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 int get_update_sysctl_factor(void)
146 unsigned int cpus = min_t(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 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 = ACCESS_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. Scan slower
1627 if (local + shared == 0) {
1628 p->numa_scan_period = min(p->numa_scan_period_max,
1629 p->numa_scan_period << 1);
1631 p->mm->numa_next_scan = jiffies +
1632 msecs_to_jiffies(p->numa_scan_period);
1638 * Prepare to scale scan period relative to the current period.
1639 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1640 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1641 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1643 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1644 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1645 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1646 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1649 diff = slot * period_slot;
1651 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1654 * Scale scan rate increases based on sharing. There is an
1655 * inverse relationship between the degree of sharing and
1656 * the adjustment made to the scanning period. Broadly
1657 * speaking the intent is that there is little point
1658 * scanning faster if shared accesses dominate as it may
1659 * simply bounce migrations uselessly
1661 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1662 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1665 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1666 task_scan_min(p), task_scan_max(p));
1667 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1671 * Get the fraction of time the task has been running since the last
1672 * NUMA placement cycle. The scheduler keeps similar statistics, but
1673 * decays those on a 32ms period, which is orders of magnitude off
1674 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1675 * stats only if the task is so new there are no NUMA statistics yet.
1677 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1679 u64 runtime, delta, now;
1680 /* Use the start of this time slice to avoid calculations. */
1681 now = p->se.exec_start;
1682 runtime = p->se.sum_exec_runtime;
1684 if (p->last_task_numa_placement) {
1685 delta = runtime - p->last_sum_exec_runtime;
1686 *period = now - p->last_task_numa_placement;
1688 delta = p->se.avg.runnable_avg_sum;
1689 *period = p->se.avg.avg_period;
1692 p->last_sum_exec_runtime = runtime;
1693 p->last_task_numa_placement = now;
1699 * Determine the preferred nid for a task in a numa_group. This needs to
1700 * be done in a way that produces consistent results with group_weight,
1701 * otherwise workloads might not converge.
1703 static int preferred_group_nid(struct task_struct *p, int nid)
1708 /* Direct connections between all NUMA nodes. */
1709 if (sched_numa_topology_type == NUMA_DIRECT)
1713 * On a system with glueless mesh NUMA topology, group_weight
1714 * scores nodes according to the number of NUMA hinting faults on
1715 * both the node itself, and on nearby nodes.
1717 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1718 unsigned long score, max_score = 0;
1719 int node, max_node = nid;
1721 dist = sched_max_numa_distance;
1723 for_each_online_node(node) {
1724 score = group_weight(p, node, dist);
1725 if (score > max_score) {
1734 * Finding the preferred nid in a system with NUMA backplane
1735 * interconnect topology is more involved. The goal is to locate
1736 * tasks from numa_groups near each other in the system, and
1737 * untangle workloads from different sides of the system. This requires
1738 * searching down the hierarchy of node groups, recursively searching
1739 * inside the highest scoring group of nodes. The nodemask tricks
1740 * keep the complexity of the search down.
1742 nodes = node_online_map;
1743 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1744 unsigned long max_faults = 0;
1745 nodemask_t max_group = NODE_MASK_NONE;
1748 /* Are there nodes at this distance from each other? */
1749 if (!find_numa_distance(dist))
1752 for_each_node_mask(a, nodes) {
1753 unsigned long faults = 0;
1754 nodemask_t this_group;
1755 nodes_clear(this_group);
1757 /* Sum group's NUMA faults; includes a==b case. */
1758 for_each_node_mask(b, nodes) {
1759 if (node_distance(a, b) < dist) {
1760 faults += group_faults(p, b);
1761 node_set(b, this_group);
1762 node_clear(b, nodes);
1766 /* Remember the top group. */
1767 if (faults > max_faults) {
1768 max_faults = faults;
1769 max_group = this_group;
1771 * subtle: at the smallest distance there is
1772 * just one node left in each "group", the
1773 * winner is the preferred nid.
1778 /* Next round, evaluate the nodes within max_group. */
1786 static void task_numa_placement(struct task_struct *p)
1788 int seq, nid, max_nid = -1, max_group_nid = -1;
1789 unsigned long max_faults = 0, max_group_faults = 0;
1790 unsigned long fault_types[2] = { 0, 0 };
1791 unsigned long total_faults;
1792 u64 runtime, period;
1793 spinlock_t *group_lock = NULL;
1795 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1796 if (p->numa_scan_seq == seq)
1798 p->numa_scan_seq = seq;
1799 p->numa_scan_period_max = task_scan_max(p);
1801 total_faults = p->numa_faults_locality[0] +
1802 p->numa_faults_locality[1];
1803 runtime = numa_get_avg_runtime(p, &period);
1805 /* If the task is part of a group prevent parallel updates to group stats */
1806 if (p->numa_group) {
1807 group_lock = &p->numa_group->lock;
1808 spin_lock_irq(group_lock);
1811 /* Find the node with the highest number of faults */
1812 for_each_online_node(nid) {
1813 /* Keep track of the offsets in numa_faults array */
1814 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1815 unsigned long faults = 0, group_faults = 0;
1818 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1819 long diff, f_diff, f_weight;
1821 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1822 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1823 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1824 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1826 /* Decay existing window, copy faults since last scan */
1827 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1828 fault_types[priv] += p->numa_faults[membuf_idx];
1829 p->numa_faults[membuf_idx] = 0;
1832 * Normalize the faults_from, so all tasks in a group
1833 * count according to CPU use, instead of by the raw
1834 * number of faults. Tasks with little runtime have
1835 * little over-all impact on throughput, and thus their
1836 * faults are less important.
1838 f_weight = div64_u64(runtime << 16, period + 1);
1839 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1841 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1842 p->numa_faults[cpubuf_idx] = 0;
1844 p->numa_faults[mem_idx] += diff;
1845 p->numa_faults[cpu_idx] += f_diff;
1846 faults += p->numa_faults[mem_idx];
1847 p->total_numa_faults += diff;
1848 if (p->numa_group) {
1850 * safe because we can only change our own group
1852 * mem_idx represents the offset for a given
1853 * nid and priv in a specific region because it
1854 * is at the beginning of the numa_faults array.
1856 p->numa_group->faults[mem_idx] += diff;
1857 p->numa_group->faults_cpu[mem_idx] += f_diff;
1858 p->numa_group->total_faults += diff;
1859 group_faults += p->numa_group->faults[mem_idx];
1863 if (faults > max_faults) {
1864 max_faults = faults;
1868 if (group_faults > max_group_faults) {
1869 max_group_faults = group_faults;
1870 max_group_nid = nid;
1874 update_task_scan_period(p, fault_types[0], fault_types[1]);
1876 if (p->numa_group) {
1877 update_numa_active_node_mask(p->numa_group);
1878 spin_unlock_irq(group_lock);
1879 max_nid = preferred_group_nid(p, max_group_nid);
1883 /* Set the new preferred node */
1884 if (max_nid != p->numa_preferred_nid)
1885 sched_setnuma(p, max_nid);
1887 if (task_node(p) != p->numa_preferred_nid)
1888 numa_migrate_preferred(p);
1892 static inline int get_numa_group(struct numa_group *grp)
1894 return atomic_inc_not_zero(&grp->refcount);
1897 static inline void put_numa_group(struct numa_group *grp)
1899 if (atomic_dec_and_test(&grp->refcount))
1900 kfree_rcu(grp, rcu);
1903 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1906 struct numa_group *grp, *my_grp;
1907 struct task_struct *tsk;
1909 int cpu = cpupid_to_cpu(cpupid);
1912 if (unlikely(!p->numa_group)) {
1913 unsigned int size = sizeof(struct numa_group) +
1914 4*nr_node_ids*sizeof(unsigned long);
1916 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1920 atomic_set(&grp->refcount, 1);
1921 spin_lock_init(&grp->lock);
1923 /* Second half of the array tracks nids where faults happen */
1924 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1927 node_set(task_node(current), grp->active_nodes);
1929 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1930 grp->faults[i] = p->numa_faults[i];
1932 grp->total_faults = p->total_numa_faults;
1935 rcu_assign_pointer(p->numa_group, grp);
1939 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1941 if (!cpupid_match_pid(tsk, cpupid))
1944 grp = rcu_dereference(tsk->numa_group);
1948 my_grp = p->numa_group;
1953 * Only join the other group if its bigger; if we're the bigger group,
1954 * the other task will join us.
1956 if (my_grp->nr_tasks > grp->nr_tasks)
1960 * Tie-break on the grp address.
1962 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1965 /* Always join threads in the same process. */
1966 if (tsk->mm == current->mm)
1969 /* Simple filter to avoid false positives due to PID collisions */
1970 if (flags & TNF_SHARED)
1973 /* Update priv based on whether false sharing was detected */
1976 if (join && !get_numa_group(grp))
1984 BUG_ON(irqs_disabled());
1985 double_lock_irq(&my_grp->lock, &grp->lock);
1987 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1988 my_grp->faults[i] -= p->numa_faults[i];
1989 grp->faults[i] += p->numa_faults[i];
1991 my_grp->total_faults -= p->total_numa_faults;
1992 grp->total_faults += p->total_numa_faults;
1997 spin_unlock(&my_grp->lock);
1998 spin_unlock_irq(&grp->lock);
2000 rcu_assign_pointer(p->numa_group, grp);
2002 put_numa_group(my_grp);
2010 void task_numa_free(struct task_struct *p)
2012 struct numa_group *grp = p->numa_group;
2013 void *numa_faults = p->numa_faults;
2014 unsigned long flags;
2018 spin_lock_irqsave(&grp->lock, flags);
2019 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2020 grp->faults[i] -= p->numa_faults[i];
2021 grp->total_faults -= p->total_numa_faults;
2024 spin_unlock_irqrestore(&grp->lock, flags);
2025 RCU_INIT_POINTER(p->numa_group, NULL);
2026 put_numa_group(grp);
2029 p->numa_faults = NULL;
2034 * Got a PROT_NONE fault for a page on @node.
2036 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2038 struct task_struct *p = current;
2039 bool migrated = flags & TNF_MIGRATED;
2040 int cpu_node = task_node(current);
2041 int local = !!(flags & TNF_FAULT_LOCAL);
2044 if (!numabalancing_enabled)
2047 /* for example, ksmd faulting in a user's mm */
2051 /* Allocate buffer to track faults on a per-node basis */
2052 if (unlikely(!p->numa_faults)) {
2053 int size = sizeof(*p->numa_faults) *
2054 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2056 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2057 if (!p->numa_faults)
2060 p->total_numa_faults = 0;
2061 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2065 * First accesses are treated as private, otherwise consider accesses
2066 * to be private if the accessing pid has not changed
2068 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2071 priv = cpupid_match_pid(p, last_cpupid);
2072 if (!priv && !(flags & TNF_NO_GROUP))
2073 task_numa_group(p, last_cpupid, flags, &priv);
2077 * If a workload spans multiple NUMA nodes, a shared fault that
2078 * occurs wholly within the set of nodes that the workload is
2079 * actively using should be counted as local. This allows the
2080 * scan rate to slow down when a workload has settled down.
2082 if (!priv && !local && p->numa_group &&
2083 node_isset(cpu_node, p->numa_group->active_nodes) &&
2084 node_isset(mem_node, p->numa_group->active_nodes))
2087 task_numa_placement(p);
2090 * Retry task to preferred node migration periodically, in case it
2091 * case it previously failed, or the scheduler moved us.
2093 if (time_after(jiffies, p->numa_migrate_retry))
2094 numa_migrate_preferred(p);
2097 p->numa_pages_migrated += pages;
2099 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2100 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2101 p->numa_faults_locality[local] += pages;
2104 static void reset_ptenuma_scan(struct task_struct *p)
2106 ACCESS_ONCE(p->mm->numa_scan_seq)++;
2107 p->mm->numa_scan_offset = 0;
2111 * The expensive part of numa migration is done from task_work context.
2112 * Triggered from task_tick_numa().
2114 void task_numa_work(struct callback_head *work)
2116 unsigned long migrate, next_scan, now = jiffies;
2117 struct task_struct *p = current;
2118 struct mm_struct *mm = p->mm;
2119 struct vm_area_struct *vma;
2120 unsigned long start, end;
2121 unsigned long nr_pte_updates = 0;
2124 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2126 work->next = work; /* protect against double add */
2128 * Who cares about NUMA placement when they're dying.
2130 * NOTE: make sure not to dereference p->mm before this check,
2131 * exit_task_work() happens _after_ exit_mm() so we could be called
2132 * without p->mm even though we still had it when we enqueued this
2135 if (p->flags & PF_EXITING)
2138 if (!mm->numa_next_scan) {
2139 mm->numa_next_scan = now +
2140 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2144 * Enforce maximal scan/migration frequency..
2146 migrate = mm->numa_next_scan;
2147 if (time_before(now, migrate))
2150 if (p->numa_scan_period == 0) {
2151 p->numa_scan_period_max = task_scan_max(p);
2152 p->numa_scan_period = task_scan_min(p);
2155 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2156 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2160 * Delay this task enough that another task of this mm will likely win
2161 * the next time around.
2163 p->node_stamp += 2 * TICK_NSEC;
2165 start = mm->numa_scan_offset;
2166 pages = sysctl_numa_balancing_scan_size;
2167 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2171 down_read(&mm->mmap_sem);
2172 vma = find_vma(mm, start);
2174 reset_ptenuma_scan(p);
2178 for (; vma; vma = vma->vm_next) {
2179 if (!vma_migratable(vma) || !vma_policy_mof(vma))
2183 * Shared library pages mapped by multiple processes are not
2184 * migrated as it is expected they are cache replicated. Avoid
2185 * hinting faults in read-only file-backed mappings or the vdso
2186 * as migrating the pages will be of marginal benefit.
2189 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2193 * Skip inaccessible VMAs to avoid any confusion between
2194 * PROT_NONE and NUMA hinting ptes
2196 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2200 start = max(start, vma->vm_start);
2201 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2202 end = min(end, vma->vm_end);
2203 nr_pte_updates += change_prot_numa(vma, start, end);
2206 * Scan sysctl_numa_balancing_scan_size but ensure that
2207 * at least one PTE is updated so that unused virtual
2208 * address space is quickly skipped.
2211 pages -= (end - start) >> PAGE_SHIFT;
2218 } while (end != vma->vm_end);
2223 * It is possible to reach the end of the VMA list but the last few
2224 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2225 * would find the !migratable VMA on the next scan but not reset the
2226 * scanner to the start so check it now.
2229 mm->numa_scan_offset = start;
2231 reset_ptenuma_scan(p);
2232 up_read(&mm->mmap_sem);
2236 * Drive the periodic memory faults..
2238 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2240 struct callback_head *work = &curr->numa_work;
2244 * We don't care about NUMA placement if we don't have memory.
2246 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2250 * Using runtime rather than walltime has the dual advantage that
2251 * we (mostly) drive the selection from busy threads and that the
2252 * task needs to have done some actual work before we bother with
2255 now = curr->se.sum_exec_runtime;
2256 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2258 if (now - curr->node_stamp > period) {
2259 if (!curr->node_stamp)
2260 curr->numa_scan_period = task_scan_min(curr);
2261 curr->node_stamp += period;
2263 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2264 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2265 task_work_add(curr, work, true);
2270 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2274 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2278 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2281 #endif /* CONFIG_NUMA_BALANCING */
2284 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2286 update_load_add(&cfs_rq->load, se->load.weight);
2287 if (!parent_entity(se))
2288 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2290 if (entity_is_task(se)) {
2291 struct rq *rq = rq_of(cfs_rq);
2293 account_numa_enqueue(rq, task_of(se));
2294 list_add(&se->group_node, &rq->cfs_tasks);
2297 cfs_rq->nr_running++;
2301 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2303 update_load_sub(&cfs_rq->load, se->load.weight);
2304 if (!parent_entity(se))
2305 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2306 if (entity_is_task(se)) {
2307 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2308 list_del_init(&se->group_node);
2310 cfs_rq->nr_running--;
2313 #ifdef CONFIG_FAIR_GROUP_SCHED
2315 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2320 * Use this CPU's actual weight instead of the last load_contribution
2321 * to gain a more accurate current total weight. See
2322 * update_cfs_rq_load_contribution().
2324 tg_weight = atomic_long_read(&tg->load_avg);
2325 tg_weight -= cfs_rq->tg_load_contrib;
2326 tg_weight += cfs_rq->load.weight;
2331 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2333 long tg_weight, load, shares;
2335 tg_weight = calc_tg_weight(tg, cfs_rq);
2336 load = cfs_rq->load.weight;
2338 shares = (tg->shares * load);
2340 shares /= tg_weight;
2342 if (shares < MIN_SHARES)
2343 shares = MIN_SHARES;
2344 if (shares > tg->shares)
2345 shares = tg->shares;
2349 # else /* CONFIG_SMP */
2350 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2354 # endif /* CONFIG_SMP */
2355 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2356 unsigned long weight)
2359 /* commit outstanding execution time */
2360 if (cfs_rq->curr == se)
2361 update_curr(cfs_rq);
2362 account_entity_dequeue(cfs_rq, se);
2365 update_load_set(&se->load, weight);
2368 account_entity_enqueue(cfs_rq, se);
2371 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2373 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2375 struct task_group *tg;
2376 struct sched_entity *se;
2380 se = tg->se[cpu_of(rq_of(cfs_rq))];
2381 if (!se || throttled_hierarchy(cfs_rq))
2384 if (likely(se->load.weight == tg->shares))
2387 shares = calc_cfs_shares(cfs_rq, tg);
2389 reweight_entity(cfs_rq_of(se), se, shares);
2391 #else /* CONFIG_FAIR_GROUP_SCHED */
2392 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2395 #endif /* CONFIG_FAIR_GROUP_SCHED */
2399 * We choose a half-life close to 1 scheduling period.
2400 * Note: The tables below are dependent on this value.
2402 #define LOAD_AVG_PERIOD 32
2403 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2404 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2406 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2407 static const u32 runnable_avg_yN_inv[] = {
2408 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2409 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2410 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2411 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2412 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2413 0x85aac367, 0x82cd8698,
2417 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2418 * over-estimates when re-combining.
2420 static const u32 runnable_avg_yN_sum[] = {
2421 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2422 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2423 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2428 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2430 static __always_inline u64 decay_load(u64 val, u64 n)
2432 unsigned int local_n;
2436 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2439 /* after bounds checking we can collapse to 32-bit */
2443 * As y^PERIOD = 1/2, we can combine
2444 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2445 * With a look-up table which covers y^n (n<PERIOD)
2447 * To achieve constant time decay_load.
