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)
1201 long src_capacity, dst_capacity;
1203 long load_a, load_b;
1208 * The load is corrected for the CPU capacity available on each node.
1211 * ------------ vs ---------
1212 * src_capacity dst_capacity
1214 src_capacity = env->src_stats.compute_capacity;
1215 dst_capacity = env->dst_stats.compute_capacity;
1217 /* We care about the slope of the imbalance, not the direction. */
1220 if (load_a < load_b)
1221 swap(load_a, load_b);
1223 /* Is the difference below the threshold? */
1224 imb = load_a * src_capacity * 100 -
1225 load_b * dst_capacity * env->imbalance_pct;
1230 * The imbalance is above the allowed threshold.
1231 * Allow a move that brings us closer to a balanced situation,
1232 * without moving things past the point of balance.
1234 orig_src_load = env->src_stats.load;
1237 * In a task swap, there will be one load moving from src to dst,
1238 * and another moving back. This is the net sum of both moves.
1239 * A simple task move will always have a positive value.
1240 * Allow the move if it brings the system closer to a balanced
1241 * situation, without crossing over the balance point.
1243 moved_load = orig_src_load - src_load;
1246 /* Moving src -> dst. Did we overshoot balance? */
1247 return src_load * dst_capacity < dst_load * src_capacity;
1249 /* Moving dst -> src. Did we overshoot balance? */
1250 return dst_load * src_capacity < src_load * dst_capacity;
1254 * This checks if the overall compute and NUMA accesses of the system would
1255 * be improved if the source tasks was migrated to the target dst_cpu taking
1256 * into account that it might be best if task running on the dst_cpu should
1257 * be exchanged with the source task
1259 static void task_numa_compare(struct task_numa_env *env,
1260 long taskimp, long groupimp)
1262 struct rq *src_rq = cpu_rq(env->src_cpu);
1263 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1264 struct task_struct *cur;
1265 long src_load, dst_load;
1267 long imp = env->p->numa_group ? groupimp : taskimp;
1269 int dist = env->dist;
1273 raw_spin_lock_irq(&dst_rq->lock);
1276 * No need to move the exiting task, and this ensures that ->curr
1277 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1278 * is safe under RCU read lock.
1279 * Note that rcu_read_lock() itself can't protect from the final
1280 * put_task_struct() after the last schedule().
1282 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1284 raw_spin_unlock_irq(&dst_rq->lock);
1287 * Because we have preemption enabled we can get migrated around and
1288 * end try selecting ourselves (current == env->p) as a swap candidate.
1294 * "imp" is the fault differential for the source task between the
1295 * source and destination node. Calculate the total differential for
1296 * the source task and potential destination task. The more negative
1297 * the value is, the more rmeote accesses that would be expected to
1298 * be incurred if the tasks were swapped.
1301 /* Skip this swap candidate if cannot move to the source cpu */
1302 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1306 * If dst and source tasks are in the same NUMA group, or not
1307 * in any group then look only at task weights.
1309 if (cur->numa_group == env->p->numa_group) {
1310 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1311 task_weight(cur, env->dst_nid, dist);
1313 * Add some hysteresis to prevent swapping the
1314 * tasks within a group over tiny differences.
1316 if (cur->numa_group)
1320 * Compare the group weights. If a task is all by
1321 * itself (not part of a group), use the task weight
1324 if (cur->numa_group)
1325 imp += group_weight(cur, env->src_nid, dist) -
1326 group_weight(cur, env->dst_nid, dist);
1328 imp += task_weight(cur, env->src_nid, dist) -
1329 task_weight(cur, env->dst_nid, dist);
1333 if (imp <= env->best_imp && moveimp <= env->best_imp)
1337 /* Is there capacity at our destination? */
1338 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1339 !env->dst_stats.has_free_capacity)
1345 /* Balance doesn't matter much if we're running a task per cpu */
1346 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1347 dst_rq->nr_running == 1)
1351 * In the overloaded case, try and keep the load balanced.
1354 load = task_h_load(env->p);
1355 dst_load = env->dst_stats.load + load;
1356 src_load = env->src_stats.load - load;
1358 if (moveimp > imp && moveimp > env->best_imp) {
1360 * If the improvement from just moving env->p direction is
1361 * better than swapping tasks around, check if a move is
1362 * possible. Store a slightly smaller score than moveimp,
1363 * so an actually idle CPU will win.
1365 if (!load_too_imbalanced(src_load, dst_load, env)) {
1372 if (imp <= env->best_imp)
1376 load = task_h_load(cur);
1381 if (load_too_imbalanced(src_load, dst_load, env))
1385 * One idle CPU per node is evaluated for a task numa move.
1386 * Call select_idle_sibling to maybe find a better one.
1389 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1392 task_numa_assign(env, cur, imp);
1397 static void task_numa_find_cpu(struct task_numa_env *env,
1398 long taskimp, long groupimp)
1402 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1403 /* Skip this CPU if the source task cannot migrate */
1404 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1408 task_numa_compare(env, taskimp, groupimp);
1412 static int task_numa_migrate(struct task_struct *p)
1414 struct task_numa_env env = {
1417 .src_cpu = task_cpu(p),
1418 .src_nid = task_node(p),
1420 .imbalance_pct = 112,
1426 struct sched_domain *sd;
1427 unsigned long taskweight, groupweight;
1429 long taskimp, groupimp;
1432 * Pick the lowest SD_NUMA domain, as that would have the smallest
1433 * imbalance and would be the first to start moving tasks about.
1435 * And we want to avoid any moving of tasks about, as that would create
1436 * random movement of tasks -- counter the numa conditions we're trying
1440 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1442 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1446 * Cpusets can break the scheduler domain tree into smaller
1447 * balance domains, some of which do not cross NUMA boundaries.
1448 * Tasks that are "trapped" in such domains cannot be migrated
1449 * elsewhere, so there is no point in (re)trying.
1451 if (unlikely(!sd)) {
1452 p->numa_preferred_nid = task_node(p);
1456 env.dst_nid = p->numa_preferred_nid;
1457 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1458 taskweight = task_weight(p, env.src_nid, dist);
1459 groupweight = group_weight(p, env.src_nid, dist);
1460 update_numa_stats(&env.src_stats, env.src_nid);
1461 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1462 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1463 update_numa_stats(&env.dst_stats, env.dst_nid);
1465 /* Try to find a spot on the preferred nid. */
1466 task_numa_find_cpu(&env, taskimp, groupimp);
1469 * Look at other nodes in these cases:
1470 * - there is no space available on the preferred_nid
1471 * - the task is part of a numa_group that is interleaved across
1472 * multiple NUMA nodes; in order to better consolidate the group,
1473 * we need to check other locations.
1475 if (env.best_cpu == -1 || (p->numa_group &&
1476 nodes_weight(p->numa_group->active_nodes) > 1)) {
1477 for_each_online_node(nid) {
1478 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1481 dist = node_distance(env.src_nid, env.dst_nid);
1482 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1484 taskweight = task_weight(p, env.src_nid, dist);
1485 groupweight = group_weight(p, env.src_nid, dist);
1488 /* Only consider nodes where both task and groups benefit */
1489 taskimp = task_weight(p, nid, dist) - taskweight;
1490 groupimp = group_weight(p, nid, dist) - groupweight;
1491 if (taskimp < 0 && groupimp < 0)
1496 update_numa_stats(&env.dst_stats, env.dst_nid);
1497 task_numa_find_cpu(&env, taskimp, groupimp);
1502 * If the task is part of a workload that spans multiple NUMA nodes,
1503 * and is migrating into one of the workload's active nodes, remember
1504 * this node as the task's preferred numa node, so the workload can
1506 * A task that migrated to a second choice node will be better off
1507 * trying for a better one later. Do not set the preferred node here.
1509 if (p->numa_group) {
1510 if (env.best_cpu == -1)
1515 if (node_isset(nid, p->numa_group->active_nodes))
1516 sched_setnuma(p, env.dst_nid);
1519 /* No better CPU than the current one was found. */
1520 if (env.best_cpu == -1)
1524 * Reset the scan period if the task is being rescheduled on an
1525 * alternative node to recheck if the tasks is now properly placed.
1527 p->numa_scan_period = task_scan_min(p);
1529 if (env.best_task == NULL) {
1530 ret = migrate_task_to(p, env.best_cpu);
1532 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1536 ret = migrate_swap(p, env.best_task);
1538 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1539 put_task_struct(env.best_task);
1543 /* Attempt to migrate a task to a CPU on the preferred node. */
1544 static void numa_migrate_preferred(struct task_struct *p)
1546 unsigned long interval = HZ;
1548 /* This task has no NUMA fault statistics yet */
1549 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1552 /* Periodically retry migrating the task to the preferred node */
1553 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1554 p->numa_migrate_retry = jiffies + interval;
1556 /* Success if task is already running on preferred CPU */
1557 if (task_node(p) == p->numa_preferred_nid)
1560 /* Otherwise, try migrate to a CPU on the preferred node */
1561 task_numa_migrate(p);
1565 * Find the nodes on which the workload is actively running. We do this by
1566 * tracking the nodes from which NUMA hinting faults are triggered. This can
1567 * be different from the set of nodes where the workload's memory is currently
1570 * The bitmask is used to make smarter decisions on when to do NUMA page
1571 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1572 * are added when they cause over 6/16 of the maximum number of faults, but
1573 * only removed when they drop below 3/16.
1575 static void update_numa_active_node_mask(struct numa_group *numa_group)
1577 unsigned long faults, max_faults = 0;
1580 for_each_online_node(nid) {
1581 faults = group_faults_cpu(numa_group, nid);
1582 if (faults > max_faults)
1583 max_faults = faults;
1586 for_each_online_node(nid) {
1587 faults = group_faults_cpu(numa_group, nid);
1588 if (!node_isset(nid, numa_group->active_nodes)) {
1589 if (faults > max_faults * 6 / 16)
1590 node_set(nid, numa_group->active_nodes);
1591 } else if (faults < max_faults * 3 / 16)
1592 node_clear(nid, numa_group->active_nodes);
1597 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1598 * increments. The more local the fault statistics are, the higher the scan
1599 * period will be for the next scan window. If local/(local+remote) ratio is
1600 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1601 * the scan period will decrease. Aim for 70% local accesses.
1603 #define NUMA_PERIOD_SLOTS 10
1604 #define NUMA_PERIOD_THRESHOLD 7
1607 * Increase the scan period (slow down scanning) if the majority of
1608 * our memory is already on our local node, or if the majority of
1609 * the page accesses are shared with other processes.
1610 * Otherwise, decrease the scan period.
1612 static void update_task_scan_period(struct task_struct *p,
1613 unsigned long shared, unsigned long private)
1615 unsigned int period_slot;
1619 unsigned long remote = p->numa_faults_locality[0];
1620 unsigned long local = p->numa_faults_locality[1];
1623 * If there were no record hinting faults then either the task is
1624 * completely idle or all activity is areas that are not of interest
1625 * to automatic numa balancing. Related to that, if there were failed
1626 * migration then it implies we are migrating too quickly or the local
1627 * node is overloaded. In either case, scan slower
1629 if (local + shared == 0 || p->numa_faults_locality[2]) {
1630 p->numa_scan_period = min(p->numa_scan_period_max,
1631 p->numa_scan_period << 1);
1633 p->mm->numa_next_scan = jiffies +
1634 msecs_to_jiffies(p->numa_scan_period);
1640 * Prepare to scale scan period relative to the current period.
1641 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1642 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1643 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1645 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1646 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1647 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1648 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1651 diff = slot * period_slot;
1653 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1656 * Scale scan rate increases based on sharing. There is an
1657 * inverse relationship between the degree of sharing and
1658 * the adjustment made to the scanning period. Broadly
1659 * speaking the intent is that there is little point
1660 * scanning faster if shared accesses dominate as it may
1661 * simply bounce migrations uselessly
1663 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1664 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1667 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1668 task_scan_min(p), task_scan_max(p));
1669 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1673 * Get the fraction of time the task has been running since the last
1674 * NUMA placement cycle. The scheduler keeps similar statistics, but
1675 * decays those on a 32ms period, which is orders of magnitude off
1676 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1677 * stats only if the task is so new there are no NUMA statistics yet.
1679 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1681 u64 runtime, delta, now;
1682 /* Use the start of this time slice to avoid calculations. */
1683 now = p->se.exec_start;
1684 runtime = p->se.sum_exec_runtime;
1686 if (p->last_task_numa_placement) {
1687 delta = runtime - p->last_sum_exec_runtime;
1688 *period = now - p->last_task_numa_placement;
1690 delta = p->se.avg.runnable_avg_sum;
1691 *period = p->se.avg.avg_period;
1694 p->last_sum_exec_runtime = runtime;
1695 p->last_task_numa_placement = now;
1701 * Determine the preferred nid for a task in a numa_group. This needs to
1702 * be done in a way that produces consistent results with group_weight,
1703 * otherwise workloads might not converge.
1705 static int preferred_group_nid(struct task_struct *p, int nid)
1710 /* Direct connections between all NUMA nodes. */
1711 if (sched_numa_topology_type == NUMA_DIRECT)
1715 * On a system with glueless mesh NUMA topology, group_weight
1716 * scores nodes according to the number of NUMA hinting faults on
1717 * both the node itself, and on nearby nodes.
1719 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1720 unsigned long score, max_score = 0;
1721 int node, max_node = nid;
1723 dist = sched_max_numa_distance;
1725 for_each_online_node(node) {
1726 score = group_weight(p, node, dist);
1727 if (score > max_score) {
1736 * Finding the preferred nid in a system with NUMA backplane
1737 * interconnect topology is more involved. The goal is to locate
1738 * tasks from numa_groups near each other in the system, and
1739 * untangle workloads from different sides of the system. This requires
1740 * searching down the hierarchy of node groups, recursively searching
1741 * inside the highest scoring group of nodes. The nodemask tricks
1742 * keep the complexity of the search down.
1744 nodes = node_online_map;
1745 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1746 unsigned long max_faults = 0;
1747 nodemask_t max_group = NODE_MASK_NONE;
1750 /* Are there nodes at this distance from each other? */
1751 if (!find_numa_distance(dist))
1754 for_each_node_mask(a, nodes) {
1755 unsigned long faults = 0;
1756 nodemask_t this_group;
1757 nodes_clear(this_group);
1759 /* Sum group's NUMA faults; includes a==b case. */
1760 for_each_node_mask(b, nodes) {
1761 if (node_distance(a, b) < dist) {
1762 faults += group_faults(p, b);
1763 node_set(b, this_group);
1764 node_clear(b, nodes);
1768 /* Remember the top group. */
1769 if (faults > max_faults) {
1770 max_faults = faults;
1771 max_group = this_group;
1773 * subtle: at the smallest distance there is
1774 * just one node left in each "group", the
1775 * winner is the preferred nid.
1780 /* Next round, evaluate the nodes within max_group. */
1788 static void task_numa_placement(struct task_struct *p)
1790 int seq, nid, max_nid = -1, max_group_nid = -1;
1791 unsigned long max_faults = 0, max_group_faults = 0;
1792 unsigned long fault_types[2] = { 0, 0 };
1793 unsigned long total_faults;
1794 u64 runtime, period;
1795 spinlock_t *group_lock = NULL;
1798 * The p->mm->numa_scan_seq field gets updated without
1799 * exclusive access. Use READ_ONCE() here to ensure
1800 * that the field is read in a single access:
1802 seq = READ_ONCE(p->mm->numa_scan_seq);
1803 if (p->numa_scan_seq == seq)
1805 p->numa_scan_seq = seq;
1806 p->numa_scan_period_max = task_scan_max(p);
1808 total_faults = p->numa_faults_locality[0] +
1809 p->numa_faults_locality[1];
1810 runtime = numa_get_avg_runtime(p, &period);
1812 /* If the task is part of a group prevent parallel updates to group stats */
1813 if (p->numa_group) {
1814 group_lock = &p->numa_group->lock;
1815 spin_lock_irq(group_lock);
1818 /* Find the node with the highest number of faults */
1819 for_each_online_node(nid) {
1820 /* Keep track of the offsets in numa_faults array */
1821 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1822 unsigned long faults = 0, group_faults = 0;
1825 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1826 long diff, f_diff, f_weight;
1828 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1829 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1830 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1831 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1833 /* Decay existing window, copy faults since last scan */
1834 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1835 fault_types[priv] += p->numa_faults[membuf_idx];
1836 p->numa_faults[membuf_idx] = 0;
1839 * Normalize the faults_from, so all tasks in a group
1840 * count according to CPU use, instead of by the raw
1841 * number of faults. Tasks with little runtime have
1842 * little over-all impact on throughput, and thus their
1843 * faults are less important.
1845 f_weight = div64_u64(runtime << 16, period + 1);
1846 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1848 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1849 p->numa_faults[cpubuf_idx] = 0;
1851 p->numa_faults[mem_idx] += diff;
1852 p->numa_faults[cpu_idx] += f_diff;
1853 faults += p->numa_faults[mem_idx];
1854 p->total_numa_faults += diff;
1855 if (p->numa_group) {
1857 * safe because we can only change our own group
1859 * mem_idx represents the offset for a given
1860 * nid and priv in a specific region because it
1861 * is at the beginning of the numa_faults array.
1863 p->numa_group->faults[mem_idx] += diff;
1864 p->numa_group->faults_cpu[mem_idx] += f_diff;
1865 p->numa_group->total_faults += diff;
1866 group_faults += p->numa_group->faults[mem_idx];
1870 if (faults > max_faults) {
1871 max_faults = faults;
1875 if (group_faults > max_group_faults) {
1876 max_group_faults = group_faults;
1877 max_group_nid = nid;
1881 update_task_scan_period(p, fault_types[0], fault_types[1]);
1883 if (p->numa_group) {
1884 update_numa_active_node_mask(p->numa_group);
1885 spin_unlock_irq(group_lock);
1886 max_nid = preferred_group_nid(p, max_group_nid);
1890 /* Set the new preferred node */
1891 if (max_nid != p->numa_preferred_nid)
1892 sched_setnuma(p, max_nid);
1894 if (task_node(p) != p->numa_preferred_nid)
1895 numa_migrate_preferred(p);
1899 static inline int get_numa_group(struct numa_group *grp)
1901 return atomic_inc_not_zero(&grp->refcount);
1904 static inline void put_numa_group(struct numa_group *grp)
1906 if (atomic_dec_and_test(&grp->refcount))
1907 kfree_rcu(grp, rcu);
1910 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1913 struct numa_group *grp, *my_grp;
1914 struct task_struct *tsk;
1916 int cpu = cpupid_to_cpu(cpupid);
1919 if (unlikely(!p->numa_group)) {
1920 unsigned int size = sizeof(struct numa_group) +
1921 4*nr_node_ids*sizeof(unsigned long);
1923 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1927 atomic_set(&grp->refcount, 1);
1928 spin_lock_init(&grp->lock);
1930 /* Second half of the array tracks nids where faults happen */
1931 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1934 node_set(task_node(current), grp->active_nodes);
1936 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1937 grp->faults[i] = p->numa_faults[i];
1939 grp->total_faults = p->total_numa_faults;
1942 rcu_assign_pointer(p->numa_group, grp);
1946 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1948 if (!cpupid_match_pid(tsk, cpupid))
1951 grp = rcu_dereference(tsk->numa_group);
1955 my_grp = p->numa_group;
1960 * Only join the other group if its bigger; if we're the bigger group,
1961 * the other task will join us.
1963 if (my_grp->nr_tasks > grp->nr_tasks)
1967 * Tie-break on the grp address.
1969 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1972 /* Always join threads in the same process. */
1973 if (tsk->mm == current->mm)
1976 /* Simple filter to avoid false positives due to PID collisions */
1977 if (flags & TNF_SHARED)
1980 /* Update priv based on whether false sharing was detected */
1983 if (join && !get_numa_group(grp))
1991 BUG_ON(irqs_disabled());
1992 double_lock_irq(&my_grp->lock, &grp->lock);
1994 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1995 my_grp->faults[i] -= p->numa_faults[i];
1996 grp->faults[i] += p->numa_faults[i];
1998 my_grp->total_faults -= p->total_numa_faults;
1999 grp->total_faults += p->total_numa_faults;
2004 spin_unlock(&my_grp->lock);
2005 spin_unlock_irq(&grp->lock);
2007 rcu_assign_pointer(p->numa_group, grp);
2009 put_numa_group(my_grp);
2017 void task_numa_free(struct task_struct *p)
2019 struct numa_group *grp = p->numa_group;
2020 void *numa_faults = p->numa_faults;
2021 unsigned long flags;
2025 spin_lock_irqsave(&grp->lock, flags);
2026 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2027 grp->faults[i] -= p->numa_faults[i];
2028 grp->total_faults -= p->total_numa_faults;
2031 spin_unlock_irqrestore(&grp->lock, flags);
2032 RCU_INIT_POINTER(p->numa_group, NULL);
2033 put_numa_group(grp);
2036 p->numa_faults = NULL;
2041 * Got a PROT_NONE fault for a page on @node.
