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 <pzijlstr@redhat.com>
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 void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
289 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
291 if (!cfs_rq->on_list) {
293 * Ensure we either appear before our parent (if already
294 * enqueued) or force our parent to appear after us when it is
295 * enqueued. The fact that we always enqueue bottom-up
296 * reduces this to two cases.
298 if (cfs_rq->tg->parent &&
299 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
300 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
303 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
304 &rq_of(cfs_rq)->leaf_cfs_rq_list);
308 /* We should have no load, but we need to update last_decay. */
309 update_cfs_rq_blocked_load(cfs_rq, 0);
313 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
315 if (cfs_rq->on_list) {
316 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
321 /* Iterate thr' all leaf cfs_rq's on a runqueue */
322 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
323 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
325 /* Do the two (enqueued) entities belong to the same group ? */
326 static inline struct cfs_rq *
327 is_same_group(struct sched_entity *se, struct sched_entity *pse)
329 if (se->cfs_rq == pse->cfs_rq)
335 static inline struct sched_entity *parent_entity(struct sched_entity *se)
341 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
343 int se_depth, pse_depth;
346 * preemption test can be made between sibling entities who are in the
347 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
348 * both tasks until we find their ancestors who are siblings of common
352 /* First walk up until both entities are at same depth */
353 se_depth = (*se)->depth;
354 pse_depth = (*pse)->depth;
356 while (se_depth > pse_depth) {
358 *se = parent_entity(*se);
361 while (pse_depth > se_depth) {
363 *pse = parent_entity(*pse);
366 while (!is_same_group(*se, *pse)) {
367 *se = parent_entity(*se);
368 *pse = parent_entity(*pse);
372 #else /* !CONFIG_FAIR_GROUP_SCHED */
374 static inline struct task_struct *task_of(struct sched_entity *se)
376 return container_of(se, struct task_struct, se);
379 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
381 return container_of(cfs_rq, struct rq, cfs);
384 #define entity_is_task(se) 1
386 #define for_each_sched_entity(se) \
387 for (; se; se = NULL)
389 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
391 return &task_rq(p)->cfs;
394 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
396 struct task_struct *p = task_of(se);
397 struct rq *rq = task_rq(p);
402 /* runqueue "owned" by this group */
403 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
408 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
412 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
416 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
417 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
419 static inline struct sched_entity *parent_entity(struct sched_entity *se)
425 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
429 #endif /* CONFIG_FAIR_GROUP_SCHED */
431 static __always_inline
432 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
434 /**************************************************************
435 * Scheduling class tree data structure manipulation methods:
438 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
440 s64 delta = (s64)(vruntime - max_vruntime);
442 max_vruntime = vruntime;
447 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
449 s64 delta = (s64)(vruntime - min_vruntime);
451 min_vruntime = vruntime;
456 static inline int entity_before(struct sched_entity *a,
457 struct sched_entity *b)
459 return (s64)(a->vruntime - b->vruntime) < 0;
462 static void update_min_vruntime(struct cfs_rq *cfs_rq)
464 u64 vruntime = cfs_rq->min_vruntime;
467 vruntime = cfs_rq->curr->vruntime;
469 if (cfs_rq->rb_leftmost) {
470 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
475 vruntime = se->vruntime;
477 vruntime = min_vruntime(vruntime, se->vruntime);
480 /* ensure we never gain time by being placed backwards. */
481 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
484 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
489 * Enqueue an entity into the rb-tree:
491 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
493 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
494 struct rb_node *parent = NULL;
495 struct sched_entity *entry;
499 * Find the right place in the rbtree:
503 entry = rb_entry(parent, struct sched_entity, run_node);
505 * We dont care about collisions. Nodes with
506 * the same key stay together.
508 if (entity_before(se, entry)) {
509 link = &parent->rb_left;
511 link = &parent->rb_right;
517 * Maintain a cache of leftmost tree entries (it is frequently
521 cfs_rq->rb_leftmost = &se->run_node;
523 rb_link_node(&se->run_node, parent, link);
524 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
527 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
529 if (cfs_rq->rb_leftmost == &se->run_node) {
530 struct rb_node *next_node;
532 next_node = rb_next(&se->run_node);
533 cfs_rq->rb_leftmost = next_node;
536 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
539 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
541 struct rb_node *left = cfs_rq->rb_leftmost;
546 return rb_entry(left, struct sched_entity, run_node);
549 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
551 struct rb_node *next = rb_next(&se->run_node);
556 return rb_entry(next, struct sched_entity, run_node);
559 #ifdef CONFIG_SCHED_DEBUG
560 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
562 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
567 return rb_entry(last, struct sched_entity, run_node);
570 /**************************************************************
571 * Scheduling class statistics methods:
574 int sched_proc_update_handler(struct ctl_table *table, int write,
575 void __user *buffer, size_t *lenp,
578 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
579 unsigned int factor = get_update_sysctl_factor();
584 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
585 sysctl_sched_min_granularity);
587 #define WRT_SYSCTL(name) \
588 (normalized_sysctl_##name = sysctl_##name / (factor))
589 WRT_SYSCTL(sched_min_granularity);
590 WRT_SYSCTL(sched_latency);
591 WRT_SYSCTL(sched_wakeup_granularity);
601 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
603 if (unlikely(se->load.weight != NICE_0_LOAD))
604 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
610 * The idea is to set a period in which each task runs once.
612 * When there are too many tasks (sched_nr_latency) we have to stretch
613 * this period because otherwise the slices get too small.
615 * p = (nr <= nl) ? l : l*nr/nl
617 static u64 __sched_period(unsigned long nr_running)
619 u64 period = sysctl_sched_latency;
620 unsigned long nr_latency = sched_nr_latency;
622 if (unlikely(nr_running > nr_latency)) {
623 period = sysctl_sched_min_granularity;
624 period *= nr_running;
631 * We calculate the wall-time slice from the period by taking a part
632 * proportional to the weight.
636 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
638 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
640 for_each_sched_entity(se) {
641 struct load_weight *load;
642 struct load_weight lw;
644 cfs_rq = cfs_rq_of(se);
645 load = &cfs_rq->load;
647 if (unlikely(!se->on_rq)) {
650 update_load_add(&lw, se->load.weight);
653 slice = __calc_delta(slice, se->load.weight, load);
659 * We calculate the vruntime slice of a to-be-inserted task.
663 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
665 return calc_delta_fair(sched_slice(cfs_rq, se), se);
669 static int select_idle_sibling(struct task_struct *p, int cpu);
670 static unsigned long task_h_load(struct task_struct *p);
672 static inline void __update_task_entity_contrib(struct sched_entity *se);
673 static inline void __update_task_entity_utilization(struct sched_entity *se);
675 /* Give new task start runnable values to heavy its load in infant time */
676 void init_task_runnable_average(struct task_struct *p)
680 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
681 p->se.avg.runnable_avg_sum = p->se.avg.running_avg_sum = slice;
682 p->se.avg.avg_period = slice;
683 __update_task_entity_contrib(&p->se);
684 __update_task_entity_utilization(&p->se);
687 void init_task_runnable_average(struct task_struct *p)
693 * Update the current task's runtime statistics.
695 static void update_curr(struct cfs_rq *cfs_rq)
697 struct sched_entity *curr = cfs_rq->curr;
698 u64 now = rq_clock_task(rq_of(cfs_rq));
704 delta_exec = now - curr->exec_start;
705 if (unlikely((s64)delta_exec <= 0))
708 curr->exec_start = now;
710 schedstat_set(curr->statistics.exec_max,
711 max(delta_exec, curr->statistics.exec_max));
713 curr->sum_exec_runtime += delta_exec;
714 schedstat_add(cfs_rq, exec_clock, delta_exec);
716 curr->vruntime += calc_delta_fair(delta_exec, curr);
717 update_min_vruntime(cfs_rq);
719 if (entity_is_task(curr)) {
720 struct task_struct *curtask = task_of(curr);
722 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
723 cpuacct_charge(curtask, delta_exec);
724 account_group_exec_runtime(curtask, delta_exec);
727 account_cfs_rq_runtime(cfs_rq, delta_exec);
730 static void update_curr_fair(struct rq *rq)
732 update_curr(cfs_rq_of(&rq->curr->se));
736 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
738 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
742 * Task is being enqueued - update stats:
744 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
747 * Are we enqueueing a waiting task? (for current tasks
748 * a dequeue/enqueue event is a NOP)
750 if (se != cfs_rq->curr)
751 update_stats_wait_start(cfs_rq, se);
755 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
757 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
758 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
759 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
760 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
761 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
762 #ifdef CONFIG_SCHEDSTATS
763 if (entity_is_task(se)) {
764 trace_sched_stat_wait(task_of(se),
765 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
768 schedstat_set(se->statistics.wait_start, 0);
772 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
775 * Mark the end of the wait period if dequeueing a
778 if (se != cfs_rq->curr)
779 update_stats_wait_end(cfs_rq, se);
783 * We are picking a new current task - update its stats:
786 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
789 * We are starting a new run period:
791 se->exec_start = rq_clock_task(rq_of(cfs_rq));
794 /**************************************************
795 * Scheduling class queueing methods:
798 #ifdef CONFIG_NUMA_BALANCING
800 * Approximate time to scan a full NUMA task in ms. The task scan period is
801 * calculated based on the tasks virtual memory size and
802 * numa_balancing_scan_size.
804 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
805 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
807 /* Portion of address space to scan in MB */
808 unsigned int sysctl_numa_balancing_scan_size = 256;
810 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
811 unsigned int sysctl_numa_balancing_scan_delay = 1000;
813 static unsigned int task_nr_scan_windows(struct task_struct *p)
815 unsigned long rss = 0;
816 unsigned long nr_scan_pages;
819 * Calculations based on RSS as non-present and empty pages are skipped
820 * by the PTE scanner and NUMA hinting faults should be trapped based
823 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
824 rss = get_mm_rss(p->mm);
828 rss = round_up(rss, nr_scan_pages);
829 return rss / nr_scan_pages;
832 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
833 #define MAX_SCAN_WINDOW 2560
835 static unsigned int task_scan_min(struct task_struct *p)
837 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
838 unsigned int scan, floor;
839 unsigned int windows = 1;
841 if (scan_size < MAX_SCAN_WINDOW)
842 windows = MAX_SCAN_WINDOW / scan_size;
843 floor = 1000 / windows;
845 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
846 return max_t(unsigned int, floor, scan);
849 static unsigned int task_scan_max(struct task_struct *p)
851 unsigned int smin = task_scan_min(p);
854 /* Watch for min being lower than max due to floor calculations */
855 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
856 return max(smin, smax);
859 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
861 rq->nr_numa_running += (p->numa_preferred_nid != -1);
862 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
865 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
867 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
868 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
874 spinlock_t lock; /* nr_tasks, tasks */
879 nodemask_t active_nodes;
880 unsigned long total_faults;
882 * Faults_cpu is used to decide whether memory should move
883 * towards the CPU. As a consequence, these stats are weighted
884 * more by CPU use than by memory faults.
886 unsigned long *faults_cpu;
887 unsigned long faults[0];
890 /* Shared or private faults. */
891 #define NR_NUMA_HINT_FAULT_TYPES 2
893 /* Memory and CPU locality */
894 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
896 /* Averaged statistics, and temporary buffers. */
897 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
899 pid_t task_numa_group_id(struct task_struct *p)
901 return p->numa_group ? p->numa_group->gid : 0;
905 * The averaged statistics, shared & private, memory & cpu,
906 * occupy the first half of the array. The second half of the
907 * array is for current counters, which are averaged into the
908 * first set by task_numa_placement.
910 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
912 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
915 static inline unsigned long task_faults(struct task_struct *p, int nid)
920 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
921 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
924 static inline unsigned long group_faults(struct task_struct *p, int nid)
929 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
930 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
933 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
935 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
936 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
939 /* Handle placement on systems where not all nodes are directly connected. */
940 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
941 int maxdist, bool task)
943 unsigned long score = 0;
947 * All nodes are directly connected, and the same distance
948 * from each other. No need for fancy placement algorithms.
950 if (sched_numa_topology_type == NUMA_DIRECT)
954 * This code is called for each node, introducing N^2 complexity,
955 * which should be ok given the number of nodes rarely exceeds 8.
957 for_each_online_node(node) {
958 unsigned long faults;
959 int dist = node_distance(nid, node);
962 * The furthest away nodes in the system are not interesting
963 * for placement; nid was already counted.
965 if (dist == sched_max_numa_distance || node == nid)
969 * On systems with a backplane NUMA topology, compare groups
970 * of nodes, and move tasks towards the group with the most
971 * memory accesses. When comparing two nodes at distance
972 * "hoplimit", only nodes closer by than "hoplimit" are part
973 * of each group. Skip other nodes.
975 if (sched_numa_topology_type == NUMA_BACKPLANE &&
979 /* Add up the faults from nearby nodes. */
981 faults = task_faults(p, node);
983 faults = group_faults(p, node);
986 * On systems with a glueless mesh NUMA topology, there are
987 * no fixed "groups of nodes". Instead, nodes that are not
988 * directly connected bounce traffic through intermediate
989 * nodes; a numa_group can occupy any set of nodes.
990 * The further away a node is, the less the faults count.
991 * This seems to result in good task placement.
993 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
994 faults *= (sched_max_numa_distance - dist);
995 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1005 * These return the fraction of accesses done by a particular task, or
1006 * task group, on a particular numa node. The group weight is given a
1007 * larger multiplier, in order to group tasks together that are almost
1008 * evenly spread out between numa nodes.
1010 static inline unsigned long task_weight(struct task_struct *p, int nid,
1013 unsigned long faults, total_faults;
1015 if (!p->numa_faults)
1018 total_faults = p->total_numa_faults;
1023 faults = task_faults(p, nid);
1024 faults += score_nearby_nodes(p, nid, dist, true);
1026 return 1000 * faults / total_faults;
1029 static inline unsigned long group_weight(struct task_struct *p, int nid,
1032 unsigned long faults, total_faults;
1037 total_faults = p->numa_group->total_faults;
1042 faults = group_faults(p, nid);
1043 faults += score_nearby_nodes(p, nid, dist, false);
1045 return 1000 * faults / total_faults;
1048 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1049 int src_nid, int dst_cpu)
1051 struct numa_group *ng = p->numa_group;
1052 int dst_nid = cpu_to_node(dst_cpu);
1053 int last_cpupid, this_cpupid;
1055 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1058 * Multi-stage node selection is used in conjunction with a periodic
1059 * migration fault to build a temporal task<->page relation. By using
1060 * a two-stage filter we remove short/unlikely relations.
1062 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1063 * a task's usage of a particular page (n_p) per total usage of this
1064 * page (n_t) (in a given time-span) to a probability.
1066 * Our periodic faults will sample this probability and getting the
1067 * same result twice in a row, given these samples are fully
1068 * independent, is then given by P(n)^2, provided our sample period
1069 * is sufficiently short compared to the usage pattern.
1071 * This quadric squishes small probabilities, making it less likely we
1072 * act on an unlikely task<->page relation.
1074 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1075 if (!cpupid_pid_unset(last_cpupid) &&
1076 cpupid_to_nid(last_cpupid) != dst_nid)
1079 /* Always allow migrate on private faults */
1080 if (cpupid_match_pid(p, last_cpupid))
1083 /* A shared fault, but p->numa_group has not been set up yet. */
1088 * Do not migrate if the destination is not a node that
1089 * is actively used by this numa group.
1091 if (!node_isset(dst_nid, ng->active_nodes))
1095 * Source is a node that is not actively used by this
1096 * numa group, while the destination is. Migrate.
1098 if (!node_isset(src_nid, ng->active_nodes))
1102 * Both source and destination are nodes in active
1103 * use by this numa group. Maximize memory bandwidth
1104 * by migrating from more heavily used groups, to less
1105 * heavily used ones, spreading the load around.
1106 * Use a 1/4 hysteresis to avoid spurious page movement.
1108 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1111 static unsigned long weighted_cpuload(const int cpu);
1112 static unsigned long source_load(int cpu, int type);
1113 static unsigned long target_load(int cpu, int type);
1114 static unsigned long capacity_of(int cpu);
1115 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1117 /* Cached statistics for all CPUs within a node */
1119 unsigned long nr_running;
1122 /* Total compute capacity of CPUs on a node */
1123 unsigned long compute_capacity;
1125 /* Approximate capacity in terms of runnable tasks on a node */
1126 unsigned long task_capacity;
1127 int has_free_capacity;
1131 * XXX borrowed from update_sg_lb_stats
1133 static void update_numa_stats(struct numa_stats *ns, int nid)
1135 int smt, cpu, cpus = 0;
1136 unsigned long capacity;
1138 memset(ns, 0, sizeof(*ns));
1139 for_each_cpu(cpu, cpumask_of_node(nid)) {
1140 struct rq *rq = cpu_rq(cpu);
1142 ns->nr_running += rq->nr_running;
1143 ns->load += weighted_cpuload(cpu);
1144 ns->compute_capacity += capacity_of(cpu);
1150 * If we raced with hotplug and there are no CPUs left in our mask
1151 * the @ns structure is NULL'ed and task_numa_compare() will
1152 * not find this node attractive.
1154 * We'll either bail at !has_free_capacity, or we'll detect a huge
1155 * imbalance and bail there.
1160 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1161 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1162 capacity = cpus / smt; /* cores */
1164 ns->task_capacity = min_t(unsigned, capacity,
1165 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1166 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1169 struct task_numa_env {
1170 struct task_struct *p;
1172 int src_cpu, src_nid;
1173 int dst_cpu, dst_nid;
1175 struct numa_stats src_stats, dst_stats;
1180 struct task_struct *best_task;
1185 static void task_numa_assign(struct task_numa_env *env,
1186 struct task_struct *p, long imp)
1189 put_task_struct(env->best_task);
1194 env->best_imp = imp;
1195 env->best_cpu = env->dst_cpu;
1198 static bool load_too_imbalanced(long src_load, long dst_load,
1199 struct task_numa_env *env)
1202 long orig_src_load, orig_dst_load;
1203 long src_capacity, dst_capacity;
1206 * The load is corrected for the CPU capacity available on each node.
1209 * ------------ vs ---------
1210 * src_capacity dst_capacity
1212 src_capacity = env->src_stats.compute_capacity;
1213 dst_capacity = env->dst_stats.compute_capacity;
1215 /* We care about the slope of the imbalance, not the direction. */
1216 if (dst_load < src_load)
1217 swap(dst_load, src_load);
1219 /* Is the difference below the threshold? */
1220 imb = dst_load * src_capacity * 100 -
1221 src_load * dst_capacity * env->imbalance_pct;
1226 * The imbalance is above the allowed threshold.
1227 * Compare it with the old imbalance.
1229 orig_src_load = env->src_stats.load;
1230 orig_dst_load = env->dst_stats.load;
1232 if (orig_dst_load < orig_src_load)
1233 swap(orig_dst_load, orig_src_load);
1235 old_imb = orig_dst_load * src_capacity * 100 -
1236 orig_src_load * dst_capacity * env->imbalance_pct;
1238 /* Would this change make things worse? */
1239 return (imb > old_imb);
1243 * This checks if the overall compute and NUMA accesses of the system would
1244 * be improved if the source tasks was migrated to the target dst_cpu taking
1245 * into account that it might be best if task running on the dst_cpu should
1246 * be exchanged with the source task
1248 static void task_numa_compare(struct task_numa_env *env,
1249 long taskimp, long groupimp)
1251 struct rq *src_rq = cpu_rq(env->src_cpu);
1252 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1253 struct task_struct *cur;
1254 long src_load, dst_load;
1256 long imp = env->p->numa_group ? groupimp : taskimp;
1258 int dist = env->dist;
1262 raw_spin_lock_irq(&dst_rq->lock);
1265 * No need to move the exiting task, and this ensures that ->curr
1266 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1267 * is safe under RCU read lock.
1268 * Note that rcu_read_lock() itself can't protect from the final
1269 * put_task_struct() after the last schedule().
1271 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1273 raw_spin_unlock_irq(&dst_rq->lock);
1276 * Because we have preemption enabled we can get migrated around and
1277 * end try selecting ourselves (current == env->p) as a swap candidate.
1283 * "imp" is the fault differential for the source task between the
1284 * source and destination node. Calculate the total differential for
1285 * the source task and potential destination task. The more negative
1286 * the value is, the more rmeote accesses that would be expected to
1287 * be incurred if the tasks were swapped.
1290 /* Skip this swap candidate if cannot move to the source cpu */
1291 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1295 * If dst and source tasks are in the same NUMA group, or not
1296 * in any group then look only at task weights.
1298 if (cur->numa_group == env->p->numa_group) {
1299 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1300 task_weight(cur, env->dst_nid, dist);
1302 * Add some hysteresis to prevent swapping the
1303 * tasks within a group over tiny differences.
1305 if (cur->numa_group)
1309 * Compare the group weights. If a task is all by
1310 * itself (not part of a group), use the task weight
1313 if (cur->numa_group)
1314 imp += group_weight(cur, env->src_nid, dist) -
1315 group_weight(cur, env->dst_nid, dist);
1317 imp += task_weight(cur, env->src_nid, dist) -
1318 task_weight(cur, env->dst_nid, dist);
1322 if (imp <= env->best_imp && moveimp <= env->best_imp)
1326 /* Is there capacity at our destination? */
1327 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1328 !env->dst_stats.has_free_capacity)
1334 /* Balance doesn't matter much if we're running a task per cpu */
1335 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1336 dst_rq->nr_running == 1)
1340 * In the overloaded case, try and keep the load balanced.
1343 load = task_h_load(env->p);
1344 dst_load = env->dst_stats.load + load;
1345 src_load = env->src_stats.load - load;
1347 if (moveimp > imp && moveimp > env->best_imp) {
1349 * If the improvement from just moving env->p direction is
1350 * better than swapping tasks around, check if a move is
1351 * possible. Store a slightly smaller score than moveimp,
1352 * so an actually idle CPU will win.
1354 if (!load_too_imbalanced(src_load, dst_load, env)) {
1361 if (imp <= env->best_imp)
1365 load = task_h_load(cur);
1370 if (load_too_imbalanced(src_load, dst_load, env))
1374 * One idle CPU per node is evaluated for a task numa move.
1375 * Call select_idle_sibling to maybe find a better one.
1378 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1381 task_numa_assign(env, cur, imp);
1386 static void task_numa_find_cpu(struct task_numa_env *env,
1387 long taskimp, long groupimp)
1391 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1392 /* Skip this CPU if the source task cannot migrate */
1393 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1397 task_numa_compare(env, taskimp, groupimp);
1401 /* Only move tasks to a NUMA node less busy than the current node. */
1402 static bool numa_has_capacity(struct task_numa_env *env)
1404 struct numa_stats *src = &env->src_stats;
1405 struct numa_stats *dst = &env->dst_stats;
1407 if (src->has_free_capacity && !dst->has_free_capacity)
1411 * Only consider a task move if the source has a higher load
1412 * than the destination, corrected for CPU capacity on each node.
