {
5817 env->loop_break += sched_nr_migrate_break;
5818 env->flags |= LBF_NEED_BREAK;
5822 if (!can_migrate_task(p, env))
5825 load = task_h_load(p);
5827 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5830 if ((load / 2) > env->imbalance)
5833 detach_task(p, env);
5834 list_add(&p->se.group_node, &env->tasks);
5837 env->imbalance -= load;
5839 #ifdef CONFIG_PREEMPT
5841 * NEWIDLE balancing is a source of latency, so preemptible
5842 * kernels will stop after the first task is detached to minimize
5843 * the critical section.
5845 if (env->idle == CPU_NEWLY_IDLE)
5850 * We only want to steal up to the prescribed amount of
5853 if (env->imbalance <= 0)
5858 list_move_tail(&p->se.group_node, tasks);
5862 * Right now, this is one of only two places we collect this stat
5863 * so we can safely collect detach_one_task() stats here rather
5864 * than inside detach_one_task().
5866 schedstat_add(env->sd, lb_gained[env->idle], detached);
5872 * attach_task() -- attach the task detached by detach_task() to its new rq.
5874 static void attach_task(struct rq *rq, struct task_struct *p)
5876 lockdep_assert_held(&rq->lock);
5878 BUG_ON(task_rq(p) != rq);
5879 p->on_rq = TASK_ON_RQ_QUEUED;
5880 activate_task(rq, p, 0);
5881 check_preempt_curr(rq, p, 0);
5885 * attach_one_task() -- attaches the task returned from detach_one_task() to
5888 static void attach_one_task(struct rq *rq, struct task_struct *p)
5890 raw_spin_lock(&rq->lock);
5892 raw_spin_unlock(&rq->lock);
5896 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5899 static void attach_tasks(struct lb_env *env)
5901 struct list_head *tasks = &env->tasks;
5902 struct task_struct *p;
5904 raw_spin_lock(&env->dst_rq->lock);
5906 while (!list_empty(tasks)) {
5907 p = list_first_entry(tasks, struct task_struct, se.group_node);
5908 list_del_init(&p->se.group_node);
5910 attach_task(env->dst_rq, p);
5913 raw_spin_unlock(&env->dst_rq->lock);
5916 #ifdef CONFIG_FAIR_GROUP_SCHED
5917 static void update_blocked_averages(int cpu)
5919 struct rq *rq = cpu_rq(cpu);
5920 struct cfs_rq *cfs_rq;
5921 unsigned long flags;
5923 raw_spin_lock_irqsave(&rq->lock, flags);
5924 update_rq_clock(rq);
5927 * Iterates the task_group tree in a bottom up fashion, see
5928 * list_add_leaf_cfs_rq() for details.
5930 for_each_leaf_cfs_rq(rq, cfs_rq) {
5931 /* throttled entities do not contribute to load */
5932 if (throttled_hierarchy(cfs_rq))
5935 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
5936 update_tg_load_avg(cfs_rq, 0);
5938 raw_spin_unlock_irqrestore(&rq->lock, flags);
5942 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5943 * This needs to be done in a top-down fashion because the load of a child
5944 * group is a fraction of its parents load.
5946 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5948 struct rq *rq = rq_of(cfs_rq);
5949 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5950 unsigned long now = jiffies;
5953 if (cfs_rq->last_h_load_update == now)
5956 cfs_rq->h_load_next = NULL;
5957 for_each_sched_entity(se) {
5958 cfs_rq = cfs_rq_of(se);
5959 cfs_rq->h_load_next = se;
5960 if (cfs_rq->last_h_load_update == now)
5965 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
5966 cfs_rq->last_h_load_update = now;
5969 while ((se = cfs_rq->h_load_next) != NULL) {
5970 load = cfs_rq->h_load;
5971 load = div64_ul(load * se->avg.load_avg,
5972 cfs_rq_load_avg(cfs_rq) + 1);
5973 cfs_rq = group_cfs_rq(se);
5974 cfs_rq->h_load = load;
5975 cfs_rq->last_h_load_update = now;
5979 static unsigned long task_h_load(struct task_struct *p)
5981 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5983 update_cfs_rq_h_load(cfs_rq);
5984 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
5985 cfs_rq_load_avg(cfs_rq) + 1);
5988 static inline void update_blocked_averages(int cpu)
5990 struct rq *rq = cpu_rq(cpu);
5991 struct cfs_rq *cfs_rq = &rq->cfs;
5992 unsigned long flags;
5994 raw_spin_lock_irqsave(&rq->lock, flags);
5995 update_rq_clock(rq);
5996 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
5997 raw_spin_unlock_irqrestore(&rq->lock, flags);
6000 static unsigned long task_h_load(struct task_struct *p)
6002 return p->se.avg.load_avg;
6006 /********** Helpers for find_busiest_group ************************/
6015 * sg_lb_stats - stats of a sched_group required for load_balancing
6017 struct sg_lb_stats {
6018 unsigned long avg_load; /*Avg load across the CPUs of the group */
6019 unsigned long group_load; /* Total load over the CPUs of the group */
6020 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6021 unsigned long load_per_task;
6022 unsigned long group_capacity;
6023 unsigned long group_util; /* Total utilization of the group */
6024 unsigned int sum_nr_running; /* Nr tasks running in the group */
6025 unsigned int idle_cpus;
6026 unsigned int group_weight;
6027 enum group_type group_type;
6028 int group_no_capacity;
6029 #ifdef CONFIG_NUMA_BALANCING
6030 unsigned int nr_numa_running;
6031 unsigned int nr_preferred_running;
6036 * sd_lb_stats - Structure to store the statistics of a sched_domain
6037 * during load balancing.
6039 struct sd_lb_stats {
6040 struct sched_group *busiest; /* Busiest group in this sd */
6041 struct sched_group *local; /* Local group in this sd */
6042 unsigned long total_load; /* Total load of all groups in sd */
6043 unsigned long total_capacity; /* Total capacity of all groups in sd */
6044 unsigned long avg_load; /* Average load across all groups in sd */
6046 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6047 struct sg_lb_stats local_stat; /* Statistics of the local group */
6050 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6053 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6054 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6055 * We must however clear busiest_stat::avg_load because
6056 * update_sd_pick_busiest() reads this before assignment.
6058 *sds = (struct sd_lb_stats){
6062 .total_capacity = 0UL,
6065 .sum_nr_running = 0,
6066 .group_type = group_other,
6072 * get_sd_load_idx - Obtain the load index for a given sched domain.
6073 * @sd: The sched_domain whose load_idx is to be obtained.
6074 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6076 * Return: The load index.
6078 static inline int get_sd_load_idx(struct sched_domain *sd,
6079 enum cpu_idle_type idle)
6085 load_idx = sd->busy_idx;
6088 case CPU_NEWLY_IDLE:
6089 load_idx = sd->newidle_idx;
6092 load_idx = sd->idle_idx;
6099 static unsigned long scale_rt_capacity(int cpu)
6101 struct rq *rq = cpu_rq(cpu);
6102 u64 total, used, age_stamp, avg;
6106 * Since we're reading these variables without serialization make sure
6107 * we read them once before doing sanity checks on them.
