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/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #define WMULT_CONST (~0U)
182 #define WMULT_SHIFT 32
184 static void __update_inv_weight(struct load_weight *lw)
188 if (likely(lw->inv_weight))
191 w = scale_load_down(lw->weight);
193 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
195 else if (unlikely(!w))
196 lw->inv_weight = WMULT_CONST;
198 lw->inv_weight = WMULT_CONST / w;
202 * delta_exec * weight / lw.weight
204 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
207 * we're guaranteed shift stays positive because inv_weight is guaranteed to
208 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
211 * weight/lw.weight <= 1, and therefore our shift will also be positive.
213 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
215 u64 fact = scale_load_down(weight);
216 int shift = WMULT_SHIFT;
218 __update_inv_weight(lw);
220 if (unlikely(fact >> 32)) {
227 /* hint to use a 32x32->64 mul */
228 fact = (u64)(u32)fact * lw->inv_weight;
235 return mul_u64_u32_shr(delta_exec, fact, shift);
239 const struct sched_class fair_sched_class;
241 /**************************************************************
242 * CFS operations on generic schedulable entities:
245 #ifdef CONFIG_FAIR_GROUP_SCHED
247 /* cpu runqueue to which this cfs_rq is attached */
248 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
253 /* An entity is a task if it doesn't "own" a runqueue */
254 #define entity_is_task(se) (!se->my_q)
256 static inline struct task_struct *task_of(struct sched_entity *se)
258 #ifdef CONFIG_SCHED_DEBUG
259 WARN_ON_ONCE(!entity_is_task(se));
261 return container_of(se, struct task_struct, se);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
285 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 if (!cfs_rq->on_list) {
292 * Ensure we either appear before our parent (if already
293 * enqueued) or force our parent to appear after us when it is
294 * enqueued. The fact that we always enqueue bottom-up
295 * reduces this to two cases.
297 if (cfs_rq->tg->parent &&
298 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
299 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
300 &rq_of(cfs_rq)->leaf_cfs_rq_list);
302 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
303 &rq_of(cfs_rq)->leaf_cfs_rq_list);
307 /* We should have no load, but we need to update last_decay. */
308 update_cfs_rq_blocked_load(cfs_rq, 0);
312 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
314 if (cfs_rq->on_list) {
315 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
320 /* Iterate thr' all leaf cfs_rq's on a runqueue */
321 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
322 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
324 /* Do the two (enqueued) entities belong to the same group ? */
325 static inline struct cfs_rq *
326 is_same_group(struct sched_entity *se, struct sched_entity *pse)
328 if (se->cfs_rq == pse->cfs_rq)
334 static inline struct sched_entity *parent_entity(struct sched_entity *se)
340 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
342 int se_depth, pse_depth;
345 * preemption test can be made between sibling entities who are in the
346 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
347 * both tasks until we find their ancestors who are siblings of common
351 /* First walk up until both entities are at same depth */
352 se_depth = (*se)->depth;
353 pse_depth = (*pse)->depth;
355 while (se_depth > pse_depth) {
357 *se = parent_entity(*se);
360 while (pse_depth > se_depth) {
362 *pse = parent_entity(*pse);
365 while (!is_same_group(*se, *pse)) {
366 *se = parent_entity(*se);
367 *pse = parent_entity(*pse);
371 #else /* !CONFIG_FAIR_GROUP_SCHED */
373 static inline struct task_struct *task_of(struct sched_entity *se)
375 return container_of(se, struct task_struct, se);
378 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
380 return container_of(cfs_rq, struct rq, cfs);
383 #define entity_is_task(se) 1
385 #define for_each_sched_entity(se) \
386 for (; se; se = NULL)
388 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
390 return &task_rq(p)->cfs;
393 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
395 struct task_struct *p = task_of(se);
396 struct rq *rq = task_rq(p);
401 /* runqueue "owned" by this group */
402 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
407 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
415 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
416 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
418 static inline struct sched_entity *parent_entity(struct sched_entity *se)
424 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
428 #endif /* CONFIG_FAIR_GROUP_SCHED */
430 static __always_inline
431 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
433 /**************************************************************
434 * Scheduling class tree data structure manipulation methods:
437 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
439 s64 delta = (s64)(vruntime - max_vruntime);
441 max_vruntime = vruntime;
446 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
448 s64 delta = (s64)(vruntime - min_vruntime);
450 min_vruntime = vruntime;
455 static inline int entity_before(struct sched_entity *a,
456 struct sched_entity *b)
458 return (s64)(a->vruntime - b->vruntime) < 0;
461 static void update_min_vruntime(struct cfs_rq *cfs_rq)
463 u64 vruntime = cfs_rq->min_vruntime;
466 vruntime = cfs_rq->curr->vruntime;
468 if (cfs_rq->rb_leftmost) {
469 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
474 vruntime = se->vruntime;
476 vruntime = min_vruntime(vruntime, se->vruntime);
479 /* ensure we never gain time by being placed backwards. */
480 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
483 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
488 * Enqueue an entity into the rb-tree:
490 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
492 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
493 struct rb_node *parent = NULL;
494 struct sched_entity *entry;
498 * Find the right place in the rbtree:
502 entry = rb_entry(parent, struct sched_entity, run_node);
504 * We dont care about collisions. Nodes with
505 * the same key stay together.
507 if (entity_before(se, entry)) {
508 link = &parent->rb_left;
510 link = &parent->rb_right;
516 * Maintain a cache of leftmost tree entries (it is frequently
520 cfs_rq->rb_leftmost = &se->run_node;
522 rb_link_node(&se->run_node, parent, link);
523 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
526 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
528 if (cfs_rq->rb_leftmost == &se->run_node) {
529 struct rb_node *next_node;
531 next_node = rb_next(&se->run_node);
532 cfs_rq->rb_leftmost = next_node;
535 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
538 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
540 struct rb_node *left = cfs_rq->rb_leftmost;
545 return rb_entry(left, struct sched_entity, run_node);
548 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
550 struct rb_node *next = rb_next(&se->run_node);
555 return rb_entry(next, struct sched_entity, run_node);
558 #ifdef CONFIG_SCHED_DEBUG
559 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
561 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
566 return rb_entry(last, struct sched_entity, run_node);
569 /**************************************************************
570 * Scheduling class statistics methods:
573 int sched_proc_update_handler(struct ctl_table *table, int write,
574 void __user *buffer, size_t *lenp,
577 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
578 int factor = get_update_sysctl_factor();
583 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
584 sysctl_sched_min_granularity);
586 #define WRT_SYSCTL(name) \
587 (normalized_sysctl_##name = sysctl_##name / (factor))
588 WRT_SYSCTL(sched_min_granularity);
589 WRT_SYSCTL(sched_latency);
590 WRT_SYSCTL(sched_wakeup_granularity);
600 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
602 if (unlikely(se->load.weight != NICE_0_LOAD))
603 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
609 * The idea is to set a period in which each task runs once.
611 * When there are too many tasks (sched_nr_latency) we have to stretch
612 * this period because otherwise the slices get too small.
614 * p = (nr <= nl) ? l : l*nr/nl
616 static u64 __sched_period(unsigned long nr_running)
618 u64 period = sysctl_sched_latency;
619 unsigned long nr_latency = sched_nr_latency;
621 if (unlikely(nr_running > nr_latency)) {
622 period = sysctl_sched_min_granularity;
623 period *= nr_running;
630 * We calculate the wall-time slice from the period by taking a part
631 * proportional to the weight.
635 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
637 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
639 for_each_sched_entity(se) {
640 struct load_weight *load;
641 struct load_weight lw;
643 cfs_rq = cfs_rq_of(se);
644 load = &cfs_rq->load;
646 if (unlikely(!se->on_rq)) {
649 update_load_add(&lw, se->load.weight);
652 slice = __calc_delta(slice, se->load.weight, load);
658 * We calculate the vruntime slice of a to-be-inserted task.
662 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
664 return calc_delta_fair(sched_slice(cfs_rq, se), se);
668 static unsigned long task_h_load(struct task_struct *p);
670 static inline void __update_task_entity_contrib(struct sched_entity *se);
672 /* Give new task start runnable values to heavy its load in infant time */
673 void init_task_runnable_average(struct task_struct *p)
677 p->se.avg.decay_count = 0;
678 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
679 p->se.avg.runnable_avg_sum = slice;
680 p->se.avg.runnable_avg_period = slice;
681 __update_task_entity_contrib(&p->se);
684 void init_task_runnable_average(struct task_struct *p)
690 * Update the current task's runtime statistics.
692 static void update_curr(struct cfs_rq *cfs_rq)
694 struct sched_entity *curr = cfs_rq->curr;
695 u64 now = rq_clock_task(rq_of(cfs_rq));
701 delta_exec = now - curr->exec_start;
702 if (unlikely((s64)delta_exec <= 0))
705 curr->exec_start = now;
707 schedstat_set(curr->statistics.exec_max,
708 max(delta_exec, curr->statistics.exec_max));
710 curr->sum_exec_runtime += delta_exec;
711 schedstat_add(cfs_rq, exec_clock, delta_exec);
713 curr->vruntime += calc_delta_fair(delta_exec, curr);
714 update_min_vruntime(cfs_rq);
716 if (entity_is_task(curr)) {
717 struct task_struct *curtask = task_of(curr);
719 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
720 cpuacct_charge(curtask, delta_exec);
721 account_group_exec_runtime(curtask, delta_exec);
724 account_cfs_rq_runtime(cfs_rq, delta_exec);
728 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
730 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
734 * Task is being enqueued - update stats:
736 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
739 * Are we enqueueing a waiting task? (for current tasks
740 * a dequeue/enqueue event is a NOP)
742 if (se != cfs_rq->curr)
743 update_stats_wait_start(cfs_rq, se);
747 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
749 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
750 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
751 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
752 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
753 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
754 #ifdef CONFIG_SCHEDSTATS
755 if (entity_is_task(se)) {
756 trace_sched_stat_wait(task_of(se),
757 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
760 schedstat_set(se->statistics.wait_start, 0);
764 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
767 * Mark the end of the wait period if dequeueing a
770 if (se != cfs_rq->curr)
771 update_stats_wait_end(cfs_rq, se);
775 * We are picking a new current task - update its stats:
778 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
781 * We are starting a new run period:
783 se->exec_start = rq_clock_task(rq_of(cfs_rq));
786 /**************************************************
787 * Scheduling class queueing methods:
790 #ifdef CONFIG_NUMA_BALANCING
792 * Approximate time to scan a full NUMA task in ms. The task scan period is
793 * calculated based on the tasks virtual memory size and
794 * numa_balancing_scan_size.
796 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
797 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
799 /* Portion of address space to scan in MB */
800 unsigned int sysctl_numa_balancing_scan_size = 256;
802 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
803 unsigned int sysctl_numa_balancing_scan_delay = 1000;
805 static unsigned int task_nr_scan_windows(struct task_struct *p)
807 unsigned long rss = 0;
808 unsigned long nr_scan_pages;
811 * Calculations based on RSS as non-present and empty pages are skipped
812 * by the PTE scanner and NUMA hinting faults should be trapped based
815 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
816 rss = get_mm_rss(p->mm);
820 rss = round_up(rss, nr_scan_pages);
821 return rss / nr_scan_pages;
824 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
825 #define MAX_SCAN_WINDOW 2560
827 static unsigned int task_scan_min(struct task_struct *p)
829 unsigned int scan, floor;
830 unsigned int windows = 1;
832 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
833 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
834 floor = 1000 / windows;
836 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
837 return max_t(unsigned int, floor, scan);
840 static unsigned int task_scan_max(struct task_struct *p)
842 unsigned int smin = task_scan_min(p);
845 /* Watch for min being lower than max due to floor calculations */
846 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
847 return max(smin, smax);
850 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
852 rq->nr_numa_running += (p->numa_preferred_nid != -1);
853 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
856 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
858 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
859 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
865 spinlock_t lock; /* nr_tasks, tasks */
868 struct list_head task_list;
871 nodemask_t active_nodes;
872 unsigned long total_faults;
874 * Faults_cpu is used to decide whether memory should move
875 * towards the CPU. As a consequence, these stats are weighted
876 * more by CPU use than by memory faults.
878 unsigned long *faults_cpu;
879 unsigned long faults[0];
882 /* Shared or private faults. */
883 #define NR_NUMA_HINT_FAULT_TYPES 2
885 /* Memory and CPU locality */
886 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
888 /* Averaged statistics, and temporary buffers. */
889 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
891 pid_t task_numa_group_id(struct task_struct *p)
893 return p->numa_group ? p->numa_group->gid : 0;
896 static inline int task_faults_idx(int nid, int priv)
898 return NR_NUMA_HINT_FAULT_TYPES * nid + priv;
901 static inline unsigned long task_faults(struct task_struct *p, int nid)
903 if (!p->numa_faults_memory)
906 return p->numa_faults_memory[task_faults_idx(nid, 0)] +
907 p->numa_faults_memory[task_faults_idx(nid, 1)];
910 static inline unsigned long group_faults(struct task_struct *p, int nid)
915 return p->numa_group->faults[task_faults_idx(nid, 0)] +
916 p->numa_group->faults[task_faults_idx(nid, 1)];
919 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
921 return group->faults_cpu[task_faults_idx(nid, 0)] +
922 group->faults_cpu[task_faults_idx(nid, 1)];
926 * These return the fraction of accesses done by a particular task, or
927 * task group, on a particular numa node. The group weight is given a
928 * larger multiplier, in order to group tasks together that are almost
929 * evenly spread out between numa nodes.
931 static inline unsigned long task_weight(struct task_struct *p, int nid)
933 unsigned long total_faults;
935 if (!p->numa_faults_memory)
938 total_faults = p->total_numa_faults;
943 return 1000 * task_faults(p, nid) / total_faults;
946 static inline unsigned long group_weight(struct task_struct *p, int nid)
948 if (!p->numa_group || !p->numa_group->total_faults)
951 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
954 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
955 int src_nid, int dst_cpu)
957 struct numa_group *ng = p->numa_group;
958 int dst_nid = cpu_to_node(dst_cpu);
959 int last_cpupid, this_cpupid;
961 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
964 * Multi-stage node selection is used in conjunction with a periodic
965 * migration fault to build a temporal task<->page relation. By using
966 * a two-stage filter we remove short/unlikely relations.
968 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
969 * a task's usage of a particular page (n_p) per total usage of this
970 * page (n_t) (in a given time-span) to a probability.
972 * Our periodic faults will sample this probability and getting the
973 * same result twice in a row, given these samples are fully
974 * independent, is then given by P(n)^2, provided our sample period
975 * is sufficiently short compared to the usage pattern.
977 * This quadric squishes small probabilities, making it less likely we
978 * act on an unlikely task<->page relation.
980 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
981 if (!cpupid_pid_unset(last_cpupid) &&
982 cpupid_to_nid(last_cpupid) != dst_nid)
985 /* Always allow migrate on private faults */
986 if (cpupid_match_pid(p, last_cpupid))
989 /* A shared fault, but p->numa_group has not been set up yet. */
994 * Do not migrate if the destination is not a node that
995 * is actively used by this numa group.
997 if (!node_isset(dst_nid, ng->active_nodes))
1001 * Source is a node that is not actively used by this
1002 * numa group, while the destination is. Migrate.
