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 */
690 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
691 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
693 void init_entity_runnable_average(struct sched_entity *se)
699 * Update the current task's runtime statistics.
701 static void update_curr(struct cfs_rq *cfs_rq)
703 struct sched_entity *curr = cfs_rq->curr;
704 u64 now = rq_clock_task(rq_of(cfs_rq));
710 delta_exec = now - curr->exec_start;
711 if (unlikely((s64)delta_exec <= 0))
714 curr->exec_start = now;
716 schedstat_set(curr->statistics.exec_max,
717 max(delta_exec, curr->statistics.exec_max));
719 curr->sum_exec_runtime += delta_exec;
720 schedstat_add(cfs_rq, exec_clock, delta_exec);
722 curr->vruntime += calc_delta_fair(delta_exec, curr);
723 update_min_vruntime(cfs_rq);
725 if (entity_is_task(curr)) {
726 struct task_struct *curtask = task_of(curr);
728 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
729 cpuacct_charge(curtask, delta_exec);
730 account_group_exec_runtime(curtask, delta_exec);
733 account_cfs_rq_runtime(cfs_rq, delta_exec);
736 static void update_curr_fair(struct rq *rq)
738 update_curr(cfs_rq_of(&rq->curr->se));
741 #ifdef CONFIG_SCHEDSTATS
743 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
745 u64 wait_start = rq_clock(rq_of(cfs_rq));
747 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
748 likely(wait_start > se->statistics.wait_start))
749 wait_start -= se->statistics.wait_start;
751 se->statistics.wait_start = wait_start;
755 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
757 struct task_struct *p;
758 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start;
760 if (entity_is_task(se)) {
762 if (task_on_rq_migrating(p)) {
764 * Preserve migrating task's wait time so wait_start
765 * time stamp can be adjusted to accumulate wait time
766 * prior to migration.
768 se->statistics.wait_start = delta;
771 trace_sched_stat_wait(p, delta);
774 se->statistics.wait_max = max(se->statistics.wait_max, delta);
775 se->statistics.wait_count++;
776 se->statistics.wait_sum += delta;
777 se->statistics.wait_start = 0;
781 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
786 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
792 * Task is being enqueued - update stats:
794 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
797 * Are we enqueueing a waiting task? (for current tasks
798 * a dequeue/enqueue event is a NOP)
800 if (se != cfs_rq->curr)
801 update_stats_wait_start(cfs_rq, se);
805 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
808 * Mark the end of the wait period if dequeueing a
811 if (se != cfs_rq->curr)
812 update_stats_wait_end(cfs_rq, se);
816 * We are picking a new current task - update its stats:
819 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
822 * We are starting a new run period:
824 se->exec_start = rq_clock_task(rq_of(cfs_rq));
827 /**************************************************
828 * Scheduling class queueing methods:
831 #ifdef CONFIG_NUMA_BALANCING
833 * Approximate time to scan a full NUMA task in ms. The task scan period is
834 * calculated based on the tasks virtual memory size and
835 * numa_balancing_scan_size.
837 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
838 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
840 /* Portion of address space to scan in MB */
841 unsigned int sysctl_numa_balancing_scan_size = 256;
843 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
844 unsigned int sysctl_numa_balancing_scan_delay = 1000;
846 static unsigned int task_nr_scan_windows(struct task_struct *p)
848 unsigned long rss = 0;
849 unsigned long nr_scan_pages;
852 * Calculations based on RSS as non-present and empty pages are skipped
853 * by the PTE scanner and NUMA hinting faults should be trapped based
856 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
857 rss = get_mm_rss(p->mm);
861 rss = round_up(rss, nr_scan_pages);
862 return rss / nr_scan_pages;
865 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
866 #define MAX_SCAN_WINDOW 2560
868 static unsigned int task_scan_min(struct task_struct *p)
870 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
871 unsigned int scan, floor;
872 unsigned int windows = 1;
874 if (scan_size < MAX_SCAN_WINDOW)
875 windows = MAX_SCAN_WINDOW / scan_size;
876 floor = 1000 / windows;
878 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
879 return max_t(unsigned int, floor, scan);
882 static unsigned int task_scan_max(struct task_struct *p)
884 unsigned int smin = task_scan_min(p);
887 /* Watch for min being lower than max due to floor calculations */
888 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
889 return max(smin, smax);
892 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
894 rq->nr_numa_running += (p->numa_preferred_nid != -1);
895 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
898 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
900 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
901 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
907 spinlock_t lock; /* nr_tasks, tasks */
912 nodemask_t active_nodes;
913 unsigned long total_faults;
915 * Faults_cpu is used to decide whether memory should move
916 * towards the CPU. As a consequence, these stats are weighted
917 * more by CPU use than by memory faults.
919 unsigned long *faults_cpu;
920 unsigned long faults[0];
923 /* Shared or private faults. */
924 #define NR_NUMA_HINT_FAULT_TYPES 2
926 /* Memory and CPU locality */
927 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
929 /* Averaged statistics, and temporary buffers. */
930 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
932 pid_t task_numa_group_id(struct task_struct *p)
934 return p->numa_group ? p->numa_group->gid : 0;
938 * The averaged statistics, shared & private, memory & cpu,
939 * occupy the first half of the array. The second half of the
940 * array is for current counters, which are averaged into the
941 * first set by task_numa_placement.
943 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
945 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
948 static inline unsigned long task_faults(struct task_struct *p, int nid)
953 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
954 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
957 static inline unsigned long group_faults(struct task_struct *p, int nid)
962 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
963 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
966 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
968 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
969 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
972 /* Handle placement on systems where not all nodes are directly connected. */
973 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
974 int maxdist, bool task)
976 unsigned long score = 0;
980 * All nodes are directly connected, and the same distance
981 * from each other. No need for fancy placement algorithms.
983 if (sched_numa_topology_type == NUMA_DIRECT)
987 * This code is called for each node, introducing N^2 complexity,
988 * which should be ok given the number of nodes rarely exceeds 8.
990 for_each_online_node(node) {
991 unsigned long faults;
992 int dist = node_distance(nid, node);
995 * The furthest away nodes in the system are not interesting
996 * for placement; nid was already counted.
998 if (dist == sched_max_numa_distance || node == nid)
1002 * On systems with a backplane NUMA topology, compare groups
1003 * of nodes, and move tasks towards the group with the most
1004 * memory accesses. When comparing two nodes at distance
1005 * "hoplimit", only nodes closer by than "hoplimit" are part
1006 * of each group. Skip other nodes.
1008 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1012 /* Add up the faults from nearby nodes. */
1014 faults = task_faults(p, node);
1016 faults = group_faults(p, node);
1019 * On systems with a glueless mesh NUMA topology, there are
1020 * no fixed "groups of nodes". Instead, nodes that are not
1021 * directly connected bounce traffic through intermediate
1022 * nodes; a numa_group can occupy any set of nodes.
1023 * The further away a node is, the less the faults count.
1024 * This seems to result in good task placement.
1026 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1027 faults *= (sched_max_numa_distance - dist);
1028 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1038 * These return the fraction of accesses done by a particular task, or
1039 * task group, on a particular numa node. The group weight is given a
1040 * larger multiplier, in order to group tasks together that are almost
1041 * evenly spread out between numa nodes.
1043 static inline unsigned long task_weight(struct task_struct *p, int nid,
1046 unsigned long faults, total_faults;
1048 if (!p->numa_faults)
1051 total_faults = p->total_numa_faults;
1056 faults = task_faults(p, nid);
1057 faults += score_nearby_nodes(p, nid, dist, true);
1059 return 1000 * faults / total_faults;
1062 static inline unsigned long group_weight(struct task_struct *p, int nid,
1065 unsigned long faults, total_faults;
1070 total_faults = p->numa_group->total_faults;
1075 faults = group_faults(p, nid);
1076 faults += score_nearby_nodes(p, nid, dist, false);
1078 return 1000 * faults / total_faults;
1081 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1082 int src_nid, int dst_cpu)
1084 struct numa_group *ng = p->numa_group;
1085 int dst_nid = cpu_to_node(dst_cpu);
1086 int last_cpupid, this_cpupid;
1088 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1091 * Multi-stage node selection is used in conjunction with a periodic
1092 * migration fault to build a temporal task<->page relation. By using
1093 * a two-stage filter we remove short/unlikely relations.
1095 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1096 * a task's usage of a particular page (n_p) per total usage of this
1097 * page (n_t) (in a given time-span) to a probability.
1099 * Our periodic faults will sample this probability and getting the
1100 * same result twice in a row, given these samples are fully
1101 * independent, is then given by P(n)^2, provided our sample period
1102 * is sufficiently short compared to the usage pattern.
1104 * This quadric squishes small probabilities, making it less likely we
1105 * act on an unlikely task<->page relation.
1107 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1108 if (!cpupid_pid_unset(last_cpupid) &&
1109 cpupid_to_nid(last_cpupid) != dst_nid)
1112 /* Always allow migrate on private faults */
1113 if (cpupid_match_pid(p, last_cpupid))
1116 /* A shared fault, but p->numa_group has not been set up yet. */
1121 * Do not migrate if the destination is not a node that
1122 * is actively used by this numa group.
1124 if (!node_isset(dst_nid, ng->active_nodes))
1128 * Source is a node that is not actively used by this
1129 * numa group, while the destination is. Migrate.
1131 if (!node_isset(src_nid, ng->active_nodes))
1135 * Both source and destination are nodes in active
1136 * use by this numa group. Maximize memory bandwidth
1137 * by migrating from more heavily used groups, to less
1138 * heavily used ones, spreading the load around.
1139 * Use a 1/4 hysteresis to avoid spurious page movement.
1141 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1144 static unsigned long weighted_cpuload(const int cpu);
1145 static unsigned long source_load(int cpu, int type);
1146 static unsigned long target_load(int cpu, int type);
1147 static unsigned long capacity_of(int cpu);
1148 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1150 /* Cached statistics for all CPUs within a node */
1152 unsigned long nr_running;
1155 /* Total compute capacity of CPUs on a node */
1156 unsigned long compute_capacity;
1158 /* Approximate capacity in terms of runnable tasks on a node */
1159 unsigned long task_capacity;
1160 int has_free_capacity;
1164 * XXX borrowed from update_sg_lb_stats
1166 static void update_numa_stats(struct numa_stats *ns, int nid)
1168 int smt, cpu, cpus = 0;
1169 unsigned long capacity;
1171 memset(ns, 0, sizeof(*ns));
1172 for_each_cpu(cpu, cpumask_of_node(nid)) {
1173 struct rq *rq = cpu_rq(cpu);
1175 ns->nr_running += rq->nr_running;
1176 ns->load += weighted_cpuload(cpu);
1177 ns->compute_capacity += capacity_of(cpu);
1183 * If we raced with hotplug and there are no CPUs left in our mask
1184 * the @ns structure is NULL'ed and task_numa_compare() will
1185 * not find this node attractive.
1187 * We'll either bail at !has_free_capacity, or we'll detect a huge
1188 * imbalance and bail there.
1193 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1194 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1195 capacity = cpus / smt; /* cores */
1197 ns->task_capacity = min_t(unsigned, capacity,
1198 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1199 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1202 struct task_numa_env {
1203 struct task_struct *p;
1205 int src_cpu, src_nid;
1206 int dst_cpu, dst_nid;
1208 struct numa_stats src_stats, dst_stats;
1213 struct task_struct *best_task;
1218 static void task_numa_assign(struct task_numa_env *env,
1219 struct task_struct *p, long imp)
1222 put_task_struct(env->best_task);
1225 env->best_imp = imp;
1226 env->best_cpu = env->dst_cpu;
1229 static bool load_too_imbalanced(long src_load, long dst_load,
1230 struct task_numa_env *env)
1233 long orig_src_load, orig_dst_load;
1234 long src_capacity, dst_capacity;
1237 * The load is corrected for the CPU capacity available on each node.
1240 * ------------ vs ---------
1241 * src_capacity dst_capacity
1243 src_capacity = env->src_stats.compute_capacity;
1244 dst_capacity = env->dst_stats.compute_capacity;
1246 /* We care about the slope of the imbalance, not the direction. */
1247 if (dst_load < src_load)
1248 swap(dst_load, src_load);
1250 /* Is the difference below the threshold? */
1251 imb = dst_load * src_capacity * 100 -
1252 src_load * dst_capacity * env->imbalance_pct;
1257 * The imbalance is above the allowed threshold.
1258 * Compare it with the old imbalance.
1260 orig_src_load = env->src_stats.load;
1261 orig_dst_load = env->dst_stats.load;
1263 if (orig_dst_load < orig_src_load)
1264 swap(orig_dst_load, orig_src_load);
1266 old_imb = orig_dst_load * src_capacity * 100 -
1267 orig_src_load * dst_capacity * env->imbalance_pct;
1269 /* Would this change make things worse? */
1270 return (imb > old_imb);
1274 * This checks if the overall compute and NUMA accesses of the system would
1275 * be improved if the source tasks was migrated to the target dst_cpu taking
1276 * into account that it might be best if task running on the dst_cpu should
1277 * be exchanged with the source task
1279 static void task_numa_compare(struct task_numa_env *env,
1280 long taskimp, long groupimp)
1282 struct rq *src_rq = cpu_rq(env->src_cpu);
1283 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1284 struct task_struct *cur;
1285 long src_load, dst_load;
1287 long imp = env->p->numa_group ? groupimp : taskimp;
1289 int dist = env->dist;
1290 bool assigned = false;
1294 raw_spin_lock_irq(&dst_rq->lock);
1297 * No need to move the exiting task or idle task.
1299 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1303 * The task_struct must be protected here to protect the
1304 * p->numa_faults access in the task_weight since the
1305 * numa_faults could already be freed in the following path:
1306 * finish_task_switch()
1307 * --> put_task_struct()
1308 * --> __put_task_struct()
1309 * --> task_numa_free()
1311 get_task_struct(cur);
1314 raw_spin_unlock_irq(&dst_rq->lock);
1317 * Because we have preemption enabled we can get migrated around and
1318 * end try selecting ourselves (current == env->p) as a swap candidate.
1324 * "imp" is the fault differential for the source task between the
1325 * source and destination node. Calculate the total differential for
1326 * the source task and potential destination task. The more negative
1327 * the value is, the more rmeote accesses that would be expected to
1328 * be incurred if the tasks were swapped.
1331 /* Skip this swap candidate if cannot move to the source cpu */
1332 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1336 * If dst and source tasks are in the same NUMA group, or not
1337 * in any group then look only at task weights.
1339 if (cur->numa_group == env->p->numa_group) {
1340 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1341 task_weight(cur, env->dst_nid, dist);
1343 * Add some hysteresis to prevent swapping the
1344 * tasks within a group over tiny differences.
1346 if (cur->numa_group)
1350 * Compare the group weights. If a task is all by
1351 * itself (not part of a group), use the task weight
1354 if (cur->numa_group)
1355 imp += group_weight(cur, env->src_nid, dist) -
1356 group_weight(cur, env->dst_nid, dist);
1358 imp += task_weight(cur, env->src_nid, dist) -
1359 task_weight(cur, env->dst_nid, dist);
1363 if (imp <= env->best_imp && moveimp <= env->best_imp)
1367 /* Is there capacity at our destination? */
1368 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1369 !env->dst_stats.has_free_capacity)
1375 /* Balance doesn't matter much if we're running a task per cpu */
1376 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1377 dst_rq->nr_running == 1)
1381 * In the overloaded case, try and keep the load balanced.
1384 load = task_h_load(env->p);
1385 dst_load = env->dst_stats.load + load;
1386 src_load = env->src_stats.load - load;
1388 if (moveimp > imp && moveimp > env->best_imp) {
1390 * If the improvement from just moving env->p direction is
1391 * better than swapping tasks around, check if a move is
1392 * possible. Store a slightly smaller score than moveimp,
1393 * so an actually idle CPU will win.
1395 if (!load_too_imbalanced(src_load, dst_load, env)) {
1397 put_task_struct(cur);
1403 if (imp <= env->best_imp)
1407 load = task_h_load(cur);
1412 if (load_too_imbalanced(src_load, dst_load, env))
1416 * One idle CPU per node is evaluated for a task numa move.
1417 * Call select_idle_sibling to maybe find a better one.
1420 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1424 task_numa_assign(env, cur, imp);
1428 * The dst_rq->curr isn't assigned. The protection for task_struct is
1431 if (cur && !assigned)
1432 put_task_struct(cur);
1435 static void task_numa_find_cpu(struct task_numa_env *env,
1436 long taskimp, long groupimp)
1440 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1441 /* Skip this CPU if the source task cannot migrate */
1442 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1446 task_numa_compare(env, taskimp, groupimp);
1450 /* Only move tasks to a NUMA node less busy than the current node. */
1451 static bool numa_has_capacity(struct task_numa_env *env)
1453 struct numa_stats *src = &env->src_stats;
1454 struct numa_stats *dst = &env->dst_stats;
1456 if (src->has_free_capacity && !dst->has_free_capacity)
1460 * Only consider a task move if the source has a higher load
1461 * than the destination, corrected for CPU capacity on each node.
1463 * src->load dst->load
1464 * --------------------- vs ---------------------
1465 * src->compute_capacity dst->compute_capacity
1467 if (src->load * dst->compute_capacity * env->imbalance_pct >
1469 dst->load * src->compute_capacity * 100)
1475 static int task_numa_migrate(struct task_struct *p)
1477 struct task_numa_env env = {
1480 .src_cpu = task_cpu(p),
1481 .src_nid = task_node(p),
1483 .imbalance_pct = 112,
1489 struct sched_domain *sd;
1490 unsigned long taskweight, groupweight;
1492 long taskimp, groupimp;
1495 * Pick the lowest SD_NUMA domain, as that would have the smallest
1496 * imbalance and would be the first to start moving tasks about.
1498 * And we want to avoid any moving of tasks about, as that would create
1499 * random movement of tasks -- counter the numa conditions we're trying
1503 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1505 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1509 * Cpusets can break the scheduler domain tree into smaller
1510 * balance domains, some of which do not cross NUMA boundaries.
1511 * Tasks that are "trapped" in such domains cannot be migrated
1512 * elsewhere, so there is no point in (re)trying.
1514 if (unlikely(!sd)) {
1515 p->numa_preferred_nid = task_node(p);
1519 env.dst_nid = p->numa_preferred_nid;
1520 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1521 taskweight = task_weight(p, env.src_nid, dist);
1522 groupweight = group_weight(p, env.src_nid, dist);
1523 update_numa_stats(&env.src_stats, env.src_nid);
1524 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1525 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1526 update_numa_stats(&env.dst_stats, env.dst_nid);
1528 /* Try to find a spot on the preferred nid. */
1529 if (numa_has_capacity(&env))
1530 task_numa_find_cpu(&env, taskimp, groupimp);
1533 * Look at other nodes in these cases:
1534 * - there is no space available on the preferred_nid
1535 * - the task is part of a numa_group that is interleaved across
1536 * multiple NUMA nodes; in order to better consolidate the group,
1537 * we need to check other locations.
1539 if (env.best_cpu == -1 || (p->numa_group &&
1540 nodes_weight(p->numa_group->active_nodes) > 1)) {
1541 for_each_online_node(nid) {
1542 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1545 dist = node_distance(env.src_nid, env.dst_nid);
1546 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1548 taskweight = task_weight(p, env.src_nid, dist);
1549 groupweight = group_weight(p, env.src_nid, dist);
1552 /* Only consider nodes where both task and groups benefit */
1553 taskimp = task_weight(p, nid, dist) - taskweight;
1554 groupimp = group_weight(p, nid, dist) - groupweight;
1555 if (taskimp < 0 && groupimp < 0)
1560 update_numa_stats(&env.dst_stats, env.dst_nid);
1561 if (numa_has_capacity(&env))
1562 task_numa_find_cpu(&env, taskimp, groupimp);
1567 * If the task is part of a workload that spans multiple NUMA nodes,
1568 * and is migrating into one of the workload's active nodes, remember
1569 * this node as the task's preferred numa node, so the workload can
1571 * A task that migrated to a second choice node will be better off
1572 * trying for a better one later. Do not set the preferred node here.
1574 if (p->numa_group) {
1575 if (env.best_cpu == -1)
1580 if (node_isset(nid, p->numa_group->active_nodes))
1581 sched_setnuma(p, env.dst_nid);
1584 /* No better CPU than the current one was found. */
1585 if (env.best_cpu == -1)
1589 * Reset the scan period if the task is being rescheduled on an
1590 * alternative node to recheck if the tasks is now properly placed.
1592 p->numa_scan_period = task_scan_min(p);
1594 if (env.best_task == NULL) {
1595 ret = migrate_task_to(p, env.best_cpu);
1597 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1601 ret = migrate_swap(p, env.best_task);
1603 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1604 put_task_struct(env.best_task);
1608 /* Attempt to migrate a task to a CPU on the preferred node. */
1609 static void numa_migrate_preferred(struct task_struct *p)
1611 unsigned long interval = HZ;
1613 /* This task has no NUMA fault statistics yet */
1614 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1617 /* Periodically retry migrating the task to the preferred node */
1618 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1619 p->numa_migrate_retry = jiffies + interval;
1621 /* Success if task is already running on preferred CPU */
1622 if (task_node(p) == p->numa_preferred_nid)
1625 /* Otherwise, try migrate to a CPU on the preferred node */
1626 task_numa_migrate(p);
1630 * Find the nodes on which the workload is actively running. We do this by
1631 * tracking the nodes from which NUMA hinting faults are triggered. This can
1632 * be different from the set of nodes where the workload's memory is currently
1635 * The bitmask is used to make smarter decisions on when to do NUMA page
1636 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1637 * are added when they cause over 6/16 of the maximum number of faults, but
1638 * only removed when they drop below 3/16.
1640 static void update_numa_active_node_mask(struct numa_group *numa_group)
1642 unsigned long faults, max_faults = 0;
1645 for_each_online_node(nid) {
1646 faults = group_faults_cpu(numa_group, nid);
1647 if (faults > max_faults)
1648 max_faults = faults;
1651 for_each_online_node(nid) {
1652 faults = group_faults_cpu(numa_group, nid);
1653 if (!node_isset(nid, numa_group->active_nodes)) {
1654 if (faults > max_faults * 6 / 16)
1655 node_set(nid, numa_group->active_nodes);
1656 } else if (faults < max_faults * 3 / 16)
1657 node_clear(nid, numa_group->active_nodes);
1662 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1663 * increments. The more local the fault statistics are, the higher the scan
1664 * period will be for the next scan window. If local/(local+remote) ratio is
1665 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1666 * the scan period will decrease. Aim for 70% local accesses.