6109 age_stamp = READ_ONCE(rq->age_stamp);
6110 avg = READ_ONCE(rq->rt_avg);
6111 delta = __rq_clock_broken(rq) - age_stamp;
6113 if (unlikely(delta < 0))
6116 total = sched_avg_period() + delta;
6118 used = div_u64(avg, total);
6120 if (likely(used < SCHED_CAPACITY_SCALE))
6121 return SCHED_CAPACITY_SCALE - used;
6126 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6128 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6129 struct sched_group *sdg = sd->groups;
6131 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6133 capacity *= scale_rt_capacity(cpu);
6134 capacity >>= SCHED_CAPACITY_SHIFT;
6139 cpu_rq(cpu)->cpu_capacity = capacity;
6140 sdg->sgc->capacity = capacity;
6143 void update_group_capacity(struct sched_domain *sd, int cpu)
6145 struct sched_domain *child = sd->child;
6146 struct sched_group *group, *sdg = sd->groups;
6147 unsigned long capacity;
6148 unsigned long interval;
6150 interval = msecs_to_jiffies(sd->balance_interval);
6151 interval = clamp(interval, 1UL, max_load_balance_interval);
6152 sdg->sgc->next_update = jiffies + interval;
6155 update_cpu_capacity(sd, cpu);
6161 if (child->flags & SD_OVERLAP) {
6163 * SD_OVERLAP domains cannot assume that child groups
6164 * span the current group.
6167 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6168 struct sched_group_capacity *sgc;
6169 struct rq *rq = cpu_rq(cpu);
6172 * build_sched_domains() -> init_sched_groups_capacity()
6173 * gets here before we've attached the domains to the
6176 * Use capacity_of(), which is set irrespective of domains
6177 * in update_cpu_capacity().
6179 * This avoids capacity from being 0 and
6180 * causing divide-by-zero issues on boot.
6182 if (unlikely(!rq->sd)) {
6183 capacity += capacity_of(cpu);
6187 sgc = rq->sd->groups->sgc;
6188 capacity += sgc->capacity;
6192 * !SD_OVERLAP domains can assume that child groups
6193 * span the current group.
6196 group = child->groups;
6198 capacity += group->sgc->capacity;
6199 group = group->next;
6200 } while (group != child->groups);
6203 sdg->sgc->capacity = capacity;
6207 * Check whether the capacity of the rq has been noticeably reduced by side
6208 * activity. The imbalance_pct is used for the threshold.
6209 * Return true is the capacity is reduced
6212 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6214 return ((rq->cpu_capacity * sd->imbalance_pct) <
6215 (rq->cpu_capacity_orig * 100));
6219 * Group imbalance indicates (and tries to solve) the problem where balancing
6220 * groups is inadequate due to tsk_cpus_allowed() constraints.
6222 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6223 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6226 * { 0 1 2 3 } { 4 5 6 7 }
6229 * If we were to balance group-wise we'd place two tasks in the first group and
6230 * two tasks in the second group. Clearly this is undesired as it will overload
6231 * cpu 3 and leave one of the cpus in the second group unused.
6233 * The current solution to this issue is detecting the skew in the first group
6234 * by noticing the lower domain failed to reach balance and had difficulty
6235 * moving tasks due to affinity constraints.
6237 * When this is so detected; this group becomes a candidate for busiest; see
6238 * update_sd_pick_busiest(). And calculate_imbalance() and
6239 * find_busiest_group() avoid some of the usual balance conditions to allow it
6240 * to create an effective group imbalance.
6242 * This is a somewhat tricky proposition since the next run might not find the
6243 * group imbalance and decide the groups need to be balanced again. A most
6244 * subtle and fragile situation.
6247 static inline int sg_imbalanced(struct sched_group *group)
6249 return group->sgc->imbalance;
6253 * group_has_capacity returns true if the group has spare capacity that could
6254 * be used by some tasks.
6255 * We consider that a group has spare capacity if the * number of task is
6256 * smaller than the number of CPUs or if the utilization is lower than the
6257 * available capacity for CFS tasks.
6258 * For the latter, we use a threshold to stabilize the state, to take into
6259 * account the variance of the tasks' load and to return true if the available
6260 * capacity in meaningful for the load balancer.
6261 * As an example, an available capacity of 1% can appear but it doesn't make
6262 * any benefit for the load balance.
6265 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6267 if (sgs->sum_nr_running < sgs->group_weight)
6270 if ((sgs->group_capacity * 100) >
6271 (sgs->group_util * env->sd->imbalance_pct))
6278 * group_is_overloaded returns true if the group has more tasks than it can
6280 * group_is_overloaded is not equals to !group_has_capacity because a group
6281 * with the exact right number of tasks, has no more spare capacity but is not
6282 * overloaded so both group_has_capacity and group_is_overloaded return
6286 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6288 if (sgs->sum_nr_running <= sgs->group_weight)
6291 if ((sgs->group_capacity * 100) <
6292 (sgs->group_util * env->sd->imbalance_pct))
6299 group_type group_classify(struct sched_group *group,
6300 struct sg_lb_stats *sgs)
6302 if (sgs->group_no_capacity)
6303 return group_overloaded;
6305 if (sg_imbalanced(group))
6306 return group_imbalanced;
6312 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6313 * @env: The load balancing environment.
6314 * @group: sched_group whose statistics are to be updated.
6315 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6316 * @local_group: Does group contain this_cpu.
6317 * @sgs: variable to hold the statistics for this group.
6318 * @overload: Indicate more than one runnable task for any CPU.
6320 static inline void update_sg_lb_stats(struct lb_env *env,
6321 struct sched_group *group, int load_idx,
6322 int local_group, struct sg_lb_stats *sgs,
6328 memset(sgs, 0, sizeof(*sgs));
6330 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6331 struct rq *rq = cpu_rq(i);
6333 /* Bias balancing toward cpus of our domain */
6335 load = target_load(i, load_idx);
6337 load = source_load(i, load_idx);
6339 sgs->group_load += load;
6340 sgs->group_util += cpu_util(i);
6341 sgs->sum_nr_running += rq->cfs.h_nr_running;
6343 if (rq->nr_running > 1)
6346 #ifdef CONFIG_NUMA_BALANCING
6347 sgs->nr_numa_running += rq->nr_numa_running;
6348 sgs->nr_preferred_running += rq->nr_preferred_running;
6350 sgs->sum_weighted_load += weighted_cpuload(i);
6355 /* Adjust by relative CPU capacity of the group */
6356 sgs->group_capacity = group->sgc->capacity;
6357 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6359 if (sgs->sum_nr_running)
6360 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6362 sgs->group_weight = group->group_weight;
6364 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6365 sgs->group_type = group_classify(group, sgs);
6369 * update_sd_pick_busiest - return 1 on busiest group
6370 * @env: The load balancing environment.
6371 * @sds: sched_domain statistics
6372 * @sg: sched_group candidate to be checked for being the busiest
6373 * @sgs: sched_group statistics
6375 * Determine if @sg is a busier group than the previously selected
6378 * Return: %true if @sg is a busier group than the previously selected
6379 * busiest group. %false otherwise.
6381 static bool update_sd_pick_busiest(struct lb_env *env,
6382 struct sd_lb_stats *sds,
6383 struct sched_group *sg,
6384 struct sg_lb_stats *sgs)
6386 struct sg_lb_stats *busiest = &sds->busiest_stat;
6388 if (sgs->group_type > busiest->group_type)
6391 if (sgs->group_type < busiest->group_type)
6394 if (sgs->avg_load <= busiest->avg_load)
6397 /* This is the busiest node in its class. */
6398 if (!(env->sd->flags & SD_ASYM_PACKING))
6402 * ASYM_PACKING needs to move all the work to the lowest
6403 * numbered CPUs in the group, therefore mark all groups
6404 * higher than ourself as busy.
6406 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6410 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6417 #ifdef CONFIG_NUMA_BALANCING
6418 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6420 if (sgs->sum_nr_running > sgs->nr_numa_running)
6422 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6427 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6429 if (rq->nr_running > rq->nr_numa_running)
6431 if (rq->nr_running > rq->nr_preferred_running)
6436 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6441 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6445 #endif /* CONFIG_NUMA_BALANCING */
6448 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6449 * @env: The load balancing environment.
