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);
2827 #ifndef CONFIG_64BIT
2828 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2830 u64 last_update_time_copy;
2831 u64 last_update_time;
2834 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2836 last_update_time = cfs_rq->avg.last_update_time;
2837 } while (last_update_time != last_update_time_copy);
2839 return last_update_time;
2842 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2844 return cfs_rq->avg.last_update_time;
2849 * Task first catches up with cfs_rq, and then subtract
2850 * itself from the cfs_rq (task must be off the queue now).
2852 void remove_entity_load_avg(struct sched_entity *se)
2854 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2855 u64 last_update_time;
2858 * Newly created task or never used group entity should not be removed
2859 * from its (source) cfs_rq
2861 if (se->avg.last_update_time == 0)
2864 last_update_time = cfs_rq_last_update_time(cfs_rq);
2866 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2867 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2868 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2872 * Update the rq's load with the elapsed running time before entering
2873 * idle. if the last scheduled task is not a CFS task, idle_enter will
2874 * be the only way to update the runnable statistic.
2876 void idle_enter_fair(struct rq *this_rq)
2881 * Update the rq's load with the elapsed idle time before a task is
2882 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2883 * be the only way to update the runnable statistic.
2885 void idle_exit_fair(struct rq *this_rq)
2889 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2891 return cfs_rq->runnable_load_avg;
2894 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2896 return cfs_rq->avg.load_avg;
2899 static int idle_balance(struct rq *this_rq);
2901 #else /* CONFIG_SMP */
2903 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2905 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2907 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2908 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2911 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2913 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2915 static inline int idle_balance(struct rq *rq)
2920 #endif /* CONFIG_SMP */
2922 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2924 #ifdef CONFIG_SCHEDSTATS
2925 struct task_struct *tsk = NULL;
2927 if (entity_is_task(se))
2930 if (se->statistics.sleep_start) {
2931 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2936 if (unlikely(delta > se->statistics.sleep_max))
2937 se->statistics.sleep_max = delta;
2939 se->statistics.sleep_start = 0;
2940 se->statistics.sum_sleep_runtime += delta;
2943 account_scheduler_latency(tsk, delta >> 10, 1);
2944 trace_sched_stat_sleep(tsk, delta);
2947 if (se->statistics.block_start) {
2948 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2953 if (unlikely(delta > se->statistics.block_max))
2954 se->statistics.block_max = delta;
2956 se->statistics.block_start = 0;
2957 se->statistics.sum_sleep_runtime += delta;
2960 if (tsk->in_iowait) {
2961 se->statistics.iowait_sum += delta;
2962 se->statistics.iowait_count++;
2963 trace_sched_stat_iowait(tsk, delta);
2966 trace_sched_stat_blocked(tsk, delta);
2967 trace_sched_blocked_reason(tsk);
2970 * Blocking time is in units of nanosecs, so shift by
2971 * 20 to get a milliseconds-range estimation of the
2972 * amount of time that the task spent sleeping:
2974 if (unlikely(prof_on == SLEEP_PROFILING)) {
2975 profile_hits(SLEEP_PROFILING,
2976 (void *)get_wchan(tsk),
2979 account_scheduler_latency(tsk, delta >> 10, 0);
2985 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2987 #ifdef CONFIG_SCHED_DEBUG
2988 s64 d = se->vruntime - cfs_rq->min_vruntime;
2993 if (d > 3*sysctl_sched_latency)
2994 schedstat_inc(cfs_rq, nr_spread_over);
2999 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3001 u64 vruntime = cfs_rq->min_vruntime;
3004 * The 'current' period is already promised to the current tasks,
3005 * however the extra weight of the new task will slow them down a
3006 * little, place the new task so that it fits in the slot that
3007 * stays open at the end.
3009 if (initial && sched_feat(START_DEBIT))
3010 vruntime += sched_vslice(cfs_rq, se);
3012 /* sleeps up to a single latency don't count. */
3014 unsigned long thresh = sysctl_sched_latency;
3017 * Halve their sleep time's effect, to allow
3018 * for a gentler effect of sleepers:
3020 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3026 /* ensure we never gain time by being placed backwards. */
3027 se->vruntime = max_vruntime(se->vruntime, vruntime);
3030 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3033 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3036 * Update the normalized vruntime before updating min_vruntime
3037 * through calling update_curr().
3039 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3040 se->vruntime += cfs_rq->min_vruntime;
3043 * Update run-time statistics of the 'current'.
3045 update_curr(cfs_rq);
3046 enqueue_entity_load_avg(cfs_rq, se);
3047 account_entity_enqueue(cfs_rq, se);
3048 update_cfs_shares(cfs_rq);
3050 if (flags & ENQUEUE_WAKEUP) {
3051 place_entity(cfs_rq, se, 0);
3052 enqueue_sleeper(cfs_rq, se);
3055 update_stats_enqueue(cfs_rq, se);
3056 check_spread(cfs_rq, se);
3057 if (se != cfs_rq->curr)
3058 __enqueue_entity(cfs_rq, se);
3061 if (cfs_rq->nr_running == 1) {
3062 list_add_leaf_cfs_rq(cfs_rq);
3063 check_enqueue_throttle(cfs_rq);
3067 static void __clear_buddies_last(struct sched_entity *se)
3069 for_each_sched_entity(se) {
3070 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3071 if (cfs_rq->last != se)
3074 cfs_rq->last = NULL;
3078 static void __clear_buddies_next(struct sched_entity *se)
3080 for_each_sched_entity(se) {
3081 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3082 if (cfs_rq->next != se)
3085 cfs_rq->next = NULL;
3089 static void __clear_buddies_skip(struct sched_entity *se)
3091 for_each_sched_entity(se) {
3092 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3093 if (cfs_rq->skip != se)
3096 cfs_rq->skip = NULL;
3100 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3102 if (cfs_rq->last == se)
3103 __clear_buddies_last(se);
3105 if (cfs_rq->next == se)
3106 __clear_buddies_next(se);
3108 if (cfs_rq->skip == se)
3109 __clear_buddies_skip(se);
3112 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3115 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3118 * Update run-time statistics of the 'current'.
3120 update_curr(cfs_rq);
3121 dequeue_entity_load_avg(cfs_rq, se);
3123 update_stats_dequeue(cfs_rq, se);
3124 if (flags & DEQUEUE_SLEEP) {
3125 #ifdef CONFIG_SCHEDSTATS
3126 if (entity_is_task(se)) {
3127 struct task_struct *tsk = task_of(se);
3129 if (tsk->state & TASK_INTERRUPTIBLE)
3130 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3131 if (tsk->state & TASK_UNINTERRUPTIBLE)
3132 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3137 clear_buddies(cfs_rq, se);
3139 if (se != cfs_rq->curr)
3140 __dequeue_entity(cfs_rq, se);
3142 account_entity_dequeue(cfs_rq, se);
3145 * Normalize the entity after updating the min_vruntime because the
3146 * update can refer to the ->curr item and we need to reflect this
3147 * movement in our normalized position.
3149 if (!(flags & DEQUEUE_SLEEP))
3150 se->vruntime -= cfs_rq->min_vruntime;
3152 /* return excess runtime on last dequeue */
3153 return_cfs_rq_runtime(cfs_rq);
3155 update_min_vruntime(cfs_rq);
3156 update_cfs_shares(cfs_rq);
3160 * Preempt the current task with a newly woken task if needed:
3163 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3165 unsigned long ideal_runtime, delta_exec;
3166 struct sched_entity *se;
3169 ideal_runtime = sched_slice(cfs_rq, curr);
3170 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3171 if (delta_exec > ideal_runtime) {
3172 resched_curr(rq_of(cfs_rq));
3174 * The current task ran long enough, ensure it doesn't get
3175 * re-elected due to buddy favours.
3177 clear_buddies(cfs_rq, curr);
3182 * Ensure that a task that missed wakeup preemption by a
3183 * narrow margin doesn't have to wait for a full slice.
3184 * This also mitigates buddy induced latencies under load.
3186 if (delta_exec < sysctl_sched_min_granularity)
3189 se = __pick_first_entity(cfs_rq);
3190 delta = curr->vruntime - se->vruntime;
3195 if (delta > ideal_runtime)
3196 resched_curr(rq_of(cfs_rq));
3200 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3202 /* 'current' is not kept within the tree. */
3205 * Any task has to be enqueued before it get to execute on
3206 * a CPU. So account for the time it spent waiting on the
3209 update_stats_wait_end(cfs_rq, se);
3210 __dequeue_entity(cfs_rq, se);
3211 update_load_avg(se, 1);
3214 update_stats_curr_start(cfs_rq, se);
3216 #ifdef CONFIG_SCHEDSTATS
3218 * Track our maximum slice length, if the CPU's load is at
3219 * least twice that of our own weight (i.e. dont track it
3220 * when there are only lesser-weight tasks around):
3222 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3223 se->statistics.slice_max = max(se->statistics.slice_max,
3224 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3227 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3231 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3234 * Pick the next process, keeping these things in mind, in this order:
3235 * 1) keep things fair between processes/task groups
3236 * 2) pick the "next" process, since someone really wants that to run
3237 * 3) pick the "last" process, for cache locality
3238 * 4) do not run the "skip" process, if something else is available
3240 static struct sched_entity *
3241 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3243 struct sched_entity *left = __pick_first_entity(cfs_rq);
3244 struct sched_entity *se;
3247 * If curr is set we have to see if its left of the leftmost entity
3248 * still in the tree, provided there was anything in the tree at all.
3250 if (!left || (curr && entity_before(curr, left)))
3253 se = left; /* ideally we run the leftmost entity */
3256 * Avoid running the skip buddy, if running something else can
3257 * be done without getting too unfair.
3259 if (cfs_rq->skip == se) {
3260 struct sched_entity *second;
3263 second = __pick_first_entity(cfs_rq);
3265 second = __pick_next_entity(se);
3266 if (!second || (curr && entity_before(curr, second)))
3270 if (second && wakeup_preempt_entity(second, left) < 1)
3275 * Prefer last buddy, try to return the CPU to a preempted task.
3277 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3281 * Someone really wants this to run. If it's not unfair, run it.
3283 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3286 clear_buddies(cfs_rq, se);
3291 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3293 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3296 * If still on the runqueue then deactivate_task()
3297 * was not called and update_curr() has to be done:
3300 update_curr(cfs_rq);
3302 /* throttle cfs_rqs exceeding runtime */
3303 check_cfs_rq_runtime(cfs_rq);
3305 check_spread(cfs_rq, prev);
3307 update_stats_wait_start(cfs_rq, prev);
3308 /* Put 'current' back into the tree. */
3309 __enqueue_entity(cfs_rq, prev);
3310 /* in !on_rq case, update occurred at dequeue */
3311 update_load_avg(prev, 0);
3313 cfs_rq->curr = NULL;
3317 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3320 * Update run-time statistics of the 'current'.
3322 update_curr(cfs_rq);
3325 * Ensure that runnable average is periodically updated.
3327 update_load_avg(curr, 1);
3328 update_cfs_shares(cfs_rq);
3330 #ifdef CONFIG_SCHED_HRTICK
3332 * queued ticks are scheduled to match the slice, so don't bother
3333 * validating it and just reschedule.
3336 resched_curr(rq_of(cfs_rq));
3340 * don't let the period tick interfere with the hrtick preemption
3342 if (!sched_feat(DOUBLE_TICK) &&
3343 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3347 if (cfs_rq->nr_running > 1)
3348 check_preempt_tick(cfs_rq, curr);
3352 /**************************************************
3353 * CFS bandwidth control machinery
3356 #ifdef CONFIG_CFS_BANDWIDTH
3358 #ifdef HAVE_JUMP_LABEL
3359 static struct static_key __cfs_bandwidth_used;
3361 static inline bool cfs_bandwidth_used(void)
3363 return static_key_false(&__cfs_bandwidth_used);
3366 void cfs_bandwidth_usage_inc(void)
3368 static_key_slow_inc(&__cfs_bandwidth_used);
3371 void cfs_bandwidth_usage_dec(void)
3373 static_key_slow_dec(&__cfs_bandwidth_used);
3375 #else /* HAVE_JUMP_LABEL */
3376 static bool cfs_bandwidth_used(void)
3381 void cfs_bandwidth_usage_inc(void) {}
3382 void cfs_bandwidth_usage_dec(void) {}
3383 #endif /* HAVE_JUMP_LABEL */
3386 * default period for cfs group bandwidth.
3387 * default: 0.1s, units: nanoseconds
3389 static inline u64 default_cfs_period(void)
3391 return 100000000ULL;
3394 static inline u64 sched_cfs_bandwidth_slice(void)
3396 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3400 * Replenish runtime according to assigned quota and update expiration time.
3401 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3402 * additional synchronization around rq->lock.
3404 * requires cfs_b->lock
3406 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3410 if (cfs_b->quota == RUNTIME_INF)
3413 now = sched_clock_cpu(smp_processor_id());
3414 cfs_b->runtime = cfs_b->quota;
3415 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3418 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3420 return &tg->cfs_bandwidth;
3423 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3424 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3426 if (unlikely(cfs_rq->throttle_count))
3427 return cfs_rq->throttled_clock_task;
3429 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3432 /* returns 0 on failure to allocate runtime */
3433 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3435 struct task_group *tg = cfs_rq->tg;
3436 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3437 u64 amount = 0, min_amount, expires;
3439 /* note: this is a positive sum as runtime_remaining <= 0 */
3440 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3442 raw_spin_lock(&cfs_b->lock);
3443 if (cfs_b->quota == RUNTIME_INF)
3444 amount = min_amount;
3446 start_cfs_bandwidth(cfs_b);
3448 if (cfs_b->runtime > 0) {
3449 amount = min(cfs_b->runtime, min_amount);
3450 cfs_b->runtime -= amount;
3454 expires = cfs_b->runtime_expires;
3455 raw_spin_unlock(&cfs_b->lock);
3457 cfs_rq->runtime_remaining += amount;
3459 * we may have advanced our local expiration to account for allowed
3460 * spread between our sched_clock and the one on which runtime was
3463 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3464 cfs_rq->runtime_expires = expires;
3466 return cfs_rq->runtime_remaining > 0;
3470 * Note: This depends on the synchronization provided by sched_clock and the
3471 * fact that rq->clock snapshots this value.
3473 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3475 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3477 /* if the deadline is ahead of our clock, nothing to do */
3478 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3481 if (cfs_rq->runtime_remaining < 0)
3485 * If the local deadline has passed we have to consider the
3486 * possibility that our sched_clock is 'fast' and the global deadline
3487 * has not truly expired.
3489 * Fortunately we can check determine whether this the case by checking
3490 * whether the global deadline has advanced. It is valid to compare
3491 * cfs_b->runtime_expires without any locks since we only care about
3492 * exact equality, so a partial write will still work.
3495 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3496 /* extend local deadline, drift is bounded above by 2 ticks */
3497 cfs_rq->runtime_expires += TICK_NSEC;
3499 /* global deadline is ahead, expiration has passed */
3500 cfs_rq->runtime_remaining = 0;
3504 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3506 /* dock delta_exec before expiring quota (as it could span periods) */
3507 cfs_rq->runtime_remaining -= delta_exec;
3508 expire_cfs_rq_runtime(cfs_rq);
3510 if (likely(cfs_rq->runtime_remaining > 0))
3514 * if we're unable to extend our runtime we resched so that the active
3515 * hierarchy can be throttled
3517 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3518 resched_curr(rq_of(cfs_rq));
3521 static __always_inline
3522 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3524 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3527 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3530 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3532 return cfs_bandwidth_used() && cfs_rq->throttled;
3535 /* check whether cfs_rq, or any parent, is throttled */
3536 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3538 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3542 * Ensure that neither of the group entities corresponding to src_cpu or
3543 * dest_cpu are members of a throttled hierarchy when performing group
3544 * load-balance operations.