2449 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2450 val >>= local_n / LOAD_AVG_PERIOD;
2451 local_n %= LOAD_AVG_PERIOD;
2454 val *= runnable_avg_yN_inv[local_n];
2455 /* We don't use SRR here since we always want to round down. */
2460 * For updates fully spanning n periods, the contribution to runnable
2461 * average will be: \Sum 1024*y^n
2463 * We can compute this reasonably efficiently by combining:
2464 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2466 static u32 __compute_runnable_contrib(u64 n)
2470 if (likely(n <= LOAD_AVG_PERIOD))
2471 return runnable_avg_yN_sum[n];
2472 else if (unlikely(n >= LOAD_AVG_MAX_N))
2473 return LOAD_AVG_MAX;
2475 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2477 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2478 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2480 n -= LOAD_AVG_PERIOD;
2481 } while (n > LOAD_AVG_PERIOD);
2483 contrib = decay_load(contrib, n);
2484 return contrib + runnable_avg_yN_sum[n];
2488 * We can represent the historical contribution to runnable average as the
2489 * coefficients of a geometric series. To do this we sub-divide our runnable
2490 * history into segments of approximately 1ms (1024us); label the segment that
2491 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2493 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2495 * (now) (~1ms ago) (~2ms ago)
2497 * Let u_i denote the fraction of p_i that the entity was runnable.
2499 * We then designate the fractions u_i as our co-efficients, yielding the
2500 * following representation of historical load:
2501 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2503 * We choose y based on the with of a reasonably scheduling period, fixing:
2506 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2507 * approximately half as much as the contribution to load within the last ms
2510 * When a period "rolls over" and we have new u_0`, multiplying the previous
2511 * sum again by y is sufficient to update:
2512 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2513 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2515 static __always_inline int __update_entity_runnable_avg(u64 now, int cpu,
2516 struct sched_avg *sa,
2521 u32 runnable_contrib;
2522 int delta_w, decayed = 0;
2523 unsigned long scale_freq = arch_scale_freq_capacity(NULL, cpu);
2525 delta = now - sa->last_runnable_update;
2527 * This should only happen when time goes backwards, which it
2528 * unfortunately does during sched clock init when we swap over to TSC.
2530 if ((s64)delta < 0) {
2531 sa->last_runnable_update = now;
2536 * Use 1024ns as the unit of measurement since it's a reasonable
2537 * approximation of 1us and fast to compute.
2542 sa->last_runnable_update = now;
2544 /* delta_w is the amount already accumulated against our next period */
2545 delta_w = sa->avg_period % 1024;
2546 if (delta + delta_w >= 1024) {
2547 /* period roll-over */
2551 * Now that we know we're crossing a period boundary, figure
2552 * out how much from delta we need to complete the current
2553 * period and accrue it.
2555 delta_w = 1024 - delta_w;
2557 sa->runnable_avg_sum += delta_w;
2559 sa->running_avg_sum += delta_w * scale_freq
2560 >> SCHED_CAPACITY_SHIFT;
2561 sa->avg_period += delta_w;
2565 /* Figure out how many additional periods this update spans */
2566 periods = delta / 1024;
2569 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2571 sa->running_avg_sum = decay_load(sa->running_avg_sum,
2573 sa->avg_period = decay_load(sa->avg_period,
2576 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2577 runnable_contrib = __compute_runnable_contrib(periods);
2579 sa->runnable_avg_sum += runnable_contrib;
2581 sa->running_avg_sum += runnable_contrib * scale_freq
2582 >> SCHED_CAPACITY_SHIFT;
2583 sa->avg_period += runnable_contrib;
2586 /* Remainder of delta accrued against u_0` */
2588 sa->runnable_avg_sum += delta;
2590 sa->running_avg_sum += delta * scale_freq
2591 >> SCHED_CAPACITY_SHIFT;
2592 sa->avg_period += delta;
2597 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2598 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2600 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2601 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2603 decays -= se->avg.decay_count;
2604 se->avg.decay_count = 0;
2608 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2609 se->avg.utilization_avg_contrib =
2610 decay_load(se->avg.utilization_avg_contrib, decays);
2615 #ifdef CONFIG_FAIR_GROUP_SCHED
2616 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2619 struct task_group *tg = cfs_rq->tg;
2622 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2623 tg_contrib -= cfs_rq->tg_load_contrib;
2628 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2629 atomic_long_add(tg_contrib, &tg->load_avg);
2630 cfs_rq->tg_load_contrib += tg_contrib;
2635 * Aggregate cfs_rq runnable averages into an equivalent task_group
2636 * representation for computing load contributions.
2638 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2639 struct cfs_rq *cfs_rq)
2641 struct task_group *tg = cfs_rq->tg;
2644 /* The fraction of a cpu used by this cfs_rq */
2645 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2646 sa->avg_period + 1);
2647 contrib -= cfs_rq->tg_runnable_contrib;
2649 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2650 atomic_add(contrib, &tg->runnable_avg);
2651 cfs_rq->tg_runnable_contrib += contrib;
2655 static inline void __update_group_entity_contrib(struct sched_entity *se)
2657 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2658 struct task_group *tg = cfs_rq->tg;
2663 contrib = cfs_rq->tg_load_contrib * tg->shares;
2664 se->avg.load_avg_contrib = div_u64(contrib,
2665 atomic_long_read(&tg->load_avg) + 1);
2668 * For group entities we need to compute a correction term in the case
2669 * that they are consuming <1 cpu so that we would contribute the same
2670 * load as a task of equal weight.
2672 * Explicitly co-ordinating this measurement would be expensive, but
2673 * fortunately the sum of each cpus contribution forms a usable
2674 * lower-bound on the true value.
2676 * Consider the aggregate of 2 contributions. Either they are disjoint
2677 * (and the sum represents true value) or they are disjoint and we are
2678 * understating by the aggregate of their overlap.
2680 * Extending this to N cpus, for a given overlap, the maximum amount we
2681 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2682 * cpus that overlap for this interval and w_i is the interval width.
2684 * On a small machine; the first term is well-bounded which bounds the
2685 * total error since w_i is a subset of the period. Whereas on a
2686 * larger machine, while this first term can be larger, if w_i is the
2687 * of consequential size guaranteed to see n_i*w_i quickly converge to
2688 * our upper bound of 1-cpu.
2690 runnable_avg = atomic_read(&tg->runnable_avg);
2691 if (runnable_avg < NICE_0_LOAD) {
2692 se->avg.load_avg_contrib *= runnable_avg;
2693 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2697 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2699 __update_entity_runnable_avg(rq_clock_task(rq), cpu_of(rq), &rq->avg,
2700 runnable, runnable);
2701 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2703 #else /* CONFIG_FAIR_GROUP_SCHED */
2704 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2705 int force_update) {}
2706 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2707 struct cfs_rq *cfs_rq) {}
2708 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2709 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2710 #endif /* CONFIG_FAIR_GROUP_SCHED */
2712 static inline void __update_task_entity_contrib(struct sched_entity *se)
2716 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2717 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2718 contrib /= (se->avg.avg_period + 1);
2719 se->avg.load_avg_contrib = scale_load(contrib);
2722 /* Compute the current contribution to load_avg by se, return any delta */
2723 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2725 long old_contrib = se->avg.load_avg_contrib;
2727 if (entity_is_task(se)) {
2728 __update_task_entity_contrib(se);
2730 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2731 __update_group_entity_contrib(se);
2734 return se->avg.load_avg_contrib - old_contrib;
2738 static inline void __update_task_entity_utilization(struct sched_entity *se)
2742 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2743 contrib = se->avg.running_avg_sum * scale_load_down(SCHED_LOAD_SCALE);
2744 contrib /= (se->avg.avg_period + 1);
2745 se->avg.utilization_avg_contrib = scale_load(contrib);
2748 static long __update_entity_utilization_avg_contrib(struct sched_entity *se)
2750 long old_contrib = se->avg.utilization_avg_contrib;
2752 if (entity_is_task(se))
2753 __update_task_entity_utilization(se);
2755 se->avg.utilization_avg_contrib =
2756 group_cfs_rq(se)->utilization_load_avg;
2758 return se->avg.utilization_avg_contrib - old_contrib;
2761 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2764 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2765 cfs_rq->blocked_load_avg -= load_contrib;
2767 cfs_rq->blocked_load_avg = 0;
2770 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2772 /* Update a sched_entity's runnable average */
2773 static inline void update_entity_load_avg(struct sched_entity *se,
2776 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2777 long contrib_delta, utilization_delta;
2778 int cpu = cpu_of(rq_of(cfs_rq));
2782 * For a group entity we need to use their owned cfs_rq_clock_task() in
2783 * case they are the parent of a throttled hierarchy.
2785 if (entity_is_task(se))
2786 now = cfs_rq_clock_task(cfs_rq);
2788 now = cfs_rq_clock_task(group_cfs_rq(se));
2790 if (!__update_entity_runnable_avg(now, cpu, &se->avg, se->on_rq,
2791 cfs_rq->curr == se))
2794 contrib_delta = __update_entity_load_avg_contrib(se);
2795 utilization_delta = __update_entity_utilization_avg_contrib(se);
2801 cfs_rq->runnable_load_avg += contrib_delta;
2802 cfs_rq->utilization_load_avg += utilization_delta;
2804 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2809 * Decay the load contributed by all blocked children and account this so that
2810 * their contribution may appropriately discounted when they wake up.
2812 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2814 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2817 decays = now - cfs_rq->last_decay;
2818 if (!decays && !force_update)
2821 if (atomic_long_read(&cfs_rq->removed_load)) {
2822 unsigned long removed_load;
2823 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2824 subtract_blocked_load_contrib(cfs_rq, removed_load);
2828 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2830 atomic64_add(decays, &cfs_rq->decay_counter);
2831 cfs_rq->last_decay = now;
2834 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2837 /* Add the load generated by se into cfs_rq's child load-average */
2838 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2839 struct sched_entity *se,
2843 * We track migrations using entity decay_count <= 0, on a wake-up
2844 * migration we use a negative decay count to track the remote decays
2845 * accumulated while sleeping.
2847 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2848 * are seen by enqueue_entity_load_avg() as a migration with an already
2849 * constructed load_avg_contrib.
2851 if (unlikely(se->avg.decay_count <= 0)) {
2852 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2853 if (se->avg.decay_count) {
2855 * In a wake-up migration we have to approximate the
2856 * time sleeping. This is because we can't synchronize
2857 * clock_task between the two cpus, and it is not
2858 * guaranteed to be read-safe. Instead, we can
2859 * approximate this using our carried decays, which are
2860 * explicitly atomically readable.
2862 se->avg.last_runnable_update -= (-se->avg.decay_count)
2864 update_entity_load_avg(se, 0);
2865 /* Indicate that we're now synchronized and on-rq */
2866 se->avg.decay_count = 0;
2870 __synchronize_entity_decay(se);
2873 /* migrated tasks did not contribute to our blocked load */
2875 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2876 update_entity_load_avg(se, 0);
2879 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2880 cfs_rq->utilization_load_avg += se->avg.utilization_avg_contrib;
2881 /* we force update consideration on load-balancer moves */
2882 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2886 * Remove se's load from this cfs_rq child load-average, if the entity is
2887 * transitioning to a blocked state we track its projected decay using
2890 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2891 struct sched_entity *se,
2894 update_entity_load_avg(se, 1);
2895 /* we force update consideration on load-balancer moves */
2896 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2898 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2899 cfs_rq->utilization_load_avg -= se->avg.utilization_avg_contrib;
2901 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2902 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2903 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2907 * Update the rq's load with the elapsed running time before entering
2908 * idle. if the last scheduled task is not a CFS task, idle_enter will
2909 * be the only way to update the runnable statistic.
2911 void idle_enter_fair(struct rq *this_rq)
2913 update_rq_runnable_avg(this_rq, 1);
2917 * Update the rq's load with the elapsed idle time before a task is
2918 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2919 * be the only way to update the runnable statistic.
2921 void idle_exit_fair(struct rq *this_rq)
2923 update_rq_runnable_avg(this_rq, 0);
2926 static int idle_balance(struct rq *this_rq);
2928 #else /* CONFIG_SMP */
2930 static inline void update_entity_load_avg(struct sched_entity *se,
2931 int update_cfs_rq) {}
2932 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2933 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2934 struct sched_entity *se,
2936 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2937 struct sched_entity *se,
2939 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2940 int force_update) {}
2942 static inline int idle_balance(struct rq *rq)
2947 #endif /* CONFIG_SMP */
2949 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2951 #ifdef CONFIG_SCHEDSTATS
2952 struct task_struct *tsk = NULL;
2954 if (entity_is_task(se))
2957 if (se->statistics.sleep_start) {
2958 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2963 if (unlikely(delta > se->statistics.sleep_max))
2964 se->statistics.sleep_max = delta;
2966 se->statistics.sleep_start = 0;
2967 se->statistics.sum_sleep_runtime += delta;
2970 account_scheduler_latency(tsk, delta >> 10, 1);
2971 trace_sched_stat_sleep(tsk, delta);
2974 if (se->statistics.block_start) {
2975 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2980 if (unlikely(delta > se->statistics.block_max))
2981 se->statistics.block_max = delta;
2983 se->statistics.block_start = 0;
2984 se->statistics.sum_sleep_runtime += delta;
2987 if (tsk->in_iowait) {
2988 se->statistics.iowait_sum += delta;
2989 se->statistics.iowait_count++;
2990 trace_sched_stat_iowait(tsk, delta);
2993 trace_sched_stat_blocked(tsk, delta);
2996 * Blocking time is in units of nanosecs, so shift by
2997 * 20 to get a milliseconds-range estimation of the
2998 * amount of time that the task spent sleeping:
3000 if (unlikely(prof_on == SLEEP_PROFILING)) {
3001 profile_hits(SLEEP_PROFILING,
3002 (void *)get_wchan(tsk),
3005 account_scheduler_latency(tsk, delta >> 10, 0);
3011 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3013 #ifdef CONFIG_SCHED_DEBUG
3014 s64 d = se->vruntime - cfs_rq->min_vruntime;
3019 if (d > 3*sysctl_sched_latency)
3020 schedstat_inc(cfs_rq, nr_spread_over);
3025 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3027 u64 vruntime = cfs_rq->min_vruntime;
3030 * The 'current' period is already promised to the current tasks,
3031 * however the extra weight of the new task will slow them down a
3032 * little, place the new task so that it fits in the slot that
3033 * stays open at the end.
3035 if (initial && sched_feat(START_DEBIT))
3036 vruntime += sched_vslice(cfs_rq, se);
3038 /* sleeps up to a single latency don't count. */
3040 unsigned long thresh = sysctl_sched_latency;
3043 * Halve their sleep time's effect, to allow
3044 * for a gentler effect of sleepers:
3046 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3052 /* ensure we never gain time by being placed backwards. */
3053 se->vruntime = max_vruntime(se->vruntime, vruntime);
3056 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3059 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3062 * Update the normalized vruntime before updating min_vruntime
3063 * through calling update_curr().
3065 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3066 se->vruntime += cfs_rq->min_vruntime;
3069 * Update run-time statistics of the 'current'.
3071 update_curr(cfs_rq);
3072 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
3073 account_entity_enqueue(cfs_rq, se);
3074 update_cfs_shares(cfs_rq);
3076 if (flags & ENQUEUE_WAKEUP) {
3077 place_entity(cfs_rq, se, 0);
3078 enqueue_sleeper(cfs_rq, se);
3081 update_stats_enqueue(cfs_rq, se);
3082 check_spread(cfs_rq, se);
3083 if (se != cfs_rq->curr)
3084 __enqueue_entity(cfs_rq, se);
3087 if (cfs_rq->nr_running == 1) {
3088 list_add_leaf_cfs_rq(cfs_rq);
3089 check_enqueue_throttle(cfs_rq);
3093 static void __clear_buddies_last(struct sched_entity *se)
3095 for_each_sched_entity(se) {
3096 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3097 if (cfs_rq->last != se)
3100 cfs_rq->last = NULL;
3104 static void __clear_buddies_next(struct sched_entity *se)
3106 for_each_sched_entity(se) {
3107 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3108 if (cfs_rq->next != se)
3111 cfs_rq->next = NULL;
3115 static void __clear_buddies_skip(struct sched_entity *se)
3117 for_each_sched_entity(se) {
3118 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3119 if (cfs_rq->skip != se)
3122 cfs_rq->skip = NULL;
3126 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3128 if (cfs_rq->last == se)
3129 __clear_buddies_last(se);
3131 if (cfs_rq->next == se)
3132 __clear_buddies_next(se);
3134 if (cfs_rq->skip == se)
3135 __clear_buddies_skip(se);
3138 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3141 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3144 * Update run-time statistics of the 'current'.
3146 update_curr(cfs_rq);
3147 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
3149 update_stats_dequeue(cfs_rq, se);
3150 if (flags & DEQUEUE_SLEEP) {
3151 #ifdef CONFIG_SCHEDSTATS
3152 if (entity_is_task(se)) {
3153 struct task_struct *tsk = task_of(se);
3155 if (tsk->state & TASK_INTERRUPTIBLE)
3156 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3157 if (tsk->state & TASK_UNINTERRUPTIBLE)
3158 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3163 clear_buddies(cfs_rq, se);
3165 if (se != cfs_rq->curr)
3166 __dequeue_entity(cfs_rq, se);
3168 account_entity_dequeue(cfs_rq, se);
3171 * Normalize the entity after updating the min_vruntime because the
3172 * update can refer to the ->curr item and we need to reflect this
3173 * movement in our normalized position.
3175 if (!(flags & DEQUEUE_SLEEP))
3176 se->vruntime -= cfs_rq->min_vruntime;
3178 /* return excess runtime on last dequeue */
3179 return_cfs_rq_runtime(cfs_rq);
3181 update_min_vruntime(cfs_rq);
3182 update_cfs_shares(cfs_rq);
3186 * Preempt the current task with a newly woken task if needed:
3189 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3191 unsigned long ideal_runtime, delta_exec;
3192 struct sched_entity *se;
3195 ideal_runtime = sched_slice(cfs_rq, curr);
3196 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3197 if (delta_exec > ideal_runtime) {
3198 resched_curr(rq_of(cfs_rq));
3200 * The current task ran long enough, ensure it doesn't get
3201 * re-elected due to buddy favours.
3203 clear_buddies(cfs_rq, curr);
3208 * Ensure that a task that missed wakeup preemption by a
3209 * narrow margin doesn't have to wait for a full slice.
3210 * This also mitigates buddy induced latencies under load.
3212 if (delta_exec < sysctl_sched_min_granularity)
3215 se = __pick_first_entity(cfs_rq);
3216 delta = curr->vruntime - se->vruntime;
3221 if (delta > ideal_runtime)
3222 resched_curr(rq_of(cfs_rq));
3226 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3228 /* 'current' is not kept within the tree. */
3231 * Any task has to be enqueued before it get to execute on
3232 * a CPU. So account for the time it spent waiting on the
3235 update_stats_wait_end(cfs_rq, se);
3236 __dequeue_entity(cfs_rq, se);
3237 update_entity_load_avg(se, 1);
3240 update_stats_curr_start(cfs_rq, se);
3242 #ifdef CONFIG_SCHEDSTATS
3244 * Track our maximum slice length, if the CPU's load is at
3245 * least twice that of our own weight (i.e. dont track it
3246 * when there are only lesser-weight tasks around):
3248 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3249 se->statistics.slice_max = max(se->statistics.slice_max,
3250 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3253 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3257 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3260 * Pick the next process, keeping these things in mind, in this order:
3261 * 1) keep things fair between processes/task groups
3262 * 2) pick the "next" process, since someone really wants that to run
3263 * 3) pick the "last" process, for cache locality
3264 * 4) do not run the "skip" process, if something else is available
3266 static struct sched_entity *
3267 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3269 struct sched_entity *left = __pick_first_entity(cfs_rq);
3270 struct sched_entity *se;
3273 * If curr is set we have to see if its left of the leftmost entity
3274 * still in the tree, provided there was anything in the tree at all.