1004 if (!node_isset(src_nid, ng->active_nodes))
1008 * Both source and destination are nodes in active
1009 * use by this numa group. Maximize memory bandwidth
1010 * by migrating from more heavily used groups, to less
1011 * heavily used ones, spreading the load around.
1012 * Use a 1/4 hysteresis to avoid spurious page movement.
1014 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1017 static unsigned long weighted_cpuload(const int cpu);
1018 static unsigned long source_load(int cpu, int type);
1019 static unsigned long target_load(int cpu, int type);
1020 static unsigned long capacity_of(int cpu);
1021 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1023 /* Cached statistics for all CPUs within a node */
1025 unsigned long nr_running;
1028 /* Total compute capacity of CPUs on a node */
1029 unsigned long compute_capacity;
1031 /* Approximate capacity in terms of runnable tasks on a node */
1032 unsigned long task_capacity;
1033 int has_free_capacity;
1037 * XXX borrowed from update_sg_lb_stats
1039 static void update_numa_stats(struct numa_stats *ns, int nid)
1041 int smt, cpu, cpus = 0;
1042 unsigned long capacity;
1044 memset(ns, 0, sizeof(*ns));
1045 for_each_cpu(cpu, cpumask_of_node(nid)) {
1046 struct rq *rq = cpu_rq(cpu);
1048 ns->nr_running += rq->nr_running;
1049 ns->load += weighted_cpuload(cpu);
1050 ns->compute_capacity += capacity_of(cpu);
1056 * If we raced with hotplug and there are no CPUs left in our mask
1057 * the @ns structure is NULL'ed and task_numa_compare() will
1058 * not find this node attractive.
1060 * We'll either bail at !has_free_capacity, or we'll detect a huge
1061 * imbalance and bail there.
1066 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1067 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1068 capacity = cpus / smt; /* cores */
1070 ns->task_capacity = min_t(unsigned, capacity,
1071 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1072 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1075 struct task_numa_env {
1076 struct task_struct *p;
1078 int src_cpu, src_nid;
1079 int dst_cpu, dst_nid;
1081 struct numa_stats src_stats, dst_stats;
1085 struct task_struct *best_task;
1090 static void task_numa_assign(struct task_numa_env *env,
1091 struct task_struct *p, long imp)
1094 put_task_struct(env->best_task);
1099 env->best_imp = imp;
1100 env->best_cpu = env->dst_cpu;
1103 static bool load_too_imbalanced(long src_load, long dst_load,
1104 struct task_numa_env *env)
1107 long orig_src_load, orig_dst_load;
1108 long src_capacity, dst_capacity;
1111 * The load is corrected for the CPU capacity available on each node.
1114 * ------------ vs ---------
1115 * src_capacity dst_capacity
1117 src_capacity = env->src_stats.compute_capacity;
1118 dst_capacity = env->dst_stats.compute_capacity;
1120 /* We care about the slope of the imbalance, not the direction. */
1121 if (dst_load < src_load)
1122 swap(dst_load, src_load);
1124 /* Is the difference below the threshold? */
1125 imb = dst_load * src_capacity * 100 -
1126 src_load * dst_capacity * env->imbalance_pct;
1131 * The imbalance is above the allowed threshold.
1132 * Compare it with the old imbalance.
1134 orig_src_load = env->src_stats.load;
1135 orig_dst_load = env->dst_stats.load;
1137 if (orig_dst_load < orig_src_load)
1138 swap(orig_dst_load, orig_src_load);
1140 old_imb = orig_dst_load * src_capacity * 100 -
1141 orig_src_load * dst_capacity * env->imbalance_pct;
1143 /* Would this change make things worse? */
1144 return (imb > old_imb);
1148 * This checks if the overall compute and NUMA accesses of the system would
1149 * be improved if the source tasks was migrated to the target dst_cpu taking
1150 * into account that it might be best if task running on the dst_cpu should
1151 * be exchanged with the source task
1153 static void task_numa_compare(struct task_numa_env *env,
1154 long taskimp, long groupimp)
1156 struct rq *src_rq = cpu_rq(env->src_cpu);
1157 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1158 struct task_struct *cur;
1159 long src_load, dst_load;
1161 long imp = env->p->numa_group ? groupimp : taskimp;
1165 cur = ACCESS_ONCE(dst_rq->curr);
1166 if (cur->pid == 0) /* idle */
1170 * "imp" is the fault differential for the source task between the
1171 * source and destination node. Calculate the total differential for
1172 * the source task and potential destination task. The more negative
1173 * the value is, the more rmeote accesses that would be expected to
1174 * be incurred if the tasks were swapped.
1177 /* Skip this swap candidate if cannot move to the source cpu */
1178 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1182 * If dst and source tasks are in the same NUMA group, or not
1183 * in any group then look only at task weights.
1185 if (cur->numa_group == env->p->numa_group) {
1186 imp = taskimp + task_weight(cur, env->src_nid) -
1187 task_weight(cur, env->dst_nid);
1189 * Add some hysteresis to prevent swapping the
1190 * tasks within a group over tiny differences.
1192 if (cur->numa_group)
1196 * Compare the group weights. If a task is all by
1197 * itself (not part of a group), use the task weight
1200 if (cur->numa_group)
1201 imp += group_weight(cur, env->src_nid) -
1202 group_weight(cur, env->dst_nid);
1204 imp += task_weight(cur, env->src_nid) -
1205 task_weight(cur, env->dst_nid);
1209 if (imp <= env->best_imp && moveimp <= env->best_imp)
1213 /* Is there capacity at our destination? */
1214 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1215 !env->dst_stats.has_free_capacity)
1221 /* Balance doesn't matter much if we're running a task per cpu */
1222 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1223 dst_rq->nr_running == 1)
1227 * In the overloaded case, try and keep the load balanced.
1230 load = task_h_load(env->p);
1231 dst_load = env->dst_stats.load + load;
1232 src_load = env->src_stats.load - load;
1234 if (moveimp > imp && moveimp > env->best_imp) {
1236 * If the improvement from just moving env->p direction is
1237 * better than swapping tasks around, check if a move is
1238 * possible. Store a slightly smaller score than moveimp,
1239 * so an actually idle CPU will win.
1241 if (!load_too_imbalanced(src_load, dst_load, env)) {
1248 if (imp <= env->best_imp)
1252 load = task_h_load(cur);
1257 if (load_too_imbalanced(src_load, dst_load, env))
1261 task_numa_assign(env, cur, imp);
1266 static void task_numa_find_cpu(struct task_numa_env *env,
1267 long taskimp, long groupimp)
1271 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1272 /* Skip this CPU if the source task cannot migrate */
1273 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1277 task_numa_compare(env, taskimp, groupimp);
1281 static int task_numa_migrate(struct task_struct *p)
1283 struct task_numa_env env = {
1286 .src_cpu = task_cpu(p),
1287 .src_nid = task_node(p),
1289 .imbalance_pct = 112,
1295 struct sched_domain *sd;
1296 unsigned long taskweight, groupweight;
1298 long taskimp, groupimp;
1301 * Pick the lowest SD_NUMA domain, as that would have the smallest
1302 * imbalance and would be the first to start moving tasks about.
1304 * And we want to avoid any moving of tasks about, as that would create
1305 * random movement of tasks -- counter the numa conditions we're trying
1309 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1311 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1315 * Cpusets can break the scheduler domain tree into smaller
1316 * balance domains, some of which do not cross NUMA boundaries.
1317 * Tasks that are "trapped" in such domains cannot be migrated
1318 * elsewhere, so there is no point in (re)trying.
1320 if (unlikely(!sd)) {
1321 p->numa_preferred_nid = task_node(p);
1325 taskweight = task_weight(p, env.src_nid);
1326 groupweight = group_weight(p, env.src_nid);
1327 update_numa_stats(&env.src_stats, env.src_nid);
1328 env.dst_nid = p->numa_preferred_nid;
1329 taskimp = task_weight(p, env.dst_nid) - taskweight;
1330 groupimp = group_weight(p, env.dst_nid) - groupweight;
1331 update_numa_stats(&env.dst_stats, env.dst_nid);
1333 /* Try to find a spot on the preferred nid. */
1334 task_numa_find_cpu(&env, taskimp, groupimp);
1336 /* No space available on the preferred nid. Look elsewhere. */
1337 if (env.best_cpu == -1) {
1338 for_each_online_node(nid) {
1339 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1342 /* Only consider nodes where both task and groups benefit */
1343 taskimp = task_weight(p, nid) - taskweight;
1344 groupimp = group_weight(p, nid) - groupweight;
1345 if (taskimp < 0 && groupimp < 0)
1349 update_numa_stats(&env.dst_stats, env.dst_nid);
1350 task_numa_find_cpu(&env, taskimp, groupimp);
1355 * If the task is part of a workload that spans multiple NUMA nodes,
1356 * and is migrating into one of the workload's active nodes, remember
1357 * this node as the task's preferred numa node, so the workload can
1359 * A task that migrated to a second choice node will be better off
1360 * trying for a better one later. Do not set the preferred node here.
1362 if (p->numa_group) {
1363 if (env.best_cpu == -1)
1368 if (node_isset(nid, p->numa_group->active_nodes))
1369 sched_setnuma(p, env.dst_nid);
1372 /* No better CPU than the current one was found. */
1373 if (env.best_cpu == -1)
1377 * Reset the scan period if the task is being rescheduled on an
1378 * alternative node to recheck if the tasks is now properly placed.
1380 p->numa_scan_period = task_scan_min(p);
1382 if (env.best_task == NULL) {
1383 ret = migrate_task_to(p, env.best_cpu);
1385 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1389 ret = migrate_swap(p, env.best_task);
1391 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1392 put_task_struct(env.best_task);
1396 /* Attempt to migrate a task to a CPU on the preferred node. */
1397 static void numa_migrate_preferred(struct task_struct *p)
1399 unsigned long interval = HZ;
1401 /* This task has no NUMA fault statistics yet */
1402 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1405 /* Periodically retry migrating the task to the preferred node */
1406 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1407 p->numa_migrate_retry = jiffies + interval;
1409 /* Success if task is already running on preferred CPU */
1410 if (task_node(p) == p->numa_preferred_nid)
1413 /* Otherwise, try migrate to a CPU on the preferred node */
1414 task_numa_migrate(p);
1418 * Find the nodes on which the workload is actively running. We do this by
1419 * tracking the nodes from which NUMA hinting faults are triggered. This can
1420 * be different from the set of nodes where the workload's memory is currently
1423 * The bitmask is used to make smarter decisions on when to do NUMA page
1424 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1425 * are added when they cause over 6/16 of the maximum number of faults, but
1426 * only removed when they drop below 3/16.
1428 static void update_numa_active_node_mask(struct numa_group *numa_group)
1430 unsigned long faults, max_faults = 0;
1433 for_each_online_node(nid) {
1434 faults = group_faults_cpu(numa_group, nid);
1435 if (faults > max_faults)
1436 max_faults = faults;
1439 for_each_online_node(nid) {
1440 faults = group_faults_cpu(numa_group, nid);
1441 if (!node_isset(nid, numa_group->active_nodes)) {
1442 if (faults > max_faults * 6 / 16)
1443 node_set(nid, numa_group->active_nodes);
1444 } else if (faults < max_faults * 3 / 16)
1445 node_clear(nid, numa_group->active_nodes);
1450 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1451 * increments. The more local the fault statistics are, the higher the scan
1452 * period will be for the next scan window. If local/(local+remote) ratio is
1453 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1454 * the scan period will decrease. Aim for 70% local accesses.
1456 #define NUMA_PERIOD_SLOTS 10
1457 #define NUMA_PERIOD_THRESHOLD 7
1460 * Increase the scan period (slow down scanning) if the majority of
1461 * our memory is already on our local node, or if the majority of
1462 * the page accesses are shared with other processes.
1463 * Otherwise, decrease the scan period.
1465 static void update_task_scan_period(struct task_struct *p,
1466 unsigned long shared, unsigned long private)
1468 unsigned int period_slot;
1472 unsigned long remote = p->numa_faults_locality[0];
1473 unsigned long local = p->numa_faults_locality[1];
1476 * If there were no record hinting faults then either the task is
1477 * completely idle or all activity is areas that are not of interest
1478 * to automatic numa balancing. Scan slower
1480 if (local + shared == 0) {
1481 p->numa_scan_period = min(p->numa_scan_period_max,
1482 p->numa_scan_period << 1);
1484 p->mm->numa_next_scan = jiffies +
1485 msecs_to_jiffies(p->numa_scan_period);
1491 * Prepare to scale scan period relative to the current period.
1492 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1493 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1494 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1496 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1497 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1498 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1499 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1502 diff = slot * period_slot;
1504 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1507 * Scale scan rate increases based on sharing. There is an
1508 * inverse relationship between the degree of sharing and
1509 * the adjustment made to the scanning period. Broadly
1510 * speaking the intent is that there is little point
1511 * scanning faster if shared accesses dominate as it may
1512 * simply bounce migrations uselessly
1514 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1515 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1518 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1519 task_scan_min(p), task_scan_max(p));
1520 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1524 * Get the fraction of time the task has been running since the last
1525 * NUMA placement cycle. The scheduler keeps similar statistics, but
1526 * decays those on a 32ms period, which is orders of magnitude off
1527 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1528 * stats only if the task is so new there are no NUMA statistics yet.
1530 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1532 u64 runtime, delta, now;
1533 /* Use the start of this time slice to avoid calculations. */
1534 now = p->se.exec_start;
1535 runtime = p->se.sum_exec_runtime;
1537 if (p->last_task_numa_placement) {
1538 delta = runtime - p->last_sum_exec_runtime;
1539 *period = now - p->last_task_numa_placement;
1541 delta = p->se.avg.runnable_avg_sum;
1542 *period = p->se.avg.runnable_avg_period;
1545 p->last_sum_exec_runtime = runtime;
1546 p->last_task_numa_placement = now;
1551 static void task_numa_placement(struct task_struct *p)
1553 int seq, nid, max_nid = -1, max_group_nid = -1;
1554 unsigned long max_faults = 0, max_group_faults = 0;
1555 unsigned long fault_types[2] = { 0, 0 };
1556 unsigned long total_faults;
1557 u64 runtime, period;
1558 spinlock_t *group_lock = NULL;
1560 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1561 if (p->numa_scan_seq == seq)
1563 p->numa_scan_seq = seq;
1564 p->numa_scan_period_max = task_scan_max(p);
1566 total_faults = p->numa_faults_locality[0] +
1567 p->numa_faults_locality[1];
1568 runtime = numa_get_avg_runtime(p, &period);
1570 /* If the task is part of a group prevent parallel updates to group stats */
1571 if (p->numa_group) {
1572 group_lock = &p->numa_group->lock;
1573 spin_lock_irq(group_lock);
1576 /* Find the node with the highest number of faults */
1577 for_each_online_node(nid) {
1578 unsigned long faults = 0, group_faults = 0;
1581 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1582 long diff, f_diff, f_weight;
1584 i = task_faults_idx(nid, priv);
1586 /* Decay existing window, copy faults since last scan */
1587 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1588 fault_types[priv] += p->numa_faults_buffer_memory[i];
1589 p->numa_faults_buffer_memory[i] = 0;
1592 * Normalize the faults_from, so all tasks in a group
1593 * count according to CPU use, instead of by the raw
1594 * number of faults. Tasks with little runtime have
1595 * little over-all impact on throughput, and thus their
1596 * faults are less important.