1414 * src->load dst->load
1415 * --------------------- vs ---------------------
1416 * src->compute_capacity dst->compute_capacity
1418 if (src->load * dst->compute_capacity >
1419 dst->load * src->compute_capacity)
1425 static int task_numa_migrate(struct task_struct *p)
1427 struct task_numa_env env = {
1430 .src_cpu = task_cpu(p),
1431 .src_nid = task_node(p),
1433 .imbalance_pct = 112,
1439 struct sched_domain *sd;
1440 unsigned long taskweight, groupweight;
1442 long taskimp, groupimp;
1445 * Pick the lowest SD_NUMA domain, as that would have the smallest
1446 * imbalance and would be the first to start moving tasks about.
1448 * And we want to avoid any moving of tasks about, as that would create
1449 * random movement of tasks -- counter the numa conditions we're trying
1453 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1455 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1459 * Cpusets can break the scheduler domain tree into smaller
1460 * balance domains, some of which do not cross NUMA boundaries.
1461 * Tasks that are "trapped" in such domains cannot be migrated
1462 * elsewhere, so there is no point in (re)trying.
1464 if (unlikely(!sd)) {
1465 p->numa_preferred_nid = task_node(p);
1469 env.dst_nid = p->numa_preferred_nid;
1470 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1471 taskweight = task_weight(p, env.src_nid, dist);
1472 groupweight = group_weight(p, env.src_nid, dist);
1473 update_numa_stats(&env.src_stats, env.src_nid);
1474 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1475 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1476 update_numa_stats(&env.dst_stats, env.dst_nid);
1478 /* Try to find a spot on the preferred nid. */
1479 if (numa_has_capacity(&env))
1480 task_numa_find_cpu(&env, taskimp, groupimp);
1483 * Look at other nodes in these cases:
1484 * - there is no space available on the preferred_nid
1485 * - the task is part of a numa_group that is interleaved across
1486 * multiple NUMA nodes; in order to better consolidate the group,
1487 * we need to check other locations.
1489 if (env.best_cpu == -1 || (p->numa_group &&
1490 nodes_weight(p->numa_group->active_nodes) > 1)) {
1491 for_each_online_node(nid) {
1492 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1495 dist = node_distance(env.src_nid, env.dst_nid);
1496 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1498 taskweight = task_weight(p, env.src_nid, dist);
1499 groupweight = group_weight(p, env.src_nid, dist);
1502 /* Only consider nodes where both task and groups benefit */
1503 taskimp = task_weight(p, nid, dist) - taskweight;
1504 groupimp = group_weight(p, nid, dist) - groupweight;
1505 if (taskimp < 0 && groupimp < 0)
1510 update_numa_stats(&env.dst_stats, env.dst_nid);
1511 if (numa_has_capacity(&env))
1512 task_numa_find_cpu(&env, taskimp, groupimp);
1517 * If the task is part of a workload that spans multiple NUMA nodes,
1518 * and is migrating into one of the workload's active nodes, remember
1519 * this node as the task's preferred numa node, so the workload can
1521 * A task that migrated to a second choice node will be better off
1522 * trying for a better one later. Do not set the preferred node here.
1524 if (p->numa_group) {
1525 if (env.best_cpu == -1)
1530 if (node_isset(nid, p->numa_group->active_nodes))
1531 sched_setnuma(p, env.dst_nid);
1534 /* No better CPU than the current one was found. */
1535 if (env.best_cpu == -1)
1539 * Reset the scan period if the task is being rescheduled on an
1540 * alternative node to recheck if the tasks is now properly placed.
1542 p->numa_scan_period = task_scan_min(p);
1544 if (env.best_task == NULL) {
1545 ret = migrate_task_to(p, env.best_cpu);
1547 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1551 ret = migrate_swap(p, env.best_task);
1553 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1554 put_task_struct(env.best_task);
1558 /* Attempt to migrate a task to a CPU on the preferred node. */
1559 static void numa_migrate_preferred(struct task_struct *p)
1561 unsigned long interval = HZ;
1563 /* This task has no NUMA fault statistics yet */
1564 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1567 /* Periodically retry migrating the task to the preferred node */
1568 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1569 p->numa_migrate_retry = jiffies + interval;
1571 /* Success if task is already running on preferred CPU */
1572 if (task_node(p) == p->numa_preferred_nid)
1575 /* Otherwise, try migrate to a CPU on the preferred node */
1576 task_numa_migrate(p);
1580 * Find the nodes on which the workload is actively running. We do this by
1581 * tracking the nodes from which NUMA hinting faults are triggered. This can
1582 * be different from the set of nodes where the workload's memory is currently
1585 * The bitmask is used to make smarter decisions on when to do NUMA page
1586 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1587 * are added when they cause over 6/16 of the maximum number of faults, but
1588 * only removed when they drop below 3/16.
1590 static void update_numa_active_node_mask(struct numa_group *numa_group)
1592 unsigned long faults, max_faults = 0;
1595 for_each_online_node(nid) {
1596 faults = group_faults_cpu(numa_group, nid);
1597 if (faults > max_faults)
1598 max_faults = faults;
1601 for_each_online_node(nid) {
1602 faults = group_faults_cpu(numa_group, nid);
1603 if (!node_isset(nid, numa_group->active_nodes)) {
1604 if (faults > max_faults * 6 / 16)
1605 node_set(nid, numa_group->active_nodes);
1606 } else if (faults < max_faults * 3 / 16)
1607 node_clear(nid, numa_group->active_nodes);
1612 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1613 * increments. The more local the fault statistics are, the higher the scan
1614 * period will be for the next scan window. If local/(local+remote) ratio is
1615 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1616 * the scan period will decrease. Aim for 70% local accesses.
1618 #define NUMA_PERIOD_SLOTS 10
1619 #define NUMA_PERIOD_THRESHOLD 7
1622 * Increase the scan period (slow down scanning) if the majority of
1623 * our memory is already on our local node, or if the majority of
1624 * the page accesses are shared with other processes.
1625 * Otherwise, decrease the scan period.
1627 static void update_task_scan_period(struct task_struct *p,
1628 unsigned long shared, unsigned long private)
1630 unsigned int period_slot;
1634 unsigned long remote = p->numa_faults_locality[0];
1635 unsigned long local = p->numa_faults_locality[1];
1638 * If there were no record hinting faults then either the task is
1639 * completely idle or all activity is areas that are not of interest
1640 * to automatic numa balancing. Related to that, if there were failed
1641 * migration then it implies we are migrating too quickly or the local
1642 * node is overloaded. In either case, scan slower
1644 if (local + shared == 0 || p->numa_faults_locality[2]) {
1645 p->numa_scan_period = min(p->numa_scan_period_max,
1646 p->numa_scan_period << 1);
1648 p->mm->numa_next_scan = jiffies +
1649 msecs_to_jiffies(p->numa_scan_period);
1655 * Prepare to scale scan period relative to the current period.
1656 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1657 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1658 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1660 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1661 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1662 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1663 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1666 diff = slot * period_slot;
1668 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1671 * Scale scan rate increases based on sharing. There is an
1672 * inverse relationship between the degree of sharing and
1673 * the adjustment made to the scanning period. Broadly
1674 * speaking the intent is that there is little point
1675 * scanning faster if shared accesses dominate as it may
1676 * simply bounce migrations uselessly
1678 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1679 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1682 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1683 task_scan_min(p), task_scan_max(p));
1684 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1688 * Get the fraction of time the task has been running since the last
1689 * NUMA placement cycle. The scheduler keeps similar statistics, but
1690 * decays those on a 32ms period, which is orders of magnitude off
1691 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1692 * stats only if the task is so new there are no NUMA statistics yet.
1694 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1696 u64 runtime, delta, now;
1697 /* Use the start of this time slice to avoid calculations. */
1698 now = p->se.exec_start;
1699 runtime = p->se.sum_exec_runtime;
1701 if (p->last_task_numa_placement) {
1702 delta = runtime - p->last_sum_exec_runtime;
1703 *period = now - p->last_task_numa_placement;
1705 delta = p->se.avg.runnable_avg_sum;
1706 *period = p->se.avg.avg_period;
1709 p->last_sum_exec_runtime = runtime;
1710 p->last_task_numa_placement = now;
1716 * Determine the preferred nid for a task in a numa_group. This needs to
1717 * be done in a way that produces consistent results with group_weight,
1718 * otherwise workloads might not converge.
1720 static int preferred_group_nid(struct task_struct *p, int nid)
1725 /* Direct connections between all NUMA nodes. */
1726 if (sched_numa_topology_type == NUMA_DIRECT)
1730 * On a system with glueless mesh NUMA topology, group_weight
1731 * scores nodes according to the number of NUMA hinting faults on
1732 * both the node itself, and on nearby nodes.
1734 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1735 unsigned long score, max_score = 0;
1736 int node, max_node = nid;
1738 dist = sched_max_numa_distance;
1740 for_each_online_node(node) {
1741 score = group_weight(p, node, dist);
1742 if (score > max_score) {
1751 * Finding the preferred nid in a system with NUMA backplane
1752 * interconnect topology is more involved. The goal is to locate
1753 * tasks from numa_groups near each other in the system, and
1754 * untangle workloads from different sides of the system. This requires
1755 * searching down the hierarchy of node groups, recursively searching
1756 * inside the highest scoring group of nodes. The nodemask tricks
1757 * keep the complexity of the search down.
1759 nodes = node_online_map;
1760 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1761 unsigned long max_faults = 0;
1762 nodemask_t max_group = NODE_MASK_NONE;
1765 /* Are there nodes at this distance from each other? */
1766 if (!find_numa_distance(dist))
1769 for_each_node_mask(a, nodes) {
1770 unsigned long faults = 0;
1771 nodemask_t this_group;
1772 nodes_clear(this_group);
1774 /* Sum group's NUMA faults; includes a==b case. */
1775 for_each_node_mask(b, nodes) {
1776 if (node_distance(a, b) < dist) {
1777 faults += group_faults(p, b);
1778 node_set(b, this_group);
1779 node_clear(b, nodes);
1783 /* Remember the top group. */
1784 if (faults > max_faults) {
1785 max_faults = faults;
1786 max_group = this_group;
1788 * subtle: at the smallest distance there is
1789 * just one node left in each "group", the
1790 * winner is the preferred nid.
1795 /* Next round, evaluate the nodes within max_group. */
1803 static void task_numa_placement(struct task_struct *p)
1805 int seq, nid, max_nid = -1, max_group_nid = -1;
1806 unsigned long max_faults = 0, max_group_faults = 0;
1807 unsigned long fault_types[2] = { 0, 0 };
1808 unsigned long total_faults;
1809 u64 runtime, period;
1810 spinlock_t *group_lock = NULL;
1813 * The p->mm->numa_scan_seq field gets updated without
1814 * exclusive access. Use READ_ONCE() here to ensure
1815 * that the field is read in a single access:
1817 seq = READ_ONCE(p->mm->numa_scan_seq);
1818 if (p->numa_scan_seq == seq)
1820 p->numa_scan_seq = seq;
1821 p->numa_scan_period_max = task_scan_max(p);
1823 total_faults = p->numa_faults_locality[0] +
1824 p->numa_faults_locality[1];
1825 runtime = numa_get_avg_runtime(p, &period);
1827 /* If the task is part of a group prevent parallel updates to group stats */
1828 if (p->numa_group) {
1829 group_lock = &p->numa_group->lock;
1830 spin_lock_irq(group_lock);
1833 /* Find the node with the highest number of faults */
1834 for_each_online_node(nid) {
1835 /* Keep track of the offsets in numa_faults array */
1836 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1837 unsigned long faults = 0, group_faults = 0;
1840 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1841 long diff, f_diff, f_weight;
1843 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1844 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1845 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1846 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1848 /* Decay existing window, copy faults since last scan */
1849 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1850 fault_types[priv] += p->numa_faults[membuf_idx];
1851 p->numa_faults[membuf_idx] = 0;
1854 * Normalize the faults_from, so all tasks in a group
1855 * count according to CPU use, instead of by the raw
1856 * number of faults. Tasks with little runtime have
1857 * little over-all impact on throughput, and thus their
1858 * faults are less important.
1860 f_weight = div64_u64(runtime << 16, period + 1);
1861 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1863 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1864 p->numa_faults[cpubuf_idx] = 0;
1866 p->numa_faults[mem_idx] += diff;
1867 p->numa_faults[cpu_idx] += f_diff;
1868 faults += p->numa_faults[mem_idx];
1869 p->total_numa_faults += diff;
1870 if (p->numa_group) {
1872 * safe because we can only change our own group
1874 * mem_idx represents the offset for a given
1875 * nid and priv in a specific region because it
1876 * is at the beginning of the numa_faults array.
1878 p->numa_group->faults[mem_idx] += diff;
1879 p->numa_group->faults_cpu[mem_idx] += f_diff;
1880 p->numa_group->total_faults += diff;
1881 group_faults += p->numa_group->faults[mem_idx];
1885 if (faults > max_faults) {
1886 max_faults = faults;
1890 if (group_faults > max_group_faults) {
1891 max_group_faults = group_faults;
1892 max_group_nid = nid;
1896 update_task_scan_period(p, fault_types[0], fault_types[1]);
1898 if (p->numa_group) {
1899 update_numa_active_node_mask(p->numa_group);
1900 spin_unlock_irq(group_lock);
1901 max_nid = preferred_group_nid(p, max_group_nid);
1905 /* Set the new preferred node */
1906 if (max_nid != p->numa_preferred_nid)
1907 sched_setnuma(p, max_nid);
1909 if (task_node(p) != p->numa_preferred_nid)
1910 numa_migrate_preferred(p);
1914 static inline int get_numa_group(struct numa_group *grp)
1916 return atomic_inc_not_zero(&grp->refcount);
1919 static inline void put_numa_group(struct numa_group *grp)
1921 if (atomic_dec_and_test(&grp->refcount))
1922 kfree_rcu(grp, rcu);
1925 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1928 struct numa_group *grp, *my_grp;
1929 struct task_struct *tsk;
1931 int cpu = cpupid_to_cpu(cpupid);
1934 if (unlikely(!p->numa_group)) {
1935 unsigned int size = sizeof(struct numa_group) +
1936 4*nr_node_ids*sizeof(unsigned long);
1938 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1942 atomic_set(&grp->refcount, 1);
1943 spin_lock_init(&grp->lock);
1945 /* Second half of the array tracks nids where faults happen */
1946 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1949 node_set(task_node(current), grp->active_nodes);
1951 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1952 grp->faults[i] = p->numa_faults[i];
1954 grp->total_faults = p->total_numa_faults;
1957 rcu_assign_pointer(p->numa_group, grp);
1961 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1963 if (!cpupid_match_pid(tsk, cpupid))
1966 grp = rcu_dereference(tsk->numa_group);
1970 my_grp = p->numa_group;
1975 * Only join the other group if its bigger; if we're the bigger group,
1976 * the other task will join us.
1978 if (my_grp->nr_tasks > grp->nr_tasks)
1982 * Tie-break on the grp address.
1984 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1987 /* Always join threads in the same process. */
1988 if (tsk->mm == current->mm)
1991 /* Simple filter to avoid false positives due to PID collisions */
1992 if (flags & TNF_SHARED)
1995 /* Update priv based on whether false sharing was detected */
1998 if (join && !get_numa_group(grp))
2006 BUG_ON(irqs_disabled());
2007 double_lock_irq(&my_grp->lock, &grp->lock);
2009 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2010 my_grp->faults[i] -= p->numa_faults[i];
2011 grp->faults[i] += p->numa_faults[i];
2013 my_grp->total_faults -= p->total_numa_faults;
2014 grp->total_faults += p->total_numa_faults;
2019 spin_unlock(&my_grp->lock);
2020 spin_unlock_irq(&grp->lock);
2022 rcu_assign_pointer(p->numa_group, grp);
2024 put_numa_group(my_grp);
2032 void task_numa_free(struct task_struct *p)
2034 struct numa_group *grp = p->numa_group;
2035 void *numa_faults = p->numa_faults;
2036 unsigned long flags;
2040 spin_lock_irqsave(&grp->lock, flags);
2041 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2042 grp->faults[i] -= p->numa_faults[i];
2043 grp->total_faults -= p->total_numa_faults;
2046 spin_unlock_irqrestore(&grp->lock, flags);
2047 RCU_INIT_POINTER(p->numa_group, NULL);
2048 put_numa_group(grp);
2051 p->numa_faults = NULL;
2056 * Got a PROT_NONE fault for a page on @node.
2058 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2060 struct task_struct *p = current;
2061 bool migrated = flags & TNF_MIGRATED;
2062 int cpu_node = task_node(current);
2063 int local = !!(flags & TNF_FAULT_LOCAL);
2066 if (!numabalancing_enabled)
2069 /* for example, ksmd faulting in a user's mm */
2073 /* Allocate buffer to track faults on a per-node basis */
2074 if (unlikely(!p->numa_faults)) {
2075 int size = sizeof(*p->numa_faults) *
2076 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2078 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2079 if (!p->numa_faults)
2082 p->total_numa_faults = 0;
2083 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2087 * First accesses are treated as private, otherwise consider accesses
2088 * to be private if the accessing pid has not changed
2090 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2093 priv = cpupid_match_pid(p, last_cpupid);
2094 if (!priv && !(flags & TNF_NO_GROUP))
2095 task_numa_group(p, last_cpupid, flags, &priv);
2099 * If a workload spans multiple NUMA nodes, a shared fault that
2100 * occurs wholly within the set of nodes that the workload is
2101 * actively using should be counted as local. This allows the
2102 * scan rate to slow down when a workload has settled down.
2104 if (!priv && !local && p->numa_group &&
2105 node_isset(cpu_node, p->numa_group->active_nodes) &&
2106 node_isset(mem_node, p->numa_group->active_nodes))
2109 task_numa_placement(p);
2112 * Retry task to preferred node migration periodically, in case it
2113 * case it previously failed, or the scheduler moved us.
2115 if (time_after(jiffies, p->numa_migrate_retry))
2116 numa_migrate_preferred(p);
2119 p->numa_pages_migrated += pages;
2120 if (flags & TNF_MIGRATE_FAIL)
2121 p->numa_faults_locality[2] += pages;
2123 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2124 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2125 p->numa_faults_locality[local] += pages;
2128 static void reset_ptenuma_scan(struct task_struct *p)
2131 * We only did a read acquisition of the mmap sem, so
2132 * p->mm->numa_scan_seq is written to without exclusive access
2133 * and the update is not guaranteed to be atomic. That's not
2134 * much of an issue though, since this is just used for
2135 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2136 * expensive, to avoid any form of compiler optimizations:
2138 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2139 p->mm->numa_scan_offset = 0;
2143 * The expensive part of numa migration is done from task_work context.
2144 * Triggered from task_tick_numa().
2146 void task_numa_work(struct callback_head *work)
2148 unsigned long migrate, next_scan, now = jiffies;
2149 struct task_struct *p = current;
2150 struct mm_struct *mm = p->mm;
2151 struct vm_area_struct *vma;
2152 unsigned long start, end;
2153 unsigned long nr_pte_updates = 0;
2156 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2158 work->next = work; /* protect against double add */
2160 * Who cares about NUMA placement when they're dying.
2162 * NOTE: make sure not to dereference p->mm before this check,
2163 * exit_task_work() happens _after_ exit_mm() so we could be called
2164 * without p->mm even though we still had it when we enqueued this
2167 if (p->flags & PF_EXITING)
2170 if (!mm->numa_next_scan) {
2171 mm->numa_next_scan = now +
2172 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2176 * Enforce maximal scan/migration frequency..
2178 migrate = mm->numa_next_scan;
2179 if (time_before(now, migrate))
2182 if (p->numa_scan_period == 0) {
2183 p->numa_scan_period_max = task_scan_max(p);
2184 p->numa_scan_period = task_scan_min(p);
2187 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2188 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2192 * Delay this task enough that another task of this mm will likely win
2193 * the next time around.
2195 p->node_stamp += 2 * TICK_NSEC;
2197 start = mm->numa_scan_offset;
2198 pages = sysctl_numa_balancing_scan_size;
2199 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2203 down_read(&mm->mmap_sem);
2204 vma = find_vma(mm, start);
2206 reset_ptenuma_scan(p);
2210 for (; vma; vma = vma->vm_next) {
2211 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2212 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2217 * Shared library pages mapped by multiple processes are not
2218 * migrated as it is expected they are cache replicated. Avoid
2219 * hinting faults in read-only file-backed mappings or the vdso
2220 * as migrating the pages will be of marginal benefit.
2223 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2227 * Skip inaccessible VMAs to avoid any confusion between
2228 * PROT_NONE and NUMA hinting ptes
2230 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2234 start = max(start, vma->vm_start);
2235 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2236 end = min(end, vma->vm_end);
2237 nr_pte_updates += change_prot_numa(vma, start, end);
2240 * Scan sysctl_numa_balancing_scan_size but ensure that
2241 * at least one PTE is updated so that unused virtual
2242 * address space is quickly skipped.
2245 pages -= (end - start) >> PAGE_SHIFT;
2252 } while (end != vma->vm_end);
2257 * It is possible to reach the end of the VMA list but the last few
2258 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2259 * would find the !migratable VMA on the next scan but not reset the
2260 * scanner to the start so check it now.
2263 mm->numa_scan_offset = start;
2265 reset_ptenuma_scan(p);
2266 up_read(&mm->mmap_sem);
2270 * Drive the periodic memory faults..
2272 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2274 struct callback_head *work = &curr->numa_work;
2278 * We don't care about NUMA placement if we don't have memory.
2280 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2284 * Using runtime rather than walltime has the dual advantage that
2285 * we (mostly) drive the selection from busy threads and that the
2286 * task needs to have done some actual work before we bother with
2289 now = curr->se.sum_exec_runtime;
2290 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2292 if (now - curr->node_stamp > period) {
2293 if (!curr->node_stamp)
2294 curr->numa_scan_period = task_scan_min(curr);
2295 curr->node_stamp += period;
2297 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2298 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2299 task_work_add(curr, work, true);
2304 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2308 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2312 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2315 #endif /* CONFIG_NUMA_BALANCING */
2318 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2320 update_load_add(&cfs_rq->load, se->load.weight);
2321 if (!parent_entity(se))
2322 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2324 if (entity_is_task(se)) {
2325 struct rq *rq = rq_of(cfs_rq);
2327 account_numa_enqueue(rq, task_of(se));
2328 list_add(&se->group_node, &rq->cfs_tasks);
2331 cfs_rq->nr_running++;
2335 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2337 update_load_sub(&cfs_rq->load, se->load.weight);
2338 if (!parent_entity(se))
2339 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2340 if (entity_is_task(se)) {
2341 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2342 list_del_init(&se->group_node);
2344 cfs_rq->nr_running--;
2347 #ifdef CONFIG_FAIR_GROUP_SCHED
2349 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2354 * Use this CPU's actual weight instead of the last load_contribution
2355 * to gain a more accurate current total weight. See
2356 * update_cfs_rq_load_contribution().