1668 #define NUMA_PERIOD_SLOTS 10
1669 #define NUMA_PERIOD_THRESHOLD 7
1672 * Increase the scan period (slow down scanning) if the majority of
1673 * our memory is already on our local node, or if the majority of
1674 * the page accesses are shared with other processes.
1675 * Otherwise, decrease the scan period.
1677 static void update_task_scan_period(struct task_struct *p,
1678 unsigned long shared, unsigned long private)
1680 unsigned int period_slot;
1684 unsigned long remote = p->numa_faults_locality[0];
1685 unsigned long local = p->numa_faults_locality[1];
1688 * If there were no record hinting faults then either the task is
1689 * completely idle or all activity is areas that are not of interest
1690 * to automatic numa balancing. Related to that, if there were failed
1691 * migration then it implies we are migrating too quickly or the local
1692 * node is overloaded. In either case, scan slower
1694 if (local + shared == 0 || p->numa_faults_locality[2]) {
1695 p->numa_scan_period = min(p->numa_scan_period_max,
1696 p->numa_scan_period << 1);
1698 p->mm->numa_next_scan = jiffies +
1699 msecs_to_jiffies(p->numa_scan_period);
1705 * Prepare to scale scan period relative to the current period.
1706 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1707 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1708 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1710 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1711 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1712 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1713 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1716 diff = slot * period_slot;
1718 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1721 * Scale scan rate increases based on sharing. There is an
1722 * inverse relationship between the degree of sharing and
1723 * the adjustment made to the scanning period. Broadly
1724 * speaking the intent is that there is little point
1725 * scanning faster if shared accesses dominate as it may
1726 * simply bounce migrations uselessly
1728 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1729 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1732 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1733 task_scan_min(p), task_scan_max(p));
1734 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1738 * Get the fraction of time the task has been running since the last
1739 * NUMA placement cycle. The scheduler keeps similar statistics, but
1740 * decays those on a 32ms period, which is orders of magnitude off
1741 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1742 * stats only if the task is so new there are no NUMA statistics yet.
1744 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1746 u64 runtime, delta, now;
1747 /* Use the start of this time slice to avoid calculations. */
1748 now = p->se.exec_start;
1749 runtime = p->se.sum_exec_runtime;
1751 if (p->last_task_numa_placement) {
1752 delta = runtime - p->last_sum_exec_runtime;
1753 *period = now - p->last_task_numa_placement;
1755 delta = p->se.avg.load_sum / p->se.load.weight;
1756 *period = LOAD_AVG_MAX;
1759 p->last_sum_exec_runtime = runtime;
1760 p->last_task_numa_placement = now;
1766 * Determine the preferred nid for a task in a numa_group. This needs to
1767 * be done in a way that produces consistent results with group_weight,
1768 * otherwise workloads might not converge.
1770 static int preferred_group_nid(struct task_struct *p, int nid)
1775 /* Direct connections between all NUMA nodes. */
1776 if (sched_numa_topology_type == NUMA_DIRECT)
1780 * On a system with glueless mesh NUMA topology, group_weight
1781 * scores nodes according to the number of NUMA hinting faults on
1782 * both the node itself, and on nearby nodes.
1784 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1785 unsigned long score, max_score = 0;
1786 int node, max_node = nid;
1788 dist = sched_max_numa_distance;
1790 for_each_online_node(node) {
1791 score = group_weight(p, node, dist);
1792 if (score > max_score) {
1801 * Finding the preferred nid in a system with NUMA backplane
1802 * interconnect topology is more involved. The goal is to locate
1803 * tasks from numa_groups near each other in the system, and
1804 * untangle workloads from different sides of the system. This requires
1805 * searching down the hierarchy of node groups, recursively searching
1806 * inside the highest scoring group of nodes. The nodemask tricks
1807 * keep the complexity of the search down.
1809 nodes = node_online_map;
1810 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1811 unsigned long max_faults = 0;
1812 nodemask_t max_group = NODE_MASK_NONE;
1815 /* Are there nodes at this distance from each other? */
1816 if (!find_numa_distance(dist))
1819 for_each_node_mask(a, nodes) {
1820 unsigned long faults = 0;
1821 nodemask_t this_group;
1822 nodes_clear(this_group);
1824 /* Sum group's NUMA faults; includes a==b case. */
1825 for_each_node_mask(b, nodes) {
1826 if (node_distance(a, b) < dist) {
1827 faults += group_faults(p, b);
1828 node_set(b, this_group);
1829 node_clear(b, nodes);
1833 /* Remember the top group. */
1834 if (faults > max_faults) {
1835 max_faults = faults;
1836 max_group = this_group;
1838 * subtle: at the smallest distance there is
1839 * just one node left in each "group", the
1840 * winner is the preferred nid.
1845 /* Next round, evaluate the nodes within max_group. */
1853 static void task_numa_placement(struct task_struct *p)
1855 int seq, nid, max_nid = -1, max_group_nid = -1;
1856 unsigned long max_faults = 0, max_group_faults = 0;
1857 unsigned long fault_types[2] = { 0, 0 };
1858 unsigned long total_faults;
1859 u64 runtime, period;
1860 spinlock_t *group_lock = NULL;
1863 * The p->mm->numa_scan_seq field gets updated without
1864 * exclusive access. Use READ_ONCE() here to ensure
1865 * that the field is read in a single access:
1867 seq = READ_ONCE(p->mm->numa_scan_seq);
1868 if (p->numa_scan_seq == seq)
1870 p->numa_scan_seq = seq;
1871 p->numa_scan_period_max = task_scan_max(p);
1873 total_faults = p->numa_faults_locality[0] +
1874 p->numa_faults_locality[1];
1875 runtime = numa_get_avg_runtime(p, &period);
1877 /* If the task is part of a group prevent parallel updates to group stats */
1878 if (p->numa_group) {
1879 group_lock = &p->numa_group->lock;
1880 spin_lock_irq(group_lock);
1883 /* Find the node with the highest number of faults */
1884 for_each_online_node(nid) {
1885 /* Keep track of the offsets in numa_faults array */
1886 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1887 unsigned long faults = 0, group_faults = 0;
1890 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1891 long diff, f_diff, f_weight;
1893 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1894 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1895 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1896 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1898 /* Decay existing window, copy faults since last scan */
1899 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1900 fault_types[priv] += p->numa_faults[membuf_idx];
1901 p->numa_faults[membuf_idx] = 0;
1904 * Normalize the faults_from, so all tasks in a group
1905 * count according to CPU use, instead of by the raw
1906 * number of faults. Tasks with little runtime have
1907 * little over-all impact on throughput, and thus their
1908 * faults are less important.
1910 f_weight = div64_u64(runtime << 16, period + 1);
1911 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1913 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1914 p->numa_faults[cpubuf_idx] = 0;
1916 p->numa_faults[mem_idx] += diff;
1917 p->numa_faults[cpu_idx] += f_diff;
1918 faults += p->numa_faults[mem_idx];
1919 p->total_numa_faults += diff;
1920 if (p->numa_group) {
1922 * safe because we can only change our own group
1924 * mem_idx represents the offset for a given
1925 * nid and priv in a specific region because it
1926 * is at the beginning of the numa_faults array.
1928 p->numa_group->faults[mem_idx] += diff;
1929 p->numa_group->faults_cpu[mem_idx] += f_diff;
1930 p->numa_group->total_faults += diff;
1931 group_faults += p->numa_group->faults[mem_idx];
1935 if (faults > max_faults) {
1936 max_faults = faults;
1940 if (group_faults > max_group_faults) {
1941 max_group_faults = group_faults;
1942 max_group_nid = nid;
1946 update_task_scan_period(p, fault_types[0], fault_types[1]);
1948 if (p->numa_group) {
1949 update_numa_active_node_mask(p->numa_group);
1950 spin_unlock_irq(group_lock);
1951 max_nid = preferred_group_nid(p, max_group_nid);
1955 /* Set the new preferred node */
1956 if (max_nid != p->numa_preferred_nid)
1957 sched_setnuma(p, max_nid);
1959 if (task_node(p) != p->numa_preferred_nid)
1960 numa_migrate_preferred(p);
1964 static inline int get_numa_group(struct numa_group *grp)
1966 return atomic_inc_not_zero(&grp->refcount);
1969 static inline void put_numa_group(struct numa_group *grp)
1971 if (atomic_dec_and_test(&grp->refcount))
1972 kfree_rcu(grp, rcu);
1975 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1978 struct numa_group *grp, *my_grp;
1979 struct task_struct *tsk;
1981 int cpu = cpupid_to_cpu(cpupid);
1984 if (unlikely(!p->numa_group)) {
1985 unsigned int size = sizeof(struct numa_group) +
1986 4*nr_node_ids*sizeof(unsigned long);
1988 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1992 atomic_set(&grp->refcount, 1);
1993 spin_lock_init(&grp->lock);
1995 /* Second half of the array tracks nids where faults happen */
1996 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1999 node_set(task_node(current), grp->active_nodes);
2001 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2002 grp->faults[i] = p->numa_faults[i];
2004 grp->total_faults = p->total_numa_faults;
2007 rcu_assign_pointer(p->numa_group, grp);
2011 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2013 if (!cpupid_match_pid(tsk, cpupid))
2016 grp = rcu_dereference(tsk->numa_group);
2020 my_grp = p->numa_group;
2025 * Only join the other group if its bigger; if we're the bigger group,
2026 * the other task will join us.
2028 if (my_grp->nr_tasks > grp->nr_tasks)
2032 * Tie-break on the grp address.
2034 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2037 /* Always join threads in the same process. */
2038 if (tsk->mm == current->mm)
2041 /* Simple filter to avoid false positives due to PID collisions */
2042 if (flags & TNF_SHARED)
2045 /* Update priv based on whether false sharing was detected */
2048 if (join && !get_numa_group(grp))
2056 BUG_ON(irqs_disabled());
2057 double_lock_irq(&my_grp->lock, &grp->lock);
2059 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2060 my_grp->faults[i] -= p->numa_faults[i];
2061 grp->faults[i] += p->numa_faults[i];
2063 my_grp->total_faults -= p->total_numa_faults;
2064 grp->total_faults += p->total_numa_faults;
2069 spin_unlock(&my_grp->lock);
2070 spin_unlock_irq(&grp->lock);
2072 rcu_assign_pointer(p->numa_group, grp);
2074 put_numa_group(my_grp);
2082 void task_numa_free(struct task_struct *p)
2084 struct numa_group *grp = p->numa_group;
2085 void *numa_faults = p->numa_faults;
2086 unsigned long flags;
2090 spin_lock_irqsave(&grp->lock, flags);
2091 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2092 grp->faults[i] -= p->numa_faults[i];
2093 grp->total_faults -= p->total_numa_faults;
2096 spin_unlock_irqrestore(&grp->lock, flags);
2097 RCU_INIT_POINTER(p->numa_group, NULL);
2098 put_numa_group(grp);
2101 p->numa_faults = NULL;
2106 * Got a PROT_NONE fault for a page on @node.
2108 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2110 struct task_struct *p = current;
2111 bool migrated = flags & TNF_MIGRATED;
2112 int cpu_node = task_node(current);
2113 int local = !!(flags & TNF_FAULT_LOCAL);
2116 if (!static_branch_likely(&sched_numa_balancing))
2119 /* for example, ksmd faulting in a user's mm */
2123 /* Allocate buffer to track faults on a per-node basis */
2124 if (unlikely(!p->numa_faults)) {
2125 int size = sizeof(*p->numa_faults) *
2126 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2128 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2129 if (!p->numa_faults)
2132 p->total_numa_faults = 0;
2133 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2137 * First accesses are treated as private, otherwise consider accesses
2138 * to be private if the accessing pid has not changed
2140 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2143 priv = cpupid_match_pid(p, last_cpupid);
2144 if (!priv && !(flags & TNF_NO_GROUP))
2145 task_numa_group(p, last_cpupid, flags, &priv);
2149 * If a workload spans multiple NUMA nodes, a shared fault that
2150 * occurs wholly within the set of nodes that the workload is
2151 * actively using should be counted as local. This allows the
2152 * scan rate to slow down when a workload has settled down.
2154 if (!priv && !local && p->numa_group &&
2155 node_isset(cpu_node, p->numa_group->active_nodes) &&
2156 node_isset(mem_node, p->numa_group->active_nodes))
2159 task_numa_placement(p);
2162 * Retry task to preferred node migration periodically, in case it
2163 * case it previously failed, or the scheduler moved us.
2165 if (time_after(jiffies, p->numa_migrate_retry))
2166 numa_migrate_preferred(p);
2169 p->numa_pages_migrated += pages;
2170 if (flags & TNF_MIGRATE_FAIL)
2171 p->numa_faults_locality[2] += pages;
2173 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2174 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2175 p->numa_faults_locality[local] += pages;
2178 static void reset_ptenuma_scan(struct task_struct *p)
2181 * We only did a read acquisition of the mmap sem, so
2182 * p->mm->numa_scan_seq is written to without exclusive access
2183 * and the update is not guaranteed to be atomic. That's not
2184 * much of an issue though, since this is just used for
2185 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2186 * expensive, to avoid any form of compiler optimizations:
2188 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2189 p->mm->numa_scan_offset = 0;
2193 * The expensive part of numa migration is done from task_work context.
2194 * Triggered from task_tick_numa().
2196 void task_numa_work(struct callback_head *work)
2198 unsigned long migrate, next_scan, now = jiffies;
2199 struct task_struct *p = current;
2200 struct mm_struct *mm = p->mm;
2201 u64 runtime = p->se.sum_exec_runtime;
2202 struct vm_area_struct *vma;
2203 unsigned long start, end;
2204 unsigned long nr_pte_updates = 0;
2205 long pages, virtpages;
2207 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2209 work->next = work; /* protect against double add */
2211 * Who cares about NUMA placement when they're dying.
2213 * NOTE: make sure not to dereference p->mm before this check,
2214 * exit_task_work() happens _after_ exit_mm() so we could be called
2215 * without p->mm even though we still had it when we enqueued this
2218 if (p->flags & PF_EXITING)
2221 if (!mm->numa_next_scan) {
2222 mm->numa_next_scan = now +
2223 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2227 * Enforce maximal scan/migration frequency..
2229 migrate = mm->numa_next_scan;
2230 if (time_before(now, migrate))
2233 if (p->numa_scan_period == 0) {
2234 p->numa_scan_period_max = task_scan_max(p);
2235 p->numa_scan_period = task_scan_min(p);
2238 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2239 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2243 * Delay this task enough that another task of this mm will likely win
2244 * the next time around.
2246 p->node_stamp += 2 * TICK_NSEC;
2248 start = mm->numa_scan_offset;
2249 pages = sysctl_numa_balancing_scan_size;
2250 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2251 virtpages = pages * 8; /* Scan up to this much virtual space */
2256 down_read(&mm->mmap_sem);
2257 vma = find_vma(mm, start);
2259 reset_ptenuma_scan(p);
2263 for (; vma; vma = vma->vm_next) {
2264 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2265 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2270 * Shared library pages mapped by multiple processes are not
2271 * migrated as it is expected they are cache replicated. Avoid
2272 * hinting faults in read-only file-backed mappings or the vdso
2273 * as migrating the pages will be of marginal benefit.
2276 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2280 * Skip inaccessible VMAs to avoid any confusion between
2281 * PROT_NONE and NUMA hinting ptes
2283 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2287 start = max(start, vma->vm_start);
2288 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2289 end = min(end, vma->vm_end);
2290 nr_pte_updates = change_prot_numa(vma, start, end);
2293 * Try to scan sysctl_numa_balancing_size worth of
2294 * hpages that have at least one present PTE that
2295 * is not already pte-numa. If the VMA contains
2296 * areas that are unused or already full of prot_numa
2297 * PTEs, scan up to virtpages, to skip through those
2301 pages -= (end - start) >> PAGE_SHIFT;
2302 virtpages -= (end - start) >> PAGE_SHIFT;
2305 if (pages <= 0 || virtpages <= 0)
2309 } while (end != vma->vm_end);
2314 * It is possible to reach the end of the VMA list but the last few
2315 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2316 * would find the !migratable VMA on the next scan but not reset the
2317 * scanner to the start so check it now.
2320 mm->numa_scan_offset = start;
2322 reset_ptenuma_scan(p);
2323 up_read(&mm->mmap_sem);
2326 * Make sure tasks use at least 32x as much time to run other code
2327 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2328 * Usually update_task_scan_period slows down scanning enough; on an
2329 * overloaded system we need to limit overhead on a per task basis.
2331 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2332 u64 diff = p->se.sum_exec_runtime - runtime;
2333 p->node_stamp += 32 * diff;
2338 * Drive the periodic memory faults..
2340 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2342 struct callback_head *work = &curr->numa_work;
2346 * We don't care about NUMA placement if we don't have memory.
2348 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2352 * Using runtime rather than walltime has the dual advantage that
2353 * we (mostly) drive the selection from busy threads and that the
2354 * task needs to have done some actual work before we bother with
2357 now = curr->se.sum_exec_runtime;
2358 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2360 if (now > curr->node_stamp + period) {
2361 if (!curr->node_stamp)
2362 curr->numa_scan_period = task_scan_min(curr);
2363 curr->node_stamp += period;
2365 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2366 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2367 task_work_add(curr, work, true);
2372 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2376 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2380 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2383 #endif /* CONFIG_NUMA_BALANCING */
2386 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2388 update_load_add(&cfs_rq->load, se->load.weight);
2389 if (!parent_entity(se))
2390 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2392 if (entity_is_task(se)) {
2393 struct rq *rq = rq_of(cfs_rq);
2395 account_numa_enqueue(rq, task_of(se));
2396 list_add(&se->group_node, &rq->cfs_tasks);
2399 cfs_rq->nr_running++;
2403 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2405 update_load_sub(&cfs_rq->load, se->load.weight);
2406 if (!parent_entity(se))
2407 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2408 if (entity_is_task(se)) {
2409 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2410 list_del_init(&se->group_node);
2412 cfs_rq->nr_running--;
2415 #ifdef CONFIG_FAIR_GROUP_SCHED
2417 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2422 * Use this CPU's real-time load instead of the last load contribution
2423 * as the updating of the contribution is delayed, and we will use the
2424 * the real-time load to calc the share. See update_tg_load_avg().
2426 tg_weight = atomic_long_read(&tg->load_avg);
2427 tg_weight -= cfs_rq->tg_load_avg_contrib;
2428 tg_weight += cfs_rq->load.weight;
2433 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2435 long tg_weight, load, shares;
2437 tg_weight = calc_tg_weight(tg, cfs_rq);
2438 load = cfs_rq->load.weight;
2440 shares = (tg->shares * load);
2442 shares /= tg_weight;
2444 if (shares < MIN_SHARES)
2445 shares = MIN_SHARES;
2446 if (shares > tg->shares)
2447 shares = tg->shares;
2451 # else /* CONFIG_SMP */
2452 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2456 # endif /* CONFIG_SMP */
2457 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2458 unsigned long weight)
2461 /* commit outstanding execution time */
2462 if (cfs_rq->curr == se)
2463 update_curr(cfs_rq);
2464 account_entity_dequeue(cfs_rq, se);
2467 update_load_set(&se->load, weight);
2470 account_entity_enqueue(cfs_rq, se);
2473 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2475 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2477 struct task_group *tg;
2478 struct sched_entity *se;
2482 se = tg->se[cpu_of(rq_of(cfs_rq))];
2483 if (!se || throttled_hierarchy(cfs_rq))
2486 if (likely(se->load.weight == tg->shares))
2489 shares = calc_cfs_shares(cfs_rq, tg);
2491 reweight_entity(cfs_rq_of(se), se, shares);
2493 #else /* CONFIG_FAIR_GROUP_SCHED */
2494 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2497 #endif /* CONFIG_FAIR_GROUP_SCHED */
2500 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2501 static const u32 runnable_avg_yN_inv[] = {
2502 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2503 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2504 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2505 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2506 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2507 0x85aac367, 0x82cd8698,
2511 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2512 * over-estimates when re-combining.
2514 static const u32 runnable_avg_yN_sum[] = {
2515 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2516 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2517 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2522 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2524 static __always_inline u64 decay_load(u64 val, u64 n)
2526 unsigned int local_n;
2530 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2533 /* after bounds checking we can collapse to 32-bit */
2537 * As y^PERIOD = 1/2, we can combine
2538 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2539 * With a look-up table which covers y^n (n<PERIOD)
2541 * To achieve constant time decay_load.
2543 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2544 val >>= local_n / LOAD_AVG_PERIOD;
2545 local_n %= LOAD_AVG_PERIOD;
2548 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2553 * For updates fully spanning n periods, the contribution to runnable
2554 * average will be: \Sum 1024*y^n
2556 * We can compute this reasonably efficiently by combining:
2557 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2559 static u32 __compute_runnable_contrib(u64 n)
2563 if (likely(n <= LOAD_AVG_PERIOD))
2564 return runnable_avg_yN_sum[n];
2565 else if (unlikely(n >= LOAD_AVG_MAX_N))
2566 return LOAD_AVG_MAX;
2568 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2570 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2571 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2573 n -= LOAD_AVG_PERIOD;
2574 } while (n > LOAD_AVG_PERIOD);
2576 contrib = decay_load(contrib, n);
2577 return contrib + runnable_avg_yN_sum[n];
2580 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2581 #error "load tracking assumes 2^10 as unit"
2584 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2587 * We can represent the historical contribution to runnable average as the
2588 * coefficients of a geometric series. To do this we sub-divide our runnable
2589 * history into segments of approximately 1ms (1024us); label the segment that
2590 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2592 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2594 * (now) (~1ms ago) (~2ms ago)
2596 * Let u_i denote the fraction of p_i that the entity was runnable.