6450 * @sds: variable to hold the statistics for this sched_domain.
6452 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6454 struct sched_domain *child = env->sd->child;
6455 struct sched_group *sg = env->sd->groups;
6456 struct sg_lb_stats tmp_sgs;
6457 int load_idx, prefer_sibling = 0;
6458 bool overload = false;
6460 if (child && child->flags & SD_PREFER_SIBLING)
6463 load_idx = get_sd_load_idx(env->sd, env->idle);
6466 struct sg_lb_stats *sgs = &tmp_sgs;
6469 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6472 sgs = &sds->local_stat;
6474 if (env->idle != CPU_NEWLY_IDLE ||
6475 time_after_eq(jiffies, sg->sgc->next_update))
6476 update_group_capacity(env->sd, env->dst_cpu);
6479 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6486 * In case the child domain prefers tasks go to siblings
6487 * first, lower the sg capacity so that we'll try
6488 * and move all the excess tasks away. We lower the capacity
6489 * of a group only if the local group has the capacity to fit
6490 * these excess tasks. The extra check prevents the case where
6491 * you always pull from the heaviest group when it is already
6492 * under-utilized (possible with a large weight task outweighs
6493 * the tasks on the system).
6495 if (prefer_sibling && sds->local &&
6496 group_has_capacity(env, &sds->local_stat) &&
6497 (sgs->sum_nr_running > 1)) {
6498 sgs->group_no_capacity = 1;
6499 sgs->group_type = group_classify(sg, sgs);
6502 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6504 sds->busiest_stat = *sgs;
6508 /* Now, start updating sd_lb_stats */
6509 sds->total_load += sgs->group_load;
6510 sds->total_capacity += sgs->group_capacity;
6513 } while (sg != env->sd->groups);
6515 if (env->sd->flags & SD_NUMA)
6516 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6518 if (!env->sd->parent) {
6519 /* update overload indicator if we are at root domain */
6520 if (env->dst_rq->rd->overload != overload)
6521 env->dst_rq->rd->overload = overload;
6527 * check_asym_packing - Check to see if the group is packed into the
6530 * This is primarily intended to used at the sibling level. Some
6531 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6532 * case of POWER7, it can move to lower SMT modes only when higher
6533 * threads are idle. When in lower SMT modes, the threads will
6534 * perform better since they share less core resources. Hence when we
6535 * have idle threads, we want them to be the higher ones.
6537 * This packing function is run on idle threads. It checks to see if
6538 * the busiest CPU in this domain (core in the P7 case) has a higher
6539 * CPU number than the packing function is being run on. Here we are
6540 * assuming lower CPU number will be equivalent to lower a SMT thread
6543 * Return: 1 when packing is required and a task should be moved to
6544 * this CPU. The amount of the imbalance is returned in *imbalance.
6546 * @env: The load balancing environment.
6547 * @sds: Statistics of the sched_domain which is to be packed
6549 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6553 if (!(env->sd->flags & SD_ASYM_PACKING))
6559 busiest_cpu = group_first_cpu(sds->busiest);
6560 if (env->dst_cpu > busiest_cpu)
6563 env->imbalance = DIV_ROUND_CLOSEST(
6564 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6565 SCHED_CAPACITY_SCALE);
6571 * fix_small_imbalance - Calculate the minor imbalance that exists
6572 * amongst the groups of a sched_domain, during
6574 * @env: The load balancing environment.
6575 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6578 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6580 unsigned long tmp, capa_now = 0, capa_move = 0;
6581 unsigned int imbn = 2;
6582 unsigned long scaled_busy_load_per_task;
6583 struct sg_lb_stats *local, *busiest;
6585 local = &sds->local_stat;
6586 busiest = &sds->busiest_stat;
6588 if (!local->sum_nr_running)
6589 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6590 else if (busiest->load_per_task > local->load_per_task)
6593 scaled_busy_load_per_task =
6594 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6595 busiest->group_capacity;
6597 if (busiest->avg_load + scaled_busy_load_per_task >=
6598 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6599 env->imbalance = busiest->load_per_task;
6604 * OK, we don't have enough imbalance to justify moving tasks,
6605 * however we may be able to increase total CPU capacity used by
6609 capa_now += busiest->group_capacity *
6610 min(busiest->load_per_task, busiest->avg_load);
6611 capa_now += local->group_capacity *
6612 min(local->load_per_task, local->avg_load);
6613 capa_now /= SCHED_CAPACITY_SCALE;
6615 /* Amount of load we'd subtract */
6616 if (busiest->avg_load > scaled_busy_load_per_task) {
6617 capa_move += busiest->group_capacity *
6618 min(busiest->load_per_task,
6619 busiest->avg_load - scaled_busy_load_per_task);
6622 /* Amount of load we'd add */
6623 if (busiest->avg_load * busiest->group_capacity <
6624 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6625 tmp = (busiest->avg_load * busiest->group_capacity) /
6626 local->group_capacity;
6628 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6629 local->group_capacity;
6631 capa_move += local->group_capacity *
6632 min(local->load_per_task, local->avg_load + tmp);
6633 capa_move /= SCHED_CAPACITY_SCALE;
6635 /* Move if we gain throughput */
6636 if (capa_move > capa_now)
6637 env->imbalance = busiest->load_per_task;
6641 * calculate_imbalance - Calculate the amount of imbalance present within the
6642 * groups of a given sched_domain during load balance.
6643 * @env: load balance environment
6644 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6646 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6648 unsigned long max_pull, load_above_capacity = ~0UL;
6649 struct sg_lb_stats *local, *busiest;
6651 local = &sds->local_stat;
6652 busiest = &sds->busiest_stat;
6654 if (busiest->group_type == group_imbalanced) {
6656 * In the group_imb case we cannot rely on group-wide averages
6657 * to ensure cpu-load equilibrium, look at wider averages. XXX
6659 busiest->load_per_task =
6660 min(busiest->load_per_task, sds->avg_load);
6664 * In the presence of smp nice balancing, certain scenarios can have
6665 * max load less than avg load(as we skip the groups at or below
6666 * its cpu_capacity, while calculating max_load..)
6668 if (busiest->avg_load <= sds->avg_load ||
6669 local->avg_load >= sds->avg_load) {
6671 return fix_small_imbalance(env, sds);
6675 * If there aren't any idle cpus, avoid creating some.
6677 if (busiest->group_type == group_overloaded &&
6678 local->group_type == group_overloaded) {
6679 load_above_capacity = busiest->sum_nr_running *
6681 if (load_above_capacity > busiest->group_capacity)
6682 load_above_capacity -= busiest->group_capacity;
6684 load_above_capacity = ~0UL;
6688 * We're trying to get all the cpus to the average_load, so we don't
6689 * want to push ourselves above the average load, nor do we wish to
6690 * reduce the max loaded cpu below the average load. At the same time,
6691 * we also don't want to reduce the group load below the group capacity
6692 * (so that we can implement power-savings policies etc). Thus we look
6693 * for the minimum possible imbalance.
6695 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6697 /* How much load to actually move to equalise the imbalance */
6698 env->imbalance = min(
6699 max_pull * busiest->group_capacity,
6700 (sds->avg_load - local->avg_load) * local->group_capacity
6701 ) / SCHED_CAPACITY_SCALE;
6704 * if *imbalance is less than the average load per runnable task
6705 * there is no guarantee that any tasks will be moved so we'll have
6706 * a think about bumping its value to force at least one task to be
6709 if (env->imbalance < busiest->load_per_task)
6710 return fix_small_imbalance(env, sds);
6713 /******* find_busiest_group() helpers end here *********************/
6716 * find_busiest_group - Returns the busiest group within the sched_domain
6717 * if there is an imbalance. If there isn't an imbalance, and
6718 * the user has opted for power-savings, it returns a group whose
6719 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6720 * such a group exists.