2043 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2045 struct task_struct *p = current;
2046 bool migrated = flags & TNF_MIGRATED;
2047 int cpu_node = task_node(current);
2048 int local = !!(flags & TNF_FAULT_LOCAL);
2051 if (!numabalancing_enabled)
2054 /* for example, ksmd faulting in a user's mm */
2058 /* Allocate buffer to track faults on a per-node basis */
2059 if (unlikely(!p->numa_faults)) {
2060 int size = sizeof(*p->numa_faults) *
2061 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2063 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2064 if (!p->numa_faults)
2067 p->total_numa_faults = 0;
2068 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2072 * First accesses are treated as private, otherwise consider accesses
2073 * to be private if the accessing pid has not changed
2075 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2078 priv = cpupid_match_pid(p, last_cpupid);
2079 if (!priv && !(flags & TNF_NO_GROUP))
2080 task_numa_group(p, last_cpupid, flags, &priv);
2084 * If a workload spans multiple NUMA nodes, a shared fault that
2085 * occurs wholly within the set of nodes that the workload is
2086 * actively using should be counted as local. This allows the
2087 * scan rate to slow down when a workload has settled down.
2089 if (!priv && !local && p->numa_group &&
2090 node_isset(cpu_node, p->numa_group->active_nodes) &&
2091 node_isset(mem_node, p->numa_group->active_nodes))
2094 task_numa_placement(p);
2097 * Retry task to preferred node migration periodically, in case it
2098 * case it previously failed, or the scheduler moved us.
2100 if (time_after(jiffies, p->numa_migrate_retry))
2101 numa_migrate_preferred(p);
2104 p->numa_pages_migrated += pages;
2105 if (flags & TNF_MIGRATE_FAIL)
2106 p->numa_faults_locality[2] += pages;
2108 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2109 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2110 p->numa_faults_locality[local] += pages;
2113 static void reset_ptenuma_scan(struct task_struct *p)
2116 * We only did a read acquisition of the mmap sem, so
2117 * p->mm->numa_scan_seq is written to without exclusive access
2118 * and the update is not guaranteed to be atomic. That's not
2119 * much of an issue though, since this is just used for
2120 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2121 * expensive, to avoid any form of compiler optimizations:
2123 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2124 p->mm->numa_scan_offset = 0;
2128 * The expensive part of numa migration is done from task_work context.
2129 * Triggered from task_tick_numa().
2131 void task_numa_work(struct callback_head *work)
2133 unsigned long migrate, next_scan, now = jiffies;
2134 struct task_struct *p = current;
2135 struct mm_struct *mm = p->mm;
2136 struct vm_area_struct *vma;
2137 unsigned long start, end;
2138 unsigned long nr_pte_updates = 0;
2141 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2143 work->next = work; /* protect against double add */
2145 * Who cares about NUMA placement when they're dying.
2147 * NOTE: make sure not to dereference p->mm before this check,
2148 * exit_task_work() happens _after_ exit_mm() so we could be called
2149 * without p->mm even though we still had it when we enqueued this
2152 if (p->flags & PF_EXITING)
2155 if (!mm->numa_next_scan) {
2156 mm->numa_next_scan = now +
2157 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2161 * Enforce maximal scan/migration frequency..
2163 migrate = mm->numa_next_scan;
2164 if (time_before(now, migrate))
2167 if (p->numa_scan_period == 0) {
2168 p->numa_scan_period_max = task_scan_max(p);
2169 p->numa_scan_period = task_scan_min(p);
2172 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2173 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2177 * Delay this task enough that another task of this mm will likely win
2178 * the next time around.
2180 p->node_stamp += 2 * TICK_NSEC;
2182 start = mm->numa_scan_offset;
2183 pages = sysctl_numa_balancing_scan_size;
2184 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2188 down_read(&mm->mmap_sem);
2189 vma = find_vma(mm, start);
2191 reset_ptenuma_scan(p);
2195 for (; vma; vma = vma->vm_next) {
2196 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2197 is_vm_hugetlb_page(vma)) {
2202 * Shared library pages mapped by multiple processes are not
2203 * migrated as it is expected they are cache replicated. Avoid
2204 * hinting faults in read-only file-backed mappings or the vdso
2205 * as migrating the pages will be of marginal benefit.
2208 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2212 * Skip inaccessible VMAs to avoid any confusion between
2213 * PROT_NONE and NUMA hinting ptes
2215 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2219 start = max(start, vma->vm_start);
2220 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2221 end = min(end, vma->vm_end);
2222 nr_pte_updates += change_prot_numa(vma, start, end);
2225 * Scan sysctl_numa_balancing_scan_size but ensure that
2226 * at least one PTE is updated so that unused virtual
2227 * address space is quickly skipped.
2230 pages -= (end - start) >> PAGE_SHIFT;
2237 } while (end != vma->vm_end);
2242 * It is possible to reach the end of the VMA list but the last few
2243 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2244 * would find the !migratable VMA on the next scan but not reset the
2245 * scanner to the start so check it now.
2248 mm->numa_scan_offset = start;
2250 reset_ptenuma_scan(p);
2251 up_read(&mm->mmap_sem);
2255 * Drive the periodic memory faults..
2257 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2259 struct callback_head *work = &curr->numa_work;
2263 * We don't care about NUMA placement if we don't have memory.
2265 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2269 * Using runtime rather than walltime has the dual advantage that
2270 * we (mostly) drive the selection from busy threads and that the
2271 * task needs to have done some actual work before we bother with
2274 now = curr->se.sum_exec_runtime;
2275 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2277 if (now - curr->node_stamp > period) {
2278 if (!curr->node_stamp)
2279 curr->numa_scan_period = task_scan_min(curr);
2280 curr->node_stamp += period;
2282 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2283 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2284 task_work_add(curr, work, true);
2289 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2293 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2297 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2300 #endif /* CONFIG_NUMA_BALANCING */
2303 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2305 update_load_add(&cfs_rq->load, se->load.weight);
2306 if (!parent_entity(se))
2307 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2309 if (entity_is_task(se)) {
2310 struct rq *rq = rq_of(cfs_rq);
2312 account_numa_enqueue(rq, task_of(se));
2313 list_add(&se->group_node, &rq->cfs_tasks);
2316 cfs_rq->nr_running++;
2320 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2322 update_load_sub(&cfs_rq->load, se->load.weight);
2323 if (!parent_entity(se))
2324 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2325 if (entity_is_task(se)) {
2326 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2327 list_del_init(&se->group_node);
2329 cfs_rq->nr_running--;
2332 #ifdef CONFIG_FAIR_GROUP_SCHED
2334 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2339 * Use this CPU's actual weight instead of the last load_contribution
2340 * to gain a more accurate current total weight. See
2341 * update_cfs_rq_load_contribution().
2343 tg_weight = atomic_long_read(&tg->load_avg);
2344 tg_weight -= cfs_rq->tg_load_contrib;
2345 tg_weight += cfs_rq->load.weight;
2350 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2352 long tg_weight, load, shares;
2354 tg_weight = calc_tg_weight(tg, cfs_rq);
2355 load = cfs_rq->load.weight;
2357 shares = (tg->shares * load);
2359 shares /= tg_weight;
2361 if (shares < MIN_SHARES)
2362 shares = MIN_SHARES;
2363 if (shares > tg->shares)
2364 shares = tg->shares;
2368 # else /* CONFIG_SMP */
2369 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2373 # endif /* CONFIG_SMP */
2374 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2375 unsigned long weight)
2378 /* commit outstanding execution time */
2379 if (cfs_rq->curr == se)
2380 update_curr(cfs_rq);
2381 account_entity_dequeue(cfs_rq, se);
2384 update_load_set(&se->load, weight);
2387 account_entity_enqueue(cfs_rq, se);
2390 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2392 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2394 struct task_group *tg;
2395 struct sched_entity *se;
2399 se = tg->se[cpu_of(rq_of(cfs_rq))];
2400 if (!se || throttled_hierarchy(cfs_rq))
2403 if (likely(se->load.weight == tg->shares))
2406 shares = calc_cfs_shares(cfs_rq, tg);
2408 reweight_entity(cfs_rq_of(se), se, shares);
2410 #else /* CONFIG_FAIR_GROUP_SCHED */
2411 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2414 #endif /* CONFIG_FAIR_GROUP_SCHED */
2418 * We choose a half-life close to 1 scheduling period.
2419 * Note: The tables below are dependent on this value.
2421 #define LOAD_AVG_PERIOD 32
2422 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2423 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2425 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2426 static const u32 runnable_avg_yN_inv[] = {
2427 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2428 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2429 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2430 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2431 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2432 0x85aac367, 0x82cd8698,
2436 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2437 * over-estimates when re-combining.
2439 static const u32 runnable_avg_yN_sum[] = {
2440 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2441 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2442 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2447 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2449 static __always_inline u64 decay_load(u64 val, u64 n)
2451 unsigned int local_n;
2455 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2458 /* after bounds checking we can collapse to 32-bit */
2462 * As y^PERIOD = 1/2, we can combine
2463 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2464 * With a look-up table which covers y^n (n<PERIOD)
2466 * To achieve constant time decay_load.
2468 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2469 val >>= local_n / LOAD_AVG_PERIOD;
2470 local_n %= LOAD_AVG_PERIOD;
2473 val *= runnable_avg_yN_inv[local_n];
2474 /* We don't use SRR here since we always want to round down. */
2479 * For updates fully spanning n periods, the contribution to runnable
2480 * average will be: \Sum 1024*y^n
2482 * We can compute this reasonably efficiently by combining:
2483 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2485 static u32 __compute_runnable_contrib(u64 n)
2489 if (likely(n <= LOAD_AVG_PERIOD))
2490 return runnable_avg_yN_sum[n];
2491 else if (unlikely(n >= LOAD_AVG_MAX_N))
2492 return LOAD_AVG_MAX;
2494 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2496 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2497 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2499 n -= LOAD_AVG_PERIOD;
2500 } while (n > LOAD_AVG_PERIOD);
2502 contrib = decay_load(contrib, n);
2503 return contrib + runnable_avg_yN_sum[n];
2507 * We can represent the historical contribution to runnable average as the
2508 * coefficients of a geometric series. To do this we sub-divide our runnable
2509 * history into segments of approximately 1ms (1024us); label the segment that
2510 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2512 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2514 * (now) (~1ms ago) (~2ms ago)
2516 * Let u_i denote the fraction of p_i that the entity was runnable.
2518 * We then designate the fractions u_i as our co-efficients, yielding the
2519 * following representation of historical load:
2520 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2522 * We choose y based on the with of a reasonably scheduling period, fixing:
2525 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2526 * approximately half as much as the contribution to load within the last ms
2529 * When a period "rolls over" and we have new u_0`, multiplying the previous
2530 * sum again by y is sufficient to update:
2531 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2532 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2534 static __always_inline int __update_entity_runnable_avg(u64 now, int cpu,
2535 struct sched_avg *sa,
2540 u32 runnable_contrib;
2541 int delta_w, decayed = 0;
2542 unsigned long scale_freq = arch_scale_freq_capacity(NULL, cpu);
2544 delta = now - sa->last_runnable_update;
2546 * This should only happen when time goes backwards, which it
2547 * unfortunately does during sched clock init when we swap over to TSC.
2549 if ((s64)delta < 0) {
2550 sa->last_runnable_update = now;
2555 * Use 1024ns as the unit of measurement since it's a reasonable
2556 * approximation of 1us and fast to compute.
2561 sa->last_runnable_update = now;
2563 /* delta_w is the amount already accumulated against our next period */
2564 delta_w = sa->avg_period % 1024;
2565 if (delta + delta_w >= 1024) {
2566 /* period roll-over */
2570 * Now that we know we're crossing a period boundary, figure
2571 * out how much from delta we need to complete the current
2572 * period and accrue it.
2574 delta_w = 1024 - delta_w;
2576 sa->runnable_avg_sum += delta_w;
2578 sa->running_avg_sum += delta_w * scale_freq
2579 >> SCHED_CAPACITY_SHIFT;
2580 sa->avg_period += delta_w;
2584 /* Figure out how many additional periods this update spans */
2585 periods = delta / 1024;
2588 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2590 sa->running_avg_sum = decay_load(sa->running_avg_sum,
2592 sa->avg_period = decay_load(sa->avg_period,
2595 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2596 runnable_contrib = __compute_runnable_contrib(periods);
2598 sa->runnable_avg_sum += runnable_contrib;
2600 sa->running_avg_sum += runnable_contrib * scale_freq
2601 >> SCHED_CAPACITY_SHIFT;
2602 sa->avg_period += runnable_contrib;
2605 /* Remainder of delta accrued against u_0` */
2607 sa->runnable_avg_sum += delta;
2609 sa->running_avg_sum += delta * scale_freq
2610 >> SCHED_CAPACITY_SHIFT;
2611 sa->avg_period += delta;
2616 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2617 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2619 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2620 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2622 decays -= se->avg.decay_count;
2623 se->avg.decay_count = 0;
2627 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2628 se->avg.utilization_avg_contrib =
2629 decay_load(se->avg.utilization_avg_contrib, decays);
2634 #ifdef CONFIG_FAIR_GROUP_SCHED
2635 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2638 struct task_group *tg = cfs_rq->tg;
2641 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2642 tg_contrib -= cfs_rq->tg_load_contrib;
2647 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2648 atomic_long_add(tg_contrib, &tg->load_avg);
2649 cfs_rq->tg_load_contrib += tg_contrib;
2654 * Aggregate cfs_rq runnable averages into an equivalent task_group
2655 * representation for computing load contributions.
2657 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2658 struct cfs_rq *cfs_rq)
2660 struct task_group *tg = cfs_rq->tg;
2663 /* The fraction of a cpu used by this cfs_rq */
2664 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2665 sa->avg_period + 1);
2666 contrib -= cfs_rq->tg_runnable_contrib;
2668 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2669 atomic_add(contrib, &tg->runnable_avg);
2670 cfs_rq->tg_runnable_contrib += contrib;
2674 static inline void __update_group_entity_contrib(struct sched_entity *se)
2676 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2677 struct task_group *tg = cfs_rq->tg;
2682 contrib = cfs_rq->tg_load_contrib * tg->shares;
2683 se->avg.load_avg_contrib = div_u64(contrib,
2684 atomic_long_read(&tg->load_avg) + 1);
2687 * For group entities we need to compute a correction term in the case
2688 * that they are consuming <1 cpu so that we would contribute the same
2689 * load as a task of equal weight.
2691 * Explicitly co-ordinating this measurement would be expensive, but
2692 * fortunately the sum of each cpus contribution forms a usable
2693 * lower-bound on the true value.
2695 * Consider the aggregate of 2 contributions. Either they are disjoint
2696 * (and the sum represents true value) or they are disjoint and we are
2697 * understating by the aggregate of their overlap.
2699 * Extending this to N cpus, for a given overlap, the maximum amount we
2700 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2701 * cpus that overlap for this interval and w_i is the interval width.
2703 * On a small machine; the first term is well-bounded which bounds the
2704 * total error since w_i is a subset of the period. Whereas on a
2705 * larger machine, while this first term can be larger, if w_i is the
2706 * of consequential size guaranteed to see n_i*w_i quickly converge to
2707 * our upper bound of 1-cpu.
2709 runnable_avg = atomic_read(&tg->runnable_avg);
2710 if (runnable_avg < NICE_0_LOAD) {
2711 se->avg.load_avg_contrib *= runnable_avg;
2712 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2716 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2718 __update_entity_runnable_avg(rq_clock_task(rq), cpu_of(rq), &rq->avg,
2719 runnable, runnable);
2720 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2722 #else /* CONFIG_FAIR_GROUP_SCHED */
2723 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2724 int force_update) {}
2725 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2726 struct cfs_rq *cfs_rq) {}
2727 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2728 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2729 #endif /* CONFIG_FAIR_GROUP_SCHED */
2731 static inline void __update_task_entity_contrib(struct sched_entity *se)
2735 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2736 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2737 contrib /= (se->avg.avg_period + 1);
2738 se->avg.load_avg_contrib = scale_load(contrib);
2741 /* Compute the current contribution to load_avg by se, return any delta */
2742 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2744 long old_contrib = se->avg.load_avg_contrib;
2746 if (entity_is_task(se)) {
2747 __update_task_entity_contrib(se);
2749 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2750 __update_group_entity_contrib(se);
2753 return se->avg.load_avg_contrib - old_contrib;
2757 static inline void __update_task_entity_utilization(struct sched_entity *se)
2761 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2762 contrib = se->avg.running_avg_sum * scale_load_down(SCHED_LOAD_SCALE);
2763 contrib /= (se->avg.avg_period + 1);
2764 se->avg.utilization_avg_contrib = scale_load(contrib);
2767 static long __update_entity_utilization_avg_contrib(struct sched_entity *se)
2769 long old_contrib = se->avg.utilization_avg_contrib;
2771 if (entity_is_task(se))
2772 __update_task_entity_utilization(se);
2774 se->avg.utilization_avg_contrib =
2775 group_cfs_rq(se)->utilization_load_avg;
2777 return se->avg.utilization_avg_contrib - old_contrib;
2780 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2783 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2784 cfs_rq->blocked_load_avg -= load_contrib;
2786 cfs_rq->blocked_load_avg = 0;
2789 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2791 /* Update a sched_entity's runnable average */
2792 static inline void update_entity_load_avg(struct sched_entity *se,
2795 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2796 long contrib_delta, utilization_delta;
2797 int cpu = cpu_of(rq_of(cfs_rq));
2801 * For a group entity we need to use their owned cfs_rq_clock_task() in
2802 * case they are the parent of a throttled hierarchy.
2804 if (entity_is_task(se))
2805 now = cfs_rq_clock_task(cfs_rq);
2807 now = cfs_rq_clock_task(group_cfs_rq(se));
2809 if (!__update_entity_runnable_avg(now, cpu, &se->avg, se->on_rq,
2810 cfs_rq->curr == se))
2813 contrib_delta = __update_entity_load_avg_contrib(se);
2814 utilization_delta = __update_entity_utilization_avg_contrib(se);
2820 cfs_rq->runnable_load_avg += contrib_delta;
2821 cfs_rq->utilization_load_avg += utilization_delta;
2823 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2828 * Decay the load contributed by all blocked children and account this so that
2829 * their contribution may appropriately discounted when they wake up.
2831 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2833 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2836 decays = now - cfs_rq->last_decay;
2837 if (!decays && !force_update)
2840 if (atomic_long_read(&cfs_rq->removed_load)) {
2841 unsigned long removed_load;
2842 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2843 subtract_blocked_load_contrib(cfs_rq, removed_load);
2847 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2849 atomic64_add(decays, &cfs_rq->decay_counter);
2850 cfs_rq->last_decay = now;
2853 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2856 /* Add the load generated by se into cfs_rq's child load-average */
2857 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2858 struct sched_entity *se,
2862 * We track migrations using entity decay_count <= 0, on a wake-up
2863 * migration we use a negative decay count to track the remote decays
2864 * accumulated while sleeping.
2866 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2867 * are seen by enqueue_entity_load_avg() as a migration with an already
2868 * constructed load_avg_contrib.
2870 if (unlikely(se->avg.decay_count <= 0)) {
2871 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2872 if (se->avg.decay_count) {
2874 * In a wake-up migration we have to approximate the
2875 * time sleeping. This is because we can't synchronize
2876 * clock_task between the two cpus, and it is not
2877 * guaranteed to be read-safe. Instead, we can
2878 * approximate this using our carried decays, which are
2879 * explicitly atomically readable.
2881 se->avg.last_runnable_update -= (-se->avg.decay_count)
2883 update_entity_load_avg(se, 0);
2884 /* Indicate that we're now synchronized and on-rq */
2885 se->avg.decay_count = 0;
2889 __synchronize_entity_decay(se);
2892 /* migrated tasks did not contribute to our blocked load */
2894 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2895 update_entity_load_avg(se, 0);
2898 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2899 cfs_rq->utilization_load_avg += se->avg.utilization_avg_contrib;
2900 /* we force update consideration on load-balancer moves */
2901 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2905 * Remove se's load from this cfs_rq child load-average, if the entity is
2906 * transitioning to a blocked state we track its projected decay using
2909 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2910 struct sched_entity *se,
2913 update_entity_load_avg(se, 1);
2914 /* we force update consideration on load-balancer moves */
2915 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2917 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2918 cfs_rq->utilization_load_avg -= se->avg.utilization_avg_contrib;
2920 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2921 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2922 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2926 * Update the rq's load with the elapsed running time before entering
2927 * idle. if the last scheduled task is not a CFS task, idle_enter will
2928 * be the only way to update the runnable statistic.
2930 void idle_enter_fair(struct rq *this_rq)
2932 update_rq_runnable_avg(this_rq, 1);
2936 * Update the rq's load with the elapsed idle time before a task is
2937 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2938 * be the only way to update the runnable statistic.