1598 f_weight = div64_u64(runtime << 16, period + 1);
1599 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1601 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1602 p->numa_faults_buffer_cpu[i] = 0;
1604 p->numa_faults_memory[i] += diff;
1605 p->numa_faults_cpu[i] += f_diff;
1606 faults += p->numa_faults_memory[i];
1607 p->total_numa_faults += diff;
1608 if (p->numa_group) {
1609 /* safe because we can only change our own group */
1610 p->numa_group->faults[i] += diff;
1611 p->numa_group->faults_cpu[i] += f_diff;
1612 p->numa_group->total_faults += diff;
1613 group_faults += p->numa_group->faults[i];
1617 if (faults > max_faults) {
1618 max_faults = faults;
1622 if (group_faults > max_group_faults) {
1623 max_group_faults = group_faults;
1624 max_group_nid = nid;
1628 update_task_scan_period(p, fault_types[0], fault_types[1]);
1630 if (p->numa_group) {
1631 update_numa_active_node_mask(p->numa_group);
1632 spin_unlock_irq(group_lock);
1633 max_nid = max_group_nid;
1637 /* Set the new preferred node */
1638 if (max_nid != p->numa_preferred_nid)
1639 sched_setnuma(p, max_nid);
1641 if (task_node(p) != p->numa_preferred_nid)
1642 numa_migrate_preferred(p);
1646 static inline int get_numa_group(struct numa_group *grp)
1648 return atomic_inc_not_zero(&grp->refcount);
1651 static inline void put_numa_group(struct numa_group *grp)
1653 if (atomic_dec_and_test(&grp->refcount))
1654 kfree_rcu(grp, rcu);
1657 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1660 struct numa_group *grp, *my_grp;
1661 struct task_struct *tsk;
1663 int cpu = cpupid_to_cpu(cpupid);
1666 if (unlikely(!p->numa_group)) {
1667 unsigned int size = sizeof(struct numa_group) +
1668 4*nr_node_ids*sizeof(unsigned long);
1670 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1674 atomic_set(&grp->refcount, 1);
1675 spin_lock_init(&grp->lock);
1676 INIT_LIST_HEAD(&grp->task_list);
1678 /* Second half of the array tracks nids where faults happen */
1679 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1682 node_set(task_node(current), grp->active_nodes);
1684 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1685 grp->faults[i] = p->numa_faults_memory[i];
1687 grp->total_faults = p->total_numa_faults;
1689 list_add(&p->numa_entry, &grp->task_list);
1691 rcu_assign_pointer(p->numa_group, grp);
1695 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1697 if (!cpupid_match_pid(tsk, cpupid))
1700 grp = rcu_dereference(tsk->numa_group);
1704 my_grp = p->numa_group;
1709 * Only join the other group if its bigger; if we're the bigger group,
1710 * the other task will join us.
1712 if (my_grp->nr_tasks > grp->nr_tasks)
1716 * Tie-break on the grp address.
1718 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1721 /* Always join threads in the same process. */
1722 if (tsk->mm == current->mm)
1725 /* Simple filter to avoid false positives due to PID collisions */
1726 if (flags & TNF_SHARED)
1729 /* Update priv based on whether false sharing was detected */
1732 if (join && !get_numa_group(grp))
1740 BUG_ON(irqs_disabled());
1741 double_lock_irq(&my_grp->lock, &grp->lock);
1743 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1744 my_grp->faults[i] -= p->numa_faults_memory[i];
1745 grp->faults[i] += p->numa_faults_memory[i];
1747 my_grp->total_faults -= p->total_numa_faults;
1748 grp->total_faults += p->total_numa_faults;
1750 list_move(&p->numa_entry, &grp->task_list);
1754 spin_unlock(&my_grp->lock);
1755 spin_unlock_irq(&grp->lock);
1757 rcu_assign_pointer(p->numa_group, grp);
1759 put_numa_group(my_grp);
1767 void task_numa_free(struct task_struct *p)
1769 struct numa_group *grp = p->numa_group;
1770 void *numa_faults = p->numa_faults_memory;
1771 unsigned long flags;
1775 spin_lock_irqsave(&grp->lock, flags);
1776 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1777 grp->faults[i] -= p->numa_faults_memory[i];
1778 grp->total_faults -= p->total_numa_faults;
1780 list_del(&p->numa_entry);
1782 spin_unlock_irqrestore(&grp->lock, flags);
1783 RCU_INIT_POINTER(p->numa_group, NULL);
1784 put_numa_group(grp);
1787 p->numa_faults_memory = NULL;
1788 p->numa_faults_buffer_memory = NULL;
1789 p->numa_faults_cpu= NULL;
1790 p->numa_faults_buffer_cpu = NULL;
1795 * Got a PROT_NONE fault for a page on @node.
1797 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1799 struct task_struct *p = current;
1800 bool migrated = flags & TNF_MIGRATED;
1801 int cpu_node = task_node(current);
1802 int local = !!(flags & TNF_FAULT_LOCAL);
1805 if (!numabalancing_enabled)
1808 /* for example, ksmd faulting in a user's mm */
1812 /* Do not worry about placement if exiting */
1813 if (p->state == TASK_DEAD)
1816 /* Allocate buffer to track faults on a per-node basis */
1817 if (unlikely(!p->numa_faults_memory)) {
1818 int size = sizeof(*p->numa_faults_memory) *
1819 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1821 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1822 if (!p->numa_faults_memory)
1825 BUG_ON(p->numa_faults_buffer_memory);
1827 * The averaged statistics, shared & private, memory & cpu,
1828 * occupy the first half of the array. The second half of the
1829 * array is for current counters, which are averaged into the
1830 * first set by task_numa_placement.
1832 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1833 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1834 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1835 p->total_numa_faults = 0;
1836 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1840 * First accesses are treated as private, otherwise consider accesses
1841 * to be private if the accessing pid has not changed
1843 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1846 priv = cpupid_match_pid(p, last_cpupid);
1847 if (!priv && !(flags & TNF_NO_GROUP))
1848 task_numa_group(p, last_cpupid, flags, &priv);
1852 * If a workload spans multiple NUMA nodes, a shared fault that
1853 * occurs wholly within the set of nodes that the workload is
1854 * actively using should be counted as local. This allows the
1855 * scan rate to slow down when a workload has settled down.
1857 if (!priv && !local && p->numa_group &&
1858 node_isset(cpu_node, p->numa_group->active_nodes) &&
1859 node_isset(mem_node, p->numa_group->active_nodes))
1862 task_numa_placement(p);
1865 * Retry task to preferred node migration periodically, in case it
1866 * case it previously failed, or the scheduler moved us.
1868 if (time_after(jiffies, p->numa_migrate_retry))
1869 numa_migrate_preferred(p);
1872 p->numa_pages_migrated += pages;
1874 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1875 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1876 p->numa_faults_locality[local] += pages;
1879 static void reset_ptenuma_scan(struct task_struct *p)
1881 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1882 p->mm->numa_scan_offset = 0;
1886 * The expensive part of numa migration is done from task_work context.
1887 * Triggered from task_tick_numa().
1889 void task_numa_work(struct callback_head *work)
1891 unsigned long migrate, next_scan, now = jiffies;
1892 struct task_struct *p = current;
1893 struct mm_struct *mm = p->mm;
1894 struct vm_area_struct *vma;
1895 unsigned long start, end;
1896 unsigned long nr_pte_updates = 0;
1899 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1901 work->next = work; /* protect against double add */
1903 * Who cares about NUMA placement when they're dying.
1905 * NOTE: make sure not to dereference p->mm before this check,
1906 * exit_task_work() happens _after_ exit_mm() so we could be called
1907 * without p->mm even though we still had it when we enqueued this
1910 if (p->flags & PF_EXITING)
1913 if (!mm->numa_next_scan) {
1914 mm->numa_next_scan = now +
1915 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1919 * Enforce maximal scan/migration frequency..
1921 migrate = mm->numa_next_scan;
1922 if (time_before(now, migrate))
1925 if (p->numa_scan_period == 0) {
1926 p->numa_scan_period_max = task_scan_max(p);
1927 p->numa_scan_period = task_scan_min(p);
1930 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1931 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1935 * Delay this task enough that another task of this mm will likely win
1936 * the next time around.
1938 p->node_stamp += 2 * TICK_NSEC;
1940 start = mm->numa_scan_offset;
1941 pages = sysctl_numa_balancing_scan_size;
1942 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1946 down_read(&mm->mmap_sem);
1947 vma = find_vma(mm, start);
1949 reset_ptenuma_scan(p);
1953 for (; vma; vma = vma->vm_next) {
1954 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1958 * Shared library pages mapped by multiple processes are not
1959 * migrated as it is expected they are cache replicated. Avoid
1960 * hinting faults in read-only file-backed mappings or the vdso
1961 * as migrating the pages will be of marginal benefit.
1964 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1968 * Skip inaccessible VMAs to avoid any confusion between
1969 * PROT_NONE and NUMA hinting ptes
1971 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1975 start = max(start, vma->vm_start);
1976 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1977 end = min(end, vma->vm_end);
1978 nr_pte_updates += change_prot_numa(vma, start, end);
1981 * Scan sysctl_numa_balancing_scan_size but ensure that
1982 * at least one PTE is updated so that unused virtual
1983 * address space is quickly skipped.
1986 pages -= (end - start) >> PAGE_SHIFT;
1993 } while (end != vma->vm_end);
1998 * It is possible to reach the end of the VMA list but the last few
1999 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2000 * would find the !migratable VMA on the next scan but not reset the
2001 * scanner to the start so check it now.
2004 mm->numa_scan_offset = start;
2006 reset_ptenuma_scan(p);
2007 up_read(&mm->mmap_sem);
2011 * Drive the periodic memory faults..
2013 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2015 struct callback_head *work = &curr->numa_work;
2019 * We don't care about NUMA placement if we don't have memory.
2021 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2025 * Using runtime rather than walltime has the dual advantage that
2026 * we (mostly) drive the selection from busy threads and that the
2027 * task needs to have done some actual work before we bother with
2030 now = curr->se.sum_exec_runtime;
2031 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2033 if (now - curr->node_stamp > period) {
2034 if (!curr->node_stamp)
2035 curr->numa_scan_period = task_scan_min(curr);
2036 curr->node_stamp += period;
2038 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2039 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2040 task_work_add(curr, work, true);
2045 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2049 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2053 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2056 #endif /* CONFIG_NUMA_BALANCING */
2059 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2061 update_load_add(&cfs_rq->load, se->load.weight);
2062 if (!parent_entity(se))
2063 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2065 if (entity_is_task(se)) {
2066 struct rq *rq = rq_of(cfs_rq);
2068 account_numa_enqueue(rq, task_of(se));
2069 list_add(&se->group_node, &rq->cfs_tasks);
2072 cfs_rq->nr_running++;
2076 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2078 update_load_sub(&cfs_rq->load, se->load.weight);
2079 if (!parent_entity(se))
2080 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2081 if (entity_is_task(se)) {
2082 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2083 list_del_init(&se->group_node);
2085 cfs_rq->nr_running--;
2088 #ifdef CONFIG_FAIR_GROUP_SCHED
2090 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2095 * Use this CPU's actual weight instead of the last load_contribution
2096 * to gain a more accurate current total weight. See
2097 * update_cfs_rq_load_contribution().
2099 tg_weight = atomic_long_read(&tg->load_avg);
2100 tg_weight -= cfs_rq->tg_load_contrib;
2101 tg_weight += cfs_rq->load.weight;
2106 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2108 long tg_weight, load, shares;
2110 tg_weight = calc_tg_weight(tg, cfs_rq);
2111 load = cfs_rq->load.weight;
2113 shares = (tg->shares * load);
2115 shares /= tg_weight;
2117 if (shares < MIN_SHARES)
2118 shares = MIN_SHARES;
2119 if (shares > tg->shares)
2120 shares = tg->shares;
2124 # else /* CONFIG_SMP */
2125 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2129 # endif /* CONFIG_SMP */
2130 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2131 unsigned long weight)
2134 /* commit outstanding execution time */
2135 if (cfs_rq->curr == se)
2136 update_curr(cfs_rq);
2137 account_entity_dequeue(cfs_rq, se);
2140 update_load_set(&se->load, weight);
2143 account_entity_enqueue(cfs_rq, se);
2146 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2148 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2150 struct task_group *tg;
2151 struct sched_entity *se;
2155 se = tg->se[cpu_of(rq_of(cfs_rq))];
2156 if (!se || throttled_hierarchy(cfs_rq))
2159 if (likely(se->load.weight == tg->shares))
2162 shares = calc_cfs_shares(cfs_rq, tg);
2164 reweight_entity(cfs_rq_of(se), se, shares);
2166 #else /* CONFIG_FAIR_GROUP_SCHED */
2167 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2170 #endif /* CONFIG_FAIR_GROUP_SCHED */
2174 * We choose a half-life close to 1 scheduling period.
2175 * Note: The tables below are dependent on this value.
2177 #define LOAD_AVG_PERIOD 32
2178 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2179 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2181 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2182 static const u32 runnable_avg_yN_inv[] = {
2183 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2184 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2185 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2186 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2187 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2188 0x85aac367, 0x82cd8698,
2192 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2193 * over-estimates when re-combining.
2195 static const u32 runnable_avg_yN_sum[] = {
2196 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2197 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2198 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2203 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2205 static __always_inline u64 decay_load(u64 val, u64 n)
2207 unsigned int local_n;
2211 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2214 /* after bounds checking we can collapse to 32-bit */
2218 * As y^PERIOD = 1/2, we can combine
2219 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2220 * With a look-up table which covers k^n (n<PERIOD)
2222 * To achieve constant time decay_load.
2224 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2225 val >>= local_n / LOAD_AVG_PERIOD;
2226 local_n %= LOAD_AVG_PERIOD;
2229 val *= runnable_avg_yN_inv[local_n];
2230 /* We don't use SRR here since we always want to round down. */
2235 * For updates fully spanning n periods, the contribution to runnable
2236 * average will be: \Sum 1024*y^n
2238 * We can compute this reasonably efficiently by combining:
2239 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2241 static u32 __compute_runnable_contrib(u64 n)
2245 if (likely(n <= LOAD_AVG_PERIOD))
2246 return runnable_avg_yN_sum[n];
2247 else if (unlikely(n >= LOAD_AVG_MAX_N))
2248 return LOAD_AVG_MAX;
2250 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2252 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2253 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2255 n -= LOAD_AVG_PERIOD;
2256 } while (n > LOAD_AVG_PERIOD);
2258 contrib = decay_load(contrib, n);
2259 return contrib + runnable_avg_yN_sum[n];
2263 * We can represent the historical contribution to runnable average as the
2264 * coefficients of a geometric series. To do this we sub-divide our runnable
2265 * history into segments of approximately 1ms (1024us); label the segment that
2266 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2268 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2270 * (now) (~1ms ago) (~2ms ago)
2272 * Let u_i denote the fraction of p_i that the entity was runnable.