6722 * Also calculates the amount of weighted load which should be moved
6723 * to restore balance.
6725 * @env: The load balancing environment.
6727 * Return: - The busiest group if imbalance exists.
6728 * - If no imbalance and user has opted for power-savings balance,
6729 * return the least loaded group whose CPUs can be
6730 * put to idle by rebalancing its tasks onto our group.
6732 static struct sched_group *find_busiest_group(struct lb_env *env)
6734 struct sg_lb_stats *local, *busiest;
6735 struct sd_lb_stats sds;
6737 init_sd_lb_stats(&sds);
6740 * Compute the various statistics relavent for load balancing at
6743 update_sd_lb_stats(env, &sds);
6744 local = &sds.local_stat;
6745 busiest = &sds.busiest_stat;
6747 /* ASYM feature bypasses nice load balance check */
6748 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6749 check_asym_packing(env, &sds))
6752 /* There is no busy sibling group to pull tasks from */
6753 if (!sds.busiest || busiest->sum_nr_running == 0)
6756 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6757 / sds.total_capacity;
6760 * If the busiest group is imbalanced the below checks don't
6761 * work because they assume all things are equal, which typically
6762 * isn't true due to cpus_allowed constraints and the like.
6764 if (busiest->group_type == group_imbalanced)
6767 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6768 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6769 busiest->group_no_capacity)
6773 * If the local group is busier than the selected busiest group
6774 * don't try and pull any tasks.
6776 if (local->avg_load >= busiest->avg_load)
6780 * Don't pull any tasks if this group is already above the domain
6783 if (local->avg_load >= sds.avg_load)
6786 if (env->idle == CPU_IDLE) {
6788 * This cpu is idle. If the busiest group is not overloaded
6789 * and there is no imbalance between this and busiest group
6790 * wrt idle cpus, it is balanced. The imbalance becomes
6791 * significant if the diff is greater than 1 otherwise we
6792 * might end up to just move the imbalance on another group
6794 if ((busiest->group_type != group_overloaded) &&
6795 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6799 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6800 * imbalance_pct to be conservative.
6802 if (100 * busiest->avg_load <=
6803 env->sd->imbalance_pct * local->avg_load)
6808 /* Looks like there is an imbalance. Compute it */
6809 calculate_imbalance(env, &sds);
6818 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6820 static struct rq *find_busiest_queue(struct lb_env *env,
6821 struct sched_group *group)
6823 struct rq *busiest = NULL, *rq;
6824 unsigned long busiest_load = 0, busiest_capacity = 1;
6827 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6828 unsigned long capacity, wl;
6832 rt = fbq_classify_rq(rq);
6835 * We classify groups/runqueues into three groups:
6836 * - regular: there are !numa tasks
6837 * - remote: there are numa tasks that run on the 'wrong' node
6838 * - all: there is no distinction
6840 * In order to avoid migrating ideally placed numa tasks,
6841 * ignore those when there's better options.
6843 * If we ignore the actual busiest queue to migrate another
6844 * task, the next balance pass can still reduce the busiest
6845 * queue by moving tasks around inside the node.
6847 * If we cannot move enough load due to this classification
6848 * the next pass will adjust the group classification and
6849 * allow migration of more tasks.
6851 * Both cases only affect the total convergence complexity.
6853 if (rt > env->fbq_type)
6856 capacity = capacity_of(i);
6858 wl = weighted_cpuload(i);
6861 * When comparing with imbalance, use weighted_cpuload()
6862 * which is not scaled with the cpu capacity.
6865 if (rq->nr_running == 1 && wl > env->imbalance &&
6866 !check_cpu_capacity(rq, env->sd))
6870 * For the load comparisons with the other cpu's, consider
6871 * the weighted_cpuload() scaled with the cpu capacity, so
6872 * that the load can be moved away from the cpu that is
6873 * potentially running at a lower capacity.
6875 * Thus we're looking for max(wl_i / capacity_i), crosswise
6876 * multiplication to rid ourselves of the division works out
6877 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6878 * our previous maximum.
6880 if (wl * busiest_capacity > busiest_load * capacity) {
6882 busiest_capacity = capacity;
6891 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6892 * so long as it is large enough.
6894 #define MAX_PINNED_INTERVAL 512
6896 /* Working cpumask for load_balance and load_balance_newidle. */
6897 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6899 static int need_active_balance(struct lb_env *env)
6901 struct sched_domain *sd = env->sd;
6903 if (env->idle == CPU_NEWLY_IDLE) {
6906 * ASYM_PACKING needs to force migrate tasks from busy but
6907 * higher numbered CPUs in order to pack all tasks in the
6908 * lowest numbered CPUs.
6910 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6915 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
6916 * It's worth migrating the task if the src_cpu's capacity is reduced
6917 * because of other sched_class or IRQs if more capacity stays
6918 * available on dst_cpu.
6920 if ((env->idle != CPU_NOT_IDLE) &&
6921 (env->src_rq->cfs.h_nr_running == 1)) {
6922 if ((check_cpu_capacity(env->src_rq, sd)) &&
6923 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
6927 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6930 static int active_load_balance_cpu_stop(void *data);
6932 static int should_we_balance(struct lb_env *env)
6934 struct sched_group *sg = env->sd->groups;
6935 struct cpumask *sg_cpus, *sg_mask;
6936 int cpu, balance_cpu = -1;
6939 * In the newly idle case, we will allow all the cpu's
6940 * to do the newly idle load balance.
6942 if (env->idle == CPU_NEWLY_IDLE)
6945 sg_cpus = sched_group_cpus(sg);
6946 sg_mask = sched_group_mask(sg);
6947 /* Try to find first idle cpu */
6948 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6949 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6956 if (balance_cpu == -1)
6957 balance_cpu = group_balance_cpu(sg);
6960 * First idle cpu or the first cpu(busiest) in this sched group
6961 * is eligible for doing load balancing at this and above domains.
6963 return balance_cpu == env->dst_cpu;
6967 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6968 * tasks if there is an imbalance.
6970 static int load_balance(int this_cpu, struct rq *this_rq,
6971 struct sched_domain *sd, enum cpu_idle_type idle,
6972 int *continue_balancing)
6974 int ld_moved, cur_ld_moved, active_balance = 0;
6975 struct sched_domain *sd_parent = sd->parent;
6976 struct sched_group *group;
6978 unsigned long flags;
6979 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
6981 struct lb_env env = {
6983 .dst_cpu = this_cpu,
6985 .dst_grpmask = sched_group_cpus(sd->groups),
6987 .loop_break = sched_nr_migrate_break,
6990 .tasks = LIST_HEAD_INIT(env.tasks),
6994 * For NEWLY_IDLE load_balancing, we don't need to consider
6995 * other cpus in our group
6997 if (idle == CPU_NEWLY_IDLE)
6998 env.dst_grpmask = NULL;
7000 cpumask_copy(cpus, cpu_active_mask);
7002 schedstat_inc(sd, lb_count[idle]);
7005 if (!should_we_balance(&env)) {
7006 *continue_balancing = 0;
7010 group = find_busiest_group(&env);
7012 schedstat_inc(sd, lb_nobusyg[idle]);
7016 busiest = find_busiest_queue(&env, group);
7018 schedstat_inc(sd, lb_nobusyq[idle]);
7022 BUG_ON(busiest == env.dst_rq);
7024 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7026 env.src_cpu = busiest->cpu;
7027 env.src_rq = busiest;
7030 if (busiest->nr_running > 1) {
7032 * Attempt to move tasks. If find_busiest_group has found
7033 * an imbalance but busiest->nr_running <= 1, the group is
7034 * still unbalanced. ld_moved simply stays zero, so it is
7035 * correctly treated as an imbalance.