3546 static inline int throttled_lb_pair(struct task_group *tg,
3547 int src_cpu, int dest_cpu)
3549 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3551 src_cfs_rq = tg->cfs_rq[src_cpu];
3552 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3554 return throttled_hierarchy(src_cfs_rq) ||
3555 throttled_hierarchy(dest_cfs_rq);
3558 /* updated child weight may affect parent so we have to do this bottom up */
3559 static int tg_unthrottle_up(struct task_group *tg, void *data)
3561 struct rq *rq = data;
3562 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3564 cfs_rq->throttle_count--;
3566 if (!cfs_rq->throttle_count) {
3567 /* adjust cfs_rq_clock_task() */
3568 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3569 cfs_rq->throttled_clock_task;
3576 static int tg_throttle_down(struct task_group *tg, void *data)
3578 struct rq *rq = data;
3579 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3581 /* group is entering throttled state, stop time */
3582 if (!cfs_rq->throttle_count)
3583 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3584 cfs_rq->throttle_count++;
3589 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3591 struct rq *rq = rq_of(cfs_rq);
3592 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3593 struct sched_entity *se;
3594 long task_delta, dequeue = 1;
3597 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3599 /* freeze hierarchy runnable averages while throttled */
3601 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3604 task_delta = cfs_rq->h_nr_running;
3605 for_each_sched_entity(se) {
3606 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3607 /* throttled entity or throttle-on-deactivate */
3612 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3613 qcfs_rq->h_nr_running -= task_delta;
3615 if (qcfs_rq->load.weight)
3620 sub_nr_running(rq, task_delta);
3622 cfs_rq->throttled = 1;
3623 cfs_rq->throttled_clock = rq_clock(rq);
3624 raw_spin_lock(&cfs_b->lock);
3625 empty = list_empty(&cfs_b->throttled_cfs_rq);
3628 * Add to the _head_ of the list, so that an already-started
3629 * distribute_cfs_runtime will not see us
3631 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3634 * If we're the first throttled task, make sure the bandwidth
3638 start_cfs_bandwidth(cfs_b);
3640 raw_spin_unlock(&cfs_b->lock);
3643 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3645 struct rq *rq = rq_of(cfs_rq);
3646 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3647 struct sched_entity *se;
3651 se = cfs_rq->tg->se[cpu_of(rq)];
3653 cfs_rq->throttled = 0;
3655 update_rq_clock(rq);
3657 raw_spin_lock(&cfs_b->lock);
3658 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3659 list_del_rcu(&cfs_rq->throttled_list);
3660 raw_spin_unlock(&cfs_b->lock);
3662 /* update hierarchical throttle state */
3663 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3665 if (!cfs_rq->load.weight)
3668 task_delta = cfs_rq->h_nr_running;
3669 for_each_sched_entity(se) {
3673 cfs_rq = cfs_rq_of(se);
3675 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3676 cfs_rq->h_nr_running += task_delta;
3678 if (cfs_rq_throttled(cfs_rq))
3683 add_nr_running(rq, task_delta);
3685 /* determine whether we need to wake up potentially idle cpu */
3686 if (rq->curr == rq->idle && rq->cfs.nr_running)
3690 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3691 u64 remaining, u64 expires)
3693 struct cfs_rq *cfs_rq;
3695 u64 starting_runtime = remaining;
3698 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3700 struct rq *rq = rq_of(cfs_rq);
3702 raw_spin_lock(&rq->lock);
3703 if (!cfs_rq_throttled(cfs_rq))
3706 runtime = -cfs_rq->runtime_remaining + 1;
3707 if (runtime > remaining)
3708 runtime = remaining;
3709 remaining -= runtime;
3711 cfs_rq->runtime_remaining += runtime;
3712 cfs_rq->runtime_expires = expires;
3714 /* we check whether we're throttled above */
3715 if (cfs_rq->runtime_remaining > 0)
3716 unthrottle_cfs_rq(cfs_rq);
3719 raw_spin_unlock(&rq->lock);
3726 return starting_runtime - remaining;
3730 * Responsible for refilling a task_group's bandwidth and unthrottling its
3731 * cfs_rqs as appropriate. If there has been no activity within the last
3732 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3733 * used to track this state.
3735 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3737 u64 runtime, runtime_expires;
3740 /* no need to continue the timer with no bandwidth constraint */
3741 if (cfs_b->quota == RUNTIME_INF)
3742 goto out_deactivate;
3744 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3745 cfs_b->nr_periods += overrun;
3748 * idle depends on !throttled (for the case of a large deficit), and if
3749 * we're going inactive then everything else can be deferred
3751 if (cfs_b->idle && !throttled)
3752 goto out_deactivate;
3754 __refill_cfs_bandwidth_runtime(cfs_b);
3757 /* mark as potentially idle for the upcoming period */
3762 /* account preceding periods in which throttling occurred */
3763 cfs_b->nr_throttled += overrun;
3765 runtime_expires = cfs_b->runtime_expires;
3768 * This check is repeated as we are holding onto the new bandwidth while
3769 * we unthrottle. This can potentially race with an unthrottled group
3770 * trying to acquire new bandwidth from the global pool. This can result
3771 * in us over-using our runtime if it is all used during this loop, but
3772 * only by limited amounts in that extreme case.
3774 while (throttled && cfs_b->runtime > 0) {
3775 runtime = cfs_b->runtime;
3776 raw_spin_unlock(&cfs_b->lock);
3777 /* we can't nest cfs_b->lock while distributing bandwidth */
3778 runtime = distribute_cfs_runtime(cfs_b, runtime,
3780 raw_spin_lock(&cfs_b->lock);
3782 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3784 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3788 * While we are ensured activity in the period following an
3789 * unthrottle, this also covers the case in which the new bandwidth is
3790 * insufficient to cover the existing bandwidth deficit. (Forcing the
3791 * timer to remain active while there are any throttled entities.)
3801 /* a cfs_rq won't donate quota below this amount */
3802 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3803 /* minimum remaining period time to redistribute slack quota */
3804 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3805 /* how long we wait to gather additional slack before distributing */
3806 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3809 * Are we near the end of the current quota period?
3811 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3812 * hrtimer base being cleared by hrtimer_start. In the case of
3813 * migrate_hrtimers, base is never cleared, so we are fine.
3815 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3817 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3820 /* if the call-back is running a quota refresh is already occurring */
3821 if (hrtimer_callback_running(refresh_timer))
3824 /* is a quota refresh about to occur? */
3825 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3826 if (remaining < min_expire)
3832 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3834 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3836 /* if there's a quota refresh soon don't bother with slack */
3837 if (runtime_refresh_within(cfs_b, min_left))
3840 hrtimer_start(&cfs_b->slack_timer,
3841 ns_to_ktime(cfs_bandwidth_slack_period),
3845 /* we know any runtime found here is valid as update_curr() precedes return */
3846 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3848 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3849 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3851 if (slack_runtime <= 0)
3854 raw_spin_lock(&cfs_b->lock);
3855 if (cfs_b->quota != RUNTIME_INF &&
3856 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3857 cfs_b->runtime += slack_runtime;
3859 /* we are under rq->lock, defer unthrottling using a timer */
3860 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3861 !list_empty(&cfs_b->throttled_cfs_rq))
3862 start_cfs_slack_bandwidth(cfs_b);
3864 raw_spin_unlock(&cfs_b->lock);
3866 /* even if it's not valid for return we don't want to try again */
3867 cfs_rq->runtime_remaining -= slack_runtime;
3870 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3872 if (!cfs_bandwidth_used())
3875 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3878 __return_cfs_rq_runtime(cfs_rq);
3882 * This is done with a timer (instead of inline with bandwidth return) since
3883 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3885 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3887 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3890 /* confirm we're still not at a refresh boundary */
3891 raw_spin_lock(&cfs_b->lock);
3892 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3893 raw_spin_unlock(&cfs_b->lock);
3897 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3898 runtime = cfs_b->runtime;
3900 expires = cfs_b->runtime_expires;
3901 raw_spin_unlock(&cfs_b->lock);
3906 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3908 raw_spin_lock(&cfs_b->lock);
3909 if (expires == cfs_b->runtime_expires)
3910 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3911 raw_spin_unlock(&cfs_b->lock);
3915 * When a group wakes up we want to make sure that its quota is not already
3916 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3917 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3919 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3921 if (!cfs_bandwidth_used())
3924 /* an active group must be handled by the update_curr()->put() path */
3925 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3928 /* ensure the group is not already throttled */
3929 if (cfs_rq_throttled(cfs_rq))
3932 /* update runtime allocation */
3933 account_cfs_rq_runtime(cfs_rq, 0);
3934 if (cfs_rq->runtime_remaining <= 0)
3935 throttle_cfs_rq(cfs_rq);
3938 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3939 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3941 if (!cfs_bandwidth_used())
3944 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3948 * it's possible for a throttled entity to be forced into a running
3949 * state (e.g. set_curr_task), in this case we're finished.
3951 if (cfs_rq_throttled(cfs_rq))
3954 throttle_cfs_rq(cfs_rq);
3958 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3960 struct cfs_bandwidth *cfs_b =
3961 container_of(timer, struct cfs_bandwidth, slack_timer);
3963 do_sched_cfs_slack_timer(cfs_b);
3965 return HRTIMER_NORESTART;
3968 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3970 struct cfs_bandwidth *cfs_b =
3971 container_of(timer, struct cfs_bandwidth, period_timer);
3975 raw_spin_lock(&cfs_b->lock);
3977 overrun = hrtimer_forward_now(timer, cfs_b->period);
3981 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3984 cfs_b->period_active = 0;
3985 raw_spin_unlock(&cfs_b->lock);
3987 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3990 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3992 raw_spin_lock_init(&cfs_b->lock);
3994 cfs_b->quota = RUNTIME_INF;
3995 cfs_b->period = ns_to_ktime(default_cfs_period());
3997 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3998 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
3999 cfs_b->period_timer.function = sched_cfs_period_timer;
4000 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4001 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4004 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4006 cfs_rq->runtime_enabled = 0;
4007 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4010 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4012 lockdep_assert_held(&cfs_b->lock);
4014 if (!cfs_b->period_active) {
4015 cfs_b->period_active = 1;
4016 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4017 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4021 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4023 /* init_cfs_bandwidth() was not called */
4024 if (!cfs_b->throttled_cfs_rq.next)
4027 hrtimer_cancel(&cfs_b->period_timer);
4028 hrtimer_cancel(&cfs_b->slack_timer);
4031 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4033 struct cfs_rq *cfs_rq;
4035 for_each_leaf_cfs_rq(rq, cfs_rq) {
4036 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4038 raw_spin_lock(&cfs_b->lock);
4039 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4040 raw_spin_unlock(&cfs_b->lock);
4044 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4046 struct cfs_rq *cfs_rq;
4048 for_each_leaf_cfs_rq(rq, cfs_rq) {
4049 if (!cfs_rq->runtime_enabled)
4053 * clock_task is not advancing so we just need to make sure
4054 * there's some valid quota amount
4056 cfs_rq->runtime_remaining = 1;
4058 * Offline rq is schedulable till cpu is completely disabled
4059 * in take_cpu_down(), so we prevent new cfs throttling here.
4061 cfs_rq->runtime_enabled = 0;
4063 if (cfs_rq_throttled(cfs_rq))
4064 unthrottle_cfs_rq(cfs_rq);
4068 #else /* CONFIG_CFS_BANDWIDTH */
4069 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4071 return rq_clock_task(rq_of(cfs_rq));
4074 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4075 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4076 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4077 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4079 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4084 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4089 static inline int throttled_lb_pair(struct task_group *tg,
4090 int src_cpu, int dest_cpu)
4095 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4097 #ifdef CONFIG_FAIR_GROUP_SCHED
4098 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4101 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4105 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4106 static inline void update_runtime_enabled(struct rq *rq) {}
4107 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4109 #endif /* CONFIG_CFS_BANDWIDTH */
4111 /**************************************************
4112 * CFS operations on tasks:
4115 #ifdef CONFIG_SCHED_HRTICK
4116 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4118 struct sched_entity *se = &p->se;
4119 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4121 WARN_ON(task_rq(p) != rq);
4123 if (cfs_rq->nr_running > 1) {
4124 u64 slice = sched_slice(cfs_rq, se);
4125 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4126 s64 delta = slice - ran;
4133 hrtick_start(rq, delta);
4138 * called from enqueue/dequeue and updates the hrtick when the
4139 * current task is from our class and nr_running is low enough
4142 static void hrtick_update(struct rq *rq)
4144 struct task_struct *curr = rq->curr;
4146 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4149 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4150 hrtick_start_fair(rq, curr);
4152 #else /* !CONFIG_SCHED_HRTICK */
4154 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4158 static inline void hrtick_update(struct rq *rq)
4163 static void update_capacity_of(int cpu)
4165 unsigned long req_cap;
4170 /* Convert scale-invariant capacity to cpu. */
4171 req_cap = cpu_util(cpu) * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4172 set_cfs_cpu_capacity(cpu, true, req_cap);
4175 static bool cpu_overutilized(int cpu);
4178 * The enqueue_task method is called before nr_running is
4179 * increased. Here we update the fair scheduling stats and
4180 * then put the task into the rbtree:
4183 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4185 struct cfs_rq *cfs_rq;
4186 struct sched_entity *se = &p->se;
4187 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4188 int task_wakeup = flags & ENQUEUE_WAKEUP;
4190 for_each_sched_entity(se) {
4193 cfs_rq = cfs_rq_of(se);
4194 enqueue_entity(cfs_rq, se, flags);
4197 * end evaluation on encountering a throttled cfs_rq
4199 * note: in the case of encountering a throttled cfs_rq we will
4200 * post the final h_nr_running increment below.
4202 if (cfs_rq_throttled(cfs_rq))
4204 cfs_rq->h_nr_running++;
4206 flags = ENQUEUE_WAKEUP;
4209 for_each_sched_entity(se) {
4210 cfs_rq = cfs_rq_of(se);
4211 cfs_rq->h_nr_running++;
4213 if (cfs_rq_throttled(cfs_rq))
4216 update_load_avg(se, 1);
4217 update_cfs_shares(cfs_rq);
4221 add_nr_running(rq, 1);
4222 if (!task_new && !rq->rd->overutilized &&
4223 cpu_overutilized(rq->cpu))
4224 rq->rd->overutilized = true;
4227 * We want to potentially trigger a freq switch
4228 * request only for tasks that are waking up; this is
4229 * because we get here also during load balancing, but
4230 * in these cases it seems wise to trigger as single
4231 * request after load balancing is done.
4233 if (task_new || task_wakeup)
4234 update_capacity_of(cpu_of(rq));
4239 static void set_next_buddy(struct sched_entity *se);
4242 * The dequeue_task method is called before nr_running is
4243 * decreased. We remove the task from the rbtree and
4244 * update the fair scheduling stats:
4246 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4248 struct cfs_rq *cfs_rq;
4249 struct sched_entity *se = &p->se;
4250 int task_sleep = flags & DEQUEUE_SLEEP;
4252 for_each_sched_entity(se) {
4253 cfs_rq = cfs_rq_of(se);
4254 dequeue_entity(cfs_rq, se, flags);
4257 * end evaluation on encountering a throttled cfs_rq
4259 * note: in the case of encountering a throttled cfs_rq we will
4260 * post the final h_nr_running decrement below.
4262 if (cfs_rq_throttled(cfs_rq))
4264 cfs_rq->h_nr_running--;
4266 /* Don't dequeue parent if it has other entities besides us */
4267 if (cfs_rq->load.weight) {
4269 * Bias pick_next to pick a task from this cfs_rq, as
4270 * p is sleeping when it is within its sched_slice.
4272 if (task_sleep && parent_entity(se))
4273 set_next_buddy(parent_entity(se));
4275 /* avoid re-evaluating load for this entity */
4276 se = parent_entity(se);
4279 flags |= DEQUEUE_SLEEP;
4282 for_each_sched_entity(se) {
4283 cfs_rq = cfs_rq_of(se);
4284 cfs_rq->h_nr_running--;
4286 if (cfs_rq_throttled(cfs_rq))
4289 update_load_avg(se, 1);
4290 update_cfs_shares(cfs_rq);
4294 sub_nr_running(rq, 1);
4297 * We want to potentially trigger a freq switch
4298 * request only for tasks that are going to sleep;
4299 * this is because we get here also during load
4300 * balancing, but in these cases it seems wise to
4301 * trigger as single request after load balancing is
4305 if (rq->cfs.nr_running)
4306 update_capacity_of(cpu_of(rq));
4307 else if (sched_freq())
4308 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4317 * per rq 'load' arrray crap; XXX kill this.
4321 * The exact cpuload at various idx values, calculated at every tick would be
4322 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4324 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4325 * on nth tick when cpu may be busy, then we have:
4326 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4327 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4329 * decay_load_missed() below does efficient calculation of
4330 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4331 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4333 * The calculation is approximated on a 128 point scale.
4334 * degrade_zero_ticks is the number of ticks after which load at any
4335 * particular idx is approximated to be zero.
4336 * degrade_factor is a precomputed table, a row for each load idx.
4337 * Each column corresponds to degradation factor for a power of two ticks,
4338 * based on 128 point scale.
4340 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4341 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4343 * With this power of 2 load factors, we can degrade the load n times
4344 * by looking at 1 bits in n and doing as many mult/shift instead of
4345 * n mult/shifts needed by the exact degradation.
4347 #define DEGRADE_SHIFT 7
4348 static const unsigned char
4349 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4350 static const unsigned char
4351 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4352 {0, 0, 0, 0, 0, 0, 0, 0},
4353 {64, 32, 8, 0, 0, 0, 0, 0},
4354 {96, 72, 40, 12, 1, 0, 0},
4355 {112, 98, 75, 43, 15, 1, 0},
4356 {120, 112, 98, 76, 45, 16, 2} };
4359 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4360 * would be when CPU is idle and so we just decay the old load without
4361 * adding any new load.
4363 static unsigned long
4364 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4368 if (!missed_updates)
4371 if (missed_updates >= degrade_zero_ticks[idx])
4375 return load >> missed_updates;
4377 while (missed_updates) {
4378 if (missed_updates % 2)
4379 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4381 missed_updates >>= 1;
4388 * Update rq->cpu_load[] statistics. This function is usually called every
4389 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4390 * every tick. We fix it up based on jiffies.
4392 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4393 unsigned long pending_updates)
4397 this_rq->nr_load_updates++;
4399 /* Update our load: */
4400 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4401 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4402 unsigned long old_load, new_load;
4404 /* scale is effectively 1 << i now, and >> i divides by scale */
4406 old_load = this_rq->cpu_load[i];
4407 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4408 new_load = this_load;
4410 * Round up the averaging division if load is increasing. This
4411 * prevents us from getting stuck on 9 if the load is 10, for
4414 if (new_load > old_load)
4415 new_load += scale - 1;
4417 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4420 sched_avg_update(this_rq);
4423 /* Used instead of source_load when we know the type == 0 */
4424 static unsigned long weighted_cpuload(const int cpu)
4426 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4429 #ifdef CONFIG_NO_HZ_COMMON
4431 * There is no sane way to deal with nohz on smp when using jiffies because the
4432 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4433 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4435 * Therefore we cannot use the delta approach from the regular tick since that
4436 * would seriously skew the load calculation. However we'll make do for those
4437 * updates happening while idle (nohz_idle_balance) or coming out of idle
4438 * (tick_nohz_idle_exit).