2274 * We then designate the fractions u_i as our co-efficients, yielding the
2275 * following representation of historical load:
2276 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2278 * We choose y based on the with of a reasonably scheduling period, fixing:
2281 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2282 * approximately half as much as the contribution to load within the last ms
2285 * When a period "rolls over" and we have new u_0`, multiplying the previous
2286 * sum again by y is sufficient to update:
2287 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2288 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2290 static __always_inline int __update_entity_runnable_avg(u64 now,
2291 struct sched_avg *sa,
2295 u32 runnable_contrib;
2296 int delta_w, decayed = 0;
2298 delta = now - sa->last_runnable_update;
2300 * This should only happen when time goes backwards, which it
2301 * unfortunately does during sched clock init when we swap over to TSC.
2303 if ((s64)delta < 0) {
2304 sa->last_runnable_update = now;
2309 * Use 1024ns as the unit of measurement since it's a reasonable
2310 * approximation of 1us and fast to compute.
2315 sa->last_runnable_update = now;
2317 /* delta_w is the amount already accumulated against our next period */
2318 delta_w = sa->runnable_avg_period % 1024;
2319 if (delta + delta_w >= 1024) {
2320 /* period roll-over */
2324 * Now that we know we're crossing a period boundary, figure
2325 * out how much from delta we need to complete the current
2326 * period and accrue it.
2328 delta_w = 1024 - delta_w;
2330 sa->runnable_avg_sum += delta_w;
2331 sa->runnable_avg_period += delta_w;
2335 /* Figure out how many additional periods this update spans */
2336 periods = delta / 1024;
2339 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2341 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2344 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2345 runnable_contrib = __compute_runnable_contrib(periods);
2347 sa->runnable_avg_sum += runnable_contrib;
2348 sa->runnable_avg_period += runnable_contrib;
2351 /* Remainder of delta accrued against u_0` */
2353 sa->runnable_avg_sum += delta;
2354 sa->runnable_avg_period += delta;
2359 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2360 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2362 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2363 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2365 decays -= se->avg.decay_count;
2369 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2370 se->avg.decay_count = 0;
2375 #ifdef CONFIG_FAIR_GROUP_SCHED
2376 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2379 struct task_group *tg = cfs_rq->tg;
2382 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2383 tg_contrib -= cfs_rq->tg_load_contrib;
2388 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2389 atomic_long_add(tg_contrib, &tg->load_avg);
2390 cfs_rq->tg_load_contrib += tg_contrib;
2395 * Aggregate cfs_rq runnable averages into an equivalent task_group
2396 * representation for computing load contributions.
2398 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2399 struct cfs_rq *cfs_rq)
2401 struct task_group *tg = cfs_rq->tg;
2404 /* The fraction of a cpu used by this cfs_rq */
2405 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2406 sa->runnable_avg_period + 1);
2407 contrib -= cfs_rq->tg_runnable_contrib;
2409 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2410 atomic_add(contrib, &tg->runnable_avg);
2411 cfs_rq->tg_runnable_contrib += contrib;
2415 static inline void __update_group_entity_contrib(struct sched_entity *se)
2417 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2418 struct task_group *tg = cfs_rq->tg;
2423 contrib = cfs_rq->tg_load_contrib * tg->shares;
2424 se->avg.load_avg_contrib = div_u64(contrib,
2425 atomic_long_read(&tg->load_avg) + 1);
2428 * For group entities we need to compute a correction term in the case
2429 * that they are consuming <1 cpu so that we would contribute the same
2430 * load as a task of equal weight.
2432 * Explicitly co-ordinating this measurement would be expensive, but
2433 * fortunately the sum of each cpus contribution forms a usable
2434 * lower-bound on the true value.
2436 * Consider the aggregate of 2 contributions. Either they are disjoint
2437 * (and the sum represents true value) or they are disjoint and we are
2438 * understating by the aggregate of their overlap.
2440 * Extending this to N cpus, for a given overlap, the maximum amount we
2441 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2442 * cpus that overlap for this interval and w_i is the interval width.
2444 * On a small machine; the first term is well-bounded which bounds the
2445 * total error since w_i is a subset of the period. Whereas on a
2446 * larger machine, while this first term can be larger, if w_i is the
2447 * of consequential size guaranteed to see n_i*w_i quickly converge to
2448 * our upper bound of 1-cpu.
2450 runnable_avg = atomic_read(&tg->runnable_avg);
2451 if (runnable_avg < NICE_0_LOAD) {
2452 se->avg.load_avg_contrib *= runnable_avg;
2453 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2457 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2459 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2460 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2462 #else /* CONFIG_FAIR_GROUP_SCHED */
2463 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2464 int force_update) {}
2465 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2466 struct cfs_rq *cfs_rq) {}
2467 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2468 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2469 #endif /* CONFIG_FAIR_GROUP_SCHED */
2471 static inline void __update_task_entity_contrib(struct sched_entity *se)
2475 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2476 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2477 contrib /= (se->avg.runnable_avg_period + 1);
2478 se->avg.load_avg_contrib = scale_load(contrib);
2481 /* Compute the current contribution to load_avg by se, return any delta */
2482 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2484 long old_contrib = se->avg.load_avg_contrib;
2486 if (entity_is_task(se)) {
2487 __update_task_entity_contrib(se);
2489 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2490 __update_group_entity_contrib(se);
2493 return se->avg.load_avg_contrib - old_contrib;
2496 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2499 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2500 cfs_rq->blocked_load_avg -= load_contrib;
2502 cfs_rq->blocked_load_avg = 0;
2505 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2507 /* Update a sched_entity's runnable average */
2508 static inline void update_entity_load_avg(struct sched_entity *se,
2511 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2516 * For a group entity we need to use their owned cfs_rq_clock_task() in
2517 * case they are the parent of a throttled hierarchy.
2519 if (entity_is_task(se))
2520 now = cfs_rq_clock_task(cfs_rq);
2522 now = cfs_rq_clock_task(group_cfs_rq(se));
2524 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2527 contrib_delta = __update_entity_load_avg_contrib(se);
2533 cfs_rq->runnable_load_avg += contrib_delta;
2535 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2539 * Decay the load contributed by all blocked children and account this so that
2540 * their contribution may appropriately discounted when they wake up.
2542 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2544 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2547 decays = now - cfs_rq->last_decay;
2548 if (!decays && !force_update)
2551 if (atomic_long_read(&cfs_rq->removed_load)) {
2552 unsigned long removed_load;
2553 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2554 subtract_blocked_load_contrib(cfs_rq, removed_load);
2558 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2560 atomic64_add(decays, &cfs_rq->decay_counter);
2561 cfs_rq->last_decay = now;
2564 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2567 /* Add the load generated by se into cfs_rq's child load-average */
2568 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2569 struct sched_entity *se,
2573 * We track migrations using entity decay_count <= 0, on a wake-up
2574 * migration we use a negative decay count to track the remote decays
2575 * accumulated while sleeping.
2577 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2578 * are seen by enqueue_entity_load_avg() as a migration with an already
2579 * constructed load_avg_contrib.
2581 if (unlikely(se->avg.decay_count <= 0)) {
2582 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2583 if (se->avg.decay_count) {
2585 * In a wake-up migration we have to approximate the
2586 * time sleeping. This is because we can't synchronize
2587 * clock_task between the two cpus, and it is not
2588 * guaranteed to be read-safe. Instead, we can
2589 * approximate this using our carried decays, which are
2590 * explicitly atomically readable.
2592 se->avg.last_runnable_update -= (-se->avg.decay_count)
2594 update_entity_load_avg(se, 0);
2595 /* Indicate that we're now synchronized and on-rq */
2596 se->avg.decay_count = 0;
2600 __synchronize_entity_decay(se);
2603 /* migrated tasks did not contribute to our blocked load */
2605 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2606 update_entity_load_avg(se, 0);
2609 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2610 /* we force update consideration on load-balancer moves */
2611 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2615 * Remove se's load from this cfs_rq child load-average, if the entity is
2616 * transitioning to a blocked state we track its projected decay using
2619 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2620 struct sched_entity *se,
2623 update_entity_load_avg(se, 1);
2624 /* we force update consideration on load-balancer moves */
2625 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2627 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2629 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2630 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2631 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2635 * Update the rq's load with the elapsed running time before entering
2636 * idle. if the last scheduled task is not a CFS task, idle_enter will
2637 * be the only way to update the runnable statistic.
2639 void idle_enter_fair(struct rq *this_rq)
2641 update_rq_runnable_avg(this_rq, 1);
2645 * Update the rq's load with the elapsed idle time before a task is
2646 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2647 * be the only way to update the runnable statistic.
2649 void idle_exit_fair(struct rq *this_rq)
2651 update_rq_runnable_avg(this_rq, 0);
2654 static int idle_balance(struct rq *this_rq);
2656 #else /* CONFIG_SMP */
2658 static inline void update_entity_load_avg(struct sched_entity *se,
2659 int update_cfs_rq) {}
2660 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2661 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2662 struct sched_entity *se,
2664 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2665 struct sched_entity *se,
2667 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2668 int force_update) {}
2670 static inline int idle_balance(struct rq *rq)
2675 #endif /* CONFIG_SMP */
2677 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2679 #ifdef CONFIG_SCHEDSTATS
2680 struct task_struct *tsk = NULL;
2682 if (entity_is_task(se))
2685 if (se->statistics.sleep_start) {
2686 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2691 if (unlikely(delta > se->statistics.sleep_max))
2692 se->statistics.sleep_max = delta;
2694 se->statistics.sleep_start = 0;
2695 se->statistics.sum_sleep_runtime += delta;
2698 account_scheduler_latency(tsk, delta >> 10, 1);
2699 trace_sched_stat_sleep(tsk, delta);
2702 if (se->statistics.block_start) {
2703 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2708 if (unlikely(delta > se->statistics.block_max))
2709 se->statistics.block_max = delta;
2711 se->statistics.block_start = 0;
2712 se->statistics.sum_sleep_runtime += delta;
2715 if (tsk->in_iowait) {
2716 se->statistics.iowait_sum += delta;
2717 se->statistics.iowait_count++;
2718 trace_sched_stat_iowait(tsk, delta);
2721 trace_sched_stat_blocked(tsk, delta);
2724 * Blocking time is in units of nanosecs, so shift by
2725 * 20 to get a milliseconds-range estimation of the
2726 * amount of time that the task spent sleeping:
2728 if (unlikely(prof_on == SLEEP_PROFILING)) {
2729 profile_hits(SLEEP_PROFILING,
2730 (void *)get_wchan(tsk),
2733 account_scheduler_latency(tsk, delta >> 10, 0);
2739 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2741 #ifdef CONFIG_SCHED_DEBUG
2742 s64 d = se->vruntime - cfs_rq->min_vruntime;
2747 if (d > 3*sysctl_sched_latency)
2748 schedstat_inc(cfs_rq, nr_spread_over);
2753 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2755 u64 vruntime = cfs_rq->min_vruntime;
2758 * The 'current' period is already promised to the current tasks,
2759 * however the extra weight of the new task will slow them down a
2760 * little, place the new task so that it fits in the slot that
2761 * stays open at the end.
2763 if (initial && sched_feat(START_DEBIT))
2764 vruntime += sched_vslice(cfs_rq, se);
2766 /* sleeps up to a single latency don't count. */
2768 unsigned long thresh = sysctl_sched_latency;
2771 * Halve their sleep time's effect, to allow
2772 * for a gentler effect of sleepers:
2774 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2780 /* ensure we never gain time by being placed backwards. */
2781 se->vruntime = max_vruntime(se->vruntime, vruntime);
2784 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2787 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2790 * Update the normalized vruntime before updating min_vruntime
2791 * through calling update_curr().
2793 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2794 se->vruntime += cfs_rq->min_vruntime;
2797 * Update run-time statistics of the 'current'.
2799 update_curr(cfs_rq);
2800 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2801 account_entity_enqueue(cfs_rq, se);
2802 update_cfs_shares(cfs_rq);
2804 if (flags & ENQUEUE_WAKEUP) {
2805 place_entity(cfs_rq, se, 0);
2806 enqueue_sleeper(cfs_rq, se);
2809 update_stats_enqueue(cfs_rq, se);
2810 check_spread(cfs_rq, se);
2811 if (se != cfs_rq->curr)
2812 __enqueue_entity(cfs_rq, se);
2815 if (cfs_rq->nr_running == 1) {
2816 list_add_leaf_cfs_rq(cfs_rq);
2817 check_enqueue_throttle(cfs_rq);
2821 static void __clear_buddies_last(struct sched_entity *se)
2823 for_each_sched_entity(se) {
2824 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2825 if (cfs_rq->last != se)
2828 cfs_rq->last = NULL;
2832 static void __clear_buddies_next(struct sched_entity *se)
2834 for_each_sched_entity(se) {
2835 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2836 if (cfs_rq->next != se)
2839 cfs_rq->next = NULL;
2843 static void __clear_buddies_skip(struct sched_entity *se)
2845 for_each_sched_entity(se) {
2846 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2847 if (cfs_rq->skip != se)
2850 cfs_rq->skip = NULL;
2854 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2856 if (cfs_rq->last == se)
2857 __clear_buddies_last(se);
2859 if (cfs_rq->next == se)
2860 __clear_buddies_next(se);
2862 if (cfs_rq->skip == se)
2863 __clear_buddies_skip(se);
2866 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2869 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2872 * Update run-time statistics of the 'current'.
2874 update_curr(cfs_rq);
2875 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2877 update_stats_dequeue(cfs_rq, se);
2878 if (flags & DEQUEUE_SLEEP) {
2879 #ifdef CONFIG_SCHEDSTATS
2880 if (entity_is_task(se)) {
2881 struct task_struct *tsk = task_of(se);
2883 if (tsk->state & TASK_INTERRUPTIBLE)
2884 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2885 if (tsk->state & TASK_UNINTERRUPTIBLE)
2886 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2891 clear_buddies(cfs_rq, se);
2893 if (se != cfs_rq->curr)
2894 __dequeue_entity(cfs_rq, se);
2896 account_entity_dequeue(cfs_rq, se);
2899 * Normalize the entity after updating the min_vruntime because the
2900 * update can refer to the ->curr item and we need to reflect this
2901 * movement in our normalized position.
2903 if (!(flags & DEQUEUE_SLEEP))
2904 se->vruntime -= cfs_rq->min_vruntime;
2906 /* return excess runtime on last dequeue */
2907 return_cfs_rq_runtime(cfs_rq);
2909 update_min_vruntime(cfs_rq);
2910 update_cfs_shares(cfs_rq);
2914 * Preempt the current task with a newly woken task if needed:
2917 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2919 unsigned long ideal_runtime, delta_exec;
2920 struct sched_entity *se;
2923 ideal_runtime = sched_slice(cfs_rq, curr);
2924 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2925 if (delta_exec > ideal_runtime) {
2926 resched_curr(rq_of(cfs_rq));
2928 * The current task ran long enough, ensure it doesn't get
2929 * re-elected due to buddy favours.
2931 clear_buddies(cfs_rq, curr);
2936 * Ensure that a task that missed wakeup preemption by a
2937 * narrow margin doesn't have to wait for a full slice.
2938 * This also mitigates buddy induced latencies under load.