7037 env.flags |= LBF_ALL_PINNED;
7038 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7041 raw_spin_lock_irqsave(&busiest->lock, flags);
7044 * cur_ld_moved - load moved in current iteration
7045 * ld_moved - cumulative load moved across iterations
7047 cur_ld_moved = detach_tasks(&env);
7050 * We've detached some tasks from busiest_rq. Every
7051 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7052 * unlock busiest->lock, and we are able to be sure
7053 * that nobody can manipulate the tasks in parallel.
7054 * See task_rq_lock() family for the details.
7057 raw_spin_unlock(&busiest->lock);
7061 ld_moved += cur_ld_moved;
7064 local_irq_restore(flags);
7066 if (env.flags & LBF_NEED_BREAK) {
7067 env.flags &= ~LBF_NEED_BREAK;
7072 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7073 * us and move them to an alternate dst_cpu in our sched_group
7074 * where they can run. The upper limit on how many times we
7075 * iterate on same src_cpu is dependent on number of cpus in our
7078 * This changes load balance semantics a bit on who can move
7079 * load to a given_cpu. In addition to the given_cpu itself
7080 * (or a ilb_cpu acting on its behalf where given_cpu is
7081 * nohz-idle), we now have balance_cpu in a position to move
7082 * load to given_cpu. In rare situations, this may cause
7083 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7084 * _independently_ and at _same_ time to move some load to
7085 * given_cpu) causing exceess load to be moved to given_cpu.
7086 * This however should not happen so much in practice and
7087 * moreover subsequent load balance cycles should correct the
7088 * excess load moved.
7090 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7092 /* Prevent to re-select dst_cpu via env's cpus */
7093 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7095 env.dst_rq = cpu_rq(env.new_dst_cpu);
7096 env.dst_cpu = env.new_dst_cpu;
7097 env.flags &= ~LBF_DST_PINNED;
7099 env.loop_break = sched_nr_migrate_break;
7102 * Go back to "more_balance" rather than "redo" since we
7103 * need to continue with same src_cpu.
7109 * We failed to reach balance because of affinity.
7112 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7114 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7115 *group_imbalance = 1;
7118 /* All tasks on this runqueue were pinned by CPU affinity */
7119 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7120 cpumask_clear_cpu(cpu_of(busiest), cpus);
7121 if (!cpumask_empty(cpus)) {
7123 env.loop_break = sched_nr_migrate_break;
7126 goto out_all_pinned;
7131 schedstat_inc(sd, lb_failed[idle]);
7133 * Increment the failure counter only on periodic balance.
7134 * We do not want newidle balance, which can be very
7135 * frequent, pollute the failure counter causing
7136 * excessive cache_hot migrations and active balances.
7138 if (idle != CPU_NEWLY_IDLE)
7139 sd->nr_balance_failed++;
7141 if (need_active_balance(&env)) {
7142 raw_spin_lock_irqsave(&busiest->lock, flags);
7144 /* don't kick the active_load_balance_cpu_stop,
7145 * if the curr task on busiest cpu can't be
7148 if (!cpumask_test_cpu(this_cpu,
7149 tsk_cpus_allowed(busiest->curr))) {
7150 raw_spin_unlock_irqrestore(&busiest->lock,
7152 env.flags |= LBF_ALL_PINNED;
7153 goto out_one_pinned;
7157 * ->active_balance synchronizes accesses to
7158 * ->active_balance_work. Once set, it's cleared
7159 * only after active load balance is finished.
7161 if (!busiest->active_balance) {
7162 busiest->active_balance = 1;
7163 busiest->push_cpu = this_cpu;
7166 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7168 if (active_balance) {
7169 stop_one_cpu_nowait(cpu_of(busiest),
7170 active_load_balance_cpu_stop, busiest,
7171 &busiest->active_balance_work);
7175 * We've kicked active balancing, reset the failure
7178 sd->nr_balance_failed = sd->cache_nice_tries+1;
7181 sd->nr_balance_failed = 0;
7183 if (likely(!active_balance)) {
7184 /* We were unbalanced, so reset the balancing interval */
7185 sd->balance_interval = sd->min_interval;
7188 * If we've begun active balancing, start to back off. This
7189 * case may not be covered by the all_pinned logic if there
7190 * is only 1 task on the busy runqueue (because we don't call
7193 if (sd->balance_interval < sd->max_interval)
7194 sd->balance_interval *= 2;
7201 * We reach balance although we may have faced some affinity
7202 * constraints. Clear the imbalance flag if it was set.
7205 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7207 if (*group_imbalance)
7208 *group_imbalance = 0;
7213 * We reach balance because all tasks are pinned at this level so
7214 * we can't migrate them. Let the imbalance flag set so parent level
7215 * can try to migrate them.
7217 schedstat_inc(sd, lb_balanced[idle]);
7219 sd->nr_balance_failed = 0;
7222 /* tune up the balancing interval */
7223 if (((env.flags & LBF_ALL_PINNED) &&
7224 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7225 (sd->balance_interval < sd->max_interval))
7226 sd->balance_interval *= 2;
7233 static inline unsigned long
7234 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7236 unsigned long interval = sd->balance_interval;
7239 interval *= sd->busy_factor;
7241 /* scale ms to jiffies */
7242 interval = msecs_to_jiffies(interval);
7243 interval = clamp(interval, 1UL, max_load_balance_interval);
7249 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7251 unsigned long interval, next;
7253 interval = get_sd_balance_interval(sd, cpu_busy);
7254 next = sd->last_balance + interval;
7256 if (time_after(*next_balance, next))
7257 *next_balance = next;
7261 * idle_balance is called by schedule() if this_cpu is about to become
7262 * idle. Attempts to pull tasks from other CPUs.
7264 static int idle_balance(struct rq *this_rq)
7266 unsigned long next_balance = jiffies + HZ;
7267 int this_cpu = this_rq->cpu;
7268 struct sched_domain *sd;
7269 int pulled_task = 0;
7272 idle_enter_fair(this_rq);
7275 * We must set idle_stamp _before_ calling idle_balance(), such that we
7276 * measure the duration of idle_balance() as idle time.
7278 this_rq->idle_stamp = rq_clock(this_rq);
7280 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7281 !this_rq->rd->overload) {
7283 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7285 update_next_balance(sd, 0, &next_balance);
7291 raw_spin_unlock(&this_rq->lock);
7293 update_blocked_averages(this_cpu);
7295 for_each_domain(this_cpu, sd) {
7296 int continue_balancing = 1;
7297 u64 t0, domain_cost;
7299 if (!(sd->flags & SD_LOAD_BALANCE))
7302 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7303 update_next_balance(sd, 0, &next_balance);
7307 if (sd->flags & SD_BALANCE_NEWIDLE) {
7308 t0 = sched_clock_cpu(this_cpu);
7310 pulled_task = load_balance(this_cpu, this_rq,
7312 &continue_balancing);
7314 domain_cost = sched_clock_cpu(this_cpu) - t0;
7315 if (domain_cost > sd->max_newidle_lb_cost)
7316 sd->max_newidle_lb_cost = domain_cost;
7318 curr_cost += domain_cost;
7321 update_next_balance(sd, 0, &next_balance);
7324 * Stop searching for tasks to pull if there are
7325 * now runnable tasks on this rq.