4440 * This means we might still be one tick off for nohz periods.
4444 * Called from nohz_idle_balance() to update the load ratings before doing the
4447 static void update_idle_cpu_load(struct rq *this_rq)
4449 unsigned long curr_jiffies = READ_ONCE(jiffies);
4450 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4451 unsigned long pending_updates;
4454 * bail if there's load or we're actually up-to-date.
4456 if (load || curr_jiffies == this_rq->last_load_update_tick)
4459 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4460 this_rq->last_load_update_tick = curr_jiffies;
4462 __update_cpu_load(this_rq, load, pending_updates);
4466 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4468 void update_cpu_load_nohz(void)
4470 struct rq *this_rq = this_rq();
4471 unsigned long curr_jiffies = READ_ONCE(jiffies);
4472 unsigned long pending_updates;
4474 if (curr_jiffies == this_rq->last_load_update_tick)
4477 raw_spin_lock(&this_rq->lock);
4478 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4479 if (pending_updates) {
4480 this_rq->last_load_update_tick = curr_jiffies;
4482 * We were idle, this means load 0, the current load might be
4483 * !0 due to remote wakeups and the sort.
4485 __update_cpu_load(this_rq, 0, pending_updates);
4487 raw_spin_unlock(&this_rq->lock);
4489 #endif /* CONFIG_NO_HZ */
4492 * Called from scheduler_tick()
4494 void update_cpu_load_active(struct rq *this_rq)
4496 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4498 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4500 this_rq->last_load_update_tick = jiffies;
4501 __update_cpu_load(this_rq, load, 1);
4505 * Return a low guess at the load of a migration-source cpu weighted
4506 * according to the scheduling class and "nice" value.
4508 * We want to under-estimate the load of migration sources, to
4509 * balance conservatively.
4511 static unsigned long source_load(int cpu, int type)
4513 struct rq *rq = cpu_rq(cpu);
4514 unsigned long total = weighted_cpuload(cpu);
4516 if (type == 0 || !sched_feat(LB_BIAS))
4519 return min(rq->cpu_load[type-1], total);
4523 * Return a high guess at the load of a migration-target cpu weighted
4524 * according to the scheduling class and "nice" value.
4526 static unsigned long target_load(int cpu, int type)
4528 struct rq *rq = cpu_rq(cpu);
4529 unsigned long total = weighted_cpuload(cpu);
4531 if (type == 0 || !sched_feat(LB_BIAS))
4534 return max(rq->cpu_load[type-1], total);
4538 static unsigned long cpu_avg_load_per_task(int cpu)
4540 struct rq *rq = cpu_rq(cpu);
4541 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4542 unsigned long load_avg = weighted_cpuload(cpu);
4545 return load_avg / nr_running;
4550 static void record_wakee(struct task_struct *p)
4553 * Rough decay (wiping) for cost saving, don't worry
4554 * about the boundary, really active task won't care
4557 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4558 current->wakee_flips >>= 1;
4559 current->wakee_flip_decay_ts = jiffies;
4562 if (current->last_wakee != p) {
4563 current->last_wakee = p;
4564 current->wakee_flips++;
4568 static void task_waking_fair(struct task_struct *p)
4570 struct sched_entity *se = &p->se;
4571 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4574 #ifndef CONFIG_64BIT
4575 u64 min_vruntime_copy;
4578 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4580 min_vruntime = cfs_rq->min_vruntime;
4581 } while (min_vruntime != min_vruntime_copy);
4583 min_vruntime = cfs_rq->min_vruntime;
4586 se->vruntime -= min_vruntime;
4590 #ifdef CONFIG_FAIR_GROUP_SCHED
4592 * effective_load() calculates the load change as seen from the root_task_group
4594 * Adding load to a group doesn't make a group heavier, but can cause movement
4595 * of group shares between cpus. Assuming the shares were perfectly aligned one
4596 * can calculate the shift in shares.
4598 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4599 * on this @cpu and results in a total addition (subtraction) of @wg to the
4600 * total group weight.
4602 * Given a runqueue weight distribution (rw_i) we can compute a shares
4603 * distribution (s_i) using:
4605 * s_i = rw_i / \Sum rw_j (1)
4607 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4608 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4609 * shares distribution (s_i):
4611 * rw_i = { 2, 4, 1, 0 }
4612 * s_i = { 2/7, 4/7, 1/7, 0 }
4614 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4615 * task used to run on and the CPU the waker is running on), we need to
4616 * compute the effect of waking a task on either CPU and, in case of a sync
4617 * wakeup, compute the effect of the current task going to sleep.
4619 * So for a change of @wl to the local @cpu with an overall group weight change
4620 * of @wl we can compute the new shares distribution (s'_i) using:
4622 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4624 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4625 * differences in waking a task to CPU 0. The additional task changes the
4626 * weight and shares distributions like:
4628 * rw'_i = { 3, 4, 1, 0 }
4629 * s'_i = { 3/8, 4/8, 1/8, 0 }
4631 * We can then compute the difference in effective weight by using:
4633 * dw_i = S * (s'_i - s_i) (3)
4635 * Where 'S' is the group weight as seen by its parent.
4637 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4638 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4639 * 4/7) times the weight of the group.
4641 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4643 struct sched_entity *se = tg->se[cpu];
4645 if (!tg->parent) /* the trivial, non-cgroup case */
4648 for_each_sched_entity(se) {
4649 struct cfs_rq *cfs_rq = se->my_q;
4650 long W, w = cfs_rq_load_avg(cfs_rq);
4655 * W = @wg + \Sum rw_j
4657 W = wg + atomic_long_read(&tg->load_avg);
4659 /* Ensure \Sum rw_j >= rw_i */
4660 W -= cfs_rq->tg_load_avg_contrib;
4669 * wl = S * s'_i; see (2)
4672 wl = (w * (long)tg->shares) / W;
4677 * Per the above, wl is the new se->load.weight value; since
4678 * those are clipped to [MIN_SHARES, ...) do so now. See
4679 * calc_cfs_shares().
4681 if (wl < MIN_SHARES)
4685 * wl = dw_i = S * (s'_i - s_i); see (3)
4687 wl -= se->avg.load_avg;
4690 * Recursively apply this logic to all parent groups to compute
4691 * the final effective load change on the root group. Since
4692 * only the @tg group gets extra weight, all parent groups can
4693 * only redistribute existing shares. @wl is the shift in shares
4694 * resulting from this level per the above.
4703 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4710 static inline bool energy_aware(void)
4712 return sched_feat(ENERGY_AWARE);
4716 struct sched_group *sg_top;
4717 struct sched_group *sg_cap;
4726 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4727 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4728 * energy calculations. Using the scale-invariant util returned by
4729 * cpu_util() and approximating scale-invariant util by:
4731 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4733 * the normalized util can be found using the specific capacity.
4735 * capacity = capacity_orig * curr_freq/max_freq
4737 * norm_util = running_time/time ~ util/capacity
4739 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4741 int util = __cpu_util(cpu, delta);
4743 if (util >= capacity)
4744 return SCHED_CAPACITY_SCALE;
4746 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4749 static int calc_util_delta(struct energy_env *eenv, int cpu)
4751 if (cpu == eenv->src_cpu)
4752 return -eenv->util_delta;
4753 if (cpu == eenv->dst_cpu)
4754 return eenv->util_delta;
4759 unsigned long group_max_util(struct energy_env *eenv)
4762 unsigned long max_util = 0;
4764 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4765 delta = calc_util_delta(eenv, i);
4766 max_util = max(max_util, __cpu_util(i, delta));
4773 * group_norm_util() returns the approximated group util relative to it's
4774 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4775 * energy calculations. Since task executions may or may not overlap in time in
4776 * the group the true normalized util is between max(cpu_norm_util(i)) and
4777 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4778 * latter is used as the estimate as it leads to a more pessimistic energy
4779 * estimate (more busy).
4782 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4785 unsigned long util_sum = 0;
4786 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4788 for_each_cpu(i, sched_group_cpus(sg)) {
4789 delta = calc_util_delta(eenv, i);
4790 util_sum += __cpu_norm_util(i, capacity, delta);
4793 if (util_sum > SCHED_CAPACITY_SCALE)
4794 return SCHED_CAPACITY_SCALE;
4798 static int find_new_capacity(struct energy_env *eenv,
4799 const struct sched_group_energy const *sge)
4802 unsigned long util = group_max_util(eenv);
4804 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4805 if (sge->cap_states[idx].cap >= util)
4809 eenv->cap_idx = idx;
4814 static int group_idle_state(struct sched_group *sg)
4816 int i, state = INT_MAX;
4818 /* Find the shallowest idle state in the sched group. */
4819 for_each_cpu(i, sched_group_cpus(sg))
4820 state = min(state, idle_get_state_idx(cpu_rq(i)));
4822 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4829 * sched_group_energy(): Computes the absolute energy consumption of cpus
4830 * belonging to the sched_group including shared resources shared only by
4831 * members of the group. Iterates over all cpus in the hierarchy below the
4832 * sched_group starting from the bottom working it's way up before going to
4833 * the next cpu until all cpus are covered at all levels. The current
4834 * implementation is likely to gather the same util statistics multiple times.
4835 * This can probably be done in a faster but more complex way.
4836 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4838 static int sched_group_energy(struct energy_env *eenv)
4840 struct sched_domain *sd;
4841 int cpu, total_energy = 0;
4842 struct cpumask visit_cpus;
4843 struct sched_group *sg;
4845 WARN_ON(!eenv->sg_top->sge);
4847 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4849 while (!cpumask_empty(&visit_cpus)) {
4850 struct sched_group *sg_shared_cap = NULL;
4852 cpu = cpumask_first(&visit_cpus);
4855 * Is the group utilization affected by cpus outside this
4858 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4862 * We most probably raced with hotplug; returning a
4863 * wrong energy estimation is better than entering an
4869 sg_shared_cap = sd->parent->groups;
4871 for_each_domain(cpu, sd) {
4874 /* Has this sched_domain already been visited? */
4875 if (sd->child && group_first_cpu(sg) != cpu)
4879 unsigned long group_util;
4880 int sg_busy_energy, sg_idle_energy;
4881 int cap_idx, idle_idx;
4883 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
4884 eenv->sg_cap = sg_shared_cap;
4888 cap_idx = find_new_capacity(eenv, sg->sge);
4889 idle_idx = group_idle_state(sg);
4890 group_util = group_norm_util(eenv, sg);
4891 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
4892 >> SCHED_CAPACITY_SHIFT;
4893 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
4894 * sg->sge->idle_states[idle_idx].power)
4895 >> SCHED_CAPACITY_SHIFT;
4897 total_energy += sg_busy_energy + sg_idle_energy;
4900 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
4902 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
4905 } while (sg = sg->next, sg != sd->groups);
4911 eenv->energy = total_energy;
4915 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
4917 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
4921 * energy_diff(): Estimate the energy impact of changing the utilization
4922 * distribution. eenv specifies the change: utilisation amount, source, and
4923 * destination cpu. Source or destination cpu may be -1 in which case the
4924 * utilization is removed from or added to the system (e.g. task wake-up). If
4925 * both are specified, the utilization is migrated.
4927 static int energy_diff(struct energy_env *eenv)
4929 struct sched_domain *sd;
4930 struct sched_group *sg;
4931 int sd_cpu = -1, energy_before = 0, energy_after = 0;
4933 struct energy_env eenv_before = {
4935 .src_cpu = eenv->src_cpu,
4936 .dst_cpu = eenv->dst_cpu,
4939 if (eenv->src_cpu == eenv->dst_cpu)
4942 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
4943 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
4946 return 0; /* Error */
4951 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
4952 eenv_before.sg_top = eenv->sg_top = sg;
4954 if (sched_group_energy(&eenv_before))
4955 return 0; /* Invalid result abort */
4956 energy_before += eenv_before.energy;
4958 if (sched_group_energy(eenv))
4959 return 0; /* Invalid result abort */
4960 energy_after += eenv->energy;
4962 } while (sg = sg->next, sg != sd->groups);
4964 return energy_after-energy_before;
4968 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4969 * A waker of many should wake a different task than the one last awakened
4970 * at a frequency roughly N times higher than one of its wakees. In order
4971 * to determine whether we should let the load spread vs consolodating to
4972 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4973 * partner, and a factor of lls_size higher frequency in the other. With
4974 * both conditions met, we can be relatively sure that the relationship is
4975 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4976 * being client/server, worker/dispatcher, interrupt source or whatever is
4977 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4979 static int wake_wide(struct task_struct *p)
4981 unsigned int master = current->wakee_flips;
4982 unsigned int slave = p->wakee_flips;
4983 int factor = this_cpu_read(sd_llc_size);
4986 swap(master, slave);
4987 if (slave < factor || master < slave * factor)
4992 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4994 s64 this_load, load;
4995 s64 this_eff_load, prev_eff_load;
4996 int idx, this_cpu, prev_cpu;
4997 struct task_group *tg;
4998 unsigned long weight;
5002 this_cpu = smp_processor_id();
5003 prev_cpu = task_cpu(p);
5004 load = source_load(prev_cpu, idx);
5005 this_load = target_load(this_cpu, idx);
5008 * If sync wakeup then subtract the (maximum possible)
5009 * effect of the currently running task from the load
5010 * of the current CPU:
5013 tg = task_group(current);
5014 weight = current->se.avg.load_avg;
5016 this_load += effective_load(tg, this_cpu, -weight, -weight);
5017 load += effective_load(tg, prev_cpu, 0, -weight);
5021 weight = p->se.avg.load_avg;
5024 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5025 * due to the sync cause above having dropped this_load to 0, we'll
5026 * always have an imbalance, but there's really nothing you can do
5027 * about that, so that's good too.
5029 * Otherwise check if either cpus are near enough in load to allow this
5030 * task to be woken on this_cpu.
5032 this_eff_load = 100;
5033 this_eff_load *= capacity_of(prev_cpu);
5035 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5036 prev_eff_load *= capacity_of(this_cpu);
5038 if (this_load > 0) {
5039 this_eff_load *= this_load +
5040 effective_load(tg, this_cpu, weight, weight);
5042 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5045 balanced = this_eff_load <= prev_eff_load;
5047 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5052 schedstat_inc(sd, ttwu_move_affine);
5053 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5058 static inline unsigned long task_util(struct task_struct *p)
5060 return p->se.avg.util_avg;
5063 unsigned int capacity_margin = 1280; /* ~20% margin */
5065 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5067 unsigned long capacity = capacity_of(cpu);
5069 util += task_util(p);
5071 return (capacity * 1024) > (util * capacity_margin);
5074 static inline bool task_fits_max(struct task_struct *p, int cpu)
5076 unsigned long capacity = capacity_of(cpu);
5077 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5079 if (capacity == max_capacity)
5082 if (capacity * capacity_margin > max_capacity * 1024)
5085 return __task_fits(p, cpu, 0);
5088 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5090 return __task_fits(p, cpu, cpu_util(cpu));
5093 static bool cpu_overutilized(int cpu)
5095 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5099 * find_idlest_group finds and returns the least busy CPU group within the
5102 static struct sched_group *
5103 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5104 int this_cpu, int sd_flag)
5106 struct sched_group *idlest = NULL, *group = sd->groups;
5107 struct sched_group *fit_group = NULL, *spare_group = NULL;
5108 unsigned long min_load = ULONG_MAX, this_load = 0;
5109 unsigned long fit_capacity = ULONG_MAX;
5110 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5111 int load_idx = sd->forkexec_idx;
5112 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5114 if (sd_flag & SD_BALANCE_WAKE)
5115 load_idx = sd->wake_idx;
5118 unsigned long load, avg_load, spare_capacity;
5122 /* Skip over this group if it has no CPUs allowed */
5123 if (!cpumask_intersects(sched_group_cpus(group),
5124 tsk_cpus_allowed(p)))
5127 local_group = cpumask_test_cpu(this_cpu,
5128 sched_group_cpus(group));
5130 /* Tally up the load of all CPUs in the group */
5133 for_each_cpu(i, sched_group_cpus(group)) {
5134 /* Bias balancing toward cpus of our domain */
5136 load = source_load(i, load_idx);
5138 load = target_load(i, load_idx);
5143 * Look for most energy-efficient group that can fit
5144 * that can fit the task.
5146 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5147 fit_capacity = capacity_of(i);
5152 * Look for group which has most spare capacity on a
5155 spare_capacity = capacity_of(i) - cpu_util(i);
5156 if (spare_capacity > max_spare_capacity) {
5157 max_spare_capacity = spare_capacity;
5158 spare_group = group;
5162 /* Adjust by relative CPU capacity of the group */
5163 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5166 this_load = avg_load;
5167 } else if (avg_load < min_load) {
5168 min_load = avg_load;
5171 } while (group = group->next, group != sd->groups);
5179 if (!idlest || 100*this_load < imbalance*min_load)
5185 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5188 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5190 unsigned long load, min_load = ULONG_MAX;
5191 unsigned int min_exit_latency = UINT_MAX;
5192 u64 latest_idle_timestamp = 0;
5193 int least_loaded_cpu = this_cpu;
5194 int shallowest_idle_cpu = -1;
5197 /* Traverse only the allowed CPUs */
5198 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5199 if (task_fits_spare(p, i)) {
5200 struct rq *rq = cpu_rq(i);
5201 struct cpuidle_state *idle = idle_get_state(rq);
5202 if (idle && idle->exit_latency < min_exit_latency) {
5204 * We give priority to a CPU whose idle state
5205 * has the smallest exit latency irrespective
5206 * of any idle timestamp.