3276 if (!left || (curr && entity_before(curr, left)))
3279 se = left; /* ideally we run the leftmost entity */
3282 * Avoid running the skip buddy, if running something else can
3283 * be done without getting too unfair.
3285 if (cfs_rq->skip == se) {
3286 struct sched_entity *second;
3289 second = __pick_first_entity(cfs_rq);
3291 second = __pick_next_entity(se);
3292 if (!second || (curr && entity_before(curr, second)))
3296 if (second && wakeup_preempt_entity(second, left) < 1)
3301 * Prefer last buddy, try to return the CPU to a preempted task.
3303 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3307 * Someone really wants this to run. If it's not unfair, run it.
3309 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3312 clear_buddies(cfs_rq, se);
3317 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3319 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3322 * If still on the runqueue then deactivate_task()
3323 * was not called and update_curr() has to be done:
3326 update_curr(cfs_rq);
3328 /* throttle cfs_rqs exceeding runtime */
3329 check_cfs_rq_runtime(cfs_rq);
3331 check_spread(cfs_rq, prev);
3333 update_stats_wait_start(cfs_rq, prev);
3334 /* Put 'current' back into the tree. */
3335 __enqueue_entity(cfs_rq, prev);
3336 /* in !on_rq case, update occurred at dequeue */
3337 update_entity_load_avg(prev, 1);
3339 cfs_rq->curr = NULL;
3343 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3346 * Update run-time statistics of the 'current'.
3348 update_curr(cfs_rq);
3351 * Ensure that runnable average is periodically updated.
3353 update_entity_load_avg(curr, 1);
3354 update_cfs_rq_blocked_load(cfs_rq, 1);
3355 update_cfs_shares(cfs_rq);
3357 #ifdef CONFIG_SCHED_HRTICK
3359 * queued ticks are scheduled to match the slice, so don't bother
3360 * validating it and just reschedule.
3363 resched_curr(rq_of(cfs_rq));
3367 * don't let the period tick interfere with the hrtick preemption
3369 if (!sched_feat(DOUBLE_TICK) &&
3370 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3374 if (cfs_rq->nr_running > 1)
3375 check_preempt_tick(cfs_rq, curr);
3379 /**************************************************
3380 * CFS bandwidth control machinery
3383 #ifdef CONFIG_CFS_BANDWIDTH
3385 #ifdef HAVE_JUMP_LABEL
3386 static struct static_key __cfs_bandwidth_used;
3388 static inline bool cfs_bandwidth_used(void)
3390 return static_key_false(&__cfs_bandwidth_used);
3393 void cfs_bandwidth_usage_inc(void)
3395 static_key_slow_inc(&__cfs_bandwidth_used);
3398 void cfs_bandwidth_usage_dec(void)
3400 static_key_slow_dec(&__cfs_bandwidth_used);
3402 #else /* HAVE_JUMP_LABEL */
3403 static bool cfs_bandwidth_used(void)
3408 void cfs_bandwidth_usage_inc(void) {}
3409 void cfs_bandwidth_usage_dec(void) {}
3410 #endif /* HAVE_JUMP_LABEL */
3413 * default period for cfs group bandwidth.
3414 * default: 0.1s, units: nanoseconds
3416 static inline u64 default_cfs_period(void)
3418 return 100000000ULL;
3421 static inline u64 sched_cfs_bandwidth_slice(void)
3423 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3427 * Replenish runtime according to assigned quota and update expiration time.
3428 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3429 * additional synchronization around rq->lock.
3431 * requires cfs_b->lock
3433 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3437 if (cfs_b->quota == RUNTIME_INF)
3440 now = sched_clock_cpu(smp_processor_id());
3441 cfs_b->runtime = cfs_b->quota;
3442 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3445 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3447 return &tg->cfs_bandwidth;
3450 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3451 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3453 if (unlikely(cfs_rq->throttle_count))
3454 return cfs_rq->throttled_clock_task;
3456 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3459 /* returns 0 on failure to allocate runtime */
3460 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3462 struct task_group *tg = cfs_rq->tg;
3463 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3464 u64 amount = 0, min_amount, expires;
3466 /* note: this is a positive sum as runtime_remaining <= 0 */
3467 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3469 raw_spin_lock(&cfs_b->lock);
3470 if (cfs_b->quota == RUNTIME_INF)
3471 amount = min_amount;
3474 * If the bandwidth pool has become inactive, then at least one
3475 * period must have elapsed since the last consumption.
3476 * Refresh the global state and ensure bandwidth timer becomes
3479 if (!cfs_b->timer_active) {
3480 __refill_cfs_bandwidth_runtime(cfs_b);
3481 __start_cfs_bandwidth(cfs_b, false);
3484 if (cfs_b->runtime > 0) {
3485 amount = min(cfs_b->runtime, min_amount);
3486 cfs_b->runtime -= amount;
3490 expires = cfs_b->runtime_expires;
3491 raw_spin_unlock(&cfs_b->lock);
3493 cfs_rq->runtime_remaining += amount;
3495 * we may have advanced our local expiration to account for allowed
3496 * spread between our sched_clock and the one on which runtime was
3499 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3500 cfs_rq->runtime_expires = expires;
3502 return cfs_rq->runtime_remaining > 0;
3506 * Note: This depends on the synchronization provided by sched_clock and the
3507 * fact that rq->clock snapshots this value.
3509 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3511 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3513 /* if the deadline is ahead of our clock, nothing to do */
3514 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3517 if (cfs_rq->runtime_remaining < 0)
3521 * If the local deadline has passed we have to consider the
3522 * possibility that our sched_clock is 'fast' and the global deadline
3523 * has not truly expired.
3525 * Fortunately we can check determine whether this the case by checking
3526 * whether the global deadline has advanced. It is valid to compare
3527 * cfs_b->runtime_expires without any locks since we only care about
3528 * exact equality, so a partial write will still work.
3531 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3532 /* extend local deadline, drift is bounded above by 2 ticks */
3533 cfs_rq->runtime_expires += TICK_NSEC;
3535 /* global deadline is ahead, expiration has passed */
3536 cfs_rq->runtime_remaining = 0;
3540 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3542 /* dock delta_exec before expiring quota (as it could span periods) */
3543 cfs_rq->runtime_remaining -= delta_exec;
3544 expire_cfs_rq_runtime(cfs_rq);
3546 if (likely(cfs_rq->runtime_remaining > 0))
3550 * if we're unable to extend our runtime we resched so that the active
3551 * hierarchy can be throttled
3553 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3554 resched_curr(rq_of(cfs_rq));
3557 static __always_inline
3558 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3560 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3563 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3566 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3568 return cfs_bandwidth_used() && cfs_rq->throttled;
3571 /* check whether cfs_rq, or any parent, is throttled */
3572 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3574 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3578 * Ensure that neither of the group entities corresponding to src_cpu or
3579 * dest_cpu are members of a throttled hierarchy when performing group
3580 * load-balance operations.
3582 static inline int throttled_lb_pair(struct task_group *tg,
3583 int src_cpu, int dest_cpu)
3585 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3587 src_cfs_rq = tg->cfs_rq[src_cpu];
3588 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3590 return throttled_hierarchy(src_cfs_rq) ||
3591 throttled_hierarchy(dest_cfs_rq);
3594 /* updated child weight may affect parent so we have to do this bottom up */
3595 static int tg_unthrottle_up(struct task_group *tg, void *data)
3597 struct rq *rq = data;
3598 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3600 cfs_rq->throttle_count--;
3602 if (!cfs_rq->throttle_count) {
3603 /* adjust cfs_rq_clock_task() */
3604 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3605 cfs_rq->throttled_clock_task;
3612 static int tg_throttle_down(struct task_group *tg, void *data)
3614 struct rq *rq = data;
3615 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3617 /* group is entering throttled state, stop time */
3618 if (!cfs_rq->throttle_count)
3619 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3620 cfs_rq->throttle_count++;
3625 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3627 struct rq *rq = rq_of(cfs_rq);
3628 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3629 struct sched_entity *se;
3630 long task_delta, dequeue = 1;
3632 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3634 /* freeze hierarchy runnable averages while throttled */
3636 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3639 task_delta = cfs_rq->h_nr_running;
3640 for_each_sched_entity(se) {
3641 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3642 /* throttled entity or throttle-on-deactivate */
3647 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3648 qcfs_rq->h_nr_running -= task_delta;
3650 if (qcfs_rq->load.weight)
3655 sub_nr_running(rq, task_delta);
3657 cfs_rq->throttled = 1;
3658 cfs_rq->throttled_clock = rq_clock(rq);
3659 raw_spin_lock(&cfs_b->lock);
3661 * Add to the _head_ of the list, so that an already-started
3662 * distribute_cfs_runtime will not see us
3664 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3665 if (!cfs_b->timer_active)
3666 __start_cfs_bandwidth(cfs_b, false);
3667 raw_spin_unlock(&cfs_b->lock);
3670 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3672 struct rq *rq = rq_of(cfs_rq);
3673 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3674 struct sched_entity *se;
3678 se = cfs_rq->tg->se[cpu_of(rq)];
3680 cfs_rq->throttled = 0;
3682 update_rq_clock(rq);
3684 raw_spin_lock(&cfs_b->lock);
3685 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3686 list_del_rcu(&cfs_rq->throttled_list);
3687 raw_spin_unlock(&cfs_b->lock);
3689 /* update hierarchical throttle state */
3690 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3692 if (!cfs_rq->load.weight)
3695 task_delta = cfs_rq->h_nr_running;
3696 for_each_sched_entity(se) {
3700 cfs_rq = cfs_rq_of(se);
3702 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3703 cfs_rq->h_nr_running += task_delta;
3705 if (cfs_rq_throttled(cfs_rq))
3710 add_nr_running(rq, task_delta);
3712 /* determine whether we need to wake up potentially idle cpu */
3713 if (rq->curr == rq->idle && rq->cfs.nr_running)
3717 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3718 u64 remaining, u64 expires)
3720 struct cfs_rq *cfs_rq;
3722 u64 starting_runtime = remaining;
3725 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3727 struct rq *rq = rq_of(cfs_rq);
3729 raw_spin_lock(&rq->lock);
3730 if (!cfs_rq_throttled(cfs_rq))
3733 runtime = -cfs_rq->runtime_remaining + 1;
3734 if (runtime > remaining)
3735 runtime = remaining;
3736 remaining -= runtime;
3738 cfs_rq->runtime_remaining += runtime;
3739 cfs_rq->runtime_expires = expires;
3741 /* we check whether we're throttled above */
3742 if (cfs_rq->runtime_remaining > 0)
3743 unthrottle_cfs_rq(cfs_rq);
3746 raw_spin_unlock(&rq->lock);
3753 return starting_runtime - remaining;
3757 * Responsible for refilling a task_group's bandwidth and unthrottling its
3758 * cfs_rqs as appropriate. If there has been no activity within the last
3759 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3760 * used to track this state.
3762 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3764 u64 runtime, runtime_expires;
3767 /* no need to continue the timer with no bandwidth constraint */
3768 if (cfs_b->quota == RUNTIME_INF)
3769 goto out_deactivate;
3771 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3772 cfs_b->nr_periods += overrun;
3775 * idle depends on !throttled (for the case of a large deficit), and if
3776 * we're going inactive then everything else can be deferred
3778 if (cfs_b->idle && !throttled)
3779 goto out_deactivate;
3782 * if we have relooped after returning idle once, we need to update our
3783 * status as actually running, so that other cpus doing
3784 * __start_cfs_bandwidth will stop trying to cancel us.
3786 cfs_b->timer_active = 1;
3788 __refill_cfs_bandwidth_runtime(cfs_b);
3791 /* mark as potentially idle for the upcoming period */
3796 /* account preceding periods in which throttling occurred */
3797 cfs_b->nr_throttled += overrun;
3799 runtime_expires = cfs_b->runtime_expires;
3802 * This check is repeated as we are holding onto the new bandwidth while
3803 * we unthrottle. This can potentially race with an unthrottled group
3804 * trying to acquire new bandwidth from the global pool. This can result
3805 * in us over-using our runtime if it is all used during this loop, but
3806 * only by limited amounts in that extreme case.
3808 while (throttled && cfs_b->runtime > 0) {
3809 runtime = cfs_b->runtime;
3810 raw_spin_unlock(&cfs_b->lock);
3811 /* we can't nest cfs_b->lock while distributing bandwidth */
3812 runtime = distribute_cfs_runtime(cfs_b, runtime,
3814 raw_spin_lock(&cfs_b->lock);
3816 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3818 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3822 * While we are ensured activity in the period following an
3823 * unthrottle, this also covers the case in which the new bandwidth is
3824 * insufficient to cover the existing bandwidth deficit. (Forcing the
3825 * timer to remain active while there are any throttled entities.)
3832 cfs_b->timer_active = 0;
3836 /* a cfs_rq won't donate quota below this amount */
3837 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3838 /* minimum remaining period time to redistribute slack quota */
3839 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3840 /* how long we wait to gather additional slack before distributing */
3841 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3844 * Are we near the end of the current quota period?
3846 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3847 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3848 * migrate_hrtimers, base is never cleared, so we are fine.
3850 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3852 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3855 /* if the call-back is running a quota refresh is already occurring */
3856 if (hrtimer_callback_running(refresh_timer))
3859 /* is a quota refresh about to occur? */
3860 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3861 if (remaining < min_expire)
3867 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3869 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3871 /* if there's a quota refresh soon don't bother with slack */
3872 if (runtime_refresh_within(cfs_b, min_left))
3875 start_bandwidth_timer(&cfs_b->slack_timer,
3876 ns_to_ktime(cfs_bandwidth_slack_period));
3879 /* we know any runtime found here is valid as update_curr() precedes return */
3880 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3882 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3883 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3885 if (slack_runtime <= 0)
3888 raw_spin_lock(&cfs_b->lock);
3889 if (cfs_b->quota != RUNTIME_INF &&
3890 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3891 cfs_b->runtime += slack_runtime;
3893 /* we are under rq->lock, defer unthrottling using a timer */
3894 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3895 !list_empty(&cfs_b->throttled_cfs_rq))
3896 start_cfs_slack_bandwidth(cfs_b);
3898 raw_spin_unlock(&cfs_b->lock);
3900 /* even if it's not valid for return we don't want to try again */
3901 cfs_rq->runtime_remaining -= slack_runtime;
3904 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3906 if (!cfs_bandwidth_used())
3909 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3912 __return_cfs_rq_runtime(cfs_rq);
3916 * This is done with a timer (instead of inline with bandwidth return) since
3917 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3919 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3921 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3924 /* confirm we're still not at a refresh boundary */
3925 raw_spin_lock(&cfs_b->lock);
3926 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3927 raw_spin_unlock(&cfs_b->lock);
3931 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3932 runtime = cfs_b->runtime;
3934 expires = cfs_b->runtime_expires;
3935 raw_spin_unlock(&cfs_b->lock);
3940 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3942 raw_spin_lock(&cfs_b->lock);
3943 if (expires == cfs_b->runtime_expires)
3944 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3945 raw_spin_unlock(&cfs_b->lock);
3949 * When a group wakes up we want to make sure that its quota is not already
3950 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3951 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3953 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3955 if (!cfs_bandwidth_used())
3958 /* an active group must be handled by the update_curr()->put() path */
3959 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3962 /* ensure the group is not already throttled */
3963 if (cfs_rq_throttled(cfs_rq))
3966 /* update runtime allocation */
3967 account_cfs_rq_runtime(cfs_rq, 0);
3968 if (cfs_rq->runtime_remaining <= 0)
3969 throttle_cfs_rq(cfs_rq);
3972 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3973 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3975 if (!cfs_bandwidth_used())
3978 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3982 * it's possible for a throttled entity to be forced into a running
3983 * state (e.g. set_curr_task), in this case we're finished.
3985 if (cfs_rq_throttled(cfs_rq))
3988 throttle_cfs_rq(cfs_rq);
3992 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3994 struct cfs_bandwidth *cfs_b =
3995 container_of(timer, struct cfs_bandwidth, slack_timer);
3996 do_sched_cfs_slack_timer(cfs_b);
3998 return HRTIMER_NORESTART;
4001 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4003 struct cfs_bandwidth *cfs_b =
4004 container_of(timer, struct cfs_bandwidth, period_timer);
4009 raw_spin_lock(&cfs_b->lock);
4011 now = hrtimer_cb_get_time(timer);
4012 overrun = hrtimer_forward(timer, now, cfs_b->period);
4017 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4019 raw_spin_unlock(&cfs_b->lock);
4021 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4024 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4026 raw_spin_lock_init(&cfs_b->lock);
4028 cfs_b->quota = RUNTIME_INF;
4029 cfs_b->period = ns_to_ktime(default_cfs_period());
4031 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4032 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4033 cfs_b->period_timer.function = sched_cfs_period_timer;
4034 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4035 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4038 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4040 cfs_rq->runtime_enabled = 0;
4041 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4044 /* requires cfs_b->lock, may release to reprogram timer */
4045 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force)
4048 * The timer may be active because we're trying to set a new bandwidth
4049 * period or because we're racing with the tear-down path
4050 * (timer_active==0 becomes visible before the hrtimer call-back
4051 * terminates). In either case we ensure that it's re-programmed
4053 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
4054 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
4055 /* bounce the lock to allow do_sched_cfs_period_timer to run */
4056 raw_spin_unlock(&cfs_b->lock);
4058 raw_spin_lock(&cfs_b->lock);
4059 /* if someone else restarted the timer then we're done */
4060 if (!force && cfs_b->timer_active)
4064 cfs_b->timer_active = 1;
4065 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
4068 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4070 /* init_cfs_bandwidth() was not called */
4071 if (!cfs_b->throttled_cfs_rq.next)
4074 hrtimer_cancel(&cfs_b->period_timer);
4075 hrtimer_cancel(&cfs_b->slack_timer);
4078 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4080 struct cfs_rq *cfs_rq;
4082 for_each_leaf_cfs_rq(rq, cfs_rq) {
4083 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4085 raw_spin_lock(&cfs_b->lock);
4086 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4087 raw_spin_unlock(&cfs_b->lock);
4091 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4093 struct cfs_rq *cfs_rq;
4095 for_each_leaf_cfs_rq(rq, cfs_rq) {
4096 if (!cfs_rq->runtime_enabled)
4100 * clock_task is not advancing so we just need to make sure
4101 * there's some valid quota amount
4103 cfs_rq->runtime_remaining = 1;
4105 * Offline rq is schedulable till cpu is completely disabled
4106 * in take_cpu_down(), so we prevent new cfs throttling here.