7327 if (pulled_task || this_rq->nr_running > 0)
7332 raw_spin_lock(&this_rq->lock);
7334 if (curr_cost > this_rq->max_idle_balance_cost)
7335 this_rq->max_idle_balance_cost = curr_cost;
7338 * While browsing the domains, we released the rq lock, a task could
7339 * have been enqueued in the meantime. Since we're not going idle,
7340 * pretend we pulled a task.
7342 if (this_rq->cfs.h_nr_running && !pulled_task)
7346 /* Move the next balance forward */
7347 if (time_after(this_rq->next_balance, next_balance))
7348 this_rq->next_balance = next_balance;
7350 /* Is there a task of a high priority class? */
7351 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7355 idle_exit_fair(this_rq);
7356 this_rq->idle_stamp = 0;
7363 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7364 * running tasks off the busiest CPU onto idle CPUs. It requires at
7365 * least 1 task to be running on each physical CPU where possible, and
7366 * avoids physical / logical imbalances.
7368 static int active_load_balance_cpu_stop(void *data)
7370 struct rq *busiest_rq = data;
7371 int busiest_cpu = cpu_of(busiest_rq);
7372 int target_cpu = busiest_rq->push_cpu;
7373 struct rq *target_rq = cpu_rq(target_cpu);
7374 struct sched_domain *sd;
7375 struct task_struct *p = NULL;
7377 raw_spin_lock_irq(&busiest_rq->lock);
7379 /* make sure the requested cpu hasn't gone down in the meantime */
7380 if (unlikely(busiest_cpu != smp_processor_id() ||
7381 !busiest_rq->active_balance))
7384 /* Is there any task to move? */
7385 if (busiest_rq->nr_running <= 1)
7389 * This condition is "impossible", if it occurs
7390 * we need to fix it. Originally reported by
7391 * Bjorn Helgaas on a 128-cpu setup.
7393 BUG_ON(busiest_rq == target_rq);
7395 /* Search for an sd spanning us and the target CPU. */
7397 for_each_domain(target_cpu, sd) {
7398 if ((sd->flags & SD_LOAD_BALANCE) &&
7399 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7404 struct lb_env env = {
7406 .dst_cpu = target_cpu,
7407 .dst_rq = target_rq,
7408 .src_cpu = busiest_rq->cpu,
7409 .src_rq = busiest_rq,
7413 schedstat_inc(sd, alb_count);
7415 p = detach_one_task(&env);
7417 schedstat_inc(sd, alb_pushed);
7419 schedstat_inc(sd, alb_failed);
7423 busiest_rq->active_balance = 0;
7424 raw_spin_unlock(&busiest_rq->lock);
7427 attach_one_task(target_rq, p);
7434 static inline int on_null_domain(struct rq *rq)
7436 return unlikely(!rcu_dereference_sched(rq->sd));
7439 #ifdef CONFIG_NO_HZ_COMMON
7441 * idle load balancing details
7442 * - When one of the busy CPUs notice that there may be an idle rebalancing
7443 * needed, they will kick the idle load balancer, which then does idle
7444 * load balancing for all the idle CPUs.
7447 cpumask_var_t idle_cpus_mask;
7449 unsigned long next_balance; /* in jiffy units */
7450 } nohz ____cacheline_aligned;
7452 static inline int find_new_ilb(void)
7454 int ilb = cpumask_first(nohz.idle_cpus_mask);
7456 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7463 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7464 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7465 * CPU (if there is one).
7467 static void nohz_balancer_kick(void)
7471 nohz.next_balance++;
7473 ilb_cpu = find_new_ilb();
7475 if (ilb_cpu >= nr_cpu_ids)
7478 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7481 * Use smp_send_reschedule() instead of resched_cpu().
7482 * This way we generate a sched IPI on the target cpu which
7483 * is idle. And the softirq performing nohz idle load balance
7484 * will be run before returning from the IPI.
7486 smp_send_reschedule(ilb_cpu);
7490 static inline void nohz_balance_exit_idle(int cpu)
7492 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7494 * Completely isolated CPUs don't ever set, so we must test.
7496 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7497 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7498 atomic_dec(&nohz.nr_cpus);
7500 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7504 static inline void set_cpu_sd_state_busy(void)
7506 struct sched_domain *sd;
7507 int cpu = smp_processor_id();
7510 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7512 if (!sd || !sd->nohz_idle)
7516 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7521 void set_cpu_sd_state_idle(void)
7523 struct sched_domain *sd;
7524 int cpu = smp_processor_id();
7527 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7529 if (!sd || sd->nohz_idle)
7533 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7539 * This routine will record that the cpu is going idle with tick stopped.
7540 * This info will be used in performing idle load balancing in the future.
7542 void nohz_balance_enter_idle(int cpu)
7545 * If this cpu is going down, then nothing needs to be done.
7547 if (!cpu_active(cpu))
7550 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7554 * If we're a completely isolated CPU, we don't play.
7556 if (on_null_domain(cpu_rq(cpu)))
7559 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7560 atomic_inc(&nohz.nr_cpus);
7561 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7564 static int sched_ilb_notifier(struct notifier_block *nfb,
7565 unsigned long action, void *hcpu)
7567 switch (action & ~CPU_TASKS_FROZEN) {
7569 nohz_balance_exit_idle(smp_processor_id());
7577 static DEFINE_SPINLOCK(balancing);
7580 * Scale the max load_balance interval with the number of CPUs in the system.
7581 * This trades load-balance latency on larger machines for less cross talk.
7583 void update_max_interval(void)
7585 max_load_balance_interval = HZ*num_online_cpus()/10;
7589 * It checks each scheduling domain to see if it is due to be balanced,
7590 * and initiates a balancing operation if so.
7592 * Balancing parameters are set up in init_sched_domains.
7594 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7596 int continue_balancing = 1;
7598 unsigned long interval;
7599 struct sched_domain *sd;
7600 /* Earliest time when we have to do rebalance again */
7601 unsigned long next_balance = jiffies + 60*HZ;
7602 int update_next_balance = 0;
7603 int need_serialize, need_decay = 0;
7606 update_blocked_averages(cpu);
7609 for_each_domain(cpu, sd) {
7611 * Decay the newidle max times here because this is a regular
7612 * visit to all the domains. Decay ~1% per second.
7614 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7615 sd->max_newidle_lb_cost =
7616 (sd->max_newidle_lb_cost * 253) / 256;
7617 sd->next_decay_max_lb_cost = jiffies + HZ;
7620 max_cost += sd->max_newidle_lb_cost;
7622 if (!(sd->flags & SD_LOAD_BALANCE))
7626 * Stop the load balance at this level. There is another
7627 * CPU in our sched group which is doing load balancing more
7630 if (!continue_balancing) {
7636 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7638 need_serialize = sd->flags & SD_SERIALIZE;
7639 if (need_serialize) {
7640 if (!spin_trylock(&balancing))
7644 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7645 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7647 * The LBF_DST_PINNED logic could have changed
7648 * env->dst_cpu, so we can't know our idle
7649 * state even if we migrated tasks. Update it.
7651 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7653 sd->last_balance = jiffies;
7654 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7657 spin_unlock(&balancing);
7659 if (time_after(next_balance, sd->last_balance + interval)) {
7660 next_balance = sd->last_balance + interval;
7661 update_next_balance = 1;
7666 * Ensure the rq-wide value also decays but keep it at a
7667 * reasonable floor to avoid funnies with rq->avg_idle.
7669 rq->max_idle_balance_cost =
7670 max((u64)sysctl_sched_migration_cost, max_cost);
7675 * next_balance will be updated only when there is a need.