5208 min_exit_latency = idle->exit_latency;
5209 latest_idle_timestamp = rq->idle_stamp;
5210 shallowest_idle_cpu = i;
5211 } else if (idle_cpu(i) &&
5212 (!idle || idle->exit_latency == min_exit_latency) &&
5213 rq->idle_stamp > latest_idle_timestamp) {
5215 * If equal or no active idle state, then
5216 * the most recently idled CPU might have
5219 latest_idle_timestamp = rq->idle_stamp;
5220 shallowest_idle_cpu = i;
5221 } else if (shallowest_idle_cpu == -1) {
5223 * If we haven't found an idle CPU yet
5224 * pick a non-idle one that can fit the task as
5227 shallowest_idle_cpu = i;
5229 } else if (shallowest_idle_cpu == -1) {
5230 load = weighted_cpuload(i);
5231 if (load < min_load || (load == min_load && i == this_cpu)) {
5233 least_loaded_cpu = i;
5238 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5242 * Try and locate an idle CPU in the sched_domain.
5244 static int select_idle_sibling(struct task_struct *p, int target)
5246 struct sched_domain *sd;
5247 struct sched_group *sg;
5248 int i = task_cpu(p);
5250 if (idle_cpu(target))
5254 * If the prevous cpu is cache affine and idle, don't be stupid.
5256 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5260 * Otherwise, iterate the domains and find an elegible idle cpu.
5262 sd = rcu_dereference(per_cpu(sd_llc, target));
5263 for_each_lower_domain(sd) {
5266 if (!cpumask_intersects(sched_group_cpus(sg),
5267 tsk_cpus_allowed(p)))
5270 for_each_cpu(i, sched_group_cpus(sg)) {
5271 if (i == target || !idle_cpu(i))
5275 target = cpumask_first_and(sched_group_cpus(sg),
5276 tsk_cpus_allowed(p));
5280 } while (sg != sd->groups);
5286 static int energy_aware_wake_cpu(struct task_struct *p, int target)
5288 struct sched_domain *sd;
5289 struct sched_group *sg, *sg_target;
5290 int target_max_cap = INT_MAX;
5291 int target_cpu = task_cpu(p);
5294 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5303 * Find group with sufficient capacity. We only get here if no cpu is
5304 * overutilized. We may end up overutilizing a cpu by adding the task,
5305 * but that should not be any worse than select_idle_sibling().
5306 * load_balance() should sort it out later as we get above the tipping
5310 /* Assuming all cpus are the same in group */
5311 int max_cap_cpu = group_first_cpu(sg);
5314 * Assume smaller max capacity means more energy-efficient.
5315 * Ideally we should query the energy model for the right
5316 * answer but it easily ends up in an exhaustive search.
5318 if (capacity_of(max_cap_cpu) < target_max_cap &&
5319 task_fits_max(p, max_cap_cpu)) {
5321 target_max_cap = capacity_of(max_cap_cpu);
5323 } while (sg = sg->next, sg != sd->groups);
5325 /* Find cpu with sufficient capacity */
5326 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5328 * p's blocked utilization is still accounted for on prev_cpu
5329 * so prev_cpu will receive a negative bias due to the double
5330 * accounting. However, the blocked utilization may be zero.
5332 int new_util = cpu_util(i) + task_util(p);
5334 if (new_util > capacity_orig_of(i))
5337 if (new_util < capacity_curr_of(i)) {
5339 if (cpu_rq(i)->nr_running)
5343 /* cpu has capacity at higher OPP, keep it as fallback */
5344 if (target_cpu == task_cpu(p))
5348 if (target_cpu != task_cpu(p)) {
5349 struct energy_env eenv = {
5350 .util_delta = task_util(p),
5351 .src_cpu = task_cpu(p),
5352 .dst_cpu = target_cpu,
5355 /* Not enough spare capacity on previous cpu */
5356 if (cpu_overutilized(task_cpu(p)))
5359 if (energy_diff(&eenv) >= 0)
5367 * select_task_rq_fair: Select target runqueue for the waking task in domains
5368 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5369 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5371 * Balances load by selecting the idlest cpu in the idlest group, or under
5372 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5374 * Returns the target cpu number.
5376 * preempt must be disabled.
5379 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5381 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5382 int cpu = smp_processor_id();
5383 int new_cpu = prev_cpu;
5384 int want_affine = 0;
5385 int sync = wake_flags & WF_SYNC;
5387 if (sd_flag & SD_BALANCE_WAKE)
5388 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5389 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5393 for_each_domain(cpu, tmp) {
5394 if (!(tmp->flags & SD_LOAD_BALANCE))
5398 * If both cpu and prev_cpu are part of this domain,
5399 * cpu is a valid SD_WAKE_AFFINE target.
5401 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5402 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5407 if (tmp->flags & sd_flag)
5409 else if (!want_affine)
5414 sd = NULL; /* Prefer wake_affine over balance flags */
5415 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5420 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5421 new_cpu = energy_aware_wake_cpu(p, prev_cpu);
5422 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5423 new_cpu = select_idle_sibling(p, new_cpu);
5426 struct sched_group *group;
5429 if (!(sd->flags & sd_flag)) {
5434 group = find_idlest_group(sd, p, cpu, sd_flag);
5440 new_cpu = find_idlest_cpu(group, p, cpu);
5441 if (new_cpu == -1 || new_cpu == cpu) {
5442 /* Now try balancing at a lower domain level of cpu */
5447 /* Now try balancing at a lower domain level of new_cpu */
5449 weight = sd->span_weight;
5451 for_each_domain(cpu, tmp) {
5452 if (weight <= tmp->span_weight)
5454 if (tmp->flags & sd_flag)
5457 /* while loop will break here if sd == NULL */
5465 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5466 * cfs_rq_of(p) references at time of call are still valid and identify the
5467 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5468 * other assumptions, including the state of rq->lock, should be made.
5470 static void migrate_task_rq_fair(struct task_struct *p)
5473 * We are supposed to update the task to "current" time, then its up to date
5474 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5475 * what current time is, so simply throw away the out-of-date time. This
5476 * will result in the wakee task is less decayed, but giving the wakee more
5477 * load sounds not bad.
5479 remove_entity_load_avg(&p->se);
5481 /* Tell new CPU we are migrated */
5482 p->se.avg.last_update_time = 0;
5484 /* We have migrated, no longer consider this task hot */
5485 p->se.exec_start = 0;
5488 static void task_dead_fair(struct task_struct *p)
5490 remove_entity_load_avg(&p->se);
5492 #endif /* CONFIG_SMP */
5494 static unsigned long
5495 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5497 unsigned long gran = sysctl_sched_wakeup_granularity;
5500 * Since its curr running now, convert the gran from real-time
5501 * to virtual-time in his units.
5503 * By using 'se' instead of 'curr' we penalize light tasks, so
5504 * they get preempted easier. That is, if 'se' < 'curr' then
5505 * the resulting gran will be larger, therefore penalizing the
5506 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5507 * be smaller, again penalizing the lighter task.
5509 * This is especially important for buddies when the leftmost
5510 * task is higher priority than the buddy.
5512 return calc_delta_fair(gran, se);
5516 * Should 'se' preempt 'curr'.
5530 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5532 s64 gran, vdiff = curr->vruntime - se->vruntime;
5537 gran = wakeup_gran(curr, se);
5544 static void set_last_buddy(struct sched_entity *se)
5546 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5549 for_each_sched_entity(se)
5550 cfs_rq_of(se)->last = se;
5553 static void set_next_buddy(struct sched_entity *se)
5555 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5558 for_each_sched_entity(se)
5559 cfs_rq_of(se)->next = se;
5562 static void set_skip_buddy(struct sched_entity *se)
5564 for_each_sched_entity(se)
5565 cfs_rq_of(se)->skip = se;
5569 * Preempt the current task with a newly woken task if needed:
5571 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5573 struct task_struct *curr = rq->curr;
5574 struct sched_entity *se = &curr->se, *pse = &p->se;
5575 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5576 int scale = cfs_rq->nr_running >= sched_nr_latency;
5577 int next_buddy_marked = 0;
5579 if (unlikely(se == pse))
5583 * This is possible from callers such as attach_tasks(), in which we
5584 * unconditionally check_prempt_curr() after an enqueue (which may have
5585 * lead to a throttle). This both saves work and prevents false
5586 * next-buddy nomination below.
5588 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5591 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5592 set_next_buddy(pse);
5593 next_buddy_marked = 1;
5597 * We can come here with TIF_NEED_RESCHED already set from new task
5600 * Note: this also catches the edge-case of curr being in a throttled
5601 * group (e.g. via set_curr_task), since update_curr() (in the
5602 * enqueue of curr) will have resulted in resched being set. This
5603 * prevents us from potentially nominating it as a false LAST_BUDDY
5606 if (test_tsk_need_resched(curr))
5609 /* Idle tasks are by definition preempted by non-idle tasks. */
5610 if (unlikely(curr->policy == SCHED_IDLE) &&
5611 likely(p->policy != SCHED_IDLE))
5615 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5616 * is driven by the tick):
5618 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5621 find_matching_se(&se, &pse);
5622 update_curr(cfs_rq_of(se));
5624 if (wakeup_preempt_entity(se, pse) == 1) {
5626 * Bias pick_next to pick the sched entity that is
5627 * triggering this preemption.
5629 if (!next_buddy_marked)
5630 set_next_buddy(pse);
5639 * Only set the backward buddy when the current task is still
5640 * on the rq. This can happen when a wakeup gets interleaved
5641 * with schedule on the ->pre_schedule() or idle_balance()
5642 * point, either of which can * drop the rq lock.
5644 * Also, during early boot the idle thread is in the fair class,
5645 * for obvious reasons its a bad idea to schedule back to it.
5647 if (unlikely(!se->on_rq || curr == rq->idle))
5650 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5654 static struct task_struct *
5655 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5657 struct cfs_rq *cfs_rq = &rq->cfs;
5658 struct sched_entity *se;
5659 struct task_struct *p;
5663 #ifdef CONFIG_FAIR_GROUP_SCHED
5664 if (!cfs_rq->nr_running)
5667 if (prev->sched_class != &fair_sched_class)
5671 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5672 * likely that a next task is from the same cgroup as the current.
5674 * Therefore attempt to avoid putting and setting the entire cgroup
5675 * hierarchy, only change the part that actually changes.
5679 struct sched_entity *curr = cfs_rq->curr;
5682 * Since we got here without doing put_prev_entity() we also
5683 * have to consider cfs_rq->curr. If it is still a runnable
5684 * entity, update_curr() will update its vruntime, otherwise
5685 * forget we've ever seen it.
5689 update_curr(cfs_rq);
5694 * This call to check_cfs_rq_runtime() will do the
5695 * throttle and dequeue its entity in the parent(s).
5696 * Therefore the 'simple' nr_running test will indeed
5699 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5703 se = pick_next_entity(cfs_rq, curr);
5704 cfs_rq = group_cfs_rq(se);
5710 * Since we haven't yet done put_prev_entity and if the selected task
5711 * is a different task than we started out with, try and touch the
5712 * least amount of cfs_rqs.
5715 struct sched_entity *pse = &prev->se;
5717 while (!(cfs_rq = is_same_group(se, pse))) {
5718 int se_depth = se->depth;
5719 int pse_depth = pse->depth;
5721 if (se_depth <= pse_depth) {
5722 put_prev_entity(cfs_rq_of(pse), pse);
5723 pse = parent_entity(pse);
5725 if (se_depth >= pse_depth) {
5726 set_next_entity(cfs_rq_of(se), se);
5727 se = parent_entity(se);
5731 put_prev_entity(cfs_rq, pse);
5732 set_next_entity(cfs_rq, se);
5735 if (hrtick_enabled(rq))
5736 hrtick_start_fair(rq, p);
5738 rq->misfit_task = !task_fits_max(p, rq->cpu);
5745 if (!cfs_rq->nr_running)
5748 put_prev_task(rq, prev);
5751 se = pick_next_entity(cfs_rq, NULL);
5752 set_next_entity(cfs_rq, se);
5753 cfs_rq = group_cfs_rq(se);
5758 if (hrtick_enabled(rq))
5759 hrtick_start_fair(rq, p);
5761 rq->misfit_task = !task_fits_max(p, rq->cpu);
5766 rq->misfit_task = 0;
5768 * This is OK, because current is on_cpu, which avoids it being picked
5769 * for load-balance and preemption/IRQs are still disabled avoiding
5770 * further scheduler activity on it and we're being very careful to
5771 * re-start the picking loop.
5773 lockdep_unpin_lock(&rq->lock);
5774 new_tasks = idle_balance(rq);
5775 lockdep_pin_lock(&rq->lock);
5777 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5778 * possible for any higher priority task to appear. In that case we
5779 * must re-start the pick_next_entity() loop.
5791 * Account for a descheduled task:
5793 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5795 struct sched_entity *se = &prev->se;
5796 struct cfs_rq *cfs_rq;
5798 for_each_sched_entity(se) {
5799 cfs_rq = cfs_rq_of(se);
5800 put_prev_entity(cfs_rq, se);
5805 * sched_yield() is very simple
5807 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5809 static void yield_task_fair(struct rq *rq)
5811 struct task_struct *curr = rq->curr;
5812 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5813 struct sched_entity *se = &curr->se;
5816 * Are we the only task in the tree?
5818 if (unlikely(rq->nr_running == 1))
5821 clear_buddies(cfs_rq, se);
5823 if (curr->policy != SCHED_BATCH) {
5824 update_rq_clock(rq);
5826 * Update run-time statistics of the 'current'.
5828 update_curr(cfs_rq);
5830 * Tell update_rq_clock() that we've just updated,
5831 * so we don't do microscopic update in schedule()
5832 * and double the fastpath cost.
5834 rq_clock_skip_update(rq, true);
5840 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5842 struct sched_entity *se = &p->se;
5844 /* throttled hierarchies are not runnable */
5845 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5848 /* Tell the scheduler that we'd really like pse to run next. */
5851 yield_task_fair(rq);
5857 /**************************************************
5858 * Fair scheduling class load-balancing methods.
5862 * The purpose of load-balancing is to achieve the same basic fairness the
5863 * per-cpu scheduler provides, namely provide a proportional amount of compute
5864 * time to each task. This is expressed in the following equation:
5866 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5868 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5869 * W_i,0 is defined as:
5871 * W_i,0 = \Sum_j w_i,j (2)
5873 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5874 * is derived from the nice value as per prio_to_weight[].
5876 * The weight average is an exponential decay average of the instantaneous
5879 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5881 * C_i is the compute capacity of cpu i, typically it is the
5882 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5883 * can also include other factors [XXX].
5885 * To achieve this balance we define a measure of imbalance which follows
5886 * directly from (1):
5888 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5890 * We them move tasks around to minimize the imbalance. In the continuous
5891 * function space it is obvious this converges, in the discrete case we get
5892 * a few fun cases generally called infeasible weight scenarios.
5895 * - infeasible weights;
5896 * - local vs global optima in the discrete case. ]
5901 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5902 * for all i,j solution, we create a tree of cpus that follows the hardware
5903 * topology where each level pairs two lower groups (or better). This results
5904 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5905 * tree to only the first of the previous level and we decrease the frequency
5906 * of load-balance at each level inv. proportional to the number of cpus in
5912 * \Sum { --- * --- * 2^i } = O(n) (5)
5914 * `- size of each group
5915 * | | `- number of cpus doing load-balance
5917 * `- sum over all levels
5919 * Coupled with a limit on how many tasks we can migrate every balance pass,
5920 * this makes (5) the runtime complexity of the balancer.
5922 * An important property here is that each CPU is still (indirectly) connected
5923 * to every other cpu in at most O(log n) steps:
5925 * The adjacency matrix of the resulting graph is given by:
5928 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5931 * And you'll find that:
5933 * A^(log_2 n)_i,j != 0 for all i,j (7)
5935 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5936 * The task movement gives a factor of O(m), giving a convergence complexity
5939 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5944 * In order to avoid CPUs going idle while there's still work to do, new idle
5945 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5946 * tree itself instead of relying on other CPUs to bring it work.
5948 * This adds some complexity to both (5) and (8) but it reduces the total idle
5956 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5959 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5964 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5966 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5968 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5971 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5972 * rewrite all of this once again.]
5975 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5977 enum fbq_type { regular, remote, all };
5986 #define LBF_ALL_PINNED 0x01
5987 #define LBF_NEED_BREAK 0x02
5988 #define LBF_DST_PINNED 0x04
5989 #define LBF_SOME_PINNED 0x08
5992 struct sched_domain *sd;
6000 struct cpumask *dst_grpmask;
6002 enum cpu_idle_type idle;
6004 unsigned int src_grp_nr_running;
6005 /* The set of CPUs under consideration for load-balancing */
6006 struct cpumask *cpus;
6011 unsigned int loop_break;
6012 unsigned int loop_max;
6014 enum fbq_type fbq_type;
6015 enum group_type busiest_group_type;
6016 struct list_head tasks;
6020 * Is this task likely cache-hot:
6022 static int task_hot(struct task_struct *p, struct lb_env *env)
6026 lockdep_assert_held(&env->src_rq->lock);
6028 if (p->sched_class != &fair_sched_class)
6031 if (unlikely(p->policy == SCHED_IDLE))
6035 * Buddy candidates are cache hot:
6037 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6038 (&p->se == cfs_rq_of(&p->se)->next ||
6039 &p->se == cfs_rq_of(&p->se)->last))
6042 if (sysctl_sched_migration_cost == -1)
6044 if (sysctl_sched_migration_cost == 0)
6047 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6049 return delta < (s64)sysctl_sched_migration_cost;
6052 #ifdef CONFIG_NUMA_BALANCING
6054 * Returns 1, if task migration degrades locality
6055 * Returns 0, if task migration improves locality i.e migration preferred.