2940 if (delta_exec < sysctl_sched_min_granularity)
2943 se = __pick_first_entity(cfs_rq);
2944 delta = curr->vruntime - se->vruntime;
2949 if (delta > ideal_runtime)
2950 resched_curr(rq_of(cfs_rq));
2954 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2956 /* 'current' is not kept within the tree. */
2959 * Any task has to be enqueued before it get to execute on
2960 * a CPU. So account for the time it spent waiting on the
2963 update_stats_wait_end(cfs_rq, se);
2964 __dequeue_entity(cfs_rq, se);
2967 update_stats_curr_start(cfs_rq, se);
2969 #ifdef CONFIG_SCHEDSTATS
2971 * Track our maximum slice length, if the CPU's load is at
2972 * least twice that of our own weight (i.e. dont track it
2973 * when there are only lesser-weight tasks around):
2975 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2976 se->statistics.slice_max = max(se->statistics.slice_max,
2977 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2980 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2984 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2987 * Pick the next process, keeping these things in mind, in this order:
2988 * 1) keep things fair between processes/task groups
2989 * 2) pick the "next" process, since someone really wants that to run
2990 * 3) pick the "last" process, for cache locality
2991 * 4) do not run the "skip" process, if something else is available
2993 static struct sched_entity *
2994 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2996 struct sched_entity *left = __pick_first_entity(cfs_rq);
2997 struct sched_entity *se;
3000 * If curr is set we have to see if its left of the leftmost entity
3001 * still in the tree, provided there was anything in the tree at all.
3003 if (!left || (curr && entity_before(curr, left)))
3006 se = left; /* ideally we run the leftmost entity */
3009 * Avoid running the skip buddy, if running something else can
3010 * be done without getting too unfair.
3012 if (cfs_rq->skip == se) {
3013 struct sched_entity *second;
3016 second = __pick_first_entity(cfs_rq);
3018 second = __pick_next_entity(se);
3019 if (!second || (curr && entity_before(curr, second)))
3023 if (second && wakeup_preempt_entity(second, left) < 1)
3028 * Prefer last buddy, try to return the CPU to a preempted task.
3030 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3034 * Someone really wants this to run. If it's not unfair, run it.
3036 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3039 clear_buddies(cfs_rq, se);
3044 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3046 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3049 * If still on the runqueue then deactivate_task()
3050 * was not called and update_curr() has to be done:
3053 update_curr(cfs_rq);
3055 /* throttle cfs_rqs exceeding runtime */
3056 check_cfs_rq_runtime(cfs_rq);
3058 check_spread(cfs_rq, prev);
3060 update_stats_wait_start(cfs_rq, prev);
3061 /* Put 'current' back into the tree. */
3062 __enqueue_entity(cfs_rq, prev);
3063 /* in !on_rq case, update occurred at dequeue */
3064 update_entity_load_avg(prev, 1);
3066 cfs_rq->curr = NULL;
3070 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3073 * Update run-time statistics of the 'current'.
3075 update_curr(cfs_rq);
3078 * Ensure that runnable average is periodically updated.
3080 update_entity_load_avg(curr, 1);
3081 update_cfs_rq_blocked_load(cfs_rq, 1);
3082 update_cfs_shares(cfs_rq);
3084 #ifdef CONFIG_SCHED_HRTICK
3086 * queued ticks are scheduled to match the slice, so don't bother
3087 * validating it and just reschedule.
3090 resched_curr(rq_of(cfs_rq));
3094 * don't let the period tick interfere with the hrtick preemption
3096 if (!sched_feat(DOUBLE_TICK) &&
3097 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3101 if (cfs_rq->nr_running > 1)
3102 check_preempt_tick(cfs_rq, curr);
3106 /**************************************************
3107 * CFS bandwidth control machinery
3110 #ifdef CONFIG_CFS_BANDWIDTH
3112 #ifdef HAVE_JUMP_LABEL
3113 static struct static_key __cfs_bandwidth_used;
3115 static inline bool cfs_bandwidth_used(void)
3117 return static_key_false(&__cfs_bandwidth_used);
3120 void cfs_bandwidth_usage_inc(void)
3122 static_key_slow_inc(&__cfs_bandwidth_used);
3125 void cfs_bandwidth_usage_dec(void)
3127 static_key_slow_dec(&__cfs_bandwidth_used);
3129 #else /* HAVE_JUMP_LABEL */
3130 static bool cfs_bandwidth_used(void)
3135 void cfs_bandwidth_usage_inc(void) {}
3136 void cfs_bandwidth_usage_dec(void) {}
3137 #endif /* HAVE_JUMP_LABEL */
3140 * default period for cfs group bandwidth.
3141 * default: 0.1s, units: nanoseconds
3143 static inline u64 default_cfs_period(void)
3145 return 100000000ULL;
3148 static inline u64 sched_cfs_bandwidth_slice(void)
3150 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3154 * Replenish runtime according to assigned quota and update expiration time.
3155 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3156 * additional synchronization around rq->lock.
3158 * requires cfs_b->lock
3160 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3164 if (cfs_b->quota == RUNTIME_INF)
3167 now = sched_clock_cpu(smp_processor_id());
3168 cfs_b->runtime = cfs_b->quota;
3169 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3172 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3174 return &tg->cfs_bandwidth;
3177 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3178 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3180 if (unlikely(cfs_rq->throttle_count))
3181 return cfs_rq->throttled_clock_task;
3183 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3186 /* returns 0 on failure to allocate runtime */
3187 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3189 struct task_group *tg = cfs_rq->tg;
3190 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3191 u64 amount = 0, min_amount, expires;
3193 /* note: this is a positive sum as runtime_remaining <= 0 */
3194 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3196 raw_spin_lock(&cfs_b->lock);
3197 if (cfs_b->quota == RUNTIME_INF)
3198 amount = min_amount;
3201 * If the bandwidth pool has become inactive, then at least one
3202 * period must have elapsed since the last consumption.
3203 * Refresh the global state and ensure bandwidth timer becomes
3206 if (!cfs_b->timer_active) {
3207 __refill_cfs_bandwidth_runtime(cfs_b);
3208 __start_cfs_bandwidth(cfs_b, false);
3211 if (cfs_b->runtime > 0) {
3212 amount = min(cfs_b->runtime, min_amount);
3213 cfs_b->runtime -= amount;
3217 expires = cfs_b->runtime_expires;
3218 raw_spin_unlock(&cfs_b->lock);
3220 cfs_rq->runtime_remaining += amount;
3222 * we may have advanced our local expiration to account for allowed
3223 * spread between our sched_clock and the one on which runtime was
3226 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3227 cfs_rq->runtime_expires = expires;
3229 return cfs_rq->runtime_remaining > 0;
3233 * Note: This depends on the synchronization provided by sched_clock and the
3234 * fact that rq->clock snapshots this value.
3236 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3238 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3240 /* if the deadline is ahead of our clock, nothing to do */
3241 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3244 if (cfs_rq->runtime_remaining < 0)
3248 * If the local deadline has passed we have to consider the
3249 * possibility that our sched_clock is 'fast' and the global deadline
3250 * has not truly expired.
3252 * Fortunately we can check determine whether this the case by checking
3253 * whether the global deadline has advanced. It is valid to compare
3254 * cfs_b->runtime_expires without any locks since we only care about
3255 * exact equality, so a partial write will still work.
3258 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3259 /* extend local deadline, drift is bounded above by 2 ticks */
3260 cfs_rq->runtime_expires += TICK_NSEC;
3262 /* global deadline is ahead, expiration has passed */
3263 cfs_rq->runtime_remaining = 0;
3267 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3269 /* dock delta_exec before expiring quota (as it could span periods) */
3270 cfs_rq->runtime_remaining -= delta_exec;
3271 expire_cfs_rq_runtime(cfs_rq);
3273 if (likely(cfs_rq->runtime_remaining > 0))
3277 * if we're unable to extend our runtime we resched so that the active
3278 * hierarchy can be throttled
3280 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3281 resched_curr(rq_of(cfs_rq));
3284 static __always_inline
3285 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3287 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3290 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3293 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3295 return cfs_bandwidth_used() && cfs_rq->throttled;
3298 /* check whether cfs_rq, or any parent, is throttled */
3299 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3301 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3305 * Ensure that neither of the group entities corresponding to src_cpu or
3306 * dest_cpu are members of a throttled hierarchy when performing group
3307 * load-balance operations.
3309 static inline int throttled_lb_pair(struct task_group *tg,
3310 int src_cpu, int dest_cpu)
3312 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3314 src_cfs_rq = tg->cfs_rq[src_cpu];
3315 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3317 return throttled_hierarchy(src_cfs_rq) ||
3318 throttled_hierarchy(dest_cfs_rq);
3321 /* updated child weight may affect parent so we have to do this bottom up */
3322 static int tg_unthrottle_up(struct task_group *tg, void *data)
3324 struct rq *rq = data;
3325 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3327 cfs_rq->throttle_count--;
3329 if (!cfs_rq->throttle_count) {
3330 /* adjust cfs_rq_clock_task() */
3331 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3332 cfs_rq->throttled_clock_task;
3339 static int tg_throttle_down(struct task_group *tg, void *data)
3341 struct rq *rq = data;
3342 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3344 /* group is entering throttled state, stop time */
3345 if (!cfs_rq->throttle_count)
3346 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3347 cfs_rq->throttle_count++;
3352 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3354 struct rq *rq = rq_of(cfs_rq);
3355 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3356 struct sched_entity *se;
3357 long task_delta, dequeue = 1;
3359 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3361 /* freeze hierarchy runnable averages while throttled */
3363 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3366 task_delta = cfs_rq->h_nr_running;
3367 for_each_sched_entity(se) {
3368 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3369 /* throttled entity or throttle-on-deactivate */
3374 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3375 qcfs_rq->h_nr_running -= task_delta;
3377 if (qcfs_rq->load.weight)
3382 sub_nr_running(rq, task_delta);
3384 cfs_rq->throttled = 1;
3385 cfs_rq->throttled_clock = rq_clock(rq);
3386 raw_spin_lock(&cfs_b->lock);
3388 * Add to the _head_ of the list, so that an already-started
3389 * distribute_cfs_runtime will not see us
3391 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3392 if (!cfs_b->timer_active)
3393 __start_cfs_bandwidth(cfs_b, false);
3394 raw_spin_unlock(&cfs_b->lock);
3397 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3399 struct rq *rq = rq_of(cfs_rq);
3400 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3401 struct sched_entity *se;
3405 se = cfs_rq->tg->se[cpu_of(rq)];
3407 cfs_rq->throttled = 0;
3409 update_rq_clock(rq);
3411 raw_spin_lock(&cfs_b->lock);
3412 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3413 list_del_rcu(&cfs_rq->throttled_list);
3414 raw_spin_unlock(&cfs_b->lock);
3416 /* update hierarchical throttle state */
3417 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3419 if (!cfs_rq->load.weight)
3422 task_delta = cfs_rq->h_nr_running;
3423 for_each_sched_entity(se) {
3427 cfs_rq = cfs_rq_of(se);
3429 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3430 cfs_rq->h_nr_running += task_delta;
3432 if (cfs_rq_throttled(cfs_rq))
3437 add_nr_running(rq, task_delta);
3439 /* determine whether we need to wake up potentially idle cpu */
3440 if (rq->curr == rq->idle && rq->cfs.nr_running)
3444 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3445 u64 remaining, u64 expires)
3447 struct cfs_rq *cfs_rq;
3449 u64 starting_runtime = remaining;
3452 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3454 struct rq *rq = rq_of(cfs_rq);
3456 raw_spin_lock(&rq->lock);
3457 if (!cfs_rq_throttled(cfs_rq))
3460 runtime = -cfs_rq->runtime_remaining + 1;
3461 if (runtime > remaining)
3462 runtime = remaining;
3463 remaining -= runtime;
3465 cfs_rq->runtime_remaining += runtime;
3466 cfs_rq->runtime_expires = expires;
3468 /* we check whether we're throttled above */
3469 if (cfs_rq->runtime_remaining > 0)
3470 unthrottle_cfs_rq(cfs_rq);
3473 raw_spin_unlock(&rq->lock);
3480 return starting_runtime - remaining;
3484 * Responsible for refilling a task_group's bandwidth and unthrottling its
3485 * cfs_rqs as appropriate. If there has been no activity within the last
3486 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3487 * used to track this state.
3489 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3491 u64 runtime, runtime_expires;
3494 /* no need to continue the timer with no bandwidth constraint */
3495 if (cfs_b->quota == RUNTIME_INF)
3496 goto out_deactivate;
3498 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3499 cfs_b->nr_periods += overrun;
3502 * idle depends on !throttled (for the case of a large deficit), and if
3503 * we're going inactive then everything else can be deferred
3505 if (cfs_b->idle && !throttled)
3506 goto out_deactivate;
3509 * if we have relooped after returning idle once, we need to update our
3510 * status as actually running, so that other cpus doing
3511 * __start_cfs_bandwidth will stop trying to cancel us.
3513 cfs_b->timer_active = 1;
3515 __refill_cfs_bandwidth_runtime(cfs_b);
3518 /* mark as potentially idle for the upcoming period */
3523 /* account preceding periods in which throttling occurred */
3524 cfs_b->nr_throttled += overrun;
3526 runtime_expires = cfs_b->runtime_expires;
3529 * This check is repeated as we are holding onto the new bandwidth while
3530 * we unthrottle. This can potentially race with an unthrottled group
3531 * trying to acquire new bandwidth from the global pool. This can result
3532 * in us over-using our runtime if it is all used during this loop, but
3533 * only by limited amounts in that extreme case.
3535 while (throttled && cfs_b->runtime > 0) {
3536 runtime = cfs_b->runtime;
3537 raw_spin_unlock(&cfs_b->lock);
3538 /* we can't nest cfs_b->lock while distributing bandwidth */
3539 runtime = distribute_cfs_runtime(cfs_b, runtime,
3541 raw_spin_lock(&cfs_b->lock);
3543 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3545 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3549 * While we are ensured activity in the period following an
3550 * unthrottle, this also covers the case in which the new bandwidth is
3551 * insufficient to cover the existing bandwidth deficit. (Forcing the
3552 * timer to remain active while there are any throttled entities.)
3559 cfs_b->timer_active = 0;
3563 /* a cfs_rq won't donate quota below this amount */
3564 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3565 /* minimum remaining period time to redistribute slack quota */
3566 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3567 /* how long we wait to gather additional slack before distributing */
3568 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3571 * Are we near the end of the current quota period?
3573 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3574 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3575 * migrate_hrtimers, base is never cleared, so we are fine.