7676 * When the cpu is attached to null domain for ex, it will not be
7679 if (likely(update_next_balance)) {
7680 rq->next_balance = next_balance;
7682 #ifdef CONFIG_NO_HZ_COMMON
7684 * If this CPU has been elected to perform the nohz idle
7685 * balance. Other idle CPUs have already rebalanced with
7686 * nohz_idle_balance() and nohz.next_balance has been
7687 * updated accordingly. This CPU is now running the idle load
7688 * balance for itself and we need to update the
7689 * nohz.next_balance accordingly.
7691 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
7692 nohz.next_balance = rq->next_balance;
7697 #ifdef CONFIG_NO_HZ_COMMON
7699 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7700 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7702 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7704 int this_cpu = this_rq->cpu;
7707 /* Earliest time when we have to do rebalance again */
7708 unsigned long next_balance = jiffies + 60*HZ;
7709 int update_next_balance = 0;
7711 if (idle != CPU_IDLE ||
7712 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7715 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7716 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7720 * If this cpu gets work to do, stop the load balancing
7721 * work being done for other cpus. Next load
7722 * balancing owner will pick it up.
7727 rq = cpu_rq(balance_cpu);
7730 * If time for next balance is due,
7733 if (time_after_eq(jiffies, rq->next_balance)) {
7734 raw_spin_lock_irq(&rq->lock);
7735 update_rq_clock(rq);
7736 update_idle_cpu_load(rq);
7737 raw_spin_unlock_irq(&rq->lock);
7738 rebalance_domains(rq, CPU_IDLE);
7741 if (time_after(next_balance, rq->next_balance)) {
7742 next_balance = rq->next_balance;
7743 update_next_balance = 1;
7748 * next_balance will be updated only when there is a need.
7749 * When the CPU is attached to null domain for ex, it will not be
7752 if (likely(update_next_balance))
7753 nohz.next_balance = next_balance;
7755 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7759 * Current heuristic for kicking the idle load balancer in the presence
7760 * of an idle cpu in the system.
7761 * - This rq has more than one task.
7762 * - This rq has at least one CFS task and the capacity of the CPU is
7763 * significantly reduced because of RT tasks or IRQs.
7764 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7765 * multiple busy cpu.
7766 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7767 * domain span are idle.
7769 static inline bool nohz_kick_needed(struct rq *rq)
7771 unsigned long now = jiffies;
7772 struct sched_domain *sd;
7773 struct sched_group_capacity *sgc;
7774 int nr_busy, cpu = rq->cpu;
7777 if (unlikely(rq->idle_balance))
7781 * We may be recently in ticked or tickless idle mode. At the first
7782 * busy tick after returning from idle, we will update the busy stats.
7784 set_cpu_sd_state_busy();
7785 nohz_balance_exit_idle(cpu);
7788 * None are in tickless mode and hence no need for NOHZ idle load
7791 if (likely(!atomic_read(&nohz.nr_cpus)))
7794 if (time_before(now, nohz.next_balance))
7797 if (rq->nr_running >= 2)
7801 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7803 sgc = sd->groups->sgc;
7804 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7813 sd = rcu_dereference(rq->sd);
7815 if ((rq->cfs.h_nr_running >= 1) &&
7816 check_cpu_capacity(rq, sd)) {
7822 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7823 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7824 sched_domain_span(sd)) < cpu)) {
7834 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7838 * run_rebalance_domains is triggered when needed from the scheduler tick.
7839 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7841 static void run_rebalance_domains(struct softirq_action *h)
7843 struct rq *this_rq = this_rq();
7844 enum cpu_idle_type idle = this_rq->idle_balance ?
7845 CPU_IDLE : CPU_NOT_IDLE;
7848 * If this cpu has a pending nohz_balance_kick, then do the
7849 * balancing on behalf of the other idle cpus whose ticks are
7850 * stopped. Do nohz_idle_balance *before* rebalance_domains to
7851 * give the idle cpus a chance to load balance. Else we may
7852 * load balance only within the local sched_domain hierarchy
7853 * and abort nohz_idle_balance altogether if we pull some load.
7855 nohz_idle_balance(this_rq, idle);
7856 rebalance_domains(this_rq, idle);
7860 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7862 void trigger_load_balance(struct rq *rq)
7864 /* Don't need to rebalance while attached to NULL domain */
7865 if (unlikely(on_null_domain(rq)))
7868 if (time_after_eq(jiffies, rq->next_balance))
7869 raise_softirq(SCHED_SOFTIRQ);
7870 #ifdef CONFIG_NO_HZ_COMMON
7871 if (nohz_kick_needed(rq))
7872 nohz_balancer_kick();
7876 static void rq_online_fair(struct rq *rq)
7880 update_runtime_enabled(rq);
7883 static void rq_offline_fair(struct rq *rq)
7887 /* Ensure any throttled groups are reachable by pick_next_task */
7888 unthrottle_offline_cfs_rqs(rq);
7891 #endif /* CONFIG_SMP */
7894 * scheduler tick hitting a task of our scheduling class:
7896 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7898 struct cfs_rq *cfs_rq;
7899 struct sched_entity *se = &curr->se;
7901 for_each_sched_entity(se) {
7902 cfs_rq = cfs_rq_of(se);
7903 entity_tick(cfs_rq, se, queued);
7906 if (static_branch_unlikely(&sched_numa_balancing))
7907 task_tick_numa(rq, curr);
7911 * called on fork with the child task as argument from the parent's context
7912 * - child not yet on the tasklist
7913 * - preemption disabled
7915 static void task_fork_fair(struct task_struct *p)
7917 struct cfs_rq *cfs_rq;
7918 struct sched_entity *se = &p->se, *curr;
7919 int this_cpu = smp_processor_id();
7920 struct rq *rq = this_rq();
7921 unsigned long flags;
7923 raw_spin_lock_irqsave(&rq->lock, flags);
7925 update_rq_clock(rq);
7927 cfs_rq = task_cfs_rq(current);
7928 curr = cfs_rq->curr;
7931 * Not only the cpu but also the task_group of the parent might have
7932 * been changed after parent->se.parent,cfs_rq were copied to
7933 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7934 * of child point to valid ones.
7937 __set_task_cpu(p, this_cpu);
7940 update_curr(cfs_rq);
7943 se->vruntime = curr->vruntime;
7944 place_entity(cfs_rq, se, 1);
7946 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7948 * Upon rescheduling, sched_class::put_prev_task() will place
7949 * 'current' within the tree based on its new key value.
7951 swap(curr->vruntime, se->vruntime);
7955 se->vruntime -= cfs_rq->min_vruntime;
7957 raw_spin_unlock_irqrestore(&rq->lock, flags);
7961 * Priority of the task has changed. Check to see if we preempt
7965 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7967 if (!task_on_rq_queued(p))
7971 * Reschedule if we are currently running on this runqueue and
7972 * our priority decreased, or if we are not currently running on
7973 * this runqueue and our priority is higher than the current's
7975 if (rq->curr == p) {
7976 if (p->prio > oldprio)
7979 check_preempt_curr(rq, p, 0);
7982 static inline bool vruntime_normalized(struct task_struct *p)
7984 struct sched_entity *se = &p->se;
7987 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
7988 * the dequeue_entity(.flags=0) will already have normalized the
7995 * When !on_rq, vruntime of the task has usually NOT been normalized.
7996 * But there are some cases where it has already been normalized:
7998 * - A forked child which is waiting for being woken up by
7999 * wake_up_new_task().
8000 * - A task which has been woken up by try_to_wake_up() and
8001 * waiting for actually being woken up by sched_ttwu_pending().
8003 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8009 static void detach_task_cfs_rq(struct task_struct *p)
8011 struct sched_entity *se = &p->se;
8012 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8014 if (!vruntime_normalized(p)) {
8016 * Fix up our vruntime so that the current sleep doesn't
8017 * cause 'unlimited' sleep bonus.