2358 tg_weight = atomic_long_read(&tg->load_avg);
2359 tg_weight -= cfs_rq->tg_load_contrib;
2360 tg_weight += cfs_rq->load.weight;
2365 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2367 long tg_weight, load, shares;
2369 tg_weight = calc_tg_weight(tg, cfs_rq);
2370 load = cfs_rq->load.weight;
2372 shares = (tg->shares * load);
2374 shares /= tg_weight;
2376 if (shares < MIN_SHARES)
2377 shares = MIN_SHARES;
2378 if (shares > tg->shares)
2379 shares = tg->shares;
2383 # else /* CONFIG_SMP */
2384 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2388 # endif /* CONFIG_SMP */
2389 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2390 unsigned long weight)
2393 /* commit outstanding execution time */
2394 if (cfs_rq->curr == se)
2395 update_curr(cfs_rq);
2396 account_entity_dequeue(cfs_rq, se);
2399 update_load_set(&se->load, weight);
2402 account_entity_enqueue(cfs_rq, se);
2405 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2407 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2409 struct task_group *tg;
2410 struct sched_entity *se;
2414 se = tg->se[cpu_of(rq_of(cfs_rq))];
2415 if (!se || throttled_hierarchy(cfs_rq))
2418 if (likely(se->load.weight == tg->shares))
2421 shares = calc_cfs_shares(cfs_rq, tg);
2423 reweight_entity(cfs_rq_of(se), se, shares);
2425 #else /* CONFIG_FAIR_GROUP_SCHED */
2426 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2429 #endif /* CONFIG_FAIR_GROUP_SCHED */
2433 * We choose a half-life close to 1 scheduling period.
2434 * Note: The tables below are dependent on this value.
2436 #define LOAD_AVG_PERIOD 32
2437 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2438 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2440 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2441 static const u32 runnable_avg_yN_inv[] = {
2442 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2443 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2444 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2445 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2446 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2447 0x85aac367, 0x82cd8698,
2451 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2452 * over-estimates when re-combining.
2454 static const u32 runnable_avg_yN_sum[] = {
2455 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2456 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2457 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2462 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2464 static __always_inline u64 decay_load(u64 val, u64 n)
2466 unsigned int local_n;
2470 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2473 /* after bounds checking we can collapse to 32-bit */
2477 * As y^PERIOD = 1/2, we can combine
2478 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2479 * With a look-up table which covers y^n (n<PERIOD)
2481 * To achieve constant time decay_load.
2483 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2484 val >>= local_n / LOAD_AVG_PERIOD;
2485 local_n %= LOAD_AVG_PERIOD;
2488 val *= runnable_avg_yN_inv[local_n];
2489 /* We don't use SRR here since we always want to round down. */
2494 * For updates fully spanning n periods, the contribution to runnable
2495 * average will be: \Sum 1024*y^n
2497 * We can compute this reasonably efficiently by combining:
2498 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2500 static u32 __compute_runnable_contrib(u64 n)
2504 if (likely(n <= LOAD_AVG_PERIOD))
2505 return runnable_avg_yN_sum[n];
2506 else if (unlikely(n >= LOAD_AVG_MAX_N))
2507 return LOAD_AVG_MAX;
2509 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2511 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2512 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2514 n -= LOAD_AVG_PERIOD;
2515 } while (n > LOAD_AVG_PERIOD);
2517 contrib = decay_load(contrib, n);
2518 return contrib + runnable_avg_yN_sum[n];
2522 * We can represent the historical contribution to runnable average as the
2523 * coefficients of a geometric series. To do this we sub-divide our runnable
2524 * history into segments of approximately 1ms (1024us); label the segment that
2525 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2527 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2529 * (now) (~1ms ago) (~2ms ago)
2531 * Let u_i denote the fraction of p_i that the entity was runnable.
2533 * We then designate the fractions u_i as our co-efficients, yielding the
2534 * following representation of historical load:
2535 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2537 * We choose y based on the with of a reasonably scheduling period, fixing:
2540 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2541 * approximately half as much as the contribution to load within the last ms
2544 * When a period "rolls over" and we have new u_0`, multiplying the previous
2545 * sum again by y is sufficient to update:
2546 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2547 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2549 static __always_inline int __update_entity_runnable_avg(u64 now, int cpu,
2550 struct sched_avg *sa,
2555 u32 runnable_contrib;
2556 int delta_w, decayed = 0;
2557 unsigned long scale_freq = arch_scale_freq_capacity(NULL, cpu);
2559 delta = now - sa->last_runnable_update;
2561 * This should only happen when time goes backwards, which it
2562 * unfortunately does during sched clock init when we swap over to TSC.
2564 if ((s64)delta < 0) {
2565 sa->last_runnable_update = now;
2570 * Use 1024ns as the unit of measurement since it's a reasonable
2571 * approximation of 1us and fast to compute.
2576 sa->last_runnable_update = now;
2578 /* delta_w is the amount already accumulated against our next period */
2579 delta_w = sa->avg_period % 1024;
2580 if (delta + delta_w >= 1024) {
2581 /* period roll-over */
2585 * Now that we know we're crossing a period boundary, figure
2586 * out how much from delta we need to complete the current
2587 * period and accrue it.
2589 delta_w = 1024 - delta_w;
2591 sa->runnable_avg_sum += delta_w;
2593 sa->running_avg_sum += delta_w * scale_freq
2594 >> SCHED_CAPACITY_SHIFT;
2595 sa->avg_period += delta_w;
2599 /* Figure out how many additional periods this update spans */
2600 periods = delta / 1024;
2603 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2605 sa->running_avg_sum = decay_load(sa->running_avg_sum,
2607 sa->avg_period = decay_load(sa->avg_period,
2610 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2611 runnable_contrib = __compute_runnable_contrib(periods);
2613 sa->runnable_avg_sum += runnable_contrib;
2615 sa->running_avg_sum += runnable_contrib * scale_freq
2616 >> SCHED_CAPACITY_SHIFT;
2617 sa->avg_period += runnable_contrib;
2620 /* Remainder of delta accrued against u_0` */
2622 sa->runnable_avg_sum += delta;
2624 sa->running_avg_sum += delta * scale_freq
2625 >> SCHED_CAPACITY_SHIFT;
2626 sa->avg_period += delta;
2631 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2632 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2634 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2635 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2637 decays -= se->avg.decay_count;
2638 se->avg.decay_count = 0;
2642 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2643 se->avg.utilization_avg_contrib =
2644 decay_load(se->avg.utilization_avg_contrib, decays);
2649 #ifdef CONFIG_FAIR_GROUP_SCHED
2650 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2653 struct task_group *tg = cfs_rq->tg;
2656 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2657 tg_contrib -= cfs_rq->tg_load_contrib;
2662 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2663 atomic_long_add(tg_contrib, &tg->load_avg);
2664 cfs_rq->tg_load_contrib += tg_contrib;
2669 * Aggregate cfs_rq runnable averages into an equivalent task_group
2670 * representation for computing load contributions.
2672 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2673 struct cfs_rq *cfs_rq)
2675 struct task_group *tg = cfs_rq->tg;
2678 /* The fraction of a cpu used by this cfs_rq */
2679 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2680 sa->avg_period + 1);
2681 contrib -= cfs_rq->tg_runnable_contrib;
2683 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2684 atomic_add(contrib, &tg->runnable_avg);
2685 cfs_rq->tg_runnable_contrib += contrib;
2689 static inline void __update_group_entity_contrib(struct sched_entity *se)
2691 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2692 struct task_group *tg = cfs_rq->tg;
2697 contrib = cfs_rq->tg_load_contrib * tg->shares;
2698 se->avg.load_avg_contrib = div_u64(contrib,
2699 atomic_long_read(&tg->load_avg) + 1);
2702 * For group entities we need to compute a correction term in the case
2703 * that they are consuming <1 cpu so that we would contribute the same
2704 * load as a task of equal weight.
2706 * Explicitly co-ordinating this measurement would be expensive, but
2707 * fortunately the sum of each cpus contribution forms a usable
2708 * lower-bound on the true value.
2710 * Consider the aggregate of 2 contributions. Either they are disjoint
2711 * (and the sum represents true value) or they are disjoint and we are
2712 * understating by the aggregate of their overlap.
2714 * Extending this to N cpus, for a given overlap, the maximum amount we
2715 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2716 * cpus that overlap for this interval and w_i is the interval width.
2718 * On a small machine; the first term is well-bounded which bounds the
2719 * total error since w_i is a subset of the period. Whereas on a
2720 * larger machine, while this first term can be larger, if w_i is the
2721 * of consequential size guaranteed to see n_i*w_i quickly converge to
2722 * our upper bound of 1-cpu.
2724 runnable_avg = atomic_read(&tg->runnable_avg);
2725 if (runnable_avg < NICE_0_LOAD) {
2726 se->avg.load_avg_contrib *= runnable_avg;
2727 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2731 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2733 __update_entity_runnable_avg(rq_clock_task(rq), cpu_of(rq), &rq->avg,
2734 runnable, runnable);
2735 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2737 #else /* CONFIG_FAIR_GROUP_SCHED */
2738 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2739 int force_update) {}
2740 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2741 struct cfs_rq *cfs_rq) {}
2742 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2743 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2744 #endif /* CONFIG_FAIR_GROUP_SCHED */
2746 static inline void __update_task_entity_contrib(struct sched_entity *se)
2750 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2751 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2752 contrib /= (se->avg.avg_period + 1);
2753 se->avg.load_avg_contrib = scale_load(contrib);
2756 /* Compute the current contribution to load_avg by se, return any delta */
2757 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2759 long old_contrib = se->avg.load_avg_contrib;
2761 if (entity_is_task(se)) {
2762 __update_task_entity_contrib(se);
2764 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2765 __update_group_entity_contrib(se);
2768 return se->avg.load_avg_contrib - old_contrib;
2772 static inline void __update_task_entity_utilization(struct sched_entity *se)
2776 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2777 contrib = se->avg.running_avg_sum * scale_load_down(SCHED_LOAD_SCALE);
2778 contrib /= (se->avg.avg_period + 1);
2779 se->avg.utilization_avg_contrib = scale_load(contrib);
2782 static long __update_entity_utilization_avg_contrib(struct sched_entity *se)
2784 long old_contrib = se->avg.utilization_avg_contrib;
2786 if (entity_is_task(se))
2787 __update_task_entity_utilization(se);
2789 se->avg.utilization_avg_contrib =
2790 group_cfs_rq(se)->utilization_load_avg;
2792 return se->avg.utilization_avg_contrib - old_contrib;
2795 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2798 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2799 cfs_rq->blocked_load_avg -= load_contrib;
2801 cfs_rq->blocked_load_avg = 0;
2804 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2806 /* Update a sched_entity's runnable average */
2807 static inline void update_entity_load_avg(struct sched_entity *se,
2810 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2811 long contrib_delta, utilization_delta;
2812 int cpu = cpu_of(rq_of(cfs_rq));
2816 * For a group entity we need to use their owned cfs_rq_clock_task() in
2817 * case they are the parent of a throttled hierarchy.
2819 if (entity_is_task(se))
2820 now = cfs_rq_clock_task(cfs_rq);
2822 now = cfs_rq_clock_task(group_cfs_rq(se));
2824 if (!__update_entity_runnable_avg(now, cpu, &se->avg, se->on_rq,
2825 cfs_rq->curr == se))
2828 contrib_delta = __update_entity_load_avg_contrib(se);
2829 utilization_delta = __update_entity_utilization_avg_contrib(se);
2835 cfs_rq->runnable_load_avg += contrib_delta;
2836 cfs_rq->utilization_load_avg += utilization_delta;
2838 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2843 * Decay the load contributed by all blocked children and account this so that
2844 * their contribution may appropriately discounted when they wake up.
2846 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2848 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2851 decays = now - cfs_rq->last_decay;
2852 if (!decays && !force_update)
2855 if (atomic_long_read(&cfs_rq->removed_load)) {
2856 unsigned long removed_load;
2857 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2858 subtract_blocked_load_contrib(cfs_rq, removed_load);
2862 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2864 atomic64_add(decays, &cfs_rq->decay_counter);
2865 cfs_rq->last_decay = now;
2868 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2871 /* Add the load generated by se into cfs_rq's child load-average */
2872 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2873 struct sched_entity *se,
2877 * We track migrations using entity decay_count <= 0, on a wake-up
2878 * migration we use a negative decay count to track the remote decays
2879 * accumulated while sleeping.
2881 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2882 * are seen by enqueue_entity_load_avg() as a migration with an already
2883 * constructed load_avg_contrib.
2885 if (unlikely(se->avg.decay_count <= 0)) {
2886 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2887 if (se->avg.decay_count) {
2889 * In a wake-up migration we have to approximate the
2890 * time sleeping. This is because we can't synchronize
2891 * clock_task between the two cpus, and it is not
2892 * guaranteed to be read-safe. Instead, we can
2893 * approximate this using our carried decays, which are
2894 * explicitly atomically readable.
2896 se->avg.last_runnable_update -= (-se->avg.decay_count)
2898 update_entity_load_avg(se, 0);
2899 /* Indicate that we're now synchronized and on-rq */
2900 se->avg.decay_count = 0;
2904 __synchronize_entity_decay(se);
2907 /* migrated tasks did not contribute to our blocked load */
2909 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2910 update_entity_load_avg(se, 0);
2913 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2914 cfs_rq->utilization_load_avg += se->avg.utilization_avg_contrib;
2915 /* we force update consideration on load-balancer moves */
2916 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2920 * Remove se's load from this cfs_rq child load-average, if the entity is
2921 * transitioning to a blocked state we track its projected decay using
2924 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2925 struct sched_entity *se,
2928 update_entity_load_avg(se, 1);
2929 /* we force update consideration on load-balancer moves */
2930 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2932 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2933 cfs_rq->utilization_load_avg -= se->avg.utilization_avg_contrib;
2935 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2936 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2937 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2941 * Update the rq's load with the elapsed running time before entering
2942 * idle. if the last scheduled task is not a CFS task, idle_enter will
2943 * be the only way to update the runnable statistic.
2945 void idle_enter_fair(struct rq *this_rq)
2947 update_rq_runnable_avg(this_rq, 1);
2951 * Update the rq's load with the elapsed idle time before a task is
2952 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2953 * be the only way to update the runnable statistic.
2955 void idle_exit_fair(struct rq *this_rq)
2957 update_rq_runnable_avg(this_rq, 0);
2960 static int idle_balance(struct rq *this_rq);
2962 #else /* CONFIG_SMP */
2964 static inline void update_entity_load_avg(struct sched_entity *se,
2965 int update_cfs_rq) {}
2966 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2967 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2968 struct sched_entity *se,
2970 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2971 struct sched_entity *se,
2973 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2974 int force_update) {}
2976 static inline int idle_balance(struct rq *rq)
2981 #endif /* CONFIG_SMP */
2983 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2985 #ifdef CONFIG_SCHEDSTATS
2986 struct task_struct *tsk = NULL;
2988 if (entity_is_task(se))
2991 if (se->statistics.sleep_start) {
2992 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2997 if (unlikely(delta > se->statistics.sleep_max))
2998 se->statistics.sleep_max = delta;
3000 se->statistics.sleep_start = 0;
3001 se->statistics.sum_sleep_runtime += delta;
3004 account_scheduler_latency(tsk, delta >> 10, 1);
3005 trace_sched_stat_sleep(tsk, delta);
3008 if (se->statistics.block_start) {
3009 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3014 if (unlikely(delta > se->statistics.block_max))
3015 se->statistics.block_max = delta;
3017 se->statistics.block_start = 0;
3018 se->statistics.sum_sleep_runtime += delta;
3021 if (tsk->in_iowait) {
3022 se->statistics.iowait_sum += delta;
3023 se->statistics.iowait_count++;
3024 trace_sched_stat_iowait(tsk, delta);
3027 trace_sched_stat_blocked(tsk, delta);
3030 * Blocking time is in units of nanosecs, so shift by
3031 * 20 to get a milliseconds-range estimation of the
3032 * amount of time that the task spent sleeping:
3034 if (unlikely(prof_on == SLEEP_PROFILING)) {
3035 profile_hits(SLEEP_PROFILING,
3036 (void *)get_wchan(tsk),
3039 account_scheduler_latency(tsk, delta >> 10, 0);
3045 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3047 #ifdef CONFIG_SCHED_DEBUG
3048 s64 d = se->vruntime - cfs_rq->min_vruntime;
3053 if (d > 3*sysctl_sched_latency)
3054 schedstat_inc(cfs_rq, nr_spread_over);
3059 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3061 u64 vruntime = cfs_rq->min_vruntime;
3064 * The 'current' period is already promised to the current tasks,
3065 * however the extra weight of the new task will slow them down a
3066 * little, place the new task so that it fits in the slot that
3067 * stays open at the end.
3069 if (initial && sched_feat(START_DEBIT))
3070 vruntime += sched_vslice(cfs_rq, se);
3072 /* sleeps up to a single latency don't count. */
3074 unsigned long thresh = sysctl_sched_latency;
3077 * Halve their sleep time's effect, to allow
3078 * for a gentler effect of sleepers:
3080 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3086 /* ensure we never gain time by being placed backwards. */
3087 se->vruntime = max_vruntime(se->vruntime, vruntime);
3090 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3093 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3096 * Update the normalized vruntime before updating min_vruntime
3097 * through calling update_curr().
3099 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3100 se->vruntime += cfs_rq->min_vruntime;
3103 * Update run-time statistics of the 'current'.
3105 update_curr(cfs_rq);
3106 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
3107 account_entity_enqueue(cfs_rq, se);
3108 update_cfs_shares(cfs_rq);
3110 if (flags & ENQUEUE_WAKEUP) {
3111 place_entity(cfs_rq, se, 0);
3112 enqueue_sleeper(cfs_rq, se);
3115 update_stats_enqueue(cfs_rq, se);
3116 check_spread(cfs_rq, se);
3117 if (se != cfs_rq->curr)
3118 __enqueue_entity(cfs_rq, se);
3121 if (cfs_rq->nr_running == 1) {
3122 list_add_leaf_cfs_rq(cfs_rq);
3123 check_enqueue_throttle(cfs_rq);
3127 static void __clear_buddies_last(struct sched_entity *se)
3129 for_each_sched_entity(se) {
3130 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3131 if (cfs_rq->last != se)
3134 cfs_rq->last = NULL;
3138 static void __clear_buddies_next(struct sched_entity *se)
3140 for_each_sched_entity(se) {
3141 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3142 if (cfs_rq->next != se)
3145 cfs_rq->next = NULL;
3149 static void __clear_buddies_skip(struct sched_entity *se)
3151 for_each_sched_entity(se) {
3152 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3153 if (cfs_rq->skip != se)
3156 cfs_rq->skip = NULL;
3160 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3162 if (cfs_rq->last == se)
3163 __clear_buddies_last(se);
3165 if (cfs_rq->next == se)
3166 __clear_buddies_next(se);
3168 if (cfs_rq->skip == se)
3169 __clear_buddies_skip(se);
3172 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3175 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3178 * Update run-time statistics of the 'current'.
3180 update_curr(cfs_rq);
3181 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
3183 update_stats_dequeue(cfs_rq, se);
3184 if (flags & DEQUEUE_SLEEP) {
3185 #ifdef CONFIG_SCHEDSTATS
3186 if (entity_is_task(se)) {
3187 struct task_struct *tsk = task_of(se);
3189 if (tsk->state & TASK_INTERRUPTIBLE)
3190 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3191 if (tsk->state & TASK_UNINTERRUPTIBLE)
3192 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3197 clear_buddies(cfs_rq, se);
3199 if (se != cfs_rq->curr)
3200 __dequeue_entity(cfs_rq, se);
3202 account_entity_dequeue(cfs_rq, se);
3205 * Normalize the entity after updating the min_vruntime because the
3206 * update can refer to the ->curr item and we need to reflect this
3207 * movement in our normalized position.
3209 if (!(flags & DEQUEUE_SLEEP))
3210 se->vruntime -= cfs_rq->min_vruntime;
3212 /* return excess runtime on last dequeue */
3213 return_cfs_rq_runtime(cfs_rq);
3215 update_min_vruntime(cfs_rq);
3216 update_cfs_shares(cfs_rq);
3220 * Preempt the current task with a newly woken task if needed:
3223 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3225 unsigned long ideal_runtime, delta_exec;
3226 struct sched_entity *se;
3229 ideal_runtime = sched_slice(cfs_rq, curr);
3230 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3231 if (delta_exec > ideal_runtime) {
3232 resched_curr(rq_of(cfs_rq));
3234 * The current task ran long enough, ensure it doesn't get
3235 * re-elected due to buddy favours.
3237 clear_buddies(cfs_rq, curr);
3242 * Ensure that a task that missed wakeup preemption by a
3243 * narrow margin doesn't have to wait for a full slice.
3244 * This also mitigates buddy induced latencies under load.
3246 if (delta_exec < sysctl_sched_min_granularity)
3249 se = __pick_first_entity(cfs_rq);
3250 delta = curr->vruntime - se->vruntime;
3255 if (delta > ideal_runtime)
3256 resched_curr(rq_of(cfs_rq));
3260 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3262 /* 'current' is not kept within the tree. */
3265 * Any task has to be enqueued before it get to execute on
3266 * a CPU. So account for the time it spent waiting on the
3269 update_stats_wait_end(cfs_rq, se);
3270 __dequeue_entity(cfs_rq, se);
3271 update_entity_load_avg(se, 1);
3274 update_stats_curr_start(cfs_rq, se);
3276 #ifdef CONFIG_SCHEDSTATS
3278 * Track our maximum slice length, if the CPU's load is at
3279 * least twice that of our own weight (i.e. dont track it
3280 * when there are only lesser-weight tasks around):
3282 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3283 se->statistics.slice_max = max(se->statistics.slice_max,
3284 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3287 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3291 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3294 * Pick the next process, keeping these things in mind, in this order:
3295 * 1) keep things fair between processes/task groups
3296 * 2) pick the "next" process, since someone really wants that to run
3297 * 3) pick the "last" process, for cache locality
3298 * 4) do not run the "skip" process, if something else is available
3300 static struct sched_entity *
3301 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3303 struct sched_entity *left = __pick_first_entity(cfs_rq);
3304 struct sched_entity *se;
3307 * If curr is set we have to see if its left of the leftmost entity
3308 * still in the tree, provided there was anything in the tree at all.
3310 if (!left || (curr && entity_before(curr, left)))
3313 se = left; /* ideally we run the leftmost entity */
3316 * Avoid running the skip buddy, if running something else can
3317 * be done without getting too unfair.