4108 cfs_rq->runtime_enabled = 0;
4110 if (cfs_rq_throttled(cfs_rq))
4111 unthrottle_cfs_rq(cfs_rq);
4115 #else /* CONFIG_CFS_BANDWIDTH */
4116 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4118 return rq_clock_task(rq_of(cfs_rq));
4121 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4122 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4123 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4124 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4126 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4131 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4136 static inline int throttled_lb_pair(struct task_group *tg,
4137 int src_cpu, int dest_cpu)
4142 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4144 #ifdef CONFIG_FAIR_GROUP_SCHED
4145 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4148 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4152 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4153 static inline void update_runtime_enabled(struct rq *rq) {}
4154 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4156 #endif /* CONFIG_CFS_BANDWIDTH */
4158 /**************************************************
4159 * CFS operations on tasks:
4162 #ifdef CONFIG_SCHED_HRTICK
4163 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4165 struct sched_entity *se = &p->se;
4166 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4168 WARN_ON(task_rq(p) != rq);
4170 if (cfs_rq->nr_running > 1) {
4171 u64 slice = sched_slice(cfs_rq, se);
4172 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4173 s64 delta = slice - ran;
4180 hrtick_start(rq, delta);
4185 * called from enqueue/dequeue and updates the hrtick when the
4186 * current task is from our class and nr_running is low enough
4189 static void hrtick_update(struct rq *rq)
4191 struct task_struct *curr = rq->curr;
4193 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4196 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4197 hrtick_start_fair(rq, curr);
4199 #else /* !CONFIG_SCHED_HRTICK */
4201 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4205 static inline void hrtick_update(struct rq *rq)
4211 * The enqueue_task method is called before nr_running is
4212 * increased. Here we update the fair scheduling stats and
4213 * then put the task into the rbtree:
4216 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4218 struct cfs_rq *cfs_rq;
4219 struct sched_entity *se = &p->se;
4221 for_each_sched_entity(se) {
4224 cfs_rq = cfs_rq_of(se);
4225 enqueue_entity(cfs_rq, se, flags);
4228 * end evaluation on encountering a throttled cfs_rq
4230 * note: in the case of encountering a throttled cfs_rq we will
4231 * post the final h_nr_running increment below.
4233 if (cfs_rq_throttled(cfs_rq))
4235 cfs_rq->h_nr_running++;
4237 flags = ENQUEUE_WAKEUP;
4240 for_each_sched_entity(se) {
4241 cfs_rq = cfs_rq_of(se);
4242 cfs_rq->h_nr_running++;
4244 if (cfs_rq_throttled(cfs_rq))
4247 update_cfs_shares(cfs_rq);
4248 update_entity_load_avg(se, 1);
4252 update_rq_runnable_avg(rq, rq->nr_running);
4253 add_nr_running(rq, 1);
4258 static void set_next_buddy(struct sched_entity *se);
4261 * The dequeue_task method is called before nr_running is
4262 * decreased. We remove the task from the rbtree and
4263 * update the fair scheduling stats:
4265 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4267 struct cfs_rq *cfs_rq;
4268 struct sched_entity *se = &p->se;
4269 int task_sleep = flags & DEQUEUE_SLEEP;
4271 for_each_sched_entity(se) {
4272 cfs_rq = cfs_rq_of(se);
4273 dequeue_entity(cfs_rq, se, flags);
4276 * end evaluation on encountering a throttled cfs_rq
4278 * note: in the case of encountering a throttled cfs_rq we will
4279 * post the final h_nr_running decrement below.
4281 if (cfs_rq_throttled(cfs_rq))
4283 cfs_rq->h_nr_running--;
4285 /* Don't dequeue parent if it has other entities besides us */
4286 if (cfs_rq->load.weight) {
4288 * Bias pick_next to pick a task from this cfs_rq, as
4289 * p is sleeping when it is within its sched_slice.
4291 if (task_sleep && parent_entity(se))
4292 set_next_buddy(parent_entity(se));
4294 /* avoid re-evaluating load for this entity */
4295 se = parent_entity(se);
4298 flags |= DEQUEUE_SLEEP;
4301 for_each_sched_entity(se) {
4302 cfs_rq = cfs_rq_of(se);
4303 cfs_rq->h_nr_running--;
4305 if (cfs_rq_throttled(cfs_rq))
4308 update_cfs_shares(cfs_rq);
4309 update_entity_load_avg(se, 1);
4313 sub_nr_running(rq, 1);
4314 update_rq_runnable_avg(rq, 1);
4320 /* Used instead of source_load when we know the type == 0 */
4321 static unsigned long weighted_cpuload(const int cpu)
4323 return cpu_rq(cpu)->cfs.runnable_load_avg;
4327 * Return a low guess at the load of a migration-source cpu weighted
4328 * according to the scheduling class and "nice" value.
4330 * We want to under-estimate the load of migration sources, to
4331 * balance conservatively.
4333 static unsigned long source_load(int cpu, int type)
4335 struct rq *rq = cpu_rq(cpu);
4336 unsigned long total = weighted_cpuload(cpu);
4338 if (type == 0 || !sched_feat(LB_BIAS))
4341 return min(rq->cpu_load[type-1], total);
4345 * Return a high guess at the load of a migration-target cpu weighted
4346 * according to the scheduling class and "nice" value.
4348 static unsigned long target_load(int cpu, int type)
4350 struct rq *rq = cpu_rq(cpu);
4351 unsigned long total = weighted_cpuload(cpu);
4353 if (type == 0 || !sched_feat(LB_BIAS))
4356 return max(rq->cpu_load[type-1], total);
4359 static unsigned long capacity_of(int cpu)
4361 return cpu_rq(cpu)->cpu_capacity;
4364 static unsigned long capacity_orig_of(int cpu)
4366 return cpu_rq(cpu)->cpu_capacity_orig;
4369 static unsigned long cpu_avg_load_per_task(int cpu)
4371 struct rq *rq = cpu_rq(cpu);
4372 unsigned long nr_running = ACCESS_ONCE(rq->cfs.h_nr_running);
4373 unsigned long load_avg = rq->cfs.runnable_load_avg;
4376 return load_avg / nr_running;
4381 static void record_wakee(struct task_struct *p)
4384 * Rough decay (wiping) for cost saving, don't worry
4385 * about the boundary, really active task won't care
4388 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4389 current->wakee_flips >>= 1;
4390 current->wakee_flip_decay_ts = jiffies;
4393 if (current->last_wakee != p) {
4394 current->last_wakee = p;
4395 current->wakee_flips++;
4399 static void task_waking_fair(struct task_struct *p)
4401 struct sched_entity *se = &p->se;
4402 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4405 #ifndef CONFIG_64BIT
4406 u64 min_vruntime_copy;
4409 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4411 min_vruntime = cfs_rq->min_vruntime;
4412 } while (min_vruntime != min_vruntime_copy);
4414 min_vruntime = cfs_rq->min_vruntime;
4417 se->vruntime -= min_vruntime;
4421 #ifdef CONFIG_FAIR_GROUP_SCHED
4423 * effective_load() calculates the load change as seen from the root_task_group
4425 * Adding load to a group doesn't make a group heavier, but can cause movement
4426 * of group shares between cpus. Assuming the shares were perfectly aligned one
4427 * can calculate the shift in shares.
4429 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4430 * on this @cpu and results in a total addition (subtraction) of @wg to the
4431 * total group weight.
4433 * Given a runqueue weight distribution (rw_i) we can compute a shares
4434 * distribution (s_i) using:
4436 * s_i = rw_i / \Sum rw_j (1)
4438 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4439 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4440 * shares distribution (s_i):
4442 * rw_i = { 2, 4, 1, 0 }
4443 * s_i = { 2/7, 4/7, 1/7, 0 }
4445 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4446 * task used to run on and the CPU the waker is running on), we need to
4447 * compute the effect of waking a task on either CPU and, in case of a sync
4448 * wakeup, compute the effect of the current task going to sleep.
4450 * So for a change of @wl to the local @cpu with an overall group weight change
4451 * of @wl we can compute the new shares distribution (s'_i) using:
4453 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4455 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4456 * differences in waking a task to CPU 0. The additional task changes the
4457 * weight and shares distributions like:
4459 * rw'_i = { 3, 4, 1, 0 }
4460 * s'_i = { 3/8, 4/8, 1/8, 0 }
4462 * We can then compute the difference in effective weight by using:
4464 * dw_i = S * (s'_i - s_i) (3)
4466 * Where 'S' is the group weight as seen by its parent.
4468 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4469 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4470 * 4/7) times the weight of the group.
4472 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4474 struct sched_entity *se = tg->se[cpu];
4476 if (!tg->parent) /* the trivial, non-cgroup case */
4479 for_each_sched_entity(se) {
4485 * W = @wg + \Sum rw_j
4487 W = wg + calc_tg_weight(tg, se->my_q);
4492 w = se->my_q->load.weight + wl;
4495 * wl = S * s'_i; see (2)
4498 wl = (w * (long)tg->shares) / W;
4503 * Per the above, wl is the new se->load.weight value; since
4504 * those are clipped to [MIN_SHARES, ...) do so now. See
4505 * calc_cfs_shares().
4507 if (wl < MIN_SHARES)
4511 * wl = dw_i = S * (s'_i - s_i); see (3)
4513 wl -= se->load.weight;
4516 * Recursively apply this logic to all parent groups to compute
4517 * the final effective load change on the root group. Since
4518 * only the @tg group gets extra weight, all parent groups can
4519 * only redistribute existing shares. @wl is the shift in shares
4520 * resulting from this level per the above.
4529 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4536 static int wake_wide(struct task_struct *p)
4538 int factor = this_cpu_read(sd_llc_size);
4541 * Yeah, it's the switching-frequency, could means many wakee or
4542 * rapidly switch, use factor here will just help to automatically
4543 * adjust the loose-degree, so bigger node will lead to more pull.
4545 if (p->wakee_flips > factor) {
4547 * wakee is somewhat hot, it needs certain amount of cpu
4548 * resource, so if waker is far more hot, prefer to leave
4551 if (current->wakee_flips > (factor * p->wakee_flips))
4558 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4560 s64 this_load, load;
4561 s64 this_eff_load, prev_eff_load;
4562 int idx, this_cpu, prev_cpu;
4563 struct task_group *tg;
4564 unsigned long weight;
4568 * If we wake multiple tasks be careful to not bounce
4569 * ourselves around too much.
4575 this_cpu = smp_processor_id();
4576 prev_cpu = task_cpu(p);
4577 load = source_load(prev_cpu, idx);
4578 this_load = target_load(this_cpu, idx);
4581 * If sync wakeup then subtract the (maximum possible)
4582 * effect of the currently running task from the load
4583 * of the current CPU:
4586 tg = task_group(current);
4587 weight = current->se.load.weight;
4589 this_load += effective_load(tg, this_cpu, -weight, -weight);
4590 load += effective_load(tg, prev_cpu, 0, -weight);
4594 weight = p->se.load.weight;
4597 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4598 * due to the sync cause above having dropped this_load to 0, we'll
4599 * always have an imbalance, but there's really nothing you can do
4600 * about that, so that's good too.
4602 * Otherwise check if either cpus are near enough in load to allow this
4603 * task to be woken on this_cpu.
4605 this_eff_load = 100;
4606 this_eff_load *= capacity_of(prev_cpu);
4608 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4609 prev_eff_load *= capacity_of(this_cpu);
4611 if (this_load > 0) {
4612 this_eff_load *= this_load +
4613 effective_load(tg, this_cpu, weight, weight);
4615 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4618 balanced = this_eff_load <= prev_eff_load;
4620 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4625 schedstat_inc(sd, ttwu_move_affine);
4626 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4632 * find_idlest_group finds and returns the least busy CPU group within the
4635 static struct sched_group *
4636 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4637 int this_cpu, int sd_flag)
4639 struct sched_group *idlest = NULL, *group = sd->groups;
4640 unsigned long min_load = ULONG_MAX, this_load = 0;
4641 int load_idx = sd->forkexec_idx;
4642 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4644 if (sd_flag & SD_BALANCE_WAKE)
4645 load_idx = sd->wake_idx;
4648 unsigned long load, avg_load;
4652 /* Skip over this group if it has no CPUs allowed */
4653 if (!cpumask_intersects(sched_group_cpus(group),
4654 tsk_cpus_allowed(p)))
4657 local_group = cpumask_test_cpu(this_cpu,
4658 sched_group_cpus(group));
4660 /* Tally up the load of all CPUs in the group */
4663 for_each_cpu(i, sched_group_cpus(group)) {
4664 /* Bias balancing toward cpus of our domain */
4666 load = source_load(i, load_idx);
4668 load = target_load(i, load_idx);
4673 /* Adjust by relative CPU capacity of the group */
4674 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4677 this_load = avg_load;
4678 } else if (avg_load < min_load) {
4679 min_load = avg_load;
4682 } while (group = group->next, group != sd->groups);
4684 if (!idlest || 100*this_load < imbalance*min_load)
4690 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4693 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4695 unsigned long load, min_load = ULONG_MAX;
4696 unsigned int min_exit_latency = UINT_MAX;
4697 u64 latest_idle_timestamp = 0;
4698 int least_loaded_cpu = this_cpu;
4699 int shallowest_idle_cpu = -1;
4702 /* Traverse only the allowed CPUs */
4703 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4705 struct rq *rq = cpu_rq(i);
4706 struct cpuidle_state *idle = idle_get_state(rq);
4707 if (idle && idle->exit_latency < min_exit_latency) {
4709 * We give priority to a CPU whose idle state
4710 * has the smallest exit latency irrespective
4711 * of any idle timestamp.
4713 min_exit_latency = idle->exit_latency;
4714 latest_idle_timestamp = rq->idle_stamp;
4715 shallowest_idle_cpu = i;
4716 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4717 rq->idle_stamp > latest_idle_timestamp) {
4719 * If equal or no active idle state, then
4720 * the most recently idled CPU might have
4723 latest_idle_timestamp = rq->idle_stamp;
4724 shallowest_idle_cpu = i;
4726 } else if (shallowest_idle_cpu == -1) {
4727 load = weighted_cpuload(i);
4728 if (load < min_load || (load == min_load && i == this_cpu)) {
4730 least_loaded_cpu = i;
4735 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4739 * Try and locate an idle CPU in the sched_domain.
4741 static int select_idle_sibling(struct task_struct *p, int target)
4743 struct sched_domain *sd;
4744 struct sched_group *sg;
4745 int i = task_cpu(p);
4747 if (idle_cpu(target))
4751 * If the prevous cpu is cache affine and idle, don't be stupid.
4753 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4757 * Otherwise, iterate the domains and find an elegible idle cpu.
4759 sd = rcu_dereference(per_cpu(sd_llc, target));
4760 for_each_lower_domain(sd) {
4763 if (!cpumask_intersects(sched_group_cpus(sg),
4764 tsk_cpus_allowed(p)))
4767 for_each_cpu(i, sched_group_cpus(sg)) {
4768 if (i == target || !idle_cpu(i))
4772 target = cpumask_first_and(sched_group_cpus(sg),
4773 tsk_cpus_allowed(p));
4777 } while (sg != sd->groups);
4783 * get_cpu_usage returns the amount of capacity of a CPU that is used by CFS
4784 * tasks. The unit of the return value must be the one of capacity so we can
4785 * compare the usage with the capacity of the CPU that is available for CFS
4786 * task (ie cpu_capacity).
4787 * cfs.utilization_load_avg is the sum of running time of runnable tasks on a
4788 * CPU. It represents the amount of utilization of a CPU in the range
4789 * [0..SCHED_LOAD_SCALE]. The usage of a CPU can't be higher than the full
4790 * capacity of the CPU because it's about the running time on this CPU.
4791 * Nevertheless, cfs.utilization_load_avg can be higher than SCHED_LOAD_SCALE
4792 * because of unfortunate rounding in avg_period and running_load_avg or just
4793 * after migrating tasks until the average stabilizes with the new running
4794 * time. So we need to check that the usage stays into the range
4795 * [0..cpu_capacity_orig] and cap if necessary.
4796 * Without capping the usage, a group could be seen as overloaded (CPU0 usage
4797 * at 121% + CPU1 usage at 80%) whereas CPU1 has 20% of available capacity
4799 static int get_cpu_usage(int cpu)
4801 unsigned long usage = cpu_rq(cpu)->cfs.utilization_load_avg;
4802 unsigned long capacity = capacity_orig_of(cpu);
4804 if (usage >= SCHED_LOAD_SCALE)
4807 return (usage * capacity) >> SCHED_LOAD_SHIFT;
4811 * select_task_rq_fair: Select target runqueue for the waking task in domains
4812 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4813 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4815 * Balances load by selecting the idlest cpu in the idlest group, or under
4816 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4818 * Returns the target cpu number.
4820 * preempt must be disabled.
4823 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4825 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4826 int cpu = smp_processor_id();
4828 int want_affine = 0;
4829 int sync = wake_flags & WF_SYNC;
4831 if (sd_flag & SD_BALANCE_WAKE)
4832 want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4835 for_each_domain(cpu, tmp) {
4836 if (!(tmp->flags & SD_LOAD_BALANCE))
4840 * If both cpu and prev_cpu are part of this domain,
4841 * cpu is a valid SD_WAKE_AFFINE target.
4843 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4844 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4849 if (tmp->flags & sd_flag)
4853 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4856 if (sd_flag & SD_BALANCE_WAKE) {
4857 new_cpu = select_idle_sibling(p, prev_cpu);
4862 struct sched_group *group;
4865 if (!(sd->flags & sd_flag)) {
4870 group = find_idlest_group(sd, p, cpu, sd_flag);
4876 new_cpu = find_idlest_cpu(group, p, cpu);
4877 if (new_cpu == -1 || new_cpu == cpu) {
4878 /* Now try balancing at a lower domain level of cpu */
4883 /* Now try balancing at a lower domain level of new_cpu */
4885 weight = sd->span_weight;
4887 for_each_domain(cpu, tmp) {
4888 if (weight <= tmp->span_weight)
4890 if (tmp->flags & sd_flag)
4893 /* while loop will break here if sd == NULL */
4902 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4903 * cfs_rq_of(p) references at time of call are still valid and identify the
4904 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4905 * other assumptions, including the state of rq->lock, should be made.
4908 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4910 struct sched_entity *se = &p->se;
4911 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4914 * Load tracking: accumulate removed load so that it can be processed
4915 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4916 * to blocked load iff they have a positive decay-count. It can never
4917 * be negative here since on-rq tasks have decay-count == 0.
4919 if (se->avg.decay_count) {
4920 se->avg.decay_count = -__synchronize_entity_decay(se);
4921 atomic_long_add(se->avg.load_avg_contrib,
4922 &cfs_rq->removed_load);
4925 /* We have migrated, no longer consider this task hot */
4928 #endif /* CONFIG_SMP */
4930 static unsigned long
4931 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4933 unsigned long gran = sysctl_sched_wakeup_granularity;
4936 * Since its curr running now, convert the gran from real-time
4937 * to virtual-time in his units.