6056 * Returns -1, if task migration is not affected by locality.
6058 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6060 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6061 unsigned long src_faults, dst_faults;
6062 int src_nid, dst_nid;
6064 if (!static_branch_likely(&sched_numa_balancing))
6067 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6070 src_nid = cpu_to_node(env->src_cpu);
6071 dst_nid = cpu_to_node(env->dst_cpu);
6073 if (src_nid == dst_nid)
6076 /* Migrating away from the preferred node is always bad. */
6077 if (src_nid == p->numa_preferred_nid) {
6078 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6084 /* Encourage migration to the preferred node. */
6085 if (dst_nid == p->numa_preferred_nid)
6089 src_faults = group_faults(p, src_nid);
6090 dst_faults = group_faults(p, dst_nid);
6092 src_faults = task_faults(p, src_nid);
6093 dst_faults = task_faults(p, dst_nid);
6096 return dst_faults < src_faults;
6100 static inline int migrate_degrades_locality(struct task_struct *p,
6108 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6111 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6115 lockdep_assert_held(&env->src_rq->lock);
6118 * We do not migrate tasks that are:
6119 * 1) throttled_lb_pair, or
6120 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6121 * 3) running (obviously), or
6122 * 4) are cache-hot on their current CPU.
6124 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6127 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6130 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6132 env->flags |= LBF_SOME_PINNED;
6135 * Remember if this task can be migrated to any other cpu in
6136 * our sched_group. We may want to revisit it if we couldn't
6137 * meet load balance goals by pulling other tasks on src_cpu.
6139 * Also avoid computing new_dst_cpu if we have already computed
6140 * one in current iteration.
6142 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6145 /* Prevent to re-select dst_cpu via env's cpus */
6146 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6147 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6148 env->flags |= LBF_DST_PINNED;
6149 env->new_dst_cpu = cpu;
6157 /* Record that we found atleast one task that could run on dst_cpu */
6158 env->flags &= ~LBF_ALL_PINNED;
6160 if (task_running(env->src_rq, p)) {
6161 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6166 * Aggressive migration if:
6167 * 1) destination numa is preferred
6168 * 2) task is cache cold, or
6169 * 3) too many balance attempts have failed.
6171 tsk_cache_hot = migrate_degrades_locality(p, env);
6172 if (tsk_cache_hot == -1)
6173 tsk_cache_hot = task_hot(p, env);
6175 if (tsk_cache_hot <= 0 ||
6176 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6177 if (tsk_cache_hot == 1) {
6178 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6179 schedstat_inc(p, se.statistics.nr_forced_migrations);
6184 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6189 * detach_task() -- detach the task for the migration specified in env
6191 static void detach_task(struct task_struct *p, struct lb_env *env)
6193 lockdep_assert_held(&env->src_rq->lock);
6195 deactivate_task(env->src_rq, p, 0);
6196 p->on_rq = TASK_ON_RQ_MIGRATING;
6197 set_task_cpu(p, env->dst_cpu);
6201 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6202 * part of active balancing operations within "domain".
6204 * Returns a task if successful and NULL otherwise.
6206 static struct task_struct *detach_one_task(struct lb_env *env)
6208 struct task_struct *p, *n;
6210 lockdep_assert_held(&env->src_rq->lock);
6212 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6213 if (!can_migrate_task(p, env))
6216 detach_task(p, env);
6219 * Right now, this is only the second place where
6220 * lb_gained[env->idle] is updated (other is detach_tasks)
6221 * so we can safely collect stats here rather than
6222 * inside detach_tasks().
6224 schedstat_inc(env->sd, lb_gained[env->idle]);
6230 static const unsigned int sched_nr_migrate_break = 32;
6233 * detach_tasks() -- tries to detach up to imbalance weighted load from
6234 * busiest_rq, as part of a balancing operation within domain "sd".
6236 * Returns number of detached tasks if successful and 0 otherwise.
6238 static int detach_tasks(struct lb_env *env)
6240 struct list_head *tasks = &env->src_rq->cfs_tasks;
6241 struct task_struct *p;
6245 lockdep_assert_held(&env->src_rq->lock);
6247 if (env->imbalance <= 0)
6250 while (!list_empty(tasks)) {
6252 * We don't want to steal all, otherwise we may be treated likewise,
6253 * which could at worst lead to a livelock crash.
6255 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6258 p = list_first_entry(tasks, struct task_struct, se.group_node);
6261 /* We've more or less seen every task there is, call it quits */
6262 if (env->loop > env->loop_max)
6265 /* take a breather every nr_migrate tasks */
6266 if (env->loop > env->loop_break) {
6267 env->loop_break += sched_nr_migrate_break;
6268 env->flags |= LBF_NEED_BREAK;
6272 if (!can_migrate_task(p, env))
6275 load = task_h_load(p);
6277 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6280 if ((load / 2) > env->imbalance)
6283 detach_task(p, env);
6284 list_add(&p->se.group_node, &env->tasks);
6287 env->imbalance -= load;
6289 #ifdef CONFIG_PREEMPT
6291 * NEWIDLE balancing is a source of latency, so preemptible
6292 * kernels will stop after the first task is detached to minimize
6293 * the critical section.
6295 if (env->idle == CPU_NEWLY_IDLE)
6300 * We only want to steal up to the prescribed amount of
6303 if (env->imbalance <= 0)
6308 list_move_tail(&p->se.group_node, tasks);
6312 * Right now, this is one of only two places we collect this stat
6313 * so we can safely collect detach_one_task() stats here rather
6314 * than inside detach_one_task().
6316 schedstat_add(env->sd, lb_gained[env->idle], detached);
6322 * attach_task() -- attach the task detached by detach_task() to its new rq.
6324 static void attach_task(struct rq *rq, struct task_struct *p)
6326 lockdep_assert_held(&rq->lock);
6328 BUG_ON(task_rq(p) != rq);
6329 p->on_rq = TASK_ON_RQ_QUEUED;
6330 activate_task(rq, p, 0);
6331 check_preempt_curr(rq, p, 0);
6335 * attach_one_task() -- attaches the task returned from detach_one_task() to
6338 static void attach_one_task(struct rq *rq, struct task_struct *p)
6340 raw_spin_lock(&rq->lock);
6343 * We want to potentially raise target_cpu's OPP.
6345 update_capacity_of(cpu_of(rq));
6346 raw_spin_unlock(&rq->lock);
6350 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6353 static void attach_tasks(struct lb_env *env)
6355 struct list_head *tasks = &env->tasks;
6356 struct task_struct *p;
6358 raw_spin_lock(&env->dst_rq->lock);
6360 while (!list_empty(tasks)) {
6361 p = list_first_entry(tasks, struct task_struct, se.group_node);
6362 list_del_init(&p->se.group_node);
6364 attach_task(env->dst_rq, p);
6368 * We want to potentially raise env.dst_cpu's OPP.
6370 update_capacity_of(env->dst_cpu);
6372 raw_spin_unlock(&env->dst_rq->lock);
6375 #ifdef CONFIG_FAIR_GROUP_SCHED
6376 static void update_blocked_averages(int cpu)
6378 struct rq *rq = cpu_rq(cpu);
6379 struct cfs_rq *cfs_rq;
6380 unsigned long flags;
6382 raw_spin_lock_irqsave(&rq->lock, flags);
6383 update_rq_clock(rq);
6386 * Iterates the task_group tree in a bottom up fashion, see
6387 * list_add_leaf_cfs_rq() for details.
6389 for_each_leaf_cfs_rq(rq, cfs_rq) {
6390 /* throttled entities do not contribute to load */
6391 if (throttled_hierarchy(cfs_rq))
6394 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6395 update_tg_load_avg(cfs_rq, 0);
6397 raw_spin_unlock_irqrestore(&rq->lock, flags);
6401 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6402 * This needs to be done in a top-down fashion because the load of a child
6403 * group is a fraction of its parents load.
6405 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6407 struct rq *rq = rq_of(cfs_rq);
6408 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6409 unsigned long now = jiffies;
6412 if (cfs_rq->last_h_load_update == now)
6415 cfs_rq->h_load_next = NULL;
6416 for_each_sched_entity(se) {
6417 cfs_rq = cfs_rq_of(se);
6418 cfs_rq->h_load_next = se;
6419 if (cfs_rq->last_h_load_update == now)
6424 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6425 cfs_rq->last_h_load_update = now;
6428 while ((se = cfs_rq->h_load_next) != NULL) {
6429 load = cfs_rq->h_load;
6430 load = div64_ul(load * se->avg.load_avg,
6431 cfs_rq_load_avg(cfs_rq) + 1);
6432 cfs_rq = group_cfs_rq(se);
6433 cfs_rq->h_load = load;
6434 cfs_rq->last_h_load_update = now;
6438 static unsigned long task_h_load(struct task_struct *p)
6440 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6442 update_cfs_rq_h_load(cfs_rq);
6443 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6444 cfs_rq_load_avg(cfs_rq) + 1);
6447 static inline void update_blocked_averages(int cpu)
6449 struct rq *rq = cpu_rq(cpu);
6450 struct cfs_rq *cfs_rq = &rq->cfs;
6451 unsigned long flags;
6453 raw_spin_lock_irqsave(&rq->lock, flags);
6454 update_rq_clock(rq);
6455 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6456 raw_spin_unlock_irqrestore(&rq->lock, flags);
6459 static unsigned long task_h_load(struct task_struct *p)
6461 return p->se.avg.load_avg;
6465 /********** Helpers for find_busiest_group ************************/
6468 * sg_lb_stats - stats of a sched_group required for load_balancing
6470 struct sg_lb_stats {
6471 unsigned long avg_load; /*Avg load across the CPUs of the group */
6472 unsigned long group_load; /* Total load over the CPUs of the group */
6473 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6474 unsigned long load_per_task;
6475 unsigned long group_capacity;
6476 unsigned long group_util; /* Total utilization of the group */
6477 unsigned int sum_nr_running; /* Nr tasks running in the group */
6478 unsigned int idle_cpus;
6479 unsigned int group_weight;
6480 enum group_type group_type;
6481 int group_no_capacity;
6482 int group_misfit_task; /* A cpu has a task too big for its capacity */
6483 #ifdef CONFIG_NUMA_BALANCING
6484 unsigned int nr_numa_running;
6485 unsigned int nr_preferred_running;
6490 * sd_lb_stats - Structure to store the statistics of a sched_domain
6491 * during load balancing.
6493 struct sd_lb_stats {
6494 struct sched_group *busiest; /* Busiest group in this sd */
6495 struct sched_group *local; /* Local group in this sd */
6496 unsigned long total_load; /* Total load of all groups in sd */
6497 unsigned long total_capacity; /* Total capacity of all groups in sd */
6498 unsigned long avg_load; /* Average load across all groups in sd */
6500 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6501 struct sg_lb_stats local_stat; /* Statistics of the local group */
6504 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6507 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6508 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6509 * We must however clear busiest_stat::avg_load because
6510 * update_sd_pick_busiest() reads this before assignment.
6512 *sds = (struct sd_lb_stats){
6516 .total_capacity = 0UL,
6519 .sum_nr_running = 0,
6520 .group_type = group_other,
6526 * get_sd_load_idx - Obtain the load index for a given sched domain.
6527 * @sd: The sched_domain whose load_idx is to be obtained.
6528 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6530 * Return: The load index.
6532 static inline int get_sd_load_idx(struct sched_domain *sd,
6533 enum cpu_idle_type idle)
6539 load_idx = sd->busy_idx;
6542 case CPU_NEWLY_IDLE:
6543 load_idx = sd->newidle_idx;
6546 load_idx = sd->idle_idx;
6553 static unsigned long scale_rt_capacity(int cpu)
6555 struct rq *rq = cpu_rq(cpu);
6556 u64 total, used, age_stamp, avg;
6560 * Since we're reading these variables without serialization make sure
6561 * we read them once before doing sanity checks on them.
6563 age_stamp = READ_ONCE(rq->age_stamp);
6564 avg = READ_ONCE(rq->rt_avg);
6565 delta = __rq_clock_broken(rq) - age_stamp;
6567 if (unlikely(delta < 0))
6570 total = sched_avg_period() + delta;
6572 used = div_u64(avg, total);
6574 if (likely(used < SCHED_CAPACITY_SCALE))
6575 return SCHED_CAPACITY_SCALE - used;
6580 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
6582 raw_spin_lock_init(&mcc->lock);
6587 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6589 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6590 struct sched_group *sdg = sd->groups;
6591 struct max_cpu_capacity *mcc;
6592 unsigned long max_capacity;
6594 unsigned long flags;
6596 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6598 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
6600 raw_spin_lock_irqsave(&mcc->lock, flags);
6601 max_capacity = mcc->val;
6602 max_cap_cpu = mcc->cpu;
6604 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
6605 (max_capacity < capacity)) {
6606 mcc->val = capacity;
6608 #ifdef CONFIG_SCHED_DEBUG
6609 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6610 pr_info("CPU%d: update max cpu_capacity %lu\n", cpu, capacity);
6614 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6616 skip_unlock: __attribute__ ((unused));
6617 capacity *= scale_rt_capacity(cpu);
6618 capacity >>= SCHED_CAPACITY_SHIFT;
6623 cpu_rq(cpu)->cpu_capacity = capacity;
6624 sdg->sgc->capacity = capacity;
6625 sdg->sgc->max_capacity = capacity;
6628 void update_group_capacity(struct sched_domain *sd, int cpu)
6630 struct sched_domain *child = sd->child;
6631 struct sched_group *group, *sdg = sd->groups;
6632 unsigned long capacity, max_capacity;
6633 unsigned long interval;
6635 interval = msecs_to_jiffies(sd->balance_interval);
6636 interval = clamp(interval, 1UL, max_load_balance_interval);
6637 sdg->sgc->next_update = jiffies + interval;
6640 update_cpu_capacity(sd, cpu);
6647 if (child->flags & SD_OVERLAP) {
6649 * SD_OVERLAP domains cannot assume that child groups
6650 * span the current group.
6653 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6654 struct sched_group_capacity *sgc;
6655 struct rq *rq = cpu_rq(cpu);
6658 * build_sched_domains() -> init_sched_groups_capacity()
6659 * gets here before we've attached the domains to the
6662 * Use capacity_of(), which is set irrespective of domains
6663 * in update_cpu_capacity().
6665 * This avoids capacity from being 0 and
6666 * causing divide-by-zero issues on boot.
6668 if (unlikely(!rq->sd)) {
6669 capacity += capacity_of(cpu);
6671 sgc = rq->sd->groups->sgc;
6672 capacity += sgc->capacity;
6675 max_capacity = max(capacity, max_capacity);
6679 * !SD_OVERLAP domains can assume that child groups
6680 * span the current group.
6683 group = child->groups;
6685 struct sched_group_capacity *sgc = group->sgc;
6687 capacity += sgc->capacity;
6688 max_capacity = max(sgc->max_capacity, max_capacity);
6689 group = group->next;
6690 } while (group != child->groups);
6693 sdg->sgc->capacity = capacity;
6694 sdg->sgc->max_capacity = max_capacity;
6698 * Check whether the capacity of the rq has been noticeably reduced by side
6699 * activity. The imbalance_pct is used for the threshold.
6700 * Return true is the capacity is reduced
6703 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6705 return ((rq->cpu_capacity * sd->imbalance_pct) <
6706 (rq->cpu_capacity_orig * 100));
6710 * Group imbalance indicates (and tries to solve) the problem where balancing
6711 * groups is inadequate due to tsk_cpus_allowed() constraints.
6713 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6714 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6717 * { 0 1 2 3 } { 4 5 6 7 }
6720 * If we were to balance group-wise we'd place two tasks in the first group and
6721 * two tasks in the second group. Clearly this is undesired as it will overload
6722 * cpu 3 and leave one of the cpus in the second group unused.
6724 * The current solution to this issue is detecting the skew in the first group
6725 * by noticing the lower domain failed to reach balance and had difficulty
6726 * moving tasks due to affinity constraints.
6728 * When this is so detected; this group becomes a candidate for busiest; see
6729 * update_sd_pick_busiest(). And calculate_imbalance() and
6730 * find_busiest_group() avoid some of the usual balance conditions to allow it
6731 * to create an effective group imbalance.
6733 * This is a somewhat tricky proposition since the next run might not find the
6734 * group imbalance and decide the groups need to be balanced again. A most
6735 * subtle and fragile situation.
6738 static inline int sg_imbalanced(struct sched_group *group)
6740 return group->sgc->imbalance;
6744 * group_has_capacity returns true if the group has spare capacity that could
6745 * be used by some tasks.
6746 * We consider that a group has spare capacity if the * number of task is
6747 * smaller than the number of CPUs or if the utilization is lower than the
6748 * available capacity for CFS tasks.
6749 * For the latter, we use a threshold to stabilize the state, to take into
6750 * account the variance of the tasks' load and to return true if the available
6751 * capacity in meaningful for the load balancer.
6752 * As an example, an available capacity of 1% can appear but it doesn't make
6753 * any benefit for the load balance.