3319 if (cfs_rq->skip == se) {
3320 struct sched_entity *second;
3323 second = __pick_first_entity(cfs_rq);
3325 second = __pick_next_entity(se);
3326 if (!second || (curr && entity_before(curr, second)))
3330 if (second && wakeup_preempt_entity(second, left) < 1)
3335 * Prefer last buddy, try to return the CPU to a preempted task.
3337 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3341 * Someone really wants this to run. If it's not unfair, run it.
3343 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3346 clear_buddies(cfs_rq, se);
3351 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3353 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3356 * If still on the runqueue then deactivate_task()
3357 * was not called and update_curr() has to be done:
3360 update_curr(cfs_rq);
3362 /* throttle cfs_rqs exceeding runtime */
3363 check_cfs_rq_runtime(cfs_rq);
3365 check_spread(cfs_rq, prev);
3367 update_stats_wait_start(cfs_rq, prev);
3368 /* Put 'current' back into the tree. */
3369 __enqueue_entity(cfs_rq, prev);
3370 /* in !on_rq case, update occurred at dequeue */
3371 update_entity_load_avg(prev, 1);
3373 cfs_rq->curr = NULL;
3377 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3380 * Update run-time statistics of the 'current'.
3382 update_curr(cfs_rq);
3385 * Ensure that runnable average is periodically updated.
3387 update_entity_load_avg(curr, 1);
3388 update_cfs_rq_blocked_load(cfs_rq, 1);
3389 update_cfs_shares(cfs_rq);
3391 #ifdef CONFIG_SCHED_HRTICK
3393 * queued ticks are scheduled to match the slice, so don't bother
3394 * validating it and just reschedule.
3397 resched_curr(rq_of(cfs_rq));
3401 * don't let the period tick interfere with the hrtick preemption
3403 if (!sched_feat(DOUBLE_TICK) &&
3404 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3408 if (cfs_rq->nr_running > 1)
3409 check_preempt_tick(cfs_rq, curr);
3413 /**************************************************
3414 * CFS bandwidth control machinery
3417 #ifdef CONFIG_CFS_BANDWIDTH
3419 #ifdef HAVE_JUMP_LABEL
3420 static struct static_key __cfs_bandwidth_used;
3422 static inline bool cfs_bandwidth_used(void)
3424 return static_key_false(&__cfs_bandwidth_used);
3427 void cfs_bandwidth_usage_inc(void)
3429 static_key_slow_inc(&__cfs_bandwidth_used);
3432 void cfs_bandwidth_usage_dec(void)
3434 static_key_slow_dec(&__cfs_bandwidth_used);
3436 #else /* HAVE_JUMP_LABEL */
3437 static bool cfs_bandwidth_used(void)
3442 void cfs_bandwidth_usage_inc(void) {}
3443 void cfs_bandwidth_usage_dec(void) {}
3444 #endif /* HAVE_JUMP_LABEL */
3447 * default period for cfs group bandwidth.
3448 * default: 0.1s, units: nanoseconds
3450 static inline u64 default_cfs_period(void)
3452 return 100000000ULL;
3455 static inline u64 sched_cfs_bandwidth_slice(void)
3457 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3461 * Replenish runtime according to assigned quota and update expiration time.
3462 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3463 * additional synchronization around rq->lock.
3465 * requires cfs_b->lock
3467 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3471 if (cfs_b->quota == RUNTIME_INF)
3474 now = sched_clock_cpu(smp_processor_id());
3475 cfs_b->runtime = cfs_b->quota;
3476 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3479 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3481 return &tg->cfs_bandwidth;
3484 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3485 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3487 if (unlikely(cfs_rq->throttle_count))
3488 return cfs_rq->throttled_clock_task;
3490 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3493 /* returns 0 on failure to allocate runtime */
3494 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3496 struct task_group *tg = cfs_rq->tg;
3497 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3498 u64 amount = 0, min_amount, expires;
3500 /* note: this is a positive sum as runtime_remaining <= 0 */
3501 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3503 raw_spin_lock(&cfs_b->lock);
3504 if (cfs_b->quota == RUNTIME_INF)
3505 amount = min_amount;
3507 start_cfs_bandwidth(cfs_b);
3509 if (cfs_b->runtime > 0) {
3510 amount = min(cfs_b->runtime, min_amount);
3511 cfs_b->runtime -= amount;
3515 expires = cfs_b->runtime_expires;
3516 raw_spin_unlock(&cfs_b->lock);
3518 cfs_rq->runtime_remaining += amount;
3520 * we may have advanced our local expiration to account for allowed
3521 * spread between our sched_clock and the one on which runtime was
3524 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3525 cfs_rq->runtime_expires = expires;
3527 return cfs_rq->runtime_remaining > 0;
3531 * Note: This depends on the synchronization provided by sched_clock and the
3532 * fact that rq->clock snapshots this value.
3534 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3536 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3538 /* if the deadline is ahead of our clock, nothing to do */
3539 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3542 if (cfs_rq->runtime_remaining < 0)
3546 * If the local deadline has passed we have to consider the
3547 * possibility that our sched_clock is 'fast' and the global deadline
3548 * has not truly expired.
3550 * Fortunately we can check determine whether this the case by checking
3551 * whether the global deadline has advanced. It is valid to compare
3552 * cfs_b->runtime_expires without any locks since we only care about
3553 * exact equality, so a partial write will still work.
3556 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3557 /* extend local deadline, drift is bounded above by 2 ticks */
3558 cfs_rq->runtime_expires += TICK_NSEC;
3560 /* global deadline is ahead, expiration has passed */
3561 cfs_rq->runtime_remaining = 0;
3565 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3567 /* dock delta_exec before expiring quota (as it could span periods) */
3568 cfs_rq->runtime_remaining -= delta_exec;
3569 expire_cfs_rq_runtime(cfs_rq);
3571 if (likely(cfs_rq->runtime_remaining > 0))
3575 * if we're unable to extend our runtime we resched so that the active
3576 * hierarchy can be throttled
3578 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3579 resched_curr(rq_of(cfs_rq));
3582 static __always_inline
3583 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3585 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3588 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3591 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3593 return cfs_bandwidth_used() && cfs_rq->throttled;
3596 /* check whether cfs_rq, or any parent, is throttled */
3597 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3599 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3603 * Ensure that neither of the group entities corresponding to src_cpu or
3604 * dest_cpu are members of a throttled hierarchy when performing group
3605 * load-balance operations.
3607 static inline int throttled_lb_pair(struct task_group *tg,
3608 int src_cpu, int dest_cpu)
3610 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3612 src_cfs_rq = tg->cfs_rq[src_cpu];
3613 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3615 return throttled_hierarchy(src_cfs_rq) ||
3616 throttled_hierarchy(dest_cfs_rq);
3619 /* updated child weight may affect parent so we have to do this bottom up */
3620 static int tg_unthrottle_up(struct task_group *tg, void *data)
3622 struct rq *rq = data;
3623 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3625 cfs_rq->throttle_count--;
3627 if (!cfs_rq->throttle_count) {
3628 /* adjust cfs_rq_clock_task() */
3629 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3630 cfs_rq->throttled_clock_task;
3637 static int tg_throttle_down(struct task_group *tg, void *data)
3639 struct rq *rq = data;
3640 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3642 /* group is entering throttled state, stop time */
3643 if (!cfs_rq->throttle_count)
3644 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3645 cfs_rq->throttle_count++;
3650 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3652 struct rq *rq = rq_of(cfs_rq);
3653 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3654 struct sched_entity *se;
3655 long task_delta, dequeue = 1;
3658 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3660 /* freeze hierarchy runnable averages while throttled */
3662 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3665 task_delta = cfs_rq->h_nr_running;
3666 for_each_sched_entity(se) {
3667 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3668 /* throttled entity or throttle-on-deactivate */
3673 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3674 qcfs_rq->h_nr_running -= task_delta;
3676 if (qcfs_rq->load.weight)
3681 sub_nr_running(rq, task_delta);
3683 cfs_rq->throttled = 1;
3684 cfs_rq->throttled_clock = rq_clock(rq);
3685 raw_spin_lock(&cfs_b->lock);
3686 empty = list_empty(&cfs_rq->throttled_list);
3689 * Add to the _head_ of the list, so that an already-started
3690 * distribute_cfs_runtime will not see us
3692 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3695 * If we're the first throttled task, make sure the bandwidth
3699 start_cfs_bandwidth(cfs_b);
3701 raw_spin_unlock(&cfs_b->lock);
3704 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3706 struct rq *rq = rq_of(cfs_rq);
3707 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3708 struct sched_entity *se;
3712 se = cfs_rq->tg->se[cpu_of(rq)];
3714 cfs_rq->throttled = 0;
3716 update_rq_clock(rq);
3718 raw_spin_lock(&cfs_b->lock);
3719 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3720 list_del_rcu(&cfs_rq->throttled_list);
3721 raw_spin_unlock(&cfs_b->lock);
3723 /* update hierarchical throttle state */
3724 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3726 if (!cfs_rq->load.weight)
3729 task_delta = cfs_rq->h_nr_running;
3730 for_each_sched_entity(se) {
3734 cfs_rq = cfs_rq_of(se);
3736 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3737 cfs_rq->h_nr_running += task_delta;
3739 if (cfs_rq_throttled(cfs_rq))
3744 add_nr_running(rq, task_delta);
3746 /* determine whether we need to wake up potentially idle cpu */
3747 if (rq->curr == rq->idle && rq->cfs.nr_running)
3751 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3752 u64 remaining, u64 expires)
3754 struct cfs_rq *cfs_rq;
3756 u64 starting_runtime = remaining;
3759 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3761 struct rq *rq = rq_of(cfs_rq);
3763 raw_spin_lock(&rq->lock);
3764 if (!cfs_rq_throttled(cfs_rq))
3767 runtime = -cfs_rq->runtime_remaining + 1;
3768 if (runtime > remaining)
3769 runtime = remaining;
3770 remaining -= runtime;
3772 cfs_rq->runtime_remaining += runtime;
3773 cfs_rq->runtime_expires = expires;
3775 /* we check whether we're throttled above */
3776 if (cfs_rq->runtime_remaining > 0)
3777 unthrottle_cfs_rq(cfs_rq);
3780 raw_spin_unlock(&rq->lock);
3787 return starting_runtime - remaining;
3791 * Responsible for refilling a task_group's bandwidth and unthrottling its
3792 * cfs_rqs as appropriate. If there has been no activity within the last
3793 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3794 * used to track this state.
3796 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3798 u64 runtime, runtime_expires;
3801 /* no need to continue the timer with no bandwidth constraint */
3802 if (cfs_b->quota == RUNTIME_INF)
3803 goto out_deactivate;
3805 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3806 cfs_b->nr_periods += overrun;
3809 * idle depends on !throttled (for the case of a large deficit), and if
3810 * we're going inactive then everything else can be deferred
3812 if (cfs_b->idle && !throttled)
3813 goto out_deactivate;
3815 __refill_cfs_bandwidth_runtime(cfs_b);
3818 /* mark as potentially idle for the upcoming period */
3823 /* account preceding periods in which throttling occurred */
3824 cfs_b->nr_throttled += overrun;
3826 runtime_expires = cfs_b->runtime_expires;
3829 * This check is repeated as we are holding onto the new bandwidth while
3830 * we unthrottle. This can potentially race with an unthrottled group
3831 * trying to acquire new bandwidth from the global pool. This can result
3832 * in us over-using our runtime if it is all used during this loop, but
3833 * only by limited amounts in that extreme case.
3835 while (throttled && cfs_b->runtime > 0) {
3836 runtime = cfs_b->runtime;
3837 raw_spin_unlock(&cfs_b->lock);
3838 /* we can't nest cfs_b->lock while distributing bandwidth */
3839 runtime = distribute_cfs_runtime(cfs_b, runtime,
3841 raw_spin_lock(&cfs_b->lock);
3843 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3845 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3849 * While we are ensured activity in the period following an
3850 * unthrottle, this also covers the case in which the new bandwidth is
3851 * insufficient to cover the existing bandwidth deficit. (Forcing the
3852 * timer to remain active while there are any throttled entities.)
3862 /* a cfs_rq won't donate quota below this amount */
3863 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3864 /* minimum remaining period time to redistribute slack quota */
3865 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3866 /* how long we wait to gather additional slack before distributing */
3867 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3870 * Are we near the end of the current quota period?
3872 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3873 * hrtimer base being cleared by hrtimer_start. In the case of
3874 * migrate_hrtimers, base is never cleared, so we are fine.
3876 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3878 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3881 /* if the call-back is running a quota refresh is already occurring */
3882 if (hrtimer_callback_running(refresh_timer))
3885 /* is a quota refresh about to occur? */
3886 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3887 if (remaining < min_expire)
3893 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3895 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3897 /* if there's a quota refresh soon don't bother with slack */
3898 if (runtime_refresh_within(cfs_b, min_left))
3901 hrtimer_start(&cfs_b->slack_timer,
3902 ns_to_ktime(cfs_bandwidth_slack_period),
3906 /* we know any runtime found here is valid as update_curr() precedes return */
3907 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3909 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3910 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3912 if (slack_runtime <= 0)
3915 raw_spin_lock(&cfs_b->lock);
3916 if (cfs_b->quota != RUNTIME_INF &&
3917 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3918 cfs_b->runtime += slack_runtime;
3920 /* we are under rq->lock, defer unthrottling using a timer */
3921 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3922 !list_empty(&cfs_b->throttled_cfs_rq))
3923 start_cfs_slack_bandwidth(cfs_b);
3925 raw_spin_unlock(&cfs_b->lock);
3927 /* even if it's not valid for return we don't want to try again */
3928 cfs_rq->runtime_remaining -= slack_runtime;
3931 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3933 if (!cfs_bandwidth_used())
3936 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3939 __return_cfs_rq_runtime(cfs_rq);
3943 * This is done with a timer (instead of inline with bandwidth return) since
3944 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3946 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3948 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3951 /* confirm we're still not at a refresh boundary */
3952 raw_spin_lock(&cfs_b->lock);
3953 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3954 raw_spin_unlock(&cfs_b->lock);
3958 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3959 runtime = cfs_b->runtime;
3961 expires = cfs_b->runtime_expires;
3962 raw_spin_unlock(&cfs_b->lock);
3967 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3969 raw_spin_lock(&cfs_b->lock);
3970 if (expires == cfs_b->runtime_expires)
3971 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3972 raw_spin_unlock(&cfs_b->lock);
3976 * When a group wakes up we want to make sure that its quota is not already
3977 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3978 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3980 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3982 if (!cfs_bandwidth_used())
3985 /* an active group must be handled by the update_curr()->put() path */
3986 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3989 /* ensure the group is not already throttled */
3990 if (cfs_rq_throttled(cfs_rq))
3993 /* update runtime allocation */
3994 account_cfs_rq_runtime(cfs_rq, 0);
3995 if (cfs_rq->runtime_remaining <= 0)
3996 throttle_cfs_rq(cfs_rq);
3999 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4000 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4002 if (!cfs_bandwidth_used())
4005 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4009 * it's possible for a throttled entity to be forced into a running
4010 * state (e.g. set_curr_task), in this case we're finished.
4012 if (cfs_rq_throttled(cfs_rq))
4015 throttle_cfs_rq(cfs_rq);
4019 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4021 struct cfs_bandwidth *cfs_b =
4022 container_of(timer, struct cfs_bandwidth, slack_timer);
4024 do_sched_cfs_slack_timer(cfs_b);
4026 return HRTIMER_NORESTART;
4029 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4031 struct cfs_bandwidth *cfs_b =
4032 container_of(timer, struct cfs_bandwidth, period_timer);
4036 raw_spin_lock(&cfs_b->lock);
4038 overrun = hrtimer_forward_now(timer, cfs_b->period);
4042 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4045 cfs_b->period_active = 0;
4046 raw_spin_unlock(&cfs_b->lock);
4048 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4051 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4053 raw_spin_lock_init(&cfs_b->lock);
4055 cfs_b->quota = RUNTIME_INF;
4056 cfs_b->period = ns_to_ktime(default_cfs_period());
4058 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4059 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4060 cfs_b->period_timer.function = sched_cfs_period_timer;
4061 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4062 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4065 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4067 cfs_rq->runtime_enabled = 0;
4068 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4071 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4073 lockdep_assert_held(&cfs_b->lock);
4075 if (!cfs_b->period_active) {
4076 cfs_b->period_active = 1;
4077 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4078 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4082 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4084 /* init_cfs_bandwidth() was not called */
4085 if (!cfs_b->throttled_cfs_rq.next)
4088 hrtimer_cancel(&cfs_b->period_timer);
4089 hrtimer_cancel(&cfs_b->slack_timer);
4092 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4094 struct cfs_rq *cfs_rq;
4096 for_each_leaf_cfs_rq(rq, cfs_rq) {
4097 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4099 raw_spin_lock(&cfs_b->lock);
4100 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4101 raw_spin_unlock(&cfs_b->lock);
4105 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4107 struct cfs_rq *cfs_rq;
4109 for_each_leaf_cfs_rq(rq, cfs_rq) {
4110 if (!cfs_rq->runtime_enabled)
4114 * clock_task is not advancing so we just need to make sure
4115 * there's some valid quota amount
4117 cfs_rq->runtime_remaining = 1;
4119 * Offline rq is schedulable till cpu is completely disabled
4120 * in take_cpu_down(), so we prevent new cfs throttling here.
4122 cfs_rq->runtime_enabled = 0;
4124 if (cfs_rq_throttled(cfs_rq))
4125 unthrottle_cfs_rq(cfs_rq);
4129 #else /* CONFIG_CFS_BANDWIDTH */
4130 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4132 return rq_clock_task(rq_of(cfs_rq));
4135 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4136 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4137 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4138 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4140 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4145 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4150 static inline int throttled_lb_pair(struct task_group *tg,
4151 int src_cpu, int dest_cpu)
4156 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4158 #ifdef CONFIG_FAIR_GROUP_SCHED
4159 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4162 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4166 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4167 static inline void update_runtime_enabled(struct rq *rq) {}
4168 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4170 #endif /* CONFIG_CFS_BANDWIDTH */
4172 /**************************************************
4173 * CFS operations on tasks:
4176 #ifdef CONFIG_SCHED_HRTICK
4177 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4179 struct sched_entity *se = &p->se;
4180 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4182 WARN_ON(task_rq(p) != rq);
4184 if (cfs_rq->nr_running > 1) {
4185 u64 slice = sched_slice(cfs_rq, se);
4186 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4187 s64 delta = slice - ran;
4194 hrtick_start(rq, delta);
4199 * called from enqueue/dequeue and updates the hrtick when the
4200 * current task is from our class and nr_running is low enough
4203 static void hrtick_update(struct rq *rq)
4205 struct task_struct *curr = rq->curr;
4207 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4210 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4211 hrtick_start_fair(rq, curr);
4213 #else /* !CONFIG_SCHED_HRTICK */
4215 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4219 static inline void hrtick_update(struct rq *rq)
4225 * The enqueue_task method is called before nr_running is
4226 * increased. Here we update the fair scheduling stats and
4227 * then put the task into the rbtree:
4230 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4232 struct cfs_rq *cfs_rq;
4233 struct sched_entity *se = &p->se;
4235 for_each_sched_entity(se) {
4238 cfs_rq = cfs_rq_of(se);
4239 enqueue_entity(cfs_rq, se, flags);
4242 * end evaluation on encountering a throttled cfs_rq
4244 * note: in the case of encountering a throttled cfs_rq we will
4245 * post the final h_nr_running increment below.
4247 if (cfs_rq_throttled(cfs_rq))
4249 cfs_rq->h_nr_running++;
4251 flags = ENQUEUE_WAKEUP;
4254 for_each_sched_entity(se) {
4255 cfs_rq = cfs_rq_of(se);
4256 cfs_rq->h_nr_running++;
4258 if (cfs_rq_throttled(cfs_rq))
4261 update_cfs_shares(cfs_rq);
4262 update_entity_load_avg(se, 1);
4266 update_rq_runnable_avg(rq, rq->nr_running);
4267 add_nr_running(rq, 1);
4272 static void set_next_buddy(struct sched_entity *se);
4275 * The dequeue_task method is called before nr_running is
4276 * decreased. We remove the task from the rbtree and
4277 * update the fair scheduling stats:
4279 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4281 struct cfs_rq *cfs_rq;
4282 struct sched_entity *se = &p->se;
4283 int task_sleep = flags & DEQUEUE_SLEEP;
4285 for_each_sched_entity(se) {
4286 cfs_rq = cfs_rq_of(se);
4287 dequeue_entity(cfs_rq, se, flags);
4290 * end evaluation on encountering a throttled cfs_rq
4292 * note: in the case of encountering a throttled cfs_rq we will
4293 * post the final h_nr_running decrement below.
4295 if (cfs_rq_throttled(cfs_rq))
4297 cfs_rq->h_nr_running--;
4299 /* Don't dequeue parent if it has other entities besides us */
4300 if (cfs_rq->load.weight) {
4302 * Bias pick_next to pick a task from this cfs_rq, as
4303 * p is sleeping when it is within its sched_slice.
4305 if (task_sleep && parent_entity(se))
4306 set_next_buddy(parent_entity(se));
4308 /* avoid re-evaluating load for this entity */
4309 se = parent_entity(se);
4312 flags |= DEQUEUE_SLEEP;
4315 for_each_sched_entity(se) {
4316 cfs_rq = cfs_rq_of(se);
4317 cfs_rq->h_nr_running--;
4319 if (cfs_rq_throttled(cfs_rq))
4322 update_cfs_shares(cfs_rq);
4323 update_entity_load_avg(se, 1);
4327 sub_nr_running(rq, 1);
4328 update_rq_runnable_avg(rq, 1);
4336 * per rq 'load' arrray crap; XXX kill this.
4340 * The exact cpuload at various idx values, calculated at every tick would be
4341 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4343 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4344 * on nth tick when cpu may be busy, then we have:
4345 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4346 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4348 * decay_load_missed() below does efficient calculation of
4349 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4350 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4352 * The calculation is approximated on a 128 point scale.
4353 * degrade_zero_ticks is the number of ticks after which load at any
4354 * particular idx is approximated to be zero.
4355 * degrade_factor is a precomputed table, a row for each load idx.
4356 * Each column corresponds to degradation factor for a power of two ticks,
4357 * based on 128 point scale.
4359 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4360 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4362 * With this power of 2 load factors, we can degrade the load n times
4363 * by looking at 1 bits in n and doing as many mult/shift instead of
4364 * n mult/shifts needed by the exact degradation.
4366 #define DEGRADE_SHIFT 7
4367 static const unsigned char
4368 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4369 static const unsigned char
4370 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4371 {0, 0, 0, 0, 0, 0, 0, 0},
4372 {64, 32, 8, 0, 0, 0, 0, 0},
4373 {96, 72, 40, 12, 1, 0, 0},
4374 {112, 98, 75, 43, 15, 1, 0},
4375 {120, 112, 98, 76, 45, 16, 2} };
4378 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4379 * would be when CPU is idle and so we just decay the old load without
4380 * adding any new load.