8019 place_entity(cfs_rq, se, 0);
8020 se->vruntime -= cfs_rq->min_vruntime;
8023 /* Catch up with the cfs_rq and remove our load when we leave */
8024 detach_entity_load_avg(cfs_rq, se);
8027 static void attach_task_cfs_rq(struct task_struct *p)
8029 struct sched_entity *se = &p->se;
8030 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8032 #ifdef CONFIG_FAIR_GROUP_SCHED
8034 * Since the real-depth could have been changed (only FAIR
8035 * class maintain depth value), reset depth properly.
8037 se->depth = se->parent ? se->parent->depth + 1 : 0;
8040 /* Synchronize task with its cfs_rq */
8041 attach_entity_load_avg(cfs_rq, se);
8043 if (!vruntime_normalized(p))
8044 se->vruntime += cfs_rq->min_vruntime;
8047 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8049 detach_task_cfs_rq(p);
8052 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8054 attach_task_cfs_rq(p);
8056 if (task_on_rq_queued(p)) {
8058 * We were most likely switched from sched_rt, so
8059 * kick off the schedule if running, otherwise just see
8060 * if we can still preempt the current task.
8065 check_preempt_curr(rq, p, 0);
8069 /* Account for a task changing its policy or group.
8071 * This routine is mostly called to set cfs_rq->curr field when a task
8072 * migrates between groups/classes.
8074 static void set_curr_task_fair(struct rq *rq)
8076 struct sched_entity *se = &rq->curr->se;
8078 for_each_sched_entity(se) {
8079 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8081 set_next_entity(cfs_rq, se);
8082 /* ensure bandwidth has been allocated on our new cfs_rq */
8083 account_cfs_rq_runtime(cfs_rq, 0);
8087 void init_cfs_rq(struct cfs_rq *cfs_rq)
8089 cfs_rq->tasks_timeline = RB_ROOT;
8090 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8091 #ifndef CONFIG_64BIT
8092 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8095 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8096 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8100 #ifdef CONFIG_FAIR_GROUP_SCHED
8101 static void task_move_group_fair(struct task_struct *p)
8103 detach_task_cfs_rq(p);
8104 set_task_rq(p, task_cpu(p));
8107 /* Tell se's cfs_rq has been changed -- migrated */
8108 p->se.avg.last_update_time = 0;
8110 attach_task_cfs_rq(p);
8113 void free_fair_sched_group(struct task_group *tg)
8117 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8119 for_each_possible_cpu(i) {
8121 kfree(tg->cfs_rq[i]);
8124 remove_entity_load_avg(tg->se[i]);
8133 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8135 struct cfs_rq *cfs_rq;
8136 struct sched_entity *se;
8139 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8142 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8146 tg->shares = NICE_0_LOAD;
8148 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8150 for_each_possible_cpu(i) {
8151 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8152 GFP_KERNEL, cpu_to_node(i));
8156 se = kzalloc_node(sizeof(struct sched_entity),
8157 GFP_KERNEL, cpu_to_node(i));
8161 init_cfs_rq(cfs_rq);
8162 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8163 init_entity_runnable_average(se);
8174 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8176 struct rq *rq = cpu_rq(cpu);
8177 unsigned long flags;
8180 * Only empty task groups can be destroyed; so we can speculatively
8181 * check on_list without danger of it being re-added.
8183 if (!tg->cfs_rq[cpu]->on_list)
8186 raw_spin_lock_irqsave(&rq->lock, flags);
8187 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8188 raw_spin_unlock_irqrestore(&rq->lock, flags);
8191 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8192 struct sched_entity *se, int cpu,
8193 struct sched_entity *parent)
8195 struct rq *rq = cpu_rq(cpu);
8199 init_cfs_rq_runtime(cfs_rq);
8201 tg->cfs_rq[cpu] = cfs_rq;
8204 /* se could be NULL for root_task_group */
8209 se->cfs_rq = &rq->cfs;
8212 se->cfs_rq = parent->my_q;
8213 se->depth = parent->depth + 1;
8217 /* guarantee group entities always have weight */
8218 update_load_set(&se->load, NICE_0_LOAD);
8219 se->parent = parent;
8222 static DEFINE_MUTEX(shares_mutex);
8224 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8227 unsigned long flags;
8230 * We can't change the weight of the root cgroup.
8235 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8237 mutex_lock(&shares_mutex);
8238 if (tg->shares == shares)
8241 tg->shares = shares;
8242 for_each_possible_cpu(i) {
8243 struct rq *rq = cpu_rq(i);
8244 struct sched_entity *se;
8247 /* Propagate contribution to hierarchy */
8248 raw_spin_lock_irqsave(&rq->lock, flags);
8250 /* Possible calls to update_curr() need rq clock */
8251 update_rq_clock(rq);
8252 for_each_sched_entity(se)
8253 update_cfs_shares(group_cfs_rq(se));
8254 raw_spin_unlock_irqrestore(&rq->lock, flags);
8258 mutex_unlock(&shares_mutex);
8261 #else /* CONFIG_FAIR_GROUP_SCHED */
8263 void free_fair_sched_group(struct task_group *tg) { }
8265 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8270 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8272 #endif /* CONFIG_FAIR_GROUP_SCHED */
8275 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8277 struct sched_entity *se = &task->se;
8278 unsigned int rr_interval = 0;
8281 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8284 if (rq->cfs.load.weight)
8285 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8291 * All the scheduling class methods:
8293 const struct sched_class fair_sched_class = {
8294 .next = &idle_sched_class,
8295 .enqueue_task = enqueue_task_fair,
8296 .dequeue_task = dequeue_task_fair,
8297 .yield_task = yield_task_fair,
8298 .yield_to_task = yield_to_task_fair,
8300 .check_preempt_curr = check_preempt_wakeup,
8302 .pick_next_task = pick_next_task_fair,
8303 .put_prev_task = put_prev_task_fair,
8306 .select_task_rq = select_task_rq_fair,
8307 .migrate_task_rq = migrate_task_rq_fair,
8309 .rq_online = rq_online_fair,
8310 .rq_offline = rq_offline_fair,
8312 .task_waking = task_waking_fair,
8313 .task_dead = task_dead_fair,
8314 .set_cpus_allowed = set_cpus_allowed_common,
8317 .set_curr_task = set_curr_task_fair,
8318 .task_tick = task_tick_fair,
8319 .task_fork = task_fork_fair,
8321 .prio_changed = prio_changed_fair,
8322 .switched_from = switched_from_fair,
8323 .switched_to = switched_to_fair,
8325 .get_rr_interval = get_rr_interval_fair,
8327 .update_curr = update_curr_fair,
8329 #ifdef CONFIG_FAIR_GROUP_SCHED
8330 .task_move_group = task_move_group_fair,
8334 #ifdef CONFIG_SCHED_DEBUG
8335 void print_cfs_stats(struct seq_file *m, int cpu)
8337 struct cfs_rq *cfs_rq;
8340 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8341 print_cfs_rq(m, cpu, cfs_rq);
8345 #ifdef CONFIG_NUMA_BALANCING
8346 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8349 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8351 for_each_online_node(node) {
8352 if (p->numa_faults) {
8353 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8354 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8356 if (p->numa_group) {
8357 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8358 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8360 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8363 #endif /* CONFIG_NUMA_BALANCING */
8364 #endif /* CONFIG_SCHED_DEBUG */
8366 __init void init_sched_fair_class(void)
8369 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8371 #ifdef CONFIG_NO_HZ_COMMON
8372 nohz.next_balance = jiffies;
8373 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8374 cpu_notifier(sched_ilb_notifier, 0);