3577 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3579 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3582 /* if the call-back is running a quota refresh is already occurring */
3583 if (hrtimer_callback_running(refresh_timer))
3586 /* is a quota refresh about to occur? */
3587 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3588 if (remaining < min_expire)
3594 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3596 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3598 /* if there's a quota refresh soon don't bother with slack */
3599 if (runtime_refresh_within(cfs_b, min_left))
3602 start_bandwidth_timer(&cfs_b->slack_timer,
3603 ns_to_ktime(cfs_bandwidth_slack_period));
3606 /* we know any runtime found here is valid as update_curr() precedes return */
3607 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3609 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3610 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3612 if (slack_runtime <= 0)
3615 raw_spin_lock(&cfs_b->lock);
3616 if (cfs_b->quota != RUNTIME_INF &&
3617 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3618 cfs_b->runtime += slack_runtime;
3620 /* we are under rq->lock, defer unthrottling using a timer */
3621 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3622 !list_empty(&cfs_b->throttled_cfs_rq))
3623 start_cfs_slack_bandwidth(cfs_b);
3625 raw_spin_unlock(&cfs_b->lock);
3627 /* even if it's not valid for return we don't want to try again */
3628 cfs_rq->runtime_remaining -= slack_runtime;
3631 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3633 if (!cfs_bandwidth_used())
3636 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3639 __return_cfs_rq_runtime(cfs_rq);
3643 * This is done with a timer (instead of inline with bandwidth return) since
3644 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3646 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3648 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3651 /* confirm we're still not at a refresh boundary */
3652 raw_spin_lock(&cfs_b->lock);
3653 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3654 raw_spin_unlock(&cfs_b->lock);
3658 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3659 runtime = cfs_b->runtime;
3661 expires = cfs_b->runtime_expires;
3662 raw_spin_unlock(&cfs_b->lock);
3667 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3669 raw_spin_lock(&cfs_b->lock);
3670 if (expires == cfs_b->runtime_expires)
3671 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3672 raw_spin_unlock(&cfs_b->lock);
3676 * When a group wakes up we want to make sure that its quota is not already
3677 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3678 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3680 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3682 if (!cfs_bandwidth_used())
3685 /* an active group must be handled by the update_curr()->put() path */
3686 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3689 /* ensure the group is not already throttled */
3690 if (cfs_rq_throttled(cfs_rq))
3693 /* update runtime allocation */
3694 account_cfs_rq_runtime(cfs_rq, 0);
3695 if (cfs_rq->runtime_remaining <= 0)
3696 throttle_cfs_rq(cfs_rq);
3699 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3700 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3702 if (!cfs_bandwidth_used())
3705 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3709 * it's possible for a throttled entity to be forced into a running
3710 * state (e.g. set_curr_task), in this case we're finished.
3712 if (cfs_rq_throttled(cfs_rq))
3715 throttle_cfs_rq(cfs_rq);
3719 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3721 struct cfs_bandwidth *cfs_b =
3722 container_of(timer, struct cfs_bandwidth, slack_timer);
3723 do_sched_cfs_slack_timer(cfs_b);
3725 return HRTIMER_NORESTART;
3728 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3730 struct cfs_bandwidth *cfs_b =
3731 container_of(timer, struct cfs_bandwidth, period_timer);
3736 raw_spin_lock(&cfs_b->lock);
3738 now = hrtimer_cb_get_time(timer);
3739 overrun = hrtimer_forward(timer, now, cfs_b->period);
3744 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3746 raw_spin_unlock(&cfs_b->lock);
3748 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3751 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3753 raw_spin_lock_init(&cfs_b->lock);
3755 cfs_b->quota = RUNTIME_INF;
3756 cfs_b->period = ns_to_ktime(default_cfs_period());
3758 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3759 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3760 cfs_b->period_timer.function = sched_cfs_period_timer;
3761 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3762 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3765 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3767 cfs_rq->runtime_enabled = 0;
3768 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3771 /* requires cfs_b->lock, may release to reprogram timer */
3772 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force)
3775 * The timer may be active because we're trying to set a new bandwidth
3776 * period or because we're racing with the tear-down path
3777 * (timer_active==0 becomes visible before the hrtimer call-back
3778 * terminates). In either case we ensure that it's re-programmed
3780 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3781 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3782 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3783 raw_spin_unlock(&cfs_b->lock);
3785 raw_spin_lock(&cfs_b->lock);
3786 /* if someone else restarted the timer then we're done */
3787 if (!force && cfs_b->timer_active)
3791 cfs_b->timer_active = 1;
3792 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3795 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3797 hrtimer_cancel(&cfs_b->period_timer);
3798 hrtimer_cancel(&cfs_b->slack_timer);
3801 static void __maybe_unused update_runtime_enabled(struct rq *rq)
3803 struct cfs_rq *cfs_rq;
3805 for_each_leaf_cfs_rq(rq, cfs_rq) {
3806 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
3808 raw_spin_lock(&cfs_b->lock);
3809 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
3810 raw_spin_unlock(&cfs_b->lock);
3814 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3816 struct cfs_rq *cfs_rq;
3818 for_each_leaf_cfs_rq(rq, cfs_rq) {
3819 if (!cfs_rq->runtime_enabled)
3823 * clock_task is not advancing so we just need to make sure
3824 * there's some valid quota amount
3826 cfs_rq->runtime_remaining = 1;
3828 * Offline rq is schedulable till cpu is completely disabled
3829 * in take_cpu_down(), so we prevent new cfs throttling here.
3831 cfs_rq->runtime_enabled = 0;
3833 if (cfs_rq_throttled(cfs_rq))
3834 unthrottle_cfs_rq(cfs_rq);
3838 #else /* CONFIG_CFS_BANDWIDTH */
3839 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3841 return rq_clock_task(rq_of(cfs_rq));
3844 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3845 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3846 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3847 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3849 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3854 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3859 static inline int throttled_lb_pair(struct task_group *tg,
3860 int src_cpu, int dest_cpu)
3865 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3867 #ifdef CONFIG_FAIR_GROUP_SCHED
3868 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3871 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3875 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3876 static inline void update_runtime_enabled(struct rq *rq) {}
3877 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3879 #endif /* CONFIG_CFS_BANDWIDTH */
3881 /**************************************************
3882 * CFS operations on tasks:
3885 #ifdef CONFIG_SCHED_HRTICK
3886 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3888 struct sched_entity *se = &p->se;
3889 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3891 WARN_ON(task_rq(p) != rq);
3893 if (cfs_rq->nr_running > 1) {
3894 u64 slice = sched_slice(cfs_rq, se);
3895 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3896 s64 delta = slice - ran;
3903 hrtick_start(rq, delta);
3908 * called from enqueue/dequeue and updates the hrtick when the
3909 * current task is from our class and nr_running is low enough
3912 static void hrtick_update(struct rq *rq)
3914 struct task_struct *curr = rq->curr;
3916 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3919 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3920 hrtick_start_fair(rq, curr);
3922 #else /* !CONFIG_SCHED_HRTICK */
3924 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3928 static inline void hrtick_update(struct rq *rq)
3934 * The enqueue_task method is called before nr_running is
3935 * increased. Here we update the fair scheduling stats and
3936 * then put the task into the rbtree:
3939 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3941 struct cfs_rq *cfs_rq;
3942 struct sched_entity *se = &p->se;
3944 for_each_sched_entity(se) {
3947 cfs_rq = cfs_rq_of(se);
3948 enqueue_entity(cfs_rq, se, flags);
3951 * end evaluation on encountering a throttled cfs_rq
3953 * note: in the case of encountering a throttled cfs_rq we will
3954 * post the final h_nr_running increment below.
3956 if (cfs_rq_throttled(cfs_rq))
3958 cfs_rq->h_nr_running++;
3960 flags = ENQUEUE_WAKEUP;
3963 for_each_sched_entity(se) {
3964 cfs_rq = cfs_rq_of(se);
3965 cfs_rq->h_nr_running++;
3967 if (cfs_rq_throttled(cfs_rq))
3970 update_cfs_shares(cfs_rq);
3971 update_entity_load_avg(se, 1);
3975 update_rq_runnable_avg(rq, rq->nr_running);
3976 add_nr_running(rq, 1);
3981 static void set_next_buddy(struct sched_entity *se);
3984 * The dequeue_task method is called before nr_running is
3985 * decreased. We remove the task from the rbtree and
3986 * update the fair scheduling stats:
3988 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3990 struct cfs_rq *cfs_rq;
3991 struct sched_entity *se = &p->se;
3992 int task_sleep = flags & DEQUEUE_SLEEP;
3994 for_each_sched_entity(se) {
3995 cfs_rq = cfs_rq_of(se);
3996 dequeue_entity(cfs_rq, se, flags);
3999 * end evaluation on encountering a throttled cfs_rq
4001 * note: in the case of encountering a throttled cfs_rq we will
4002 * post the final h_nr_running decrement below.
4004 if (cfs_rq_throttled(cfs_rq))
4006 cfs_rq->h_nr_running--;
4008 /* Don't dequeue parent if it has other entities besides us */
4009 if (cfs_rq->load.weight) {
4011 * Bias pick_next to pick a task from this cfs_rq, as
4012 * p is sleeping when it is within its sched_slice.
4014 if (task_sleep && parent_entity(se))
4015 set_next_buddy(parent_entity(se));
4017 /* avoid re-evaluating load for this entity */
4018 se = parent_entity(se);
4021 flags |= DEQUEUE_SLEEP;
4024 for_each_sched_entity(se) {
4025 cfs_rq = cfs_rq_of(se);
4026 cfs_rq->h_nr_running--;
4028 if (cfs_rq_throttled(cfs_rq))
4031 update_cfs_shares(cfs_rq);
4032 update_entity_load_avg(se, 1);
4036 sub_nr_running(rq, 1);
4037 update_rq_runnable_avg(rq, 1);
4043 /* Used instead of source_load when we know the type == 0 */
4044 static unsigned long weighted_cpuload(const int cpu)
4046 return cpu_rq(cpu)->cfs.runnable_load_avg;
4050 * Return a low guess at the load of a migration-source cpu weighted
4051 * according to the scheduling class and "nice" value.
4053 * We want to under-estimate the load of migration sources, to
4054 * balance conservatively.
4056 static unsigned long source_load(int cpu, int type)
4058 struct rq *rq = cpu_rq(cpu);
4059 unsigned long total = weighted_cpuload(cpu);
4061 if (type == 0 || !sched_feat(LB_BIAS))
4064 return min(rq->cpu_load[type-1], total);
4068 * Return a high guess at the load of a migration-target cpu weighted
4069 * according to the scheduling class and "nice" value.
4071 static unsigned long target_load(int cpu, int type)
4073 struct rq *rq = cpu_rq(cpu);
4074 unsigned long total = weighted_cpuload(cpu);
4076 if (type == 0 || !sched_feat(LB_BIAS))
4079 return max(rq->cpu_load[type-1], total);
4082 static unsigned long capacity_of(int cpu)
4084 return cpu_rq(cpu)->cpu_capacity;
4087 static unsigned long cpu_avg_load_per_task(int cpu)
4089 struct rq *rq = cpu_rq(cpu);
4090 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
4091 unsigned long load_avg = rq->cfs.runnable_load_avg;
4094 return load_avg / nr_running;
4099 static void record_wakee(struct task_struct *p)
4102 * Rough decay (wiping) for cost saving, don't worry
4103 * about the boundary, really active task won't care
4106 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4107 current->wakee_flips >>= 1;
4108 current->wakee_flip_decay_ts = jiffies;
4111 if (current->last_wakee != p) {
4112 current->last_wakee = p;
4113 current->wakee_flips++;
4117 static void task_waking_fair(struct task_struct *p)
4119 struct sched_entity *se = &p->se;
4120 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4123 #ifndef CONFIG_64BIT
4124 u64 min_vruntime_copy;
4127 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4129 min_vruntime = cfs_rq->min_vruntime;
4130 } while (min_vruntime != min_vruntime_copy);
4132 min_vruntime = cfs_rq->min_vruntime;
4135 se->vruntime -= min_vruntime;
4139 #ifdef CONFIG_FAIR_GROUP_SCHED
4141 * effective_load() calculates the load change as seen from the root_task_group
4143 * Adding load to a group doesn't make a group heavier, but can cause movement
4144 * of group shares between cpus. Assuming the shares were perfectly aligned one
4145 * can calculate the shift in shares.
4147 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4148 * on this @cpu and results in a total addition (subtraction) of @wg to the
4149 * total group weight.
4151 * Given a runqueue weight distribution (rw_i) we can compute a shares
4152 * distribution (s_i) using:
4154 * s_i = rw_i / \Sum rw_j (1)
4156 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4157 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4158 * shares distribution (s_i):
4160 * rw_i = { 2, 4, 1, 0 }
4161 * s_i = { 2/7, 4/7, 1/7, 0 }
4163 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4164 * task used to run on and the CPU the waker is running on), we need to
4165 * compute the effect of waking a task on either CPU and, in case of a sync
4166 * wakeup, compute the effect of the current task going to sleep.
4168 * So for a change of @wl to the local @cpu with an overall group weight change
4169 * of @wl we can compute the new shares distribution (s'_i) using:
4171 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4173 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4174 * differences in waking a task to CPU 0. The additional task changes the
4175 * weight and shares distributions like:
4177 * rw'_i = { 3, 4, 1, 0 }
4178 * s'_i = { 3/8, 4/8, 1/8, 0 }
4180 * We can then compute the difference in effective weight by using:
4182 * dw_i = S * (s'_i - s_i) (3)
4184 * Where 'S' is the group weight as seen by its parent.
4186 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4187 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4188 * 4/7) times the weight of the group.
4190 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4192 struct sched_entity *se = tg->se[cpu];
4194 if (!tg->parent) /* the trivial, non-cgroup case */
4197 for_each_sched_entity(se) {
4203 * W = @wg + \Sum rw_j
4205 W = wg + calc_tg_weight(tg, se->my_q);
4210 w = se->my_q->load.weight + wl;
4213 * wl = S * s'_i; see (2)
4216 wl = (w * tg->shares) / W;
4221 * Per the above, wl is the new se->load.weight value; since
4222 * those are clipped to [MIN_SHARES, ...) do so now. See
4223 * calc_cfs_shares().
4225 if (wl < MIN_SHARES)
4229 * wl = dw_i = S * (s'_i - s_i); see (3)
4231 wl -= se->load.weight;
4234 * Recursively apply this logic to all parent groups to compute
4235 * the final effective load change on the root group. Since
4236 * only the @tg group gets extra weight, all parent groups can
4237 * only redistribute existing shares. @wl is the shift in shares
4238 * resulting from this level per the above.
4247 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4254 static int wake_wide(struct task_struct *p)
4256 int factor = this_cpu_read(sd_llc_size);
4259 * Yeah, it's the switching-frequency, could means many wakee or
4260 * rapidly switch, use factor here will just help to automatically
4261 * adjust the loose-degree, so bigger node will lead to more pull.
4263 if (p->wakee_flips > factor) {
4265 * wakee is somewhat hot, it needs certain amount of cpu
4266 * resource, so if waker is far more hot, prefer to leave
4269 if (current->wakee_flips > (factor * p->wakee_flips))
4276 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4278 s64 this_load, load;
4279 int idx, this_cpu, prev_cpu;
4280 unsigned long tl_per_task;
4281 struct task_group *tg;
4282 unsigned long weight;
4286 * If we wake multiple tasks be careful to not bounce
4287 * ourselves around too much.
4293 this_cpu = smp_processor_id();
4294 prev_cpu = task_cpu(p);
4295 load = source_load(prev_cpu, idx);
4296 this_load = target_load(this_cpu, idx);
4299 * If sync wakeup then subtract the (maximum possible)
4300 * effect of the currently running task from the load
4301 * of the current CPU:
4304 tg = task_group(current);
4305 weight = current->se.load.weight;
4307 this_load += effective_load(tg, this_cpu, -weight, -weight);
4308 load += effective_load(tg, prev_cpu, 0, -weight);
4312 weight = p->se.load.weight;
4315 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4316 * due to the sync cause above having dropped this_load to 0, we'll
4317 * always have an imbalance, but there's really nothing you can do
4318 * about that, so that's good too.
4320 * Otherwise check if either cpus are near enough in load to allow this
4321 * task to be woken on this_cpu.