4382 static unsigned long
4383 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4387 if (!missed_updates)
4390 if (missed_updates >= degrade_zero_ticks[idx])
4394 return load >> missed_updates;
4396 while (missed_updates) {
4397 if (missed_updates % 2)
4398 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4400 missed_updates >>= 1;
4407 * Update rq->cpu_load[] statistics. This function is usually called every
4408 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4409 * every tick. We fix it up based on jiffies.
4411 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4412 unsigned long pending_updates)
4416 this_rq->nr_load_updates++;
4418 /* Update our load: */
4419 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4420 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4421 unsigned long old_load, new_load;
4423 /* scale is effectively 1 << i now, and >> i divides by scale */
4425 old_load = this_rq->cpu_load[i];
4426 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4427 new_load = this_load;
4429 * Round up the averaging division if load is increasing. This
4430 * prevents us from getting stuck on 9 if the load is 10, for
4433 if (new_load > old_load)
4434 new_load += scale - 1;
4436 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4439 sched_avg_update(this_rq);
4442 #ifdef CONFIG_NO_HZ_COMMON
4444 * There is no sane way to deal with nohz on smp when using jiffies because the
4445 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4446 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4448 * Therefore we cannot use the delta approach from the regular tick since that
4449 * would seriously skew the load calculation. However we'll make do for those
4450 * updates happening while idle (nohz_idle_balance) or coming out of idle
4451 * (tick_nohz_idle_exit).
4453 * This means we might still be one tick off for nohz periods.
4457 * Called from nohz_idle_balance() to update the load ratings before doing the
4460 static void update_idle_cpu_load(struct rq *this_rq)
4462 unsigned long curr_jiffies = READ_ONCE(jiffies);
4463 unsigned long load = this_rq->cfs.runnable_load_avg;
4464 unsigned long pending_updates;
4467 * bail if there's load or we're actually up-to-date.
4469 if (load || curr_jiffies == this_rq->last_load_update_tick)
4472 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4473 this_rq->last_load_update_tick = curr_jiffies;
4475 __update_cpu_load(this_rq, load, pending_updates);
4479 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4481 void update_cpu_load_nohz(void)
4483 struct rq *this_rq = this_rq();
4484 unsigned long curr_jiffies = READ_ONCE(jiffies);
4485 unsigned long pending_updates;
4487 if (curr_jiffies == this_rq->last_load_update_tick)
4490 raw_spin_lock(&this_rq->lock);
4491 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4492 if (pending_updates) {
4493 this_rq->last_load_update_tick = curr_jiffies;
4495 * We were idle, this means load 0, the current load might be
4496 * !0 due to remote wakeups and the sort.
4498 __update_cpu_load(this_rq, 0, pending_updates);
4500 raw_spin_unlock(&this_rq->lock);
4502 #endif /* CONFIG_NO_HZ */
4505 * Called from scheduler_tick()
4507 void update_cpu_load_active(struct rq *this_rq)
4509 unsigned long load = this_rq->cfs.runnable_load_avg;
4511 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4513 this_rq->last_load_update_tick = jiffies;
4514 __update_cpu_load(this_rq, load, 1);
4517 /* Used instead of source_load when we know the type == 0 */
4518 static unsigned long weighted_cpuload(const int cpu)
4520 return cpu_rq(cpu)->cfs.runnable_load_avg;
4524 * Return a low guess at the load of a migration-source cpu weighted
4525 * according to the scheduling class and "nice" value.
4527 * We want to under-estimate the load of migration sources, to
4528 * balance conservatively.
4530 static unsigned long source_load(int cpu, int type)
4532 struct rq *rq = cpu_rq(cpu);
4533 unsigned long total = weighted_cpuload(cpu);
4535 if (type == 0 || !sched_feat(LB_BIAS))
4538 return min(rq->cpu_load[type-1], total);
4542 * Return a high guess at the load of a migration-target cpu weighted
4543 * according to the scheduling class and "nice" value.
4545 static unsigned long target_load(int cpu, int type)
4547 struct rq *rq = cpu_rq(cpu);
4548 unsigned long total = weighted_cpuload(cpu);
4550 if (type == 0 || !sched_feat(LB_BIAS))
4553 return max(rq->cpu_load[type-1], total);
4556 static unsigned long capacity_of(int cpu)
4558 return cpu_rq(cpu)->cpu_capacity;
4561 static unsigned long capacity_orig_of(int cpu)
4563 return cpu_rq(cpu)->cpu_capacity_orig;
4566 static unsigned long cpu_avg_load_per_task(int cpu)
4568 struct rq *rq = cpu_rq(cpu);
4569 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4570 unsigned long load_avg = rq->cfs.runnable_load_avg;
4573 return load_avg / nr_running;
4578 static void record_wakee(struct task_struct *p)
4581 * Rough decay (wiping) for cost saving, don't worry
4582 * about the boundary, really active task won't care
4585 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4586 current->wakee_flips >>= 1;
4587 current->wakee_flip_decay_ts = jiffies;
4590 if (current->last_wakee != p) {
4591 current->last_wakee = p;
4592 current->wakee_flips++;
4596 static void task_waking_fair(struct task_struct *p)
4598 struct sched_entity *se = &p->se;
4599 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4602 #ifndef CONFIG_64BIT
4603 u64 min_vruntime_copy;
4606 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4608 min_vruntime = cfs_rq->min_vruntime;
4609 } while (min_vruntime != min_vruntime_copy);
4611 min_vruntime = cfs_rq->min_vruntime;
4614 se->vruntime -= min_vruntime;
4618 #ifdef CONFIG_FAIR_GROUP_SCHED
4620 * effective_load() calculates the load change as seen from the root_task_group
4622 * Adding load to a group doesn't make a group heavier, but can cause movement
4623 * of group shares between cpus. Assuming the shares were perfectly aligned one
4624 * can calculate the shift in shares.
4626 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4627 * on this @cpu and results in a total addition (subtraction) of @wg to the
4628 * total group weight.
4630 * Given a runqueue weight distribution (rw_i) we can compute a shares
4631 * distribution (s_i) using:
4633 * s_i = rw_i / \Sum rw_j (1)
4635 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4636 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4637 * shares distribution (s_i):
4639 * rw_i = { 2, 4, 1, 0 }
4640 * s_i = { 2/7, 4/7, 1/7, 0 }
4642 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4643 * task used to run on and the CPU the waker is running on), we need to
4644 * compute the effect of waking a task on either CPU and, in case of a sync
4645 * wakeup, compute the effect of the current task going to sleep.
4647 * So for a change of @wl to the local @cpu with an overall group weight change
4648 * of @wl we can compute the new shares distribution (s'_i) using:
4650 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4652 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4653 * differences in waking a task to CPU 0. The additional task changes the
4654 * weight and shares distributions like:
4656 * rw'_i = { 3, 4, 1, 0 }
4657 * s'_i = { 3/8, 4/8, 1/8, 0 }
4659 * We can then compute the difference in effective weight by using:
4661 * dw_i = S * (s'_i - s_i) (3)
4663 * Where 'S' is the group weight as seen by its parent.
4665 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4666 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4667 * 4/7) times the weight of the group.
4669 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4671 struct sched_entity *se = tg->se[cpu];
4673 if (!tg->parent) /* the trivial, non-cgroup case */
4676 for_each_sched_entity(se) {
4682 * W = @wg + \Sum rw_j
4684 W = wg + calc_tg_weight(tg, se->my_q);
4689 w = se->my_q->load.weight + wl;
4692 * wl = S * s'_i; see (2)
4695 wl = (w * (long)tg->shares) / W;
4700 * Per the above, wl is the new se->load.weight value; since
4701 * those are clipped to [MIN_SHARES, ...) do so now. See
4702 * calc_cfs_shares().
4704 if (wl < MIN_SHARES)
4708 * wl = dw_i = S * (s'_i - s_i); see (3)
4710 wl -= se->load.weight;
4713 * Recursively apply this logic to all parent groups to compute
4714 * the final effective load change on the root group. Since
4715 * only the @tg group gets extra weight, all parent groups can
4716 * only redistribute existing shares. @wl is the shift in shares
4717 * resulting from this level per the above.
4726 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4733 static int wake_wide(struct task_struct *p)
4735 int factor = this_cpu_read(sd_llc_size);
4738 * Yeah, it's the switching-frequency, could means many wakee or
4739 * rapidly switch, use factor here will just help to automatically
4740 * adjust the loose-degree, so bigger node will lead to more pull.
4742 if (p->wakee_flips > factor) {
4744 * wakee is somewhat hot, it needs certain amount of cpu
4745 * resource, so if waker is far more hot, prefer to leave
4748 if (current->wakee_flips > (factor * p->wakee_flips))
4755 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4757 s64 this_load, load;
4758 s64 this_eff_load, prev_eff_load;
4759 int idx, this_cpu, prev_cpu;
4760 struct task_group *tg;
4761 unsigned long weight;
4765 * If we wake multiple tasks be careful to not bounce
4766 * ourselves around too much.
4772 this_cpu = smp_processor_id();
4773 prev_cpu = task_cpu(p);
4774 load = source_load(prev_cpu, idx);
4775 this_load = target_load(this_cpu, idx);
4778 * If sync wakeup then subtract the (maximum possible)
4779 * effect of the currently running task from the load
4780 * of the current CPU:
4783 tg = task_group(current);
4784 weight = current->se.load.weight;
4786 this_load += effective_load(tg, this_cpu, -weight, -weight);
4787 load += effective_load(tg, prev_cpu, 0, -weight);
4791 weight = p->se.load.weight;
4794 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4795 * due to the sync cause above having dropped this_load to 0, we'll
4796 * always have an imbalance, but there's really nothing you can do
4797 * about that, so that's good too.
4799 * Otherwise check if either cpus are near enough in load to allow this
4800 * task to be woken on this_cpu.
4802 this_eff_load = 100;
4803 this_eff_load *= capacity_of(prev_cpu);
4805 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4806 prev_eff_load *= capacity_of(this_cpu);
4808 if (this_load > 0) {
4809 this_eff_load *= this_load +
4810 effective_load(tg, this_cpu, weight, weight);
4812 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4815 balanced = this_eff_load <= prev_eff_load;
4817 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4822 schedstat_inc(sd, ttwu_move_affine);
4823 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4829 * find_idlest_group finds and returns the least busy CPU group within the
4832 static struct sched_group *
4833 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4834 int this_cpu, int sd_flag)
4836 struct sched_group *idlest = NULL, *group = sd->groups;
4837 unsigned long min_load = ULONG_MAX, this_load = 0;
4838 int load_idx = sd->forkexec_idx;
4839 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4841 if (sd_flag & SD_BALANCE_WAKE)
4842 load_idx = sd->wake_idx;
4845 unsigned long load, avg_load;
4849 /* Skip over this group if it has no CPUs allowed */
4850 if (!cpumask_intersects(sched_group_cpus(group),
4851 tsk_cpus_allowed(p)))
4854 local_group = cpumask_test_cpu(this_cpu,
4855 sched_group_cpus(group));
4857 /* Tally up the load of all CPUs in the group */
4860 for_each_cpu(i, sched_group_cpus(group)) {
4861 /* Bias balancing toward cpus of our domain */
4863 load = source_load(i, load_idx);
4865 load = target_load(i, load_idx);
4870 /* Adjust by relative CPU capacity of the group */
4871 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4874 this_load = avg_load;
4875 } else if (avg_load < min_load) {
4876 min_load = avg_load;
4879 } while (group = group->next, group != sd->groups);
4881 if (!idlest || 100*this_load < imbalance*min_load)
4887 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4890 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4892 unsigned long load, min_load = ULONG_MAX;
4893 unsigned int min_exit_latency = UINT_MAX;
4894 u64 latest_idle_timestamp = 0;
4895 int least_loaded_cpu = this_cpu;
4896 int shallowest_idle_cpu = -1;
4899 /* Traverse only the allowed CPUs */
4900 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4902 struct rq *rq = cpu_rq(i);
4903 struct cpuidle_state *idle = idle_get_state(rq);
4904 if (idle && idle->exit_latency < min_exit_latency) {
4906 * We give priority to a CPU whose idle state
4907 * has the smallest exit latency irrespective
4908 * of any idle timestamp.
4910 min_exit_latency = idle->exit_latency;
4911 latest_idle_timestamp = rq->idle_stamp;
4912 shallowest_idle_cpu = i;
4913 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4914 rq->idle_stamp > latest_idle_timestamp) {
4916 * If equal or no active idle state, then
4917 * the most recently idled CPU might have
4920 latest_idle_timestamp = rq->idle_stamp;
4921 shallowest_idle_cpu = i;
4923 } else if (shallowest_idle_cpu == -1) {
4924 load = weighted_cpuload(i);
4925 if (load < min_load || (load == min_load && i == this_cpu)) {
4927 least_loaded_cpu = i;
4932 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4936 * Try and locate an idle CPU in the sched_domain.
4938 static int select_idle_sibling(struct task_struct *p, int target)
4940 struct sched_domain *sd;
4941 struct sched_group *sg;
4942 int i = task_cpu(p);
4944 if (idle_cpu(target))
4948 * If the prevous cpu is cache affine and idle, don't be stupid.
4950 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4954 * Otherwise, iterate the domains and find an elegible idle cpu.
4956 sd = rcu_dereference(per_cpu(sd_llc, target));
4957 for_each_lower_domain(sd) {
4960 if (!cpumask_intersects(sched_group_cpus(sg),
4961 tsk_cpus_allowed(p)))
4964 for_each_cpu(i, sched_group_cpus(sg)) {
4965 if (i == target || !idle_cpu(i))
4969 target = cpumask_first_and(sched_group_cpus(sg),
4970 tsk_cpus_allowed(p));
4974 } while (sg != sd->groups);
4980 * get_cpu_usage returns the amount of capacity of a CPU that is used by CFS
4981 * tasks. The unit of the return value must be the one of capacity so we can
4982 * compare the usage with the capacity of the CPU that is available for CFS
4983 * task (ie cpu_capacity).
4984 * cfs.utilization_load_avg is the sum of running time of runnable tasks on a
4985 * CPU. It represents the amount of utilization of a CPU in the range
4986 * [0..SCHED_LOAD_SCALE]. The usage of a CPU can't be higher than the full
4987 * capacity of the CPU because it's about the running time on this CPU.
4988 * Nevertheless, cfs.utilization_load_avg can be higher than SCHED_LOAD_SCALE
4989 * because of unfortunate rounding in avg_period and running_load_avg or just
4990 * after migrating tasks until the average stabilizes with the new running
4991 * time. So we need to check that the usage stays into the range
4992 * [0..cpu_capacity_orig] and cap if necessary.
4993 * Without capping the usage, a group could be seen as overloaded (CPU0 usage
4994 * at 121% + CPU1 usage at 80%) whereas CPU1 has 20% of available capacity
4996 static int get_cpu_usage(int cpu)
4998 unsigned long usage = cpu_rq(cpu)->cfs.utilization_load_avg;
4999 unsigned long capacity = capacity_orig_of(cpu);
5001 if (usage >= SCHED_LOAD_SCALE)
5004 return (usage * capacity) >> SCHED_LOAD_SHIFT;
5008 * select_task_rq_fair: Select target runqueue for the waking task in domains
5009 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5010 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5012 * Balances load by selecting the idlest cpu in the idlest group, or under
5013 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5015 * Returns the target cpu number.
5017 * preempt must be disabled.
5020 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5022 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5023 int cpu = smp_processor_id();
5025 int want_affine = 0;
5026 int sync = wake_flags & WF_SYNC;
5028 if (sd_flag & SD_BALANCE_WAKE)
5029 want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5032 for_each_domain(cpu, tmp) {
5033 if (!(tmp->flags & SD_LOAD_BALANCE))
5037 * If both cpu and prev_cpu are part of this domain,
5038 * cpu is a valid SD_WAKE_AFFINE target.
5040 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5041 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5046 if (tmp->flags & sd_flag)
5050 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5053 if (sd_flag & SD_BALANCE_WAKE) {
5054 new_cpu = select_idle_sibling(p, prev_cpu);
5059 struct sched_group *group;
5062 if (!(sd->flags & sd_flag)) {
5067 group = find_idlest_group(sd, p, cpu, sd_flag);
5073 new_cpu = find_idlest_cpu(group, p, cpu);
5074 if (new_cpu == -1 || new_cpu == cpu) {
5075 /* Now try balancing at a lower domain level of cpu */
5080 /* Now try balancing at a lower domain level of new_cpu */
5082 weight = sd->span_weight;
5084 for_each_domain(cpu, tmp) {
5085 if (weight <= tmp->span_weight)
5087 if (tmp->flags & sd_flag)
5090 /* while loop will break here if sd == NULL */
5099 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5100 * cfs_rq_of(p) references at time of call are still valid and identify the
5101 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5102 * other assumptions, including the state of rq->lock, should be made.
5105 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
5107 struct sched_entity *se = &p->se;
5108 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5111 * Load tracking: accumulate removed load so that it can be processed
5112 * when we next update owning cfs_rq under rq->lock. Tasks contribute
5113 * to blocked load iff they have a positive decay-count. It can never
5114 * be negative here since on-rq tasks have decay-count == 0.
5116 if (se->avg.decay_count) {
5117 se->avg.decay_count = -__synchronize_entity_decay(se);
5118 atomic_long_add(se->avg.load_avg_contrib,
5119 &cfs_rq->removed_load);
5122 /* We have migrated, no longer consider this task hot */
5125 #endif /* CONFIG_SMP */
5127 static unsigned long
5128 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5130 unsigned long gran = sysctl_sched_wakeup_granularity;
5133 * Since its curr running now, convert the gran from real-time
5134 * to virtual-time in his units.
5136 * By using 'se' instead of 'curr' we penalize light tasks, so
5137 * they get preempted easier. That is, if 'se' < 'curr' then
5138 * the resulting gran will be larger, therefore penalizing the
5139 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5140 * be smaller, again penalizing the lighter task.
5142 * This is especially important for buddies when the leftmost
5143 * task is higher priority than the buddy.
5145 return calc_delta_fair(gran, se);
5149 * Should 'se' preempt 'curr'.
5163 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5165 s64 gran, vdiff = curr->vruntime - se->vruntime;
5170 gran = wakeup_gran(curr, se);
5177 static void set_last_buddy(struct sched_entity *se)
5179 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5182 for_each_sched_entity(se)
5183 cfs_rq_of(se)->last = se;
5186 static void set_next_buddy(struct sched_entity *se)
5188 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5191 for_each_sched_entity(se)
5192 cfs_rq_of(se)->next = se;
5195 static void set_skip_buddy(struct sched_entity *se)
5197 for_each_sched_entity(se)
5198 cfs_rq_of(se)->skip = se;
5202 * Preempt the current task with a newly woken task if needed:
5204 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5206 struct task_struct *curr = rq->curr;
5207 struct sched_entity *se = &curr->se, *pse = &p->se;
5208 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5209 int scale = cfs_rq->nr_running >= sched_nr_latency;
5210 int next_buddy_marked = 0;
5212 if (unlikely(se == pse))
5216 * This is possible from callers such as attach_tasks(), in which we
5217 * unconditionally check_prempt_curr() after an enqueue (which may have
5218 * lead to a throttle). This both saves work and prevents false
5219 * next-buddy nomination below.
5221 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5224 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5225 set_next_buddy(pse);
5226 next_buddy_marked = 1;
5230 * We can come here with TIF_NEED_RESCHED already set from new task
5233 * Note: this also catches the edge-case of curr being in a throttled
5234 * group (e.g. via set_curr_task), since update_curr() (in the
5235 * enqueue of curr) will have resulted in resched being set. This
5236 * prevents us from potentially nominating it as a false LAST_BUDDY
5239 if (test_tsk_need_resched(curr))
5242 /* Idle tasks are by definition preempted by non-idle tasks. */
5243 if (unlikely(curr->policy == SCHED_IDLE) &&
5244 likely(p->policy != SCHED_IDLE))
5248 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5249 * is driven by the tick):
5251 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5254 find_matching_se(&se, &pse);
5255 update_curr(cfs_rq_of(se));
5257 if (wakeup_preempt_entity(se, pse) == 1) {
5259 * Bias pick_next to pick the sched entity that is
5260 * triggering this preemption.
5262 if (!next_buddy_marked)
5263 set_next_buddy(pse);
5272 * Only set the backward buddy when the current task is still
5273 * on the rq. This can happen when a wakeup gets interleaved
5274 * with schedule on the ->pre_schedule() or idle_balance()
5275 * point, either of which can * drop the rq lock.
5277 * Also, during early boot the idle thread is in the fair class,
5278 * for obvious reasons its a bad idea to schedule back to it.
5280 if (unlikely(!se->on_rq || curr == rq->idle))
5283 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5287 static struct task_struct *
5288 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5290 struct cfs_rq *cfs_rq = &rq->cfs;
5291 struct sched_entity *se;
5292 struct task_struct *p;
5296 #ifdef CONFIG_FAIR_GROUP_SCHED
5297 if (!cfs_rq->nr_running)
5300 if (prev->sched_class != &fair_sched_class)
5304 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5305 * likely that a next task is from the same cgroup as the current.
5307 * Therefore attempt to avoid putting and setting the entire cgroup
5308 * hierarchy, only change the part that actually changes.
5312 struct sched_entity *curr = cfs_rq->curr;
5315 * Since we got here without doing put_prev_entity() we also
5316 * have to consider cfs_rq->curr. If it is still a runnable
5317 * entity, update_curr() will update its vruntime, otherwise
5318 * forget we've ever seen it.
5322 update_curr(cfs_rq);
5327 * This call to check_cfs_rq_runtime() will do the
5328 * throttle and dequeue its entity in the parent(s).
5329 * Therefore the 'simple' nr_running test will indeed
5332 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5336 se = pick_next_entity(cfs_rq, curr);
5337 cfs_rq = group_cfs_rq(se);
5343 * Since we haven't yet done put_prev_entity and if the selected task
5344 * is a different task than we started out with, try and touch the
5345 * least amount of cfs_rqs.
5348 struct sched_entity *pse = &prev->se;
5350 while (!(cfs_rq = is_same_group(se, pse))) {
5351 int se_depth = se->depth;
5352 int pse_depth = pse->depth;
5354 if (se_depth <= pse_depth) {
5355 put_prev_entity(cfs_rq_of(pse), pse);
5356 pse = parent_entity(pse);
5358 if (se_depth >= pse_depth) {
5359 set_next_entity(cfs_rq_of(se), se);
5360 se = parent_entity(se);
5364 put_prev_entity(cfs_rq, pse);
5365 set_next_entity(cfs_rq, se);
5368 if (hrtick_enabled(rq))
5369 hrtick_start_fair(rq, p);
5376 if (!cfs_rq->nr_running)
5379 put_prev_task(rq, prev);
5382 se = pick_next_entity(cfs_rq, NULL);
5383 set_next_entity(cfs_rq, se);
5384 cfs_rq = group_cfs_rq(se);
5389 if (hrtick_enabled(rq))
5390 hrtick_start_fair(rq, p);
5395 new_tasks = idle_balance(rq);
5397 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5398 * possible for any higher priority task to appear. In that case we
5399 * must re-start the pick_next_entity() loop.