2598 * We then designate the fractions u_i as our co-efficients, yielding the
2599 * following representation of historical load:
2600 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2602 * We choose y based on the with of a reasonably scheduling period, fixing:
2605 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2606 * approximately half as much as the contribution to load within the last ms
2609 * When a period "rolls over" and we have new u_0`, multiplying the previous
2610 * sum again by y is sufficient to update:
2611 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2612 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2614 static __always_inline int
2615 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2616 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2618 u64 delta, scaled_delta, periods;
2620 unsigned int delta_w, scaled_delta_w, decayed = 0;
2621 unsigned long scale_freq, scale_cpu;
2623 delta = now - sa->last_update_time;
2625 * This should only happen when time goes backwards, which it
2626 * unfortunately does during sched clock init when we swap over to TSC.
2628 if ((s64)delta < 0) {
2629 sa->last_update_time = now;
2634 * Use 1024ns as the unit of measurement since it's a reasonable
2635 * approximation of 1us and fast to compute.
2640 sa->last_update_time = now;
2642 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2643 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2645 /* delta_w is the amount already accumulated against our next period */
2646 delta_w = sa->period_contrib;
2647 if (delta + delta_w >= 1024) {
2650 /* how much left for next period will start over, we don't know yet */
2651 sa->period_contrib = 0;
2654 * Now that we know we're crossing a period boundary, figure
2655 * out how much from delta we need to complete the current
2656 * period and accrue it.
2658 delta_w = 1024 - delta_w;
2659 scaled_delta_w = cap_scale(delta_w, scale_freq);
2661 sa->load_sum += weight * scaled_delta_w;
2663 cfs_rq->runnable_load_sum +=
2664 weight * scaled_delta_w;
2668 sa->util_sum += scaled_delta_w * scale_cpu;
2672 /* Figure out how many additional periods this update spans */
2673 periods = delta / 1024;
2676 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2678 cfs_rq->runnable_load_sum =
2679 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2681 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2683 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2684 contrib = __compute_runnable_contrib(periods);
2685 contrib = cap_scale(contrib, scale_freq);
2687 sa->load_sum += weight * contrib;
2689 cfs_rq->runnable_load_sum += weight * contrib;
2692 sa->util_sum += contrib * scale_cpu;
2695 /* Remainder of delta accrued against u_0` */
2696 scaled_delta = cap_scale(delta, scale_freq);
2698 sa->load_sum += weight * scaled_delta;
2700 cfs_rq->runnable_load_sum += weight * scaled_delta;
2703 sa->util_sum += scaled_delta * scale_cpu;
2705 sa->period_contrib += delta;
2708 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2710 cfs_rq->runnable_load_avg =
2711 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2713 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2719 #ifdef CONFIG_FAIR_GROUP_SCHED
2721 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2722 * and effective_load (which is not done because it is too costly).
2724 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2726 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2729 * No need to update load_avg for root_task_group as it is not used.
2731 if (cfs_rq->tg == &root_task_group)
2734 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2735 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2736 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2741 * Called within set_task_rq() right before setting a task's cpu. The
2742 * caller only guarantees p->pi_lock is held; no other assumptions,
2743 * including the state of rq->lock, should be made.
2745 void set_task_rq_fair(struct sched_entity *se,
2746 struct cfs_rq *prev, struct cfs_rq *next)
2748 if (!sched_feat(ATTACH_AGE_LOAD))
2752 * We are supposed to update the task to "current" time, then its up to
2753 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2754 * getting what current time is, so simply throw away the out-of-date
2755 * time. This will result in the wakee task is less decayed, but giving
2756 * the wakee more load sounds not bad.
2758 if (se->avg.last_update_time && prev) {
2759 u64 p_last_update_time;
2760 u64 n_last_update_time;
2762 #ifndef CONFIG_64BIT
2763 u64 p_last_update_time_copy;
2764 u64 n_last_update_time_copy;
2767 p_last_update_time_copy = prev->load_last_update_time_copy;
2768 n_last_update_time_copy = next->load_last_update_time_copy;
2772 p_last_update_time = prev->avg.last_update_time;
2773 n_last_update_time = next->avg.last_update_time;
2775 } while (p_last_update_time != p_last_update_time_copy ||
2776 n_last_update_time != n_last_update_time_copy);
2778 p_last_update_time = prev->avg.last_update_time;
2779 n_last_update_time = next->avg.last_update_time;
2781 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2782 &se->avg, 0, 0, NULL);
2783 se->avg.last_update_time = n_last_update_time;
2786 #else /* CONFIG_FAIR_GROUP_SCHED */
2787 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2788 #endif /* CONFIG_FAIR_GROUP_SCHED */
2790 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2792 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2793 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2795 struct sched_avg *sa = &cfs_rq->avg;
2796 int decayed, removed = 0;
2798 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2799 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2800 sa->load_avg = max_t(long, sa->load_avg - r, 0);
2801 sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
2805 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2806 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2807 sa->util_avg = max_t(long, sa->util_avg - r, 0);
2808 sa->util_sum = max_t(s32, sa->util_sum - r * LOAD_AVG_MAX, 0);
2811 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2812 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2814 #ifndef CONFIG_64BIT
2816 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2819 return decayed || removed;
2822 /* Update task and its cfs_rq load average */
2823 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2825 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2826 u64 now = cfs_rq_clock_task(cfs_rq);
2827 int cpu = cpu_of(rq_of(cfs_rq));
2830 * Track task load average for carrying it to new CPU after migrated, and
2831 * track group sched_entity load average for task_h_load calc in migration
2833 __update_load_avg(now, cpu, &se->avg,
2834 se->on_rq * scale_load_down(se->load.weight),
2835 cfs_rq->curr == se, NULL);
2837 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2838 update_tg_load_avg(cfs_rq, 0);
2841 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2843 if (!sched_feat(ATTACH_AGE_LOAD))
2847 * If we got migrated (either between CPUs or between cgroups) we'll
2848 * have aged the average right before clearing @last_update_time.
2850 if (se->avg.last_update_time) {
2851 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2852 &se->avg, 0, 0, NULL);
2855 * XXX: we could have just aged the entire load away if we've been
2856 * absent from the fair class for too long.
2861 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2862 cfs_rq->avg.load_avg += se->avg.load_avg;
2863 cfs_rq->avg.load_sum += se->avg.load_sum;
2864 cfs_rq->avg.util_avg += se->avg.util_avg;
2865 cfs_rq->avg.util_sum += se->avg.util_sum;
2868 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2870 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2871 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2872 cfs_rq->curr == se, NULL);
2874 cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
2875 cfs_rq->avg.load_sum = max_t(s64, cfs_rq->avg.load_sum - se->avg.load_sum, 0);
2876 cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
2877 cfs_rq->avg.util_sum = max_t(s32, cfs_rq->avg.util_sum - se->avg.util_sum, 0);
2880 /* Add the load generated by se into cfs_rq's load average */
2882 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2884 struct sched_avg *sa = &se->avg;
2885 u64 now = cfs_rq_clock_task(cfs_rq);
2886 int migrated, decayed;
2888 migrated = !sa->last_update_time;
2890 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2891 se->on_rq * scale_load_down(se->load.weight),
2892 cfs_rq->curr == se, NULL);
2895 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2897 cfs_rq->runnable_load_avg += sa->load_avg;
2898 cfs_rq->runnable_load_sum += sa->load_sum;
2901 attach_entity_load_avg(cfs_rq, se);
2903 if (decayed || migrated)
2904 update_tg_load_avg(cfs_rq, 0);
2907 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2909 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2911 update_load_avg(se, 1);
2913 cfs_rq->runnable_load_avg =
2914 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2915 cfs_rq->runnable_load_sum =
2916 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2919 #ifndef CONFIG_64BIT
2920 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2922 u64 last_update_time_copy;
2923 u64 last_update_time;
2926 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2928 last_update_time = cfs_rq->avg.last_update_time;
2929 } while (last_update_time != last_update_time_copy);
2931 return last_update_time;
2934 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2936 return cfs_rq->avg.last_update_time;
2941 * Task first catches up with cfs_rq, and then subtract
2942 * itself from the cfs_rq (task must be off the queue now).
2944 void remove_entity_load_avg(struct sched_entity *se)
2946 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2947 u64 last_update_time;
2950 * Newly created task or never used group entity should not be removed
2951 * from its (source) cfs_rq
2953 if (se->avg.last_update_time == 0)
2956 last_update_time = cfs_rq_last_update_time(cfs_rq);
2958 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2959 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2960 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2963 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2965 return cfs_rq->runnable_load_avg;
2968 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2970 return cfs_rq->avg.load_avg;
2973 static int idle_balance(struct rq *this_rq);
2975 #else /* CONFIG_SMP */
2977 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2979 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2981 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2982 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2985 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2987 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2989 static inline int idle_balance(struct rq *rq)
2994 #endif /* CONFIG_SMP */
2996 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2998 #ifdef CONFIG_SCHEDSTATS
2999 struct task_struct *tsk = NULL;
3001 if (entity_is_task(se))
3004 if (se->statistics.sleep_start) {
3005 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3010 if (unlikely(delta > se->statistics.sleep_max))
3011 se->statistics.sleep_max = delta;
3013 se->statistics.sleep_start = 0;
3014 se->statistics.sum_sleep_runtime += delta;
3017 account_scheduler_latency(tsk, delta >> 10, 1);
3018 trace_sched_stat_sleep(tsk, delta);
3021 if (se->statistics.block_start) {
3022 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3027 if (unlikely(delta > se->statistics.block_max))
3028 se->statistics.block_max = delta;
3030 se->statistics.block_start = 0;
3031 se->statistics.sum_sleep_runtime += delta;
3034 if (tsk->in_iowait) {
3035 se->statistics.iowait_sum += delta;
3036 se->statistics.iowait_count++;
3037 trace_sched_stat_iowait(tsk, delta);
3040 trace_sched_stat_blocked(tsk, delta);
3043 * Blocking time is in units of nanosecs, so shift by
3044 * 20 to get a milliseconds-range estimation of the
3045 * amount of time that the task spent sleeping:
3047 if (unlikely(prof_on == SLEEP_PROFILING)) {
3048 profile_hits(SLEEP_PROFILING,
3049 (void *)get_wchan(tsk),
3052 account_scheduler_latency(tsk, delta >> 10, 0);
3058 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3060 #ifdef CONFIG_SCHED_DEBUG
3061 s64 d = se->vruntime - cfs_rq->min_vruntime;
3066 if (d > 3*sysctl_sched_latency)
3067 schedstat_inc(cfs_rq, nr_spread_over);
3072 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3074 u64 vruntime = cfs_rq->min_vruntime;
3077 * The 'current' period is already promised to the current tasks,
3078 * however the extra weight of the new task will slow them down a
3079 * little, place the new task so that it fits in the slot that
3080 * stays open at the end.
3082 if (initial && sched_feat(START_DEBIT))
3083 vruntime += sched_vslice(cfs_rq, se);
3085 /* sleeps up to a single latency don't count. */
3087 unsigned long thresh = sysctl_sched_latency;
3090 * Halve their sleep time's effect, to allow
3091 * for a gentler effect of sleepers:
3093 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3099 /* ensure we never gain time by being placed backwards. */
3100 se->vruntime = max_vruntime(se->vruntime, vruntime);
3103 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3106 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3109 * Update the normalized vruntime before updating min_vruntime
3110 * through calling update_curr().
3112 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3113 se->vruntime += cfs_rq->min_vruntime;
3116 * Update run-time statistics of the 'current'.
3118 update_curr(cfs_rq);
3119 enqueue_entity_load_avg(cfs_rq, se);
3120 account_entity_enqueue(cfs_rq, se);
3121 update_cfs_shares(cfs_rq);
3123 if (flags & ENQUEUE_WAKEUP) {
3124 place_entity(cfs_rq, se, 0);
3125 enqueue_sleeper(cfs_rq, se);
3128 update_stats_enqueue(cfs_rq, se);
3129 check_spread(cfs_rq, se);
3130 if (se != cfs_rq->curr)
3131 __enqueue_entity(cfs_rq, se);
3134 if (cfs_rq->nr_running == 1) {
3135 list_add_leaf_cfs_rq(cfs_rq);
3136 check_enqueue_throttle(cfs_rq);
3140 static void __clear_buddies_last(struct sched_entity *se)
3142 for_each_sched_entity(se) {
3143 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3144 if (cfs_rq->last != se)
3147 cfs_rq->last = NULL;
3151 static void __clear_buddies_next(struct sched_entity *se)
3153 for_each_sched_entity(se) {
3154 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3155 if (cfs_rq->next != se)
3158 cfs_rq->next = NULL;
3162 static void __clear_buddies_skip(struct sched_entity *se)
3164 for_each_sched_entity(se) {
3165 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3166 if (cfs_rq->skip != se)
3169 cfs_rq->skip = NULL;
3173 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3175 if (cfs_rq->last == se)
3176 __clear_buddies_last(se);
3178 if (cfs_rq->next == se)
3179 __clear_buddies_next(se);
3181 if (cfs_rq->skip == se)
3182 __clear_buddies_skip(se);
3185 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3188 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3191 * Update run-time statistics of the 'current'.
3193 update_curr(cfs_rq);
3194 dequeue_entity_load_avg(cfs_rq, se);
3196 update_stats_dequeue(cfs_rq, se);
3197 if (flags & DEQUEUE_SLEEP) {
3198 #ifdef CONFIG_SCHEDSTATS
3199 if (entity_is_task(se)) {
3200 struct task_struct *tsk = task_of(se);
3202 if (tsk->state & TASK_INTERRUPTIBLE)
3203 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3204 if (tsk->state & TASK_UNINTERRUPTIBLE)
3205 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3210 clear_buddies(cfs_rq, se);
3212 if (se != cfs_rq->curr)
3213 __dequeue_entity(cfs_rq, se);
3215 account_entity_dequeue(cfs_rq, se);
3218 * Normalize the entity after updating the min_vruntime because the
3219 * update can refer to the ->curr item and we need to reflect this
3220 * movement in our normalized position.
3222 if (!(flags & DEQUEUE_SLEEP))
3223 se->vruntime -= cfs_rq->min_vruntime;
3225 /* return excess runtime on last dequeue */
3226 return_cfs_rq_runtime(cfs_rq);
3228 update_min_vruntime(cfs_rq);
3229 update_cfs_shares(cfs_rq);
3233 * Preempt the current task with a newly woken task if needed:
3236 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3238 unsigned long ideal_runtime, delta_exec;
3239 struct sched_entity *se;
3242 ideal_runtime = sched_slice(cfs_rq, curr);
3243 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3244 if (delta_exec > ideal_runtime) {
3245 resched_curr(rq_of(cfs_rq));
3247 * The current task ran long enough, ensure it doesn't get
3248 * re-elected due to buddy favours.
3250 clear_buddies(cfs_rq, curr);
3255 * Ensure that a task that missed wakeup preemption by a
3256 * narrow margin doesn't have to wait for a full slice.
3257 * This also mitigates buddy induced latencies under load.
3259 if (delta_exec < sysctl_sched_min_granularity)
3262 se = __pick_first_entity(cfs_rq);
3263 delta = curr->vruntime - se->vruntime;
3268 if (delta > ideal_runtime)
3269 resched_curr(rq_of(cfs_rq));
3273 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3275 /* 'current' is not kept within the tree. */
3278 * Any task has to be enqueued before it get to execute on
3279 * a CPU. So account for the time it spent waiting on the
3282 update_stats_wait_end(cfs_rq, se);
3283 __dequeue_entity(cfs_rq, se);
3284 update_load_avg(se, 1);
3287 update_stats_curr_start(cfs_rq, se);
3289 #ifdef CONFIG_SCHEDSTATS
3291 * Track our maximum slice length, if the CPU's load is at
3292 * least twice that of our own weight (i.e. dont track it
3293 * when there are only lesser-weight tasks around):
3295 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3296 se->statistics.slice_max = max(se->statistics.slice_max,
3297 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3300 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3304 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3307 * Pick the next process, keeping these things in mind, in this order:
3308 * 1) keep things fair between processes/task groups
3309 * 2) pick the "next" process, since someone really wants that to run
3310 * 3) pick the "last" process, for cache locality
3311 * 4) do not run the "skip" process, if something else is available
3313 static struct sched_entity *
3314 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3316 struct sched_entity *left = __pick_first_entity(cfs_rq);
3317 struct sched_entity *se;
3320 * If curr is set we have to see if its left of the leftmost entity
3321 * still in the tree, provided there was anything in the tree at all.
3323 if (!left || (curr && entity_before(curr, left)))
3326 se = left; /* ideally we run the leftmost entity */
3329 * Avoid running the skip buddy, if running something else can
3330 * be done without getting too unfair.
3332 if (cfs_rq->skip == se) {
3333 struct sched_entity *second;
3336 second = __pick_first_entity(cfs_rq);
3338 second = __pick_next_entity(se);
3339 if (!second || (curr && entity_before(curr, second)))
3343 if (second && wakeup_preempt_entity(second, left) < 1)
3348 * Prefer last buddy, try to return the CPU to a preempted task.
3350 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3354 * Someone really wants this to run. If it's not unfair, run it.
3356 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3359 clear_buddies(cfs_rq, se);
3364 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3366 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3369 * If still on the runqueue then deactivate_task()
3370 * was not called and update_curr() has to be done:
3373 update_curr(cfs_rq);
3375 /* throttle cfs_rqs exceeding runtime */
3376 check_cfs_rq_runtime(cfs_rq);
3378 check_spread(cfs_rq, prev);
3380 update_stats_wait_start(cfs_rq, prev);
3381 /* Put 'current' back into the tree. */
3382 __enqueue_entity(cfs_rq, prev);
3383 /* in !on_rq case, update occurred at dequeue */
3384 update_load_avg(prev, 0);
3386 cfs_rq->curr = NULL;
3390 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3393 * Update run-time statistics of the 'current'.
3395 update_curr(cfs_rq);
3398 * Ensure that runnable average is periodically updated.
3400 update_load_avg(curr, 1);
3401 update_cfs_shares(cfs_rq);
3403 #ifdef CONFIG_SCHED_HRTICK
3405 * queued ticks are scheduled to match the slice, so don't bother
3406 * validating it and just reschedule.
3409 resched_curr(rq_of(cfs_rq));
3413 * don't let the period tick interfere with the hrtick preemption
3415 if (!sched_feat(DOUBLE_TICK) &&
3416 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3420 if (cfs_rq->nr_running > 1)
3421 check_preempt_tick(cfs_rq, curr);
3425 /**************************************************
3426 * CFS bandwidth control machinery
3429 #ifdef CONFIG_CFS_BANDWIDTH
3431 #ifdef HAVE_JUMP_LABEL
3432 static struct static_key __cfs_bandwidth_used;
3434 static inline bool cfs_bandwidth_used(void)
3436 return static_key_false(&__cfs_bandwidth_used);
3439 void cfs_bandwidth_usage_inc(void)
3441 static_key_slow_inc(&__cfs_bandwidth_used);
3444 void cfs_bandwidth_usage_dec(void)
3446 static_key_slow_dec(&__cfs_bandwidth_used);
3448 #else /* HAVE_JUMP_LABEL */
3449 static bool cfs_bandwidth_used(void)
3454 void cfs_bandwidth_usage_inc(void) {}
3455 void cfs_bandwidth_usage_dec(void) {}
3456 #endif /* HAVE_JUMP_LABEL */
3459 * default period for cfs group bandwidth.
3460 * default: 0.1s, units: nanoseconds
3462 static inline u64 default_cfs_period(void)
3464 return 100000000ULL;
3467 static inline u64 sched_cfs_bandwidth_slice(void)
3469 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3473 * Replenish runtime according to assigned quota and update expiration time.
3474 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3475 * additional synchronization around rq->lock.
3477 * requires cfs_b->lock
3479 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3483 if (cfs_b->quota == RUNTIME_INF)
3486 now = sched_clock_cpu(smp_processor_id());
3487 cfs_b->runtime = cfs_b->quota;
3488 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3491 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3493 return &tg->cfs_bandwidth;
3496 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3497 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3499 if (unlikely(cfs_rq->throttle_count))
3500 return cfs_rq->throttled_clock_task;
3502 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3505 /* returns 0 on failure to allocate runtime */
3506 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3508 struct task_group *tg = cfs_rq->tg;
3509 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3510 u64 amount = 0, min_amount, expires;
3512 /* note: this is a positive sum as runtime_remaining <= 0 */
3513 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3515 raw_spin_lock(&cfs_b->lock);
3516 if (cfs_b->quota == RUNTIME_INF)
3517 amount = min_amount;
3519 start_cfs_bandwidth(cfs_b);
3521 if (cfs_b->runtime > 0) {
3522 amount = min(cfs_b->runtime, min_amount);
3523 cfs_b->runtime -= amount;
3527 expires = cfs_b->runtime_expires;
3528 raw_spin_unlock(&cfs_b->lock);
3530 cfs_rq->runtime_remaining += amount;
3532 * we may have advanced our local expiration to account for allowed
3533 * spread between our sched_clock and the one on which runtime was
3536 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3537 cfs_rq->runtime_expires = expires;
3539 return cfs_rq->runtime_remaining > 0;
3543 * Note: This depends on the synchronization provided by sched_clock and the
3544 * fact that rq->clock snapshots this value.
3546 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3548 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3550 /* if the deadline is ahead of our clock, nothing to do */
3551 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3554 if (cfs_rq->runtime_remaining < 0)
3558 * If the local deadline has passed we have to consider the
3559 * possibility that our sched_clock is 'fast' and the global deadline
3560 * has not truly expired.
3562 * Fortunately we can check determine whether this the case by checking
3563 * whether the global deadline has advanced. It is valid to compare
3564 * cfs_b->runtime_expires without any locks since we only care about
3565 * exact equality, so a partial write will still work.