6756 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6758 if (sgs->sum_nr_running < sgs->group_weight)
6761 if ((sgs->group_capacity * 100) >
6762 (sgs->group_util * env->sd->imbalance_pct))
6769 * group_is_overloaded returns true if the group has more tasks than it can
6771 * group_is_overloaded is not equals to !group_has_capacity because a group
6772 * with the exact right number of tasks, has no more spare capacity but is not
6773 * overloaded so both group_has_capacity and group_is_overloaded return
6777 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6779 if (sgs->sum_nr_running <= sgs->group_weight)
6782 if ((sgs->group_capacity * 100) <
6783 (sgs->group_util * env->sd->imbalance_pct))
6791 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
6792 * per-cpu capacity than sched_group ref.
6795 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
6797 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
6798 ref->sgc->max_capacity;
6802 group_type group_classify(struct sched_group *group,
6803 struct sg_lb_stats *sgs)
6805 if (sgs->group_no_capacity)
6806 return group_overloaded;
6808 if (sg_imbalanced(group))
6809 return group_imbalanced;
6811 if (sgs->group_misfit_task)
6812 return group_misfit_task;
6818 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6819 * @env: The load balancing environment.
6820 * @group: sched_group whose statistics are to be updated.
6821 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6822 * @local_group: Does group contain this_cpu.
6823 * @sgs: variable to hold the statistics for this group.
6824 * @overload: Indicate more than one runnable task for any CPU.
6825 * @overutilized: Indicate overutilization for any CPU.
6827 static inline void update_sg_lb_stats(struct lb_env *env,
6828 struct sched_group *group, int load_idx,
6829 int local_group, struct sg_lb_stats *sgs,
6830 bool *overload, bool *overutilized)
6835 memset(sgs, 0, sizeof(*sgs));
6837 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6838 struct rq *rq = cpu_rq(i);
6840 /* Bias balancing toward cpus of our domain */
6842 load = target_load(i, load_idx);
6844 load = source_load(i, load_idx);
6846 sgs->group_load += load;
6847 sgs->group_util += cpu_util(i);
6848 sgs->sum_nr_running += rq->cfs.h_nr_running;
6850 if (rq->nr_running > 1)
6853 #ifdef CONFIG_NUMA_BALANCING
6854 sgs->nr_numa_running += rq->nr_numa_running;
6855 sgs->nr_preferred_running += rq->nr_preferred_running;
6857 sgs->sum_weighted_load += weighted_cpuload(i);
6861 if (cpu_overutilized(i)) {
6862 *overutilized = true;
6863 if (!sgs->group_misfit_task && rq->misfit_task)
6864 sgs->group_misfit_task = capacity_of(i);
6868 /* Adjust by relative CPU capacity of the group */
6869 sgs->group_capacity = group->sgc->capacity;
6870 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6872 if (sgs->sum_nr_running)
6873 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6875 sgs->group_weight = group->group_weight;
6877 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6878 sgs->group_type = group_classify(group, sgs);
6882 * update_sd_pick_busiest - return 1 on busiest group
6883 * @env: The load balancing environment.
6884 * @sds: sched_domain statistics
6885 * @sg: sched_group candidate to be checked for being the busiest
6886 * @sgs: sched_group statistics
6888 * Determine if @sg is a busier group than the previously selected
6891 * Return: %true if @sg is a busier group than the previously selected
6892 * busiest group. %false otherwise.
6894 static bool update_sd_pick_busiest(struct lb_env *env,
6895 struct sd_lb_stats *sds,
6896 struct sched_group *sg,
6897 struct sg_lb_stats *sgs)
6899 struct sg_lb_stats *busiest = &sds->busiest_stat;
6901 if (sgs->group_type > busiest->group_type)
6904 if (sgs->group_type < busiest->group_type)
6908 * Candidate sg doesn't face any serious load-balance problems
6909 * so don't pick it if the local sg is already filled up.
6911 if (sgs->group_type == group_other &&
6912 !group_has_capacity(env, &sds->local_stat))
6915 if (sgs->avg_load <= busiest->avg_load)
6919 * Candiate sg has no more than one task per cpu and has higher
6920 * per-cpu capacity. No reason to pull tasks to less capable cpus.
6922 if (sgs->sum_nr_running <= sgs->group_weight &&
6923 group_smaller_cpu_capacity(sds->local, sg))
6926 /* This is the busiest node in its class. */
6927 if (!(env->sd->flags & SD_ASYM_PACKING))
6931 * ASYM_PACKING needs to move all the work to the lowest
6932 * numbered CPUs in the group, therefore mark all groups
6933 * higher than ourself as busy.
6935 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6939 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6946 #ifdef CONFIG_NUMA_BALANCING
6947 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6949 if (sgs->sum_nr_running > sgs->nr_numa_running)
6951 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6956 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6958 if (rq->nr_running > rq->nr_numa_running)
6960 if (rq->nr_running > rq->nr_preferred_running)
6965 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6970 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6974 #endif /* CONFIG_NUMA_BALANCING */
6977 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6978 * @env: The load balancing environment.
6979 * @sds: variable to hold the statistics for this sched_domain.
6981 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6983 struct sched_domain *child = env->sd->child;
6984 struct sched_group *sg = env->sd->groups;
6985 struct sg_lb_stats tmp_sgs;
6986 int load_idx, prefer_sibling = 0;
6987 bool overload = false, overutilized = false;
6989 if (child && child->flags & SD_PREFER_SIBLING)
6992 load_idx = get_sd_load_idx(env->sd, env->idle);
6995 struct sg_lb_stats *sgs = &tmp_sgs;
6998 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7001 sgs = &sds->local_stat;
7003 if (env->idle != CPU_NEWLY_IDLE ||
7004 time_after_eq(jiffies, sg->sgc->next_update))
7005 update_group_capacity(env->sd, env->dst_cpu);
7008 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7009 &overload, &overutilized);
7015 * In case the child domain prefers tasks go to siblings
7016 * first, lower the sg capacity so that we'll try
7017 * and move all the excess tasks away. We lower the capacity
7018 * of a group only if the local group has the capacity to fit
7019 * these excess tasks. The extra check prevents the case where
7020 * you always pull from the heaviest group when it is already
7021 * under-utilized (possible with a large weight task outweighs
7022 * the tasks on the system).
7024 if (prefer_sibling && sds->local &&
7025 group_has_capacity(env, &sds->local_stat) &&
7026 (sgs->sum_nr_running > 1)) {
7027 sgs->group_no_capacity = 1;
7028 sgs->group_type = group_classify(sg, sgs);
7032 * Ignore task groups with misfit tasks if local group has no
7033 * capacity or if per-cpu capacity isn't higher.
7035 if (sgs->group_type == group_misfit_task &&
7036 (!group_has_capacity(env, &sds->local_stat) ||
7037 !group_smaller_cpu_capacity(sg, sds->local)))
7038 sgs->group_type = group_other;
7040 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7042 sds->busiest_stat = *sgs;
7046 /* Now, start updating sd_lb_stats */
7047 sds->total_load += sgs->group_load;
7048 sds->total_capacity += sgs->group_capacity;
7051 } while (sg != env->sd->groups);
7053 if (env->sd->flags & SD_NUMA)
7054 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7056 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7058 if (!env->sd->parent) {
7059 /* update overload indicator if we are at root domain */
7060 if (env->dst_rq->rd->overload != overload)
7061 env->dst_rq->rd->overload = overload;
7063 /* Update over-utilization (tipping point, U >= 0) indicator */
7064 if (env->dst_rq->rd->overutilized != overutilized)
7065 env->dst_rq->rd->overutilized = overutilized;
7067 if (!env->dst_rq->rd->overutilized && overutilized)
7068 env->dst_rq->rd->overutilized = true;
7073 * check_asym_packing - Check to see if the group is packed into the
7076 * This is primarily intended to used at the sibling level. Some
7077 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7078 * case of POWER7, it can move to lower SMT modes only when higher
7079 * threads are idle. When in lower SMT modes, the threads will
7080 * perform better since they share less core resources. Hence when we
7081 * have idle threads, we want them to be the higher ones.
7083 * This packing function is run on idle threads. It checks to see if
7084 * the busiest CPU in this domain (core in the P7 case) has a higher
7085 * CPU number than the packing function is being run on. Here we are
7086 * assuming lower CPU number will be equivalent to lower a SMT thread
7089 * Return: 1 when packing is required and a task should be moved to
7090 * this CPU. The amount of the imbalance is returned in *imbalance.
7092 * @env: The load balancing environment.
7093 * @sds: Statistics of the sched_domain which is to be packed
7095 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7099 if (!(env->sd->flags & SD_ASYM_PACKING))
7105 busiest_cpu = group_first_cpu(sds->busiest);
7106 if (env->dst_cpu > busiest_cpu)
7109 env->imbalance = DIV_ROUND_CLOSEST(
7110 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7111 SCHED_CAPACITY_SCALE);
7117 * fix_small_imbalance - Calculate the minor imbalance that exists
7118 * amongst the groups of a sched_domain, during
7120 * @env: The load balancing environment.
7121 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7124 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7126 unsigned long tmp, capa_now = 0, capa_move = 0;
7127 unsigned int imbn = 2;
7128 unsigned long scaled_busy_load_per_task;
7129 struct sg_lb_stats *local, *busiest;
7131 local = &sds->local_stat;
7132 busiest = &sds->busiest_stat;
7134 if (!local->sum_nr_running)
7135 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7136 else if (busiest->load_per_task > local->load_per_task)
7139 scaled_busy_load_per_task =
7140 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7141 busiest->group_capacity;
7143 if (busiest->avg_load + scaled_busy_load_per_task >=
7144 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7145 env->imbalance = busiest->load_per_task;
7150 * OK, we don't have enough imbalance to justify moving tasks,
7151 * however we may be able to increase total CPU capacity used by
7155 capa_now += busiest->group_capacity *
7156 min(busiest->load_per_task, busiest->avg_load);
7157 capa_now += local->group_capacity *
7158 min(local->load_per_task, local->avg_load);
7159 capa_now /= SCHED_CAPACITY_SCALE;
7161 /* Amount of load we'd subtract */
7162 if (busiest->avg_load > scaled_busy_load_per_task) {
7163 capa_move += busiest->group_capacity *
7164 min(busiest->load_per_task,
7165 busiest->avg_load - scaled_busy_load_per_task);
7168 /* Amount of load we'd add */
7169 if (busiest->avg_load * busiest->group_capacity <
7170 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7171 tmp = (busiest->avg_load * busiest->group_capacity) /
7172 local->group_capacity;
7174 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7175 local->group_capacity;
7177 capa_move += local->group_capacity *
7178 min(local->load_per_task, local->avg_load + tmp);
7179 capa_move /= SCHED_CAPACITY_SCALE;
7181 /* Move if we gain throughput */
7182 if (capa_move > capa_now)
7183 env->imbalance = busiest->load_per_task;
7187 * calculate_imbalance - Calculate the amount of imbalance present within the
7188 * groups of a given sched_domain during load balance.
7189 * @env: load balance environment
7190 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7192 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7194 unsigned long max_pull, load_above_capacity = ~0UL;
7195 struct sg_lb_stats *local, *busiest;
7197 local = &sds->local_stat;
7198 busiest = &sds->busiest_stat;
7200 if (busiest->group_type == group_imbalanced) {
7202 * In the group_imb case we cannot rely on group-wide averages
7203 * to ensure cpu-load equilibrium, look at wider averages. XXX
7205 busiest->load_per_task =
7206 min(busiest->load_per_task, sds->avg_load);
7210 * In the presence of smp nice balancing, certain scenarios can have
7211 * max load less than avg load(as we skip the groups at or below
7212 * its cpu_capacity, while calculating max_load..)
7214 if (busiest->avg_load <= sds->avg_load ||
7215 local->avg_load >= sds->avg_load) {
7216 /* Misfitting tasks should be migrated in any case */
7217 if (busiest->group_type == group_misfit_task) {
7218 env->imbalance = busiest->group_misfit_task;
7223 * Busiest group is overloaded, local is not, use the spare
7224 * cycles to maximize throughput
7226 if (busiest->group_type == group_overloaded &&
7227 local->group_type <= group_misfit_task) {
7228 env->imbalance = busiest->load_per_task;
7233 return fix_small_imbalance(env, sds);
7237 * If there aren't any idle cpus, avoid creating some.
7239 if (busiest->group_type == group_overloaded &&
7240 local->group_type == group_overloaded) {
7241 load_above_capacity = busiest->sum_nr_running *
7243 if (load_above_capacity > busiest->group_capacity)
7244 load_above_capacity -= busiest->group_capacity;
7246 load_above_capacity = ~0UL;
7250 * We're trying to get all the cpus to the average_load, so we don't
7251 * want to push ourselves above the average load, nor do we wish to
7252 * reduce the max loaded cpu below the average load. At the same time,
7253 * we also don't want to reduce the group load below the group capacity
7254 * (so that we can implement power-savings policies etc). Thus we look
7255 * for the minimum possible imbalance.
7257 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7259 /* How much load to actually move to equalise the imbalance */
7260 env->imbalance = min(
7261 max_pull * busiest->group_capacity,
7262 (sds->avg_load - local->avg_load) * local->group_capacity
7263 ) / SCHED_CAPACITY_SCALE;
7265 /* Boost imbalance to allow misfit task to be balanced. */
7266 if (busiest->group_type == group_misfit_task)
7267 env->imbalance = max_t(long, env->imbalance,
7268 busiest->group_misfit_task);
7271 * if *imbalance is less than the average load per runnable task
7272 * there is no guarantee that any tasks will be moved so we'll have
7273 * a think about bumping its value to force at least one task to be
7276 if (env->imbalance < busiest->load_per_task)
7277 return fix_small_imbalance(env, sds);
7280 /******* find_busiest_group() helpers end here *********************/
7283 * find_busiest_group - Returns the busiest group within the sched_domain
7284 * if there is an imbalance. If there isn't an imbalance, and
7285 * the user has opted for power-savings, it returns a group whose
7286 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7287 * such a group exists.
7289 * Also calculates the amount of weighted load which should be moved
7290 * to restore balance.
7292 * @env: The load balancing environment.
7294 * Return: - The busiest group if imbalance exists.
7295 * - If no imbalance and user has opted for power-savings balance,
7296 * return the least loaded group whose CPUs can be
7297 * put to idle by rebalancing its tasks onto our group.
7299 static struct sched_group *find_busiest_group(struct lb_env *env)
7301 struct sg_lb_stats *local, *busiest;
7302 struct sd_lb_stats sds;
7304 init_sd_lb_stats(&sds);
7307 * Compute the various statistics relavent for load balancing at
7310 update_sd_lb_stats(env, &sds);
7312 if (energy_aware() && !env->dst_rq->rd->overutilized)
7315 local = &sds.local_stat;
7316 busiest = &sds.busiest_stat;
7318 /* ASYM feature bypasses nice load balance check */
7319 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7320 check_asym_packing(env, &sds))
7323 /* There is no busy sibling group to pull tasks from */
7324 if (!sds.busiest || busiest->sum_nr_running == 0)
7327 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7328 / sds.total_capacity;
7331 * If the busiest group is imbalanced the below checks don't
7332 * work because they assume all things are equal, which typically
7333 * isn't true due to cpus_allowed constraints and the like.
7335 if (busiest->group_type == group_imbalanced)
7338 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7339 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7340 busiest->group_no_capacity)
7343 /* Misfitting tasks should be dealt with regardless of the avg load */
7344 if (busiest->group_type == group_misfit_task) {
7349 * If the local group is busier than the selected busiest group
7350 * don't try and pull any tasks.
7352 if (local->avg_load >= busiest->avg_load)
7356 * Don't pull any tasks if this group is already above the domain
7359 if (local->avg_load >= sds.avg_load)
7362 if (env->idle == CPU_IDLE) {
7364 * This cpu is idle. If the busiest group is not overloaded
7365 * and there is no imbalance between this and busiest group
7366 * wrt idle cpus, it is balanced. The imbalance becomes
7367 * significant if the diff is greater than 1 otherwise we
7368 * might end up to just move the imbalance on another group
7370 if ((busiest->group_type != group_overloaded) &&
7371 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7372 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7376 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7377 * imbalance_pct to be conservative.
7379 if (100 * busiest->avg_load <=
7380 env->sd->imbalance_pct * local->avg_load)
7385 env->busiest_group_type = busiest->group_type;
7386 /* Looks like there is an imbalance. Compute it */
7387 calculate_imbalance(env, &sds);
7396 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7398 static struct rq *find_busiest_queue(struct lb_env *env,
7399 struct sched_group *group)
7401 struct rq *busiest = NULL, *rq;
7402 unsigned long busiest_load = 0, busiest_capacity = 1;
7405 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7406 unsigned long capacity, wl;
7410 rt = fbq_classify_rq(rq);
7413 * We classify groups/runqueues into three groups:
7414 * - regular: there are !numa tasks
7415 * - remote: there are numa tasks that run on the 'wrong' node
7416 * - all: there is no distinction
7418 * In order to avoid migrating ideally placed numa tasks,
7419 * ignore those when there's better options.
7421 * If we ignore the actual busiest queue to migrate another
7422 * task, the next balance pass can still reduce the busiest
7423 * queue by moving tasks around inside the node.
7425 * If we cannot move enough load due to this classification
7426 * the next pass will adjust the group classification and
7427 * allow migration of more tasks.
7429 * Both cases only affect the total convergence complexity.