4323 if (this_load > 0) {
4324 s64 this_eff_load, prev_eff_load;
4326 this_eff_load = 100;
4327 this_eff_load *= capacity_of(prev_cpu);
4328 this_eff_load *= this_load +
4329 effective_load(tg, this_cpu, weight, weight);
4331 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4332 prev_eff_load *= capacity_of(this_cpu);
4333 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4335 balanced = this_eff_load <= prev_eff_load;
4340 * If the currently running task will sleep within
4341 * a reasonable amount of time then attract this newly
4344 if (sync && balanced)
4347 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4348 tl_per_task = cpu_avg_load_per_task(this_cpu);
4351 (this_load <= load &&
4352 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4354 * This domain has SD_WAKE_AFFINE and
4355 * p is cache cold in this domain, and
4356 * there is no bad imbalance.
4358 schedstat_inc(sd, ttwu_move_affine);
4359 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4367 * find_idlest_group finds and returns the least busy CPU group within the
4370 static struct sched_group *
4371 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4372 int this_cpu, int sd_flag)
4374 struct sched_group *idlest = NULL, *group = sd->groups;
4375 unsigned long min_load = ULONG_MAX, this_load = 0;
4376 int load_idx = sd->forkexec_idx;
4377 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4379 if (sd_flag & SD_BALANCE_WAKE)
4380 load_idx = sd->wake_idx;
4383 unsigned long load, avg_load;
4387 /* Skip over this group if it has no CPUs allowed */
4388 if (!cpumask_intersects(sched_group_cpus(group),
4389 tsk_cpus_allowed(p)))
4392 local_group = cpumask_test_cpu(this_cpu,
4393 sched_group_cpus(group));
4395 /* Tally up the load of all CPUs in the group */
4398 for_each_cpu(i, sched_group_cpus(group)) {
4399 /* Bias balancing toward cpus of our domain */
4401 load = source_load(i, load_idx);
4403 load = target_load(i, load_idx);
4408 /* Adjust by relative CPU capacity of the group */
4409 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4412 this_load = avg_load;
4413 } else if (avg_load < min_load) {
4414 min_load = avg_load;
4417 } while (group = group->next, group != sd->groups);
4419 if (!idlest || 100*this_load < imbalance*min_load)
4425 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4428 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4430 unsigned long load, min_load = ULONG_MAX;
4434 /* Traverse only the allowed CPUs */
4435 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4436 load = weighted_cpuload(i);
4438 if (load < min_load || (load == min_load && i == this_cpu)) {
4448 * Try and locate an idle CPU in the sched_domain.
4450 static int select_idle_sibling(struct task_struct *p, int target)
4452 struct sched_domain *sd;
4453 struct sched_group *sg;
4454 int i = task_cpu(p);
4456 if (idle_cpu(target))
4460 * If the prevous cpu is cache affine and idle, don't be stupid.
4462 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4466 * Otherwise, iterate the domains and find an elegible idle cpu.
4468 sd = rcu_dereference(per_cpu(sd_llc, target));
4469 for_each_lower_domain(sd) {
4472 if (!cpumask_intersects(sched_group_cpus(sg),
4473 tsk_cpus_allowed(p)))
4476 for_each_cpu(i, sched_group_cpus(sg)) {
4477 if (i == target || !idle_cpu(i))
4481 target = cpumask_first_and(sched_group_cpus(sg),
4482 tsk_cpus_allowed(p));
4486 } while (sg != sd->groups);
4493 * select_task_rq_fair: Select target runqueue for the waking task in domains
4494 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4495 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4497 * Balances load by selecting the idlest cpu in the idlest group, or under
4498 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4500 * Returns the target cpu number.
4502 * preempt must be disabled.
4505 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4507 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4508 int cpu = smp_processor_id();
4510 int want_affine = 0;
4511 int sync = wake_flags & WF_SYNC;
4513 if (p->nr_cpus_allowed == 1)
4516 if (sd_flag & SD_BALANCE_WAKE) {
4517 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4523 for_each_domain(cpu, tmp) {
4524 if (!(tmp->flags & SD_LOAD_BALANCE))
4528 * If both cpu and prev_cpu are part of this domain,
4529 * cpu is a valid SD_WAKE_AFFINE target.
4531 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4532 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4537 if (tmp->flags & sd_flag)
4541 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4544 if (sd_flag & SD_BALANCE_WAKE) {
4545 new_cpu = select_idle_sibling(p, prev_cpu);
4550 struct sched_group *group;
4553 if (!(sd->flags & sd_flag)) {
4558 group = find_idlest_group(sd, p, cpu, sd_flag);
4564 new_cpu = find_idlest_cpu(group, p, cpu);
4565 if (new_cpu == -1 || new_cpu == cpu) {
4566 /* Now try balancing at a lower domain level of cpu */
4571 /* Now try balancing at a lower domain level of new_cpu */
4573 weight = sd->span_weight;
4575 for_each_domain(cpu, tmp) {
4576 if (weight <= tmp->span_weight)
4578 if (tmp->flags & sd_flag)
4581 /* while loop will break here if sd == NULL */
4590 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4591 * cfs_rq_of(p) references at time of call are still valid and identify the
4592 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4593 * other assumptions, including the state of rq->lock, should be made.
4596 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4598 struct sched_entity *se = &p->se;
4599 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4602 * Load tracking: accumulate removed load so that it can be processed
4603 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4604 * to blocked load iff they have a positive decay-count. It can never
4605 * be negative here since on-rq tasks have decay-count == 0.
4607 if (se->avg.decay_count) {
4608 se->avg.decay_count = -__synchronize_entity_decay(se);
4609 atomic_long_add(se->avg.load_avg_contrib,
4610 &cfs_rq->removed_load);
4613 /* We have migrated, no longer consider this task hot */
4616 #endif /* CONFIG_SMP */
4618 static unsigned long
4619 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4621 unsigned long gran = sysctl_sched_wakeup_granularity;
4624 * Since its curr running now, convert the gran from real-time
4625 * to virtual-time in his units.
4627 * By using 'se' instead of 'curr' we penalize light tasks, so
4628 * they get preempted easier. That is, if 'se' < 'curr' then
4629 * the resulting gran will be larger, therefore penalizing the
4630 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4631 * be smaller, again penalizing the lighter task.
4633 * This is especially important for buddies when the leftmost
4634 * task is higher priority than the buddy.
4636 return calc_delta_fair(gran, se);
4640 * Should 'se' preempt 'curr'.
4654 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4656 s64 gran, vdiff = curr->vruntime - se->vruntime;
4661 gran = wakeup_gran(curr, se);
4668 static void set_last_buddy(struct sched_entity *se)
4670 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4673 for_each_sched_entity(se)
4674 cfs_rq_of(se)->last = se;
4677 static void set_next_buddy(struct sched_entity *se)
4679 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4682 for_each_sched_entity(se)
4683 cfs_rq_of(se)->next = se;
4686 static void set_skip_buddy(struct sched_entity *se)
4688 for_each_sched_entity(se)
4689 cfs_rq_of(se)->skip = se;
4693 * Preempt the current task with a newly woken task if needed:
4695 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4697 struct task_struct *curr = rq->curr;
4698 struct sched_entity *se = &curr->se, *pse = &p->se;
4699 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4700 int scale = cfs_rq->nr_running >= sched_nr_latency;
4701 int next_buddy_marked = 0;
4703 if (unlikely(se == pse))
4707 * This is possible from callers such as attach_tasks(), in which we
4708 * unconditionally check_prempt_curr() after an enqueue (which may have
4709 * lead to a throttle). This both saves work and prevents false
4710 * next-buddy nomination below.
4712 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4715 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4716 set_next_buddy(pse);
4717 next_buddy_marked = 1;
4721 * We can come here with TIF_NEED_RESCHED already set from new task
4724 * Note: this also catches the edge-case of curr being in a throttled
4725 * group (e.g. via set_curr_task), since update_curr() (in the
4726 * enqueue of curr) will have resulted in resched being set. This
4727 * prevents us from potentially nominating it as a false LAST_BUDDY
4730 if (test_tsk_need_resched(curr))
4733 /* Idle tasks are by definition preempted by non-idle tasks. */
4734 if (unlikely(curr->policy == SCHED_IDLE) &&
4735 likely(p->policy != SCHED_IDLE))
4739 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4740 * is driven by the tick):
4742 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4745 find_matching_se(&se, &pse);
4746 update_curr(cfs_rq_of(se));
4748 if (wakeup_preempt_entity(se, pse) == 1) {
4750 * Bias pick_next to pick the sched entity that is
4751 * triggering this preemption.
4753 if (!next_buddy_marked)
4754 set_next_buddy(pse);
4763 * Only set the backward buddy when the current task is still
4764 * on the rq. This can happen when a wakeup gets interleaved
4765 * with schedule on the ->pre_schedule() or idle_balance()
4766 * point, either of which can * drop the rq lock.
4768 * Also, during early boot the idle thread is in the fair class,
4769 * for obvious reasons its a bad idea to schedule back to it.
4771 if (unlikely(!se->on_rq || curr == rq->idle))
4774 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4778 static struct task_struct *
4779 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4781 struct cfs_rq *cfs_rq = &rq->cfs;
4782 struct sched_entity *se;
4783 struct task_struct *p;
4787 #ifdef CONFIG_FAIR_GROUP_SCHED
4788 if (!cfs_rq->nr_running)
4791 if (prev->sched_class != &fair_sched_class)
4795 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4796 * likely that a next task is from the same cgroup as the current.
4798 * Therefore attempt to avoid putting and setting the entire cgroup
4799 * hierarchy, only change the part that actually changes.
4803 struct sched_entity *curr = cfs_rq->curr;
4806 * Since we got here without doing put_prev_entity() we also
4807 * have to consider cfs_rq->curr. If it is still a runnable
4808 * entity, update_curr() will update its vruntime, otherwise
4809 * forget we've ever seen it.
4811 if (curr && curr->on_rq)
4812 update_curr(cfs_rq);
4817 * This call to check_cfs_rq_runtime() will do the throttle and
4818 * dequeue its entity in the parent(s). Therefore the 'simple'
4819 * nr_running test will indeed be correct.
4821 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4824 se = pick_next_entity(cfs_rq, curr);
4825 cfs_rq = group_cfs_rq(se);
4831 * Since we haven't yet done put_prev_entity and if the selected task
4832 * is a different task than we started out with, try and touch the
4833 * least amount of cfs_rqs.
4836 struct sched_entity *pse = &prev->se;
4838 while (!(cfs_rq = is_same_group(se, pse))) {
4839 int se_depth = se->depth;
4840 int pse_depth = pse->depth;
4842 if (se_depth <= pse_depth) {
4843 put_prev_entity(cfs_rq_of(pse), pse);
4844 pse = parent_entity(pse);
4846 if (se_depth >= pse_depth) {
4847 set_next_entity(cfs_rq_of(se), se);
4848 se = parent_entity(se);
4852 put_prev_entity(cfs_rq, pse);
4853 set_next_entity(cfs_rq, se);
4856 if (hrtick_enabled(rq))
4857 hrtick_start_fair(rq, p);
4864 if (!cfs_rq->nr_running)
4867 put_prev_task(rq, prev);
4870 se = pick_next_entity(cfs_rq, NULL);
4871 set_next_entity(cfs_rq, se);
4872 cfs_rq = group_cfs_rq(se);
4877 if (hrtick_enabled(rq))
4878 hrtick_start_fair(rq, p);
4883 new_tasks = idle_balance(rq);
4885 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4886 * possible for any higher priority task to appear. In that case we
4887 * must re-start the pick_next_entity() loop.
4899 * Account for a descheduled task:
4901 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4903 struct sched_entity *se = &prev->se;
4904 struct cfs_rq *cfs_rq;
4906 for_each_sched_entity(se) {
4907 cfs_rq = cfs_rq_of(se);
4908 put_prev_entity(cfs_rq, se);
4913 * sched_yield() is very simple
4915 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4917 static void yield_task_fair(struct rq *rq)
4919 struct task_struct *curr = rq->curr;
4920 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4921 struct sched_entity *se = &curr->se;
4924 * Are we the only task in the tree?
4926 if (unlikely(rq->nr_running == 1))
4929 clear_buddies(cfs_rq, se);
4931 if (curr->policy != SCHED_BATCH) {
4932 update_rq_clock(rq);
4934 * Update run-time statistics of the 'current'.
4936 update_curr(cfs_rq);
4938 * Tell update_rq_clock() that we've just updated,
4939 * so we don't do microscopic update in schedule()
4940 * and double the fastpath cost.
4942 rq->skip_clock_update = 1;
4948 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4950 struct sched_entity *se = &p->se;
4952 /* throttled hierarchies are not runnable */
4953 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4956 /* Tell the scheduler that we'd really like pse to run next. */
4959 yield_task_fair(rq);
4965 /**************************************************
4966 * Fair scheduling class load-balancing methods.
4970 * The purpose of load-balancing is to achieve the same basic fairness the
4971 * per-cpu scheduler provides, namely provide a proportional amount of compute
4972 * time to each task. This is expressed in the following equation:
4974 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4976 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4977 * W_i,0 is defined as:
4979 * W_i,0 = \Sum_j w_i,j (2)
4981 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4982 * is derived from the nice value as per prio_to_weight[].
4984 * The weight average is an exponential decay average of the instantaneous
4987 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4989 * C_i is the compute capacity of cpu i, typically it is the
4990 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4991 * can also include other factors [XXX].
4993 * To achieve this balance we define a measure of imbalance which follows
4994 * directly from (1):
4996 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
4998 * We them move tasks around to minimize the imbalance. In the continuous
4999 * function space it is obvious this converges, in the discrete case we get
5000 * a few fun cases generally called infeasible weight scenarios.
5003 * - infeasible weights;
5004 * - local vs global optima in the discrete case. ]
5009 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5010 * for all i,j solution, we create a tree of cpus that follows the hardware
5011 * topology where each level pairs two lower groups (or better). This results
5012 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5013 * tree to only the first of the previous level and we decrease the frequency
5014 * of load-balance at each level inv. proportional to the number of cpus in
5020 * \Sum { --- * --- * 2^i } = O(n) (5)
5022 * `- size of each group
5023 * | | `- number of cpus doing load-balance
5025 * `- sum over all levels
5027 * Coupled with a limit on how many tasks we can migrate every balance pass,
5028 * this makes (5) the runtime complexity of the balancer.
5030 * An important property here is that each CPU is still (indirectly) connected
5031 * to every other cpu in at most O(log n) steps:
5033 * The adjacency matrix of the resulting graph is given by:
5036 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5039 * And you'll find that:
5041 * A^(log_2 n)_i,j != 0 for all i,j (7)
5043 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5044 * The task movement gives a factor of O(m), giving a convergence complexity
5047 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5052 * In order to avoid CPUs going idle while there's still work to do, new idle
5053 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5054 * tree itself instead of relying on other CPUs to bring it work.
5056 * This adds some complexity to both (5) and (8) but it reduces the total idle
5064 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5067 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5072 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5074 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5076 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5079 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5080 * rewrite all of this once again.]