3568 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3569 /* extend local deadline, drift is bounded above by 2 ticks */
3570 cfs_rq->runtime_expires += TICK_NSEC;
3572 /* global deadline is ahead, expiration has passed */
3573 cfs_rq->runtime_remaining = 0;
3577 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3579 /* dock delta_exec before expiring quota (as it could span periods) */
3580 cfs_rq->runtime_remaining -= delta_exec;
3581 expire_cfs_rq_runtime(cfs_rq);
3583 if (likely(cfs_rq->runtime_remaining > 0))
3587 * if we're unable to extend our runtime we resched so that the active
3588 * hierarchy can be throttled
3590 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3591 resched_curr(rq_of(cfs_rq));
3594 static __always_inline
3595 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3597 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3600 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3603 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3605 return cfs_bandwidth_used() && cfs_rq->throttled;
3608 /* check whether cfs_rq, or any parent, is throttled */
3609 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3611 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3615 * Ensure that neither of the group entities corresponding to src_cpu or
3616 * dest_cpu are members of a throttled hierarchy when performing group
3617 * load-balance operations.
3619 static inline int throttled_lb_pair(struct task_group *tg,
3620 int src_cpu, int dest_cpu)
3622 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3624 src_cfs_rq = tg->cfs_rq[src_cpu];
3625 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3627 return throttled_hierarchy(src_cfs_rq) ||
3628 throttled_hierarchy(dest_cfs_rq);
3631 /* updated child weight may affect parent so we have to do this bottom up */
3632 static int tg_unthrottle_up(struct task_group *tg, void *data)
3634 struct rq *rq = data;
3635 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3637 cfs_rq->throttle_count--;
3639 if (!cfs_rq->throttle_count) {
3640 /* adjust cfs_rq_clock_task() */
3641 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3642 cfs_rq->throttled_clock_task;
3649 static int tg_throttle_down(struct task_group *tg, void *data)
3651 struct rq *rq = data;
3652 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3654 /* group is entering throttled state, stop time */
3655 if (!cfs_rq->throttle_count)
3656 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3657 cfs_rq->throttle_count++;
3662 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3664 struct rq *rq = rq_of(cfs_rq);
3665 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3666 struct sched_entity *se;
3667 long task_delta, dequeue = 1;
3670 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3672 /* freeze hierarchy runnable averages while throttled */
3674 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3677 task_delta = cfs_rq->h_nr_running;
3678 for_each_sched_entity(se) {
3679 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3680 /* throttled entity or throttle-on-deactivate */
3685 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3686 qcfs_rq->h_nr_running -= task_delta;
3688 if (qcfs_rq->load.weight)
3693 sub_nr_running(rq, task_delta);
3695 cfs_rq->throttled = 1;
3696 cfs_rq->throttled_clock = rq_clock(rq);
3697 raw_spin_lock(&cfs_b->lock);
3698 empty = list_empty(&cfs_b->throttled_cfs_rq);
3701 * Add to the _head_ of the list, so that an already-started
3702 * distribute_cfs_runtime will not see us
3704 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3707 * If we're the first throttled task, make sure the bandwidth
3711 start_cfs_bandwidth(cfs_b);
3713 raw_spin_unlock(&cfs_b->lock);
3716 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3718 struct rq *rq = rq_of(cfs_rq);
3719 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3720 struct sched_entity *se;
3724 se = cfs_rq->tg->se[cpu_of(rq)];
3726 cfs_rq->throttled = 0;
3728 update_rq_clock(rq);
3730 raw_spin_lock(&cfs_b->lock);
3731 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3732 list_del_rcu(&cfs_rq->throttled_list);
3733 raw_spin_unlock(&cfs_b->lock);
3735 /* update hierarchical throttle state */
3736 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3738 if (!cfs_rq->load.weight)
3741 task_delta = cfs_rq->h_nr_running;
3742 for_each_sched_entity(se) {
3746 cfs_rq = cfs_rq_of(se);
3748 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3749 cfs_rq->h_nr_running += task_delta;
3751 if (cfs_rq_throttled(cfs_rq))
3756 add_nr_running(rq, task_delta);
3758 /* determine whether we need to wake up potentially idle cpu */
3759 if (rq->curr == rq->idle && rq->cfs.nr_running)
3763 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3764 u64 remaining, u64 expires)
3766 struct cfs_rq *cfs_rq;
3768 u64 starting_runtime = remaining;
3771 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3773 struct rq *rq = rq_of(cfs_rq);
3775 raw_spin_lock(&rq->lock);
3776 if (!cfs_rq_throttled(cfs_rq))
3779 runtime = -cfs_rq->runtime_remaining + 1;
3780 if (runtime > remaining)
3781 runtime = remaining;
3782 remaining -= runtime;
3784 cfs_rq->runtime_remaining += runtime;
3785 cfs_rq->runtime_expires = expires;
3787 /* we check whether we're throttled above */
3788 if (cfs_rq->runtime_remaining > 0)
3789 unthrottle_cfs_rq(cfs_rq);
3792 raw_spin_unlock(&rq->lock);
3799 return starting_runtime - remaining;
3803 * Responsible for refilling a task_group's bandwidth and unthrottling its
3804 * cfs_rqs as appropriate. If there has been no activity within the last
3805 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3806 * used to track this state.
3808 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3810 u64 runtime, runtime_expires;
3813 /* no need to continue the timer with no bandwidth constraint */
3814 if (cfs_b->quota == RUNTIME_INF)
3815 goto out_deactivate;
3817 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3818 cfs_b->nr_periods += overrun;
3821 * idle depends on !throttled (for the case of a large deficit), and if
3822 * we're going inactive then everything else can be deferred
3824 if (cfs_b->idle && !throttled)
3825 goto out_deactivate;
3827 __refill_cfs_bandwidth_runtime(cfs_b);
3830 /* mark as potentially idle for the upcoming period */
3835 /* account preceding periods in which throttling occurred */
3836 cfs_b->nr_throttled += overrun;
3838 runtime_expires = cfs_b->runtime_expires;
3841 * This check is repeated as we are holding onto the new bandwidth while
3842 * we unthrottle. This can potentially race with an unthrottled group
3843 * trying to acquire new bandwidth from the global pool. This can result
3844 * in us over-using our runtime if it is all used during this loop, but
3845 * only by limited amounts in that extreme case.
3847 while (throttled && cfs_b->runtime > 0) {
3848 runtime = cfs_b->runtime;
3849 raw_spin_unlock(&cfs_b->lock);
3850 /* we can't nest cfs_b->lock while distributing bandwidth */
3851 runtime = distribute_cfs_runtime(cfs_b, runtime,
3853 raw_spin_lock(&cfs_b->lock);
3855 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3857 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3861 * While we are ensured activity in the period following an
3862 * unthrottle, this also covers the case in which the new bandwidth is
3863 * insufficient to cover the existing bandwidth deficit. (Forcing the
3864 * timer to remain active while there are any throttled entities.)
3874 /* a cfs_rq won't donate quota below this amount */
3875 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3876 /* minimum remaining period time to redistribute slack quota */
3877 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3878 /* how long we wait to gather additional slack before distributing */
3879 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3882 * Are we near the end of the current quota period?
3884 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3885 * hrtimer base being cleared by hrtimer_start. In the case of
3886 * migrate_hrtimers, base is never cleared, so we are fine.
3888 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3890 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3893 /* if the call-back is running a quota refresh is already occurring */
3894 if (hrtimer_callback_running(refresh_timer))
3897 /* is a quota refresh about to occur? */
3898 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3899 if (remaining < min_expire)
3905 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3907 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3909 /* if there's a quota refresh soon don't bother with slack */
3910 if (runtime_refresh_within(cfs_b, min_left))
3913 hrtimer_start(&cfs_b->slack_timer,
3914 ns_to_ktime(cfs_bandwidth_slack_period),
3918 /* we know any runtime found here is valid as update_curr() precedes return */
3919 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3921 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3922 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3924 if (slack_runtime <= 0)
3927 raw_spin_lock(&cfs_b->lock);
3928 if (cfs_b->quota != RUNTIME_INF &&
3929 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3930 cfs_b->runtime += slack_runtime;
3932 /* we are under rq->lock, defer unthrottling using a timer */
3933 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3934 !list_empty(&cfs_b->throttled_cfs_rq))
3935 start_cfs_slack_bandwidth(cfs_b);
3937 raw_spin_unlock(&cfs_b->lock);
3939 /* even if it's not valid for return we don't want to try again */
3940 cfs_rq->runtime_remaining -= slack_runtime;
3943 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3945 if (!cfs_bandwidth_used())
3948 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3951 __return_cfs_rq_runtime(cfs_rq);
3955 * This is done with a timer (instead of inline with bandwidth return) since
3956 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3958 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3960 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3963 /* confirm we're still not at a refresh boundary */
3964 raw_spin_lock(&cfs_b->lock);
3965 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3966 raw_spin_unlock(&cfs_b->lock);
3970 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3971 runtime = cfs_b->runtime;
3973 expires = cfs_b->runtime_expires;
3974 raw_spin_unlock(&cfs_b->lock);
3979 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3981 raw_spin_lock(&cfs_b->lock);
3982 if (expires == cfs_b->runtime_expires)
3983 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3984 raw_spin_unlock(&cfs_b->lock);
3988 * When a group wakes up we want to make sure that its quota is not already
3989 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3990 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3992 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3994 if (!cfs_bandwidth_used())
3997 /* an active group must be handled by the update_curr()->put() path */
3998 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4001 /* ensure the group is not already throttled */
4002 if (cfs_rq_throttled(cfs_rq))
4005 /* update runtime allocation */
4006 account_cfs_rq_runtime(cfs_rq, 0);
4007 if (cfs_rq->runtime_remaining <= 0)
4008 throttle_cfs_rq(cfs_rq);
4011 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4012 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4014 if (!cfs_bandwidth_used())
4017 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4021 * it's possible for a throttled entity to be forced into a running
4022 * state (e.g. set_curr_task), in this case we're finished.
4024 if (cfs_rq_throttled(cfs_rq))
4027 throttle_cfs_rq(cfs_rq);
4031 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4033 struct cfs_bandwidth *cfs_b =
4034 container_of(timer, struct cfs_bandwidth, slack_timer);
4036 do_sched_cfs_slack_timer(cfs_b);
4038 return HRTIMER_NORESTART;
4041 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4043 struct cfs_bandwidth *cfs_b =
4044 container_of(timer, struct cfs_bandwidth, period_timer);
4048 raw_spin_lock(&cfs_b->lock);
4050 overrun = hrtimer_forward_now(timer, cfs_b->period);
4054 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4057 cfs_b->period_active = 0;
4058 raw_spin_unlock(&cfs_b->lock);
4060 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4063 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4065 raw_spin_lock_init(&cfs_b->lock);
4067 cfs_b->quota = RUNTIME_INF;
4068 cfs_b->period = ns_to_ktime(default_cfs_period());
4070 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4071 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4072 cfs_b->period_timer.function = sched_cfs_period_timer;
4073 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4074 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4077 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4079 cfs_rq->runtime_enabled = 0;
4080 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4083 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4085 lockdep_assert_held(&cfs_b->lock);
4087 if (!cfs_b->period_active) {
4088 cfs_b->period_active = 1;
4089 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4090 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4094 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4096 /* init_cfs_bandwidth() was not called */
4097 if (!cfs_b->throttled_cfs_rq.next)
4100 hrtimer_cancel(&cfs_b->period_timer);
4101 hrtimer_cancel(&cfs_b->slack_timer);
4104 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4106 struct cfs_rq *cfs_rq;
4108 for_each_leaf_cfs_rq(rq, cfs_rq) {
4109 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4111 raw_spin_lock(&cfs_b->lock);
4112 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4113 raw_spin_unlock(&cfs_b->lock);
4117 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4119 struct cfs_rq *cfs_rq;
4121 for_each_leaf_cfs_rq(rq, cfs_rq) {
4122 if (!cfs_rq->runtime_enabled)
4126 * clock_task is not advancing so we just need to make sure
4127 * there's some valid quota amount
4129 cfs_rq->runtime_remaining = 1;
4131 * Offline rq is schedulable till cpu is completely disabled
4132 * in take_cpu_down(), so we prevent new cfs throttling here.
4134 cfs_rq->runtime_enabled = 0;
4136 if (cfs_rq_throttled(cfs_rq))
4137 unthrottle_cfs_rq(cfs_rq);
4141 #else /* CONFIG_CFS_BANDWIDTH */
4142 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4144 return rq_clock_task(rq_of(cfs_rq));
4147 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4148 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4149 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4150 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4152 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4157 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4162 static inline int throttled_lb_pair(struct task_group *tg,
4163 int src_cpu, int dest_cpu)
4168 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4170 #ifdef CONFIG_FAIR_GROUP_SCHED
4171 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4174 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4178 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4179 static inline void update_runtime_enabled(struct rq *rq) {}
4180 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4182 #endif /* CONFIG_CFS_BANDWIDTH */
4184 /**************************************************
4185 * CFS operations on tasks:
4188 #ifdef CONFIG_SCHED_HRTICK
4189 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4191 struct sched_entity *se = &p->se;
4192 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4194 WARN_ON(task_rq(p) != rq);
4196 if (cfs_rq->nr_running > 1) {
4197 u64 slice = sched_slice(cfs_rq, se);
4198 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4199 s64 delta = slice - ran;
4206 hrtick_start(rq, delta);
4211 * called from enqueue/dequeue and updates the hrtick when the
4212 * current task is from our class and nr_running is low enough
4215 static void hrtick_update(struct rq *rq)
4217 struct task_struct *curr = rq->curr;
4219 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4222 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4223 hrtick_start_fair(rq, curr);
4225 #else /* !CONFIG_SCHED_HRTICK */
4227 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4231 static inline void hrtick_update(struct rq *rq)
4237 * The enqueue_task method is called before nr_running is
4238 * increased. Here we update the fair scheduling stats and
4239 * then put the task into the rbtree:
4242 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4244 struct cfs_rq *cfs_rq;
4245 struct sched_entity *se = &p->se;
4247 for_each_sched_entity(se) {
4250 cfs_rq = cfs_rq_of(se);
4251 enqueue_entity(cfs_rq, se, flags);
4254 * end evaluation on encountering a throttled cfs_rq
4256 * note: in the case of encountering a throttled cfs_rq we will
4257 * post the final h_nr_running increment below.
4259 if (cfs_rq_throttled(cfs_rq))
4261 cfs_rq->h_nr_running++;
4263 flags = ENQUEUE_WAKEUP;
4266 for_each_sched_entity(se) {
4267 cfs_rq = cfs_rq_of(se);
4268 cfs_rq->h_nr_running++;
4270 if (cfs_rq_throttled(cfs_rq))
4273 update_load_avg(se, 1);
4274 update_cfs_shares(cfs_rq);
4278 add_nr_running(rq, 1);
4283 static void set_next_buddy(struct sched_entity *se);
4286 * The dequeue_task method is called before nr_running is
4287 * decreased. We remove the task from the rbtree and
4288 * update the fair scheduling stats:
4290 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4292 struct cfs_rq *cfs_rq;
4293 struct sched_entity *se = &p->se;
4294 int task_sleep = flags & DEQUEUE_SLEEP;
4296 for_each_sched_entity(se) {
4297 cfs_rq = cfs_rq_of(se);
4298 dequeue_entity(cfs_rq, se, flags);
4301 * end evaluation on encountering a throttled cfs_rq
4303 * note: in the case of encountering a throttled cfs_rq we will
4304 * post the final h_nr_running decrement below.
4306 if (cfs_rq_throttled(cfs_rq))
4308 cfs_rq->h_nr_running--;
4310 /* Don't dequeue parent if it has other entities besides us */
4311 if (cfs_rq->load.weight) {
4313 * Bias pick_next to pick a task from this cfs_rq, as
4314 * p is sleeping when it is within its sched_slice.
4316 if (task_sleep && parent_entity(se))
4317 set_next_buddy(parent_entity(se));
4319 /* avoid re-evaluating load for this entity */
4320 se = parent_entity(se);
4323 flags |= DEQUEUE_SLEEP;
4326 for_each_sched_entity(se) {
4327 cfs_rq = cfs_rq_of(se);
4328 cfs_rq->h_nr_running--;
4330 if (cfs_rq_throttled(cfs_rq))
4333 update_load_avg(se, 1);
4334 update_cfs_shares(cfs_rq);
4338 sub_nr_running(rq, 1);
4346 * per rq 'load' arrray crap; XXX kill this.
4350 * The exact cpuload calculated at every tick would be:
4352 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4354 * If a cpu misses updates for n ticks (as it was idle) and update gets
4355 * called on the n+1-th tick when cpu may be busy, then we have:
4357 * load_n = (1 - 1/2^i)^n * load_0
4358 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4360 * decay_load_missed() below does efficient calculation of
4362 * load' = (1 - 1/2^i)^n * load
4364 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4365 * This allows us to precompute the above in said factors, thereby allowing the
4366 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4367 * fixed_power_int())
4369 * The calculation is approximated on a 128 point scale.
4371 #define DEGRADE_SHIFT 7
4373 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4374 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4375 { 0, 0, 0, 0, 0, 0, 0, 0 },
4376 { 64, 32, 8, 0, 0, 0, 0, 0 },
4377 { 96, 72, 40, 12, 1, 0, 0, 0 },
4378 { 112, 98, 75, 43, 15, 1, 0, 0 },
4379 { 120, 112, 98, 76, 45, 16, 2, 0 }
4383 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4384 * would be when CPU is idle and so we just decay the old load without
4385 * adding any new load.
4387 static unsigned long
4388 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4392 if (!missed_updates)
4395 if (missed_updates >= degrade_zero_ticks[idx])
4399 return load >> missed_updates;
4401 while (missed_updates) {
4402 if (missed_updates % 2)
4403 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4405 missed_updates >>= 1;
4412 * __update_cpu_load - update the rq->cpu_load[] statistics
4413 * @this_rq: The rq to update statistics for
4414 * @this_load: The current load
4415 * @pending_updates: The number of missed updates
4416 * @active: !0 for NOHZ_FULL
4418 * Update rq->cpu_load[] statistics. This function is usually called every
4419 * scheduler tick (TICK_NSEC).
4421 * This function computes a decaying average:
4423 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4425 * Because of NOHZ it might not get called on every tick which gives need for
4426 * the @pending_updates argument.
4428 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4429 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4430 * = A * (A * load[i]_n-2 + B) + B
4431 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4432 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4433 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4434 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4435 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4437 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4438 * any change in load would have resulted in the tick being turned back on.
4440 * For regular NOHZ, this reduces to:
4442 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4444 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4445 * term. See the @active paramter.
4447 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4448 unsigned long pending_updates, int active)
4450 unsigned long tickless_load = active ? this_rq->cpu_load[0] : 0;
4453 this_rq->nr_load_updates++;
4455 /* Update our load: */
4456 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4457 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4458 unsigned long old_load, new_load;
4460 /* scale is effectively 1 << i now, and >> i divides by scale */
4462 old_load = this_rq->cpu_load[i] - tickless_load;
4463 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4464 old_load += tickless_load;
4465 new_load = this_load;
4467 * Round up the averaging division if load is increasing. This
4468 * prevents us from getting stuck on 9 if the load is 10, for
4471 if (new_load > old_load)
4472 new_load += scale - 1;
4474 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4477 sched_avg_update(this_rq);
4480 /* Used instead of source_load when we know the type == 0 */
4481 static unsigned long weighted_cpuload(const int cpu)
4483 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4486 #ifdef CONFIG_NO_HZ_COMMON
4488 * There is no sane way to deal with nohz on smp when using jiffies because the
4489 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4490 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4492 * Therefore we cannot use the delta approach from the regular tick since that
4493 * would seriously skew the load calculation. However we'll make do for those
4494 * updates happening while idle (nohz_idle_balance) or coming out of idle
4495 * (tick_nohz_idle_exit).
4497 * This means we might still be one tick off for nohz periods.
4501 * Called from nohz_idle_balance() to update the load ratings before doing the
4504 static void update_idle_cpu_load(struct rq *this_rq)
4506 unsigned long curr_jiffies = READ_ONCE(jiffies);
4507 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4508 unsigned long pending_updates;
4511 * bail if there's load or we're actually up-to-date.
4513 if (load || curr_jiffies == this_rq->last_load_update_tick)
4516 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4517 this_rq->last_load_update_tick = curr_jiffies;
4519 __update_cpu_load(this_rq, load, pending_updates, 0);
4523 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4525 void update_cpu_load_nohz(int active)
4527 struct rq *this_rq = this_rq();
4528 unsigned long curr_jiffies = READ_ONCE(jiffies);
4529 unsigned long load = active ? weighted_cpuload(cpu_of(this_rq)) : 0;
4530 unsigned long pending_updates;
4532 if (curr_jiffies == this_rq->last_load_update_tick)
4535 raw_spin_lock(&this_rq->lock);
4536 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4537 if (pending_updates) {
4538 this_rq->last_load_update_tick = curr_jiffies;
4540 * In the regular NOHZ case, we were idle, this means load 0.
4541 * In the NOHZ_FULL case, we were non-idle, we should consider
4542 * its weighted load.
4544 __update_cpu_load(this_rq, load, pending_updates, active);
4546 raw_spin_unlock(&this_rq->lock);
4548 #endif /* CONFIG_NO_HZ */
4551 * Called from scheduler_tick()
4553 void update_cpu_load_active(struct rq *this_rq)
4555 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4557 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4559 this_rq->last_load_update_tick = jiffies;
4560 __update_cpu_load(this_rq, load, 1, 1);
4564 * Return a low guess at the load of a migration-source cpu weighted
4565 * according to the scheduling class and "nice" value.
4567 * We want to under-estimate the load of migration sources, to
4568 * balance conservatively.
4570 static unsigned long source_load(int cpu, int type)
4572 struct rq *rq = cpu_rq(cpu);
4573 unsigned long total = weighted_cpuload(cpu);
4575 if (type == 0 || !sched_feat(LB_BIAS))
4578 return min(rq->cpu_load[type-1], total);
4582 * Return a high guess at the load of a migration-target cpu weighted
4583 * according to the scheduling class and "nice" value.