2940 void idle_exit_fair(struct rq *this_rq)
2942 update_rq_runnable_avg(this_rq, 0);
2945 static int idle_balance(struct rq *this_rq);
2947 #else /* CONFIG_SMP */
2949 static inline void update_entity_load_avg(struct sched_entity *se,
2950 int update_cfs_rq) {}
2951 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2952 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2953 struct sched_entity *se,
2955 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2956 struct sched_entity *se,
2958 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2959 int force_update) {}
2961 static inline int idle_balance(struct rq *rq)
2966 #endif /* CONFIG_SMP */
2968 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2970 #ifdef CONFIG_SCHEDSTATS
2971 struct task_struct *tsk = NULL;
2973 if (entity_is_task(se))
2976 if (se->statistics.sleep_start) {
2977 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2982 if (unlikely(delta > se->statistics.sleep_max))
2983 se->statistics.sleep_max = delta;
2985 se->statistics.sleep_start = 0;
2986 se->statistics.sum_sleep_runtime += delta;
2989 account_scheduler_latency(tsk, delta >> 10, 1);
2990 trace_sched_stat_sleep(tsk, delta);
2993 if (se->statistics.block_start) {
2994 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2999 if (unlikely(delta > se->statistics.block_max))
3000 se->statistics.block_max = delta;
3002 se->statistics.block_start = 0;
3003 se->statistics.sum_sleep_runtime += delta;
3006 if (tsk->in_iowait) {
3007 se->statistics.iowait_sum += delta;
3008 se->statistics.iowait_count++;
3009 trace_sched_stat_iowait(tsk, delta);
3012 trace_sched_stat_blocked(tsk, delta);
3015 * Blocking time is in units of nanosecs, so shift by
3016 * 20 to get a milliseconds-range estimation of the
3017 * amount of time that the task spent sleeping:
3019 if (unlikely(prof_on == SLEEP_PROFILING)) {
3020 profile_hits(SLEEP_PROFILING,
3021 (void *)get_wchan(tsk),
3024 account_scheduler_latency(tsk, delta >> 10, 0);
3030 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3032 #ifdef CONFIG_SCHED_DEBUG
3033 s64 d = se->vruntime - cfs_rq->min_vruntime;
3038 if (d > 3*sysctl_sched_latency)
3039 schedstat_inc(cfs_rq, nr_spread_over);
3044 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3046 u64 vruntime = cfs_rq->min_vruntime;
3049 * The 'current' period is already promised to the current tasks,
3050 * however the extra weight of the new task will slow them down a
3051 * little, place the new task so that it fits in the slot that
3052 * stays open at the end.
3054 if (initial && sched_feat(START_DEBIT))
3055 vruntime += sched_vslice(cfs_rq, se);
3057 /* sleeps up to a single latency don't count. */
3059 unsigned long thresh = sysctl_sched_latency;
3062 * Halve their sleep time's effect, to allow
3063 * for a gentler effect of sleepers:
3065 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3071 /* ensure we never gain time by being placed backwards. */
3072 se->vruntime = max_vruntime(se->vruntime, vruntime);
3075 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3078 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3081 * Update the normalized vruntime before updating min_vruntime
3082 * through calling update_curr().
3084 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3085 se->vruntime += cfs_rq->min_vruntime;
3088 * Update run-time statistics of the 'current'.
3090 update_curr(cfs_rq);
3091 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
3092 account_entity_enqueue(cfs_rq, se);
3093 update_cfs_shares(cfs_rq);
3095 if (flags & ENQUEUE_WAKEUP) {
3096 place_entity(cfs_rq, se, 0);
3097 enqueue_sleeper(cfs_rq, se);
3100 update_stats_enqueue(cfs_rq, se);
3101 check_spread(cfs_rq, se);
3102 if (se != cfs_rq->curr)
3103 __enqueue_entity(cfs_rq, se);
3106 if (cfs_rq->nr_running == 1) {
3107 list_add_leaf_cfs_rq(cfs_rq);
3108 check_enqueue_throttle(cfs_rq);
3112 static void __clear_buddies_last(struct sched_entity *se)
3114 for_each_sched_entity(se) {
3115 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3116 if (cfs_rq->last != se)
3119 cfs_rq->last = NULL;
3123 static void __clear_buddies_next(struct sched_entity *se)
3125 for_each_sched_entity(se) {
3126 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3127 if (cfs_rq->next != se)
3130 cfs_rq->next = NULL;
3134 static void __clear_buddies_skip(struct sched_entity *se)
3136 for_each_sched_entity(se) {
3137 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3138 if (cfs_rq->skip != se)
3141 cfs_rq->skip = NULL;
3145 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3147 if (cfs_rq->last == se)
3148 __clear_buddies_last(se);
3150 if (cfs_rq->next == se)
3151 __clear_buddies_next(se);
3153 if (cfs_rq->skip == se)
3154 __clear_buddies_skip(se);
3157 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3160 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3163 * Update run-time statistics of the 'current'.
3165 update_curr(cfs_rq);
3166 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
3168 update_stats_dequeue(cfs_rq, se);
3169 if (flags & DEQUEUE_SLEEP) {
3170 #ifdef CONFIG_SCHEDSTATS
3171 if (entity_is_task(se)) {
3172 struct task_struct *tsk = task_of(se);
3174 if (tsk->state & TASK_INTERRUPTIBLE)
3175 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3176 if (tsk->state & TASK_UNINTERRUPTIBLE)
3177 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3182 clear_buddies(cfs_rq, se);
3184 if (se != cfs_rq->curr)
3185 __dequeue_entity(cfs_rq, se);
3187 account_entity_dequeue(cfs_rq, se);
3190 * Normalize the entity after updating the min_vruntime because the
3191 * update can refer to the ->curr item and we need to reflect this
3192 * movement in our normalized position.
3194 if (!(flags & DEQUEUE_SLEEP))
3195 se->vruntime -= cfs_rq->min_vruntime;
3197 /* return excess runtime on last dequeue */
3198 return_cfs_rq_runtime(cfs_rq);
3200 update_min_vruntime(cfs_rq);
3201 update_cfs_shares(cfs_rq);
3205 * Preempt the current task with a newly woken task if needed:
3208 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3210 unsigned long ideal_runtime, delta_exec;
3211 struct sched_entity *se;
3214 ideal_runtime = sched_slice(cfs_rq, curr);
3215 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3216 if (delta_exec > ideal_runtime) {
3217 resched_curr(rq_of(cfs_rq));
3219 * The current task ran long enough, ensure it doesn't get
3220 * re-elected due to buddy favours.
3222 clear_buddies(cfs_rq, curr);
3227 * Ensure that a task that missed wakeup preemption by a
3228 * narrow margin doesn't have to wait for a full slice.
3229 * This also mitigates buddy induced latencies under load.
3231 if (delta_exec < sysctl_sched_min_granularity)
3234 se = __pick_first_entity(cfs_rq);
3235 delta = curr->vruntime - se->vruntime;
3240 if (delta > ideal_runtime)
3241 resched_curr(rq_of(cfs_rq));
3245 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3247 /* 'current' is not kept within the tree. */
3250 * Any task has to be enqueued before it get to execute on
3251 * a CPU. So account for the time it spent waiting on the
3254 update_stats_wait_end(cfs_rq, se);
3255 __dequeue_entity(cfs_rq, se);
3256 update_entity_load_avg(se, 1);
3259 update_stats_curr_start(cfs_rq, se);
3261 #ifdef CONFIG_SCHEDSTATS
3263 * Track our maximum slice length, if the CPU's load is at
3264 * least twice that of our own weight (i.e. dont track it
3265 * when there are only lesser-weight tasks around):
3267 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3268 se->statistics.slice_max = max(se->statistics.slice_max,
3269 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3272 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3276 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3279 * Pick the next process, keeping these things in mind, in this order:
3280 * 1) keep things fair between processes/task groups
3281 * 2) pick the "next" process, since someone really wants that to run
3282 * 3) pick the "last" process, for cache locality
3283 * 4) do not run the "skip" process, if something else is available
3285 static struct sched_entity *
3286 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3288 struct sched_entity *left = __pick_first_entity(cfs_rq);
3289 struct sched_entity *se;
3292 * If curr is set we have to see if its left of the leftmost entity
3293 * still in the tree, provided there was anything in the tree at all.
3295 if (!left || (curr && entity_before(curr, left)))
3298 se = left; /* ideally we run the leftmost entity */
3301 * Avoid running the skip buddy, if running something else can
3302 * be done without getting too unfair.
3304 if (cfs_rq->skip == se) {
3305 struct sched_entity *second;
3308 second = __pick_first_entity(cfs_rq);
3310 second = __pick_next_entity(se);
3311 if (!second || (curr && entity_before(curr, second)))
3315 if (second && wakeup_preempt_entity(second, left) < 1)
3320 * Prefer last buddy, try to return the CPU to a preempted task.
3322 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3326 * Someone really wants this to run. If it's not unfair, run it.
3328 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3331 clear_buddies(cfs_rq, se);
3336 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3338 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3341 * If still on the runqueue then deactivate_task()
3342 * was not called and update_curr() has to be done:
3345 update_curr(cfs_rq);
3347 /* throttle cfs_rqs exceeding runtime */
3348 check_cfs_rq_runtime(cfs_rq);
3350 check_spread(cfs_rq, prev);
3352 update_stats_wait_start(cfs_rq, prev);
3353 /* Put 'current' back into the tree. */
3354 __enqueue_entity(cfs_rq, prev);
3355 /* in !on_rq case, update occurred at dequeue */
3356 update_entity_load_avg(prev, 1);
3358 cfs_rq->curr = NULL;
3362 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3365 * Update run-time statistics of the 'current'.
3367 update_curr(cfs_rq);
3370 * Ensure that runnable average is periodically updated.
3372 update_entity_load_avg(curr, 1);
3373 update_cfs_rq_blocked_load(cfs_rq, 1);
3374 update_cfs_shares(cfs_rq);
3376 #ifdef CONFIG_SCHED_HRTICK
3378 * queued ticks are scheduled to match the slice, so don't bother
3379 * validating it and just reschedule.
3382 resched_curr(rq_of(cfs_rq));
3386 * don't let the period tick interfere with the hrtick preemption
3388 if (!sched_feat(DOUBLE_TICK) &&
3389 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3393 if (cfs_rq->nr_running > 1)
3394 check_preempt_tick(cfs_rq, curr);
3398 /**************************************************
3399 * CFS bandwidth control machinery
3402 #ifdef CONFIG_CFS_BANDWIDTH
3404 #ifdef HAVE_JUMP_LABEL
3405 static struct static_key __cfs_bandwidth_used;
3407 static inline bool cfs_bandwidth_used(void)
3409 return static_key_false(&__cfs_bandwidth_used);
3412 void cfs_bandwidth_usage_inc(void)
3414 static_key_slow_inc(&__cfs_bandwidth_used);
3417 void cfs_bandwidth_usage_dec(void)
3419 static_key_slow_dec(&__cfs_bandwidth_used);
3421 #else /* HAVE_JUMP_LABEL */
3422 static bool cfs_bandwidth_used(void)
3427 void cfs_bandwidth_usage_inc(void) {}
3428 void cfs_bandwidth_usage_dec(void) {}
3429 #endif /* HAVE_JUMP_LABEL */
3432 * default period for cfs group bandwidth.
3433 * default: 0.1s, units: nanoseconds
3435 static inline u64 default_cfs_period(void)
3437 return 100000000ULL;
3440 static inline u64 sched_cfs_bandwidth_slice(void)
3442 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3446 * Replenish runtime according to assigned quota and update expiration time.
3447 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3448 * additional synchronization around rq->lock.
3450 * requires cfs_b->lock
3452 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3456 if (cfs_b->quota == RUNTIME_INF)
3459 now = sched_clock_cpu(smp_processor_id());
3460 cfs_b->runtime = cfs_b->quota;
3461 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3464 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3466 return &tg->cfs_bandwidth;
3469 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3470 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3472 if (unlikely(cfs_rq->throttle_count))
3473 return cfs_rq->throttled_clock_task;
3475 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3478 /* returns 0 on failure to allocate runtime */
3479 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3481 struct task_group *tg = cfs_rq->tg;
3482 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3483 u64 amount = 0, min_amount, expires;
3485 /* note: this is a positive sum as runtime_remaining <= 0 */
3486 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3488 raw_spin_lock(&cfs_b->lock);
3489 if (cfs_b->quota == RUNTIME_INF)
3490 amount = min_amount;
3493 * If the bandwidth pool has become inactive, then at least one
3494 * period must have elapsed since the last consumption.
3495 * Refresh the global state and ensure bandwidth timer becomes
3498 if (!cfs_b->timer_active) {
3499 __refill_cfs_bandwidth_runtime(cfs_b);
3500 __start_cfs_bandwidth(cfs_b, false);
3503 if (cfs_b->runtime > 0) {
3504 amount = min(cfs_b->runtime, min_amount);
3505 cfs_b->runtime -= amount;
3509 expires = cfs_b->runtime_expires;
3510 raw_spin_unlock(&cfs_b->lock);
3512 cfs_rq->runtime_remaining += amount;
3514 * we may have advanced our local expiration to account for allowed
3515 * spread between our sched_clock and the one on which runtime was
3518 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3519 cfs_rq->runtime_expires = expires;
3521 return cfs_rq->runtime_remaining > 0;
3525 * Note: This depends on the synchronization provided by sched_clock and the
3526 * fact that rq->clock snapshots this value.
3528 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3530 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3532 /* if the deadline is ahead of our clock, nothing to do */
3533 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3536 if (cfs_rq->runtime_remaining < 0)
3540 * If the local deadline has passed we have to consider the
3541 * possibility that our sched_clock is 'fast' and the global deadline
3542 * has not truly expired.
3544 * Fortunately we can check determine whether this the case by checking
3545 * whether the global deadline has advanced. It is valid to compare
3546 * cfs_b->runtime_expires without any locks since we only care about
3547 * exact equality, so a partial write will still work.
3550 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3551 /* extend local deadline, drift is bounded above by 2 ticks */
3552 cfs_rq->runtime_expires += TICK_NSEC;
3554 /* global deadline is ahead, expiration has passed */
3555 cfs_rq->runtime_remaining = 0;
3559 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3561 /* dock delta_exec before expiring quota (as it could span periods) */
3562 cfs_rq->runtime_remaining -= delta_exec;
3563 expire_cfs_rq_runtime(cfs_rq);
3565 if (likely(cfs_rq->runtime_remaining > 0))
3569 * if we're unable to extend our runtime we resched so that the active
3570 * hierarchy can be throttled
3572 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3573 resched_curr(rq_of(cfs_rq));
3576 static __always_inline
3577 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3579 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3582 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3585 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3587 return cfs_bandwidth_used() && cfs_rq->throttled;
3590 /* check whether cfs_rq, or any parent, is throttled */
3591 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3593 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3597 * Ensure that neither of the group entities corresponding to src_cpu or
3598 * dest_cpu are members of a throttled hierarchy when performing group
3599 * load-balance operations.
3601 static inline int throttled_lb_pair(struct task_group *tg,
3602 int src_cpu, int dest_cpu)
3604 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3606 src_cfs_rq = tg->cfs_rq[src_cpu];
3607 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3609 return throttled_hierarchy(src_cfs_rq) ||
3610 throttled_hierarchy(dest_cfs_rq);
3613 /* updated child weight may affect parent so we have to do this bottom up */
3614 static int tg_unthrottle_up(struct task_group *tg, void *data)
3616 struct rq *rq = data;
3617 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3619 cfs_rq->throttle_count--;
3621 if (!cfs_rq->throttle_count) {
3622 /* adjust cfs_rq_clock_task() */
3623 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3624 cfs_rq->throttled_clock_task;
3631 static int tg_throttle_down(struct task_group *tg, void *data)
3633 struct rq *rq = data;
3634 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3636 /* group is entering throttled state, stop time */
3637 if (!cfs_rq->throttle_count)
3638 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3639 cfs_rq->throttle_count++;
3644 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3646 struct rq *rq = rq_of(cfs_rq);
3647 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3648 struct sched_entity *se;
3649 long task_delta, dequeue = 1;
3651 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3653 /* freeze hierarchy runnable averages while throttled */
3655 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3658 task_delta = cfs_rq->h_nr_running;
3659 for_each_sched_entity(se) {
3660 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3661 /* throttled entity or throttle-on-deactivate */
3666 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3667 qcfs_rq->h_nr_running -= task_delta;
3669 if (qcfs_rq->load.weight)
3674 sub_nr_running(rq, task_delta);
3676 cfs_rq->throttled = 1;
3677 cfs_rq->throttled_clock = rq_clock(rq);
3678 raw_spin_lock(&cfs_b->lock);
3680 * Add to the _head_ of the list, so that an already-started
3681 * distribute_cfs_runtime will not see us
3683 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3684 if (!cfs_b->timer_active)
3685 __start_cfs_bandwidth(cfs_b, false);
3686 raw_spin_unlock(&cfs_b->lock);
3689 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3691 struct rq *rq = rq_of(cfs_rq);
3692 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3693 struct sched_entity *se;
3697 se = cfs_rq->tg->se[cpu_of(rq)];
3699 cfs_rq->throttled = 0;
3701 update_rq_clock(rq);
3703 raw_spin_lock(&cfs_b->lock);
3704 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3705 list_del_rcu(&cfs_rq->throttled_list);
3706 raw_spin_unlock(&cfs_b->lock);
3708 /* update hierarchical throttle state */
3709 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3711 if (!cfs_rq->load.weight)
3714 task_delta = cfs_rq->h_nr_running;
3715 for_each_sched_entity(se) {
3719 cfs_rq = cfs_rq_of(se);
3721 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3722 cfs_rq->h_nr_running += task_delta;
3724 if (cfs_rq_throttled(cfs_rq))
3729 add_nr_running(rq, task_delta);
3731 /* determine whether we need to wake up potentially idle cpu */
3732 if (rq->curr == rq->idle && rq->cfs.nr_running)
3736 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3737 u64 remaining, u64 expires)
3739 struct cfs_rq *cfs_rq;
3741 u64 starting_runtime = remaining;
3744 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3746 struct rq *rq = rq_of(cfs_rq);
3748 raw_spin_lock(&rq->lock);
3749 if (!cfs_rq_throttled(cfs_rq))
3752 runtime = -cfs_rq->runtime_remaining + 1;
3753 if (runtime > remaining)
3754 runtime = remaining;
3755 remaining -= runtime;
3757 cfs_rq->runtime_remaining += runtime;
3758 cfs_rq->runtime_expires = expires;
3760 /* we check whether we're throttled above */
3761 if (cfs_rq->runtime_remaining > 0)
3762 unthrottle_cfs_rq(cfs_rq);
3765 raw_spin_unlock(&rq->lock);
3772 return starting_runtime - remaining;
3776 * Responsible for refilling a task_group's bandwidth and unthrottling its
3777 * cfs_rqs as appropriate. If there has been no activity within the last
3778 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3779 * used to track this state.
3781 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3783 u64 runtime, runtime_expires;
3786 /* no need to continue the timer with no bandwidth constraint */
3787 if (cfs_b->quota == RUNTIME_INF)
3788 goto out_deactivate;
3790 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3791 cfs_b->nr_periods += overrun;
3794 * idle depends on !throttled (for the case of a large deficit), and if
3795 * we're going inactive then everything else can be deferred
3797 if (cfs_b->idle && !throttled)
3798 goto out_deactivate;
3801 * if we have relooped after returning idle once, we need to update our
3802 * status as actually running, so that other cpus doing
3803 * __start_cfs_bandwidth will stop trying to cancel us.
3805 cfs_b->timer_active = 1;
3807 __refill_cfs_bandwidth_runtime(cfs_b);
3810 /* mark as potentially idle for the upcoming period */
3815 /* account preceding periods in which throttling occurred */
3816 cfs_b->nr_throttled += overrun;
3818 runtime_expires = cfs_b->runtime_expires;
3821 * This check is repeated as we are holding onto the new bandwidth while
3822 * we unthrottle. This can potentially race with an unthrottled group
3823 * trying to acquire new bandwidth from the global pool. This can result
3824 * in us over-using our runtime if it is all used during this loop, but
3825 * only by limited amounts in that extreme case.
3827 while (throttled && cfs_b->runtime > 0) {
3828 runtime = cfs_b->runtime;
3829 raw_spin_unlock(&cfs_b->lock);
3830 /* we can't nest cfs_b->lock while distributing bandwidth */
3831 runtime = distribute_cfs_runtime(cfs_b, runtime,
3833 raw_spin_lock(&cfs_b->lock);
3835 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3837 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3841 * While we are ensured activity in the period following an
3842 * unthrottle, this also covers the case in which the new bandwidth is
3843 * insufficient to cover the existing bandwidth deficit. (Forcing the
3844 * timer to remain active while there are any throttled entities.)
3851 cfs_b->timer_active = 0;
3855 /* a cfs_rq won't donate quota below this amount */
3856 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3857 /* minimum remaining period time to redistribute slack quota */
3858 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3859 /* how long we wait to gather additional slack before distributing */
3860 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3863 * Are we near the end of the current quota period?
3865 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3866 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3867 * migrate_hrtimers, base is never cleared, so we are fine.