5411 * Account for a descheduled task:
5413 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5415 struct sched_entity *se = &prev->se;
5416 struct cfs_rq *cfs_rq;
5418 for_each_sched_entity(se) {
5419 cfs_rq = cfs_rq_of(se);
5420 put_prev_entity(cfs_rq, se);
5425 * sched_yield() is very simple
5427 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5429 static void yield_task_fair(struct rq *rq)
5431 struct task_struct *curr = rq->curr;
5432 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5433 struct sched_entity *se = &curr->se;
5436 * Are we the only task in the tree?
5438 if (unlikely(rq->nr_running == 1))
5441 clear_buddies(cfs_rq, se);
5443 if (curr->policy != SCHED_BATCH) {
5444 update_rq_clock(rq);
5446 * Update run-time statistics of the 'current'.
5448 update_curr(cfs_rq);
5450 * Tell update_rq_clock() that we've just updated,
5451 * so we don't do microscopic update in schedule()
5452 * and double the fastpath cost.
5454 rq_clock_skip_update(rq, true);
5460 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5462 struct sched_entity *se = &p->se;
5464 /* throttled hierarchies are not runnable */
5465 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5468 /* Tell the scheduler that we'd really like pse to run next. */
5471 yield_task_fair(rq);
5477 /**************************************************
5478 * Fair scheduling class load-balancing methods.
5482 * The purpose of load-balancing is to achieve the same basic fairness the
5483 * per-cpu scheduler provides, namely provide a proportional amount of compute
5484 * time to each task. This is expressed in the following equation:
5486 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5488 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5489 * W_i,0 is defined as:
5491 * W_i,0 = \Sum_j w_i,j (2)
5493 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5494 * is derived from the nice value as per prio_to_weight[].
5496 * The weight average is an exponential decay average of the instantaneous
5499 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5501 * C_i is the compute capacity of cpu i, typically it is the
5502 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5503 * can also include other factors [XXX].
5505 * To achieve this balance we define a measure of imbalance which follows
5506 * directly from (1):
5508 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5510 * We them move tasks around to minimize the imbalance. In the continuous
5511 * function space it is obvious this converges, in the discrete case we get
5512 * a few fun cases generally called infeasible weight scenarios.
5515 * - infeasible weights;
5516 * - local vs global optima in the discrete case. ]
5521 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5522 * for all i,j solution, we create a tree of cpus that follows the hardware
5523 * topology where each level pairs two lower groups (or better). This results
5524 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5525 * tree to only the first of the previous level and we decrease the frequency
5526 * of load-balance at each level inv. proportional to the number of cpus in
5532 * \Sum { --- * --- * 2^i } = O(n) (5)
5534 * `- size of each group
5535 * | | `- number of cpus doing load-balance
5537 * `- sum over all levels
5539 * Coupled with a limit on how many tasks we can migrate every balance pass,
5540 * this makes (5) the runtime complexity of the balancer.
5542 * An important property here is that each CPU is still (indirectly) connected
5543 * to every other cpu in at most O(log n) steps:
5545 * The adjacency matrix of the resulting graph is given by:
5548 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5551 * And you'll find that:
5553 * A^(log_2 n)_i,j != 0 for all i,j (7)
5555 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5556 * The task movement gives a factor of O(m), giving a convergence complexity
5559 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5564 * In order to avoid CPUs going idle while there's still work to do, new idle
5565 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5566 * tree itself instead of relying on other CPUs to bring it work.
5568 * This adds some complexity to both (5) and (8) but it reduces the total idle
5576 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5579 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5584 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5586 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5588 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5591 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5592 * rewrite all of this once again.]
5595 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5597 enum fbq_type { regular, remote, all };
5599 #define LBF_ALL_PINNED 0x01
5600 #define LBF_NEED_BREAK 0x02
5601 #define LBF_DST_PINNED 0x04
5602 #define LBF_SOME_PINNED 0x08
5605 struct sched_domain *sd;
5613 struct cpumask *dst_grpmask;
5615 enum cpu_idle_type idle;
5617 /* The set of CPUs under consideration for load-balancing */
5618 struct cpumask *cpus;
5623 unsigned int loop_break;
5624 unsigned int loop_max;
5626 enum fbq_type fbq_type;
5627 struct list_head tasks;
5631 * Is this task likely cache-hot:
5633 static int task_hot(struct task_struct *p, struct lb_env *env)
5637 lockdep_assert_held(&env->src_rq->lock);
5639 if (p->sched_class != &fair_sched_class)
5642 if (unlikely(p->policy == SCHED_IDLE))
5646 * Buddy candidates are cache hot:
5648 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5649 (&p->se == cfs_rq_of(&p->se)->next ||
5650 &p->se == cfs_rq_of(&p->se)->last))
5653 if (sysctl_sched_migration_cost == -1)
5655 if (sysctl_sched_migration_cost == 0)
5658 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5660 return delta < (s64)sysctl_sched_migration_cost;
5663 #ifdef CONFIG_NUMA_BALANCING
5665 * Returns true if the destination node is the preferred node.
5666 * Needs to match fbq_classify_rq(): if there is a runnable task
5667 * that is not on its preferred node, we should identify it.
5669 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5671 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5672 unsigned long src_faults, dst_faults;
5673 int src_nid, dst_nid;
5675 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
5676 !(env->sd->flags & SD_NUMA)) {
5680 src_nid = cpu_to_node(env->src_cpu);
5681 dst_nid = cpu_to_node(env->dst_cpu);
5683 if (src_nid == dst_nid)
5686 /* Encourage migration to the preferred node. */
5687 if (dst_nid == p->numa_preferred_nid)
5690 /* Migrating away from the preferred node is bad. */
5691 if (src_nid == p->numa_preferred_nid)
5695 src_faults = group_faults(p, src_nid);
5696 dst_faults = group_faults(p, dst_nid);
5698 src_faults = task_faults(p, src_nid);
5699 dst_faults = task_faults(p, dst_nid);
5702 return dst_faults > src_faults;
5706 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5708 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5709 unsigned long src_faults, dst_faults;
5710 int src_nid, dst_nid;
5712 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5715 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5718 src_nid = cpu_to_node(env->src_cpu);
5719 dst_nid = cpu_to_node(env->dst_cpu);
5721 if (src_nid == dst_nid)
5724 /* Migrating away from the preferred node is bad. */
5725 if (src_nid == p->numa_preferred_nid)
5728 /* Encourage migration to the preferred node. */
5729 if (dst_nid == p->numa_preferred_nid)
5733 src_faults = group_faults(p, src_nid);
5734 dst_faults = group_faults(p, dst_nid);
5736 src_faults = task_faults(p, src_nid);
5737 dst_faults = task_faults(p, dst_nid);
5740 return dst_faults < src_faults;
5744 static inline bool migrate_improves_locality(struct task_struct *p,
5750 static inline bool migrate_degrades_locality(struct task_struct *p,
5758 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5761 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5763 int tsk_cache_hot = 0;
5765 lockdep_assert_held(&env->src_rq->lock);
5768 * We do not migrate tasks that are:
5769 * 1) throttled_lb_pair, or
5770 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5771 * 3) running (obviously), or
5772 * 4) are cache-hot on their current CPU.
5774 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5777 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5780 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5782 env->flags |= LBF_SOME_PINNED;
5785 * Remember if this task can be migrated to any other cpu in
5786 * our sched_group. We may want to revisit it if we couldn't
5787 * meet load balance goals by pulling other tasks on src_cpu.
5789 * Also avoid computing new_dst_cpu if we have already computed
5790 * one in current iteration.
5792 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5795 /* Prevent to re-select dst_cpu via env's cpus */
5796 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5797 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5798 env->flags |= LBF_DST_PINNED;
5799 env->new_dst_cpu = cpu;
5807 /* Record that we found atleast one task that could run on dst_cpu */
5808 env->flags &= ~LBF_ALL_PINNED;
5810 if (task_running(env->src_rq, p)) {
5811 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5816 * Aggressive migration if:
5817 * 1) destination numa is preferred
5818 * 2) task is cache cold, or
5819 * 3) too many balance attempts have failed.
5821 tsk_cache_hot = task_hot(p, env);
5823 tsk_cache_hot = migrate_degrades_locality(p, env);
5825 if (migrate_improves_locality(p, env) || !tsk_cache_hot ||
5826 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5827 if (tsk_cache_hot) {
5828 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5829 schedstat_inc(p, se.statistics.nr_forced_migrations);
5834 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5839 * detach_task() -- detach the task for the migration specified in env
5841 static void detach_task(struct task_struct *p, struct lb_env *env)
5843 lockdep_assert_held(&env->src_rq->lock);
5845 deactivate_task(env->src_rq, p, 0);
5846 p->on_rq = TASK_ON_RQ_MIGRATING;
5847 set_task_cpu(p, env->dst_cpu);
5851 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5852 * part of active balancing operations within "domain".
5854 * Returns a task if successful and NULL otherwise.
5856 static struct task_struct *detach_one_task(struct lb_env *env)
5858 struct task_struct *p, *n;
5860 lockdep_assert_held(&env->src_rq->lock);
5862 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5863 if (!can_migrate_task(p, env))
5866 detach_task(p, env);
5869 * Right now, this is only the second place where
5870 * lb_gained[env->idle] is updated (other is detach_tasks)
5871 * so we can safely collect stats here rather than
5872 * inside detach_tasks().
5874 schedstat_inc(env->sd, lb_gained[env->idle]);
5880 static const unsigned int sched_nr_migrate_break = 32;
5883 * detach_tasks() -- tries to detach up to imbalance weighted load from
5884 * busiest_rq, as part of a balancing operation within domain "sd".
5886 * Returns number of detached tasks if successful and 0 otherwise.
5888 static int detach_tasks(struct lb_env *env)
5890 struct list_head *tasks = &env->src_rq->cfs_tasks;
5891 struct task_struct *p;
5895 lockdep_assert_held(&env->src_rq->lock);
5897 if (env->imbalance <= 0)
5900 while (!list_empty(tasks)) {
5901 p = list_first_entry(tasks, struct task_struct, se.group_node);
5904 /* We've more or less seen every task there is, call it quits */
5905 if (env->loop > env->loop_max)
5908 /* take a breather every nr_migrate tasks */
5909 if (env->loop > env->loop_break) {
5910 env->loop_break += sched_nr_migrate_break;
5911 env->flags |= LBF_NEED_BREAK;
5915 if (!can_migrate_task(p, env))
5918 load = task_h_load(p);
5920 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5923 if ((load / 2) > env->imbalance)
5926 detach_task(p, env);
5927 list_add(&p->se.group_node, &env->tasks);
5930 env->imbalance -= load;
5932 #ifdef CONFIG_PREEMPT
5934 * NEWIDLE balancing is a source of latency, so preemptible
5935 * kernels will stop after the first task is detached to minimize
5936 * the critical section.
5938 if (env->idle == CPU_NEWLY_IDLE)
5943 * We only want to steal up to the prescribed amount of
5946 if (env->imbalance <= 0)
5951 list_move_tail(&p->se.group_node, tasks);
5955 * Right now, this is one of only two places we collect this stat
5956 * so we can safely collect detach_one_task() stats here rather
5957 * than inside detach_one_task().
5959 schedstat_add(env->sd, lb_gained[env->idle], detached);
5965 * attach_task() -- attach the task detached by detach_task() to its new rq.
5967 static void attach_task(struct rq *rq, struct task_struct *p)
5969 lockdep_assert_held(&rq->lock);
5971 BUG_ON(task_rq(p) != rq);
5972 p->on_rq = TASK_ON_RQ_QUEUED;
5973 activate_task(rq, p, 0);
5974 check_preempt_curr(rq, p, 0);
5978 * attach_one_task() -- attaches the task returned from detach_one_task() to
5981 static void attach_one_task(struct rq *rq, struct task_struct *p)
5983 raw_spin_lock(&rq->lock);
5985 raw_spin_unlock(&rq->lock);
5989 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5992 static void attach_tasks(struct lb_env *env)
5994 struct list_head *tasks = &env->tasks;
5995 struct task_struct *p;
5997 raw_spin_lock(&env->dst_rq->lock);
5999 while (!list_empty(tasks)) {
6000 p = list_first_entry(tasks, struct task_struct, se.group_node);
6001 list_del_init(&p->se.group_node);
6003 attach_task(env->dst_rq, p);
6006 raw_spin_unlock(&env->dst_rq->lock);
6009 #ifdef CONFIG_FAIR_GROUP_SCHED
6011 * update tg->load_weight by folding this cpu's load_avg
6013 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
6015 struct sched_entity *se = tg->se[cpu];
6016 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
6018 /* throttled entities do not contribute to load */
6019 if (throttled_hierarchy(cfs_rq))
6022 update_cfs_rq_blocked_load(cfs_rq, 1);
6025 update_entity_load_avg(se, 1);
6027 * We pivot on our runnable average having decayed to zero for
6028 * list removal. This generally implies that all our children
6029 * have also been removed (modulo rounding error or bandwidth
6030 * control); however, such cases are rare and we can fix these
6033 * TODO: fix up out-of-order children on enqueue.
6035 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
6036 list_del_leaf_cfs_rq(cfs_rq);
6038 struct rq *rq = rq_of(cfs_rq);
6039 update_rq_runnable_avg(rq, rq->nr_running);
6043 static void update_blocked_averages(int cpu)
6045 struct rq *rq = cpu_rq(cpu);
6046 struct cfs_rq *cfs_rq;
6047 unsigned long flags;
6049 raw_spin_lock_irqsave(&rq->lock, flags);
6050 update_rq_clock(rq);
6052 * Iterates the task_group tree in a bottom up fashion, see
6053 * list_add_leaf_cfs_rq() for details.
6055 for_each_leaf_cfs_rq(rq, cfs_rq) {
6057 * Note: We may want to consider periodically releasing
6058 * rq->lock about these updates so that creating many task
6059 * groups does not result in continually extending hold time.
6061 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
6064 raw_spin_unlock_irqrestore(&rq->lock, flags);
6068 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6069 * This needs to be done in a top-down fashion because the load of a child
6070 * group is a fraction of its parents load.
6072 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6074 struct rq *rq = rq_of(cfs_rq);
6075 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6076 unsigned long now = jiffies;
6079 if (cfs_rq->last_h_load_update == now)
6082 cfs_rq->h_load_next = NULL;
6083 for_each_sched_entity(se) {
6084 cfs_rq = cfs_rq_of(se);
6085 cfs_rq->h_load_next = se;
6086 if (cfs_rq->last_h_load_update == now)
6091 cfs_rq->h_load = cfs_rq->runnable_load_avg;
6092 cfs_rq->last_h_load_update = now;
6095 while ((se = cfs_rq->h_load_next) != NULL) {
6096 load = cfs_rq->h_load;
6097 load = div64_ul(load * se->avg.load_avg_contrib,
6098 cfs_rq->runnable_load_avg + 1);
6099 cfs_rq = group_cfs_rq(se);
6100 cfs_rq->h_load = load;
6101 cfs_rq->last_h_load_update = now;
6105 static unsigned long task_h_load(struct task_struct *p)
6107 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6109 update_cfs_rq_h_load(cfs_rq);
6110 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
6111 cfs_rq->runnable_load_avg + 1);
6114 static inline void update_blocked_averages(int cpu)
6118 static unsigned long task_h_load(struct task_struct *p)
6120 return p->se.avg.load_avg_contrib;
6124 /********** Helpers for find_busiest_group ************************/
6133 * sg_lb_stats - stats of a sched_group required for load_balancing
6135 struct sg_lb_stats {
6136 unsigned long avg_load; /*Avg load across the CPUs of the group */
6137 unsigned long group_load; /* Total load over the CPUs of the group */
6138 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6139 unsigned long load_per_task;
6140 unsigned long group_capacity;
6141 unsigned long group_usage; /* Total usage of the group */
6142 unsigned int sum_nr_running; /* Nr tasks running in the group */
6143 unsigned int idle_cpus;
6144 unsigned int group_weight;
6145 enum group_type group_type;
6146 int group_no_capacity;
6147 #ifdef CONFIG_NUMA_BALANCING
6148 unsigned int nr_numa_running;
6149 unsigned int nr_preferred_running;
6154 * sd_lb_stats - Structure to store the statistics of a sched_domain
6155 * during load balancing.
6157 struct sd_lb_stats {
6158 struct sched_group *busiest; /* Busiest group in this sd */
6159 struct sched_group *local; /* Local group in this sd */
6160 unsigned long total_load; /* Total load of all groups in sd */
6161 unsigned long total_capacity; /* Total capacity of all groups in sd */
6162 unsigned long avg_load; /* Average load across all groups in sd */
6164 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6165 struct sg_lb_stats local_stat; /* Statistics of the local group */
6168 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6171 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6172 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6173 * We must however clear busiest_stat::avg_load because
6174 * update_sd_pick_busiest() reads this before assignment.
6176 *sds = (struct sd_lb_stats){
6180 .total_capacity = 0UL,
6183 .sum_nr_running = 0,
6184 .group_type = group_other,
6190 * get_sd_load_idx - Obtain the load index for a given sched domain.
6191 * @sd: The sched_domain whose load_idx is to be obtained.
6192 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6194 * Return: The load index.
6196 static inline int get_sd_load_idx(struct sched_domain *sd,
6197 enum cpu_idle_type idle)
6203 load_idx = sd->busy_idx;
6206 case CPU_NEWLY_IDLE:
6207 load_idx = sd->newidle_idx;
6210 load_idx = sd->idle_idx;
6217 static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu)
6219 if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1))
6220 return sd->smt_gain / sd->span_weight;
6222 return SCHED_CAPACITY_SCALE;
6225 unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
6227 return default_scale_cpu_capacity(sd, cpu);
6230 static unsigned long scale_rt_capacity(int cpu)
6232 struct rq *rq = cpu_rq(cpu);
6233 u64 total, used, age_stamp, avg;
6237 * Since we're reading these variables without serialization make sure
6238 * we read them once before doing sanity checks on them.
6240 age_stamp = READ_ONCE(rq->age_stamp);
6241 avg = READ_ONCE(rq->rt_avg);
6242 delta = __rq_clock_broken(rq) - age_stamp;
6244 if (unlikely(delta < 0))
6247 total = sched_avg_period() + delta;
6249 used = div_u64(avg, total);
6251 if (likely(used < SCHED_CAPACITY_SCALE))
6252 return SCHED_CAPACITY_SCALE - used;
6257 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6259 unsigned long capacity = SCHED_CAPACITY_SCALE;
6260 struct sched_group *sdg = sd->groups;
6262 if (sched_feat(ARCH_CAPACITY))
6263 capacity *= arch_scale_cpu_capacity(sd, cpu);
6265 capacity *= default_scale_cpu_capacity(sd, cpu);
6267 capacity >>= SCHED_CAPACITY_SHIFT;
6269 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6271 capacity *= scale_rt_capacity(cpu);
6272 capacity >>= SCHED_CAPACITY_SHIFT;
6277 cpu_rq(cpu)->cpu_capacity = capacity;
6278 sdg->sgc->capacity = capacity;
6281 void update_group_capacity(struct sched_domain *sd, int cpu)
6283 struct sched_domain *child = sd->child;
6284 struct sched_group *group, *sdg = sd->groups;
6285 unsigned long capacity;
6286 unsigned long interval;
6288 interval = msecs_to_jiffies(sd->balance_interval);
6289 interval = clamp(interval, 1UL, max_load_balance_interval);
6290 sdg->sgc->next_update = jiffies + interval;
6293 update_cpu_capacity(sd, cpu);
6299 if (child->flags & SD_OVERLAP) {
6301 * SD_OVERLAP domains cannot assume that child groups
6302 * span the current group.
6305 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6306 struct sched_group_capacity *sgc;
6307 struct rq *rq = cpu_rq(cpu);
6310 * build_sched_domains() -> init_sched_groups_capacity()
6311 * gets here before we've attached the domains to the
6314 * Use capacity_of(), which is set irrespective of domains
6315 * in update_cpu_capacity().
6317 * This avoids capacity from being 0 and
6318 * causing divide-by-zero issues on boot.
6320 if (unlikely(!rq->sd)) {
6321 capacity += capacity_of(cpu);
6325 sgc = rq->sd->groups->sgc;
6326 capacity += sgc->capacity;
6330 * !SD_OVERLAP domains can assume that child groups
6331 * span the current group.
6334 group = child->groups;
6336 capacity += group->sgc->capacity;
6337 group = group->next;
6338 } while (group != child->groups);
6341 sdg->sgc->capacity = capacity;
6345 * Check whether the capacity of the rq has been noticeably reduced by side
6346 * activity. The imbalance_pct is used for the threshold.
6347 * Return true is the capacity is reduced
6350 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6352 return ((rq->cpu_capacity * sd->imbalance_pct) <
6353 (rq->cpu_capacity_orig * 100));
6357 * Group imbalance indicates (and tries to solve) the problem where balancing
6358 * groups is inadequate due to tsk_cpus_allowed() constraints.
6360 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6361 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6364 * { 0 1 2 3 } { 4 5 6 7 }
6367 * If we were to balance group-wise we'd place two tasks in the first group and
6368 * two tasks in the second group. Clearly this is undesired as it will overload
6369 * cpu 3 and leave one of the cpus in the second group unused.
6371 * The current solution to this issue is detecting the skew in the first group
6372 * by noticing the lower domain failed to reach balance and had difficulty
6373 * moving tasks due to affinity constraints.
6375 * When this is so detected; this group becomes a candidate for busiest; see
6376 * update_sd_pick_busiest(). And calculate_imbalance() and
6377 * find_busiest_group() avoid some of the usual balance conditions to allow it
6378 * to create an effective group imbalance.
6380 * This is a somewhat tricky proposition since the next run might not find the
6381 * group imbalance and decide the groups need to be balanced again. A most
6382 * subtle and fragile situation.
6385 static inline int sg_imbalanced(struct sched_group *group)
6387 return group->sgc->imbalance;
6391 * group_has_capacity returns true if the group has spare capacity that could
6392 * be used by some tasks.
6393 * We consider that a group has spare capacity if the * number of task is
6394 * smaller than the number of CPUs or if the usage is lower than the available
6395 * capacity for CFS tasks.
6396 * For the latter, we use a threshold to stabilize the state, to take into
6397 * account the variance of the tasks' load and to return true if the available
6398 * capacity in meaningful for the load balancer.