4939 * By using 'se' instead of 'curr' we penalize light tasks, so
4940 * they get preempted easier. That is, if 'se' < 'curr' then
4941 * the resulting gran will be larger, therefore penalizing the
4942 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4943 * be smaller, again penalizing the lighter task.
4945 * This is especially important for buddies when the leftmost
4946 * task is higher priority than the buddy.
4948 return calc_delta_fair(gran, se);
4952 * Should 'se' preempt 'curr'.
4966 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4968 s64 gran, vdiff = curr->vruntime - se->vruntime;
4973 gran = wakeup_gran(curr, se);
4980 static void set_last_buddy(struct sched_entity *se)
4982 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4985 for_each_sched_entity(se)
4986 cfs_rq_of(se)->last = se;
4989 static void set_next_buddy(struct sched_entity *se)
4991 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4994 for_each_sched_entity(se)
4995 cfs_rq_of(se)->next = se;
4998 static void set_skip_buddy(struct sched_entity *se)
5000 for_each_sched_entity(se)
5001 cfs_rq_of(se)->skip = se;
5005 * Preempt the current task with a newly woken task if needed:
5007 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5009 struct task_struct *curr = rq->curr;
5010 struct sched_entity *se = &curr->se, *pse = &p->se;
5011 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5012 int scale = cfs_rq->nr_running >= sched_nr_latency;
5013 int next_buddy_marked = 0;
5015 if (unlikely(se == pse))
5019 * This is possible from callers such as attach_tasks(), in which we
5020 * unconditionally check_prempt_curr() after an enqueue (which may have
5021 * lead to a throttle). This both saves work and prevents false
5022 * next-buddy nomination below.
5024 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5027 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5028 set_next_buddy(pse);
5029 next_buddy_marked = 1;
5033 * We can come here with TIF_NEED_RESCHED already set from new task
5036 * Note: this also catches the edge-case of curr being in a throttled
5037 * group (e.g. via set_curr_task), since update_curr() (in the
5038 * enqueue of curr) will have resulted in resched being set. This
5039 * prevents us from potentially nominating it as a false LAST_BUDDY
5042 if (test_tsk_need_resched(curr))
5045 /* Idle tasks are by definition preempted by non-idle tasks. */
5046 if (unlikely(curr->policy == SCHED_IDLE) &&
5047 likely(p->policy != SCHED_IDLE))
5051 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5052 * is driven by the tick):
5054 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5057 find_matching_se(&se, &pse);
5058 update_curr(cfs_rq_of(se));
5060 if (wakeup_preempt_entity(se, pse) == 1) {
5062 * Bias pick_next to pick the sched entity that is
5063 * triggering this preemption.
5065 if (!next_buddy_marked)
5066 set_next_buddy(pse);
5075 * Only set the backward buddy when the current task is still
5076 * on the rq. This can happen when a wakeup gets interleaved
5077 * with schedule on the ->pre_schedule() or idle_balance()
5078 * point, either of which can * drop the rq lock.
5080 * Also, during early boot the idle thread is in the fair class,
5081 * for obvious reasons its a bad idea to schedule back to it.
5083 if (unlikely(!se->on_rq || curr == rq->idle))
5086 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5090 static struct task_struct *
5091 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5093 struct cfs_rq *cfs_rq = &rq->cfs;
5094 struct sched_entity *se;
5095 struct task_struct *p;
5099 #ifdef CONFIG_FAIR_GROUP_SCHED
5100 if (!cfs_rq->nr_running)
5103 if (prev->sched_class != &fair_sched_class)
5107 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5108 * likely that a next task is from the same cgroup as the current.
5110 * Therefore attempt to avoid putting and setting the entire cgroup
5111 * hierarchy, only change the part that actually changes.
5115 struct sched_entity *curr = cfs_rq->curr;
5118 * Since we got here without doing put_prev_entity() we also
5119 * have to consider cfs_rq->curr. If it is still a runnable
5120 * entity, update_curr() will update its vruntime, otherwise
5121 * forget we've ever seen it.
5123 if (curr && curr->on_rq)
5124 update_curr(cfs_rq);
5129 * This call to check_cfs_rq_runtime() will do the throttle and
5130 * dequeue its entity in the parent(s). Therefore the 'simple'
5131 * nr_running test will indeed be correct.
5133 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5136 se = pick_next_entity(cfs_rq, curr);
5137 cfs_rq = group_cfs_rq(se);
5143 * Since we haven't yet done put_prev_entity and if the selected task
5144 * is a different task than we started out with, try and touch the
5145 * least amount of cfs_rqs.
5148 struct sched_entity *pse = &prev->se;
5150 while (!(cfs_rq = is_same_group(se, pse))) {
5151 int se_depth = se->depth;
5152 int pse_depth = pse->depth;
5154 if (se_depth <= pse_depth) {
5155 put_prev_entity(cfs_rq_of(pse), pse);
5156 pse = parent_entity(pse);
5158 if (se_depth >= pse_depth) {
5159 set_next_entity(cfs_rq_of(se), se);
5160 se = parent_entity(se);
5164 put_prev_entity(cfs_rq, pse);
5165 set_next_entity(cfs_rq, se);
5168 if (hrtick_enabled(rq))
5169 hrtick_start_fair(rq, p);
5176 if (!cfs_rq->nr_running)
5179 put_prev_task(rq, prev);
5182 se = pick_next_entity(cfs_rq, NULL);
5183 set_next_entity(cfs_rq, se);
5184 cfs_rq = group_cfs_rq(se);
5189 if (hrtick_enabled(rq))
5190 hrtick_start_fair(rq, p);
5195 new_tasks = idle_balance(rq);
5197 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5198 * possible for any higher priority task to appear. In that case we
5199 * must re-start the pick_next_entity() loop.
5211 * Account for a descheduled task:
5213 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5215 struct sched_entity *se = &prev->se;
5216 struct cfs_rq *cfs_rq;
5218 for_each_sched_entity(se) {
5219 cfs_rq = cfs_rq_of(se);
5220 put_prev_entity(cfs_rq, se);
5225 * sched_yield() is very simple
5227 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5229 static void yield_task_fair(struct rq *rq)
5231 struct task_struct *curr = rq->curr;
5232 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5233 struct sched_entity *se = &curr->se;
5236 * Are we the only task in the tree?
5238 if (unlikely(rq->nr_running == 1))
5241 clear_buddies(cfs_rq, se);
5243 if (curr->policy != SCHED_BATCH) {
5244 update_rq_clock(rq);
5246 * Update run-time statistics of the 'current'.
5248 update_curr(cfs_rq);
5250 * Tell update_rq_clock() that we've just updated,
5251 * so we don't do microscopic update in schedule()
5252 * and double the fastpath cost.
5254 rq_clock_skip_update(rq, true);
5260 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5262 struct sched_entity *se = &p->se;
5264 /* throttled hierarchies are not runnable */
5265 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5268 /* Tell the scheduler that we'd really like pse to run next. */
5271 yield_task_fair(rq);
5277 /**************************************************
5278 * Fair scheduling class load-balancing methods.
5282 * The purpose of load-balancing is to achieve the same basic fairness the
5283 * per-cpu scheduler provides, namely provide a proportional amount of compute
5284 * time to each task. This is expressed in the following equation:
5286 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5288 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5289 * W_i,0 is defined as:
5291 * W_i,0 = \Sum_j w_i,j (2)
5293 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5294 * is derived from the nice value as per prio_to_weight[].
5296 * The weight average is an exponential decay average of the instantaneous
5299 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5301 * C_i is the compute capacity of cpu i, typically it is the
5302 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5303 * can also include other factors [XXX].
5305 * To achieve this balance we define a measure of imbalance which follows
5306 * directly from (1):
5308 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5310 * We them move tasks around to minimize the imbalance. In the continuous
5311 * function space it is obvious this converges, in the discrete case we get
5312 * a few fun cases generally called infeasible weight scenarios.
5315 * - infeasible weights;
5316 * - local vs global optima in the discrete case. ]
5321 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5322 * for all i,j solution, we create a tree of cpus that follows the hardware
5323 * topology where each level pairs two lower groups (or better). This results
5324 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5325 * tree to only the first of the previous level and we decrease the frequency
5326 * of load-balance at each level inv. proportional to the number of cpus in
5332 * \Sum { --- * --- * 2^i } = O(n) (5)
5334 * `- size of each group
5335 * | | `- number of cpus doing load-balance
5337 * `- sum over all levels
5339 * Coupled with a limit on how many tasks we can migrate every balance pass,
5340 * this makes (5) the runtime complexity of the balancer.
5342 * An important property here is that each CPU is still (indirectly) connected
5343 * to every other cpu in at most O(log n) steps:
5345 * The adjacency matrix of the resulting graph is given by:
5348 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5351 * And you'll find that:
5353 * A^(log_2 n)_i,j != 0 for all i,j (7)
5355 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5356 * The task movement gives a factor of O(m), giving a convergence complexity
5359 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5364 * In order to avoid CPUs going idle while there's still work to do, new idle
5365 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5366 * tree itself instead of relying on other CPUs to bring it work.
5368 * This adds some complexity to both (5) and (8) but it reduces the total idle
5376 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5379 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5384 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5386 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5388 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5391 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5392 * rewrite all of this once again.]
5395 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5397 enum fbq_type { regular, remote, all };
5399 #define LBF_ALL_PINNED 0x01
5400 #define LBF_NEED_BREAK 0x02
5401 #define LBF_DST_PINNED 0x04
5402 #define LBF_SOME_PINNED 0x08
5405 struct sched_domain *sd;
5413 struct cpumask *dst_grpmask;
5415 enum cpu_idle_type idle;
5417 /* The set of CPUs under consideration for load-balancing */
5418 struct cpumask *cpus;
5423 unsigned int loop_break;
5424 unsigned int loop_max;
5426 enum fbq_type fbq_type;
5427 struct list_head tasks;
5431 * Is this task likely cache-hot:
5433 static int task_hot(struct task_struct *p, struct lb_env *env)
5437 lockdep_assert_held(&env->src_rq->lock);
5439 if (p->sched_class != &fair_sched_class)
5442 if (unlikely(p->policy == SCHED_IDLE))
5446 * Buddy candidates are cache hot:
5448 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5449 (&p->se == cfs_rq_of(&p->se)->next ||
5450 &p->se == cfs_rq_of(&p->se)->last))
5453 if (sysctl_sched_migration_cost == -1)
5455 if (sysctl_sched_migration_cost == 0)
5458 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5460 return delta < (s64)sysctl_sched_migration_cost;
5463 #ifdef CONFIG_NUMA_BALANCING
5464 /* Returns true if the destination node has incurred more faults */
5465 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5467 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5468 int src_nid, dst_nid;
5470 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
5471 !(env->sd->flags & SD_NUMA)) {
5475 src_nid = cpu_to_node(env->src_cpu);
5476 dst_nid = cpu_to_node(env->dst_cpu);
5478 if (src_nid == dst_nid)
5482 /* Task is already in the group's interleave set. */
5483 if (node_isset(src_nid, numa_group->active_nodes))
5486 /* Task is moving into the group's interleave set. */
5487 if (node_isset(dst_nid, numa_group->active_nodes))
5490 return group_faults(p, dst_nid) > group_faults(p, src_nid);
5493 /* Encourage migration to the preferred node. */
5494 if (dst_nid == p->numa_preferred_nid)
5497 return task_faults(p, dst_nid) > task_faults(p, src_nid);
5501 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5503 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5504 int src_nid, dst_nid;
5506 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5509 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5512 src_nid = cpu_to_node(env->src_cpu);
5513 dst_nid = cpu_to_node(env->dst_cpu);
5515 if (src_nid == dst_nid)
5519 /* Task is moving within/into the group's interleave set. */
5520 if (node_isset(dst_nid, numa_group->active_nodes))
5523 /* Task is moving out of the group's interleave set. */
5524 if (node_isset(src_nid, numa_group->active_nodes))
5527 return group_faults(p, dst_nid) < group_faults(p, src_nid);
5530 /* Migrating away from the preferred node is always bad. */
5531 if (src_nid == p->numa_preferred_nid)
5534 return task_faults(p, dst_nid) < task_faults(p, src_nid);
5538 static inline bool migrate_improves_locality(struct task_struct *p,
5544 static inline bool migrate_degrades_locality(struct task_struct *p,
5552 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5555 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5557 int tsk_cache_hot = 0;
5559 lockdep_assert_held(&env->src_rq->lock);
5562 * We do not migrate tasks that are:
5563 * 1) throttled_lb_pair, or
5564 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5565 * 3) running (obviously), or
5566 * 4) are cache-hot on their current CPU.
5568 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5571 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5574 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5576 env->flags |= LBF_SOME_PINNED;
5579 * Remember if this task can be migrated to any other cpu in
5580 * our sched_group. We may want to revisit it if we couldn't
5581 * meet load balance goals by pulling other tasks on src_cpu.
5583 * Also avoid computing new_dst_cpu if we have already computed
5584 * one in current iteration.
5586 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5589 /* Prevent to re-select dst_cpu via env's cpus */
5590 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5591 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5592 env->flags |= LBF_DST_PINNED;
5593 env->new_dst_cpu = cpu;
5601 /* Record that we found atleast one task that could run on dst_cpu */
5602 env->flags &= ~LBF_ALL_PINNED;
5604 if (task_running(env->src_rq, p)) {
5605 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5610 * Aggressive migration if:
5611 * 1) destination numa is preferred
5612 * 2) task is cache cold, or
5613 * 3) too many balance attempts have failed.
5615 tsk_cache_hot = task_hot(p, env);
5617 tsk_cache_hot = migrate_degrades_locality(p, env);
5619 if (migrate_improves_locality(p, env) || !tsk_cache_hot ||
5620 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5621 if (tsk_cache_hot) {
5622 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5623 schedstat_inc(p, se.statistics.nr_forced_migrations);
5628 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5633 * detach_task() -- detach the task for the migration specified in env
5635 static void detach_task(struct task_struct *p, struct lb_env *env)
5637 lockdep_assert_held(&env->src_rq->lock);
5639 deactivate_task(env->src_rq, p, 0);
5640 p->on_rq = TASK_ON_RQ_MIGRATING;
5641 set_task_cpu(p, env->dst_cpu);
5645 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5646 * part of active balancing operations within "domain".
5648 * Returns a task if successful and NULL otherwise.
5650 static struct task_struct *detach_one_task(struct lb_env *env)
5652 struct task_struct *p, *n;
5654 lockdep_assert_held(&env->src_rq->lock);
5656 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5657 if (!can_migrate_task(p, env))
5660 detach_task(p, env);
5663 * Right now, this is only the second place where
5664 * lb_gained[env->idle] is updated (other is detach_tasks)
5665 * so we can safely collect stats here rather than
5666 * inside detach_tasks().
5668 schedstat_inc(env->sd, lb_gained[env->idle]);
5674 static const unsigned int sched_nr_migrate_break = 32;
5677 * detach_tasks() -- tries to detach up to imbalance weighted load from
5678 * busiest_rq, as part of a balancing operation within domain "sd".
5680 * Returns number of detached tasks if successful and 0 otherwise.
5682 static int detach_tasks(struct lb_env *env)
5684 struct list_head *tasks = &env->src_rq->cfs_tasks;
5685 struct task_struct *p;
5689 lockdep_assert_held(&env->src_rq->lock);
5691 if (env->imbalance <= 0)
5694 while (!list_empty(tasks)) {
5695 p = list_first_entry(tasks, struct task_struct, se.group_node);
5698 /* We've more or less seen every task there is, call it quits */
5699 if (env->loop > env->loop_max)
5702 /* take a breather every nr_migrate tasks */
5703 if (env->loop > env->loop_break) {
5704 env->loop_break += sched_nr_migrate_break;
5705 env->flags |= LBF_NEED_BREAK;
5709 if (!can_migrate_task(p, env))
5712 load = task_h_load(p);
5714 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5717 if ((load / 2) > env->imbalance)
5720 detach_task(p, env);
5721 list_add(&p->se.group_node, &env->tasks);
5724 env->imbalance -= load;
5726 #ifdef CONFIG_PREEMPT
5728 * NEWIDLE balancing is a source of latency, so preemptible
5729 * kernels will stop after the first task is detached to minimize
5730 * the critical section.
5732 if (env->idle == CPU_NEWLY_IDLE)
5737 * We only want to steal up to the prescribed amount of
5740 if (env->imbalance <= 0)
5745 list_move_tail(&p->se.group_node, tasks);
5749 * Right now, this is one of only two places we collect this stat
5750 * so we can safely collect detach_one_task() stats here rather
5751 * than inside detach_one_task().
5753 schedstat_add(env->sd, lb_gained[env->idle], detached);
5759 * attach_task() -- attach the task detached by detach_task() to its new rq.
5761 static void attach_task(struct rq *rq, struct task_struct *p)
5763 lockdep_assert_held(&rq->lock);
5765 BUG_ON(task_rq(p) != rq);
5766 p->on_rq = TASK_ON_RQ_QUEUED;
5767 activate_task(rq, p, 0);
5768 check_preempt_curr(rq, p, 0);
5772 * attach_one_task() -- attaches the task returned from detach_one_task() to
5775 static void attach_one_task(struct rq *rq, struct task_struct *p)
5777 raw_spin_lock(&rq->lock);
5779 raw_spin_unlock(&rq->lock);
5783 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5786 static void attach_tasks(struct lb_env *env)
5788 struct list_head *tasks = &env->tasks;
5789 struct task_struct *p;
5791 raw_spin_lock(&env->dst_rq->lock);
5793 while (!list_empty(tasks)) {
5794 p = list_first_entry(tasks, struct task_struct, se.group_node);
5795 list_del_init(&p->se.group_node);
5797 attach_task(env->dst_rq, p);
5800 raw_spin_unlock(&env->dst_rq->lock);
5803 #ifdef CONFIG_FAIR_GROUP_SCHED
5805 * update tg->load_weight by folding this cpu's load_avg
5807 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5809 struct sched_entity *se = tg->se[cpu];
5810 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5812 /* throttled entities do not contribute to load */
5813 if (throttled_hierarchy(cfs_rq))
5816 update_cfs_rq_blocked_load(cfs_rq, 1);
5819 update_entity_load_avg(se, 1);
5821 * We pivot on our runnable average having decayed to zero for
5822 * list removal. This generally implies that all our children
5823 * have also been removed (modulo rounding error or bandwidth
5824 * control); however, such cases are rare and we can fix these
5827 * TODO: fix up out-of-order children on enqueue.
5829 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5830 list_del_leaf_cfs_rq(cfs_rq);
5832 struct rq *rq = rq_of(cfs_rq);
5833 update_rq_runnable_avg(rq, rq->nr_running);
5837 static void update_blocked_averages(int cpu)
5839 struct rq *rq = cpu_rq(cpu);
5840 struct cfs_rq *cfs_rq;
5841 unsigned long flags;
5843 raw_spin_lock_irqsave(&rq->lock, flags);
5844 update_rq_clock(rq);
5846 * Iterates the task_group tree in a bottom up fashion, see
5847 * list_add_leaf_cfs_rq() for details.