3869 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3871 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3874 /* if the call-back is running a quota refresh is already occurring */
3875 if (hrtimer_callback_running(refresh_timer))
3878 /* is a quota refresh about to occur? */
3879 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3880 if (remaining < min_expire)
3886 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3888 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3890 /* if there's a quota refresh soon don't bother with slack */
3891 if (runtime_refresh_within(cfs_b, min_left))
3894 start_bandwidth_timer(&cfs_b->slack_timer,
3895 ns_to_ktime(cfs_bandwidth_slack_period));
3898 /* we know any runtime found here is valid as update_curr() precedes return */
3899 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3901 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3902 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3904 if (slack_runtime <= 0)
3907 raw_spin_lock(&cfs_b->lock);
3908 if (cfs_b->quota != RUNTIME_INF &&
3909 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3910 cfs_b->runtime += slack_runtime;
3912 /* we are under rq->lock, defer unthrottling using a timer */
3913 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3914 !list_empty(&cfs_b->throttled_cfs_rq))
3915 start_cfs_slack_bandwidth(cfs_b);
3917 raw_spin_unlock(&cfs_b->lock);
3919 /* even if it's not valid for return we don't want to try again */
3920 cfs_rq->runtime_remaining -= slack_runtime;
3923 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3925 if (!cfs_bandwidth_used())
3928 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3931 __return_cfs_rq_runtime(cfs_rq);
3935 * This is done with a timer (instead of inline with bandwidth return) since
3936 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3938 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3940 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3943 /* confirm we're still not at a refresh boundary */
3944 raw_spin_lock(&cfs_b->lock);
3945 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3946 raw_spin_unlock(&cfs_b->lock);
3950 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3951 runtime = cfs_b->runtime;
3953 expires = cfs_b->runtime_expires;
3954 raw_spin_unlock(&cfs_b->lock);
3959 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3961 raw_spin_lock(&cfs_b->lock);
3962 if (expires == cfs_b->runtime_expires)
3963 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3964 raw_spin_unlock(&cfs_b->lock);
3968 * When a group wakes up we want to make sure that its quota is not already
3969 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3970 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3972 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3974 if (!cfs_bandwidth_used())
3977 /* an active group must be handled by the update_curr()->put() path */
3978 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3981 /* ensure the group is not already throttled */
3982 if (cfs_rq_throttled(cfs_rq))
3985 /* update runtime allocation */
3986 account_cfs_rq_runtime(cfs_rq, 0);
3987 if (cfs_rq->runtime_remaining <= 0)
3988 throttle_cfs_rq(cfs_rq);
3991 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3992 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3994 if (!cfs_bandwidth_used())
3997 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4001 * it's possible for a throttled entity to be forced into a running
4002 * state (e.g. set_curr_task), in this case we're finished.
4004 if (cfs_rq_throttled(cfs_rq))
4007 throttle_cfs_rq(cfs_rq);
4011 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4013 struct cfs_bandwidth *cfs_b =
4014 container_of(timer, struct cfs_bandwidth, slack_timer);
4015 do_sched_cfs_slack_timer(cfs_b);
4017 return HRTIMER_NORESTART;
4020 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4022 struct cfs_bandwidth *cfs_b =
4023 container_of(timer, struct cfs_bandwidth, period_timer);
4028 raw_spin_lock(&cfs_b->lock);
4030 now = hrtimer_cb_get_time(timer);
4031 overrun = hrtimer_forward(timer, now, cfs_b->period);
4036 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4038 raw_spin_unlock(&cfs_b->lock);
4040 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4043 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4045 raw_spin_lock_init(&cfs_b->lock);
4047 cfs_b->quota = RUNTIME_INF;
4048 cfs_b->period = ns_to_ktime(default_cfs_period());
4050 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4051 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4052 cfs_b->period_timer.function = sched_cfs_period_timer;
4053 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4054 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4057 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4059 cfs_rq->runtime_enabled = 0;
4060 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4063 /* requires cfs_b->lock, may release to reprogram timer */
4064 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force)
4067 * The timer may be active because we're trying to set a new bandwidth
4068 * period or because we're racing with the tear-down path
4069 * (timer_active==0 becomes visible before the hrtimer call-back
4070 * terminates). In either case we ensure that it's re-programmed
4072 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
4073 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
4074 /* bounce the lock to allow do_sched_cfs_period_timer to run */
4075 raw_spin_unlock(&cfs_b->lock);
4077 raw_spin_lock(&cfs_b->lock);
4078 /* if someone else restarted the timer then we're done */
4079 if (!force && cfs_b->timer_active)
4083 cfs_b->timer_active = 1;
4084 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
4087 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4089 /* init_cfs_bandwidth() was not called */
4090 if (!cfs_b->throttled_cfs_rq.next)
4093 hrtimer_cancel(&cfs_b->period_timer);
4094 hrtimer_cancel(&cfs_b->slack_timer);
4097 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4099 struct cfs_rq *cfs_rq;
4101 for_each_leaf_cfs_rq(rq, cfs_rq) {
4102 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4104 raw_spin_lock(&cfs_b->lock);
4105 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4106 raw_spin_unlock(&cfs_b->lock);
4110 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4112 struct cfs_rq *cfs_rq;
4114 for_each_leaf_cfs_rq(rq, cfs_rq) {
4115 if (!cfs_rq->runtime_enabled)
4119 * clock_task is not advancing so we just need to make sure
4120 * there's some valid quota amount
4122 cfs_rq->runtime_remaining = 1;
4124 * Offline rq is schedulable till cpu is completely disabled
4125 * in take_cpu_down(), so we prevent new cfs throttling here.
4127 cfs_rq->runtime_enabled = 0;
4129 if (cfs_rq_throttled(cfs_rq))
4130 unthrottle_cfs_rq(cfs_rq);
4134 #else /* CONFIG_CFS_BANDWIDTH */
4135 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4137 return rq_clock_task(rq_of(cfs_rq));
4140 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4141 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4142 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4143 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4145 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4150 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4155 static inline int throttled_lb_pair(struct task_group *tg,
4156 int src_cpu, int dest_cpu)
4161 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4163 #ifdef CONFIG_FAIR_GROUP_SCHED
4164 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4167 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4171 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4172 static inline void update_runtime_enabled(struct rq *rq) {}
4173 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4175 #endif /* CONFIG_CFS_BANDWIDTH */
4177 /**************************************************
4178 * CFS operations on tasks:
4181 #ifdef CONFIG_SCHED_HRTICK
4182 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4184 struct sched_entity *se = &p->se;
4185 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4187 WARN_ON(task_rq(p) != rq);
4189 if (cfs_rq->nr_running > 1) {
4190 u64 slice = sched_slice(cfs_rq, se);
4191 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4192 s64 delta = slice - ran;
4199 hrtick_start(rq, delta);
4204 * called from enqueue/dequeue and updates the hrtick when the
4205 * current task is from our class and nr_running is low enough
4208 static void hrtick_update(struct rq *rq)
4210 struct task_struct *curr = rq->curr;
4212 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4215 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4216 hrtick_start_fair(rq, curr);
4218 #else /* !CONFIG_SCHED_HRTICK */
4220 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4224 static inline void hrtick_update(struct rq *rq)
4230 * The enqueue_task method is called before nr_running is
4231 * increased. Here we update the fair scheduling stats and
4232 * then put the task into the rbtree:
4235 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4237 struct cfs_rq *cfs_rq;
4238 struct sched_entity *se = &p->se;
4240 for_each_sched_entity(se) {
4243 cfs_rq = cfs_rq_of(se);
4244 enqueue_entity(cfs_rq, se, flags);
4247 * end evaluation on encountering a throttled cfs_rq
4249 * note: in the case of encountering a throttled cfs_rq we will
4250 * post the final h_nr_running increment below.
4252 if (cfs_rq_throttled(cfs_rq))
4254 cfs_rq->h_nr_running++;
4256 flags = ENQUEUE_WAKEUP;
4259 for_each_sched_entity(se) {
4260 cfs_rq = cfs_rq_of(se);
4261 cfs_rq->h_nr_running++;
4263 if (cfs_rq_throttled(cfs_rq))
4266 update_cfs_shares(cfs_rq);
4267 update_entity_load_avg(se, 1);
4271 update_rq_runnable_avg(rq, rq->nr_running);
4272 add_nr_running(rq, 1);
4277 static void set_next_buddy(struct sched_entity *se);
4280 * The dequeue_task method is called before nr_running is
4281 * decreased. We remove the task from the rbtree and
4282 * update the fair scheduling stats:
4284 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4286 struct cfs_rq *cfs_rq;
4287 struct sched_entity *se = &p->se;
4288 int task_sleep = flags & DEQUEUE_SLEEP;
4290 for_each_sched_entity(se) {
4291 cfs_rq = cfs_rq_of(se);
4292 dequeue_entity(cfs_rq, se, flags);
4295 * end evaluation on encountering a throttled cfs_rq
4297 * note: in the case of encountering a throttled cfs_rq we will
4298 * post the final h_nr_running decrement below.
4300 if (cfs_rq_throttled(cfs_rq))
4302 cfs_rq->h_nr_running--;
4304 /* Don't dequeue parent if it has other entities besides us */
4305 if (cfs_rq->load.weight) {
4307 * Bias pick_next to pick a task from this cfs_rq, as
4308 * p is sleeping when it is within its sched_slice.
4310 if (task_sleep && parent_entity(se))
4311 set_next_buddy(parent_entity(se));
4313 /* avoid re-evaluating load for this entity */
4314 se = parent_entity(se);
4317 flags |= DEQUEUE_SLEEP;
4320 for_each_sched_entity(se) {
4321 cfs_rq = cfs_rq_of(se);
4322 cfs_rq->h_nr_running--;
4324 if (cfs_rq_throttled(cfs_rq))
4327 update_cfs_shares(cfs_rq);
4328 update_entity_load_avg(se, 1);
4332 sub_nr_running(rq, 1);
4333 update_rq_runnable_avg(rq, 1);
4341 * per rq 'load' arrray crap; XXX kill this.
4345 * The exact cpuload at various idx values, calculated at every tick would be
4346 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4348 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4349 * on nth tick when cpu may be busy, then we have:
4350 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4351 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4353 * decay_load_missed() below does efficient calculation of
4354 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4355 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4357 * The calculation is approximated on a 128 point scale.
4358 * degrade_zero_ticks is the number of ticks after which load at any
4359 * particular idx is approximated to be zero.
4360 * degrade_factor is a precomputed table, a row for each load idx.
4361 * Each column corresponds to degradation factor for a power of two ticks,
4362 * based on 128 point scale.
4364 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4365 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4367 * With this power of 2 load factors, we can degrade the load n times
4368 * by looking at 1 bits in n and doing as many mult/shift instead of
4369 * n mult/shifts needed by the exact degradation.
4371 #define DEGRADE_SHIFT 7
4372 static const unsigned char
4373 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4374 static const unsigned char
4375 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4376 {0, 0, 0, 0, 0, 0, 0, 0},
4377 {64, 32, 8, 0, 0, 0, 0, 0},
4378 {96, 72, 40, 12, 1, 0, 0},
4379 {112, 98, 75, 43, 15, 1, 0},
4380 {120, 112, 98, 76, 45, 16, 2} };
4383 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4384 * would be when CPU is idle and so we just decay the old load without
4385 * adding any new load.
4387 static unsigned long
4388 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4392 if (!missed_updates)
4395 if (missed_updates >= degrade_zero_ticks[idx])
4399 return load >> missed_updates;
4401 while (missed_updates) {
4402 if (missed_updates % 2)
4403 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4405 missed_updates >>= 1;
4412 * Update rq->cpu_load[] statistics. This function is usually called every
4413 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4414 * every tick. We fix it up based on jiffies.
4416 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4417 unsigned long pending_updates)
4421 this_rq->nr_load_updates++;
4423 /* Update our load: */
4424 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4425 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4426 unsigned long old_load, new_load;
4428 /* scale is effectively 1 << i now, and >> i divides by scale */
4430 old_load = this_rq->cpu_load[i];
4431 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4432 new_load = this_load;
4434 * Round up the averaging division if load is increasing. This
4435 * prevents us from getting stuck on 9 if the load is 10, for
4438 if (new_load > old_load)
4439 new_load += scale - 1;
4441 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4444 sched_avg_update(this_rq);
4447 #ifdef CONFIG_NO_HZ_COMMON
4449 * There is no sane way to deal with nohz on smp when using jiffies because the
4450 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4451 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4453 * Therefore we cannot use the delta approach from the regular tick since that
4454 * would seriously skew the load calculation. However we'll make do for those
4455 * updates happening while idle (nohz_idle_balance) or coming out of idle
4456 * (tick_nohz_idle_exit).
4458 * This means we might still be one tick off for nohz periods.
4462 * Called from nohz_idle_balance() to update the load ratings before doing the
4465 static void update_idle_cpu_load(struct rq *this_rq)
4467 unsigned long curr_jiffies = READ_ONCE(jiffies);
4468 unsigned long load = this_rq->cfs.runnable_load_avg;
4469 unsigned long pending_updates;
4472 * bail if there's load or we're actually up-to-date.
4474 if (load || curr_jiffies == this_rq->last_load_update_tick)
4477 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4478 this_rq->last_load_update_tick = curr_jiffies;
4480 __update_cpu_load(this_rq, load, pending_updates);
4484 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4486 void update_cpu_load_nohz(void)
4488 struct rq *this_rq = this_rq();
4489 unsigned long curr_jiffies = READ_ONCE(jiffies);
4490 unsigned long pending_updates;
4492 if (curr_jiffies == this_rq->last_load_update_tick)
4495 raw_spin_lock(&this_rq->lock);
4496 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4497 if (pending_updates) {
4498 this_rq->last_load_update_tick = curr_jiffies;
4500 * We were idle, this means load 0, the current load might be
4501 * !0 due to remote wakeups and the sort.
4503 __update_cpu_load(this_rq, 0, pending_updates);
4505 raw_spin_unlock(&this_rq->lock);
4507 #endif /* CONFIG_NO_HZ */
4510 * Called from scheduler_tick()
4512 void update_cpu_load_active(struct rq *this_rq)
4514 unsigned long load = this_rq->cfs.runnable_load_avg;
4516 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4518 this_rq->last_load_update_tick = jiffies;
4519 __update_cpu_load(this_rq, load, 1);
4522 /* Used instead of source_load when we know the type == 0 */
4523 static unsigned long weighted_cpuload(const int cpu)
4525 return cpu_rq(cpu)->cfs.runnable_load_avg;
4529 * Return a low guess at the load of a migration-source cpu weighted
4530 * according to the scheduling class and "nice" value.
4532 * We want to under-estimate the load of migration sources, to
4533 * balance conservatively.
4535 static unsigned long source_load(int cpu, int type)
4537 struct rq *rq = cpu_rq(cpu);
4538 unsigned long total = weighted_cpuload(cpu);
4540 if (type == 0 || !sched_feat(LB_BIAS))
4543 return min(rq->cpu_load[type-1], total);
4547 * Return a high guess at the load of a migration-target cpu weighted
4548 * according to the scheduling class and "nice" value.
4550 static unsigned long target_load(int cpu, int type)
4552 struct rq *rq = cpu_rq(cpu);
4553 unsigned long total = weighted_cpuload(cpu);
4555 if (type == 0 || !sched_feat(LB_BIAS))
4558 return max(rq->cpu_load[type-1], total);
4561 static unsigned long capacity_of(int cpu)
4563 return cpu_rq(cpu)->cpu_capacity;
4566 static unsigned long capacity_orig_of(int cpu)
4568 return cpu_rq(cpu)->cpu_capacity_orig;
4571 static unsigned long cpu_avg_load_per_task(int cpu)
4573 struct rq *rq = cpu_rq(cpu);
4574 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4575 unsigned long load_avg = rq->cfs.runnable_load_avg;
4578 return load_avg / nr_running;
4583 static void record_wakee(struct task_struct *p)
4586 * Rough decay (wiping) for cost saving, don't worry
4587 * about the boundary, really active task won't care
4590 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4591 current->wakee_flips >>= 1;
4592 current->wakee_flip_decay_ts = jiffies;
4595 if (current->last_wakee != p) {
4596 current->last_wakee = p;
4597 current->wakee_flips++;
4601 static void task_waking_fair(struct task_struct *p)
4603 struct sched_entity *se = &p->se;
4604 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4607 #ifndef CONFIG_64BIT
4608 u64 min_vruntime_copy;
4611 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4613 min_vruntime = cfs_rq->min_vruntime;
4614 } while (min_vruntime != min_vruntime_copy);
4616 min_vruntime = cfs_rq->min_vruntime;
4619 se->vruntime -= min_vruntime;
4623 #ifdef CONFIG_FAIR_GROUP_SCHED
4625 * effective_load() calculates the load change as seen from the root_task_group
4627 * Adding load to a group doesn't make a group heavier, but can cause movement
4628 * of group shares between cpus. Assuming the shares were perfectly aligned one
4629 * can calculate the shift in shares.
4631 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4632 * on this @cpu and results in a total addition (subtraction) of @wg to the
4633 * total group weight.
4635 * Given a runqueue weight distribution (rw_i) we can compute a shares
4636 * distribution (s_i) using:
4638 * s_i = rw_i / \Sum rw_j (1)
4640 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4641 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4642 * shares distribution (s_i):
4644 * rw_i = { 2, 4, 1, 0 }
4645 * s_i = { 2/7, 4/7, 1/7, 0 }
4647 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4648 * task used to run on and the CPU the waker is running on), we need to
4649 * compute the effect of waking a task on either CPU and, in case of a sync
4650 * wakeup, compute the effect of the current task going to sleep.
4652 * So for a change of @wl to the local @cpu with an overall group weight change
4653 * of @wl we can compute the new shares distribution (s'_i) using:
4655 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4657 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4658 * differences in waking a task to CPU 0. The additional task changes the
4659 * weight and shares distributions like:
4661 * rw'_i = { 3, 4, 1, 0 }
4662 * s'_i = { 3/8, 4/8, 1/8, 0 }
4664 * We can then compute the difference in effective weight by using:
4666 * dw_i = S * (s'_i - s_i) (3)
4668 * Where 'S' is the group weight as seen by its parent.
4670 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4671 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4672 * 4/7) times the weight of the group.
4674 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4676 struct sched_entity *se = tg->se[cpu];
4678 if (!tg->parent) /* the trivial, non-cgroup case */
4681 for_each_sched_entity(se) {
4687 * W = @wg + \Sum rw_j
4689 W = wg + calc_tg_weight(tg, se->my_q);
4694 w = se->my_q->load.weight + wl;
4697 * wl = S * s'_i; see (2)
4700 wl = (w * (long)tg->shares) / W;
4705 * Per the above, wl is the new se->load.weight value; since
4706 * those are clipped to [MIN_SHARES, ...) do so now. See
4707 * calc_cfs_shares().
4709 if (wl < MIN_SHARES)
4713 * wl = dw_i = S * (s'_i - s_i); see (3)
4715 wl -= se->load.weight;
4718 * Recursively apply this logic to all parent groups to compute
4719 * the final effective load change on the root group. Since
4720 * only the @tg group gets extra weight, all parent groups can
4721 * only redistribute existing shares. @wl is the shift in shares
4722 * resulting from this level per the above.
4731 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4738 static int wake_wide(struct task_struct *p)
4740 int factor = this_cpu_read(sd_llc_size);
4743 * Yeah, it's the switching-frequency, could means many wakee or
4744 * rapidly switch, use factor here will just help to automatically
4745 * adjust the loose-degree, so bigger node will lead to more pull.
4747 if (p->wakee_flips > factor) {
4749 * wakee is somewhat hot, it needs certain amount of cpu
4750 * resource, so if waker is far more hot, prefer to leave
4753 if (current->wakee_flips > (factor * p->wakee_flips))
4760 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4762 s64 this_load, load;
4763 s64 this_eff_load, prev_eff_load;
4764 int idx, this_cpu, prev_cpu;
4765 struct task_group *tg;
4766 unsigned long weight;
4770 * If we wake multiple tasks be careful to not bounce
4771 * ourselves around too much.
4777 this_cpu = smp_processor_id();
4778 prev_cpu = task_cpu(p);
4779 load = source_load(prev_cpu, idx);
4780 this_load = target_load(this_cpu, idx);
4783 * If sync wakeup then subtract the (maximum possible)
4784 * effect of the currently running task from the load
4785 * of the current CPU:
4788 tg = task_group(current);
4789 weight = current->se.load.weight;
4791 this_load += effective_load(tg, this_cpu, -weight, -weight);
4792 load += effective_load(tg, prev_cpu, 0, -weight);
4796 weight = p->se.load.weight;
4799 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4800 * due to the sync cause above having dropped this_load to 0, we'll
4801 * always have an imbalance, but there's really nothing you can do
4802 * about that, so that's good too.
4804 * Otherwise check if either cpus are near enough in load to allow this
4805 * task to be woken on this_cpu.