4585 static unsigned long target_load(int cpu, int type)
4587 struct rq *rq = cpu_rq(cpu);
4588 unsigned long total = weighted_cpuload(cpu);
4590 if (type == 0 || !sched_feat(LB_BIAS))
4593 return max(rq->cpu_load[type-1], total);
4596 static unsigned long capacity_of(int cpu)
4598 return cpu_rq(cpu)->cpu_capacity;
4601 static unsigned long capacity_orig_of(int cpu)
4603 return cpu_rq(cpu)->cpu_capacity_orig;
4606 static unsigned long cpu_avg_load_per_task(int cpu)
4608 struct rq *rq = cpu_rq(cpu);
4609 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4610 unsigned long load_avg = weighted_cpuload(cpu);
4613 return load_avg / nr_running;
4618 static void record_wakee(struct task_struct *p)
4621 * Rough decay (wiping) for cost saving, don't worry
4622 * about the boundary, really active task won't care
4625 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4626 current->wakee_flips >>= 1;
4627 current->wakee_flip_decay_ts = jiffies;
4630 if (current->last_wakee != p) {
4631 current->last_wakee = p;
4632 current->wakee_flips++;
4636 static void task_waking_fair(struct task_struct *p)
4638 struct sched_entity *se = &p->se;
4639 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4642 #ifndef CONFIG_64BIT
4643 u64 min_vruntime_copy;
4646 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4648 min_vruntime = cfs_rq->min_vruntime;
4649 } while (min_vruntime != min_vruntime_copy);
4651 min_vruntime = cfs_rq->min_vruntime;
4654 se->vruntime -= min_vruntime;
4658 #ifdef CONFIG_FAIR_GROUP_SCHED
4660 * effective_load() calculates the load change as seen from the root_task_group
4662 * Adding load to a group doesn't make a group heavier, but can cause movement
4663 * of group shares between cpus. Assuming the shares were perfectly aligned one
4664 * can calculate the shift in shares.
4666 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4667 * on this @cpu and results in a total addition (subtraction) of @wg to the
4668 * total group weight.
4670 * Given a runqueue weight distribution (rw_i) we can compute a shares
4671 * distribution (s_i) using:
4673 * s_i = rw_i / \Sum rw_j (1)
4675 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4676 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4677 * shares distribution (s_i):
4679 * rw_i = { 2, 4, 1, 0 }
4680 * s_i = { 2/7, 4/7, 1/7, 0 }
4682 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4683 * task used to run on and the CPU the waker is running on), we need to
4684 * compute the effect of waking a task on either CPU and, in case of a sync
4685 * wakeup, compute the effect of the current task going to sleep.
4687 * So for a change of @wl to the local @cpu with an overall group weight change
4688 * of @wl we can compute the new shares distribution (s'_i) using:
4690 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4692 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4693 * differences in waking a task to CPU 0. The additional task changes the
4694 * weight and shares distributions like:
4696 * rw'_i = { 3, 4, 1, 0 }
4697 * s'_i = { 3/8, 4/8, 1/8, 0 }
4699 * We can then compute the difference in effective weight by using:
4701 * dw_i = S * (s'_i - s_i) (3)
4703 * Where 'S' is the group weight as seen by its parent.
4705 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4706 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4707 * 4/7) times the weight of the group.
4709 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4711 struct sched_entity *se = tg->se[cpu];
4713 if (!tg->parent) /* the trivial, non-cgroup case */
4716 for_each_sched_entity(se) {
4722 * W = @wg + \Sum rw_j
4724 W = wg + calc_tg_weight(tg, se->my_q);
4729 w = cfs_rq_load_avg(se->my_q) + wl;
4732 * wl = S * s'_i; see (2)
4735 wl = (w * (long)tg->shares) / W;
4740 * Per the above, wl is the new se->load.weight value; since
4741 * those are clipped to [MIN_SHARES, ...) do so now. See
4742 * calc_cfs_shares().
4744 if (wl < MIN_SHARES)
4748 * wl = dw_i = S * (s'_i - s_i); see (3)
4750 wl -= se->avg.load_avg;
4753 * Recursively apply this logic to all parent groups to compute
4754 * the final effective load change on the root group. Since
4755 * only the @tg group gets extra weight, all parent groups can
4756 * only redistribute existing shares. @wl is the shift in shares
4757 * resulting from this level per the above.
4766 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4774 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4775 * A waker of many should wake a different task than the one last awakened
4776 * at a frequency roughly N times higher than one of its wakees. In order
4777 * to determine whether we should let the load spread vs consolodating to
4778 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4779 * partner, and a factor of lls_size higher frequency in the other. With
4780 * both conditions met, we can be relatively sure that the relationship is
4781 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4782 * being client/server, worker/dispatcher, interrupt source or whatever is
4783 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4785 static int wake_wide(struct task_struct *p)
4787 unsigned int master = current->wakee_flips;
4788 unsigned int slave = p->wakee_flips;
4789 int factor = this_cpu_read(sd_llc_size);
4792 swap(master, slave);
4793 if (slave < factor || master < slave * factor)
4798 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4800 s64 this_load, load;
4801 s64 this_eff_load, prev_eff_load;
4802 int idx, this_cpu, prev_cpu;
4803 struct task_group *tg;
4804 unsigned long weight;
4808 this_cpu = smp_processor_id();
4809 prev_cpu = task_cpu(p);
4810 load = source_load(prev_cpu, idx);
4811 this_load = target_load(this_cpu, idx);
4814 * If sync wakeup then subtract the (maximum possible)
4815 * effect of the currently running task from the load
4816 * of the current CPU:
4819 tg = task_group(current);
4820 weight = current->se.avg.load_avg;
4822 this_load += effective_load(tg, this_cpu, -weight, -weight);
4823 load += effective_load(tg, prev_cpu, 0, -weight);
4827 weight = p->se.avg.load_avg;
4830 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4831 * due to the sync cause above having dropped this_load to 0, we'll
4832 * always have an imbalance, but there's really nothing you can do
4833 * about that, so that's good too.
4835 * Otherwise check if either cpus are near enough in load to allow this
4836 * task to be woken on this_cpu.
4838 this_eff_load = 100;
4839 this_eff_load *= capacity_of(prev_cpu);
4841 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4842 prev_eff_load *= capacity_of(this_cpu);
4844 if (this_load > 0) {
4845 this_eff_load *= this_load +
4846 effective_load(tg, this_cpu, weight, weight);
4848 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4851 balanced = this_eff_load <= prev_eff_load;
4853 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4858 schedstat_inc(sd, ttwu_move_affine);
4859 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4865 * find_idlest_group finds and returns the least busy CPU group within the
4868 static struct sched_group *
4869 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4870 int this_cpu, int sd_flag)
4872 struct sched_group *idlest = NULL, *group = sd->groups;
4873 unsigned long min_load = ULONG_MAX, this_load = 0;
4874 int load_idx = sd->forkexec_idx;
4875 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4877 if (sd_flag & SD_BALANCE_WAKE)
4878 load_idx = sd->wake_idx;
4881 unsigned long load, avg_load;
4885 /* Skip over this group if it has no CPUs allowed */
4886 if (!cpumask_intersects(sched_group_cpus(group),
4887 tsk_cpus_allowed(p)))
4890 local_group = cpumask_test_cpu(this_cpu,
4891 sched_group_cpus(group));
4893 /* Tally up the load of all CPUs in the group */
4896 for_each_cpu(i, sched_group_cpus(group)) {
4897 /* Bias balancing toward cpus of our domain */
4899 load = source_load(i, load_idx);
4901 load = target_load(i, load_idx);
4906 /* Adjust by relative CPU capacity of the group */
4907 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4910 this_load = avg_load;
4911 } else if (avg_load < min_load) {
4912 min_load = avg_load;
4915 } while (group = group->next, group != sd->groups);
4917 if (!idlest || 100*this_load < imbalance*min_load)
4923 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4926 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4928 unsigned long load, min_load = ULONG_MAX;
4929 unsigned int min_exit_latency = UINT_MAX;
4930 u64 latest_idle_timestamp = 0;
4931 int least_loaded_cpu = this_cpu;
4932 int shallowest_idle_cpu = -1;
4935 /* Traverse only the allowed CPUs */
4936 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4938 struct rq *rq = cpu_rq(i);
4939 struct cpuidle_state *idle = idle_get_state(rq);
4940 if (idle && idle->exit_latency < min_exit_latency) {
4942 * We give priority to a CPU whose idle state
4943 * has the smallest exit latency irrespective
4944 * of any idle timestamp.
4946 min_exit_latency = idle->exit_latency;
4947 latest_idle_timestamp = rq->idle_stamp;
4948 shallowest_idle_cpu = i;
4949 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4950 rq->idle_stamp > latest_idle_timestamp) {
4952 * If equal or no active idle state, then
4953 * the most recently idled CPU might have
4956 latest_idle_timestamp = rq->idle_stamp;
4957 shallowest_idle_cpu = i;
4959 } else if (shallowest_idle_cpu == -1) {
4960 load = weighted_cpuload(i);
4961 if (load < min_load || (load == min_load && i == this_cpu)) {
4963 least_loaded_cpu = i;
4968 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4972 * Try and locate an idle CPU in the sched_domain.
4974 static int select_idle_sibling(struct task_struct *p, int target)
4976 struct sched_domain *sd;
4977 struct sched_group *sg;
4978 int i = task_cpu(p);
4980 if (idle_cpu(target))
4984 * If the prevous cpu is cache affine and idle, don't be stupid.
4986 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4990 * Otherwise, iterate the domains and find an elegible idle cpu.
4992 sd = rcu_dereference(per_cpu(sd_llc, target));
4993 for_each_lower_domain(sd) {
4996 if (!cpumask_intersects(sched_group_cpus(sg),
4997 tsk_cpus_allowed(p)))
5000 for_each_cpu(i, sched_group_cpus(sg)) {
5001 if (i == target || !idle_cpu(i))
5005 target = cpumask_first_and(sched_group_cpus(sg),
5006 tsk_cpus_allowed(p));
5010 } while (sg != sd->groups);
5017 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5018 * tasks. The unit of the return value must be the one of capacity so we can
5019 * compare the utilization with the capacity of the CPU that is available for
5020 * CFS task (ie cpu_capacity).
5022 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5023 * recent utilization of currently non-runnable tasks on a CPU. It represents
5024 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5025 * capacity_orig is the cpu_capacity available at the highest frequency
5026 * (arch_scale_freq_capacity()).
5027 * The utilization of a CPU converges towards a sum equal to or less than the
5028 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5029 * the running time on this CPU scaled by capacity_curr.
5031 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5032 * higher than capacity_orig because of unfortunate rounding in
5033 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5034 * the average stabilizes with the new running time. We need to check that the
5035 * utilization stays within the range of [0..capacity_orig] and cap it if
5036 * necessary. Without utilization capping, a group could be seen as overloaded
5037 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5038 * available capacity. We allow utilization to overshoot capacity_curr (but not
5039 * capacity_orig) as it useful for predicting the capacity required after task
5040 * migrations (scheduler-driven DVFS).
5042 static int cpu_util(int cpu)
5044 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5045 unsigned long capacity = capacity_orig_of(cpu);
5047 return (util >= capacity) ? capacity : util;
5051 * select_task_rq_fair: Select target runqueue for the waking task in domains
5052 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5053 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5055 * Balances load by selecting the idlest cpu in the idlest group, or under
5056 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5058 * Returns the target cpu number.
5060 * preempt must be disabled.
5063 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5065 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5066 int cpu = smp_processor_id();
5067 int new_cpu = prev_cpu;
5068 int want_affine = 0;
5069 int sync = wake_flags & WF_SYNC;
5071 if (sd_flag & SD_BALANCE_WAKE)
5072 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5075 for_each_domain(cpu, tmp) {
5076 if (!(tmp->flags & SD_LOAD_BALANCE))
5080 * If both cpu and prev_cpu are part of this domain,
5081 * cpu is a valid SD_WAKE_AFFINE target.
5083 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5084 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5089 if (tmp->flags & sd_flag)
5091 else if (!want_affine)
5096 sd = NULL; /* Prefer wake_affine over balance flags */
5097 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5102 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5103 new_cpu = select_idle_sibling(p, new_cpu);
5106 struct sched_group *group;
5109 if (!(sd->flags & sd_flag)) {
5114 group = find_idlest_group(sd, p, cpu, sd_flag);
5120 new_cpu = find_idlest_cpu(group, p, cpu);
5121 if (new_cpu == -1 || new_cpu == cpu) {
5122 /* Now try balancing at a lower domain level of cpu */
5127 /* Now try balancing at a lower domain level of new_cpu */
5129 weight = sd->span_weight;
5131 for_each_domain(cpu, tmp) {
5132 if (weight <= tmp->span_weight)
5134 if (tmp->flags & sd_flag)
5137 /* while loop will break here if sd == NULL */
5145 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5146 * cfs_rq_of(p) references at time of call are still valid and identify the
5147 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5149 static void migrate_task_rq_fair(struct task_struct *p)
5152 * We are supposed to update the task to "current" time, then its up to date
5153 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5154 * what current time is, so simply throw away the out-of-date time. This
5155 * will result in the wakee task is less decayed, but giving the wakee more
5156 * load sounds not bad.
5158 remove_entity_load_avg(&p->se);
5160 /* Tell new CPU we are migrated */
5161 p->se.avg.last_update_time = 0;
5163 /* We have migrated, no longer consider this task hot */
5164 p->se.exec_start = 0;
5167 static void task_dead_fair(struct task_struct *p)
5169 remove_entity_load_avg(&p->se);
5171 #endif /* CONFIG_SMP */
5173 static unsigned long
5174 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5176 unsigned long gran = sysctl_sched_wakeup_granularity;
5179 * Since its curr running now, convert the gran from real-time
5180 * to virtual-time in his units.
5182 * By using 'se' instead of 'curr' we penalize light tasks, so
5183 * they get preempted easier. That is, if 'se' < 'curr' then
5184 * the resulting gran will be larger, therefore penalizing the
5185 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5186 * be smaller, again penalizing the lighter task.
5188 * This is especially important for buddies when the leftmost
5189 * task is higher priority than the buddy.
5191 return calc_delta_fair(gran, se);
5195 * Should 'se' preempt 'curr'.
5209 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5211 s64 gran, vdiff = curr->vruntime - se->vruntime;
5216 gran = wakeup_gran(curr, se);
5223 static void set_last_buddy(struct sched_entity *se)
5225 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5228 for_each_sched_entity(se)
5229 cfs_rq_of(se)->last = se;
5232 static void set_next_buddy(struct sched_entity *se)
5234 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5237 for_each_sched_entity(se)
5238 cfs_rq_of(se)->next = se;
5241 static void set_skip_buddy(struct sched_entity *se)
5243 for_each_sched_entity(se)
5244 cfs_rq_of(se)->skip = se;
5248 * Preempt the current task with a newly woken task if needed:
5250 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5252 struct task_struct *curr = rq->curr;
5253 struct sched_entity *se = &curr->se, *pse = &p->se;
5254 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5255 int scale = cfs_rq->nr_running >= sched_nr_latency;
5256 int next_buddy_marked = 0;
5258 if (unlikely(se == pse))
5262 * This is possible from callers such as attach_tasks(), in which we
5263 * unconditionally check_prempt_curr() after an enqueue (which may have
5264 * lead to a throttle). This both saves work and prevents false
5265 * next-buddy nomination below.
5267 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5270 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5271 set_next_buddy(pse);
5272 next_buddy_marked = 1;
5276 * We can come here with TIF_NEED_RESCHED already set from new task
5279 * Note: this also catches the edge-case of curr being in a throttled
5280 * group (e.g. via set_curr_task), since update_curr() (in the
5281 * enqueue of curr) will have resulted in resched being set. This
5282 * prevents us from potentially nominating it as a false LAST_BUDDY
5285 if (test_tsk_need_resched(curr))
5288 /* Idle tasks are by definition preempted by non-idle tasks. */
5289 if (unlikely(curr->policy == SCHED_IDLE) &&
5290 likely(p->policy != SCHED_IDLE))
5294 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5295 * is driven by the tick):
5297 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5300 find_matching_se(&se, &pse);
5301 update_curr(cfs_rq_of(se));
5303 if (wakeup_preempt_entity(se, pse) == 1) {
5305 * Bias pick_next to pick the sched entity that is
5306 * triggering this preemption.
5308 if (!next_buddy_marked)
5309 set_next_buddy(pse);
5318 * Only set the backward buddy when the current task is still
5319 * on the rq. This can happen when a wakeup gets interleaved
5320 * with schedule on the ->pre_schedule() or idle_balance()
5321 * point, either of which can * drop the rq lock.
5323 * Also, during early boot the idle thread is in the fair class,
5324 * for obvious reasons its a bad idea to schedule back to it.
5326 if (unlikely(!se->on_rq || curr == rq->idle))
5329 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5333 static struct task_struct *
5334 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5336 struct cfs_rq *cfs_rq = &rq->cfs;
5337 struct sched_entity *se;
5338 struct task_struct *p;
5342 #ifdef CONFIG_FAIR_GROUP_SCHED
5343 if (!cfs_rq->nr_running)
5346 if (prev->sched_class != &fair_sched_class)
5350 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5351 * likely that a next task is from the same cgroup as the current.
5353 * Therefore attempt to avoid putting and setting the entire cgroup
5354 * hierarchy, only change the part that actually changes.
5358 struct sched_entity *curr = cfs_rq->curr;
5361 * Since we got here without doing put_prev_entity() we also
5362 * have to consider cfs_rq->curr. If it is still a runnable
5363 * entity, update_curr() will update its vruntime, otherwise
5364 * forget we've ever seen it.
5368 update_curr(cfs_rq);
5373 * This call to check_cfs_rq_runtime() will do the
5374 * throttle and dequeue its entity in the parent(s).
5375 * Therefore the 'simple' nr_running test will indeed
5378 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5382 se = pick_next_entity(cfs_rq, curr);
5383 cfs_rq = group_cfs_rq(se);
5389 * Since we haven't yet done put_prev_entity and if the selected task
5390 * is a different task than we started out with, try and touch the
5391 * least amount of cfs_rqs.
5394 struct sched_entity *pse = &prev->se;
5396 while (!(cfs_rq = is_same_group(se, pse))) {
5397 int se_depth = se->depth;
5398 int pse_depth = pse->depth;
5400 if (se_depth <= pse_depth) {
5401 put_prev_entity(cfs_rq_of(pse), pse);
5402 pse = parent_entity(pse);
5404 if (se_depth >= pse_depth) {
5405 set_next_entity(cfs_rq_of(se), se);
5406 se = parent_entity(se);
5410 put_prev_entity(cfs_rq, pse);
5411 set_next_entity(cfs_rq, se);
5414 if (hrtick_enabled(rq))
5415 hrtick_start_fair(rq, p);
5422 if (!cfs_rq->nr_running)
5425 put_prev_task(rq, prev);
5428 se = pick_next_entity(cfs_rq, NULL);
5429 set_next_entity(cfs_rq, se);
5430 cfs_rq = group_cfs_rq(se);
5435 if (hrtick_enabled(rq))
5436 hrtick_start_fair(rq, p);
5442 * This is OK, because current is on_cpu, which avoids it being picked
5443 * for load-balance and preemption/IRQs are still disabled avoiding
5444 * further scheduler activity on it and we're being very careful to
5445 * re-start the picking loop.
5447 lockdep_unpin_lock(&rq->lock);
5448 new_tasks = idle_balance(rq);
5449 lockdep_pin_lock(&rq->lock);
5451 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5452 * possible for any higher priority task to appear. In that case we
5453 * must re-start the pick_next_entity() loop.
5465 * Account for a descheduled task:
5467 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5469 struct sched_entity *se = &prev->se;
5470 struct cfs_rq *cfs_rq;
5472 for_each_sched_entity(se) {
5473 cfs_rq = cfs_rq_of(se);
5474 put_prev_entity(cfs_rq, se);
5479 * sched_yield() is very simple
5481 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5483 static void yield_task_fair(struct rq *rq)
5485 struct task_struct *curr = rq->curr;
5486 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5487 struct sched_entity *se = &curr->se;
5490 * Are we the only task in the tree?
5492 if (unlikely(rq->nr_running == 1))
5495 clear_buddies(cfs_rq, se);
5497 if (curr->policy != SCHED_BATCH) {
5498 update_rq_clock(rq);
5500 * Update run-time statistics of the 'current'.
5502 update_curr(cfs_rq);
5504 * Tell update_rq_clock() that we've just updated,
5505 * so we don't do microscopic update in schedule()
5506 * and double the fastpath cost.
5508 rq_clock_skip_update(rq, true);
5514 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5516 struct sched_entity *se = &p->se;
5518 /* throttled hierarchies are not runnable */
5519 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5522 /* Tell the scheduler that we'd really like pse to run next. */
5525 yield_task_fair(rq);
5531 /**************************************************
5532 * Fair scheduling class load-balancing methods.
5536 * The purpose of load-balancing is to achieve the same basic fairness the
5537 * per-cpu scheduler provides, namely provide a proportional amount of compute
5538 * time to each task. This is expressed in the following equation:
5540 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5542 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5543 * W_i,0 is defined as:
5545 * W_i,0 = \Sum_j w_i,j (2)
5547 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5548 * is derived from the nice value as per prio_to_weight[].
5550 * The weight average is an exponential decay average of the instantaneous
5553 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5555 * C_i is the compute capacity of cpu i, typically it is the
5556 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5557 * can also include other factors [XXX].
5559 * To achieve this balance we define a measure of imbalance which follows
5560 * directly from (1):
5562 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5564 * We them move tasks around to minimize the imbalance. In the continuous
5565 * function space it is obvious this converges, in the discrete case we get
5566 * a few fun cases generally called infeasible weight scenarios.
5569 * - infeasible weights;
5570 * - local vs global optima in the discrete case. ]
5575 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5576 * for all i,j solution, we create a tree of cpus that follows the hardware
5577 * topology where each level pairs two lower groups (or better). This results
5578 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5579 * tree to only the first of the previous level and we decrease the frequency
5580 * of load-balance at each level inv. proportional to the number of cpus in
5586 * \Sum { --- * --- * 2^i } = O(n) (5)
5588 * `- size of each group
5589 * | | `- number of cpus doing load-balance
5591 * `- sum over all levels
5593 * Coupled with a limit on how many tasks we can migrate every balance pass,
5594 * this makes (5) the runtime complexity of the balancer.