6399 * As an example, an available capacity of 1% can appear but it doesn't make
6400 * any benefit for the load balance.
6403 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6405 if (sgs->sum_nr_running < sgs->group_weight)
6408 if ((sgs->group_capacity * 100) >
6409 (sgs->group_usage * env->sd->imbalance_pct))
6416 * group_is_overloaded returns true if the group has more tasks than it can
6418 * group_is_overloaded is not equals to !group_has_capacity because a group
6419 * with the exact right number of tasks, has no more spare capacity but is not
6420 * overloaded so both group_has_capacity and group_is_overloaded return
6424 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6426 if (sgs->sum_nr_running <= sgs->group_weight)
6429 if ((sgs->group_capacity * 100) <
6430 (sgs->group_usage * env->sd->imbalance_pct))
6436 static enum group_type group_classify(struct lb_env *env,
6437 struct sched_group *group,
6438 struct sg_lb_stats *sgs)
6440 if (sgs->group_no_capacity)
6441 return group_overloaded;
6443 if (sg_imbalanced(group))
6444 return group_imbalanced;
6450 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6451 * @env: The load balancing environment.
6452 * @group: sched_group whose statistics are to be updated.
6453 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6454 * @local_group: Does group contain this_cpu.
6455 * @sgs: variable to hold the statistics for this group.
6456 * @overload: Indicate more than one runnable task for any CPU.
6458 static inline void update_sg_lb_stats(struct lb_env *env,
6459 struct sched_group *group, int load_idx,
6460 int local_group, struct sg_lb_stats *sgs,
6466 memset(sgs, 0, sizeof(*sgs));
6468 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6469 struct rq *rq = cpu_rq(i);
6471 /* Bias balancing toward cpus of our domain */
6473 load = target_load(i, load_idx);
6475 load = source_load(i, load_idx);
6477 sgs->group_load += load;
6478 sgs->group_usage += get_cpu_usage(i);
6479 sgs->sum_nr_running += rq->cfs.h_nr_running;
6481 if (rq->nr_running > 1)
6484 #ifdef CONFIG_NUMA_BALANCING
6485 sgs->nr_numa_running += rq->nr_numa_running;
6486 sgs->nr_preferred_running += rq->nr_preferred_running;
6488 sgs->sum_weighted_load += weighted_cpuload(i);
6493 /* Adjust by relative CPU capacity of the group */
6494 sgs->group_capacity = group->sgc->capacity;
6495 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6497 if (sgs->sum_nr_running)
6498 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6500 sgs->group_weight = group->group_weight;
6502 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6503 sgs->group_type = group_classify(env, group, sgs);
6507 * update_sd_pick_busiest - return 1 on busiest group
6508 * @env: The load balancing environment.
6509 * @sds: sched_domain statistics
6510 * @sg: sched_group candidate to be checked for being the busiest
6511 * @sgs: sched_group statistics
6513 * Determine if @sg is a busier group than the previously selected
6516 * Return: %true if @sg is a busier group than the previously selected
6517 * busiest group. %false otherwise.
6519 static bool update_sd_pick_busiest(struct lb_env *env,
6520 struct sd_lb_stats *sds,
6521 struct sched_group *sg,
6522 struct sg_lb_stats *sgs)
6524 struct sg_lb_stats *busiest = &sds->busiest_stat;
6526 if (sgs->group_type > busiest->group_type)
6529 if (sgs->group_type < busiest->group_type)
6532 if (sgs->avg_load <= busiest->avg_load)
6535 /* This is the busiest node in its class. */
6536 if (!(env->sd->flags & SD_ASYM_PACKING))
6540 * ASYM_PACKING needs to move all the work to the lowest
6541 * numbered CPUs in the group, therefore mark all groups
6542 * higher than ourself as busy.
6544 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6548 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6555 #ifdef CONFIG_NUMA_BALANCING
6556 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6558 if (sgs->sum_nr_running > sgs->nr_numa_running)
6560 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6565 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6567 if (rq->nr_running > rq->nr_numa_running)
6569 if (rq->nr_running > rq->nr_preferred_running)
6574 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6579 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6583 #endif /* CONFIG_NUMA_BALANCING */
6586 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6587 * @env: The load balancing environment.
6588 * @sds: variable to hold the statistics for this sched_domain.
6590 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6592 struct sched_domain *child = env->sd->child;
6593 struct sched_group *sg = env->sd->groups;
6594 struct sg_lb_stats tmp_sgs;
6595 int load_idx, prefer_sibling = 0;
6596 bool overload = false;
6598 if (child && child->flags & SD_PREFER_SIBLING)
6601 load_idx = get_sd_load_idx(env->sd, env->idle);
6604 struct sg_lb_stats *sgs = &tmp_sgs;
6607 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6610 sgs = &sds->local_stat;
6612 if (env->idle != CPU_NEWLY_IDLE ||
6613 time_after_eq(jiffies, sg->sgc->next_update))
6614 update_group_capacity(env->sd, env->dst_cpu);
6617 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6624 * In case the child domain prefers tasks go to siblings
6625 * first, lower the sg capacity so that we'll try
6626 * and move all the excess tasks away. We lower the capacity
6627 * of a group only if the local group has the capacity to fit
6628 * these excess tasks. The extra check prevents the case where
6629 * you always pull from the heaviest group when it is already
6630 * under-utilized (possible with a large weight task outweighs
6631 * the tasks on the system).
6633 if (prefer_sibling && sds->local &&
6634 group_has_capacity(env, &sds->local_stat) &&
6635 (sgs->sum_nr_running > 1)) {
6636 sgs->group_no_capacity = 1;
6637 sgs->group_type = group_overloaded;
6640 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6642 sds->busiest_stat = *sgs;
6646 /* Now, start updating sd_lb_stats */
6647 sds->total_load += sgs->group_load;
6648 sds->total_capacity += sgs->group_capacity;
6651 } while (sg != env->sd->groups);
6653 if (env->sd->flags & SD_NUMA)
6654 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6656 if (!env->sd->parent) {
6657 /* update overload indicator if we are at root domain */
6658 if (env->dst_rq->rd->overload != overload)
6659 env->dst_rq->rd->overload = overload;
6665 * check_asym_packing - Check to see if the group is packed into the
6668 * This is primarily intended to used at the sibling level. Some
6669 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6670 * case of POWER7, it can move to lower SMT modes only when higher
6671 * threads are idle. When in lower SMT modes, the threads will
6672 * perform better since they share less core resources. Hence when we
6673 * have idle threads, we want them to be the higher ones.
6675 * This packing function is run on idle threads. It checks to see if
6676 * the busiest CPU in this domain (core in the P7 case) has a higher
6677 * CPU number than the packing function is being run on. Here we are
6678 * assuming lower CPU number will be equivalent to lower a SMT thread
6681 * Return: 1 when packing is required and a task should be moved to
6682 * this CPU. The amount of the imbalance is returned in *imbalance.
6684 * @env: The load balancing environment.
6685 * @sds: Statistics of the sched_domain which is to be packed
6687 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6691 if (!(env->sd->flags & SD_ASYM_PACKING))
6697 busiest_cpu = group_first_cpu(sds->busiest);
6698 if (env->dst_cpu > busiest_cpu)
6701 env->imbalance = DIV_ROUND_CLOSEST(
6702 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6703 SCHED_CAPACITY_SCALE);
6709 * fix_small_imbalance - Calculate the minor imbalance that exists
6710 * amongst the groups of a sched_domain, during
6712 * @env: The load balancing environment.
6713 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6716 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6718 unsigned long tmp, capa_now = 0, capa_move = 0;
6719 unsigned int imbn = 2;
6720 unsigned long scaled_busy_load_per_task;
6721 struct sg_lb_stats *local, *busiest;
6723 local = &sds->local_stat;
6724 busiest = &sds->busiest_stat;
6726 if (!local->sum_nr_running)
6727 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6728 else if (busiest->load_per_task > local->load_per_task)
6731 scaled_busy_load_per_task =
6732 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6733 busiest->group_capacity;
6735 if (busiest->avg_load + scaled_busy_load_per_task >=
6736 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6737 env->imbalance = busiest->load_per_task;
6742 * OK, we don't have enough imbalance to justify moving tasks,
6743 * however we may be able to increase total CPU capacity used by
6747 capa_now += busiest->group_capacity *
6748 min(busiest->load_per_task, busiest->avg_load);
6749 capa_now += local->group_capacity *
6750 min(local->load_per_task, local->avg_load);
6751 capa_now /= SCHED_CAPACITY_SCALE;
6753 /* Amount of load we'd subtract */
6754 if (busiest->avg_load > scaled_busy_load_per_task) {
6755 capa_move += busiest->group_capacity *
6756 min(busiest->load_per_task,
6757 busiest->avg_load - scaled_busy_load_per_task);
6760 /* Amount of load we'd add */
6761 if (busiest->avg_load * busiest->group_capacity <
6762 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6763 tmp = (busiest->avg_load * busiest->group_capacity) /
6764 local->group_capacity;
6766 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6767 local->group_capacity;
6769 capa_move += local->group_capacity *
6770 min(local->load_per_task, local->avg_load + tmp);
6771 capa_move /= SCHED_CAPACITY_SCALE;
6773 /* Move if we gain throughput */
6774 if (capa_move > capa_now)
6775 env->imbalance = busiest->load_per_task;
6779 * calculate_imbalance - Calculate the amount of imbalance present within the
6780 * groups of a given sched_domain during load balance.
6781 * @env: load balance environment
6782 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6784 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6786 unsigned long max_pull, load_above_capacity = ~0UL;
6787 struct sg_lb_stats *local, *busiest;
6789 local = &sds->local_stat;
6790 busiest = &sds->busiest_stat;
6792 if (busiest->group_type == group_imbalanced) {
6794 * In the group_imb case we cannot rely on group-wide averages
6795 * to ensure cpu-load equilibrium, look at wider averages. XXX
6797 busiest->load_per_task =
6798 min(busiest->load_per_task, sds->avg_load);
6802 * In the presence of smp nice balancing, certain scenarios can have
6803 * max load less than avg load(as we skip the groups at or below
6804 * its cpu_capacity, while calculating max_load..)
6806 if (busiest->avg_load <= sds->avg_load ||
6807 local->avg_load >= sds->avg_load) {
6809 return fix_small_imbalance(env, sds);
6813 * If there aren't any idle cpus, avoid creating some.
6815 if (busiest->group_type == group_overloaded &&
6816 local->group_type == group_overloaded) {
6817 load_above_capacity = busiest->sum_nr_running *
6819 if (load_above_capacity > busiest->group_capacity)
6820 load_above_capacity -= busiest->group_capacity;
6822 load_above_capacity = ~0UL;
6826 * We're trying to get all the cpus to the average_load, so we don't
6827 * want to push ourselves above the average load, nor do we wish to
6828 * reduce the max loaded cpu below the average load. At the same time,
6829 * we also don't want to reduce the group load below the group capacity
6830 * (so that we can implement power-savings policies etc). Thus we look
6831 * for the minimum possible imbalance.
6833 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6835 /* How much load to actually move to equalise the imbalance */
6836 env->imbalance = min(
6837 max_pull * busiest->group_capacity,
6838 (sds->avg_load - local->avg_load) * local->group_capacity
6839 ) / SCHED_CAPACITY_SCALE;
6842 * if *imbalance is less than the average load per runnable task
6843 * there is no guarantee that any tasks will be moved so we'll have
6844 * a think about bumping its value to force at least one task to be
6847 if (env->imbalance < busiest->load_per_task)
6848 return fix_small_imbalance(env, sds);
6851 /******* find_busiest_group() helpers end here *********************/
6854 * find_busiest_group - Returns the busiest group within the sched_domain
6855 * if there is an imbalance. If there isn't an imbalance, and
6856 * the user has opted for power-savings, it returns a group whose
6857 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6858 * such a group exists.
6860 * Also calculates the amount of weighted load which should be moved
6861 * to restore balance.
6863 * @env: The load balancing environment.
6865 * Return: - The busiest group if imbalance exists.
6866 * - If no imbalance and user has opted for power-savings balance,
6867 * return the least loaded group whose CPUs can be
6868 * put to idle by rebalancing its tasks onto our group.
6870 static struct sched_group *find_busiest_group(struct lb_env *env)
6872 struct sg_lb_stats *local, *busiest;
6873 struct sd_lb_stats sds;
6875 init_sd_lb_stats(&sds);
6878 * Compute the various statistics relavent for load balancing at
6881 update_sd_lb_stats(env, &sds);
6882 local = &sds.local_stat;
6883 busiest = &sds.busiest_stat;
6885 /* ASYM feature bypasses nice load balance check */
6886 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6887 check_asym_packing(env, &sds))
6890 /* There is no busy sibling group to pull tasks from */
6891 if (!sds.busiest || busiest->sum_nr_running == 0)
6894 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6895 / sds.total_capacity;
6898 * If the busiest group is imbalanced the below checks don't
6899 * work because they assume all things are equal, which typically
6900 * isn't true due to cpus_allowed constraints and the like.
6902 if (busiest->group_type == group_imbalanced)
6905 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6906 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6907 busiest->group_no_capacity)
6911 * If the local group is busier than the selected busiest group
6912 * don't try and pull any tasks.
6914 if (local->avg_load >= busiest->avg_load)
6918 * Don't pull any tasks if this group is already above the domain
6921 if (local->avg_load >= sds.avg_load)
6924 if (env->idle == CPU_IDLE) {
6926 * This cpu is idle. If the busiest group is not overloaded
6927 * and there is no imbalance between this and busiest group
6928 * wrt idle cpus, it is balanced. The imbalance becomes
6929 * significant if the diff is greater than 1 otherwise we
6930 * might end up to just move the imbalance on another group
6932 if ((busiest->group_type != group_overloaded) &&
6933 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6937 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6938 * imbalance_pct to be conservative.
6940 if (100 * busiest->avg_load <=
6941 env->sd->imbalance_pct * local->avg_load)
6946 /* Looks like there is an imbalance. Compute it */
6947 calculate_imbalance(env, &sds);
6956 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6958 static struct rq *find_busiest_queue(struct lb_env *env,
6959 struct sched_group *group)
6961 struct rq *busiest = NULL, *rq;
6962 unsigned long busiest_load = 0, busiest_capacity = 1;
6965 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6966 unsigned long capacity, wl;
6970 rt = fbq_classify_rq(rq);
6973 * We classify groups/runqueues into three groups:
6974 * - regular: there are !numa tasks
6975 * - remote: there are numa tasks that run on the 'wrong' node
6976 * - all: there is no distinction
6978 * In order to avoid migrating ideally placed numa tasks,
6979 * ignore those when there's better options.
6981 * If we ignore the actual busiest queue to migrate another
6982 * task, the next balance pass can still reduce the busiest
6983 * queue by moving tasks around inside the node.
6985 * If we cannot move enough load due to this classification
6986 * the next pass will adjust the group classification and
6987 * allow migration of more tasks.
6989 * Both cases only affect the total convergence complexity.
6991 if (rt > env->fbq_type)
6994 capacity = capacity_of(i);
6996 wl = weighted_cpuload(i);
6999 * When comparing with imbalance, use weighted_cpuload()
7000 * which is not scaled with the cpu capacity.
7003 if (rq->nr_running == 1 && wl > env->imbalance &&
7004 !check_cpu_capacity(rq, env->sd))
7008 * For the load comparisons with the other cpu's, consider
7009 * the weighted_cpuload() scaled with the cpu capacity, so
7010 * that the load can be moved away from the cpu that is
7011 * potentially running at a lower capacity.
7013 * Thus we're looking for max(wl_i / capacity_i), crosswise
7014 * multiplication to rid ourselves of the division works out
7015 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7016 * our previous maximum.
7018 if (wl * busiest_capacity > busiest_load * capacity) {
7020 busiest_capacity = capacity;
7029 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7030 * so long as it is large enough.
7032 #define MAX_PINNED_INTERVAL 512
7034 /* Working cpumask for load_balance and load_balance_newidle. */
7035 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7037 static int need_active_balance(struct lb_env *env)
7039 struct sched_domain *sd = env->sd;
7041 if (env->idle == CPU_NEWLY_IDLE) {
7044 * ASYM_PACKING needs to force migrate tasks from busy but
7045 * higher numbered CPUs in order to pack all tasks in the
7046 * lowest numbered CPUs.
7048 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7053 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7054 * It's worth migrating the task if the src_cpu's capacity is reduced
7055 * because of other sched_class or IRQs if more capacity stays
7056 * available on dst_cpu.
7058 if ((env->idle != CPU_NOT_IDLE) &&
7059 (env->src_rq->cfs.h_nr_running == 1)) {
7060 if ((check_cpu_capacity(env->src_rq, sd)) &&
7061 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7065 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7068 static int active_load_balance_cpu_stop(void *data);
7070 static int should_we_balance(struct lb_env *env)
7072 struct sched_group *sg = env->sd->groups;
7073 struct cpumask *sg_cpus, *sg_mask;
7074 int cpu, balance_cpu = -1;
7077 * In the newly idle case, we will allow all the cpu's
7078 * to do the newly idle load balance.
7080 if (env->idle == CPU_NEWLY_IDLE)
7083 sg_cpus = sched_group_cpus(sg);
7084 sg_mask = sched_group_mask(sg);
7085 /* Try to find first idle cpu */
7086 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7087 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7094 if (balance_cpu == -1)
7095 balance_cpu = group_balance_cpu(sg);
7098 * First idle cpu or the first cpu(busiest) in this sched group
7099 * is eligible for doing load balancing at this and above domains.
7101 return balance_cpu == env->dst_cpu;
7105 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7106 * tasks if there is an imbalance.
7108 static int load_balance(int this_cpu, struct rq *this_rq,
7109 struct sched_domain *sd, enum cpu_idle_type idle,
7110 int *continue_balancing)
7112 int ld_moved, cur_ld_moved, active_balance = 0;
7113 struct sched_domain *sd_parent = sd->parent;
7114 struct sched_group *group;
7116 unsigned long flags;
7117 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7119 struct lb_env env = {
7121 .dst_cpu = this_cpu,
7123 .dst_grpmask = sched_group_cpus(sd->groups),
7125 .loop_break = sched_nr_migrate_break,
7128 .tasks = LIST_HEAD_INIT(env.tasks),
7132 * For NEWLY_IDLE load_balancing, we don't need to consider
7133 * other cpus in our group
7135 if (idle == CPU_NEWLY_IDLE)
7136 env.dst_grpmask = NULL;
7138 cpumask_copy(cpus, cpu_active_mask);
7140 schedstat_inc(sd, lb_count[idle]);
7143 if (!should_we_balance(&env)) {
7144 *continue_balancing = 0;
7148 group = find_busiest_group(&env);
7150 schedstat_inc(sd, lb_nobusyg[idle]);
7154 busiest = find_busiest_queue(&env, group);
7156 schedstat_inc(sd, lb_nobusyq[idle]);
7160 BUG_ON(busiest == env.dst_rq);
7162 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7164 env.src_cpu = busiest->cpu;
7165 env.src_rq = busiest;
7168 if (busiest->nr_running > 1) {
7170 * Attempt to move tasks. If find_busiest_group has found
7171 * an imbalance but busiest->nr_running <= 1, the group is
7172 * still unbalanced. ld_moved simply stays zero, so it is
7173 * correctly treated as an imbalance.
7175 env.flags |= LBF_ALL_PINNED;
7176 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7179 raw_spin_lock_irqsave(&busiest->lock, flags);
7182 * cur_ld_moved - load moved in current iteration
7183 * ld_moved - cumulative load moved across iterations
7185 cur_ld_moved = detach_tasks(&env);
7188 * We've detached some tasks from busiest_rq. Every
7189 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7190 * unlock busiest->lock, and we are able to be sure
7191 * that nobody can manipulate the tasks in parallel.
7192 * See task_rq_lock() family for the details.
7195 raw_spin_unlock(&busiest->lock);
7199 ld_moved += cur_ld_moved;
7202 local_irq_restore(flags);
7204 if (env.flags & LBF_NEED_BREAK) {
7205 env.flags &= ~LBF_NEED_BREAK;
7210 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7211 * us and move them to an alternate dst_cpu in our sched_group
7212 * where they can run. The upper limit on how many times we
7213 * iterate on same src_cpu is dependent on number of cpus in our
7216 * This changes load balance semantics a bit on who can move
7217 * load to a given_cpu. In addition to the given_cpu itself
7218 * (or a ilb_cpu acting on its behalf where given_cpu is
7219 * nohz-idle), we now have balance_cpu in a position to move
7220 * load to given_cpu. In rare situations, this may cause
7221 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7222 * _independently_ and at _same_ time to move some load to
7223 * given_cpu) causing exceess load to be moved to given_cpu.
7224 * This however should not happen so much in practice and
7225 * moreover subsequent load balance cycles should correct the
7226 * excess load moved.
7228 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7230 /* Prevent to re-select dst_cpu via env's cpus */
7231 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7233 env.dst_rq = cpu_rq(env.new_dst_cpu);
7234 env.dst_cpu = env.new_dst_cpu;
7235 env.flags &= ~LBF_DST_PINNED;
7237 env.loop_break = sched_nr_migrate_break;
7240 * Go back to "more_balance" rather than "redo" since we
7241 * need to continue with same src_cpu.
7247 * We failed to reach balance because of affinity.
7250 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7252 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7253 *group_imbalance = 1;
7256 /* All tasks on this runqueue were pinned by CPU affinity */
7257 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7258 cpumask_clear_cpu(cpu_of(busiest), cpus);
7259 if (!cpumask_empty(cpus)) {
7261 env.loop_break = sched_nr_migrate_break;
7264 goto out_all_pinned;
7269 schedstat_inc(sd, lb_failed[idle]);
7271 * Increment the failure counter only on periodic balance.
7272 * We do not want newidle balance, which can be very
7273 * frequent, pollute the failure counter causing
7274 * excessive cache_hot migrations and active balances.
7276 if (idle != CPU_NEWLY_IDLE)
7277 sd->nr_balance_failed++;
7279 if (need_active_balance(&env)) {
7280 raw_spin_lock_irqsave(&busiest->lock, flags);
7282 /* don't kick the active_load_balance_cpu_stop,
7283 * if the curr task on busiest cpu can't be
7286 if (!cpumask_test_cpu(this_cpu,
7287 tsk_cpus_allowed(busiest->curr))) {
7288 raw_spin_unlock_irqrestore(&busiest->lock,
7290 env.flags |= LBF_ALL_PINNED;
7291 goto out_one_pinned;
7295 * ->active_balance synchronizes accesses to
7296 * ->active_balance_work. Once set, it's cleared
7297 * only after active load balance is finished.