4807 this_eff_load = 100;
4808 this_eff_load *= capacity_of(prev_cpu);
4810 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4811 prev_eff_load *= capacity_of(this_cpu);
4813 if (this_load > 0) {
4814 this_eff_load *= this_load +
4815 effective_load(tg, this_cpu, weight, weight);
4817 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4820 balanced = this_eff_load <= prev_eff_load;
4822 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4827 schedstat_inc(sd, ttwu_move_affine);
4828 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4834 * find_idlest_group finds and returns the least busy CPU group within the
4837 static struct sched_group *
4838 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4839 int this_cpu, int sd_flag)
4841 struct sched_group *idlest = NULL, *group = sd->groups;
4842 unsigned long min_load = ULONG_MAX, this_load = 0;
4843 int load_idx = sd->forkexec_idx;
4844 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4846 if (sd_flag & SD_BALANCE_WAKE)
4847 load_idx = sd->wake_idx;
4850 unsigned long load, avg_load;
4854 /* Skip over this group if it has no CPUs allowed */
4855 if (!cpumask_intersects(sched_group_cpus(group),
4856 tsk_cpus_allowed(p)))
4859 local_group = cpumask_test_cpu(this_cpu,
4860 sched_group_cpus(group));
4862 /* Tally up the load of all CPUs in the group */
4865 for_each_cpu(i, sched_group_cpus(group)) {
4866 /* Bias balancing toward cpus of our domain */
4868 load = source_load(i, load_idx);
4870 load = target_load(i, load_idx);
4875 /* Adjust by relative CPU capacity of the group */
4876 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4879 this_load = avg_load;
4880 } else if (avg_load < min_load) {
4881 min_load = avg_load;
4884 } while (group = group->next, group != sd->groups);
4886 if (!idlest || 100*this_load < imbalance*min_load)
4892 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4895 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4897 unsigned long load, min_load = ULONG_MAX;
4898 unsigned int min_exit_latency = UINT_MAX;
4899 u64 latest_idle_timestamp = 0;
4900 int least_loaded_cpu = this_cpu;
4901 int shallowest_idle_cpu = -1;
4904 /* Traverse only the allowed CPUs */
4905 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4907 struct rq *rq = cpu_rq(i);
4908 struct cpuidle_state *idle = idle_get_state(rq);
4909 if (idle && idle->exit_latency < min_exit_latency) {
4911 * We give priority to a CPU whose idle state
4912 * has the smallest exit latency irrespective
4913 * of any idle timestamp.
4915 min_exit_latency = idle->exit_latency;
4916 latest_idle_timestamp = rq->idle_stamp;
4917 shallowest_idle_cpu = i;
4918 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4919 rq->idle_stamp > latest_idle_timestamp) {
4921 * If equal or no active idle state, then
4922 * the most recently idled CPU might have
4925 latest_idle_timestamp = rq->idle_stamp;
4926 shallowest_idle_cpu = i;
4928 } else if (shallowest_idle_cpu == -1) {
4929 load = weighted_cpuload(i);
4930 if (load < min_load || (load == min_load && i == this_cpu)) {
4932 least_loaded_cpu = i;
4937 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4941 * Try and locate an idle CPU in the sched_domain.
4943 static int select_idle_sibling(struct task_struct *p, int target)
4945 struct sched_domain *sd;
4946 struct sched_group *sg;
4947 int i = task_cpu(p);
4949 if (idle_cpu(target))
4953 * If the prevous cpu is cache affine and idle, don't be stupid.
4955 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4959 * Otherwise, iterate the domains and find an elegible idle cpu.
4961 sd = rcu_dereference(per_cpu(sd_llc, target));
4962 for_each_lower_domain(sd) {
4965 if (!cpumask_intersects(sched_group_cpus(sg),
4966 tsk_cpus_allowed(p)))
4969 for_each_cpu(i, sched_group_cpus(sg)) {
4970 if (i == target || !idle_cpu(i))
4974 target = cpumask_first_and(sched_group_cpus(sg),
4975 tsk_cpus_allowed(p));
4979 } while (sg != sd->groups);
4985 * get_cpu_usage returns the amount of capacity of a CPU that is used by CFS
4986 * tasks. The unit of the return value must be the one of capacity so we can
4987 * compare the usage with the capacity of the CPU that is available for CFS
4988 * task (ie cpu_capacity).
4989 * cfs.utilization_load_avg is the sum of running time of runnable tasks on a
4990 * CPU. It represents the amount of utilization of a CPU in the range
4991 * [0..SCHED_LOAD_SCALE]. The usage of a CPU can't be higher than the full
4992 * capacity of the CPU because it's about the running time on this CPU.
4993 * Nevertheless, cfs.utilization_load_avg can be higher than SCHED_LOAD_SCALE
4994 * because of unfortunate rounding in avg_period and running_load_avg or just
4995 * after migrating tasks until the average stabilizes with the new running
4996 * time. So we need to check that the usage stays into the range
4997 * [0..cpu_capacity_orig] and cap if necessary.
4998 * Without capping the usage, a group could be seen as overloaded (CPU0 usage
4999 * at 121% + CPU1 usage at 80%) whereas CPU1 has 20% of available capacity
5001 static int get_cpu_usage(int cpu)
5003 unsigned long usage = cpu_rq(cpu)->cfs.utilization_load_avg;
5004 unsigned long capacity = capacity_orig_of(cpu);
5006 if (usage >= SCHED_LOAD_SCALE)
5009 return (usage * capacity) >> SCHED_LOAD_SHIFT;
5013 * select_task_rq_fair: Select target runqueue for the waking task in domains
5014 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5015 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5017 * Balances load by selecting the idlest cpu in the idlest group, or under
5018 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5020 * Returns the target cpu number.
5022 * preempt must be disabled.
5025 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5027 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5028 int cpu = smp_processor_id();
5030 int want_affine = 0;
5031 int sync = wake_flags & WF_SYNC;
5033 if (sd_flag & SD_BALANCE_WAKE)
5034 want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5037 for_each_domain(cpu, tmp) {
5038 if (!(tmp->flags & SD_LOAD_BALANCE))
5042 * If both cpu and prev_cpu are part of this domain,
5043 * cpu is a valid SD_WAKE_AFFINE target.
5045 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5046 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5051 if (tmp->flags & sd_flag)
5055 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5058 if (sd_flag & SD_BALANCE_WAKE) {
5059 new_cpu = select_idle_sibling(p, prev_cpu);
5064 struct sched_group *group;
5067 if (!(sd->flags & sd_flag)) {
5072 group = find_idlest_group(sd, p, cpu, sd_flag);
5078 new_cpu = find_idlest_cpu(group, p, cpu);
5079 if (new_cpu == -1 || new_cpu == cpu) {
5080 /* Now try balancing at a lower domain level of cpu */
5085 /* Now try balancing at a lower domain level of new_cpu */
5087 weight = sd->span_weight;
5089 for_each_domain(cpu, tmp) {
5090 if (weight <= tmp->span_weight)
5092 if (tmp->flags & sd_flag)
5095 /* while loop will break here if sd == NULL */
5104 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5105 * cfs_rq_of(p) references at time of call are still valid and identify the
5106 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5107 * other assumptions, including the state of rq->lock, should be made.
5110 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
5112 struct sched_entity *se = &p->se;
5113 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5116 * Load tracking: accumulate removed load so that it can be processed
5117 * when we next update owning cfs_rq under rq->lock. Tasks contribute
5118 * to blocked load iff they have a positive decay-count. It can never
5119 * be negative here since on-rq tasks have decay-count == 0.
5121 if (se->avg.decay_count) {
5122 se->avg.decay_count = -__synchronize_entity_decay(se);
5123 atomic_long_add(se->avg.load_avg_contrib,
5124 &cfs_rq->removed_load);
5127 /* We have migrated, no longer consider this task hot */
5130 #endif /* CONFIG_SMP */
5132 static unsigned long
5133 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5135 unsigned long gran = sysctl_sched_wakeup_granularity;
5138 * Since its curr running now, convert the gran from real-time
5139 * to virtual-time in his units.
5141 * By using 'se' instead of 'curr' we penalize light tasks, so
5142 * they get preempted easier. That is, if 'se' < 'curr' then
5143 * the resulting gran will be larger, therefore penalizing the
5144 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5145 * be smaller, again penalizing the lighter task.
5147 * This is especially important for buddies when the leftmost
5148 * task is higher priority than the buddy.
5150 return calc_delta_fair(gran, se);
5154 * Should 'se' preempt 'curr'.
5168 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5170 s64 gran, vdiff = curr->vruntime - se->vruntime;
5175 gran = wakeup_gran(curr, se);
5182 static void set_last_buddy(struct sched_entity *se)
5184 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5187 for_each_sched_entity(se)
5188 cfs_rq_of(se)->last = se;
5191 static void set_next_buddy(struct sched_entity *se)
5193 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5196 for_each_sched_entity(se)
5197 cfs_rq_of(se)->next = se;
5200 static void set_skip_buddy(struct sched_entity *se)
5202 for_each_sched_entity(se)
5203 cfs_rq_of(se)->skip = se;
5207 * Preempt the current task with a newly woken task if needed:
5209 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5211 struct task_struct *curr = rq->curr;
5212 struct sched_entity *se = &curr->se, *pse = &p->se;
5213 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5214 int scale = cfs_rq->nr_running >= sched_nr_latency;
5215 int next_buddy_marked = 0;
5217 if (unlikely(se == pse))
5221 * This is possible from callers such as attach_tasks(), in which we
5222 * unconditionally check_prempt_curr() after an enqueue (which may have
5223 * lead to a throttle). This both saves work and prevents false
5224 * next-buddy nomination below.
5226 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5229 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5230 set_next_buddy(pse);
5231 next_buddy_marked = 1;
5235 * We can come here with TIF_NEED_RESCHED already set from new task
5238 * Note: this also catches the edge-case of curr being in a throttled
5239 * group (e.g. via set_curr_task), since update_curr() (in the
5240 * enqueue of curr) will have resulted in resched being set. This
5241 * prevents us from potentially nominating it as a false LAST_BUDDY
5244 if (test_tsk_need_resched(curr))
5247 /* Idle tasks are by definition preempted by non-idle tasks. */
5248 if (unlikely(curr->policy == SCHED_IDLE) &&
5249 likely(p->policy != SCHED_IDLE))
5253 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5254 * is driven by the tick):
5256 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5259 find_matching_se(&se, &pse);
5260 update_curr(cfs_rq_of(se));
5262 if (wakeup_preempt_entity(se, pse) == 1) {
5264 * Bias pick_next to pick the sched entity that is
5265 * triggering this preemption.
5267 if (!next_buddy_marked)
5268 set_next_buddy(pse);
5277 * Only set the backward buddy when the current task is still
5278 * on the rq. This can happen when a wakeup gets interleaved
5279 * with schedule on the ->pre_schedule() or idle_balance()
5280 * point, either of which can * drop the rq lock.
5282 * Also, during early boot the idle thread is in the fair class,
5283 * for obvious reasons its a bad idea to schedule back to it.
5285 if (unlikely(!se->on_rq || curr == rq->idle))
5288 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5292 static struct task_struct *
5293 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5295 struct cfs_rq *cfs_rq = &rq->cfs;
5296 struct sched_entity *se;
5297 struct task_struct *p;
5301 #ifdef CONFIG_FAIR_GROUP_SCHED
5302 if (!cfs_rq->nr_running)
5305 if (prev->sched_class != &fair_sched_class)
5309 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5310 * likely that a next task is from the same cgroup as the current.
5312 * Therefore attempt to avoid putting and setting the entire cgroup
5313 * hierarchy, only change the part that actually changes.
5317 struct sched_entity *curr = cfs_rq->curr;
5320 * Since we got here without doing put_prev_entity() we also
5321 * have to consider cfs_rq->curr. If it is still a runnable
5322 * entity, update_curr() will update its vruntime, otherwise
5323 * forget we've ever seen it.
5325 if (curr && curr->on_rq)
5326 update_curr(cfs_rq);
5331 * This call to check_cfs_rq_runtime() will do the throttle and
5332 * dequeue its entity in the parent(s). Therefore the 'simple'
5333 * nr_running test will indeed be correct.
5335 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5338 se = pick_next_entity(cfs_rq, curr);
5339 cfs_rq = group_cfs_rq(se);
5345 * Since we haven't yet done put_prev_entity and if the selected task
5346 * is a different task than we started out with, try and touch the
5347 * least amount of cfs_rqs.
5350 struct sched_entity *pse = &prev->se;
5352 while (!(cfs_rq = is_same_group(se, pse))) {
5353 int se_depth = se->depth;
5354 int pse_depth = pse->depth;
5356 if (se_depth <= pse_depth) {
5357 put_prev_entity(cfs_rq_of(pse), pse);
5358 pse = parent_entity(pse);
5360 if (se_depth >= pse_depth) {
5361 set_next_entity(cfs_rq_of(se), se);
5362 se = parent_entity(se);
5366 put_prev_entity(cfs_rq, pse);
5367 set_next_entity(cfs_rq, se);
5370 if (hrtick_enabled(rq))
5371 hrtick_start_fair(rq, p);
5378 if (!cfs_rq->nr_running)
5381 put_prev_task(rq, prev);
5384 se = pick_next_entity(cfs_rq, NULL);
5385 set_next_entity(cfs_rq, se);
5386 cfs_rq = group_cfs_rq(se);
5391 if (hrtick_enabled(rq))
5392 hrtick_start_fair(rq, p);
5397 new_tasks = idle_balance(rq);
5399 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5400 * possible for any higher priority task to appear. In that case we
5401 * must re-start the pick_next_entity() loop.
5413 * Account for a descheduled task:
5415 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5417 struct sched_entity *se = &prev->se;
5418 struct cfs_rq *cfs_rq;
5420 for_each_sched_entity(se) {
5421 cfs_rq = cfs_rq_of(se);
5422 put_prev_entity(cfs_rq, se);
5427 * sched_yield() is very simple
5429 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5431 static void yield_task_fair(struct rq *rq)
5433 struct task_struct *curr = rq->curr;
5434 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5435 struct sched_entity *se = &curr->se;
5438 * Are we the only task in the tree?
5440 if (unlikely(rq->nr_running == 1))
5443 clear_buddies(cfs_rq, se);
5445 if (curr->policy != SCHED_BATCH) {
5446 update_rq_clock(rq);
5448 * Update run-time statistics of the 'current'.
5450 update_curr(cfs_rq);
5452 * Tell update_rq_clock() that we've just updated,
5453 * so we don't do microscopic update in schedule()
5454 * and double the fastpath cost.
5456 rq_clock_skip_update(rq, true);
5462 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5464 struct sched_entity *se = &p->se;
5466 /* throttled hierarchies are not runnable */
5467 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5470 /* Tell the scheduler that we'd really like pse to run next. */
5473 yield_task_fair(rq);
5479 /**************************************************
5480 * Fair scheduling class load-balancing methods.
5484 * The purpose of load-balancing is to achieve the same basic fairness the
5485 * per-cpu scheduler provides, namely provide a proportional amount of compute
5486 * time to each task. This is expressed in the following equation:
5488 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5490 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5491 * W_i,0 is defined as:
5493 * W_i,0 = \Sum_j w_i,j (2)
5495 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5496 * is derived from the nice value as per prio_to_weight[].
5498 * The weight average is an exponential decay average of the instantaneous
5501 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5503 * C_i is the compute capacity of cpu i, typically it is the
5504 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5505 * can also include other factors [XXX].
5507 * To achieve this balance we define a measure of imbalance which follows
5508 * directly from (1):
5510 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5512 * We them move tasks around to minimize the imbalance. In the continuous
5513 * function space it is obvious this converges, in the discrete case we get
5514 * a few fun cases generally called infeasible weight scenarios.
5517 * - infeasible weights;
5518 * - local vs global optima in the discrete case. ]
5523 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5524 * for all i,j solution, we create a tree of cpus that follows the hardware
5525 * topology where each level pairs two lower groups (or better). This results
5526 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5527 * tree to only the first of the previous level and we decrease the frequency
5528 * of load-balance at each level inv. proportional to the number of cpus in
5534 * \Sum { --- * --- * 2^i } = O(n) (5)
5536 * `- size of each group
5537 * | | `- number of cpus doing load-balance
5539 * `- sum over all levels
5541 * Coupled with a limit on how many tasks we can migrate every balance pass,
5542 * this makes (5) the runtime complexity of the balancer.
5544 * An important property here is that each CPU is still (indirectly) connected
5545 * to every other cpu in at most O(log n) steps:
5547 * The adjacency matrix of the resulting graph is given by:
5550 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5553 * And you'll find that:
5555 * A^(log_2 n)_i,j != 0 for all i,j (7)
5557 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5558 * The task movement gives a factor of O(m), giving a convergence complexity
5561 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5566 * In order to avoid CPUs going idle while there's still work to do, new idle
5567 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5568 * tree itself instead of relying on other CPUs to bring it work.
5570 * This adds some complexity to both (5) and (8) but it reduces the total idle
5578 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5581 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5586 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5588 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5590 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5593 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5594 * rewrite all of this once again.]
5597 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5599 enum fbq_type { regular, remote, all };
5601 #define LBF_ALL_PINNED 0x01
5602 #define LBF_NEED_BREAK 0x02
5603 #define LBF_DST_PINNED 0x04
5604 #define LBF_SOME_PINNED 0x08
5607 struct sched_domain *sd;
5615 struct cpumask *dst_grpmask;
5617 enum cpu_idle_type idle;
5619 /* The set of CPUs under consideration for load-balancing */
5620 struct cpumask *cpus;
5625 unsigned int loop_break;
5626 unsigned int loop_max;
5628 enum fbq_type fbq_type;
5629 struct list_head tasks;
5633 * Is this task likely cache-hot:
5635 static int task_hot(struct task_struct *p, struct lb_env *env)
5639 lockdep_assert_held(&env->src_rq->lock);
5641 if (p->sched_class != &fair_sched_class)
5644 if (unlikely(p->policy == SCHED_IDLE))
5648 * Buddy candidates are cache hot:
5650 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5651 (&p->se == cfs_rq_of(&p->se)->next ||
5652 &p->se == cfs_rq_of(&p->se)->last))
5655 if (sysctl_sched_migration_cost == -1)
5657 if (sysctl_sched_migration_cost == 0)
5660 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5662 return delta < (s64)sysctl_sched_migration_cost;
5665 #ifdef CONFIG_NUMA_BALANCING
5666 /* Returns true if the destination node has incurred more faults */
5667 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5669 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5670 int src_nid, dst_nid;
5672 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
5673 !(env->sd->flags & SD_NUMA)) {
5677 src_nid = cpu_to_node(env->src_cpu);
5678 dst_nid = cpu_to_node(env->dst_cpu);
5680 if (src_nid == dst_nid)
5684 /* Task is already in the group's interleave set. */
5685 if (node_isset(src_nid, numa_group->active_nodes))
5688 /* Task is moving into the group's interleave set. */
5689 if (node_isset(dst_nid, numa_group->active_nodes))
5692 return group_faults(p, dst_nid) > group_faults(p, src_nid);
5695 /* Encourage migration to the preferred node. */
5696 if (dst_nid == p->numa_preferred_nid)
5699 return task_faults(p, dst_nid) > task_faults(p, src_nid);
5703 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5705 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5706 int src_nid, dst_nid;
5708 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5711 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5714 src_nid = cpu_to_node(env->src_cpu);
5715 dst_nid = cpu_to_node(env->dst_cpu);
5717 if (src_nid == dst_nid)
5721 /* Task is moving within/into the group's interleave set. */
5722 if (node_isset(dst_nid, numa_group->active_nodes))
5725 /* Task is moving out of the group's interleave set. */
5726 if (node_isset(src_nid, numa_group->active_nodes))
5729 return group_faults(p, dst_nid) < group_faults(p, src_nid);
5732 /* Migrating away from the preferred node is always bad. */
5733 if (src_nid == p->numa_preferred_nid)
5736 return task_faults(p, dst_nid) < task_faults(p, src_nid);
5740 static inline bool migrate_improves_locality(struct task_struct *p,
5746 static inline bool migrate_degrades_locality(struct task_struct *p,
5754 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5757 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5759 int tsk_cache_hot = 0;
5761 lockdep_assert_held(&env->src_rq->lock);
5764 * We do not migrate tasks that are:
5765 * 1) throttled_lb_pair, or
5766 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5767 * 3) running (obviously), or
5768 * 4) are cache-hot on their current CPU.
5770 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5773 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5776 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5778 env->flags |= LBF_SOME_PINNED;
5781 * Remember if this task can be migrated to any other cpu in
5782 * our sched_group. We may want to revisit it if we couldn't
5783 * meet load balance goals by pulling other tasks on src_cpu.
5785 * Also avoid computing new_dst_cpu if we have already computed
5786 * one in current iteration.
5788 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5791 /* Prevent to re-select dst_cpu via env's cpus */
5792 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5793 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5794 env->flags |= LBF_DST_PINNED;
5795 env->new_dst_cpu = cpu;
5803 /* Record that we found atleast one task that could run on dst_cpu */
5804 env->flags &= ~LBF_ALL_PINNED;
5806 if (task_running(env->src_rq, p)) {
5807 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5812 * Aggressive migration if:
5813 * 1) destination numa is preferred
5814 * 2) task is cache cold, or
5815 * 3) too many balance attempts have failed.