5596 * An important property here is that each CPU is still (indirectly) connected
5597 * to every other cpu in at most O(log n) steps:
5599 * The adjacency matrix of the resulting graph is given by:
5602 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5605 * And you'll find that:
5607 * A^(log_2 n)_i,j != 0 for all i,j (7)
5609 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5610 * The task movement gives a factor of O(m), giving a convergence complexity
5613 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5618 * In order to avoid CPUs going idle while there's still work to do, new idle
5619 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5620 * tree itself instead of relying on other CPUs to bring it work.
5622 * This adds some complexity to both (5) and (8) but it reduces the total idle
5630 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5633 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5638 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5640 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5642 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5645 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5646 * rewrite all of this once again.]
5649 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5651 enum fbq_type { regular, remote, all };
5653 #define LBF_ALL_PINNED 0x01
5654 #define LBF_NEED_BREAK 0x02
5655 #define LBF_DST_PINNED 0x04
5656 #define LBF_SOME_PINNED 0x08
5659 struct sched_domain *sd;
5667 struct cpumask *dst_grpmask;
5669 enum cpu_idle_type idle;
5671 /* The set of CPUs under consideration for load-balancing */
5672 struct cpumask *cpus;
5677 unsigned int loop_break;
5678 unsigned int loop_max;
5680 enum fbq_type fbq_type;
5681 struct list_head tasks;
5685 * Is this task likely cache-hot:
5687 static int task_hot(struct task_struct *p, struct lb_env *env)
5691 lockdep_assert_held(&env->src_rq->lock);
5693 if (p->sched_class != &fair_sched_class)
5696 if (unlikely(p->policy == SCHED_IDLE))
5700 * Buddy candidates are cache hot:
5702 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5703 (&p->se == cfs_rq_of(&p->se)->next ||
5704 &p->se == cfs_rq_of(&p->se)->last))
5707 if (sysctl_sched_migration_cost == -1)
5709 if (sysctl_sched_migration_cost == 0)
5712 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5714 return delta < (s64)sysctl_sched_migration_cost;
5717 #ifdef CONFIG_NUMA_BALANCING
5719 * Returns 1, if task migration degrades locality
5720 * Returns 0, if task migration improves locality i.e migration preferred.
5721 * Returns -1, if task migration is not affected by locality.
5723 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5725 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5726 unsigned long src_faults, dst_faults;
5727 int src_nid, dst_nid;
5729 if (!static_branch_likely(&sched_numa_balancing))
5732 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5735 src_nid = cpu_to_node(env->src_cpu);
5736 dst_nid = cpu_to_node(env->dst_cpu);
5738 if (src_nid == dst_nid)
5741 /* Migrating away from the preferred node is always bad. */
5742 if (src_nid == p->numa_preferred_nid) {
5743 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5749 /* Encourage migration to the preferred node. */
5750 if (dst_nid == p->numa_preferred_nid)
5754 src_faults = group_faults(p, src_nid);
5755 dst_faults = group_faults(p, dst_nid);
5757 src_faults = task_faults(p, src_nid);
5758 dst_faults = task_faults(p, dst_nid);
5761 return dst_faults < src_faults;
5765 static inline int migrate_degrades_locality(struct task_struct *p,
5773 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5776 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5780 lockdep_assert_held(&env->src_rq->lock);
5783 * We do not migrate tasks that are:
5784 * 1) throttled_lb_pair, or
5785 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5786 * 3) running (obviously), or
5787 * 4) are cache-hot on their current CPU.
5789 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5792 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5795 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5797 env->flags |= LBF_SOME_PINNED;
5800 * Remember if this task can be migrated to any other cpu in
5801 * our sched_group. We may want to revisit it if we couldn't
5802 * meet load balance goals by pulling other tasks on src_cpu.
5804 * Also avoid computing new_dst_cpu if we have already computed
5805 * one in current iteration.
5807 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5810 /* Prevent to re-select dst_cpu via env's cpus */
5811 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5812 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5813 env->flags |= LBF_DST_PINNED;
5814 env->new_dst_cpu = cpu;
5822 /* Record that we found atleast one task that could run on dst_cpu */
5823 env->flags &= ~LBF_ALL_PINNED;
5825 if (task_running(env->src_rq, p)) {
5826 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5831 * Aggressive migration if:
5832 * 1) destination numa is preferred
5833 * 2) task is cache cold, or
5834 * 3) too many balance attempts have failed.
5836 tsk_cache_hot = migrate_degrades_locality(p, env);
5837 if (tsk_cache_hot == -1)
5838 tsk_cache_hot = task_hot(p, env);
5840 if (tsk_cache_hot <= 0 ||
5841 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5842 if (tsk_cache_hot == 1) {
5843 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5844 schedstat_inc(p, se.statistics.nr_forced_migrations);
5849 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5854 * detach_task() -- detach the task for the migration specified in env
5856 static void detach_task(struct task_struct *p, struct lb_env *env)
5858 lockdep_assert_held(&env->src_rq->lock);
5860 p->on_rq = TASK_ON_RQ_MIGRATING;
5861 deactivate_task(env->src_rq, p, 0);
5862 set_task_cpu(p, env->dst_cpu);
5866 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5867 * part of active balancing operations within "domain".
5869 * Returns a task if successful and NULL otherwise.
5871 static struct task_struct *detach_one_task(struct lb_env *env)
5873 struct task_struct *p, *n;
5875 lockdep_assert_held(&env->src_rq->lock);
5877 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5878 if (!can_migrate_task(p, env))
5881 detach_task(p, env);
5884 * Right now, this is only the second place where
5885 * lb_gained[env->idle] is updated (other is detach_tasks)
5886 * so we can safely collect stats here rather than
5887 * inside detach_tasks().
5889 schedstat_inc(env->sd, lb_gained[env->idle]);
5895 static const unsigned int sched_nr_migrate_break = 32;
5898 * detach_tasks() -- tries to detach up to imbalance weighted load from
5899 * busiest_rq, as part of a balancing operation within domain "sd".
5901 * Returns number of detached tasks if successful and 0 otherwise.
5903 static int detach_tasks(struct lb_env *env)
5905 struct list_head *tasks = &env->src_rq->cfs_tasks;
5906 struct task_struct *p;
5910 lockdep_assert_held(&env->src_rq->lock);
5912 if (env->imbalance <= 0)
5915 while (!list_empty(tasks)) {
5917 * We don't want to steal all, otherwise we may be treated likewise,
5918 * which could at worst lead to a livelock crash.
5920 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5923 p = list_first_entry(tasks, struct task_struct, se.group_node);
5926 /* We've more or less seen every task there is, call it quits */
5927 if (env->loop > env->loop_max)
5930 /* take a breather every nr_migrate tasks */
5931 if (env->loop > env->loop_break) {
5932 env->loop_break += sched_nr_migrate_break;
5933 env->flags |= LBF_NEED_BREAK;
5937 if (!can_migrate_task(p, env))
5940 load = task_h_load(p);
5942 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5945 if ((load / 2) > env->imbalance)
5948 detach_task(p, env);
5949 list_add(&p->se.group_node, &env->tasks);
5952 env->imbalance -= load;
5954 #ifdef CONFIG_PREEMPT
5956 * NEWIDLE balancing is a source of latency, so preemptible
5957 * kernels will stop after the first task is detached to minimize
5958 * the critical section.
5960 if (env->idle == CPU_NEWLY_IDLE)
5965 * We only want to steal up to the prescribed amount of
5968 if (env->imbalance <= 0)
5973 list_move_tail(&p->se.group_node, tasks);
5977 * Right now, this is one of only two places we collect this stat
5978 * so we can safely collect detach_one_task() stats here rather
5979 * than inside detach_one_task().
5981 schedstat_add(env->sd, lb_gained[env->idle], detached);
5987 * attach_task() -- attach the task detached by detach_task() to its new rq.
5989 static void attach_task(struct rq *rq, struct task_struct *p)
5991 lockdep_assert_held(&rq->lock);
5993 BUG_ON(task_rq(p) != rq);
5994 activate_task(rq, p, 0);
5995 p->on_rq = TASK_ON_RQ_QUEUED;
5996 check_preempt_curr(rq, p, 0);
6000 * attach_one_task() -- attaches the task returned from detach_one_task() to
6003 static void attach_one_task(struct rq *rq, struct task_struct *p)
6005 raw_spin_lock(&rq->lock);
6007 raw_spin_unlock(&rq->lock);
6011 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6014 static void attach_tasks(struct lb_env *env)
6016 struct list_head *tasks = &env->tasks;
6017 struct task_struct *p;
6019 raw_spin_lock(&env->dst_rq->lock);
6021 while (!list_empty(tasks)) {
6022 p = list_first_entry(tasks, struct task_struct, se.group_node);
6023 list_del_init(&p->se.group_node);
6025 attach_task(env->dst_rq, p);
6028 raw_spin_unlock(&env->dst_rq->lock);
6031 #ifdef CONFIG_FAIR_GROUP_SCHED
6032 static void update_blocked_averages(int cpu)
6034 struct rq *rq = cpu_rq(cpu);
6035 struct cfs_rq *cfs_rq;
6036 unsigned long flags;
6038 raw_spin_lock_irqsave(&rq->lock, flags);
6039 update_rq_clock(rq);
6042 * Iterates the task_group tree in a bottom up fashion, see
6043 * list_add_leaf_cfs_rq() for details.
6045 for_each_leaf_cfs_rq(rq, cfs_rq) {
6046 /* throttled entities do not contribute to load */
6047 if (throttled_hierarchy(cfs_rq))
6050 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6051 update_tg_load_avg(cfs_rq, 0);
6053 raw_spin_unlock_irqrestore(&rq->lock, flags);
6057 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6058 * This needs to be done in a top-down fashion because the load of a child
6059 * group is a fraction of its parents load.
6061 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6063 struct rq *rq = rq_of(cfs_rq);
6064 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6065 unsigned long now = jiffies;
6068 if (cfs_rq->last_h_load_update == now)
6071 cfs_rq->h_load_next = NULL;
6072 for_each_sched_entity(se) {
6073 cfs_rq = cfs_rq_of(se);
6074 cfs_rq->h_load_next = se;
6075 if (cfs_rq->last_h_load_update == now)
6080 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6081 cfs_rq->last_h_load_update = now;
6084 while ((se = cfs_rq->h_load_next) != NULL) {
6085 load = cfs_rq->h_load;
6086 load = div64_ul(load * se->avg.load_avg,
6087 cfs_rq_load_avg(cfs_rq) + 1);
6088 cfs_rq = group_cfs_rq(se);
6089 cfs_rq->h_load = load;
6090 cfs_rq->last_h_load_update = now;
6094 static unsigned long task_h_load(struct task_struct *p)
6096 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6098 update_cfs_rq_h_load(cfs_rq);
6099 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6100 cfs_rq_load_avg(cfs_rq) + 1);
6103 static inline void update_blocked_averages(int cpu)
6105 struct rq *rq = cpu_rq(cpu);
6106 struct cfs_rq *cfs_rq = &rq->cfs;
6107 unsigned long flags;
6109 raw_spin_lock_irqsave(&rq->lock, flags);
6110 update_rq_clock(rq);
6111 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6112 raw_spin_unlock_irqrestore(&rq->lock, flags);
6115 static unsigned long task_h_load(struct task_struct *p)
6117 return p->se.avg.load_avg;
6121 /********** Helpers for find_busiest_group ************************/
6130 * sg_lb_stats - stats of a sched_group required for load_balancing
6132 struct sg_lb_stats {
6133 unsigned long avg_load; /*Avg load across the CPUs of the group */
6134 unsigned long group_load; /* Total load over the CPUs of the group */
6135 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6136 unsigned long load_per_task;
6137 unsigned long group_capacity;
6138 unsigned long group_util; /* Total utilization of the group */
6139 unsigned int sum_nr_running; /* Nr tasks running in the group */
6140 unsigned int idle_cpus;
6141 unsigned int group_weight;
6142 enum group_type group_type;
6143 int group_no_capacity;
6144 #ifdef CONFIG_NUMA_BALANCING
6145 unsigned int nr_numa_running;
6146 unsigned int nr_preferred_running;
6151 * sd_lb_stats - Structure to store the statistics of a sched_domain
6152 * during load balancing.
6154 struct sd_lb_stats {
6155 struct sched_group *busiest; /* Busiest group in this sd */
6156 struct sched_group *local; /* Local group in this sd */
6157 unsigned long total_load; /* Total load of all groups in sd */
6158 unsigned long total_capacity; /* Total capacity of all groups in sd */
6159 unsigned long avg_load; /* Average load across all groups in sd */
6161 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6162 struct sg_lb_stats local_stat; /* Statistics of the local group */
6165 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6168 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6169 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6170 * We must however clear busiest_stat::avg_load because
6171 * update_sd_pick_busiest() reads this before assignment.
6173 *sds = (struct sd_lb_stats){
6177 .total_capacity = 0UL,
6180 .sum_nr_running = 0,
6181 .group_type = group_other,
6187 * get_sd_load_idx - Obtain the load index for a given sched domain.
6188 * @sd: The sched_domain whose load_idx is to be obtained.
6189 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6191 * Return: The load index.
6193 static inline int get_sd_load_idx(struct sched_domain *sd,
6194 enum cpu_idle_type idle)
6200 load_idx = sd->busy_idx;
6203 case CPU_NEWLY_IDLE:
6204 load_idx = sd->newidle_idx;
6207 load_idx = sd->idle_idx;
6214 static unsigned long scale_rt_capacity(int cpu)
6216 struct rq *rq = cpu_rq(cpu);
6217 u64 total, used, age_stamp, avg;
6221 * Since we're reading these variables without serialization make sure
6222 * we read them once before doing sanity checks on them.
6224 age_stamp = READ_ONCE(rq->age_stamp);
6225 avg = READ_ONCE(rq->rt_avg);
6226 delta = __rq_clock_broken(rq) - age_stamp;
6228 if (unlikely(delta < 0))
6231 total = sched_avg_period() + delta;
6233 used = div_u64(avg, total);
6235 if (likely(used < SCHED_CAPACITY_SCALE))
6236 return SCHED_CAPACITY_SCALE - used;
6241 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6243 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6244 struct sched_group *sdg = sd->groups;
6246 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6248 capacity *= scale_rt_capacity(cpu);
6249 capacity >>= SCHED_CAPACITY_SHIFT;
6254 cpu_rq(cpu)->cpu_capacity = capacity;
6255 sdg->sgc->capacity = capacity;
6258 void update_group_capacity(struct sched_domain *sd, int cpu)
6260 struct sched_domain *child = sd->child;
6261 struct sched_group *group, *sdg = sd->groups;
6262 unsigned long capacity;
6263 unsigned long interval;
6265 interval = msecs_to_jiffies(sd->balance_interval);
6266 interval = clamp(interval, 1UL, max_load_balance_interval);
6267 sdg->sgc->next_update = jiffies + interval;
6270 update_cpu_capacity(sd, cpu);
6276 if (child->flags & SD_OVERLAP) {
6278 * SD_OVERLAP domains cannot assume that child groups
6279 * span the current group.
6282 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6283 struct sched_group_capacity *sgc;
6284 struct rq *rq = cpu_rq(cpu);
6287 * build_sched_domains() -> init_sched_groups_capacity()
6288 * gets here before we've attached the domains to the
6291 * Use capacity_of(), which is set irrespective of domains
6292 * in update_cpu_capacity().
6294 * This avoids capacity from being 0 and
6295 * causing divide-by-zero issues on boot.
6297 if (unlikely(!rq->sd)) {
6298 capacity += capacity_of(cpu);
6302 sgc = rq->sd->groups->sgc;
6303 capacity += sgc->capacity;
6307 * !SD_OVERLAP domains can assume that child groups
6308 * span the current group.
6311 group = child->groups;
6313 capacity += group->sgc->capacity;
6314 group = group->next;
6315 } while (group != child->groups);
6318 sdg->sgc->capacity = capacity;
6322 * Check whether the capacity of the rq has been noticeably reduced by side
6323 * activity. The imbalance_pct is used for the threshold.
6324 * Return true is the capacity is reduced
6327 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6329 return ((rq->cpu_capacity * sd->imbalance_pct) <
6330 (rq->cpu_capacity_orig * 100));
6334 * Group imbalance indicates (and tries to solve) the problem where balancing
6335 * groups is inadequate due to tsk_cpus_allowed() constraints.
6337 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6338 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6341 * { 0 1 2 3 } { 4 5 6 7 }
6344 * If we were to balance group-wise we'd place two tasks in the first group and
6345 * two tasks in the second group. Clearly this is undesired as it will overload
6346 * cpu 3 and leave one of the cpus in the second group unused.
6348 * The current solution to this issue is detecting the skew in the first group
6349 * by noticing the lower domain failed to reach balance and had difficulty
6350 * moving tasks due to affinity constraints.
6352 * When this is so detected; this group becomes a candidate for busiest; see
6353 * update_sd_pick_busiest(). And calculate_imbalance() and
6354 * find_busiest_group() avoid some of the usual balance conditions to allow it
6355 * to create an effective group imbalance.
6357 * This is a somewhat tricky proposition since the next run might not find the
6358 * group imbalance and decide the groups need to be balanced again. A most
6359 * subtle and fragile situation.
6362 static inline int sg_imbalanced(struct sched_group *group)
6364 return group->sgc->imbalance;
6368 * group_has_capacity returns true if the group has spare capacity that could
6369 * be used by some tasks.
6370 * We consider that a group has spare capacity if the * number of task is
6371 * smaller than the number of CPUs or if the utilization is lower than the
6372 * available capacity for CFS tasks.
6373 * For the latter, we use a threshold to stabilize the state, to take into
6374 * account the variance of the tasks' load and to return true if the available
6375 * capacity in meaningful for the load balancer.
6376 * As an example, an available capacity of 1% can appear but it doesn't make
6377 * any benefit for the load balance.
6380 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6382 if (sgs->sum_nr_running < sgs->group_weight)
6385 if ((sgs->group_capacity * 100) >
6386 (sgs->group_util * env->sd->imbalance_pct))
6393 * group_is_overloaded returns true if the group has more tasks than it can
6395 * group_is_overloaded is not equals to !group_has_capacity because a group
6396 * with the exact right number of tasks, has no more spare capacity but is not
6397 * overloaded so both group_has_capacity and group_is_overloaded return
6401 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6403 if (sgs->sum_nr_running <= sgs->group_weight)
6406 if ((sgs->group_capacity * 100) <
6407 (sgs->group_util * env->sd->imbalance_pct))
6414 group_type group_classify(struct sched_group *group,
6415 struct sg_lb_stats *sgs)
6417 if (sgs->group_no_capacity)
6418 return group_overloaded;
6420 if (sg_imbalanced(group))
6421 return group_imbalanced;
6427 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6428 * @env: The load balancing environment.
6429 * @group: sched_group whose statistics are to be updated.
6430 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6431 * @local_group: Does group contain this_cpu.
6432 * @sgs: variable to hold the statistics for this group.
6433 * @overload: Indicate more than one runnable task for any CPU.
6435 static inline void update_sg_lb_stats(struct lb_env *env,
6436 struct sched_group *group, int load_idx,
6437 int local_group, struct sg_lb_stats *sgs,
6443 memset(sgs, 0, sizeof(*sgs));
6445 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6446 struct rq *rq = cpu_rq(i);
6448 /* Bias balancing toward cpus of our domain */
6450 load = target_load(i, load_idx);
6452 load = source_load(i, load_idx);
6454 sgs->group_load += load;
6455 sgs->group_util += cpu_util(i);
6456 sgs->sum_nr_running += rq->cfs.h_nr_running;
6458 nr_running = rq->nr_running;
6462 #ifdef CONFIG_NUMA_BALANCING
6463 sgs->nr_numa_running += rq->nr_numa_running;
6464 sgs->nr_preferred_running += rq->nr_preferred_running;
6466 sgs->sum_weighted_load += weighted_cpuload(i);
6468 * No need to call idle_cpu() if nr_running is not 0
6470 if (!nr_running && idle_cpu(i))
6474 /* Adjust by relative CPU capacity of the group */
6475 sgs->group_capacity = group->sgc->capacity;
6476 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6478 if (sgs->sum_nr_running)
6479 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6481 sgs->group_weight = group->group_weight;
6483 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6484 sgs->group_type = group_classify(group, sgs);
6488 * update_sd_pick_busiest - return 1 on busiest group
6489 * @env: The load balancing environment.
6490 * @sds: sched_domain statistics
6491 * @sg: sched_group candidate to be checked for being the busiest
6492 * @sgs: sched_group statistics
6494 * Determine if @sg is a busier group than the previously selected
6497 * Return: %true if @sg is a busier group than the previously selected
6498 * busiest group. %false otherwise.
6500 static bool update_sd_pick_busiest(struct lb_env *env,
6501 struct sd_lb_stats *sds,
6502 struct sched_group *sg,
6503 struct sg_lb_stats *sgs)
6505 struct sg_lb_stats *busiest = &sds->busiest_stat;
6507 if (sgs->group_type > busiest->group_type)
6510 if (sgs->group_type < busiest->group_type)
6513 if (sgs->avg_load <= busiest->avg_load)
6516 /* This is the busiest node in its class. */
6517 if (!(env->sd->flags & SD_ASYM_PACKING))
6521 * ASYM_PACKING needs to move all the work to the lowest
6522 * numbered CPUs in the group, therefore mark all groups
6523 * higher than ourself as busy.
6525 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6529 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6536 #ifdef CONFIG_NUMA_BALANCING
6537 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6539 if (sgs->sum_nr_running > sgs->nr_numa_running)
6541 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6546 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6548 if (rq->nr_running > rq->nr_numa_running)
6550 if (rq->nr_running > rq->nr_preferred_running)
6555 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6560 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6564 #endif /* CONFIG_NUMA_BALANCING */
6567 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6568 * @env: The load balancing environment.
6569 * @sds: variable to hold the statistics for this sched_domain.