7431 if (rt > env->fbq_type)
7434 capacity = capacity_of(i);
7436 wl = weighted_cpuload(i);
7439 * When comparing with imbalance, use weighted_cpuload()
7440 * which is not scaled with the cpu capacity.
7443 if (rq->nr_running == 1 && wl > env->imbalance &&
7444 !check_cpu_capacity(rq, env->sd) &&
7445 env->busiest_group_type != group_misfit_task)
7449 * For the load comparisons with the other cpu's, consider
7450 * the weighted_cpuload() scaled with the cpu capacity, so
7451 * that the load can be moved away from the cpu that is
7452 * potentially running at a lower capacity.
7454 * Thus we're looking for max(wl_i / capacity_i), crosswise
7455 * multiplication to rid ourselves of the division works out
7456 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7457 * our previous maximum.
7459 if (wl * busiest_capacity > busiest_load * capacity) {
7461 busiest_capacity = capacity;
7470 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7471 * so long as it is large enough.
7473 #define MAX_PINNED_INTERVAL 512
7475 /* Working cpumask for load_balance and load_balance_newidle. */
7476 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7478 static int need_active_balance(struct lb_env *env)
7480 struct sched_domain *sd = env->sd;
7482 if (env->idle == CPU_NEWLY_IDLE) {
7485 * ASYM_PACKING needs to force migrate tasks from busy but
7486 * higher numbered CPUs in order to pack all tasks in the
7487 * lowest numbered CPUs.
7489 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7494 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7495 * It's worth migrating the task if the src_cpu's capacity is reduced
7496 * because of other sched_class or IRQs if more capacity stays
7497 * available on dst_cpu.
7499 if ((env->idle != CPU_NOT_IDLE) &&
7500 (env->src_rq->cfs.h_nr_running == 1)) {
7501 if ((check_cpu_capacity(env->src_rq, sd)) &&
7502 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7506 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
7507 env->src_rq->cfs.h_nr_running == 1 &&
7508 cpu_overutilized(env->src_cpu) &&
7509 !cpu_overutilized(env->dst_cpu)) {
7513 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7516 static int active_load_balance_cpu_stop(void *data);
7518 static int should_we_balance(struct lb_env *env)
7520 struct sched_group *sg = env->sd->groups;
7521 struct cpumask *sg_cpus, *sg_mask;
7522 int cpu, balance_cpu = -1;
7525 * In the newly idle case, we will allow all the cpu's
7526 * to do the newly idle load balance.
7528 if (env->idle == CPU_NEWLY_IDLE)
7531 sg_cpus = sched_group_cpus(sg);
7532 sg_mask = sched_group_mask(sg);
7533 /* Try to find first idle cpu */
7534 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7535 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7542 if (balance_cpu == -1)
7543 balance_cpu = group_balance_cpu(sg);
7546 * First idle cpu or the first cpu(busiest) in this sched group
7547 * is eligible for doing load balancing at this and above domains.
7549 return balance_cpu == env->dst_cpu;
7553 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7554 * tasks if there is an imbalance.
7556 static int load_balance(int this_cpu, struct rq *this_rq,
7557 struct sched_domain *sd, enum cpu_idle_type idle,
7558 int *continue_balancing)
7560 int ld_moved, cur_ld_moved, active_balance = 0;
7561 struct sched_domain *sd_parent = sd->parent;
7562 struct sched_group *group;
7564 unsigned long flags;
7565 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7567 struct lb_env env = {
7569 .dst_cpu = this_cpu,
7571 .dst_grpmask = sched_group_cpus(sd->groups),
7573 .loop_break = sched_nr_migrate_break,
7576 .tasks = LIST_HEAD_INIT(env.tasks),
7580 * For NEWLY_IDLE load_balancing, we don't need to consider
7581 * other cpus in our group
7583 if (idle == CPU_NEWLY_IDLE)
7584 env.dst_grpmask = NULL;
7586 cpumask_copy(cpus, cpu_active_mask);
7588 schedstat_inc(sd, lb_count[idle]);
7591 if (!should_we_balance(&env)) {
7592 *continue_balancing = 0;
7596 group = find_busiest_group(&env);
7598 schedstat_inc(sd, lb_nobusyg[idle]);
7602 busiest = find_busiest_queue(&env, group);
7604 schedstat_inc(sd, lb_nobusyq[idle]);
7608 BUG_ON(busiest == env.dst_rq);
7610 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7612 env.src_cpu = busiest->cpu;
7613 env.src_rq = busiest;
7616 if (busiest->nr_running > 1) {
7618 * Attempt to move tasks. If find_busiest_group has found
7619 * an imbalance but busiest->nr_running <= 1, the group is
7620 * still unbalanced. ld_moved simply stays zero, so it is
7621 * correctly treated as an imbalance.
7623 env.flags |= LBF_ALL_PINNED;
7624 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7627 raw_spin_lock_irqsave(&busiest->lock, flags);
7630 * cur_ld_moved - load moved in current iteration
7631 * ld_moved - cumulative load moved across iterations
7633 cur_ld_moved = detach_tasks(&env);
7635 * We want to potentially lower env.src_cpu's OPP.
7638 update_capacity_of(env.src_cpu);
7641 * We've detached some tasks from busiest_rq. Every
7642 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7643 * unlock busiest->lock, and we are able to be sure
7644 * that nobody can manipulate the tasks in parallel.
7645 * See task_rq_lock() family for the details.
7648 raw_spin_unlock(&busiest->lock);
7652 ld_moved += cur_ld_moved;
7655 local_irq_restore(flags);
7657 if (env.flags & LBF_NEED_BREAK) {
7658 env.flags &= ~LBF_NEED_BREAK;
7663 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7664 * us and move them to an alternate dst_cpu in our sched_group
7665 * where they can run. The upper limit on how many times we
7666 * iterate on same src_cpu is dependent on number of cpus in our
7669 * This changes load balance semantics a bit on who can move
7670 * load to a given_cpu. In addition to the given_cpu itself
7671 * (or a ilb_cpu acting on its behalf where given_cpu is
7672 * nohz-idle), we now have balance_cpu in a position to move
7673 * load to given_cpu. In rare situations, this may cause
7674 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7675 * _independently_ and at _same_ time to move some load to
7676 * given_cpu) causing exceess load to be moved to given_cpu.
7677 * This however should not happen so much in practice and
7678 * moreover subsequent load balance cycles should correct the
7679 * excess load moved.
7681 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7683 /* Prevent to re-select dst_cpu via env's cpus */
7684 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7686 env.dst_rq = cpu_rq(env.new_dst_cpu);
7687 env.dst_cpu = env.new_dst_cpu;
7688 env.flags &= ~LBF_DST_PINNED;
7690 env.loop_break = sched_nr_migrate_break;
7693 * Go back to "more_balance" rather than "redo" since we
7694 * need to continue with same src_cpu.
7700 * We failed to reach balance because of affinity.
7703 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7705 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7706 *group_imbalance = 1;
7709 /* All tasks on this runqueue were pinned by CPU affinity */
7710 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7711 cpumask_clear_cpu(cpu_of(busiest), cpus);
7712 if (!cpumask_empty(cpus)) {
7714 env.loop_break = sched_nr_migrate_break;
7717 goto out_all_pinned;
7722 schedstat_inc(sd, lb_failed[idle]);
7724 * Increment the failure counter only on periodic balance.
7725 * We do not want newidle balance, which can be very
7726 * frequent, pollute the failure counter causing
7727 * excessive cache_hot migrations and active balances.
7729 if (idle != CPU_NEWLY_IDLE)
7730 if (env.src_grp_nr_running > 1)
7731 sd->nr_balance_failed++;
7733 if (need_active_balance(&env)) {
7734 raw_spin_lock_irqsave(&busiest->lock, flags);
7736 /* don't kick the active_load_balance_cpu_stop,
7737 * if the curr task on busiest cpu can't be
7740 if (!cpumask_test_cpu(this_cpu,
7741 tsk_cpus_allowed(busiest->curr))) {
7742 raw_spin_unlock_irqrestore(&busiest->lock,
7744 env.flags |= LBF_ALL_PINNED;
7745 goto out_one_pinned;
7749 * ->active_balance synchronizes accesses to
7750 * ->active_balance_work. Once set, it's cleared
7751 * only after active load balance is finished.
7753 if (!busiest->active_balance) {
7754 busiest->active_balance = 1;
7755 busiest->push_cpu = this_cpu;
7758 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7760 if (active_balance) {
7761 stop_one_cpu_nowait(cpu_of(busiest),
7762 active_load_balance_cpu_stop, busiest,
7763 &busiest->active_balance_work);
7767 * We've kicked active balancing, reset the failure
7770 sd->nr_balance_failed = sd->cache_nice_tries+1;
7773 sd->nr_balance_failed = 0;
7775 if (likely(!active_balance)) {
7776 /* We were unbalanced, so reset the balancing interval */
7777 sd->balance_interval = sd->min_interval;
7780 * If we've begun active balancing, start to back off. This
7781 * case may not be covered by the all_pinned logic if there
7782 * is only 1 task on the busy runqueue (because we don't call
7785 if (sd->balance_interval < sd->max_interval)
7786 sd->balance_interval *= 2;
7793 * We reach balance although we may have faced some affinity
7794 * constraints. Clear the imbalance flag if it was set.
7797 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7799 if (*group_imbalance)
7800 *group_imbalance = 0;
7805 * We reach balance because all tasks are pinned at this level so
7806 * we can't migrate them. Let the imbalance flag set so parent level
7807 * can try to migrate them.
7809 schedstat_inc(sd, lb_balanced[idle]);
7811 sd->nr_balance_failed = 0;
7814 /* tune up the balancing interval */
7815 if (((env.flags & LBF_ALL_PINNED) &&
7816 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7817 (sd->balance_interval < sd->max_interval))
7818 sd->balance_interval *= 2;
7825 static inline unsigned long
7826 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7828 unsigned long interval = sd->balance_interval;
7831 interval *= sd->busy_factor;
7833 /* scale ms to jiffies */
7834 interval = msecs_to_jiffies(interval);
7835 interval = clamp(interval, 1UL, max_load_balance_interval);
7841 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7843 unsigned long interval, next;
7845 interval = get_sd_balance_interval(sd, cpu_busy);
7846 next = sd->last_balance + interval;
7848 if (time_after(*next_balance, next))
7849 *next_balance = next;
7853 * idle_balance is called by schedule() if this_cpu is about to become
7854 * idle. Attempts to pull tasks from other CPUs.
7856 static int idle_balance(struct rq *this_rq)
7858 unsigned long next_balance = jiffies + HZ;
7859 int this_cpu = this_rq->cpu;
7860 struct sched_domain *sd;
7861 int pulled_task = 0;
7864 idle_enter_fair(this_rq);
7867 * We must set idle_stamp _before_ calling idle_balance(), such that we
7868 * measure the duration of idle_balance() as idle time.
7870 this_rq->idle_stamp = rq_clock(this_rq);
7872 if (!energy_aware() &&
7873 (this_rq->avg_idle < sysctl_sched_migration_cost ||
7874 !this_rq->rd->overload)) {
7876 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7878 update_next_balance(sd, 0, &next_balance);
7884 raw_spin_unlock(&this_rq->lock);
7886 update_blocked_averages(this_cpu);
7888 for_each_domain(this_cpu, sd) {
7889 int continue_balancing = 1;
7890 u64 t0, domain_cost;
7892 if (!(sd->flags & SD_LOAD_BALANCE))
7895 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7896 update_next_balance(sd, 0, &next_balance);
7900 if (sd->flags & SD_BALANCE_NEWIDLE) {
7901 t0 = sched_clock_cpu(this_cpu);
7903 pulled_task = load_balance(this_cpu, this_rq,
7905 &continue_balancing);
7907 domain_cost = sched_clock_cpu(this_cpu) - t0;
7908 if (domain_cost > sd->max_newidle_lb_cost)
7909 sd->max_newidle_lb_cost = domain_cost;
7911 curr_cost += domain_cost;
7914 update_next_balance(sd, 0, &next_balance);
7917 * Stop searching for tasks to pull if there are
7918 * now runnable tasks on this rq.
7920 if (pulled_task || this_rq->nr_running > 0)
7925 raw_spin_lock(&this_rq->lock);
7927 if (curr_cost > this_rq->max_idle_balance_cost)
7928 this_rq->max_idle_balance_cost = curr_cost;
7931 * While browsing the domains, we released the rq lock, a task could
7932 * have been enqueued in the meantime. Since we're not going idle,
7933 * pretend we pulled a task.
7935 if (this_rq->cfs.h_nr_running && !pulled_task)
7939 /* Move the next balance forward */
7940 if (time_after(this_rq->next_balance, next_balance))
7941 this_rq->next_balance = next_balance;
7943 /* Is there a task of a high priority class? */
7944 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7948 idle_exit_fair(this_rq);
7949 this_rq->idle_stamp = 0;
7956 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7957 * running tasks off the busiest CPU onto idle CPUs. It requires at
7958 * least 1 task to be running on each physical CPU where possible, and
7959 * avoids physical / logical imbalances.
7961 static int active_load_balance_cpu_stop(void *data)
7963 struct rq *busiest_rq = data;
7964 int busiest_cpu = cpu_of(busiest_rq);
7965 int target_cpu = busiest_rq->push_cpu;
7966 struct rq *target_rq = cpu_rq(target_cpu);
7967 struct sched_domain *sd;
7968 struct task_struct *p = NULL;
7970 raw_spin_lock_irq(&busiest_rq->lock);
7972 /* make sure the requested cpu hasn't gone down in the meantime */
7973 if (unlikely(busiest_cpu != smp_processor_id() ||
7974 !busiest_rq->active_balance))
7977 /* Is there any task to move? */
7978 if (busiest_rq->nr_running <= 1)
7982 * This condition is "impossible", if it occurs
7983 * we need to fix it. Originally reported by
7984 * Bjorn Helgaas on a 128-cpu setup.
7986 BUG_ON(busiest_rq == target_rq);
7988 /* Search for an sd spanning us and the target CPU. */
7990 for_each_domain(target_cpu, sd) {
7991 if ((sd->flags & SD_LOAD_BALANCE) &&
7992 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7997 struct lb_env env = {
7999 .dst_cpu = target_cpu,
8000 .dst_rq = target_rq,
8001 .src_cpu = busiest_rq->cpu,
8002 .src_rq = busiest_rq,
8006 schedstat_inc(sd, alb_count);
8008 p = detach_one_task(&env);
8010 schedstat_inc(sd, alb_pushed);
8012 * We want to potentially lower env.src_cpu's OPP.
8014 update_capacity_of(env.src_cpu);
8017 schedstat_inc(sd, alb_failed);
8021 busiest_rq->active_balance = 0;
8022 raw_spin_unlock(&busiest_rq->lock);
8025 attach_one_task(target_rq, p);
8032 static inline int on_null_domain(struct rq *rq)
8034 return unlikely(!rcu_dereference_sched(rq->sd));
8037 #ifdef CONFIG_NO_HZ_COMMON
8039 * idle load balancing details
8040 * - When one of the busy CPUs notice that there may be an idle rebalancing
8041 * needed, they will kick the idle load balancer, which then does idle
8042 * load balancing for all the idle CPUs.
8045 cpumask_var_t idle_cpus_mask;
8047 unsigned long next_balance; /* in jiffy units */
8048 } nohz ____cacheline_aligned;
8050 static inline int find_new_ilb(void)
8052 int ilb = cpumask_first(nohz.idle_cpus_mask);
8054 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8061 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8062 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8063 * CPU (if there is one).
8065 static void nohz_balancer_kick(void)
8069 nohz.next_balance++;
8071 ilb_cpu = find_new_ilb();
8073 if (ilb_cpu >= nr_cpu_ids)
8076 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8079 * Use smp_send_reschedule() instead of resched_cpu().
8080 * This way we generate a sched IPI on the target cpu which
8081 * is idle. And the softirq performing nohz idle load balance
8082 * will be run before returning from the IPI.
8084 smp_send_reschedule(ilb_cpu);
8088 static inline void nohz_balance_exit_idle(int cpu)
8090 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8092 * Completely isolated CPUs don't ever set, so we must test.
8094 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8095 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8096 atomic_dec(&nohz.nr_cpus);
8098 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8102 static inline void set_cpu_sd_state_busy(void)
8104 struct sched_domain *sd;
8105 int cpu = smp_processor_id();
8108 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8110 if (!sd || !sd->nohz_idle)
8114 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8119 void set_cpu_sd_state_idle(void)
8121 struct sched_domain *sd;
8122 int cpu = smp_processor_id();
8125 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8127 if (!sd || sd->nohz_idle)
8131 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8137 * This routine will record that the cpu is going idle with tick stopped.
8138 * This info will be used in performing idle load balancing in the future.
8140 void nohz_balance_enter_idle(int cpu)
8143 * If this cpu is going down, then nothing needs to be done.
8145 if (!cpu_active(cpu))
8148 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8152 * If we're a completely isolated CPU, we don't play.
8154 if (on_null_domain(cpu_rq(cpu)))
8157 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8158 atomic_inc(&nohz.nr_cpus);
8159 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8162 static int sched_ilb_notifier(struct notifier_block *nfb,
8163 unsigned long action, void *hcpu)
8165 switch (action & ~CPU_TASKS_FROZEN) {
8167 nohz_balance_exit_idle(smp_processor_id());
8175 static DEFINE_SPINLOCK(balancing);
8178 * Scale the max load_balance interval with the number of CPUs in the system.
8179 * This trades load-balance latency on larger machines for less cross talk.