5817 tsk_cache_hot = task_hot(p, env);
5819 tsk_cache_hot = migrate_degrades_locality(p, env);
5821 if (migrate_improves_locality(p, env) || !tsk_cache_hot ||
5822 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5823 if (tsk_cache_hot) {
5824 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5825 schedstat_inc(p, se.statistics.nr_forced_migrations);
5830 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5835 * detach_task() -- detach the task for the migration specified in env
5837 static void detach_task(struct task_struct *p, struct lb_env *env)
5839 lockdep_assert_held(&env->src_rq->lock);
5841 deactivate_task(env->src_rq, p, 0);
5842 p->on_rq = TASK_ON_RQ_MIGRATING;
5843 set_task_cpu(p, env->dst_cpu);
5847 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5848 * part of active balancing operations within "domain".
5850 * Returns a task if successful and NULL otherwise.
5852 static struct task_struct *detach_one_task(struct lb_env *env)
5854 struct task_struct *p, *n;
5856 lockdep_assert_held(&env->src_rq->lock);
5858 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5859 if (!can_migrate_task(p, env))
5862 detach_task(p, env);
5865 * Right now, this is only the second place where
5866 * lb_gained[env->idle] is updated (other is detach_tasks)
5867 * so we can safely collect stats here rather than
5868 * inside detach_tasks().
5870 schedstat_inc(env->sd, lb_gained[env->idle]);
5876 static const unsigned int sched_nr_migrate_break = 32;
5879 * detach_tasks() -- tries to detach up to imbalance weighted load from
5880 * busiest_rq, as part of a balancing operation within domain "sd".
5882 * Returns number of detached tasks if successful and 0 otherwise.
5884 static int detach_tasks(struct lb_env *env)
5886 struct list_head *tasks = &env->src_rq->cfs_tasks;
5887 struct task_struct *p;
5891 lockdep_assert_held(&env->src_rq->lock);
5893 if (env->imbalance <= 0)
5896 while (!list_empty(tasks)) {
5897 p = list_first_entry(tasks, struct task_struct, se.group_node);
5900 /* We've more or less seen every task there is, call it quits */
5901 if (env->loop > env->loop_max)
5904 /* take a breather every nr_migrate tasks */
5905 if (env->loop > env->loop_break) {
5906 env->loop_break += sched_nr_migrate_break;
5907 env->flags |= LBF_NEED_BREAK;
5911 if (!can_migrate_task(p, env))
5914 load = task_h_load(p);
5916 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5919 if ((load / 2) > env->imbalance)
5922 detach_task(p, env);
5923 list_add(&p->se.group_node, &env->tasks);
5926 env->imbalance -= load;
5928 #ifdef CONFIG_PREEMPT
5930 * NEWIDLE balancing is a source of latency, so preemptible
5931 * kernels will stop after the first task is detached to minimize
5932 * the critical section.
5934 if (env->idle == CPU_NEWLY_IDLE)
5939 * We only want to steal up to the prescribed amount of
5942 if (env->imbalance <= 0)
5947 list_move_tail(&p->se.group_node, tasks);
5951 * Right now, this is one of only two places we collect this stat
5952 * so we can safely collect detach_one_task() stats here rather
5953 * than inside detach_one_task().
5955 schedstat_add(env->sd, lb_gained[env->idle], detached);
5961 * attach_task() -- attach the task detached by detach_task() to its new rq.
5963 static void attach_task(struct rq *rq, struct task_struct *p)
5965 lockdep_assert_held(&rq->lock);
5967 BUG_ON(task_rq(p) != rq);
5968 p->on_rq = TASK_ON_RQ_QUEUED;
5969 activate_task(rq, p, 0);
5970 check_preempt_curr(rq, p, 0);
5974 * attach_one_task() -- attaches the task returned from detach_one_task() to
5977 static void attach_one_task(struct rq *rq, struct task_struct *p)
5979 raw_spin_lock(&rq->lock);
5981 raw_spin_unlock(&rq->lock);
5985 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5988 static void attach_tasks(struct lb_env *env)
5990 struct list_head *tasks = &env->tasks;
5991 struct task_struct *p;
5993 raw_spin_lock(&env->dst_rq->lock);
5995 while (!list_empty(tasks)) {
5996 p = list_first_entry(tasks, struct task_struct, se.group_node);
5997 list_del_init(&p->se.group_node);
5999 attach_task(env->dst_rq, p);
6002 raw_spin_unlock(&env->dst_rq->lock);
6005 #ifdef CONFIG_FAIR_GROUP_SCHED
6007 * update tg->load_weight by folding this cpu's load_avg
6009 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
6011 struct sched_entity *se = tg->se[cpu];
6012 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
6014 /* throttled entities do not contribute to load */
6015 if (throttled_hierarchy(cfs_rq))
6018 update_cfs_rq_blocked_load(cfs_rq, 1);
6021 update_entity_load_avg(se, 1);
6023 * We pivot on our runnable average having decayed to zero for
6024 * list removal. This generally implies that all our children
6025 * have also been removed (modulo rounding error or bandwidth
6026 * control); however, such cases are rare and we can fix these
6029 * TODO: fix up out-of-order children on enqueue.
6031 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
6032 list_del_leaf_cfs_rq(cfs_rq);
6034 struct rq *rq = rq_of(cfs_rq);
6035 update_rq_runnable_avg(rq, rq->nr_running);
6039 static void update_blocked_averages(int cpu)
6041 struct rq *rq = cpu_rq(cpu);
6042 struct cfs_rq *cfs_rq;
6043 unsigned long flags;
6045 raw_spin_lock_irqsave(&rq->lock, flags);
6046 update_rq_clock(rq);
6048 * Iterates the task_group tree in a bottom up fashion, see
6049 * list_add_leaf_cfs_rq() for details.
6051 for_each_leaf_cfs_rq(rq, cfs_rq) {
6053 * Note: We may want to consider periodically releasing
6054 * rq->lock about these updates so that creating many task
6055 * groups does not result in continually extending hold time.
6057 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
6060 raw_spin_unlock_irqrestore(&rq->lock, flags);
6064 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6065 * This needs to be done in a top-down fashion because the load of a child
6066 * group is a fraction of its parents load.
6068 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6070 struct rq *rq = rq_of(cfs_rq);
6071 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6072 unsigned long now = jiffies;
6075 if (cfs_rq->last_h_load_update == now)
6078 cfs_rq->h_load_next = NULL;
6079 for_each_sched_entity(se) {
6080 cfs_rq = cfs_rq_of(se);
6081 cfs_rq->h_load_next = se;
6082 if (cfs_rq->last_h_load_update == now)
6087 cfs_rq->h_load = cfs_rq->runnable_load_avg;
6088 cfs_rq->last_h_load_update = now;
6091 while ((se = cfs_rq->h_load_next) != NULL) {
6092 load = cfs_rq->h_load;
6093 load = div64_ul(load * se->avg.load_avg_contrib,
6094 cfs_rq->runnable_load_avg + 1);
6095 cfs_rq = group_cfs_rq(se);
6096 cfs_rq->h_load = load;
6097 cfs_rq->last_h_load_update = now;
6101 static unsigned long task_h_load(struct task_struct *p)
6103 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6105 update_cfs_rq_h_load(cfs_rq);
6106 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
6107 cfs_rq->runnable_load_avg + 1);
6110 static inline void update_blocked_averages(int cpu)
6114 static unsigned long task_h_load(struct task_struct *p)
6116 return p->se.avg.load_avg_contrib;
6120 /********** Helpers for find_busiest_group ************************/
6129 * sg_lb_stats - stats of a sched_group required for load_balancing
6131 struct sg_lb_stats {
6132 unsigned long avg_load; /*Avg load across the CPUs of the group */
6133 unsigned long group_load; /* Total load over the CPUs of the group */
6134 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6135 unsigned long load_per_task;
6136 unsigned long group_capacity;
6137 unsigned long group_usage; /* Total usage of the group */
6138 unsigned int sum_nr_running; /* Nr tasks running in the group */
6139 unsigned int idle_cpus;
6140 unsigned int group_weight;
6141 enum group_type group_type;
6142 int group_no_capacity;
6143 #ifdef CONFIG_NUMA_BALANCING
6144 unsigned int nr_numa_running;
6145 unsigned int nr_preferred_running;
6150 * sd_lb_stats - Structure to store the statistics of a sched_domain
6151 * during load balancing.
6153 struct sd_lb_stats {
6154 struct sched_group *busiest; /* Busiest group in this sd */
6155 struct sched_group *local; /* Local group in this sd */
6156 unsigned long total_load; /* Total load of all groups in sd */
6157 unsigned long total_capacity; /* Total capacity of all groups in sd */
6158 unsigned long avg_load; /* Average load across all groups in sd */
6160 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6161 struct sg_lb_stats local_stat; /* Statistics of the local group */
6164 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6167 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6168 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6169 * We must however clear busiest_stat::avg_load because
6170 * update_sd_pick_busiest() reads this before assignment.
6172 *sds = (struct sd_lb_stats){
6176 .total_capacity = 0UL,
6179 .sum_nr_running = 0,
6180 .group_type = group_other,
6186 * get_sd_load_idx - Obtain the load index for a given sched domain.
6187 * @sd: The sched_domain whose load_idx is to be obtained.
6188 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6190 * Return: The load index.
6192 static inline int get_sd_load_idx(struct sched_domain *sd,
6193 enum cpu_idle_type idle)
6199 load_idx = sd->busy_idx;
6202 case CPU_NEWLY_IDLE:
6203 load_idx = sd->newidle_idx;
6206 load_idx = sd->idle_idx;
6213 static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu)
6215 if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1))
6216 return sd->smt_gain / sd->span_weight;
6218 return SCHED_CAPACITY_SCALE;
6221 unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
6223 return default_scale_cpu_capacity(sd, cpu);
6226 static unsigned long scale_rt_capacity(int cpu)
6228 struct rq *rq = cpu_rq(cpu);
6229 u64 total, used, age_stamp, avg;
6233 * Since we're reading these variables without serialization make sure
6234 * we read them once before doing sanity checks on them.
6236 age_stamp = READ_ONCE(rq->age_stamp);
6237 avg = READ_ONCE(rq->rt_avg);
6238 delta = __rq_clock_broken(rq) - age_stamp;
6240 if (unlikely(delta < 0))
6243 total = sched_avg_period() + delta;
6245 used = div_u64(avg, total);
6247 if (likely(used < SCHED_CAPACITY_SCALE))
6248 return SCHED_CAPACITY_SCALE - used;
6253 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6255 unsigned long capacity = SCHED_CAPACITY_SCALE;
6256 struct sched_group *sdg = sd->groups;
6258 if (sched_feat(ARCH_CAPACITY))
6259 capacity *= arch_scale_cpu_capacity(sd, cpu);
6261 capacity *= default_scale_cpu_capacity(sd, cpu);
6263 capacity >>= SCHED_CAPACITY_SHIFT;
6265 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6267 capacity *= scale_rt_capacity(cpu);
6268 capacity >>= SCHED_CAPACITY_SHIFT;
6273 cpu_rq(cpu)->cpu_capacity = capacity;
6274 sdg->sgc->capacity = capacity;
6277 void update_group_capacity(struct sched_domain *sd, int cpu)
6279 struct sched_domain *child = sd->child;
6280 struct sched_group *group, *sdg = sd->groups;
6281 unsigned long capacity;
6282 unsigned long interval;
6284 interval = msecs_to_jiffies(sd->balance_interval);
6285 interval = clamp(interval, 1UL, max_load_balance_interval);
6286 sdg->sgc->next_update = jiffies + interval;
6289 update_cpu_capacity(sd, cpu);
6295 if (child->flags & SD_OVERLAP) {
6297 * SD_OVERLAP domains cannot assume that child groups
6298 * span the current group.
6301 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6302 struct sched_group_capacity *sgc;
6303 struct rq *rq = cpu_rq(cpu);
6306 * build_sched_domains() -> init_sched_groups_capacity()
6307 * gets here before we've attached the domains to the
6310 * Use capacity_of(), which is set irrespective of domains
6311 * in update_cpu_capacity().
6313 * This avoids capacity from being 0 and
6314 * causing divide-by-zero issues on boot.
6316 if (unlikely(!rq->sd)) {
6317 capacity += capacity_of(cpu);
6321 sgc = rq->sd->groups->sgc;
6322 capacity += sgc->capacity;
6326 * !SD_OVERLAP domains can assume that child groups
6327 * span the current group.
6330 group = child->groups;
6332 capacity += group->sgc->capacity;
6333 group = group->next;
6334 } while (group != child->groups);
6337 sdg->sgc->capacity = capacity;
6341 * Check whether the capacity of the rq has been noticeably reduced by side
6342 * activity. The imbalance_pct is used for the threshold.
6343 * Return true is the capacity is reduced
6346 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6348 return ((rq->cpu_capacity * sd->imbalance_pct) <
6349 (rq->cpu_capacity_orig * 100));
6353 * Group imbalance indicates (and tries to solve) the problem where balancing
6354 * groups is inadequate due to tsk_cpus_allowed() constraints.
6356 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6357 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6360 * { 0 1 2 3 } { 4 5 6 7 }
6363 * If we were to balance group-wise we'd place two tasks in the first group and
6364 * two tasks in the second group. Clearly this is undesired as it will overload
6365 * cpu 3 and leave one of the cpus in the second group unused.
6367 * The current solution to this issue is detecting the skew in the first group
6368 * by noticing the lower domain failed to reach balance and had difficulty
6369 * moving tasks due to affinity constraints.
6371 * When this is so detected; this group becomes a candidate for busiest; see
6372 * update_sd_pick_busiest(). And calculate_imbalance() and
6373 * find_busiest_group() avoid some of the usual balance conditions to allow it
6374 * to create an effective group imbalance.
6376 * This is a somewhat tricky proposition since the next run might not find the
6377 * group imbalance and decide the groups need to be balanced again. A most
6378 * subtle and fragile situation.
6381 static inline int sg_imbalanced(struct sched_group *group)
6383 return group->sgc->imbalance;
6387 * group_has_capacity returns true if the group has spare capacity that could
6388 * be used by some tasks.
6389 * We consider that a group has spare capacity if the * number of task is
6390 * smaller than the number of CPUs or if the usage is lower than the available
6391 * capacity for CFS tasks.
6392 * For the latter, we use a threshold to stabilize the state, to take into
6393 * account the variance of the tasks' load and to return true if the available
6394 * capacity in meaningful for the load balancer.
6395 * As an example, an available capacity of 1% can appear but it doesn't make
6396 * any benefit for the load balance.
6399 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6401 if (sgs->sum_nr_running < sgs->group_weight)
6404 if ((sgs->group_capacity * 100) >
6405 (sgs->group_usage * env->sd->imbalance_pct))
6412 * group_is_overloaded returns true if the group has more tasks than it can
6414 * group_is_overloaded is not equals to !group_has_capacity because a group
6415 * with the exact right number of tasks, has no more spare capacity but is not
6416 * overloaded so both group_has_capacity and group_is_overloaded return
6420 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6422 if (sgs->sum_nr_running <= sgs->group_weight)
6425 if ((sgs->group_capacity * 100) <
6426 (sgs->group_usage * env->sd->imbalance_pct))
6432 static enum group_type group_classify(struct lb_env *env,
6433 struct sched_group *group,
6434 struct sg_lb_stats *sgs)
6436 if (sgs->group_no_capacity)
6437 return group_overloaded;
6439 if (sg_imbalanced(group))
6440 return group_imbalanced;
6446 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6447 * @env: The load balancing environment.
6448 * @group: sched_group whose statistics are to be updated.
6449 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6450 * @local_group: Does group contain this_cpu.
6451 * @sgs: variable to hold the statistics for this group.
6452 * @overload: Indicate more than one runnable task for any CPU.
6454 static inline void update_sg_lb_stats(struct lb_env *env,
6455 struct sched_group *group, int load_idx,
6456 int local_group, struct sg_lb_stats *sgs,
6462 memset(sgs, 0, sizeof(*sgs));
6464 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6465 struct rq *rq = cpu_rq(i);
6467 /* Bias balancing toward cpus of our domain */
6469 load = target_load(i, load_idx);
6471 load = source_load(i, load_idx);
6473 sgs->group_load += load;
6474 sgs->group_usage += get_cpu_usage(i);
6475 sgs->sum_nr_running += rq->cfs.h_nr_running;
6477 if (rq->nr_running > 1)
6480 #ifdef CONFIG_NUMA_BALANCING
6481 sgs->nr_numa_running += rq->nr_numa_running;
6482 sgs->nr_preferred_running += rq->nr_preferred_running;
6484 sgs->sum_weighted_load += weighted_cpuload(i);
6489 /* Adjust by relative CPU capacity of the group */
6490 sgs->group_capacity = group->sgc->capacity;
6491 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6493 if (sgs->sum_nr_running)
6494 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6496 sgs->group_weight = group->group_weight;
6498 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6499 sgs->group_type = group_classify(env, group, sgs);
6503 * update_sd_pick_busiest - return 1 on busiest group
6504 * @env: The load balancing environment.
6505 * @sds: sched_domain statistics
6506 * @sg: sched_group candidate to be checked for being the busiest
6507 * @sgs: sched_group statistics
6509 * Determine if @sg is a busier group than the previously selected
6512 * Return: %true if @sg is a busier group than the previously selected
6513 * busiest group. %false otherwise.
6515 static bool update_sd_pick_busiest(struct lb_env *env,
6516 struct sd_lb_stats *sds,
6517 struct sched_group *sg,
6518 struct sg_lb_stats *sgs)
6520 struct sg_lb_stats *busiest = &sds->busiest_stat;
6522 if (sgs->group_type > busiest->group_type)
6525 if (sgs->group_type < busiest->group_type)
6528 if (sgs->avg_load <= busiest->avg_load)
6531 /* This is the busiest node in its class. */
6532 if (!(env->sd->flags & SD_ASYM_PACKING))
6536 * ASYM_PACKING needs to move all the work to the lowest
6537 * numbered CPUs in the group, therefore mark all groups
6538 * higher than ourself as busy.
6540 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6544 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6551 #ifdef CONFIG_NUMA_BALANCING
6552 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6554 if (sgs->sum_nr_running > sgs->nr_numa_running)
6556 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6561 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6563 if (rq->nr_running > rq->nr_numa_running)
6565 if (rq->nr_running > rq->nr_preferred_running)
6570 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6575 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6579 #endif /* CONFIG_NUMA_BALANCING */
6582 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6583 * @env: The load balancing environment.
6584 * @sds: variable to hold the statistics for this sched_domain.
6586 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6588 struct sched_domain *child = env->sd->child;
6589 struct sched_group *sg = env->sd->groups;
6590 struct sg_lb_stats tmp_sgs;
6591 int load_idx, prefer_sibling = 0;
6592 bool overload = false;
6594 if (child && child->flags & SD_PREFER_SIBLING)
6597 load_idx = get_sd_load_idx(env->sd, env->idle);
6600 struct sg_lb_stats *sgs = &tmp_sgs;
6603 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6606 sgs = &sds->local_stat;
6608 if (env->idle != CPU_NEWLY_IDLE ||
6609 time_after_eq(jiffies, sg->sgc->next_update))
6610 update_group_capacity(env->sd, env->dst_cpu);
6613 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6620 * In case the child domain prefers tasks go to siblings
6621 * first, lower the sg capacity so that we'll try
6622 * and move all the excess tasks away. We lower the capacity
6623 * of a group only if the local group has the capacity to fit
6624 * these excess tasks. The extra check prevents the case where
6625 * you always pull from the heaviest group when it is already
6626 * under-utilized (possible with a large weight task outweighs
6627 * the tasks on the system).
6629 if (prefer_sibling && sds->local &&
6630 group_has_capacity(env, &sds->local_stat) &&
6631 (sgs->sum_nr_running > 1)) {
6632 sgs->group_no_capacity = 1;
6633 sgs->group_type = group_overloaded;
6636 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6638 sds->busiest_stat = *sgs;
6642 /* Now, start updating sd_lb_stats */
6643 sds->total_load += sgs->group_load;
6644 sds->total_capacity += sgs->group_capacity;
6647 } while (sg != env->sd->groups);
6649 if (env->sd->flags & SD_NUMA)
6650 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6652 if (!env->sd->parent) {
6653 /* update overload indicator if we are at root domain */
6654 if (env->dst_rq->rd->overload != overload)
6655 env->dst_rq->rd->overload = overload;
6661 * check_asym_packing - Check to see if the group is packed into the
6664 * This is primarily intended to used at the sibling level. Some
6665 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6666 * case of POWER7, it can move to lower SMT modes only when higher
6667 * threads are idle. When in lower SMT modes, the threads will
6668 * perform better since they share less core resources. Hence when we
6669 * have idle threads, we want them to be the higher ones.
6671 * This packing function is run on idle threads. It checks to see if
6672 * the busiest CPU in this domain (core in the P7 case) has a higher
6673 * CPU number than the packing function is being run on. Here we are
6674 * assuming lower CPU number will be equivalent to lower a SMT thread
6677 * Return: 1 when packing is required and a task should be moved to
6678 * this CPU. The amount of the imbalance is returned in *imbalance.