5849 for_each_leaf_cfs_rq(rq, cfs_rq) {
5851 * Note: We may want to consider periodically releasing
5852 * rq->lock about these updates so that creating many task
5853 * groups does not result in continually extending hold time.
5855 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5858 raw_spin_unlock_irqrestore(&rq->lock, flags);
5862 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5863 * This needs to be done in a top-down fashion because the load of a child
5864 * group is a fraction of its parents load.
5866 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5868 struct rq *rq = rq_of(cfs_rq);
5869 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5870 unsigned long now = jiffies;
5873 if (cfs_rq->last_h_load_update == now)
5876 cfs_rq->h_load_next = NULL;
5877 for_each_sched_entity(se) {
5878 cfs_rq = cfs_rq_of(se);
5879 cfs_rq->h_load_next = se;
5880 if (cfs_rq->last_h_load_update == now)
5885 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5886 cfs_rq->last_h_load_update = now;
5889 while ((se = cfs_rq->h_load_next) != NULL) {
5890 load = cfs_rq->h_load;
5891 load = div64_ul(load * se->avg.load_avg_contrib,
5892 cfs_rq->runnable_load_avg + 1);
5893 cfs_rq = group_cfs_rq(se);
5894 cfs_rq->h_load = load;
5895 cfs_rq->last_h_load_update = now;
5899 static unsigned long task_h_load(struct task_struct *p)
5901 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5903 update_cfs_rq_h_load(cfs_rq);
5904 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5905 cfs_rq->runnable_load_avg + 1);
5908 static inline void update_blocked_averages(int cpu)
5912 static unsigned long task_h_load(struct task_struct *p)
5914 return p->se.avg.load_avg_contrib;
5918 /********** Helpers for find_busiest_group ************************/
5927 * sg_lb_stats - stats of a sched_group required for load_balancing
5929 struct sg_lb_stats {
5930 unsigned long avg_load; /*Avg load across the CPUs of the group */
5931 unsigned long group_load; /* Total load over the CPUs of the group */
5932 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5933 unsigned long load_per_task;
5934 unsigned long group_capacity;
5935 unsigned long group_usage; /* Total usage of the group */
5936 unsigned int sum_nr_running; /* Nr tasks running in the group */
5937 unsigned int idle_cpus;
5938 unsigned int group_weight;
5939 enum group_type group_type;
5940 int group_no_capacity;
5941 #ifdef CONFIG_NUMA_BALANCING
5942 unsigned int nr_numa_running;
5943 unsigned int nr_preferred_running;
5948 * sd_lb_stats - Structure to store the statistics of a sched_domain
5949 * during load balancing.
5951 struct sd_lb_stats {
5952 struct sched_group *busiest; /* Busiest group in this sd */
5953 struct sched_group *local; /* Local group in this sd */
5954 unsigned long total_load; /* Total load of all groups in sd */
5955 unsigned long total_capacity; /* Total capacity of all groups in sd */
5956 unsigned long avg_load; /* Average load across all groups in sd */
5958 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5959 struct sg_lb_stats local_stat; /* Statistics of the local group */
5962 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5965 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5966 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5967 * We must however clear busiest_stat::avg_load because
5968 * update_sd_pick_busiest() reads this before assignment.
5970 *sds = (struct sd_lb_stats){
5974 .total_capacity = 0UL,
5977 .sum_nr_running = 0,
5978 .group_type = group_other,
5984 * get_sd_load_idx - Obtain the load index for a given sched domain.
5985 * @sd: The sched_domain whose load_idx is to be obtained.
5986 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5988 * Return: The load index.
5990 static inline int get_sd_load_idx(struct sched_domain *sd,
5991 enum cpu_idle_type idle)
5997 load_idx = sd->busy_idx;
6000 case CPU_NEWLY_IDLE:
6001 load_idx = sd->newidle_idx;
6004 load_idx = sd->idle_idx;
6011 static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu)
6013 if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1))
6014 return sd->smt_gain / sd->span_weight;
6016 return SCHED_CAPACITY_SCALE;
6019 unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
6021 return default_scale_cpu_capacity(sd, cpu);
6024 static unsigned long scale_rt_capacity(int cpu)
6026 struct rq *rq = cpu_rq(cpu);
6027 u64 total, used, age_stamp, avg;
6031 * Since we're reading these variables without serialization make sure
6032 * we read them once before doing sanity checks on them.
6034 age_stamp = ACCESS_ONCE(rq->age_stamp);
6035 avg = ACCESS_ONCE(rq->rt_avg);
6036 delta = __rq_clock_broken(rq) - age_stamp;
6038 if (unlikely(delta < 0))
6041 total = sched_avg_period() + delta;
6043 used = div_u64(avg, total);
6045 if (likely(used < SCHED_CAPACITY_SCALE))
6046 return SCHED_CAPACITY_SCALE - used;
6051 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6053 unsigned long capacity = SCHED_CAPACITY_SCALE;
6054 struct sched_group *sdg = sd->groups;
6056 if (sched_feat(ARCH_CAPACITY))
6057 capacity *= arch_scale_cpu_capacity(sd, cpu);
6059 capacity *= default_scale_cpu_capacity(sd, cpu);
6061 capacity >>= SCHED_CAPACITY_SHIFT;
6063 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6065 capacity *= scale_rt_capacity(cpu);
6066 capacity >>= SCHED_CAPACITY_SHIFT;
6071 cpu_rq(cpu)->cpu_capacity = capacity;
6072 sdg->sgc->capacity = capacity;
6075 void update_group_capacity(struct sched_domain *sd, int cpu)
6077 struct sched_domain *child = sd->child;
6078 struct sched_group *group, *sdg = sd->groups;
6079 unsigned long capacity;
6080 unsigned long interval;
6082 interval = msecs_to_jiffies(sd->balance_interval);
6083 interval = clamp(interval, 1UL, max_load_balance_interval);
6084 sdg->sgc->next_update = jiffies + interval;
6087 update_cpu_capacity(sd, cpu);
6093 if (child->flags & SD_OVERLAP) {
6095 * SD_OVERLAP domains cannot assume that child groups
6096 * span the current group.
6099 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6100 struct sched_group_capacity *sgc;
6101 struct rq *rq = cpu_rq(cpu);
6104 * build_sched_domains() -> init_sched_groups_capacity()
6105 * gets here before we've attached the domains to the
6108 * Use capacity_of(), which is set irrespective of domains
6109 * in update_cpu_capacity().
6111 * This avoids capacity from being 0 and
6112 * causing divide-by-zero issues on boot.
6114 if (unlikely(!rq->sd)) {
6115 capacity += capacity_of(cpu);
6119 sgc = rq->sd->groups->sgc;
6120 capacity += sgc->capacity;
6124 * !SD_OVERLAP domains can assume that child groups
6125 * span the current group.
6128 group = child->groups;
6130 capacity += group->sgc->capacity;
6131 group = group->next;
6132 } while (group != child->groups);
6135 sdg->sgc->capacity = capacity;
6139 * Check whether the capacity of the rq has been noticeably reduced by side
6140 * activity. The imbalance_pct is used for the threshold.
6141 * Return true is the capacity is reduced
6144 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6146 return ((rq->cpu_capacity * sd->imbalance_pct) <
6147 (rq->cpu_capacity_orig * 100));
6151 * Group imbalance indicates (and tries to solve) the problem where balancing
6152 * groups is inadequate due to tsk_cpus_allowed() constraints.
6154 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6155 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6158 * { 0 1 2 3 } { 4 5 6 7 }
6161 * If we were to balance group-wise we'd place two tasks in the first group and
6162 * two tasks in the second group. Clearly this is undesired as it will overload
6163 * cpu 3 and leave one of the cpus in the second group unused.
6165 * The current solution to this issue is detecting the skew in the first group
6166 * by noticing the lower domain failed to reach balance and had difficulty
6167 * moving tasks due to affinity constraints.
6169 * When this is so detected; this group becomes a candidate for busiest; see
6170 * update_sd_pick_busiest(). And calculate_imbalance() and
6171 * find_busiest_group() avoid some of the usual balance conditions to allow it
6172 * to create an effective group imbalance.
6174 * This is a somewhat tricky proposition since the next run might not find the
6175 * group imbalance and decide the groups need to be balanced again. A most
6176 * subtle and fragile situation.
6179 static inline int sg_imbalanced(struct sched_group *group)
6181 return group->sgc->imbalance;
6185 * group_has_capacity returns true if the group has spare capacity that could
6186 * be used by some tasks.
6187 * We consider that a group has spare capacity if the * number of task is
6188 * smaller than the number of CPUs or if the usage is lower than the available
6189 * capacity for CFS tasks.
6190 * For the latter, we use a threshold to stabilize the state, to take into
6191 * account the variance of the tasks' load and to return true if the available
6192 * capacity in meaningful for the load balancer.
6193 * As an example, an available capacity of 1% can appear but it doesn't make
6194 * any benefit for the load balance.
6197 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6199 if (sgs->sum_nr_running < sgs->group_weight)
6202 if ((sgs->group_capacity * 100) >
6203 (sgs->group_usage * env->sd->imbalance_pct))
6210 * group_is_overloaded returns true if the group has more tasks than it can
6212 * group_is_overloaded is not equals to !group_has_capacity because a group
6213 * with the exact right number of tasks, has no more spare capacity but is not
6214 * overloaded so both group_has_capacity and group_is_overloaded return
6218 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6220 if (sgs->sum_nr_running <= sgs->group_weight)
6223 if ((sgs->group_capacity * 100) <
6224 (sgs->group_usage * env->sd->imbalance_pct))
6230 static enum group_type group_classify(struct lb_env *env,
6231 struct sched_group *group,
6232 struct sg_lb_stats *sgs)
6234 if (sgs->group_no_capacity)
6235 return group_overloaded;
6237 if (sg_imbalanced(group))
6238 return group_imbalanced;
6244 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6245 * @env: The load balancing environment.
6246 * @group: sched_group whose statistics are to be updated.
6247 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6248 * @local_group: Does group contain this_cpu.
6249 * @sgs: variable to hold the statistics for this group.
6250 * @overload: Indicate more than one runnable task for any CPU.
6252 static inline void update_sg_lb_stats(struct lb_env *env,
6253 struct sched_group *group, int load_idx,
6254 int local_group, struct sg_lb_stats *sgs,
6260 memset(sgs, 0, sizeof(*sgs));
6262 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6263 struct rq *rq = cpu_rq(i);
6265 /* Bias balancing toward cpus of our domain */
6267 load = target_load(i, load_idx);
6269 load = source_load(i, load_idx);
6271 sgs->group_load += load;
6272 sgs->group_usage += get_cpu_usage(i);
6273 sgs->sum_nr_running += rq->cfs.h_nr_running;
6275 if (rq->nr_running > 1)
6278 #ifdef CONFIG_NUMA_BALANCING
6279 sgs->nr_numa_running += rq->nr_numa_running;
6280 sgs->nr_preferred_running += rq->nr_preferred_running;
6282 sgs->sum_weighted_load += weighted_cpuload(i);
6287 /* Adjust by relative CPU capacity of the group */
6288 sgs->group_capacity = group->sgc->capacity;
6289 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6291 if (sgs->sum_nr_running)
6292 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6294 sgs->group_weight = group->group_weight;
6296 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6297 sgs->group_type = group_classify(env, group, sgs);
6301 * update_sd_pick_busiest - return 1 on busiest group
6302 * @env: The load balancing environment.
6303 * @sds: sched_domain statistics
6304 * @sg: sched_group candidate to be checked for being the busiest
6305 * @sgs: sched_group statistics
6307 * Determine if @sg is a busier group than the previously selected
6310 * Return: %true if @sg is a busier group than the previously selected
6311 * busiest group. %false otherwise.
6313 static bool update_sd_pick_busiest(struct lb_env *env,
6314 struct sd_lb_stats *sds,
6315 struct sched_group *sg,
6316 struct sg_lb_stats *sgs)
6318 struct sg_lb_stats *busiest = &sds->busiest_stat;
6320 if (sgs->group_type > busiest->group_type)
6323 if (sgs->group_type < busiest->group_type)
6326 if (sgs->avg_load <= busiest->avg_load)
6329 /* This is the busiest node in its class. */
6330 if (!(env->sd->flags & SD_ASYM_PACKING))
6334 * ASYM_PACKING needs to move all the work to the lowest
6335 * numbered CPUs in the group, therefore mark all groups
6336 * higher than ourself as busy.
6338 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6342 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6349 #ifdef CONFIG_NUMA_BALANCING
6350 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6352 if (sgs->sum_nr_running > sgs->nr_numa_running)
6354 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6359 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6361 if (rq->nr_running > rq->nr_numa_running)
6363 if (rq->nr_running > rq->nr_preferred_running)
6368 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6373 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6377 #endif /* CONFIG_NUMA_BALANCING */
6380 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6381 * @env: The load balancing environment.
6382 * @sds: variable to hold the statistics for this sched_domain.
6384 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6386 struct sched_domain *child = env->sd->child;
6387 struct sched_group *sg = env->sd->groups;
6388 struct sg_lb_stats tmp_sgs;
6389 int load_idx, prefer_sibling = 0;
6390 bool overload = false;
6392 if (child && child->flags & SD_PREFER_SIBLING)
6395 load_idx = get_sd_load_idx(env->sd, env->idle);
6398 struct sg_lb_stats *sgs = &tmp_sgs;
6401 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6404 sgs = &sds->local_stat;
6406 if (env->idle != CPU_NEWLY_IDLE ||
6407 time_after_eq(jiffies, sg->sgc->next_update))
6408 update_group_capacity(env->sd, env->dst_cpu);
6411 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6418 * In case the child domain prefers tasks go to siblings
6419 * first, lower the sg capacity so that we'll try
6420 * and move all the excess tasks away. We lower the capacity
6421 * of a group only if the local group has the capacity to fit
6422 * these excess tasks. The extra check prevents the case where
6423 * you always pull from the heaviest group when it is already
6424 * under-utilized (possible with a large weight task outweighs
6425 * the tasks on the system).
6427 if (prefer_sibling && sds->local &&
6428 group_has_capacity(env, &sds->local_stat) &&
6429 (sgs->sum_nr_running > 1)) {
6430 sgs->group_no_capacity = 1;
6431 sgs->group_type = group_overloaded;
6434 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6436 sds->busiest_stat = *sgs;
6440 /* Now, start updating sd_lb_stats */
6441 sds->total_load += sgs->group_load;
6442 sds->total_capacity += sgs->group_capacity;
6445 } while (sg != env->sd->groups);
6447 if (env->sd->flags & SD_NUMA)
6448 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6450 if (!env->sd->parent) {
6451 /* update overload indicator if we are at root domain */
6452 if (env->dst_rq->rd->overload != overload)
6453 env->dst_rq->rd->overload = overload;
6459 * check_asym_packing - Check to see if the group is packed into the
6462 * This is primarily intended to used at the sibling level. Some
6463 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6464 * case of POWER7, it can move to lower SMT modes only when higher
6465 * threads are idle. When in lower SMT modes, the threads will
6466 * perform better since they share less core resources. Hence when we
6467 * have idle threads, we want them to be the higher ones.
6469 * This packing function is run on idle threads. It checks to see if
6470 * the busiest CPU in this domain (core in the P7 case) has a higher
6471 * CPU number than the packing function is being run on. Here we are
6472 * assuming lower CPU number will be equivalent to lower a SMT thread
6475 * Return: 1 when packing is required and a task should be moved to
6476 * this CPU. The amount of the imbalance is returned in *imbalance.
6478 * @env: The load balancing environment.
6479 * @sds: Statistics of the sched_domain which is to be packed
6481 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6485 if (!(env->sd->flags & SD_ASYM_PACKING))
6491 busiest_cpu = group_first_cpu(sds->busiest);
6492 if (env->dst_cpu > busiest_cpu)
6495 env->imbalance = DIV_ROUND_CLOSEST(
6496 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6497 SCHED_CAPACITY_SCALE);
6503 * fix_small_imbalance - Calculate the minor imbalance that exists
6504 * amongst the groups of a sched_domain, during
6506 * @env: The load balancing environment.
6507 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6510 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6512 unsigned long tmp, capa_now = 0, capa_move = 0;
6513 unsigned int imbn = 2;
6514 unsigned long scaled_busy_load_per_task;
6515 struct sg_lb_stats *local, *busiest;
6517 local = &sds->local_stat;
6518 busiest = &sds->busiest_stat;
6520 if (!local->sum_nr_running)
6521 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6522 else if (busiest->load_per_task > local->load_per_task)
6525 scaled_busy_load_per_task =
6526 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6527 busiest->group_capacity;
6529 if (busiest->avg_load + scaled_busy_load_per_task >=
6530 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6531 env->imbalance = busiest->load_per_task;
6536 * OK, we don't have enough imbalance to justify moving tasks,
6537 * however we may be able to increase total CPU capacity used by
6541 capa_now += busiest->group_capacity *
6542 min(busiest->load_per_task, busiest->avg_load);
6543 capa_now += local->group_capacity *
6544 min(local->load_per_task, local->avg_load);
6545 capa_now /= SCHED_CAPACITY_SCALE;
6547 /* Amount of load we'd subtract */
6548 if (busiest->avg_load > scaled_busy_load_per_task) {
6549 capa_move += busiest->group_capacity *
6550 min(busiest->load_per_task,
6551 busiest->avg_load - scaled_busy_load_per_task);
6554 /* Amount of load we'd add */
6555 if (busiest->avg_load * busiest->group_capacity <
6556 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6557 tmp = (busiest->avg_load * busiest->group_capacity) /
6558 local->group_capacity;
6560 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6561 local->group_capacity;
6563 capa_move += local->group_capacity *
6564 min(local->load_per_task, local->avg_load + tmp);
6565 capa_move /= SCHED_CAPACITY_SCALE;
6567 /* Move if we gain throughput */
6568 if (capa_move > capa_now)
6569 env->imbalance = busiest->load_per_task;
6573 * calculate_imbalance - Calculate the amount of imbalance present within the
6574 * groups of a given sched_domain during load balance.
6575 * @env: load balance environment
6576 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6578 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6580 unsigned long max_pull, load_above_capacity = ~0UL;
6581 struct sg_lb_stats *local, *busiest;
6583 local = &sds->local_stat;
6584 busiest = &sds->busiest_stat;
6586 if (busiest->group_type == group_imbalanced) {
6588 * In the group_imb case we cannot rely on group-wide averages
6589 * to ensure cpu-load equilibrium, look at wider averages. XXX
6591 busiest->load_per_task =
6592 min(busiest->load_per_task, sds->avg_load);
6596 * In the presence of smp nice balancing, certain scenarios can have
6597 * max load less than avg load(as we skip the groups at or below
6598 * its cpu_capacity, while calculating max_load..)
6600 if (busiest->avg_load <= sds->avg_load ||
6601 local->avg_load >= sds->avg_load) {
6603 return fix_small_imbalance(env, sds);
6607 * If there aren't any idle cpus, avoid creating some.