8181 void update_max_interval(void)
8183 max_load_balance_interval = HZ*num_online_cpus()/10;
8187 * It checks each scheduling domain to see if it is due to be balanced,
8188 * and initiates a balancing operation if so.
8190 * Balancing parameters are set up in init_sched_domains.
8192 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8194 int continue_balancing = 1;
8196 unsigned long interval;
8197 struct sched_domain *sd;
8198 /* Earliest time when we have to do rebalance again */
8199 unsigned long next_balance = jiffies + 60*HZ;
8200 int update_next_balance = 0;
8201 int need_serialize, need_decay = 0;
8204 update_blocked_averages(cpu);
8207 for_each_domain(cpu, sd) {
8209 * Decay the newidle max times here because this is a regular
8210 * visit to all the domains. Decay ~1% per second.
8212 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8213 sd->max_newidle_lb_cost =
8214 (sd->max_newidle_lb_cost * 253) / 256;
8215 sd->next_decay_max_lb_cost = jiffies + HZ;
8218 max_cost += sd->max_newidle_lb_cost;
8220 if (!(sd->flags & SD_LOAD_BALANCE))
8224 * Stop the load balance at this level. There is another
8225 * CPU in our sched group which is doing load balancing more
8228 if (!continue_balancing) {
8234 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8236 need_serialize = sd->flags & SD_SERIALIZE;
8237 if (need_serialize) {
8238 if (!spin_trylock(&balancing))
8242 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8243 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8245 * The LBF_DST_PINNED logic could have changed
8246 * env->dst_cpu, so we can't know our idle
8247 * state even if we migrated tasks. Update it.
8249 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8251 sd->last_balance = jiffies;
8252 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8255 spin_unlock(&balancing);
8257 if (time_after(next_balance, sd->last_balance + interval)) {
8258 next_balance = sd->last_balance + interval;
8259 update_next_balance = 1;
8264 * Ensure the rq-wide value also decays but keep it at a
8265 * reasonable floor to avoid funnies with rq->avg_idle.
8267 rq->max_idle_balance_cost =
8268 max((u64)sysctl_sched_migration_cost, max_cost);
8273 * next_balance will be updated only when there is a need.
8274 * When the cpu is attached to null domain for ex, it will not be
8277 if (likely(update_next_balance)) {
8278 rq->next_balance = next_balance;
8280 #ifdef CONFIG_NO_HZ_COMMON
8282 * If this CPU has been elected to perform the nohz idle
8283 * balance. Other idle CPUs have already rebalanced with
8284 * nohz_idle_balance() and nohz.next_balance has been
8285 * updated accordingly. This CPU is now running the idle load
8286 * balance for itself and we need to update the
8287 * nohz.next_balance accordingly.
8289 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8290 nohz.next_balance = rq->next_balance;
8295 #ifdef CONFIG_NO_HZ_COMMON
8297 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8298 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8300 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8302 int this_cpu = this_rq->cpu;
8305 /* Earliest time when we have to do rebalance again */
8306 unsigned long next_balance = jiffies + 60*HZ;
8307 int update_next_balance = 0;
8309 if (idle != CPU_IDLE ||
8310 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8313 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8314 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8318 * If this cpu gets work to do, stop the load balancing
8319 * work being done for other cpus. Next load
8320 * balancing owner will pick it up.
8325 rq = cpu_rq(balance_cpu);
8328 * If time for next balance is due,
8331 if (time_after_eq(jiffies, rq->next_balance)) {
8332 raw_spin_lock_irq(&rq->lock);
8333 update_rq_clock(rq);
8334 update_idle_cpu_load(rq);
8335 raw_spin_unlock_irq(&rq->lock);
8336 rebalance_domains(rq, CPU_IDLE);
8339 if (time_after(next_balance, rq->next_balance)) {
8340 next_balance = rq->next_balance;
8341 update_next_balance = 1;
8346 * next_balance will be updated only when there is a need.
8347 * When the CPU is attached to null domain for ex, it will not be
8350 if (likely(update_next_balance))
8351 nohz.next_balance = next_balance;
8353 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8357 * Current heuristic for kicking the idle load balancer in the presence
8358 * of an idle cpu in the system.
8359 * - This rq has more than one task.
8360 * - This rq has at least one CFS task and the capacity of the CPU is
8361 * significantly reduced because of RT tasks or IRQs.
8362 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8363 * multiple busy cpu.
8364 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8365 * domain span are idle.
8367 static inline bool nohz_kick_needed(struct rq *rq)
8369 unsigned long now = jiffies;
8370 struct sched_domain *sd;
8371 struct sched_group_capacity *sgc;
8372 int nr_busy, cpu = rq->cpu;
8375 if (unlikely(rq->idle_balance))
8379 * We may be recently in ticked or tickless idle mode. At the first
8380 * busy tick after returning from idle, we will update the busy stats.
8382 set_cpu_sd_state_busy();
8383 nohz_balance_exit_idle(cpu);
8386 * None are in tickless mode and hence no need for NOHZ idle load
8389 if (likely(!atomic_read(&nohz.nr_cpus)))
8392 if (time_before(now, nohz.next_balance))
8395 if (rq->nr_running >= 2 &&
8396 (!energy_aware() || cpu_overutilized(cpu)))
8400 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8401 if (sd && !energy_aware()) {
8402 sgc = sd->groups->sgc;
8403 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8412 sd = rcu_dereference(rq->sd);
8414 if ((rq->cfs.h_nr_running >= 1) &&
8415 check_cpu_capacity(rq, sd)) {
8421 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8422 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8423 sched_domain_span(sd)) < cpu)) {
8433 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8437 * run_rebalance_domains is triggered when needed from the scheduler tick.
8438 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8440 static void run_rebalance_domains(struct softirq_action *h)
8442 struct rq *this_rq = this_rq();
8443 enum cpu_idle_type idle = this_rq->idle_balance ?
8444 CPU_IDLE : CPU_NOT_IDLE;
8447 * If this cpu has a pending nohz_balance_kick, then do the
8448 * balancing on behalf of the other idle cpus whose ticks are
8449 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8450 * give the idle cpus a chance to load balance. Else we may
8451 * load balance only within the local sched_domain hierarchy
8452 * and abort nohz_idle_balance altogether if we pull some load.
8454 nohz_idle_balance(this_rq, idle);
8455 rebalance_domains(this_rq, idle);
8459 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8461 void trigger_load_balance(struct rq *rq)
8463 /* Don't need to rebalance while attached to NULL domain */
8464 if (unlikely(on_null_domain(rq)))
8467 if (time_after_eq(jiffies, rq->next_balance))
8468 raise_softirq(SCHED_SOFTIRQ);
8469 #ifdef CONFIG_NO_HZ_COMMON
8470 if (nohz_kick_needed(rq))
8471 nohz_balancer_kick();
8475 static void rq_online_fair(struct rq *rq)
8479 update_runtime_enabled(rq);
8482 static void rq_offline_fair(struct rq *rq)
8486 /* Ensure any throttled groups are reachable by pick_next_task */
8487 unthrottle_offline_cfs_rqs(rq);
8490 #endif /* CONFIG_SMP */
8493 * scheduler tick hitting a task of our scheduling class:
8495 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8497 struct cfs_rq *cfs_rq;
8498 struct sched_entity *se = &curr->se;
8500 for_each_sched_entity(se) {
8501 cfs_rq = cfs_rq_of(se);
8502 entity_tick(cfs_rq, se, queued);
8505 if (static_branch_unlikely(&sched_numa_balancing))
8506 task_tick_numa(rq, curr);
8508 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr)))
8509 rq->rd->overutilized = true;
8511 rq->misfit_task = !task_fits_max(curr, rq->cpu);
8515 * called on fork with the child task as argument from the parent's context
8516 * - child not yet on the tasklist
8517 * - preemption disabled
8519 static void task_fork_fair(struct task_struct *p)
8521 struct cfs_rq *cfs_rq;
8522 struct sched_entity *se = &p->se, *curr;
8523 int this_cpu = smp_processor_id();
8524 struct rq *rq = this_rq();
8525 unsigned long flags;
8527 raw_spin_lock_irqsave(&rq->lock, flags);
8529 update_rq_clock(rq);
8531 cfs_rq = task_cfs_rq(current);
8532 curr = cfs_rq->curr;
8535 * Not only the cpu but also the task_group of the parent might have
8536 * been changed after parent->se.parent,cfs_rq were copied to
8537 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8538 * of child point to valid ones.
8541 __set_task_cpu(p, this_cpu);
8544 update_curr(cfs_rq);
8547 se->vruntime = curr->vruntime;
8548 place_entity(cfs_rq, se, 1);
8550 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8552 * Upon rescheduling, sched_class::put_prev_task() will place
8553 * 'current' within the tree based on its new key value.
8555 swap(curr->vruntime, se->vruntime);
8559 se->vruntime -= cfs_rq->min_vruntime;
8561 raw_spin_unlock_irqrestore(&rq->lock, flags);
8565 * Priority of the task has changed. Check to see if we preempt
8569 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8571 if (!task_on_rq_queued(p))
8575 * Reschedule if we are currently running on this runqueue and
8576 * our priority decreased, or if we are not currently running on
8577 * this runqueue and our priority is higher than the current's
8579 if (rq->curr == p) {
8580 if (p->prio > oldprio)
8583 check_preempt_curr(rq, p, 0);
8586 static inline bool vruntime_normalized(struct task_struct *p)
8588 struct sched_entity *se = &p->se;
8591 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8592 * the dequeue_entity(.flags=0) will already have normalized the
8599 * When !on_rq, vruntime of the task has usually NOT been normalized.
8600 * But there are some cases where it has already been normalized:
8602 * - A forked child which is waiting for being woken up by
8603 * wake_up_new_task().
8604 * - A task which has been woken up by try_to_wake_up() and
8605 * waiting for actually being woken up by sched_ttwu_pending().
8607 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8613 static void detach_task_cfs_rq(struct task_struct *p)
8615 struct sched_entity *se = &p->se;
8616 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8618 if (!vruntime_normalized(p)) {
8620 * Fix up our vruntime so that the current sleep doesn't
8621 * cause 'unlimited' sleep bonus.
8623 place_entity(cfs_rq, se, 0);
8624 se->vruntime -= cfs_rq->min_vruntime;
8627 /* Catch up with the cfs_rq and remove our load when we leave */
8628 detach_entity_load_avg(cfs_rq, se);
8631 static void attach_task_cfs_rq(struct task_struct *p)
8633 struct sched_entity *se = &p->se;
8634 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8636 #ifdef CONFIG_FAIR_GROUP_SCHED
8638 * Since the real-depth could have been changed (only FAIR
8639 * class maintain depth value), reset depth properly.
8641 se->depth = se->parent ? se->parent->depth + 1 : 0;
8644 /* Synchronize task with its cfs_rq */
8645 attach_entity_load_avg(cfs_rq, se);
8647 if (!vruntime_normalized(p))
8648 se->vruntime += cfs_rq->min_vruntime;
8651 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8653 detach_task_cfs_rq(p);
8656 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8658 attach_task_cfs_rq(p);
8660 if (task_on_rq_queued(p)) {
8662 * We were most likely switched from sched_rt, so
8663 * kick off the schedule if running, otherwise just see
8664 * if we can still preempt the current task.
8669 check_preempt_curr(rq, p, 0);
8673 /* Account for a task changing its policy or group.
8675 * This routine is mostly called to set cfs_rq->curr field when a task
8676 * migrates between groups/classes.
8678 static void set_curr_task_fair(struct rq *rq)
8680 struct sched_entity *se = &rq->curr->se;
8682 for_each_sched_entity(se) {
8683 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8685 set_next_entity(cfs_rq, se);
8686 /* ensure bandwidth has been allocated on our new cfs_rq */
8687 account_cfs_rq_runtime(cfs_rq, 0);
8691 void init_cfs_rq(struct cfs_rq *cfs_rq)
8693 cfs_rq->tasks_timeline = RB_ROOT;
8694 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8695 #ifndef CONFIG_64BIT
8696 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8699 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8700 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8704 #ifdef CONFIG_FAIR_GROUP_SCHED
8705 static void task_move_group_fair(struct task_struct *p)
8707 detach_task_cfs_rq(p);
8708 set_task_rq(p, task_cpu(p));
8711 /* Tell se's cfs_rq has been changed -- migrated */
8712 p->se.avg.last_update_time = 0;
8714 attach_task_cfs_rq(p);
8717 void free_fair_sched_group(struct task_group *tg)
8721 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8723 for_each_possible_cpu(i) {
8725 kfree(tg->cfs_rq[i]);
8728 remove_entity_load_avg(tg->se[i]);
8737 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8739 struct cfs_rq *cfs_rq;
8740 struct sched_entity *se;
8743 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8746 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8750 tg->shares = NICE_0_LOAD;
8752 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8754 for_each_possible_cpu(i) {
8755 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8756 GFP_KERNEL, cpu_to_node(i));
8760 se = kzalloc_node(sizeof(struct sched_entity),
8761 GFP_KERNEL, cpu_to_node(i));
8765 init_cfs_rq(cfs_rq);
8766 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8767 init_entity_runnable_average(se);
8778 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8780 struct rq *rq = cpu_rq(cpu);
8781 unsigned long flags;
8784 * Only empty task groups can be destroyed; so we can speculatively
8785 * check on_list without danger of it being re-added.
8787 if (!tg->cfs_rq[cpu]->on_list)
8790 raw_spin_lock_irqsave(&rq->lock, flags);
8791 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8792 raw_spin_unlock_irqrestore(&rq->lock, flags);
8795 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8796 struct sched_entity *se, int cpu,
8797 struct sched_entity *parent)
8799 struct rq *rq = cpu_rq(cpu);
8803 init_cfs_rq_runtime(cfs_rq);
8805 tg->cfs_rq[cpu] = cfs_rq;
8808 /* se could be NULL for root_task_group */
8813 se->cfs_rq = &rq->cfs;
8816 se->cfs_rq = parent->my_q;
8817 se->depth = parent->depth + 1;
8821 /* guarantee group entities always have weight */
8822 update_load_set(&se->load, NICE_0_LOAD);
8823 se->parent = parent;
8826 static DEFINE_MUTEX(shares_mutex);
8828 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8831 unsigned long flags;
8834 * We can't change the weight of the root cgroup.
8839 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8841 mutex_lock(&shares_mutex);
8842 if (tg->shares == shares)
8845 tg->shares = shares;
8846 for_each_possible_cpu(i) {
8847 struct rq *rq = cpu_rq(i);
8848 struct sched_entity *se;
8851 /* Propagate contribution to hierarchy */
8852 raw_spin_lock_irqsave(&rq->lock, flags);
8854 /* Possible calls to update_curr() need rq clock */
8855 update_rq_clock(rq);
8856 for_each_sched_entity(se)
8857 update_cfs_shares(group_cfs_rq(se));
8858 raw_spin_unlock_irqrestore(&rq->lock, flags);
8862 mutex_unlock(&shares_mutex);
8865 #else /* CONFIG_FAIR_GROUP_SCHED */
8867 void free_fair_sched_group(struct task_group *tg) { }
8869 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8874 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8876 #endif /* CONFIG_FAIR_GROUP_SCHED */
8879 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8881 struct sched_entity *se = &task->se;
8882 unsigned int rr_interval = 0;
8885 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8888 if (rq->cfs.load.weight)
8889 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8895 * All the scheduling class methods:
8897 const struct sched_class fair_sched_class = {
8898 .next = &idle_sched_class,
8899 .enqueue_task = enqueue_task_fair,
8900 .dequeue_task = dequeue_task_fair,
8901 .yield_task = yield_task_fair,
8902 .yield_to_task = yield_to_task_fair,
8904 .check_preempt_curr = check_preempt_wakeup,
8906 .pick_next_task = pick_next_task_fair,
8907 .put_prev_task = put_prev_task_fair,
8910 .select_task_rq = select_task_rq_fair,
8911 .migrate_task_rq = migrate_task_rq_fair,
8913 .rq_online = rq_online_fair,
8914 .rq_offline = rq_offline_fair,
8916 .task_waking = task_waking_fair,
8917 .task_dead = task_dead_fair,
8918 .set_cpus_allowed = set_cpus_allowed_common,
8921 .set_curr_task = set_curr_task_fair,
8922 .task_tick = task_tick_fair,
8923 .task_fork = task_fork_fair,
8925 .prio_changed = prio_changed_fair,
8926 .switched_from = switched_from_fair,
8927 .switched_to = switched_to_fair,
8929 .get_rr_interval = get_rr_interval_fair,
8931 .update_curr = update_curr_fair,
8933 #ifdef CONFIG_FAIR_GROUP_SCHED
8934 .task_move_group = task_move_group_fair,
8938 #ifdef CONFIG_SCHED_DEBUG
8939 void print_cfs_stats(struct seq_file *m, int cpu)
8941 struct cfs_rq *cfs_rq;
8944 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8945 print_cfs_rq(m, cpu, cfs_rq);
8949 #ifdef CONFIG_NUMA_BALANCING
8950 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8953 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8955 for_each_online_node(node) {
8956 if (p->numa_faults) {
8957 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8958 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8960 if (p->numa_group) {
8961 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8962 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8964 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8967 #endif /* CONFIG_NUMA_BALANCING */
8968 #endif /* CONFIG_SCHED_DEBUG */
8970 __init void init_sched_fair_class(void)
8973 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8975 #ifdef CONFIG_NO_HZ_COMMON
8976 nohz.next_balance = jiffies;
8977 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8978 cpu_notifier(sched_ilb_notifier, 0);