6680 * @env: The load balancing environment.
6681 * @sds: Statistics of the sched_domain which is to be packed
6683 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6687 if (!(env->sd->flags & SD_ASYM_PACKING))
6693 busiest_cpu = group_first_cpu(sds->busiest);
6694 if (env->dst_cpu > busiest_cpu)
6697 env->imbalance = DIV_ROUND_CLOSEST(
6698 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6699 SCHED_CAPACITY_SCALE);
6705 * fix_small_imbalance - Calculate the minor imbalance that exists
6706 * amongst the groups of a sched_domain, during
6708 * @env: The load balancing environment.
6709 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6712 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6714 unsigned long tmp, capa_now = 0, capa_move = 0;
6715 unsigned int imbn = 2;
6716 unsigned long scaled_busy_load_per_task;
6717 struct sg_lb_stats *local, *busiest;
6719 local = &sds->local_stat;
6720 busiest = &sds->busiest_stat;
6722 if (!local->sum_nr_running)
6723 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6724 else if (busiest->load_per_task > local->load_per_task)
6727 scaled_busy_load_per_task =
6728 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6729 busiest->group_capacity;
6731 if (busiest->avg_load + scaled_busy_load_per_task >=
6732 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6733 env->imbalance = busiest->load_per_task;
6738 * OK, we don't have enough imbalance to justify moving tasks,
6739 * however we may be able to increase total CPU capacity used by
6743 capa_now += busiest->group_capacity *
6744 min(busiest->load_per_task, busiest->avg_load);
6745 capa_now += local->group_capacity *
6746 min(local->load_per_task, local->avg_load);
6747 capa_now /= SCHED_CAPACITY_SCALE;
6749 /* Amount of load we'd subtract */
6750 if (busiest->avg_load > scaled_busy_load_per_task) {
6751 capa_move += busiest->group_capacity *
6752 min(busiest->load_per_task,
6753 busiest->avg_load - scaled_busy_load_per_task);
6756 /* Amount of load we'd add */
6757 if (busiest->avg_load * busiest->group_capacity <
6758 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6759 tmp = (busiest->avg_load * busiest->group_capacity) /
6760 local->group_capacity;
6762 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6763 local->group_capacity;
6765 capa_move += local->group_capacity *
6766 min(local->load_per_task, local->avg_load + tmp);
6767 capa_move /= SCHED_CAPACITY_SCALE;
6769 /* Move if we gain throughput */
6770 if (capa_move > capa_now)
6771 env->imbalance = busiest->load_per_task;
6775 * calculate_imbalance - Calculate the amount of imbalance present within the
6776 * groups of a given sched_domain during load balance.
6777 * @env: load balance environment
6778 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6780 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6782 unsigned long max_pull, load_above_capacity = ~0UL;
6783 struct sg_lb_stats *local, *busiest;
6785 local = &sds->local_stat;
6786 busiest = &sds->busiest_stat;
6788 if (busiest->group_type == group_imbalanced) {
6790 * In the group_imb case we cannot rely on group-wide averages
6791 * to ensure cpu-load equilibrium, look at wider averages. XXX
6793 busiest->load_per_task =
6794 min(busiest->load_per_task, sds->avg_load);
6798 * In the presence of smp nice balancing, certain scenarios can have
6799 * max load less than avg load(as we skip the groups at or below
6800 * its cpu_capacity, while calculating max_load..)
6802 if (busiest->avg_load <= sds->avg_load ||
6803 local->avg_load >= sds->avg_load) {
6805 return fix_small_imbalance(env, sds);
6809 * If there aren't any idle cpus, avoid creating some.
6811 if (busiest->group_type == group_overloaded &&
6812 local->group_type == group_overloaded) {
6813 load_above_capacity = busiest->sum_nr_running *
6815 if (load_above_capacity > busiest->group_capacity)
6816 load_above_capacity -= busiest->group_capacity;
6818 load_above_capacity = ~0UL;
6822 * We're trying to get all the cpus to the average_load, so we don't
6823 * want to push ourselves above the average load, nor do we wish to
6824 * reduce the max loaded cpu below the average load. At the same time,
6825 * we also don't want to reduce the group load below the group capacity
6826 * (so that we can implement power-savings policies etc). Thus we look
6827 * for the minimum possible imbalance.
6829 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6831 /* How much load to actually move to equalise the imbalance */
6832 env->imbalance = min(
6833 max_pull * busiest->group_capacity,
6834 (sds->avg_load - local->avg_load) * local->group_capacity
6835 ) / SCHED_CAPACITY_SCALE;
6838 * if *imbalance is less than the average load per runnable task
6839 * there is no guarantee that any tasks will be moved so we'll have
6840 * a think about bumping its value to force at least one task to be
6843 if (env->imbalance < busiest->load_per_task)
6844 return fix_small_imbalance(env, sds);
6847 /******* find_busiest_group() helpers end here *********************/
6850 * find_busiest_group - Returns the busiest group within the sched_domain
6851 * if there is an imbalance. If there isn't an imbalance, and
6852 * the user has opted for power-savings, it returns a group whose
6853 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6854 * such a group exists.
6856 * Also calculates the amount of weighted load which should be moved
6857 * to restore balance.
6859 * @env: The load balancing environment.
6861 * Return: - The busiest group if imbalance exists.
6862 * - If no imbalance and user has opted for power-savings balance,
6863 * return the least loaded group whose CPUs can be
6864 * put to idle by rebalancing its tasks onto our group.
6866 static struct sched_group *find_busiest_group(struct lb_env *env)
6868 struct sg_lb_stats *local, *busiest;
6869 struct sd_lb_stats sds;
6871 init_sd_lb_stats(&sds);
6874 * Compute the various statistics relavent for load balancing at
6877 update_sd_lb_stats(env, &sds);
6878 local = &sds.local_stat;
6879 busiest = &sds.busiest_stat;
6881 /* ASYM feature bypasses nice load balance check */
6882 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6883 check_asym_packing(env, &sds))
6886 /* There is no busy sibling group to pull tasks from */
6887 if (!sds.busiest || busiest->sum_nr_running == 0)
6890 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6891 / sds.total_capacity;
6894 * If the busiest group is imbalanced the below checks don't
6895 * work because they assume all things are equal, which typically
6896 * isn't true due to cpus_allowed constraints and the like.
6898 if (busiest->group_type == group_imbalanced)
6901 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6902 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6903 busiest->group_no_capacity)
6907 * If the local group is busier than the selected busiest group
6908 * don't try and pull any tasks.
6910 if (local->avg_load >= busiest->avg_load)
6914 * Don't pull any tasks if this group is already above the domain
6917 if (local->avg_load >= sds.avg_load)
6920 if (env->idle == CPU_IDLE) {
6922 * This cpu is idle. If the busiest group is not overloaded
6923 * and there is no imbalance between this and busiest group
6924 * wrt idle cpus, it is balanced. The imbalance becomes
6925 * significant if the diff is greater than 1 otherwise we
6926 * might end up to just move the imbalance on another group
6928 if ((busiest->group_type != group_overloaded) &&
6929 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6933 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6934 * imbalance_pct to be conservative.
6936 if (100 * busiest->avg_load <=
6937 env->sd->imbalance_pct * local->avg_load)
6942 /* Looks like there is an imbalance. Compute it */
6943 calculate_imbalance(env, &sds);
6952 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6954 static struct rq *find_busiest_queue(struct lb_env *env,
6955 struct sched_group *group)
6957 struct rq *busiest = NULL, *rq;
6958 unsigned long busiest_load = 0, busiest_capacity = 1;
6961 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6962 unsigned long capacity, wl;
6966 rt = fbq_classify_rq(rq);
6969 * We classify groups/runqueues into three groups:
6970 * - regular: there are !numa tasks
6971 * - remote: there are numa tasks that run on the 'wrong' node
6972 * - all: there is no distinction
6974 * In order to avoid migrating ideally placed numa tasks,
6975 * ignore those when there's better options.
6977 * If we ignore the actual busiest queue to migrate another
6978 * task, the next balance pass can still reduce the busiest
6979 * queue by moving tasks around inside the node.
6981 * If we cannot move enough load due to this classification
6982 * the next pass will adjust the group classification and
6983 * allow migration of more tasks.
6985 * Both cases only affect the total convergence complexity.
6987 if (rt > env->fbq_type)
6990 capacity = capacity_of(i);
6992 wl = weighted_cpuload(i);
6995 * When comparing with imbalance, use weighted_cpuload()
6996 * which is not scaled with the cpu capacity.
6999 if (rq->nr_running == 1 && wl > env->imbalance &&
7000 !check_cpu_capacity(rq, env->sd))
7004 * For the load comparisons with the other cpu's, consider
7005 * the weighted_cpuload() scaled with the cpu capacity, so
7006 * that the load can be moved away from the cpu that is
7007 * potentially running at a lower capacity.
7009 * Thus we're looking for max(wl_i / capacity_i), crosswise
7010 * multiplication to rid ourselves of the division works out
7011 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7012 * our previous maximum.
7014 if (wl * busiest_capacity > busiest_load * capacity) {
7016 busiest_capacity = capacity;
7025 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7026 * so long as it is large enough.
7028 #define MAX_PINNED_INTERVAL 512
7030 /* Working cpumask for load_balance and load_balance_newidle. */
7031 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7033 static int need_active_balance(struct lb_env *env)
7035 struct sched_domain *sd = env->sd;
7037 if (env->idle == CPU_NEWLY_IDLE) {
7040 * ASYM_PACKING needs to force migrate tasks from busy but
7041 * higher numbered CPUs in order to pack all tasks in the
7042 * lowest numbered CPUs.
7044 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7049 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7050 * It's worth migrating the task if the src_cpu's capacity is reduced
7051 * because of other sched_class or IRQs if more capacity stays
7052 * available on dst_cpu.
7054 if ((env->idle != CPU_NOT_IDLE) &&
7055 (env->src_rq->cfs.h_nr_running == 1)) {
7056 if ((check_cpu_capacity(env->src_rq, sd)) &&
7057 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7061 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7064 static int active_load_balance_cpu_stop(void *data);
7066 static int should_we_balance(struct lb_env *env)
7068 struct sched_group *sg = env->sd->groups;
7069 struct cpumask *sg_cpus, *sg_mask;
7070 int cpu, balance_cpu = -1;
7073 * In the newly idle case, we will allow all the cpu's
7074 * to do the newly idle load balance.
7076 if (env->idle == CPU_NEWLY_IDLE)
7079 sg_cpus = sched_group_cpus(sg);
7080 sg_mask = sched_group_mask(sg);
7081 /* Try to find first idle cpu */
7082 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7083 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7090 if (balance_cpu == -1)
7091 balance_cpu = group_balance_cpu(sg);
7094 * First idle cpu or the first cpu(busiest) in this sched group
7095 * is eligible for doing load balancing at this and above domains.
7097 return balance_cpu == env->dst_cpu;
7101 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7102 * tasks if there is an imbalance.
7104 static int load_balance(int this_cpu, struct rq *this_rq,
7105 struct sched_domain *sd, enum cpu_idle_type idle,
7106 int *continue_balancing)
7108 int ld_moved, cur_ld_moved, active_balance = 0;
7109 struct sched_domain *sd_parent = sd->parent;
7110 struct sched_group *group;
7112 unsigned long flags;
7113 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7115 struct lb_env env = {
7117 .dst_cpu = this_cpu,
7119 .dst_grpmask = sched_group_cpus(sd->groups),
7121 .loop_break = sched_nr_migrate_break,
7124 .tasks = LIST_HEAD_INIT(env.tasks),
7128 * For NEWLY_IDLE load_balancing, we don't need to consider
7129 * other cpus in our group
7131 if (idle == CPU_NEWLY_IDLE)
7132 env.dst_grpmask = NULL;
7134 cpumask_copy(cpus, cpu_active_mask);
7136 schedstat_inc(sd, lb_count[idle]);
7139 if (!should_we_balance(&env)) {
7140 *continue_balancing = 0;
7144 group = find_busiest_group(&env);
7146 schedstat_inc(sd, lb_nobusyg[idle]);
7150 busiest = find_busiest_queue(&env, group);
7152 schedstat_inc(sd, lb_nobusyq[idle]);
7156 BUG_ON(busiest == env.dst_rq);
7158 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7160 env.src_cpu = busiest->cpu;
7161 env.src_rq = busiest;
7164 if (busiest->nr_running > 1) {
7166 * Attempt to move tasks. If find_busiest_group has found
7167 * an imbalance but busiest->nr_running <= 1, the group is
7168 * still unbalanced. ld_moved simply stays zero, so it is
7169 * correctly treated as an imbalance.
7171 env.flags |= LBF_ALL_PINNED;
7172 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7175 raw_spin_lock_irqsave(&busiest->lock, flags);
7178 * cur_ld_moved - load moved in current iteration
7179 * ld_moved - cumulative load moved across iterations
7181 cur_ld_moved = detach_tasks(&env);
7184 * We've detached some tasks from busiest_rq. Every
7185 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7186 * unlock busiest->lock, and we are able to be sure
7187 * that nobody can manipulate the tasks in parallel.
7188 * See task_rq_lock() family for the details.
7191 raw_spin_unlock(&busiest->lock);
7195 ld_moved += cur_ld_moved;
7198 local_irq_restore(flags);
7200 if (env.flags & LBF_NEED_BREAK) {
7201 env.flags &= ~LBF_NEED_BREAK;
7206 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7207 * us and move them to an alternate dst_cpu in our sched_group
7208 * where they can run. The upper limit on how many times we
7209 * iterate on same src_cpu is dependent on number of cpus in our
7212 * This changes load balance semantics a bit on who can move
7213 * load to a given_cpu. In addition to the given_cpu itself
7214 * (or a ilb_cpu acting on its behalf where given_cpu is
7215 * nohz-idle), we now have balance_cpu in a position to move
7216 * load to given_cpu. In rare situations, this may cause
7217 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7218 * _independently_ and at _same_ time to move some load to
7219 * given_cpu) causing exceess load to be moved to given_cpu.
7220 * This however should not happen so much in practice and
7221 * moreover subsequent load balance cycles should correct the
7222 * excess load moved.
7224 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7226 /* Prevent to re-select dst_cpu via env's cpus */
7227 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7229 env.dst_rq = cpu_rq(env.new_dst_cpu);
7230 env.dst_cpu = env.new_dst_cpu;
7231 env.flags &= ~LBF_DST_PINNED;
7233 env.loop_break = sched_nr_migrate_break;
7236 * Go back to "more_balance" rather than "redo" since we
7237 * need to continue with same src_cpu.
7243 * We failed to reach balance because of affinity.
7246 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7248 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7249 *group_imbalance = 1;
7252 /* All tasks on this runqueue were pinned by CPU affinity */
7253 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7254 cpumask_clear_cpu(cpu_of(busiest), cpus);
7255 if (!cpumask_empty(cpus)) {
7257 env.loop_break = sched_nr_migrate_break;
7260 goto out_all_pinned;
7265 schedstat_inc(sd, lb_failed[idle]);
7267 * Increment the failure counter only on periodic balance.
7268 * We do not want newidle balance, which can be very
7269 * frequent, pollute the failure counter causing
7270 * excessive cache_hot migrations and active balances.
7272 if (idle != CPU_NEWLY_IDLE)
7273 sd->nr_balance_failed++;
7275 if (need_active_balance(&env)) {
7276 raw_spin_lock_irqsave(&busiest->lock, flags);
7278 /* don't kick the active_load_balance_cpu_stop,
7279 * if the curr task on busiest cpu can't be
7282 if (!cpumask_test_cpu(this_cpu,
7283 tsk_cpus_allowed(busiest->curr))) {
7284 raw_spin_unlock_irqrestore(&busiest->lock,
7286 env.flags |= LBF_ALL_PINNED;
7287 goto out_one_pinned;
7291 * ->active_balance synchronizes accesses to
7292 * ->active_balance_work. Once set, it's cleared
7293 * only after active load balance is finished.
7295 if (!busiest->active_balance) {
7296 busiest->active_balance = 1;
7297 busiest->push_cpu = this_cpu;
7300 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7302 if (active_balance) {
7303 stop_one_cpu_nowait(cpu_of(busiest),
7304 active_load_balance_cpu_stop, busiest,
7305 &busiest->active_balance_work);
7309 * We've kicked active balancing, reset the failure
7312 sd->nr_balance_failed = sd->cache_nice_tries+1;
7315 sd->nr_balance_failed = 0;
7317 if (likely(!active_balance)) {
7318 /* We were unbalanced, so reset the balancing interval */
7319 sd->balance_interval = sd->min_interval;
7322 * If we've begun active balancing, start to back off. This
7323 * case may not be covered by the all_pinned logic if there
7324 * is only 1 task on the busy runqueue (because we don't call
7327 if (sd->balance_interval < sd->max_interval)
7328 sd->balance_interval *= 2;
7335 * We reach balance although we may have faced some affinity
7336 * constraints. Clear the imbalance flag if it was set.
7339 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7341 if (*group_imbalance)
7342 *group_imbalance = 0;
7347 * We reach balance because all tasks are pinned at this level so
7348 * we can't migrate them. Let the imbalance flag set so parent level
7349 * can try to migrate them.
7351 schedstat_inc(sd, lb_balanced[idle]);
7353 sd->nr_balance_failed = 0;
7356 /* tune up the balancing interval */
7357 if (((env.flags & LBF_ALL_PINNED) &&
7358 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7359 (sd->balance_interval < sd->max_interval))
7360 sd->balance_interval *= 2;
7367 static inline unsigned long
7368 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7370 unsigned long interval = sd->balance_interval;
7373 interval *= sd->busy_factor;
7375 /* scale ms to jiffies */
7376 interval = msecs_to_jiffies(interval);
7377 interval = clamp(interval, 1UL, max_load_balance_interval);
7383 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7385 unsigned long interval, next;
7387 interval = get_sd_balance_interval(sd, cpu_busy);
7388 next = sd->last_balance + interval;
7390 if (time_after(*next_balance, next))
7391 *next_balance = next;
7395 * idle_balance is called by schedule() if this_cpu is about to become
7396 * idle. Attempts to pull tasks from other CPUs.
7398 static int idle_balance(struct rq *this_rq)
7400 unsigned long next_balance = jiffies + HZ;
7401 int this_cpu = this_rq->cpu;
7402 struct sched_domain *sd;
7403 int pulled_task = 0;
7406 idle_enter_fair(this_rq);
7409 * We must set idle_stamp _before_ calling idle_balance(), such that we
7410 * measure the duration of idle_balance() as idle time.
7412 this_rq->idle_stamp = rq_clock(this_rq);
7414 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7415 !this_rq->rd->overload) {
7417 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7419 update_next_balance(sd, 0, &next_balance);
7426 * Drop the rq->lock, but keep IRQ/preempt disabled.
7428 raw_spin_unlock(&this_rq->lock);
7430 update_blocked_averages(this_cpu);
7432 for_each_domain(this_cpu, sd) {
7433 int continue_balancing = 1;
7434 u64 t0, domain_cost;
7436 if (!(sd->flags & SD_LOAD_BALANCE))
7439 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7440 update_next_balance(sd, 0, &next_balance);
7444 if (sd->flags & SD_BALANCE_NEWIDLE) {
7445 t0 = sched_clock_cpu(this_cpu);
7447 pulled_task = load_balance(this_cpu, this_rq,
7449 &continue_balancing);
7451 domain_cost = sched_clock_cpu(this_cpu) - t0;
7452 if (domain_cost > sd->max_newidle_lb_cost)
7453 sd->max_newidle_lb_cost = domain_cost;
7455 curr_cost += domain_cost;
7458 update_next_balance(sd, 0, &next_balance);
7461 * Stop searching for tasks to pull if there are
7462 * now runnable tasks on this rq.
7464 if (pulled_task || this_rq->nr_running > 0)
7469 raw_spin_lock(&this_rq->lock);
7471 if (curr_cost > this_rq->max_idle_balance_cost)
7472 this_rq->max_idle_balance_cost = curr_cost;
7475 * While browsing the domains, we released the rq lock, a task could
7476 * have been enqueued in the meantime. Since we're not going idle,
7477 * pretend we pulled a task.
7479 if (this_rq->cfs.h_nr_running && !pulled_task)
7483 /* Move the next balance forward */
7484 if (time_after(this_rq->next_balance, next_balance))
7485 this_rq->next_balance = next_balance;
7487 /* Is there a task of a high priority class? */
7488 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7492 idle_exit_fair(this_rq);
7493 this_rq->idle_stamp = 0;
7500 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7501 * running tasks off the busiest CPU onto idle CPUs. It requires at
7502 * least 1 task to be running on each physical CPU where possible, and
7503 * avoids physical / logical imbalances.