6609 if (busiest->group_type == group_overloaded &&
6610 local->group_type == group_overloaded) {
6611 load_above_capacity = busiest->sum_nr_running *
6613 if (load_above_capacity > busiest->group_capacity)
6614 load_above_capacity -= busiest->group_capacity;
6616 load_above_capacity = ~0UL;
6620 * We're trying to get all the cpus to the average_load, so we don't
6621 * want to push ourselves above the average load, nor do we wish to
6622 * reduce the max loaded cpu below the average load. At the same time,
6623 * we also don't want to reduce the group load below the group capacity
6624 * (so that we can implement power-savings policies etc). Thus we look
6625 * for the minimum possible imbalance.
6627 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6629 /* How much load to actually move to equalise the imbalance */
6630 env->imbalance = min(
6631 max_pull * busiest->group_capacity,
6632 (sds->avg_load - local->avg_load) * local->group_capacity
6633 ) / SCHED_CAPACITY_SCALE;
6636 * if *imbalance is less than the average load per runnable task
6637 * there is no guarantee that any tasks will be moved so we'll have
6638 * a think about bumping its value to force at least one task to be
6641 if (env->imbalance < busiest->load_per_task)
6642 return fix_small_imbalance(env, sds);
6645 /******* find_busiest_group() helpers end here *********************/
6648 * find_busiest_group - Returns the busiest group within the sched_domain
6649 * if there is an imbalance. If there isn't an imbalance, and
6650 * the user has opted for power-savings, it returns a group whose
6651 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6652 * such a group exists.
6654 * Also calculates the amount of weighted load which should be moved
6655 * to restore balance.
6657 * @env: The load balancing environment.
6659 * Return: - The busiest group if imbalance exists.
6660 * - If no imbalance and user has opted for power-savings balance,
6661 * return the least loaded group whose CPUs can be
6662 * put to idle by rebalancing its tasks onto our group.
6664 static struct sched_group *find_busiest_group(struct lb_env *env)
6666 struct sg_lb_stats *local, *busiest;
6667 struct sd_lb_stats sds;
6669 init_sd_lb_stats(&sds);
6672 * Compute the various statistics relavent for load balancing at
6675 update_sd_lb_stats(env, &sds);
6676 local = &sds.local_stat;
6677 busiest = &sds.busiest_stat;
6679 /* ASYM feature bypasses nice load balance check */
6680 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6681 check_asym_packing(env, &sds))
6684 /* There is no busy sibling group to pull tasks from */
6685 if (!sds.busiest || busiest->sum_nr_running == 0)
6688 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6689 / sds.total_capacity;
6692 * If the busiest group is imbalanced the below checks don't
6693 * work because they assume all things are equal, which typically
6694 * isn't true due to cpus_allowed constraints and the like.
6696 if (busiest->group_type == group_imbalanced)
6699 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6700 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6701 busiest->group_no_capacity)
6705 * If the local group is busier than the selected busiest group
6706 * don't try and pull any tasks.
6708 if (local->avg_load >= busiest->avg_load)
6712 * Don't pull any tasks if this group is already above the domain
6715 if (local->avg_load >= sds.avg_load)
6718 if (env->idle == CPU_IDLE) {
6720 * This cpu is idle. If the busiest group is not overloaded
6721 * and there is no imbalance between this and busiest group
6722 * wrt idle cpus, it is balanced. The imbalance becomes
6723 * significant if the diff is greater than 1 otherwise we
6724 * might end up to just move the imbalance on another group
6726 if ((busiest->group_type != group_overloaded) &&
6727 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6731 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6732 * imbalance_pct to be conservative.
6734 if (100 * busiest->avg_load <=
6735 env->sd->imbalance_pct * local->avg_load)
6740 /* Looks like there is an imbalance. Compute it */
6741 calculate_imbalance(env, &sds);
6750 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6752 static struct rq *find_busiest_queue(struct lb_env *env,
6753 struct sched_group *group)
6755 struct rq *busiest = NULL, *rq;
6756 unsigned long busiest_load = 0, busiest_capacity = 1;
6759 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6760 unsigned long capacity, wl;
6764 rt = fbq_classify_rq(rq);
6767 * We classify groups/runqueues into three groups:
6768 * - regular: there are !numa tasks
6769 * - remote: there are numa tasks that run on the 'wrong' node
6770 * - all: there is no distinction
6772 * In order to avoid migrating ideally placed numa tasks,
6773 * ignore those when there's better options.
6775 * If we ignore the actual busiest queue to migrate another
6776 * task, the next balance pass can still reduce the busiest
6777 * queue by moving tasks around inside the node.
6779 * If we cannot move enough load due to this classification
6780 * the next pass will adjust the group classification and
6781 * allow migration of more tasks.
6783 * Both cases only affect the total convergence complexity.
6785 if (rt > env->fbq_type)
6788 capacity = capacity_of(i);
6790 wl = weighted_cpuload(i);
6793 * When comparing with imbalance, use weighted_cpuload()
6794 * which is not scaled with the cpu capacity.
6797 if (rq->nr_running == 1 && wl > env->imbalance &&
6798 !check_cpu_capacity(rq, env->sd))
6802 * For the load comparisons with the other cpu's, consider
6803 * the weighted_cpuload() scaled with the cpu capacity, so
6804 * that the load can be moved away from the cpu that is
6805 * potentially running at a lower capacity.
6807 * Thus we're looking for max(wl_i / capacity_i), crosswise
6808 * multiplication to rid ourselves of the division works out
6809 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6810 * our previous maximum.
6812 if (wl * busiest_capacity > busiest_load * capacity) {
6814 busiest_capacity = capacity;
6823 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6824 * so long as it is large enough.
6826 #define MAX_PINNED_INTERVAL 512
6828 /* Working cpumask for load_balance and load_balance_newidle. */
6829 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6831 static int need_active_balance(struct lb_env *env)
6833 struct sched_domain *sd = env->sd;
6835 if (env->idle == CPU_NEWLY_IDLE) {
6838 * ASYM_PACKING needs to force migrate tasks from busy but
6839 * higher numbered CPUs in order to pack all tasks in the
6840 * lowest numbered CPUs.
6842 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6847 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
6848 * It's worth migrating the task if the src_cpu's capacity is reduced
6849 * because of other sched_class or IRQs if more capacity stays
6850 * available on dst_cpu.
6852 if ((env->idle != CPU_NOT_IDLE) &&
6853 (env->src_rq->cfs.h_nr_running == 1)) {
6854 if ((check_cpu_capacity(env->src_rq, sd)) &&
6855 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
6859 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6862 static int active_load_balance_cpu_stop(void *data);
6864 static int should_we_balance(struct lb_env *env)
6866 struct sched_group *sg = env->sd->groups;
6867 struct cpumask *sg_cpus, *sg_mask;
6868 int cpu, balance_cpu = -1;
6871 * In the newly idle case, we will allow all the cpu's
6872 * to do the newly idle load balance.
6874 if (env->idle == CPU_NEWLY_IDLE)
6877 sg_cpus = sched_group_cpus(sg);
6878 sg_mask = sched_group_mask(sg);
6879 /* Try to find first idle cpu */
6880 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6881 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6888 if (balance_cpu == -1)
6889 balance_cpu = group_balance_cpu(sg);
6892 * First idle cpu or the first cpu(busiest) in this sched group
6893 * is eligible for doing load balancing at this and above domains.
6895 return balance_cpu == env->dst_cpu;
6899 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6900 * tasks if there is an imbalance.
6902 static int load_balance(int this_cpu, struct rq *this_rq,
6903 struct sched_domain *sd, enum cpu_idle_type idle,
6904 int *continue_balancing)
6906 int ld_moved, cur_ld_moved, active_balance = 0;
6907 struct sched_domain *sd_parent = sd->parent;
6908 struct sched_group *group;
6910 unsigned long flags;
6911 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
6913 struct lb_env env = {
6915 .dst_cpu = this_cpu,
6917 .dst_grpmask = sched_group_cpus(sd->groups),
6919 .loop_break = sched_nr_migrate_break,
6922 .tasks = LIST_HEAD_INIT(env.tasks),
6926 * For NEWLY_IDLE load_balancing, we don't need to consider
6927 * other cpus in our group
6929 if (idle == CPU_NEWLY_IDLE)
6930 env.dst_grpmask = NULL;
6932 cpumask_copy(cpus, cpu_active_mask);
6934 schedstat_inc(sd, lb_count[idle]);
6937 if (!should_we_balance(&env)) {
6938 *continue_balancing = 0;
6942 group = find_busiest_group(&env);
6944 schedstat_inc(sd, lb_nobusyg[idle]);
6948 busiest = find_busiest_queue(&env, group);
6950 schedstat_inc(sd, lb_nobusyq[idle]);
6954 BUG_ON(busiest == env.dst_rq);
6956 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6958 env.src_cpu = busiest->cpu;
6959 env.src_rq = busiest;
6962 if (busiest->nr_running > 1) {
6964 * Attempt to move tasks. If find_busiest_group has found
6965 * an imbalance but busiest->nr_running <= 1, the group is
6966 * still unbalanced. ld_moved simply stays zero, so it is
6967 * correctly treated as an imbalance.
6969 env.flags |= LBF_ALL_PINNED;
6970 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6973 raw_spin_lock_irqsave(&busiest->lock, flags);
6976 * cur_ld_moved - load moved in current iteration
6977 * ld_moved - cumulative load moved across iterations
6979 cur_ld_moved = detach_tasks(&env);
6982 * We've detached some tasks from busiest_rq. Every
6983 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6984 * unlock busiest->lock, and we are able to be sure
6985 * that nobody can manipulate the tasks in parallel.
6986 * See task_rq_lock() family for the details.
6989 raw_spin_unlock(&busiest->lock);
6993 ld_moved += cur_ld_moved;
6996 local_irq_restore(flags);
6998 if (env.flags & LBF_NEED_BREAK) {
6999 env.flags &= ~LBF_NEED_BREAK;
7004 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7005 * us and move them to an alternate dst_cpu in our sched_group
7006 * where they can run. The upper limit on how many times we
7007 * iterate on same src_cpu is dependent on number of cpus in our
7010 * This changes load balance semantics a bit on who can move
7011 * load to a given_cpu. In addition to the given_cpu itself
7012 * (or a ilb_cpu acting on its behalf where given_cpu is
7013 * nohz-idle), we now have balance_cpu in a position to move
7014 * load to given_cpu. In rare situations, this may cause
7015 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7016 * _independently_ and at _same_ time to move some load to
7017 * given_cpu) causing exceess load to be moved to given_cpu.
7018 * This however should not happen so much in practice and
7019 * moreover subsequent load balance cycles should correct the
7020 * excess load moved.
7022 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7024 /* Prevent to re-select dst_cpu via env's cpus */
7025 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7027 env.dst_rq = cpu_rq(env.new_dst_cpu);
7028 env.dst_cpu = env.new_dst_cpu;
7029 env.flags &= ~LBF_DST_PINNED;
7031 env.loop_break = sched_nr_migrate_break;
7034 * Go back to "more_balance" rather than "redo" since we
7035 * need to continue with same src_cpu.
7041 * We failed to reach balance because of affinity.
7044 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7046 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7047 *group_imbalance = 1;
7050 /* All tasks on this runqueue were pinned by CPU affinity */
7051 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7052 cpumask_clear_cpu(cpu_of(busiest), cpus);
7053 if (!cpumask_empty(cpus)) {
7055 env.loop_break = sched_nr_migrate_break;
7058 goto out_all_pinned;
7063 schedstat_inc(sd, lb_failed[idle]);
7065 * Increment the failure counter only on periodic balance.
7066 * We do not want newidle balance, which can be very
7067 * frequent, pollute the failure counter causing
7068 * excessive cache_hot migrations and active balances.
7070 if (idle != CPU_NEWLY_IDLE)
7071 sd->nr_balance_failed++;
7073 if (need_active_balance(&env)) {
7074 raw_spin_lock_irqsave(&busiest->lock, flags);
7076 /* don't kick the active_load_balance_cpu_stop,
7077 * if the curr task on busiest cpu can't be
7080 if (!cpumask_test_cpu(this_cpu,
7081 tsk_cpus_allowed(busiest->curr))) {
7082 raw_spin_unlock_irqrestore(&busiest->lock,
7084 env.flags |= LBF_ALL_PINNED;
7085 goto out_one_pinned;
7089 * ->active_balance synchronizes accesses to
7090 * ->active_balance_work. Once set, it's cleared
7091 * only after active load balance is finished.
7093 if (!busiest->active_balance) {
7094 busiest->active_balance = 1;
7095 busiest->push_cpu = this_cpu;
7098 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7100 if (active_balance) {
7101 stop_one_cpu_nowait(cpu_of(busiest),
7102 active_load_balance_cpu_stop, busiest,
7103 &busiest->active_balance_work);
7107 * We've kicked active balancing, reset the failure
7110 sd->nr_balance_failed = sd->cache_nice_tries+1;
7113 sd->nr_balance_failed = 0;
7115 if (likely(!active_balance)) {
7116 /* We were unbalanced, so reset the balancing interval */
7117 sd->balance_interval = sd->min_interval;
7120 * If we've begun active balancing, start to back off. This
7121 * case may not be covered by the all_pinned logic if there
7122 * is only 1 task on the busy runqueue (because we don't call
7125 if (sd->balance_interval < sd->max_interval)
7126 sd->balance_interval *= 2;
7133 * We reach balance although we may have faced some affinity
7134 * constraints. Clear the imbalance flag if it was set.
7137 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7139 if (*group_imbalance)
7140 *group_imbalance = 0;
7145 * We reach balance because all tasks are pinned at this level so
7146 * we can't migrate them. Let the imbalance flag set so parent level
7147 * can try to migrate them.
7149 schedstat_inc(sd, lb_balanced[idle]);
7151 sd->nr_balance_failed = 0;
7154 /* tune up the balancing interval */
7155 if (((env.flags & LBF_ALL_PINNED) &&
7156 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7157 (sd->balance_interval < sd->max_interval))
7158 sd->balance_interval *= 2;
7165 static inline unsigned long
7166 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7168 unsigned long interval = sd->balance_interval;
7171 interval *= sd->busy_factor;
7173 /* scale ms to jiffies */
7174 interval = msecs_to_jiffies(interval);
7175 interval = clamp(interval, 1UL, max_load_balance_interval);
7181 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7183 unsigned long interval, next;
7185 interval = get_sd_balance_interval(sd, cpu_busy);
7186 next = sd->last_balance + interval;
7188 if (time_after(*next_balance, next))
7189 *next_balance = next;
7193 * idle_balance is called by schedule() if this_cpu is about to become
7194 * idle. Attempts to pull tasks from other CPUs.
7196 static int idle_balance(struct rq *this_rq)
7198 unsigned long next_balance = jiffies + HZ;
7199 int this_cpu = this_rq->cpu;
7200 struct sched_domain *sd;
7201 int pulled_task = 0;
7204 idle_enter_fair(this_rq);
7207 * We must set idle_stamp _before_ calling idle_balance(), such that we
7208 * measure the duration of idle_balance() as idle time.
7210 this_rq->idle_stamp = rq_clock(this_rq);
7212 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7213 !this_rq->rd->overload) {
7215 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7217 update_next_balance(sd, 0, &next_balance);
7224 * Drop the rq->lock, but keep IRQ/preempt disabled.
7226 raw_spin_unlock(&this_rq->lock);
7228 update_blocked_averages(this_cpu);
7230 for_each_domain(this_cpu, sd) {
7231 int continue_balancing = 1;
7232 u64 t0, domain_cost;
7234 if (!(sd->flags & SD_LOAD_BALANCE))
7237 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7238 update_next_balance(sd, 0, &next_balance);
7242 if (sd->flags & SD_BALANCE_NEWIDLE) {
7243 t0 = sched_clock_cpu(this_cpu);
7245 pulled_task = load_balance(this_cpu, this_rq,
7247 &continue_balancing);
7249 domain_cost = sched_clock_cpu(this_cpu) - t0;
7250 if (domain_cost > sd->max_newidle_lb_cost)
7251 sd->max_newidle_lb_cost = domain_cost;
7253 curr_cost += domain_cost;
7256 update_next_balance(sd, 0, &next_balance);
7259 * Stop searching for tasks to pull if there are
7260 * now runnable tasks on this rq.
7262 if (pulled_task || this_rq->nr_running > 0)
7267 raw_spin_lock(&this_rq->lock);
7269 if (curr_cost > this_rq->max_idle_balance_cost)
7270 this_rq->max_idle_balance_cost = curr_cost;
7273 * While browsing the domains, we released the rq lock, a task could
7274 * have been enqueued in the meantime. Since we're not going idle,
7275 * pretend we pulled a task.
7277 if (this_rq->cfs.h_nr_running && !pulled_task)
7281 /* Move the next balance forward */
7282 if (time_after(this_rq->next_balance, next_balance))
7283 this_rq->next_balance = next_balance;
7285 /* Is there a task of a high priority class? */
7286 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7290 idle_exit_fair(this_rq);
7291 this_rq->idle_stamp = 0;
7298 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7299 * running tasks off the busiest CPU onto idle CPUs. It requires at
7300 * least 1 task to be running on each physical CPU where possible, and
7301 * avoids physical / logical imbalances.
7303 static int active_load_balance_cpu_stop(void *data)
7305 struct rq *busiest_rq = data;
7306 int busiest_cpu = cpu_of(busiest_rq);
7307 int target_cpu = busiest_rq->push_cpu;
7308 struct rq *target_rq = cpu_rq(target_cpu);
7309 struct sched_domain *sd;
7310 struct task_struct *p = NULL;
7312 raw_spin_lock_irq(&busiest_rq->lock);
7314 /* make sure the requested cpu hasn't gone down in the meantime */
7315 if (unlikely(busiest_cpu != smp_processor_id() ||
7316 !busiest_rq->active_balance))
7319 /* Is there any task to move? */
7320 if (busiest_rq->nr_running <= 1)
7324 * This condition is "impossible", if it occurs
7325 * we need to fix it. Originally reported by
7326 * Bjorn Helgaas on a 128-cpu setup.
7328 BUG_ON(busiest_rq == target_rq);
7330 /* Search for an sd spanning us and the target CPU. */
7332 for_each_domain(target_cpu, sd) {
7333 if ((sd->flags & SD_LOAD_BALANCE) &&
7334 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7339 struct lb_env env = {
7341 .dst_cpu = target_cpu,
7342 .dst_rq = target_rq,
7343 .src_cpu = busiest_rq->cpu,
7344 .src_rq = busiest_rq,
7348 schedstat_inc(sd, alb_count);
7350 p = detach_one_task(&env);
7352 schedstat_inc(sd, alb_pushed);
7354 schedstat_inc(sd, alb_failed);
7358 busiest_rq->active_balance = 0;
7359 raw_spin_unlock(&busiest_rq->lock);
7362 attach_one_task(target_rq, p);
7369 static inline int on_null_domain(struct rq *rq)
7371 return unlikely(!rcu_dereference_sched(rq->sd));
7374 #ifdef CONFIG_NO_HZ_COMMON
7376 * idle load balancing details
7377 * - When one of the busy CPUs notice that there may be an idle rebalancing
7378 * needed, they will kick the idle load balancer, which then does idle
7379 * load balancing for all the idle CPUs.