7299 if (!busiest->active_balance) {
7300 busiest->active_balance = 1;
7301 busiest->push_cpu = this_cpu;
7304 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7306 if (active_balance) {
7307 stop_one_cpu_nowait(cpu_of(busiest),
7308 active_load_balance_cpu_stop, busiest,
7309 &busiest->active_balance_work);
7313 * We've kicked active balancing, reset the failure
7316 sd->nr_balance_failed = sd->cache_nice_tries+1;
7319 sd->nr_balance_failed = 0;
7321 if (likely(!active_balance)) {
7322 /* We were unbalanced, so reset the balancing interval */
7323 sd->balance_interval = sd->min_interval;
7326 * If we've begun active balancing, start to back off. This
7327 * case may not be covered by the all_pinned logic if there
7328 * is only 1 task on the busy runqueue (because we don't call
7331 if (sd->balance_interval < sd->max_interval)
7332 sd->balance_interval *= 2;
7339 * We reach balance although we may have faced some affinity
7340 * constraints. Clear the imbalance flag if it was set.
7343 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7345 if (*group_imbalance)
7346 *group_imbalance = 0;
7351 * We reach balance because all tasks are pinned at this level so
7352 * we can't migrate them. Let the imbalance flag set so parent level
7353 * can try to migrate them.
7355 schedstat_inc(sd, lb_balanced[idle]);
7357 sd->nr_balance_failed = 0;
7360 /* tune up the balancing interval */
7361 if (((env.flags & LBF_ALL_PINNED) &&
7362 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7363 (sd->balance_interval < sd->max_interval))
7364 sd->balance_interval *= 2;
7371 static inline unsigned long
7372 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7374 unsigned long interval = sd->balance_interval;
7377 interval *= sd->busy_factor;
7379 /* scale ms to jiffies */
7380 interval = msecs_to_jiffies(interval);
7381 interval = clamp(interval, 1UL, max_load_balance_interval);
7387 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7389 unsigned long interval, next;
7391 interval = get_sd_balance_interval(sd, cpu_busy);
7392 next = sd->last_balance + interval;
7394 if (time_after(*next_balance, next))
7395 *next_balance = next;
7399 * idle_balance is called by schedule() if this_cpu is about to become
7400 * idle. Attempts to pull tasks from other CPUs.
7402 static int idle_balance(struct rq *this_rq)
7404 unsigned long next_balance = jiffies + HZ;
7405 int this_cpu = this_rq->cpu;
7406 struct sched_domain *sd;
7407 int pulled_task = 0;
7410 idle_enter_fair(this_rq);
7413 * We must set idle_stamp _before_ calling idle_balance(), such that we
7414 * measure the duration of idle_balance() as idle time.
7416 this_rq->idle_stamp = rq_clock(this_rq);
7418 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7419 !this_rq->rd->overload) {
7421 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7423 update_next_balance(sd, 0, &next_balance);
7430 * Drop the rq->lock, but keep IRQ/preempt disabled.
7432 raw_spin_unlock(&this_rq->lock);
7434 update_blocked_averages(this_cpu);
7436 for_each_domain(this_cpu, sd) {
7437 int continue_balancing = 1;
7438 u64 t0, domain_cost;
7440 if (!(sd->flags & SD_LOAD_BALANCE))
7443 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7444 update_next_balance(sd, 0, &next_balance);
7448 if (sd->flags & SD_BALANCE_NEWIDLE) {
7449 t0 = sched_clock_cpu(this_cpu);
7451 pulled_task = load_balance(this_cpu, this_rq,
7453 &continue_balancing);
7455 domain_cost = sched_clock_cpu(this_cpu) - t0;
7456 if (domain_cost > sd->max_newidle_lb_cost)
7457 sd->max_newidle_lb_cost = domain_cost;
7459 curr_cost += domain_cost;
7462 update_next_balance(sd, 0, &next_balance);
7465 * Stop searching for tasks to pull if there are
7466 * now runnable tasks on this rq.
7468 if (pulled_task || this_rq->nr_running > 0)
7473 raw_spin_lock(&this_rq->lock);
7475 if (curr_cost > this_rq->max_idle_balance_cost)
7476 this_rq->max_idle_balance_cost = curr_cost;
7479 * While browsing the domains, we released the rq lock, a task could
7480 * have been enqueued in the meantime. Since we're not going idle,
7481 * pretend we pulled a task.
7483 if (this_rq->cfs.h_nr_running && !pulled_task)
7487 /* Move the next balance forward */
7488 if (time_after(this_rq->next_balance, next_balance))
7489 this_rq->next_balance = next_balance;
7491 /* Is there a task of a high priority class? */
7492 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7496 idle_exit_fair(this_rq);
7497 this_rq->idle_stamp = 0;
7504 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7505 * running tasks off the busiest CPU onto idle CPUs. It requires at
7506 * least 1 task to be running on each physical CPU where possible, and
7507 * avoids physical / logical imbalances.
7509 static int active_load_balance_cpu_stop(void *data)
7511 struct rq *busiest_rq = data;
7512 int busiest_cpu = cpu_of(busiest_rq);
7513 int target_cpu = busiest_rq->push_cpu;
7514 struct rq *target_rq = cpu_rq(target_cpu);
7515 struct sched_domain *sd;
7516 struct task_struct *p = NULL;
7518 raw_spin_lock_irq(&busiest_rq->lock);
7520 /* make sure the requested cpu hasn't gone down in the meantime */
7521 if (unlikely(busiest_cpu != smp_processor_id() ||
7522 !busiest_rq->active_balance))
7525 /* Is there any task to move? */
7526 if (busiest_rq->nr_running <= 1)
7530 * This condition is "impossible", if it occurs
7531 * we need to fix it. Originally reported by
7532 * Bjorn Helgaas on a 128-cpu setup.
7534 BUG_ON(busiest_rq == target_rq);
7536 /* Search for an sd spanning us and the target CPU. */
7538 for_each_domain(target_cpu, sd) {
7539 if ((sd->flags & SD_LOAD_BALANCE) &&
7540 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7545 struct lb_env env = {
7547 .dst_cpu = target_cpu,
7548 .dst_rq = target_rq,
7549 .src_cpu = busiest_rq->cpu,
7550 .src_rq = busiest_rq,
7554 schedstat_inc(sd, alb_count);
7556 p = detach_one_task(&env);
7558 schedstat_inc(sd, alb_pushed);
7560 schedstat_inc(sd, alb_failed);
7564 busiest_rq->active_balance = 0;
7565 raw_spin_unlock(&busiest_rq->lock);
7568 attach_one_task(target_rq, p);
7575 static inline int on_null_domain(struct rq *rq)
7577 return unlikely(!rcu_dereference_sched(rq->sd));
7580 #ifdef CONFIG_NO_HZ_COMMON
7582 * idle load balancing details
7583 * - When one of the busy CPUs notice that there may be an idle rebalancing
7584 * needed, they will kick the idle load balancer, which then does idle
7585 * load balancing for all the idle CPUs.
7588 cpumask_var_t idle_cpus_mask;
7590 unsigned long next_balance; /* in jiffy units */
7591 } nohz ____cacheline_aligned;
7593 static inline int find_new_ilb(void)
7595 int ilb = cpumask_first(nohz.idle_cpus_mask);
7597 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7604 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7605 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7606 * CPU (if there is one).
7608 static void nohz_balancer_kick(void)
7612 nohz.next_balance++;
7614 ilb_cpu = find_new_ilb();
7616 if (ilb_cpu >= nr_cpu_ids)
7619 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7622 * Use smp_send_reschedule() instead of resched_cpu().
7623 * This way we generate a sched IPI on the target cpu which
7624 * is idle. And the softirq performing nohz idle load balance
7625 * will be run before returning from the IPI.
7627 smp_send_reschedule(ilb_cpu);
7631 static inline void nohz_balance_exit_idle(int cpu)
7633 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7635 * Completely isolated CPUs don't ever set, so we must test.
7637 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7638 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7639 atomic_dec(&nohz.nr_cpus);
7641 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7645 static inline void set_cpu_sd_state_busy(void)
7647 struct sched_domain *sd;
7648 int cpu = smp_processor_id();
7651 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7653 if (!sd || !sd->nohz_idle)
7657 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7662 void set_cpu_sd_state_idle(void)
7664 struct sched_domain *sd;
7665 int cpu = smp_processor_id();
7668 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7670 if (!sd || sd->nohz_idle)
7674 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7680 * This routine will record that the cpu is going idle with tick stopped.
7681 * This info will be used in performing idle load balancing in the future.
7683 void nohz_balance_enter_idle(int cpu)
7686 * If this cpu is going down, then nothing needs to be done.
7688 if (!cpu_active(cpu))
7691 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7695 * If we're a completely isolated CPU, we don't play.
7697 if (on_null_domain(cpu_rq(cpu)))
7700 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7701 atomic_inc(&nohz.nr_cpus);
7702 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7705 static int sched_ilb_notifier(struct notifier_block *nfb,
7706 unsigned long action, void *hcpu)
7708 switch (action & ~CPU_TASKS_FROZEN) {
7710 nohz_balance_exit_idle(smp_processor_id());
7718 static DEFINE_SPINLOCK(balancing);
7721 * Scale the max load_balance interval with the number of CPUs in the system.
7722 * This trades load-balance latency on larger machines for less cross talk.
7724 void update_max_interval(void)
7726 max_load_balance_interval = HZ*num_online_cpus()/10;
7730 * It checks each scheduling domain to see if it is due to be balanced,
7731 * and initiates a balancing operation if so.
7733 * Balancing parameters are set up in init_sched_domains.
7735 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7737 int continue_balancing = 1;
7739 unsigned long interval;
7740 struct sched_domain *sd;
7741 /* Earliest time when we have to do rebalance again */
7742 unsigned long next_balance = jiffies + 60*HZ;
7743 int update_next_balance = 0;
7744 int need_serialize, need_decay = 0;
7747 update_blocked_averages(cpu);
7750 for_each_domain(cpu, sd) {
7752 * Decay the newidle max times here because this is a regular
7753 * visit to all the domains. Decay ~1% per second.
7755 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7756 sd->max_newidle_lb_cost =
7757 (sd->max_newidle_lb_cost * 253) / 256;
7758 sd->next_decay_max_lb_cost = jiffies + HZ;
7761 max_cost += sd->max_newidle_lb_cost;
7763 if (!(sd->flags & SD_LOAD_BALANCE))
7767 * Stop the load balance at this level. There is another
7768 * CPU in our sched group which is doing load balancing more
7771 if (!continue_balancing) {
7777 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7779 need_serialize = sd->flags & SD_SERIALIZE;
7780 if (need_serialize) {
7781 if (!spin_trylock(&balancing))
7785 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7786 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7788 * The LBF_DST_PINNED logic could have changed
7789 * env->dst_cpu, so we can't know our idle
7790 * state even if we migrated tasks. Update it.
7792 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7794 sd->last_balance = jiffies;
7795 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7798 spin_unlock(&balancing);
7800 if (time_after(next_balance, sd->last_balance + interval)) {
7801 next_balance = sd->last_balance + interval;
7802 update_next_balance = 1;
7807 * Ensure the rq-wide value also decays but keep it at a
7808 * reasonable floor to avoid funnies with rq->avg_idle.
7810 rq->max_idle_balance_cost =
7811 max((u64)sysctl_sched_migration_cost, max_cost);
7816 * next_balance will be updated only when there is a need.
7817 * When the cpu is attached to null domain for ex, it will not be
7820 if (likely(update_next_balance))
7821 rq->next_balance = next_balance;
7824 #ifdef CONFIG_NO_HZ_COMMON
7826 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7827 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7829 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7831 int this_cpu = this_rq->cpu;
7835 if (idle != CPU_IDLE ||
7836 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7839 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7840 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7844 * If this cpu gets work to do, stop the load balancing
7845 * work being done for other cpus. Next load
7846 * balancing owner will pick it up.
7851 rq = cpu_rq(balance_cpu);
7854 * If time for next balance is due,
7857 if (time_after_eq(jiffies, rq->next_balance)) {
7858 raw_spin_lock_irq(&rq->lock);
7859 update_rq_clock(rq);
7860 update_idle_cpu_load(rq);
7861 raw_spin_unlock_irq(&rq->lock);
7862 rebalance_domains(rq, CPU_IDLE);
7865 if (time_after(this_rq->next_balance, rq->next_balance))
7866 this_rq->next_balance = rq->next_balance;
7868 nohz.next_balance = this_rq->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 (numabalancing_enabled)
8022 task_tick_numa(rq, curr);
8024 update_rq_runnable_avg(rq, 1);
8028 * called on fork with the child task as argument from the parent's context
8029 * - child not yet on the tasklist
8030 * - preemption disabled
8032 static void task_fork_fair(struct task_struct *p)
8034 struct cfs_rq *cfs_rq;
8035 struct sched_entity *se = &p->se, *curr;
8036 int this_cpu = smp_processor_id();
8037 struct rq *rq = this_rq();
8038 unsigned long flags;
8040 raw_spin_lock_irqsave(&rq->lock, flags);
8042 update_rq_clock(rq);
8044 cfs_rq = task_cfs_rq(current);
8045 curr = cfs_rq->curr;
8048 * Not only the cpu but also the task_group of the parent might have
8049 * been changed after parent->se.parent,cfs_rq were copied to
8050 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8051 * of child point to valid ones.
8054 __set_task_cpu(p, this_cpu);
8057 update_curr(cfs_rq);
8060 se->vruntime = curr->vruntime;
8061 place_entity(cfs_rq, se, 1);
8063 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8065 * Upon rescheduling, sched_class::put_prev_task() will place
8066 * 'current' within the tree based on its new key value.
8068 swap(curr->vruntime, se->vruntime);
8072 se->vruntime -= cfs_rq->min_vruntime;
8074 raw_spin_unlock_irqrestore(&rq->lock, flags);
8078 * Priority of the task has changed. Check to see if we preempt
8082 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8084 if (!task_on_rq_queued(p))
8088 * Reschedule if we are currently running on this runqueue and
8089 * our priority decreased, or if we are not currently running on
8090 * this runqueue and our priority is higher than the current's
8092 if (rq->curr == p) {
8093 if (p->prio > oldprio)
8096 check_preempt_curr(rq, p, 0);
8099 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8101 struct sched_entity *se = &p->se;
8102 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8105 * Ensure the task's vruntime is normalized, so that when it's
8106 * switched back to the fair class the enqueue_entity(.flags=0) will
8107 * do the right thing.
8109 * If it's queued, then the dequeue_entity(.flags=0) will already
8110 * have normalized the vruntime, if it's !queued, then only when
8111 * the task is sleeping will it still have non-normalized vruntime.
8113 if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) {
8115 * Fix up our vruntime so that the current sleep doesn't
8116 * cause 'unlimited' sleep bonus.
8118 place_entity(cfs_rq, se, 0);
8119 se->vruntime -= cfs_rq->min_vruntime;
8124 * Remove our load from contribution when we leave sched_fair
8125 * and ensure we don't carry in an old decay_count if we
8128 if (se->avg.decay_count) {
8129 __synchronize_entity_decay(se);
8130 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
8136 * We switched to the sched_fair class.
8138 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8140 #ifdef CONFIG_FAIR_GROUP_SCHED
8141 struct sched_entity *se = &p->se;
8143 * Since the real-depth could have been changed (only FAIR
8144 * class maintain depth value), reset depth properly.
8146 se->depth = se->parent ? se->parent->depth + 1 : 0;
8148 if (!task_on_rq_queued(p))
8152 * We were most likely switched from sched_rt, so
8153 * kick off the schedule if running, otherwise just see
8154 * if we can still preempt the current task.
8159 check_preempt_curr(rq, p, 0);
8162 /* Account for a task changing its policy or group.
8164 * This routine is mostly called to set cfs_rq->curr field when a task
8165 * migrates between groups/classes.
8167 static void set_curr_task_fair(struct rq *rq)
8169 struct sched_entity *se = &rq->curr->se;
8171 for_each_sched_entity(se) {
8172 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8174 set_next_entity(cfs_rq, se);
8175 /* ensure bandwidth has been allocated on our new cfs_rq */
8176 account_cfs_rq_runtime(cfs_rq, 0);
8180 void init_cfs_rq(struct cfs_rq *cfs_rq)
8182 cfs_rq->tasks_timeline = RB_ROOT;
8183 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8184 #ifndef CONFIG_64BIT
8185 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8188 atomic64_set(&cfs_rq->decay_counter, 1);
8189 atomic_long_set(&cfs_rq->removed_load, 0);
8193 #ifdef CONFIG_FAIR_GROUP_SCHED
8194 static void task_move_group_fair(struct task_struct *p, int queued)
8196 struct sched_entity *se = &p->se;
8197 struct cfs_rq *cfs_rq;
8200 * If the task was not on the rq at the time of this cgroup movement
8201 * it must have been asleep, sleeping tasks keep their ->vruntime
8202 * absolute on their old rq until wakeup (needed for the fair sleeper
8203 * bonus in place_entity()).
8205 * If it was on the rq, we've just 'preempted' it, which does convert
8206 * ->vruntime to a relative base.
8208 * Make sure both cases convert their relative position when migrating
8209 * to another cgroup's rq. This does somewhat interfere with the
8210 * fair sleeper stuff for the first placement, but who cares.
8213 * When !queued, vruntime of the task has usually NOT been normalized.
8214 * But there are some cases where it has already been normalized:
8216 * - Moving a forked child which is waiting for being woken up by
8217 * wake_up_new_task().
8218 * - Moving a task which has been woken up by try_to_wake_up() and
8219 * waiting for actually being woken up by sched_ttwu_pending().
8221 * To prevent boost or penalty in the new cfs_rq caused by delta
8222 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
8224 if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING))
8228 se->vruntime -= cfs_rq_of(se)->min_vruntime;
8229 set_task_rq(p, task_cpu(p));
8230 se->depth = se->parent ? se->parent->depth + 1 : 0;
8232 cfs_rq = cfs_rq_of(se);
8233 se->vruntime += cfs_rq->min_vruntime;
8236 * migrate_task_rq_fair() will have removed our previous
8237 * contribution, but we must synchronize for ongoing future
8240 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
8241 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
8246 void free_fair_sched_group(struct task_group *tg)
8250 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8252 for_each_possible_cpu(i) {
8254 kfree(tg->cfs_rq[i]);
8263 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8265 struct cfs_rq *cfs_rq;
8266 struct sched_entity *se;
8269 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8272 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8276 tg->shares = NICE_0_LOAD;
8278 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8280 for_each_possible_cpu(i) {
8281 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8282 GFP_KERNEL, cpu_to_node(i));
8286 se = kzalloc_node(sizeof(struct sched_entity),
8287 GFP_KERNEL, cpu_to_node(i));
8291 init_cfs_rq(cfs_rq);
8292 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8303 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8305 struct rq *rq = cpu_rq(cpu);
8306 unsigned long flags;
8309 * Only empty task groups can be destroyed; so we can speculatively
8310 * check on_list without danger of it being re-added.
8312 if (!tg->cfs_rq[cpu]->on_list)
8315 raw_spin_lock_irqsave(&rq->lock, flags);
8316 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8317 raw_spin_unlock_irqrestore(&rq->lock, flags);
8320 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8321 struct sched_entity *se, int cpu,
8322 struct sched_entity *parent)
8324 struct rq *rq = cpu_rq(cpu);
8328 init_cfs_rq_runtime(cfs_rq);
8330 tg->cfs_rq[cpu] = cfs_rq;
8333 /* se could be NULL for root_task_group */
8338 se->cfs_rq = &rq->cfs;
8341 se->cfs_rq = parent->my_q;
8342 se->depth = parent->depth + 1;
8346 /* guarantee group entities always have weight */
8347 update_load_set(&se->load, NICE_0_LOAD);
8348 se->parent = parent;
8351 static DEFINE_MUTEX(shares_mutex);
8353 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8356 unsigned long flags;
8359 * We can't change the weight of the root cgroup.
8364 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8366 mutex_lock(&shares_mutex);
8367 if (tg->shares == shares)
8370 tg->shares = shares;
8371 for_each_possible_cpu(i) {
8372 struct rq *rq = cpu_rq(i);
8373 struct sched_entity *se;
8376 /* Propagate contribution to hierarchy */
8377 raw_spin_lock_irqsave(&rq->lock, flags);
8379 /* Possible calls to update_curr() need rq clock */
8380 update_rq_clock(rq);
8381 for_each_sched_entity(se)
8382 update_cfs_shares(group_cfs_rq(se));
8383 raw_spin_unlock_irqrestore(&rq->lock, flags);
8387 mutex_unlock(&shares_mutex);
8390 #else /* CONFIG_FAIR_GROUP_SCHED */
8392 void free_fair_sched_group(struct task_group *tg) { }
8394 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8399 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8401 #endif /* CONFIG_FAIR_GROUP_SCHED */
8404 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8406 struct sched_entity *se = &task->se;
8407 unsigned int rr_interval = 0;
8410 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8413 if (rq->cfs.load.weight)
8414 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8420 * All the scheduling class methods:
8422 const struct sched_class fair_sched_class = {
8423 .next = &idle_sched_class,
8424 .enqueue_task = enqueue_task_fair,
8425 .dequeue_task = dequeue_task_fair,
8426 .yield_task = yield_task_fair,
8427 .yield_to_task = yield_to_task_fair,
8429 .check_preempt_curr = check_preempt_wakeup,
8431 .pick_next_task = pick_next_task_fair,
8432 .put_prev_task = put_prev_task_fair,
8435 .select_task_rq = select_task_rq_fair,
8436 .migrate_task_rq = migrate_task_rq_fair,
8438 .rq_online = rq_online_fair,
8439 .rq_offline = rq_offline_fair,
8441 .task_waking = task_waking_fair,
8444 .set_curr_task = set_curr_task_fair,
8445 .task_tick = task_tick_fair,
8446 .task_fork = task_fork_fair,
8448 .prio_changed = prio_changed_fair,
8449 .switched_from = switched_from_fair,
8450 .switched_to = switched_to_fair,
8452 .get_rr_interval = get_rr_interval_fair,
8454 .update_curr = update_curr_fair,
8456 #ifdef CONFIG_FAIR_GROUP_SCHED
8457 .task_move_group = task_move_group_fair,
8461 #ifdef CONFIG_SCHED_DEBUG
8462 void print_cfs_stats(struct seq_file *m, int cpu)
8464 struct cfs_rq *cfs_rq;
8467 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8468 print_cfs_rq(m, cpu, cfs_rq);
8472 #ifdef CONFIG_NUMA_BALANCING
8473 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8476 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8478 for_each_online_node(node) {
8479 if (p->numa_faults) {
8480 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8481 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8483 if (p->numa_group) {
8484 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8485 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8487 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8490 #endif /* CONFIG_NUMA_BALANCING */
8491 #endif /* CONFIG_SCHED_DEBUG */
8493 __init void init_sched_fair_class(void)
8496 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8498 #ifdef CONFIG_NO_HZ_COMMON
8499 nohz.next_balance = jiffies;
8500 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8501 cpu_notifier(sched_ilb_notifier, 0);