7505 static int active_load_balance_cpu_stop(void *data)
7507 struct rq *busiest_rq = data;
7508 int busiest_cpu = cpu_of(busiest_rq);
7509 int target_cpu = busiest_rq->push_cpu;
7510 struct rq *target_rq = cpu_rq(target_cpu);
7511 struct sched_domain *sd;
7512 struct task_struct *p = NULL;
7514 raw_spin_lock_irq(&busiest_rq->lock);
7516 /* make sure the requested cpu hasn't gone down in the meantime */
7517 if (unlikely(busiest_cpu != smp_processor_id() ||
7518 !busiest_rq->active_balance))
7521 /* Is there any task to move? */
7522 if (busiest_rq->nr_running <= 1)
7526 * This condition is "impossible", if it occurs
7527 * we need to fix it. Originally reported by
7528 * Bjorn Helgaas on a 128-cpu setup.
7530 BUG_ON(busiest_rq == target_rq);
7532 /* Search for an sd spanning us and the target CPU. */
7534 for_each_domain(target_cpu, sd) {
7535 if ((sd->flags & SD_LOAD_BALANCE) &&
7536 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7541 struct lb_env env = {
7543 .dst_cpu = target_cpu,
7544 .dst_rq = target_rq,
7545 .src_cpu = busiest_rq->cpu,
7546 .src_rq = busiest_rq,
7550 schedstat_inc(sd, alb_count);
7552 p = detach_one_task(&env);
7554 schedstat_inc(sd, alb_pushed);
7556 schedstat_inc(sd, alb_failed);
7560 busiest_rq->active_balance = 0;
7561 raw_spin_unlock(&busiest_rq->lock);
7564 attach_one_task(target_rq, p);
7571 static inline int on_null_domain(struct rq *rq)
7573 return unlikely(!rcu_dereference_sched(rq->sd));
7576 #ifdef CONFIG_NO_HZ_COMMON
7578 * idle load balancing details
7579 * - When one of the busy CPUs notice that there may be an idle rebalancing
7580 * needed, they will kick the idle load balancer, which then does idle
7581 * load balancing for all the idle CPUs.
7584 cpumask_var_t idle_cpus_mask;
7586 unsigned long next_balance; /* in jiffy units */
7587 } nohz ____cacheline_aligned;
7589 static inline int find_new_ilb(void)
7591 int ilb = cpumask_first(nohz.idle_cpus_mask);
7593 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7600 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7601 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7602 * CPU (if there is one).
7604 static void nohz_balancer_kick(void)
7608 nohz.next_balance++;
7610 ilb_cpu = find_new_ilb();
7612 if (ilb_cpu >= nr_cpu_ids)
7615 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7618 * Use smp_send_reschedule() instead of resched_cpu().
7619 * This way we generate a sched IPI on the target cpu which
7620 * is idle. And the softirq performing nohz idle load balance
7621 * will be run before returning from the IPI.
7623 smp_send_reschedule(ilb_cpu);
7627 static inline void nohz_balance_exit_idle(int cpu)
7629 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7631 * Completely isolated CPUs don't ever set, so we must test.
7633 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7634 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7635 atomic_dec(&nohz.nr_cpus);
7637 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7641 static inline void set_cpu_sd_state_busy(void)
7643 struct sched_domain *sd;
7644 int cpu = smp_processor_id();
7647 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7649 if (!sd || !sd->nohz_idle)
7653 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7658 void set_cpu_sd_state_idle(void)
7660 struct sched_domain *sd;
7661 int cpu = smp_processor_id();
7664 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7666 if (!sd || sd->nohz_idle)
7670 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7676 * This routine will record that the cpu is going idle with tick stopped.
7677 * This info will be used in performing idle load balancing in the future.
7679 void nohz_balance_enter_idle(int cpu)
7682 * If this cpu is going down, then nothing needs to be done.
7684 if (!cpu_active(cpu))
7687 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7691 * If we're a completely isolated CPU, we don't play.
7693 if (on_null_domain(cpu_rq(cpu)))
7696 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7697 atomic_inc(&nohz.nr_cpus);
7698 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7701 static int sched_ilb_notifier(struct notifier_block *nfb,
7702 unsigned long action, void *hcpu)
7704 switch (action & ~CPU_TASKS_FROZEN) {
7706 nohz_balance_exit_idle(smp_processor_id());
7714 static DEFINE_SPINLOCK(balancing);
7717 * Scale the max load_balance interval with the number of CPUs in the system.
7718 * This trades load-balance latency on larger machines for less cross talk.
7720 void update_max_interval(void)
7722 max_load_balance_interval = HZ*num_online_cpus()/10;
7726 * It checks each scheduling domain to see if it is due to be balanced,
7727 * and initiates a balancing operation if so.
7729 * Balancing parameters are set up in init_sched_domains.
7731 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7733 int continue_balancing = 1;
7735 unsigned long interval;
7736 struct sched_domain *sd;
7737 /* Earliest time when we have to do rebalance again */
7738 unsigned long next_balance = jiffies + 60*HZ;
7739 int update_next_balance = 0;
7740 int need_serialize, need_decay = 0;
7743 update_blocked_averages(cpu);
7746 for_each_domain(cpu, sd) {
7748 * Decay the newidle max times here because this is a regular
7749 * visit to all the domains. Decay ~1% per second.
7751 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7752 sd->max_newidle_lb_cost =
7753 (sd->max_newidle_lb_cost * 253) / 256;
7754 sd->next_decay_max_lb_cost = jiffies + HZ;
7757 max_cost += sd->max_newidle_lb_cost;
7759 if (!(sd->flags & SD_LOAD_BALANCE))
7763 * Stop the load balance at this level. There is another
7764 * CPU in our sched group which is doing load balancing more
7767 if (!continue_balancing) {
7773 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7775 need_serialize = sd->flags & SD_SERIALIZE;
7776 if (need_serialize) {
7777 if (!spin_trylock(&balancing))
7781 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7782 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7784 * The LBF_DST_PINNED logic could have changed
7785 * env->dst_cpu, so we can't know our idle
7786 * state even if we migrated tasks. Update it.
7788 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7790 sd->last_balance = jiffies;
7791 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7794 spin_unlock(&balancing);
7796 if (time_after(next_balance, sd->last_balance + interval)) {
7797 next_balance = sd->last_balance + interval;
7798 update_next_balance = 1;
7803 * Ensure the rq-wide value also decays but keep it at a
7804 * reasonable floor to avoid funnies with rq->avg_idle.
7806 rq->max_idle_balance_cost =
7807 max((u64)sysctl_sched_migration_cost, max_cost);
7812 * next_balance will be updated only when there is a need.
7813 * When the cpu is attached to null domain for ex, it will not be
7816 if (likely(update_next_balance))
7817 rq->next_balance = next_balance;
7820 #ifdef CONFIG_NO_HZ_COMMON
7822 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7823 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7825 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7827 int this_cpu = this_rq->cpu;
7831 if (idle != CPU_IDLE ||
7832 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7835 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7836 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7840 * If this cpu gets work to do, stop the load balancing
7841 * work being done for other cpus. Next load
7842 * balancing owner will pick it up.
7847 rq = cpu_rq(balance_cpu);
7850 * If time for next balance is due,
7853 if (time_after_eq(jiffies, rq->next_balance)) {
7854 raw_spin_lock_irq(&rq->lock);
7855 update_rq_clock(rq);
7856 update_idle_cpu_load(rq);
7857 raw_spin_unlock_irq(&rq->lock);
7858 rebalance_domains(rq, CPU_IDLE);
7861 if (time_after(this_rq->next_balance, rq->next_balance))
7862 this_rq->next_balance = rq->next_balance;
7864 nohz.next_balance = this_rq->next_balance;
7866 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7870 * Current heuristic for kicking the idle load balancer in the presence
7871 * of an idle cpu in the system.
7872 * - This rq has more than one task.
7873 * - This rq has at least one CFS task and the capacity of the CPU is
7874 * significantly reduced because of RT tasks or IRQs.
7875 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7876 * multiple busy cpu.
7877 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7878 * domain span are idle.
7880 static inline bool nohz_kick_needed(struct rq *rq)
7882 unsigned long now = jiffies;
7883 struct sched_domain *sd;
7884 struct sched_group_capacity *sgc;
7885 int nr_busy, cpu = rq->cpu;
7888 if (unlikely(rq->idle_balance))
7892 * We may be recently in ticked or tickless idle mode. At the first
7893 * busy tick after returning from idle, we will update the busy stats.
7895 set_cpu_sd_state_busy();
7896 nohz_balance_exit_idle(cpu);
7899 * None are in tickless mode and hence no need for NOHZ idle load
7902 if (likely(!atomic_read(&nohz.nr_cpus)))
7905 if (time_before(now, nohz.next_balance))
7908 if (rq->nr_running >= 2)
7912 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7914 sgc = sd->groups->sgc;
7915 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7924 sd = rcu_dereference(rq->sd);
7926 if ((rq->cfs.h_nr_running >= 1) &&
7927 check_cpu_capacity(rq, sd)) {
7933 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7934 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7935 sched_domain_span(sd)) < cpu)) {
7945 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7949 * run_rebalance_domains is triggered when needed from the scheduler tick.
7950 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7952 static void run_rebalance_domains(struct softirq_action *h)
7954 struct rq *this_rq = this_rq();
7955 enum cpu_idle_type idle = this_rq->idle_balance ?
7956 CPU_IDLE : CPU_NOT_IDLE;
7959 * If this cpu has a pending nohz_balance_kick, then do the
7960 * balancing on behalf of the other idle cpus whose ticks are
7961 * stopped. Do nohz_idle_balance *before* rebalance_domains to
7962 * give the idle cpus a chance to load balance. Else we may
7963 * load balance only within the local sched_domain hierarchy
7964 * and abort nohz_idle_balance altogether if we pull some load.
7966 nohz_idle_balance(this_rq, idle);
7967 rebalance_domains(this_rq, idle);
7971 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7973 void trigger_load_balance(struct rq *rq)
7975 /* Don't need to rebalance while attached to NULL domain */
7976 if (unlikely(on_null_domain(rq)))
7979 if (time_after_eq(jiffies, rq->next_balance))
7980 raise_softirq(SCHED_SOFTIRQ);
7981 #ifdef CONFIG_NO_HZ_COMMON
7982 if (nohz_kick_needed(rq))
7983 nohz_balancer_kick();
7987 static void rq_online_fair(struct rq *rq)
7991 update_runtime_enabled(rq);
7994 static void rq_offline_fair(struct rq *rq)
7998 /* Ensure any throttled groups are reachable by pick_next_task */
7999 unthrottle_offline_cfs_rqs(rq);
8002 #endif /* CONFIG_SMP */
8005 * scheduler tick hitting a task of our scheduling class:
8007 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8009 struct cfs_rq *cfs_rq;
8010 struct sched_entity *se = &curr->se;
8012 for_each_sched_entity(se) {
8013 cfs_rq = cfs_rq_of(se);
8014 entity_tick(cfs_rq, se, queued);
8017 if (numabalancing_enabled)
8018 task_tick_numa(rq, curr);
8020 update_rq_runnable_avg(rq, 1);
8024 * called on fork with the child task as argument from the parent's context
8025 * - child not yet on the tasklist
8026 * - preemption disabled
8028 static void task_fork_fair(struct task_struct *p)
8030 struct cfs_rq *cfs_rq;
8031 struct sched_entity *se = &p->se, *curr;
8032 int this_cpu = smp_processor_id();
8033 struct rq *rq = this_rq();
8034 unsigned long flags;
8036 raw_spin_lock_irqsave(&rq->lock, flags);
8038 update_rq_clock(rq);
8040 cfs_rq = task_cfs_rq(current);
8041 curr = cfs_rq->curr;
8044 * Not only the cpu but also the task_group of the parent might have
8045 * been changed after parent->se.parent,cfs_rq were copied to
8046 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8047 * of child point to valid ones.
8050 __set_task_cpu(p, this_cpu);
8053 update_curr(cfs_rq);
8056 se->vruntime = curr->vruntime;
8057 place_entity(cfs_rq, se, 1);
8059 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8061 * Upon rescheduling, sched_class::put_prev_task() will place
8062 * 'current' within the tree based on its new key value.
8064 swap(curr->vruntime, se->vruntime);
8068 se->vruntime -= cfs_rq->min_vruntime;
8070 raw_spin_unlock_irqrestore(&rq->lock, flags);
8074 * Priority of the task has changed. Check to see if we preempt
8078 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8080 if (!task_on_rq_queued(p))
8084 * Reschedule if we are currently running on this runqueue and
8085 * our priority decreased, or if we are not currently running on
8086 * this runqueue and our priority is higher than the current's
8088 if (rq->curr == p) {
8089 if (p->prio > oldprio)
8092 check_preempt_curr(rq, p, 0);
8095 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8097 struct sched_entity *se = &p->se;
8098 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8101 * Ensure the task's vruntime is normalized, so that when it's
8102 * switched back to the fair class the enqueue_entity(.flags=0) will
8103 * do the right thing.
8105 * If it's queued, then the dequeue_entity(.flags=0) will already
8106 * have normalized the vruntime, if it's !queued, then only when
8107 * the task is sleeping will it still have non-normalized vruntime.
8109 if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) {
8111 * Fix up our vruntime so that the current sleep doesn't
8112 * cause 'unlimited' sleep bonus.
8114 place_entity(cfs_rq, se, 0);
8115 se->vruntime -= cfs_rq->min_vruntime;
8120 * Remove our load from contribution when we leave sched_fair
8121 * and ensure we don't carry in an old decay_count if we
8124 if (se->avg.decay_count) {
8125 __synchronize_entity_decay(se);
8126 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
8132 * We switched to the sched_fair class.
8134 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8136 #ifdef CONFIG_FAIR_GROUP_SCHED
8137 struct sched_entity *se = &p->se;
8139 * Since the real-depth could have been changed (only FAIR
8140 * class maintain depth value), reset depth properly.
8142 se->depth = se->parent ? se->parent->depth + 1 : 0;
8144 if (!task_on_rq_queued(p))
8148 * We were most likely switched from sched_rt, so
8149 * kick off the schedule if running, otherwise just see
8150 * if we can still preempt the current task.
8155 check_preempt_curr(rq, p, 0);
8158 /* Account for a task changing its policy or group.
8160 * This routine is mostly called to set cfs_rq->curr field when a task
8161 * migrates between groups/classes.
8163 static void set_curr_task_fair(struct rq *rq)
8165 struct sched_entity *se = &rq->curr->se;
8167 for_each_sched_entity(se) {
8168 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8170 set_next_entity(cfs_rq, se);
8171 /* ensure bandwidth has been allocated on our new cfs_rq */
8172 account_cfs_rq_runtime(cfs_rq, 0);
8176 void init_cfs_rq(struct cfs_rq *cfs_rq)
8178 cfs_rq->tasks_timeline = RB_ROOT;
8179 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8180 #ifndef CONFIG_64BIT
8181 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8184 atomic64_set(&cfs_rq->decay_counter, 1);
8185 atomic_long_set(&cfs_rq->removed_load, 0);
8189 #ifdef CONFIG_FAIR_GROUP_SCHED
8190 static void task_move_group_fair(struct task_struct *p, int queued)
8192 struct sched_entity *se = &p->se;
8193 struct cfs_rq *cfs_rq;
8196 * If the task was not on the rq at the time of this cgroup movement
8197 * it must have been asleep, sleeping tasks keep their ->vruntime
8198 * absolute on their old rq until wakeup (needed for the fair sleeper
8199 * bonus in place_entity()).
8201 * If it was on the rq, we've just 'preempted' it, which does convert
8202 * ->vruntime to a relative base.
8204 * Make sure both cases convert their relative position when migrating
8205 * to another cgroup's rq. This does somewhat interfere with the
8206 * fair sleeper stuff for the first placement, but who cares.
8209 * When !queued, vruntime of the task has usually NOT been normalized.
8210 * But there are some cases where it has already been normalized:
8212 * - Moving a forked child which is waiting for being woken up by
8213 * wake_up_new_task().
8214 * - Moving a task which has been woken up by try_to_wake_up() and
8215 * waiting for actually being woken up by sched_ttwu_pending().
8217 * To prevent boost or penalty in the new cfs_rq caused by delta
8218 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
8220 if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING))
8224 se->vruntime -= cfs_rq_of(se)->min_vruntime;
8225 set_task_rq(p, task_cpu(p));
8226 se->depth = se->parent ? se->parent->depth + 1 : 0;
8228 cfs_rq = cfs_rq_of(se);
8229 se->vruntime += cfs_rq->min_vruntime;
8232 * migrate_task_rq_fair() will have removed our previous
8233 * contribution, but we must synchronize for ongoing future
8236 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
8237 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
8242 void free_fair_sched_group(struct task_group *tg)
8246 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8248 for_each_possible_cpu(i) {
8250 kfree(tg->cfs_rq[i]);
8259 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8261 struct cfs_rq *cfs_rq;
8262 struct sched_entity *se;
8265 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8268 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8272 tg->shares = NICE_0_LOAD;
8274 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8276 for_each_possible_cpu(i) {
8277 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8278 GFP_KERNEL, cpu_to_node(i));
8282 se = kzalloc_node(sizeof(struct sched_entity),
8283 GFP_KERNEL, cpu_to_node(i));
8287 init_cfs_rq(cfs_rq);
8288 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8299 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8301 struct rq *rq = cpu_rq(cpu);
8302 unsigned long flags;
8305 * Only empty task groups can be destroyed; so we can speculatively
8306 * check on_list without danger of it being re-added.
8308 if (!tg->cfs_rq[cpu]->on_list)
8311 raw_spin_lock_irqsave(&rq->lock, flags);
8312 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8313 raw_spin_unlock_irqrestore(&rq->lock, flags);
8316 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8317 struct sched_entity *se, int cpu,
8318 struct sched_entity *parent)
8320 struct rq *rq = cpu_rq(cpu);
8324 init_cfs_rq_runtime(cfs_rq);
8326 tg->cfs_rq[cpu] = cfs_rq;
8329 /* se could be NULL for root_task_group */
8334 se->cfs_rq = &rq->cfs;
8337 se->cfs_rq = parent->my_q;
8338 se->depth = parent->depth + 1;
8342 /* guarantee group entities always have weight */
8343 update_load_set(&se->load, NICE_0_LOAD);
8344 se->parent = parent;
8347 static DEFINE_MUTEX(shares_mutex);
8349 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8352 unsigned long flags;
8355 * We can't change the weight of the root cgroup.
8360 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8362 mutex_lock(&shares_mutex);
8363 if (tg->shares == shares)
8366 tg->shares = shares;
8367 for_each_possible_cpu(i) {
8368 struct rq *rq = cpu_rq(i);
8369 struct sched_entity *se;
8372 /* Propagate contribution to hierarchy */
8373 raw_spin_lock_irqsave(&rq->lock, flags);
8375 /* Possible calls to update_curr() need rq clock */
8376 update_rq_clock(rq);
8377 for_each_sched_entity(se)
8378 update_cfs_shares(group_cfs_rq(se));
8379 raw_spin_unlock_irqrestore(&rq->lock, flags);
8383 mutex_unlock(&shares_mutex);
8386 #else /* CONFIG_FAIR_GROUP_SCHED */
8388 void free_fair_sched_group(struct task_group *tg) { }
8390 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8395 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8397 #endif /* CONFIG_FAIR_GROUP_SCHED */
8400 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8402 struct sched_entity *se = &task->se;
8403 unsigned int rr_interval = 0;
8406 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8409 if (rq->cfs.load.weight)
8410 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8416 * All the scheduling class methods:
8418 const struct sched_class fair_sched_class = {
8419 .next = &idle_sched_class,
8420 .enqueue_task = enqueue_task_fair,
8421 .dequeue_task = dequeue_task_fair,
8422 .yield_task = yield_task_fair,
8423 .yield_to_task = yield_to_task_fair,
8425 .check_preempt_curr = check_preempt_wakeup,
8427 .pick_next_task = pick_next_task_fair,
8428 .put_prev_task = put_prev_task_fair,
8431 .select_task_rq = select_task_rq_fair,
8432 .migrate_task_rq = migrate_task_rq_fair,
8434 .rq_online = rq_online_fair,
8435 .rq_offline = rq_offline_fair,
8437 .task_waking = task_waking_fair,
8440 .set_curr_task = set_curr_task_fair,
8441 .task_tick = task_tick_fair,
8442 .task_fork = task_fork_fair,
8444 .prio_changed = prio_changed_fair,
8445 .switched_from = switched_from_fair,
8446 .switched_to = switched_to_fair,
8448 .get_rr_interval = get_rr_interval_fair,
8450 .update_curr = update_curr_fair,
8452 #ifdef CONFIG_FAIR_GROUP_SCHED
8453 .task_move_group = task_move_group_fair,
8457 #ifdef CONFIG_SCHED_DEBUG
8458 void print_cfs_stats(struct seq_file *m, int cpu)
8460 struct cfs_rq *cfs_rq;
8463 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8464 print_cfs_rq(m, cpu, cfs_rq);
8469 __init void init_sched_fair_class(void)
8472 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8474 #ifdef CONFIG_NO_HZ_COMMON
8475 nohz.next_balance = jiffies;
8476 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8477 cpu_notifier(sched_ilb_notifier, 0);