6571 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6573 struct sched_domain *child = env->sd->child;
6574 struct sched_group *sg = env->sd->groups;
6575 struct sg_lb_stats tmp_sgs;
6576 int load_idx, prefer_sibling = 0;
6577 bool overload = false;
6579 if (child && child->flags & SD_PREFER_SIBLING)
6582 load_idx = get_sd_load_idx(env->sd, env->idle);
6585 struct sg_lb_stats *sgs = &tmp_sgs;
6588 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6591 sgs = &sds->local_stat;
6593 if (env->idle != CPU_NEWLY_IDLE ||
6594 time_after_eq(jiffies, sg->sgc->next_update))
6595 update_group_capacity(env->sd, env->dst_cpu);
6598 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6605 * In case the child domain prefers tasks go to siblings
6606 * first, lower the sg capacity so that we'll try
6607 * and move all the excess tasks away. We lower the capacity
6608 * of a group only if the local group has the capacity to fit
6609 * these excess tasks. The extra check prevents the case where
6610 * you always pull from the heaviest group when it is already
6611 * under-utilized (possible with a large weight task outweighs
6612 * the tasks on the system).
6614 if (prefer_sibling && sds->local &&
6615 group_has_capacity(env, &sds->local_stat) &&
6616 (sgs->sum_nr_running > 1)) {
6617 sgs->group_no_capacity = 1;
6618 sgs->group_type = group_classify(sg, sgs);
6621 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6623 sds->busiest_stat = *sgs;
6627 /* Now, start updating sd_lb_stats */
6628 sds->total_load += sgs->group_load;
6629 sds->total_capacity += sgs->group_capacity;
6632 } while (sg != env->sd->groups);
6634 if (env->sd->flags & SD_NUMA)
6635 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6637 if (!env->sd->parent) {
6638 /* update overload indicator if we are at root domain */
6639 if (env->dst_rq->rd->overload != overload)
6640 env->dst_rq->rd->overload = overload;
6646 * check_asym_packing - Check to see if the group is packed into the
6649 * This is primarily intended to used at the sibling level. Some
6650 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6651 * case of POWER7, it can move to lower SMT modes only when higher
6652 * threads are idle. When in lower SMT modes, the threads will
6653 * perform better since they share less core resources. Hence when we
6654 * have idle threads, we want them to be the higher ones.
6656 * This packing function is run on idle threads. It checks to see if
6657 * the busiest CPU in this domain (core in the P7 case) has a higher
6658 * CPU number than the packing function is being run on. Here we are
6659 * assuming lower CPU number will be equivalent to lower a SMT thread
6662 * Return: 1 when packing is required and a task should be moved to
6663 * this CPU. The amount of the imbalance is returned in *imbalance.
6665 * @env: The load balancing environment.
6666 * @sds: Statistics of the sched_domain which is to be packed
6668 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6672 if (!(env->sd->flags & SD_ASYM_PACKING))
6678 busiest_cpu = group_first_cpu(sds->busiest);
6679 if (env->dst_cpu > busiest_cpu)
6682 env->imbalance = DIV_ROUND_CLOSEST(
6683 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6684 SCHED_CAPACITY_SCALE);
6690 * fix_small_imbalance - Calculate the minor imbalance that exists
6691 * amongst the groups of a sched_domain, during
6693 * @env: The load balancing environment.
6694 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6697 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6699 unsigned long tmp, capa_now = 0, capa_move = 0;
6700 unsigned int imbn = 2;
6701 unsigned long scaled_busy_load_per_task;
6702 struct sg_lb_stats *local, *busiest;
6704 local = &sds->local_stat;
6705 busiest = &sds->busiest_stat;
6707 if (!local->sum_nr_running)
6708 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6709 else if (busiest->load_per_task > local->load_per_task)
6712 scaled_busy_load_per_task =
6713 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6714 busiest->group_capacity;
6716 if (busiest->avg_load + scaled_busy_load_per_task >=
6717 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6718 env->imbalance = busiest->load_per_task;
6723 * OK, we don't have enough imbalance to justify moving tasks,
6724 * however we may be able to increase total CPU capacity used by
6728 capa_now += busiest->group_capacity *
6729 min(busiest->load_per_task, busiest->avg_load);
6730 capa_now += local->group_capacity *
6731 min(local->load_per_task, local->avg_load);
6732 capa_now /= SCHED_CAPACITY_SCALE;
6734 /* Amount of load we'd subtract */
6735 if (busiest->avg_load > scaled_busy_load_per_task) {
6736 capa_move += busiest->group_capacity *
6737 min(busiest->load_per_task,
6738 busiest->avg_load - scaled_busy_load_per_task);
6741 /* Amount of load we'd add */
6742 if (busiest->avg_load * busiest->group_capacity <
6743 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6744 tmp = (busiest->avg_load * busiest->group_capacity) /
6745 local->group_capacity;
6747 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6748 local->group_capacity;
6750 capa_move += local->group_capacity *
6751 min(local->load_per_task, local->avg_load + tmp);
6752 capa_move /= SCHED_CAPACITY_SCALE;
6754 /* Move if we gain throughput */
6755 if (capa_move > capa_now)
6756 env->imbalance = busiest->load_per_task;
6760 * calculate_imbalance - Calculate the amount of imbalance present within the
6761 * groups of a given sched_domain during load balance.
6762 * @env: load balance environment
6763 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6765 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6767 unsigned long max_pull, load_above_capacity = ~0UL;
6768 struct sg_lb_stats *local, *busiest;
6770 local = &sds->local_stat;
6771 busiest = &sds->busiest_stat;
6773 if (busiest->group_type == group_imbalanced) {
6775 * In the group_imb case we cannot rely on group-wide averages
6776 * to ensure cpu-load equilibrium, look at wider averages. XXX
6778 busiest->load_per_task =
6779 min(busiest->load_per_task, sds->avg_load);
6783 * In the presence of smp nice balancing, certain scenarios can have
6784 * max load less than avg load(as we skip the groups at or below
6785 * its cpu_capacity, while calculating max_load..)
6787 if (busiest->avg_load <= sds->avg_load ||
6788 local->avg_load >= sds->avg_load) {
6790 return fix_small_imbalance(env, sds);
6794 * If there aren't any idle cpus, avoid creating some.
6796 if (busiest->group_type == group_overloaded &&
6797 local->group_type == group_overloaded) {
6798 load_above_capacity = busiest->sum_nr_running *
6800 if (load_above_capacity > busiest->group_capacity)
6801 load_above_capacity -= busiest->group_capacity;
6803 load_above_capacity = ~0UL;
6807 * We're trying to get all the cpus to the average_load, so we don't
6808 * want to push ourselves above the average load, nor do we wish to
6809 * reduce the max loaded cpu below the average load. At the same time,
6810 * we also don't want to reduce the group load below the group capacity
6811 * (so that we can implement power-savings policies etc). Thus we look
6812 * for the minimum possible imbalance.
6814 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6816 /* How much load to actually move to equalise the imbalance */
6817 env->imbalance = min(
6818 max_pull * busiest->group_capacity,
6819 (sds->avg_load - local->avg_load) * local->group_capacity
6820 ) / SCHED_CAPACITY_SCALE;
6823 * if *imbalance is less than the average load per runnable task
6824 * there is no guarantee that any tasks will be moved so we'll have
6825 * a think about bumping its value to force at least one task to be
6828 if (env->imbalance < busiest->load_per_task)
6829 return fix_small_imbalance(env, sds);
6832 /******* find_busiest_group() helpers end here *********************/
6835 * find_busiest_group - Returns the busiest group within the sched_domain
6836 * if there is an imbalance. If there isn't an imbalance, and
6837 * the user has opted for power-savings, it returns a group whose
6838 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6839 * such a group exists.
6841 * Also calculates the amount of weighted load which should be moved
6842 * to restore balance.
6844 * @env: The load balancing environment.
6846 * Return: - The busiest group if imbalance exists.
6847 * - If no imbalance and user has opted for power-savings balance,
6848 * return the least loaded group whose CPUs can be
6849 * put to idle by rebalancing its tasks onto our group.
6851 static struct sched_group *find_busiest_group(struct lb_env *env)
6853 struct sg_lb_stats *local, *busiest;
6854 struct sd_lb_stats sds;
6856 init_sd_lb_stats(&sds);
6859 * Compute the various statistics relavent for load balancing at
6862 update_sd_lb_stats(env, &sds);
6863 local = &sds.local_stat;
6864 busiest = &sds.busiest_stat;
6866 /* ASYM feature bypasses nice load balance check */
6867 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6868 check_asym_packing(env, &sds))
6871 /* There is no busy sibling group to pull tasks from */
6872 if (!sds.busiest || busiest->sum_nr_running == 0)
6875 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6876 / sds.total_capacity;
6879 * If the busiest group is imbalanced the below checks don't
6880 * work because they assume all things are equal, which typically
6881 * isn't true due to cpus_allowed constraints and the like.
6883 if (busiest->group_type == group_imbalanced)
6886 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6887 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6888 busiest->group_no_capacity)
6892 * If the local group is busier than the selected busiest group
6893 * don't try and pull any tasks.
6895 if (local->avg_load >= busiest->avg_load)
6899 * Don't pull any tasks if this group is already above the domain
6902 if (local->avg_load >= sds.avg_load)
6905 if (env->idle == CPU_IDLE) {
6907 * This cpu is idle. If the busiest group is not overloaded
6908 * and there is no imbalance between this and busiest group
6909 * wrt idle cpus, it is balanced. The imbalance becomes
6910 * significant if the diff is greater than 1 otherwise we
6911 * might end up to just move the imbalance on another group
6913 if ((busiest->group_type != group_overloaded) &&
6914 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6918 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6919 * imbalance_pct to be conservative.
6921 if (100 * busiest->avg_load <=
6922 env->sd->imbalance_pct * local->avg_load)
6927 /* Looks like there is an imbalance. Compute it */
6928 calculate_imbalance(env, &sds);
6937 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6939 static struct rq *find_busiest_queue(struct lb_env *env,
6940 struct sched_group *group)
6942 struct rq *busiest = NULL, *rq;
6943 unsigned long busiest_load = 0, busiest_capacity = 1;
6946 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6947 unsigned long capacity, wl;
6951 rt = fbq_classify_rq(rq);
6954 * We classify groups/runqueues into three groups:
6955 * - regular: there are !numa tasks
6956 * - remote: there are numa tasks that run on the 'wrong' node
6957 * - all: there is no distinction
6959 * In order to avoid migrating ideally placed numa tasks,
6960 * ignore those when there's better options.
6962 * If we ignore the actual busiest queue to migrate another
6963 * task, the next balance pass can still reduce the busiest
6964 * queue by moving tasks around inside the node.
6966 * If we cannot move enough load due to this classification
6967 * the next pass will adjust the group classification and
6968 * allow migration of more tasks.
6970 * Both cases only affect the total convergence complexity.
6972 if (rt > env->fbq_type)
6975 capacity = capacity_of(i);
6977 wl = weighted_cpuload(i);
6980 * When comparing with imbalance, use weighted_cpuload()
6981 * which is not scaled with the cpu capacity.
6984 if (rq->nr_running == 1 && wl > env->imbalance &&
6985 !check_cpu_capacity(rq, env->sd))
6989 * For the load comparisons with the other cpu's, consider
6990 * the weighted_cpuload() scaled with the cpu capacity, so
6991 * that the load can be moved away from the cpu that is
6992 * potentially running at a lower capacity.
6994 * Thus we're looking for max(wl_i / capacity_i), crosswise
6995 * multiplication to rid ourselves of the division works out
6996 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6997 * our previous maximum.
6999 if (wl * busiest_capacity > busiest_load * capacity) {
7001 busiest_capacity = capacity;
7010 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7011 * so long as it is large enough.
7013 #define MAX_PINNED_INTERVAL 512
7015 /* Working cpumask for load_balance and load_balance_newidle. */
7016 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7018 static int need_active_balance(struct lb_env *env)
7020 struct sched_domain *sd = env->sd;
7022 if (env->idle == CPU_NEWLY_IDLE) {
7025 * ASYM_PACKING needs to force migrate tasks from busy but
7026 * higher numbered CPUs in order to pack all tasks in the
7027 * lowest numbered CPUs.
7029 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7034 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7035 * It's worth migrating the task if the src_cpu's capacity is reduced
7036 * because of other sched_class or IRQs if more capacity stays
7037 * available on dst_cpu.
7039 if ((env->idle != CPU_NOT_IDLE) &&
7040 (env->src_rq->cfs.h_nr_running == 1)) {
7041 if ((check_cpu_capacity(env->src_rq, sd)) &&
7042 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7046 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7049 static int active_load_balance_cpu_stop(void *data);
7051 static int should_we_balance(struct lb_env *env)
7053 struct sched_group *sg = env->sd->groups;
7054 struct cpumask *sg_cpus, *sg_mask;
7055 int cpu, balance_cpu = -1;
7058 * In the newly idle case, we will allow all the cpu's
7059 * to do the newly idle load balance.
7061 if (env->idle == CPU_NEWLY_IDLE)
7064 sg_cpus = sched_group_cpus(sg);
7065 sg_mask = sched_group_mask(sg);
7066 /* Try to find first idle cpu */
7067 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7068 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7075 if (balance_cpu == -1)
7076 balance_cpu = group_balance_cpu(sg);
7079 * First idle cpu or the first cpu(busiest) in this sched group
7080 * is eligible for doing load balancing at this and above domains.
7082 return balance_cpu == env->dst_cpu;
7086 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7087 * tasks if there is an imbalance.
7089 static int load_balance(int this_cpu, struct rq *this_rq,
7090 struct sched_domain *sd, enum cpu_idle_type idle,
7091 int *continue_balancing)
7093 int ld_moved, cur_ld_moved, active_balance = 0;
7094 struct sched_domain *sd_parent = sd->parent;
7095 struct sched_group *group;
7097 unsigned long flags;
7098 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7100 struct lb_env env = {
7102 .dst_cpu = this_cpu,
7104 .dst_grpmask = sched_group_cpus(sd->groups),
7106 .loop_break = sched_nr_migrate_break,
7109 .tasks = LIST_HEAD_INIT(env.tasks),
7113 * For NEWLY_IDLE load_balancing, we don't need to consider
7114 * other cpus in our group
7116 if (idle == CPU_NEWLY_IDLE)
7117 env.dst_grpmask = NULL;
7119 cpumask_copy(cpus, cpu_active_mask);
7121 schedstat_inc(sd, lb_count[idle]);
7124 if (!should_we_balance(&env)) {
7125 *continue_balancing = 0;
7129 group = find_busiest_group(&env);
7131 schedstat_inc(sd, lb_nobusyg[idle]);
7135 busiest = find_busiest_queue(&env, group);
7137 schedstat_inc(sd, lb_nobusyq[idle]);
7141 BUG_ON(busiest == env.dst_rq);
7143 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7145 env.src_cpu = busiest->cpu;
7146 env.src_rq = busiest;
7149 if (busiest->nr_running > 1) {
7151 * Attempt to move tasks. If find_busiest_group has found
7152 * an imbalance but busiest->nr_running <= 1, the group is
7153 * still unbalanced. ld_moved simply stays zero, so it is
7154 * correctly treated as an imbalance.
7156 env.flags |= LBF_ALL_PINNED;
7157 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7160 raw_spin_lock_irqsave(&busiest->lock, flags);
7163 * cur_ld_moved - load moved in current iteration
7164 * ld_moved - cumulative load moved across iterations
7166 cur_ld_moved = detach_tasks(&env);
7169 * We've detached some tasks from busiest_rq. Every
7170 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7171 * unlock busiest->lock, and we are able to be sure
7172 * that nobody can manipulate the tasks in parallel.
7173 * See task_rq_lock() family for the details.
7176 raw_spin_unlock(&busiest->lock);
7180 ld_moved += cur_ld_moved;
7183 local_irq_restore(flags);
7185 if (env.flags & LBF_NEED_BREAK) {
7186 env.flags &= ~LBF_NEED_BREAK;
7191 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7192 * us and move them to an alternate dst_cpu in our sched_group
7193 * where they can run. The upper limit on how many times we
7194 * iterate on same src_cpu is dependent on number of cpus in our
7197 * This changes load balance semantics a bit on who can move
7198 * load to a given_cpu. In addition to the given_cpu itself
7199 * (or a ilb_cpu acting on its behalf where given_cpu is
7200 * nohz-idle), we now have balance_cpu in a position to move
7201 * load to given_cpu. In rare situations, this may cause
7202 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7203 * _independently_ and at _same_ time to move some load to
7204 * given_cpu) causing exceess load to be moved to given_cpu.
7205 * This however should not happen so much in practice and
7206 * moreover subsequent load balance cycles should correct the
7207 * excess load moved.
7209 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7211 /* Prevent to re-select dst_cpu via env's cpus */
7212 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7214 env.dst_rq = cpu_rq(env.new_dst_cpu);
7215 env.dst_cpu = env.new_dst_cpu;
7216 env.flags &= ~LBF_DST_PINNED;
7218 env.loop_break = sched_nr_migrate_break;
7221 * Go back to "more_balance" rather than "redo" since we
7222 * need to continue with same src_cpu.
7228 * We failed to reach balance because of affinity.
7231 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7233 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7234 *group_imbalance = 1;
7237 /* All tasks on this runqueue were pinned by CPU affinity */
7238 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7239 cpumask_clear_cpu(cpu_of(busiest), cpus);
7240 if (!cpumask_empty(cpus)) {
7242 env.loop_break = sched_nr_migrate_break;
7245 goto out_all_pinned;
7250 schedstat_inc(sd, lb_failed[idle]);
7252 * Increment the failure counter only on periodic balance.
7253 * We do not want newidle balance, which can be very
7254 * frequent, pollute the failure counter causing
7255 * excessive cache_hot migrations and active balances.
7257 if (idle != CPU_NEWLY_IDLE)
7258 sd->nr_balance_failed++;
7260 if (need_active_balance(&env)) {
7261 raw_spin_lock_irqsave(&busiest->lock, flags);
7263 /* don't kick the active_load_balance_cpu_stop,
7264 * if the curr task on busiest cpu can't be
7267 if (!cpumask_test_cpu(this_cpu,
7268 tsk_cpus_allowed(busiest->curr))) {
7269 raw_spin_unlock_irqrestore(&busiest->lock,
7271 env.flags |= LBF_ALL_PINNED;
7272 goto out_one_pinned;
7276 * ->active_balance synchronizes accesses to
7277 * ->active_balance_work. Once set, it's cleared
7278 * only after active load balance is finished.
7280 if (!busiest->active_balance) {
7281 busiest->active_balance = 1;
7282 busiest->push_cpu = this_cpu;
7285 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7287 if (active_balance) {
7288 stop_one_cpu_nowait(cpu_of(busiest),
7289 active_load_balance_cpu_stop, busiest,
7290 &busiest->active_balance_work);
7294 * We've kicked active balancing, reset the failure
7297 sd->nr_balance_failed = sd->cache_nice_tries+1;
7300 sd->nr_balance_failed = 0;
7302 if (likely(!active_balance)) {
7303 /* We were unbalanced, so reset the balancing interval */
7304 sd->balance_interval = sd->min_interval;
7307 * If we've begun active balancing, start to back off. This
7308 * case may not be covered by the all_pinned logic if there
7309 * is only 1 task on the busy runqueue (because we don't call
7312 if (sd->balance_interval < sd->max_interval)
7313 sd->balance_interval *= 2;
7320 * We reach balance although we may have faced some affinity
7321 * constraints. Clear the imbalance flag if it was set.
7324 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7326 if (*group_imbalance)
7327 *group_imbalance = 0;
7332 * We reach balance because all tasks are pinned at this level so
7333 * we can't migrate them. Let the imbalance flag set so parent level
7334 * can try to migrate them.
7336 schedstat_inc(sd, lb_balanced[idle]);
7338 sd->nr_balance_failed = 0;
7341 /* tune up the balancing interval */
7342 if (((env.flags & LBF_ALL_PINNED) &&
7343 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7344 (sd->balance_interval < sd->max_interval))
7345 sd->balance_interval *= 2;
7352 static inline unsigned long
7353 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7355 unsigned long interval = sd->balance_interval;
7358 interval *= sd->busy_factor;
7360 /* scale ms to jiffies */
7361 interval = msecs_to_jiffies(interval);
7362 interval = clamp(interval, 1UL, max_load_balance_interval);
7368 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7370 unsigned long interval, next;
7372 interval = get_sd_balance_interval(sd, cpu_busy);
7373 next = sd->last_balance + interval;
7375 if (time_after(*next_balance, next))
7376 *next_balance = next;
7380 * idle_balance is called by schedule() if this_cpu is about to become
7381 * idle. Attempts to pull tasks from other CPUs.
7383 static int idle_balance(struct rq *this_rq)
7385 unsigned long next_balance = jiffies + HZ;
7386 int this_cpu = this_rq->cpu;
7387 struct sched_domain *sd;
7388 int pulled_task = 0;
7392 * We must set idle_stamp _before_ calling idle_balance(), such that we
7393 * measure the duration of idle_balance() as idle time.
7395 this_rq->idle_stamp = rq_clock(this_rq);
7397 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7398 !this_rq->rd->overload) {
7400 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7402 update_next_balance(sd, 0, &next_balance);
7408 raw_spin_unlock(&this_rq->lock);
7410 update_blocked_averages(this_cpu);
7412 for_each_domain(this_cpu, sd) {
7413 int continue_balancing = 1;
7414 u64 t0, domain_cost;
7416 if (!(sd->flags & SD_LOAD_BALANCE))
7419 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7420 update_next_balance(sd, 0, &next_balance);
7424 if (sd->flags & SD_BALANCE_NEWIDLE) {
7425 t0 = sched_clock_cpu(this_cpu);
7427 pulled_task = load_balance(this_cpu, this_rq,
7429 &continue_balancing);
7431 domain_cost = sched_clock_cpu(this_cpu) - t0;
7432 if (domain_cost > sd->max_newidle_lb_cost)
7433 sd->max_newidle_lb_cost = domain_cost;
7435 curr_cost += domain_cost;
7438 update_next_balance(sd, 0, &next_balance);
7441 * Stop searching for tasks to pull if there are
7442 * now runnable tasks on this rq.
7444 if (pulled_task || this_rq->nr_running > 0)
7449 raw_spin_lock(&this_rq->lock);
7451 if (curr_cost > this_rq->max_idle_balance_cost)
7452 this_rq->max_idle_balance_cost = curr_cost;
7455 * While browsing the domains, we released the rq lock, a task could
7456 * have been enqueued in the meantime. Since we're not going idle,
7457 * pretend we pulled a task.