7382 cpumask_var_t idle_cpus_mask;
7384 unsigned long next_balance; /* in jiffy units */
7385 } nohz ____cacheline_aligned;
7387 static inline int find_new_ilb(void)
7389 int ilb = cpumask_first(nohz.idle_cpus_mask);
7391 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7398 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7399 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7400 * CPU (if there is one).
7402 static void nohz_balancer_kick(void)
7406 nohz.next_balance++;
7408 ilb_cpu = find_new_ilb();
7410 if (ilb_cpu >= nr_cpu_ids)
7413 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7416 * Use smp_send_reschedule() instead of resched_cpu().
7417 * This way we generate a sched IPI on the target cpu which
7418 * is idle. And the softirq performing nohz idle load balance
7419 * will be run before returning from the IPI.
7421 smp_send_reschedule(ilb_cpu);
7425 static inline void nohz_balance_exit_idle(int cpu)
7427 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7429 * Completely isolated CPUs don't ever set, so we must test.
7431 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7432 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7433 atomic_dec(&nohz.nr_cpus);
7435 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7439 static inline void set_cpu_sd_state_busy(void)
7441 struct sched_domain *sd;
7442 int cpu = smp_processor_id();
7445 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7447 if (!sd || !sd->nohz_idle)
7451 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7456 void set_cpu_sd_state_idle(void)
7458 struct sched_domain *sd;
7459 int cpu = smp_processor_id();
7462 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7464 if (!sd || sd->nohz_idle)
7468 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7474 * This routine will record that the cpu is going idle with tick stopped.
7475 * This info will be used in performing idle load balancing in the future.
7477 void nohz_balance_enter_idle(int cpu)
7480 * If this cpu is going down, then nothing needs to be done.
7482 if (!cpu_active(cpu))
7485 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7489 * If we're a completely isolated CPU, we don't play.
7491 if (on_null_domain(cpu_rq(cpu)))
7494 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7495 atomic_inc(&nohz.nr_cpus);
7496 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7499 static int sched_ilb_notifier(struct notifier_block *nfb,
7500 unsigned long action, void *hcpu)
7502 switch (action & ~CPU_TASKS_FROZEN) {
7504 nohz_balance_exit_idle(smp_processor_id());
7512 static DEFINE_SPINLOCK(balancing);
7515 * Scale the max load_balance interval with the number of CPUs in the system.
7516 * This trades load-balance latency on larger machines for less cross talk.
7518 void update_max_interval(void)
7520 max_load_balance_interval = HZ*num_online_cpus()/10;
7524 * It checks each scheduling domain to see if it is due to be balanced,
7525 * and initiates a balancing operation if so.
7527 * Balancing parameters are set up in init_sched_domains.
7529 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7531 int continue_balancing = 1;
7533 unsigned long interval;
7534 struct sched_domain *sd;
7535 /* Earliest time when we have to do rebalance again */
7536 unsigned long next_balance = jiffies + 60*HZ;
7537 int update_next_balance = 0;
7538 int need_serialize, need_decay = 0;
7541 update_blocked_averages(cpu);
7544 for_each_domain(cpu, sd) {
7546 * Decay the newidle max times here because this is a regular
7547 * visit to all the domains. Decay ~1% per second.
7549 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7550 sd->max_newidle_lb_cost =
7551 (sd->max_newidle_lb_cost * 253) / 256;
7552 sd->next_decay_max_lb_cost = jiffies + HZ;
7555 max_cost += sd->max_newidle_lb_cost;
7557 if (!(sd->flags & SD_LOAD_BALANCE))
7561 * Stop the load balance at this level. There is another
7562 * CPU in our sched group which is doing load balancing more
7565 if (!continue_balancing) {
7571 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7573 need_serialize = sd->flags & SD_SERIALIZE;
7574 if (need_serialize) {
7575 if (!spin_trylock(&balancing))
7579 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7580 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7582 * The LBF_DST_PINNED logic could have changed
7583 * env->dst_cpu, so we can't know our idle
7584 * state even if we migrated tasks. Update it.
7586 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7588 sd->last_balance = jiffies;
7589 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7592 spin_unlock(&balancing);
7594 if (time_after(next_balance, sd->last_balance + interval)) {
7595 next_balance = sd->last_balance + interval;
7596 update_next_balance = 1;
7601 * Ensure the rq-wide value also decays but keep it at a
7602 * reasonable floor to avoid funnies with rq->avg_idle.
7604 rq->max_idle_balance_cost =
7605 max((u64)sysctl_sched_migration_cost, max_cost);
7610 * next_balance will be updated only when there is a need.
7611 * When the cpu is attached to null domain for ex, it will not be
7614 if (likely(update_next_balance))
7615 rq->next_balance = next_balance;
7618 #ifdef CONFIG_NO_HZ_COMMON
7620 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7621 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7623 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7625 int this_cpu = this_rq->cpu;
7629 if (idle != CPU_IDLE ||
7630 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7633 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7634 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7638 * If this cpu gets work to do, stop the load balancing
7639 * work being done for other cpus. Next load
7640 * balancing owner will pick it up.
7645 rq = cpu_rq(balance_cpu);
7648 * If time for next balance is due,
7651 if (time_after_eq(jiffies, rq->next_balance)) {
7652 raw_spin_lock_irq(&rq->lock);
7653 update_rq_clock(rq);
7654 update_idle_cpu_load(rq);
7655 raw_spin_unlock_irq(&rq->lock);
7656 rebalance_domains(rq, CPU_IDLE);
7659 if (time_after(this_rq->next_balance, rq->next_balance))
7660 this_rq->next_balance = rq->next_balance;
7662 nohz.next_balance = this_rq->next_balance;
7664 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7668 * Current heuristic for kicking the idle load balancer in the presence
7669 * of an idle cpu in the system.
7670 * - This rq has more than one task.
7671 * - This rq has at least one CFS task and the capacity of the CPU is
7672 * significantly reduced because of RT tasks or IRQs.
7673 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7674 * multiple busy cpu.
7675 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7676 * domain span are idle.
7678 static inline bool nohz_kick_needed(struct rq *rq)
7680 unsigned long now = jiffies;
7681 struct sched_domain *sd;
7682 struct sched_group_capacity *sgc;
7683 int nr_busy, cpu = rq->cpu;
7686 if (unlikely(rq->idle_balance))
7690 * We may be recently in ticked or tickless idle mode. At the first
7691 * busy tick after returning from idle, we will update the busy stats.
7693 set_cpu_sd_state_busy();
7694 nohz_balance_exit_idle(cpu);
7697 * None are in tickless mode and hence no need for NOHZ idle load
7700 if (likely(!atomic_read(&nohz.nr_cpus)))
7703 if (time_before(now, nohz.next_balance))
7706 if (rq->nr_running >= 2)
7710 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7712 sgc = sd->groups->sgc;
7713 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7722 sd = rcu_dereference(rq->sd);
7724 if ((rq->cfs.h_nr_running >= 1) &&
7725 check_cpu_capacity(rq, sd)) {
7731 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7732 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7733 sched_domain_span(sd)) < cpu)) {
7743 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7747 * run_rebalance_domains is triggered when needed from the scheduler tick.
7748 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7750 static void run_rebalance_domains(struct softirq_action *h)
7752 struct rq *this_rq = this_rq();
7753 enum cpu_idle_type idle = this_rq->idle_balance ?
7754 CPU_IDLE : CPU_NOT_IDLE;
7756 rebalance_domains(this_rq, idle);
7759 * If this cpu has a pending nohz_balance_kick, then do the
7760 * balancing on behalf of the other idle cpus whose ticks are
7763 nohz_idle_balance(this_rq, idle);
7767 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7769 void trigger_load_balance(struct rq *rq)
7771 /* Don't need to rebalance while attached to NULL domain */
7772 if (unlikely(on_null_domain(rq)))
7775 if (time_after_eq(jiffies, rq->next_balance))
7776 raise_softirq(SCHED_SOFTIRQ);
7777 #ifdef CONFIG_NO_HZ_COMMON
7778 if (nohz_kick_needed(rq))
7779 nohz_balancer_kick();
7783 static void rq_online_fair(struct rq *rq)
7787 update_runtime_enabled(rq);
7790 static void rq_offline_fair(struct rq *rq)
7794 /* Ensure any throttled groups are reachable by pick_next_task */
7795 unthrottle_offline_cfs_rqs(rq);
7798 #endif /* CONFIG_SMP */
7801 * scheduler tick hitting a task of our scheduling class:
7803 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7805 struct cfs_rq *cfs_rq;
7806 struct sched_entity *se = &curr->se;
7808 for_each_sched_entity(se) {
7809 cfs_rq = cfs_rq_of(se);
7810 entity_tick(cfs_rq, se, queued);
7813 if (numabalancing_enabled)
7814 task_tick_numa(rq, curr);
7816 update_rq_runnable_avg(rq, 1);
7820 * called on fork with the child task as argument from the parent's context
7821 * - child not yet on the tasklist
7822 * - preemption disabled
7824 static void task_fork_fair(struct task_struct *p)
7826 struct cfs_rq *cfs_rq;
7827 struct sched_entity *se = &p->se, *curr;
7828 int this_cpu = smp_processor_id();
7829 struct rq *rq = this_rq();
7830 unsigned long flags;
7832 raw_spin_lock_irqsave(&rq->lock, flags);
7834 update_rq_clock(rq);
7836 cfs_rq = task_cfs_rq(current);
7837 curr = cfs_rq->curr;
7840 * Not only the cpu but also the task_group of the parent might have
7841 * been changed after parent->se.parent,cfs_rq were copied to
7842 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7843 * of child point to valid ones.
7846 __set_task_cpu(p, this_cpu);
7849 update_curr(cfs_rq);
7852 se->vruntime = curr->vruntime;
7853 place_entity(cfs_rq, se, 1);
7855 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7857 * Upon rescheduling, sched_class::put_prev_task() will place
7858 * 'current' within the tree based on its new key value.
7860 swap(curr->vruntime, se->vruntime);
7864 se->vruntime -= cfs_rq->min_vruntime;
7866 raw_spin_unlock_irqrestore(&rq->lock, flags);
7870 * Priority of the task has changed. Check to see if we preempt
7874 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7876 if (!task_on_rq_queued(p))
7880 * Reschedule if we are currently running on this runqueue and
7881 * our priority decreased, or if we are not currently running on
7882 * this runqueue and our priority is higher than the current's
7884 if (rq->curr == p) {
7885 if (p->prio > oldprio)
7888 check_preempt_curr(rq, p, 0);
7891 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7893 struct sched_entity *se = &p->se;
7894 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7897 * Ensure the task's vruntime is normalized, so that when it's
7898 * switched back to the fair class the enqueue_entity(.flags=0) will
7899 * do the right thing.
7901 * If it's queued, then the dequeue_entity(.flags=0) will already
7902 * have normalized the vruntime, if it's !queued, then only when
7903 * the task is sleeping will it still have non-normalized vruntime.
7905 if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) {
7907 * Fix up our vruntime so that the current sleep doesn't
7908 * cause 'unlimited' sleep bonus.
7910 place_entity(cfs_rq, se, 0);
7911 se->vruntime -= cfs_rq->min_vruntime;
7916 * Remove our load from contribution when we leave sched_fair
7917 * and ensure we don't carry in an old decay_count if we
7920 if (se->avg.decay_count) {
7921 __synchronize_entity_decay(se);
7922 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7928 * We switched to the sched_fair class.
7930 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7932 #ifdef CONFIG_FAIR_GROUP_SCHED
7933 struct sched_entity *se = &p->se;
7935 * Since the real-depth could have been changed (only FAIR
7936 * class maintain depth value), reset depth properly.
7938 se->depth = se->parent ? se->parent->depth + 1 : 0;
7940 if (!task_on_rq_queued(p))
7944 * We were most likely switched from sched_rt, so
7945 * kick off the schedule if running, otherwise just see
7946 * if we can still preempt the current task.
7951 check_preempt_curr(rq, p, 0);
7954 /* Account for a task changing its policy or group.
7956 * This routine is mostly called to set cfs_rq->curr field when a task
7957 * migrates between groups/classes.
7959 static void set_curr_task_fair(struct rq *rq)
7961 struct sched_entity *se = &rq->curr->se;
7963 for_each_sched_entity(se) {
7964 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7966 set_next_entity(cfs_rq, se);
7967 /* ensure bandwidth has been allocated on our new cfs_rq */
7968 account_cfs_rq_runtime(cfs_rq, 0);
7972 void init_cfs_rq(struct cfs_rq *cfs_rq)
7974 cfs_rq->tasks_timeline = RB_ROOT;
7975 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7976 #ifndef CONFIG_64BIT
7977 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7980 atomic64_set(&cfs_rq->decay_counter, 1);
7981 atomic_long_set(&cfs_rq->removed_load, 0);
7985 #ifdef CONFIG_FAIR_GROUP_SCHED
7986 static void task_move_group_fair(struct task_struct *p, int queued)
7988 struct sched_entity *se = &p->se;
7989 struct cfs_rq *cfs_rq;
7992 * If the task was not on the rq at the time of this cgroup movement
7993 * it must have been asleep, sleeping tasks keep their ->vruntime
7994 * absolute on their old rq until wakeup (needed for the fair sleeper
7995 * bonus in place_entity()).
7997 * If it was on the rq, we've just 'preempted' it, which does convert
7998 * ->vruntime to a relative base.
8000 * Make sure both cases convert their relative position when migrating
8001 * to another cgroup's rq. This does somewhat interfere with the
8002 * fair sleeper stuff for the first placement, but who cares.
8005 * When !queued, vruntime of the task has usually NOT been normalized.
8006 * But there are some cases where it has already been normalized:
8008 * - Moving a forked child which is waiting for being woken up by
8009 * wake_up_new_task().
8010 * - Moving a task which has been woken up by try_to_wake_up() and
8011 * waiting for actually being woken up by sched_ttwu_pending().
8013 * To prevent boost or penalty in the new cfs_rq caused by delta
8014 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
8016 if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING))
8020 se->vruntime -= cfs_rq_of(se)->min_vruntime;
8021 set_task_rq(p, task_cpu(p));
8022 se->depth = se->parent ? se->parent->depth + 1 : 0;
8024 cfs_rq = cfs_rq_of(se);
8025 se->vruntime += cfs_rq->min_vruntime;
8028 * migrate_task_rq_fair() will have removed our previous
8029 * contribution, but we must synchronize for ongoing future
8032 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
8033 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
8038 void free_fair_sched_group(struct task_group *tg)
8042 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8044 for_each_possible_cpu(i) {
8046 kfree(tg->cfs_rq[i]);
8055 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8057 struct cfs_rq *cfs_rq;
8058 struct sched_entity *se;
8061 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8064 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8068 tg->shares = NICE_0_LOAD;
8070 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8072 for_each_possible_cpu(i) {
8073 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8074 GFP_KERNEL, cpu_to_node(i));
8078 se = kzalloc_node(sizeof(struct sched_entity),
8079 GFP_KERNEL, cpu_to_node(i));
8083 init_cfs_rq(cfs_rq);
8084 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8095 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8097 struct rq *rq = cpu_rq(cpu);
8098 unsigned long flags;
8101 * Only empty task groups can be destroyed; so we can speculatively
8102 * check on_list without danger of it being re-added.
8104 if (!tg->cfs_rq[cpu]->on_list)
8107 raw_spin_lock_irqsave(&rq->lock, flags);
8108 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8109 raw_spin_unlock_irqrestore(&rq->lock, flags);
8112 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8113 struct sched_entity *se, int cpu,
8114 struct sched_entity *parent)
8116 struct rq *rq = cpu_rq(cpu);
8120 init_cfs_rq_runtime(cfs_rq);
8122 tg->cfs_rq[cpu] = cfs_rq;
8125 /* se could be NULL for root_task_group */
8130 se->cfs_rq = &rq->cfs;
8133 se->cfs_rq = parent->my_q;
8134 se->depth = parent->depth + 1;
8138 /* guarantee group entities always have weight */
8139 update_load_set(&se->load, NICE_0_LOAD);
8140 se->parent = parent;
8143 static DEFINE_MUTEX(shares_mutex);
8145 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8148 unsigned long flags;
8151 * We can't change the weight of the root cgroup.
8156 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8158 mutex_lock(&shares_mutex);
8159 if (tg->shares == shares)
8162 tg->shares = shares;
8163 for_each_possible_cpu(i) {
8164 struct rq *rq = cpu_rq(i);
8165 struct sched_entity *se;
8168 /* Propagate contribution to hierarchy */
8169 raw_spin_lock_irqsave(&rq->lock, flags);
8171 /* Possible calls to update_curr() need rq clock */
8172 update_rq_clock(rq);
8173 for_each_sched_entity(se)
8174 update_cfs_shares(group_cfs_rq(se));
8175 raw_spin_unlock_irqrestore(&rq->lock, flags);
8179 mutex_unlock(&shares_mutex);
8182 #else /* CONFIG_FAIR_GROUP_SCHED */
8184 void free_fair_sched_group(struct task_group *tg) { }
8186 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8191 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8193 #endif /* CONFIG_FAIR_GROUP_SCHED */
8196 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8198 struct sched_entity *se = &task->se;
8199 unsigned int rr_interval = 0;
8202 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8205 if (rq->cfs.load.weight)
8206 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8212 * All the scheduling class methods:
8214 const struct sched_class fair_sched_class = {
8215 .next = &idle_sched_class,
8216 .enqueue_task = enqueue_task_fair,
8217 .dequeue_task = dequeue_task_fair,
8218 .yield_task = yield_task_fair,
8219 .yield_to_task = yield_to_task_fair,
8221 .check_preempt_curr = check_preempt_wakeup,
8223 .pick_next_task = pick_next_task_fair,
8224 .put_prev_task = put_prev_task_fair,
8227 .select_task_rq = select_task_rq_fair,
8228 .migrate_task_rq = migrate_task_rq_fair,
8230 .rq_online = rq_online_fair,
8231 .rq_offline = rq_offline_fair,
8233 .task_waking = task_waking_fair,
8236 .set_curr_task = set_curr_task_fair,
8237 .task_tick = task_tick_fair,
8238 .task_fork = task_fork_fair,
8240 .prio_changed = prio_changed_fair,
8241 .switched_from = switched_from_fair,
8242 .switched_to = switched_to_fair,
8244 .get_rr_interval = get_rr_interval_fair,
8246 .update_curr = update_curr_fair,
8248 #ifdef CONFIG_FAIR_GROUP_SCHED
8249 .task_move_group = task_move_group_fair,
8253 #ifdef CONFIG_SCHED_DEBUG
8254 void print_cfs_stats(struct seq_file *m, int cpu)
8256 struct cfs_rq *cfs_rq;
8259 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8260 print_cfs_rq(m, cpu, cfs_rq);
8265 __init void init_sched_fair_class(void)
8268 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8270 #ifdef CONFIG_NO_HZ_COMMON
8271 nohz.next_balance = jiffies;
8272 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8273 cpu_notifier(sched_ilb_notifier, 0);