7459 if (this_rq->cfs.h_nr_running && !pulled_task)
7463 /* Move the next balance forward */
7464 if (time_after(this_rq->next_balance, next_balance))
7465 this_rq->next_balance = next_balance;
7467 /* Is there a task of a high priority class? */
7468 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7472 this_rq->idle_stamp = 0;
7478 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7479 * running tasks off the busiest CPU onto idle CPUs. It requires at
7480 * least 1 task to be running on each physical CPU where possible, and
7481 * avoids physical / logical imbalances.
7483 static int active_load_balance_cpu_stop(void *data)
7485 struct rq *busiest_rq = data;
7486 int busiest_cpu = cpu_of(busiest_rq);
7487 int target_cpu = busiest_rq->push_cpu;
7488 struct rq *target_rq = cpu_rq(target_cpu);
7489 struct sched_domain *sd;
7490 struct task_struct *p = NULL;
7492 raw_spin_lock_irq(&busiest_rq->lock);
7494 /* make sure the requested cpu hasn't gone down in the meantime */
7495 if (unlikely(busiest_cpu != smp_processor_id() ||
7496 !busiest_rq->active_balance))
7499 /* Is there any task to move? */
7500 if (busiest_rq->nr_running <= 1)
7504 * This condition is "impossible", if it occurs
7505 * we need to fix it. Originally reported by
7506 * Bjorn Helgaas on a 128-cpu setup.
7508 BUG_ON(busiest_rq == target_rq);
7510 /* Search for an sd spanning us and the target CPU. */
7512 for_each_domain(target_cpu, sd) {
7513 if ((sd->flags & SD_LOAD_BALANCE) &&
7514 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7519 struct lb_env env = {
7521 .dst_cpu = target_cpu,
7522 .dst_rq = target_rq,
7523 .src_cpu = busiest_rq->cpu,
7524 .src_rq = busiest_rq,
7528 schedstat_inc(sd, alb_count);
7530 p = detach_one_task(&env);
7532 schedstat_inc(sd, alb_pushed);
7534 schedstat_inc(sd, alb_failed);
7538 busiest_rq->active_balance = 0;
7539 raw_spin_unlock(&busiest_rq->lock);
7542 attach_one_task(target_rq, p);
7549 static inline int on_null_domain(struct rq *rq)
7551 return unlikely(!rcu_dereference_sched(rq->sd));
7554 #ifdef CONFIG_NO_HZ_COMMON
7556 * idle load balancing details
7557 * - When one of the busy CPUs notice that there may be an idle rebalancing
7558 * needed, they will kick the idle load balancer, which then does idle
7559 * load balancing for all the idle CPUs.
7562 cpumask_var_t idle_cpus_mask;
7564 unsigned long next_balance; /* in jiffy units */
7565 } nohz ____cacheline_aligned;
7567 static inline int find_new_ilb(void)
7569 int ilb = cpumask_first(nohz.idle_cpus_mask);
7571 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7578 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7579 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7580 * CPU (if there is one).
7582 static void nohz_balancer_kick(void)
7586 nohz.next_balance++;
7588 ilb_cpu = find_new_ilb();
7590 if (ilb_cpu >= nr_cpu_ids)
7593 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7596 * Use smp_send_reschedule() instead of resched_cpu().
7597 * This way we generate a sched IPI on the target cpu which
7598 * is idle. And the softirq performing nohz idle load balance
7599 * will be run before returning from the IPI.
7601 smp_send_reschedule(ilb_cpu);
7605 static inline void nohz_balance_exit_idle(int cpu)
7607 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7609 * Completely isolated CPUs don't ever set, so we must test.
7611 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7612 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7613 atomic_dec(&nohz.nr_cpus);
7615 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7619 static inline void set_cpu_sd_state_busy(void)
7621 struct sched_domain *sd;
7622 int cpu = smp_processor_id();
7625 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7627 if (!sd || !sd->nohz_idle)
7631 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7636 void set_cpu_sd_state_idle(void)
7638 struct sched_domain *sd;
7639 int cpu = smp_processor_id();
7642 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7644 if (!sd || sd->nohz_idle)
7648 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7654 * This routine will record that the cpu is going idle with tick stopped.
7655 * This info will be used in performing idle load balancing in the future.
7657 void nohz_balance_enter_idle(int cpu)
7660 * If this cpu is going down, then nothing needs to be done.
7662 if (!cpu_active(cpu))
7665 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7669 * If we're a completely isolated CPU, we don't play.
7671 if (on_null_domain(cpu_rq(cpu)))
7674 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7675 atomic_inc(&nohz.nr_cpus);
7676 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7679 static int sched_ilb_notifier(struct notifier_block *nfb,
7680 unsigned long action, void *hcpu)
7682 switch (action & ~CPU_TASKS_FROZEN) {
7684 nohz_balance_exit_idle(smp_processor_id());
7692 static DEFINE_SPINLOCK(balancing);
7695 * Scale the max load_balance interval with the number of CPUs in the system.
7696 * This trades load-balance latency on larger machines for less cross talk.
7698 void update_max_interval(void)
7700 max_load_balance_interval = HZ*num_online_cpus()/10;
7704 * It checks each scheduling domain to see if it is due to be balanced,
7705 * and initiates a balancing operation if so.
7707 * Balancing parameters are set up in init_sched_domains.
7709 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7711 int continue_balancing = 1;
7713 unsigned long interval;
7714 struct sched_domain *sd;
7715 /* Earliest time when we have to do rebalance again */
7716 unsigned long next_balance = jiffies + 60*HZ;
7717 int update_next_balance = 0;
7718 int need_serialize, need_decay = 0;
7721 update_blocked_averages(cpu);
7724 for_each_domain(cpu, sd) {
7726 * Decay the newidle max times here because this is a regular
7727 * visit to all the domains. Decay ~1% per second.
7729 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7730 sd->max_newidle_lb_cost =
7731 (sd->max_newidle_lb_cost * 253) / 256;
7732 sd->next_decay_max_lb_cost = jiffies + HZ;
7735 max_cost += sd->max_newidle_lb_cost;
7737 if (!(sd->flags & SD_LOAD_BALANCE))
7741 * Stop the load balance at this level. There is another
7742 * CPU in our sched group which is doing load balancing more
7745 if (!continue_balancing) {
7751 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7753 need_serialize = sd->flags & SD_SERIALIZE;
7754 if (need_serialize) {
7755 if (!spin_trylock(&balancing))
7759 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7760 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7762 * The LBF_DST_PINNED logic could have changed
7763 * env->dst_cpu, so we can't know our idle
7764 * state even if we migrated tasks. Update it.
7766 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7768 sd->last_balance = jiffies;
7769 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7772 spin_unlock(&balancing);
7774 if (time_after(next_balance, sd->last_balance + interval)) {
7775 next_balance = sd->last_balance + interval;
7776 update_next_balance = 1;
7781 * Ensure the rq-wide value also decays but keep it at a
7782 * reasonable floor to avoid funnies with rq->avg_idle.
7784 rq->max_idle_balance_cost =
7785 max((u64)sysctl_sched_migration_cost, max_cost);
7790 * next_balance will be updated only when there is a need.
7791 * When the cpu is attached to null domain for ex, it will not be
7794 if (likely(update_next_balance)) {
7795 rq->next_balance = next_balance;
7797 #ifdef CONFIG_NO_HZ_COMMON
7799 * If this CPU has been elected to perform the nohz idle
7800 * balance. Other idle CPUs have already rebalanced with
7801 * nohz_idle_balance() and nohz.next_balance has been
7802 * updated accordingly. This CPU is now running the idle load
7803 * balance for itself and we need to update the
7804 * nohz.next_balance accordingly.
7806 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
7807 nohz.next_balance = rq->next_balance;
7812 #ifdef CONFIG_NO_HZ_COMMON
7814 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7815 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7817 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7819 int this_cpu = this_rq->cpu;
7822 /* Earliest time when we have to do rebalance again */
7823 unsigned long next_balance = jiffies + 60*HZ;
7824 int update_next_balance = 0;
7826 if (idle != CPU_IDLE ||
7827 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7830 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7831 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7835 * If this cpu gets work to do, stop the load balancing
7836 * work being done for other cpus. Next load
7837 * balancing owner will pick it up.
7842 rq = cpu_rq(balance_cpu);
7845 * If time for next balance is due,
7848 if (time_after_eq(jiffies, rq->next_balance)) {
7849 raw_spin_lock_irq(&rq->lock);
7850 update_rq_clock(rq);
7851 update_idle_cpu_load(rq);
7852 raw_spin_unlock_irq(&rq->lock);
7853 rebalance_domains(rq, CPU_IDLE);
7856 if (time_after(next_balance, rq->next_balance)) {
7857 next_balance = rq->next_balance;
7858 update_next_balance = 1;
7863 * next_balance will be updated only when there is a need.
7864 * When the CPU is attached to null domain for ex, it will not be
7867 if (likely(update_next_balance))
7868 nohz.next_balance = next_balance;
7870 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7874 * Current heuristic for kicking the idle load balancer in the presence
7875 * of an idle cpu in the system.
7876 * - This rq has more than one task.
7877 * - This rq has at least one CFS task and the capacity of the CPU is
7878 * significantly reduced because of RT tasks or IRQs.
7879 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7880 * multiple busy cpu.
7881 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7882 * domain span are idle.
7884 static inline bool nohz_kick_needed(struct rq *rq)
7886 unsigned long now = jiffies;
7887 struct sched_domain *sd;
7888 struct sched_group_capacity *sgc;
7889 int nr_busy, cpu = rq->cpu;
7892 if (unlikely(rq->idle_balance))
7896 * We may be recently in ticked or tickless idle mode. At the first
7897 * busy tick after returning from idle, we will update the busy stats.
7899 set_cpu_sd_state_busy();
7900 nohz_balance_exit_idle(cpu);
7903 * None are in tickless mode and hence no need for NOHZ idle load
7906 if (likely(!atomic_read(&nohz.nr_cpus)))
7909 if (time_before(now, nohz.next_balance))
7912 if (rq->nr_running >= 2)
7916 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7918 sgc = sd->groups->sgc;
7919 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7928 sd = rcu_dereference(rq->sd);
7930 if ((rq->cfs.h_nr_running >= 1) &&
7931 check_cpu_capacity(rq, sd)) {
7937 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7938 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7939 sched_domain_span(sd)) < cpu)) {
7949 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7953 * run_rebalance_domains is triggered when needed from the scheduler tick.
7954 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7956 static void run_rebalance_domains(struct softirq_action *h)
7958 struct rq *this_rq = this_rq();
7959 enum cpu_idle_type idle = this_rq->idle_balance ?
7960 CPU_IDLE : CPU_NOT_IDLE;
7963 * If this cpu has a pending nohz_balance_kick, then do the
7964 * balancing on behalf of the other idle cpus whose ticks are
7965 * stopped. Do nohz_idle_balance *before* rebalance_domains to
7966 * give the idle cpus a chance to load balance. Else we may
7967 * load balance only within the local sched_domain hierarchy
7968 * and abort nohz_idle_balance altogether if we pull some load.
7970 nohz_idle_balance(this_rq, idle);
7971 rebalance_domains(this_rq, idle);
7975 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7977 void trigger_load_balance(struct rq *rq)
7979 /* Don't need to rebalance while attached to NULL domain */
7980 if (unlikely(on_null_domain(rq)))
7983 if (time_after_eq(jiffies, rq->next_balance))
7984 raise_softirq(SCHED_SOFTIRQ);
7985 #ifdef CONFIG_NO_HZ_COMMON
7986 if (nohz_kick_needed(rq))
7987 nohz_balancer_kick();
7991 static void rq_online_fair(struct rq *rq)
7995 update_runtime_enabled(rq);
7998 static void rq_offline_fair(struct rq *rq)
8002 /* Ensure any throttled groups are reachable by pick_next_task */
8003 unthrottle_offline_cfs_rqs(rq);
8006 #endif /* CONFIG_SMP */
8009 * scheduler tick hitting a task of our scheduling class:
8011 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8013 struct cfs_rq *cfs_rq;
8014 struct sched_entity *se = &curr->se;
8016 for_each_sched_entity(se) {
8017 cfs_rq = cfs_rq_of(se);
8018 entity_tick(cfs_rq, se, queued);
8021 if (static_branch_unlikely(&sched_numa_balancing))
8022 task_tick_numa(rq, curr);
8026 * called on fork with the child task as argument from the parent's context
8027 * - child not yet on the tasklist
8028 * - preemption disabled
8030 static void task_fork_fair(struct task_struct *p)
8032 struct cfs_rq *cfs_rq;
8033 struct sched_entity *se = &p->se, *curr;
8034 int this_cpu = smp_processor_id();
8035 struct rq *rq = this_rq();
8036 unsigned long flags;
8038 raw_spin_lock_irqsave(&rq->lock, flags);
8040 update_rq_clock(rq);
8042 cfs_rq = task_cfs_rq(current);
8043 curr = cfs_rq->curr;
8046 * Not only the cpu but also the task_group of the parent might have
8047 * been changed after parent->se.parent,cfs_rq were copied to
8048 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8049 * of child point to valid ones.
8052 __set_task_cpu(p, this_cpu);
8055 update_curr(cfs_rq);
8058 se->vruntime = curr->vruntime;
8059 place_entity(cfs_rq, se, 1);
8061 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8063 * Upon rescheduling, sched_class::put_prev_task() will place
8064 * 'current' within the tree based on its new key value.
8066 swap(curr->vruntime, se->vruntime);
8070 se->vruntime -= cfs_rq->min_vruntime;
8072 raw_spin_unlock_irqrestore(&rq->lock, flags);
8076 * Priority of the task has changed. Check to see if we preempt
8080 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8082 if (!task_on_rq_queued(p))
8086 * Reschedule if we are currently running on this runqueue and
8087 * our priority decreased, or if we are not currently running on
8088 * this runqueue and our priority is higher than the current's
8090 if (rq->curr == p) {
8091 if (p->prio > oldprio)
8094 check_preempt_curr(rq, p, 0);
8097 static inline bool vruntime_normalized(struct task_struct *p)
8099 struct sched_entity *se = &p->se;
8102 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8103 * the dequeue_entity(.flags=0) will already have normalized the
8110 * When !on_rq, vruntime of the task has usually NOT been normalized.
8111 * But there are some cases where it has already been normalized:
8113 * - A forked child which is waiting for being woken up by
8114 * wake_up_new_task().
8115 * - A task which has been woken up by try_to_wake_up() and
8116 * waiting for actually being woken up by sched_ttwu_pending().
8118 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8124 static void detach_task_cfs_rq(struct task_struct *p)
8126 struct sched_entity *se = &p->se;
8127 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8129 if (!vruntime_normalized(p)) {
8131 * Fix up our vruntime so that the current sleep doesn't
8132 * cause 'unlimited' sleep bonus.
8134 place_entity(cfs_rq, se, 0);
8135 se->vruntime -= cfs_rq->min_vruntime;
8138 /* Catch up with the cfs_rq and remove our load when we leave */
8139 detach_entity_load_avg(cfs_rq, se);
8142 static void attach_task_cfs_rq(struct task_struct *p)
8144 struct sched_entity *se = &p->se;
8145 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8147 #ifdef CONFIG_FAIR_GROUP_SCHED
8149 * Since the real-depth could have been changed (only FAIR
8150 * class maintain depth value), reset depth properly.
8152 se->depth = se->parent ? se->parent->depth + 1 : 0;
8155 /* Synchronize task with its cfs_rq */
8156 attach_entity_load_avg(cfs_rq, se);
8158 if (!vruntime_normalized(p))
8159 se->vruntime += cfs_rq->min_vruntime;
8162 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8164 detach_task_cfs_rq(p);
8167 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8169 attach_task_cfs_rq(p);
8171 if (task_on_rq_queued(p)) {
8173 * We were most likely switched from sched_rt, so
8174 * kick off the schedule if running, otherwise just see
8175 * if we can still preempt the current task.
8180 check_preempt_curr(rq, p, 0);
8184 /* Account for a task changing its policy or group.
8186 * This routine is mostly called to set cfs_rq->curr field when a task
8187 * migrates between groups/classes.
8189 static void set_curr_task_fair(struct rq *rq)
8191 struct sched_entity *se = &rq->curr->se;
8193 for_each_sched_entity(se) {
8194 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8196 set_next_entity(cfs_rq, se);
8197 /* ensure bandwidth has been allocated on our new cfs_rq */
8198 account_cfs_rq_runtime(cfs_rq, 0);
8202 void init_cfs_rq(struct cfs_rq *cfs_rq)
8204 cfs_rq->tasks_timeline = RB_ROOT;
8205 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8206 #ifndef CONFIG_64BIT
8207 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8210 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8211 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8215 #ifdef CONFIG_FAIR_GROUP_SCHED
8216 static void task_move_group_fair(struct task_struct *p)
8218 detach_task_cfs_rq(p);
8219 set_task_rq(p, task_cpu(p));
8222 /* Tell se's cfs_rq has been changed -- migrated */
8223 p->se.avg.last_update_time = 0;
8225 attach_task_cfs_rq(p);
8228 void free_fair_sched_group(struct task_group *tg)
8232 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8234 for_each_possible_cpu(i) {
8236 kfree(tg->cfs_rq[i]);
8239 remove_entity_load_avg(tg->se[i]);
8248 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8250 struct cfs_rq *cfs_rq;
8251 struct sched_entity *se;
8254 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8257 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8261 tg->shares = NICE_0_LOAD;
8263 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8265 for_each_possible_cpu(i) {
8266 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8267 GFP_KERNEL, cpu_to_node(i));
8271 se = kzalloc_node(sizeof(struct sched_entity),
8272 GFP_KERNEL, cpu_to_node(i));
8276 init_cfs_rq(cfs_rq);
8277 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8278 init_entity_runnable_average(se);
8289 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8291 struct rq *rq = cpu_rq(cpu);
8292 unsigned long flags;
8295 * Only empty task groups can be destroyed; so we can speculatively
8296 * check on_list without danger of it being re-added.
8298 if (!tg->cfs_rq[cpu]->on_list)
8301 raw_spin_lock_irqsave(&rq->lock, flags);
8302 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8303 raw_spin_unlock_irqrestore(&rq->lock, flags);
8306 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8307 struct sched_entity *se, int cpu,
8308 struct sched_entity *parent)
8310 struct rq *rq = cpu_rq(cpu);
8314 init_cfs_rq_runtime(cfs_rq);
8316 tg->cfs_rq[cpu] = cfs_rq;
8319 /* se could be NULL for root_task_group */
8324 se->cfs_rq = &rq->cfs;
8327 se->cfs_rq = parent->my_q;
8328 se->depth = parent->depth + 1;
8332 /* guarantee group entities always have weight */
8333 update_load_set(&se->load, NICE_0_LOAD);
8334 se->parent = parent;
8337 static DEFINE_MUTEX(shares_mutex);
8339 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8342 unsigned long flags;
8345 * We can't change the weight of the root cgroup.
8350 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8352 mutex_lock(&shares_mutex);
8353 if (tg->shares == shares)
8356 tg->shares = shares;
8357 for_each_possible_cpu(i) {
8358 struct rq *rq = cpu_rq(i);
8359 struct sched_entity *se;
8362 /* Propagate contribution to hierarchy */
8363 raw_spin_lock_irqsave(&rq->lock, flags);
8365 /* Possible calls to update_curr() need rq clock */
8366 update_rq_clock(rq);
8367 for_each_sched_entity(se)
8368 update_cfs_shares(group_cfs_rq(se));
8369 raw_spin_unlock_irqrestore(&rq->lock, flags);
8373 mutex_unlock(&shares_mutex);
8376 #else /* CONFIG_FAIR_GROUP_SCHED */
8378 void free_fair_sched_group(struct task_group *tg) { }
8380 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8385 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8387 #endif /* CONFIG_FAIR_GROUP_SCHED */
8390 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8392 struct sched_entity *se = &task->se;
8393 unsigned int rr_interval = 0;
8396 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8399 if (rq->cfs.load.weight)
8400 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8406 * All the scheduling class methods:
8408 const struct sched_class fair_sched_class = {
8409 .next = &idle_sched_class,
8410 .enqueue_task = enqueue_task_fair,
8411 .dequeue_task = dequeue_task_fair,
8412 .yield_task = yield_task_fair,
8413 .yield_to_task = yield_to_task_fair,
8415 .check_preempt_curr = check_preempt_wakeup,
8417 .pick_next_task = pick_next_task_fair,
8418 .put_prev_task = put_prev_task_fair,
8421 .select_task_rq = select_task_rq_fair,
8422 .migrate_task_rq = migrate_task_rq_fair,
8424 .rq_online = rq_online_fair,
8425 .rq_offline = rq_offline_fair,
8427 .task_waking = task_waking_fair,
8428 .task_dead = task_dead_fair,
8429 .set_cpus_allowed = set_cpus_allowed_common,
8432 .set_curr_task = set_curr_task_fair,
8433 .task_tick = task_tick_fair,
8434 .task_fork = task_fork_fair,
8436 .prio_changed = prio_changed_fair,
8437 .switched_from = switched_from_fair,
8438 .switched_to = switched_to_fair,
8440 .get_rr_interval = get_rr_interval_fair,
8442 .update_curr = update_curr_fair,
8444 #ifdef CONFIG_FAIR_GROUP_SCHED
8445 .task_move_group = task_move_group_fair,
8449 #ifdef CONFIG_SCHED_DEBUG
8450 void print_cfs_stats(struct seq_file *m, int cpu)
8452 struct cfs_rq *cfs_rq;
8455 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8456 print_cfs_rq(m, cpu, cfs_rq);
8460 #ifdef CONFIG_NUMA_BALANCING
8461 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8464 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8466 for_each_online_node(node) {
8467 if (p->numa_faults) {
8468 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8469 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8471 if (p->numa_group) {
8472 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8473 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8475 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8478 #endif /* CONFIG_NUMA_BALANCING */
8479 #endif /* CONFIG_SCHED_DEBUG */
8481 __init void init_sched_fair_class(void)
8484 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8486 #ifdef CONFIG_NO_HZ_COMMON
8487 nohz.next_balance = jiffies;
8488 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8489 cpu_notifier(sched_ilb_notifier, 0);