5083 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5085 enum fbq_type { regular, remote, all };
5087 #define LBF_ALL_PINNED 0x01
5088 #define LBF_NEED_BREAK 0x02
5089 #define LBF_DST_PINNED 0x04
5090 #define LBF_SOME_PINNED 0x08
5093 struct sched_domain *sd;
5101 struct cpumask *dst_grpmask;
5103 enum cpu_idle_type idle;
5105 /* The set of CPUs under consideration for load-balancing */
5106 struct cpumask *cpus;
5111 unsigned int loop_break;
5112 unsigned int loop_max;
5114 enum fbq_type fbq_type;
5115 struct list_head tasks;
5119 * Is this task likely cache-hot:
5121 static int task_hot(struct task_struct *p, struct lb_env *env)
5125 lockdep_assert_held(&env->src_rq->lock);
5127 if (p->sched_class != &fair_sched_class)
5130 if (unlikely(p->policy == SCHED_IDLE))
5134 * Buddy candidates are cache hot:
5136 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5137 (&p->se == cfs_rq_of(&p->se)->next ||
5138 &p->se == cfs_rq_of(&p->se)->last))
5141 if (sysctl_sched_migration_cost == -1)
5143 if (sysctl_sched_migration_cost == 0)
5146 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5148 return delta < (s64)sysctl_sched_migration_cost;
5151 #ifdef CONFIG_NUMA_BALANCING
5152 /* Returns true if the destination node has incurred more faults */
5153 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5155 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5156 int src_nid, dst_nid;
5158 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5159 !(env->sd->flags & SD_NUMA)) {
5163 src_nid = cpu_to_node(env->src_cpu);
5164 dst_nid = cpu_to_node(env->dst_cpu);
5166 if (src_nid == dst_nid)
5170 /* Task is already in the group's interleave set. */
5171 if (node_isset(src_nid, numa_group->active_nodes))
5174 /* Task is moving into the group's interleave set. */
5175 if (node_isset(dst_nid, numa_group->active_nodes))
5178 return group_faults(p, dst_nid) > group_faults(p, src_nid);
5181 /* Encourage migration to the preferred node. */
5182 if (dst_nid == p->numa_preferred_nid)
5185 return task_faults(p, dst_nid) > task_faults(p, src_nid);
5189 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5191 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5192 int src_nid, dst_nid;
5194 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5197 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5200 src_nid = cpu_to_node(env->src_cpu);
5201 dst_nid = cpu_to_node(env->dst_cpu);
5203 if (src_nid == dst_nid)
5207 /* Task is moving within/into the group's interleave set. */
5208 if (node_isset(dst_nid, numa_group->active_nodes))
5211 /* Task is moving out of the group's interleave set. */
5212 if (node_isset(src_nid, numa_group->active_nodes))
5215 return group_faults(p, dst_nid) < group_faults(p, src_nid);
5218 /* Migrating away from the preferred node is always bad. */
5219 if (src_nid == p->numa_preferred_nid)
5222 return task_faults(p, dst_nid) < task_faults(p, src_nid);
5226 static inline bool migrate_improves_locality(struct task_struct *p,
5232 static inline bool migrate_degrades_locality(struct task_struct *p,
5240 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5243 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5245 int tsk_cache_hot = 0;
5247 lockdep_assert_held(&env->src_rq->lock);
5250 * We do not migrate tasks that are:
5251 * 1) throttled_lb_pair, or
5252 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5253 * 3) running (obviously), or
5254 * 4) are cache-hot on their current CPU.
5256 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5259 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5262 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5264 env->flags |= LBF_SOME_PINNED;
5267 * Remember if this task can be migrated to any other cpu in
5268 * our sched_group. We may want to revisit it if we couldn't
5269 * meet load balance goals by pulling other tasks on src_cpu.
5271 * Also avoid computing new_dst_cpu if we have already computed
5272 * one in current iteration.
5274 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5277 /* Prevent to re-select dst_cpu via env's cpus */
5278 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5279 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5280 env->flags |= LBF_DST_PINNED;
5281 env->new_dst_cpu = cpu;
5289 /* Record that we found atleast one task that could run on dst_cpu */
5290 env->flags &= ~LBF_ALL_PINNED;
5292 if (task_running(env->src_rq, p)) {
5293 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5298 * Aggressive migration if:
5299 * 1) destination numa is preferred
5300 * 2) task is cache cold, or
5301 * 3) too many balance attempts have failed.
5303 tsk_cache_hot = task_hot(p, env);
5305 tsk_cache_hot = migrate_degrades_locality(p, env);
5307 if (migrate_improves_locality(p, env)) {
5308 #ifdef CONFIG_SCHEDSTATS
5309 if (tsk_cache_hot) {
5310 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5311 schedstat_inc(p, se.statistics.nr_forced_migrations);
5317 if (!tsk_cache_hot ||
5318 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5320 if (tsk_cache_hot) {
5321 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5322 schedstat_inc(p, se.statistics.nr_forced_migrations);
5328 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5333 * detach_task() -- detach the task for the migration specified in env
5335 static void detach_task(struct task_struct *p, struct lb_env *env)
5337 lockdep_assert_held(&env->src_rq->lock);
5339 deactivate_task(env->src_rq, p, 0);
5340 p->on_rq = TASK_ON_RQ_MIGRATING;
5341 set_task_cpu(p, env->dst_cpu);
5345 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5346 * part of active balancing operations within "domain".
5348 * Returns a task if successful and NULL otherwise.
5350 static struct task_struct *detach_one_task(struct lb_env *env)
5352 struct task_struct *p, *n;
5354 lockdep_assert_held(&env->src_rq->lock);
5356 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5357 if (!can_migrate_task(p, env))
5360 detach_task(p, env);
5363 * Right now, this is only the second place where
5364 * lb_gained[env->idle] is updated (other is detach_tasks)
5365 * so we can safely collect stats here rather than
5366 * inside detach_tasks().
5368 schedstat_inc(env->sd, lb_gained[env->idle]);
5374 static const unsigned int sched_nr_migrate_break = 32;
5377 * detach_tasks() -- tries to detach up to imbalance weighted load from
5378 * busiest_rq, as part of a balancing operation within domain "sd".
5380 * Returns number of detached tasks if successful and 0 otherwise.
5382 static int detach_tasks(struct lb_env *env)
5384 struct list_head *tasks = &env->src_rq->cfs_tasks;
5385 struct task_struct *p;
5389 lockdep_assert_held(&env->src_rq->lock);
5391 if (env->imbalance <= 0)
5394 while (!list_empty(tasks)) {
5395 p = list_first_entry(tasks, struct task_struct, se.group_node);
5398 /* We've more or less seen every task there is, call it quits */
5399 if (env->loop > env->loop_max)
5402 /* take a breather every nr_migrate tasks */
5403 if (env->loop > env->loop_break) {
5404 env->loop_break += sched_nr_migrate_break;
5405 env->flags |= LBF_NEED_BREAK;
5409 if (!can_migrate_task(p, env))
5412 load = task_h_load(p);
5414 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5417 if ((load / 2) > env->imbalance)
5420 detach_task(p, env);
5421 list_add(&p->se.group_node, &env->tasks);
5424 env->imbalance -= load;
5426 #ifdef CONFIG_PREEMPT
5428 * NEWIDLE balancing is a source of latency, so preemptible
5429 * kernels will stop after the first task is detached to minimize
5430 * the critical section.
5432 if (env->idle == CPU_NEWLY_IDLE)
5437 * We only want to steal up to the prescribed amount of
5440 if (env->imbalance <= 0)
5445 list_move_tail(&p->se.group_node, tasks);
5449 * Right now, this is one of only two places we collect this stat
5450 * so we can safely collect detach_one_task() stats here rather
5451 * than inside detach_one_task().
5453 schedstat_add(env->sd, lb_gained[env->idle], detached);
5459 * attach_task() -- attach the task detached by detach_task() to its new rq.
5461 static void attach_task(struct rq *rq, struct task_struct *p)
5463 lockdep_assert_held(&rq->lock);
5465 BUG_ON(task_rq(p) != rq);
5466 p->on_rq = TASK_ON_RQ_QUEUED;
5467 activate_task(rq, p, 0);
5468 check_preempt_curr(rq, p, 0);
5472 * attach_one_task() -- attaches the task returned from detach_one_task() to
5475 static void attach_one_task(struct rq *rq, struct task_struct *p)
5477 raw_spin_lock(&rq->lock);
5479 raw_spin_unlock(&rq->lock);
5483 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5486 static void attach_tasks(struct lb_env *env)
5488 struct list_head *tasks = &env->tasks;
5489 struct task_struct *p;
5491 raw_spin_lock(&env->dst_rq->lock);
5493 while (!list_empty(tasks)) {
5494 p = list_first_entry(tasks, struct task_struct, se.group_node);
5495 list_del_init(&p->se.group_node);
5497 attach_task(env->dst_rq, p);
5500 raw_spin_unlock(&env->dst_rq->lock);
5503 #ifdef CONFIG_FAIR_GROUP_SCHED
5505 * update tg->load_weight by folding this cpu's load_avg
5507 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5509 struct sched_entity *se = tg->se[cpu];
5510 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5512 /* throttled entities do not contribute to load */
5513 if (throttled_hierarchy(cfs_rq))
5516 update_cfs_rq_blocked_load(cfs_rq, 1);
5519 update_entity_load_avg(se, 1);
5521 * We pivot on our runnable average having decayed to zero for
5522 * list removal. This generally implies that all our children
5523 * have also been removed (modulo rounding error or bandwidth
5524 * control); however, such cases are rare and we can fix these
5527 * TODO: fix up out-of-order children on enqueue.
5529 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5530 list_del_leaf_cfs_rq(cfs_rq);
5532 struct rq *rq = rq_of(cfs_rq);
5533 update_rq_runnable_avg(rq, rq->nr_running);
5537 static void update_blocked_averages(int cpu)
5539 struct rq *rq = cpu_rq(cpu);
5540 struct cfs_rq *cfs_rq;
5541 unsigned long flags;
5543 raw_spin_lock_irqsave(&rq->lock, flags);
5544 update_rq_clock(rq);
5546 * Iterates the task_group tree in a bottom up fashion, see
5547 * list_add_leaf_cfs_rq() for details.
5549 for_each_leaf_cfs_rq(rq, cfs_rq) {
5551 * Note: We may want to consider periodically releasing
5552 * rq->lock about these updates so that creating many task
5553 * groups does not result in continually extending hold time.
5555 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5558 raw_spin_unlock_irqrestore(&rq->lock, flags);
5562 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5563 * This needs to be done in a top-down fashion because the load of a child
5564 * group is a fraction of its parents load.
5566 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5568 struct rq *rq = rq_of(cfs_rq);
5569 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5570 unsigned long now = jiffies;
5573 if (cfs_rq->last_h_load_update == now)
5576 cfs_rq->h_load_next = NULL;
5577 for_each_sched_entity(se) {
5578 cfs_rq = cfs_rq_of(se);
5579 cfs_rq->h_load_next = se;
5580 if (cfs_rq->last_h_load_update == now)
5585 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5586 cfs_rq->last_h_load_update = now;
5589 while ((se = cfs_rq->h_load_next) != NULL) {
5590 load = cfs_rq->h_load;
5591 load = div64_ul(load * se->avg.load_avg_contrib,
5592 cfs_rq->runnable_load_avg + 1);
5593 cfs_rq = group_cfs_rq(se);
5594 cfs_rq->h_load = load;
5595 cfs_rq->last_h_load_update = now;
5599 static unsigned long task_h_load(struct task_struct *p)
5601 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5603 update_cfs_rq_h_load(cfs_rq);
5604 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5605 cfs_rq->runnable_load_avg + 1);
5608 static inline void update_blocked_averages(int cpu)
5612 static unsigned long task_h_load(struct task_struct *p)
5614 return p->se.avg.load_avg_contrib;
5618 /********** Helpers for find_busiest_group ************************/
5627 * sg_lb_stats - stats of a sched_group required for load_balancing
5629 struct sg_lb_stats {
5630 unsigned long avg_load; /*Avg load across the CPUs of the group */
5631 unsigned long group_load; /* Total load over the CPUs of the group */
5632 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5633 unsigned long load_per_task;
5634 unsigned long group_capacity;
5635 unsigned int sum_nr_running; /* Nr tasks running in the group */
5636 unsigned int group_capacity_factor;
5637 unsigned int idle_cpus;
5638 unsigned int group_weight;
5639 enum group_type group_type;
5640 int group_has_free_capacity;
5641 #ifdef CONFIG_NUMA_BALANCING
5642 unsigned int nr_numa_running;
5643 unsigned int nr_preferred_running;
5648 * sd_lb_stats - Structure to store the statistics of a sched_domain
5649 * during load balancing.
5651 struct sd_lb_stats {
5652 struct sched_group *busiest; /* Busiest group in this sd */
5653 struct sched_group *local; /* Local group in this sd */
5654 unsigned long total_load; /* Total load of all groups in sd */
5655 unsigned long total_capacity; /* Total capacity of all groups in sd */
5656 unsigned long avg_load; /* Average load across all groups in sd */
5658 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5659 struct sg_lb_stats local_stat; /* Statistics of the local group */
5662 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5665 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5666 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5667 * We must however clear busiest_stat::avg_load because
5668 * update_sd_pick_busiest() reads this before assignment.
5670 *sds = (struct sd_lb_stats){
5674 .total_capacity = 0UL,
5677 .sum_nr_running = 0,
5678 .group_type = group_other,
5684 * get_sd_load_idx - Obtain the load index for a given sched domain.
5685 * @sd: The sched_domain whose load_idx is to be obtained.
5686 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5688 * Return: The load index.
5690 static inline int get_sd_load_idx(struct sched_domain *sd,
5691 enum cpu_idle_type idle)
5697 load_idx = sd->busy_idx;
5700 case CPU_NEWLY_IDLE:
5701 load_idx = sd->newidle_idx;
5704 load_idx = sd->idle_idx;
5711 static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu)
5713 return SCHED_CAPACITY_SCALE;
5716 unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu)
5718 return default_scale_capacity(sd, cpu);
5721 static unsigned long default_scale_smt_capacity(struct sched_domain *sd, int cpu)
5723 unsigned long weight = sd->span_weight;
5724 unsigned long smt_gain = sd->smt_gain;
5731 unsigned long __weak arch_scale_smt_capacity(struct sched_domain *sd, int cpu)
5733 return default_scale_smt_capacity(sd, cpu);
5736 static unsigned long scale_rt_capacity(int cpu)
5738 struct rq *rq = cpu_rq(cpu);
5739 u64 total, available, age_stamp, avg;
5743 * Since we're reading these variables without serialization make sure
5744 * we read them once before doing sanity checks on them.
5746 age_stamp = ACCESS_ONCE(rq->age_stamp);
5747 avg = ACCESS_ONCE(rq->rt_avg);
5749 delta = rq_clock(rq) - age_stamp;
5750 if (unlikely(delta < 0))
5753 total = sched_avg_period() + delta;
5755 if (unlikely(total < avg)) {
5756 /* Ensures that capacity won't end up being negative */
5759 available = total - avg;
5762 if (unlikely((s64)total < SCHED_CAPACITY_SCALE))
5763 total = SCHED_CAPACITY_SCALE;
5765 total >>= SCHED_CAPACITY_SHIFT;
5767 return div_u64(available, total);
5770 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
5772 unsigned long weight = sd->span_weight;
5773 unsigned long capacity = SCHED_CAPACITY_SCALE;
5774 struct sched_group *sdg = sd->groups;
5776 if ((sd->flags & SD_SHARE_CPUCAPACITY) && weight > 1) {
5777 if (sched_feat(ARCH_CAPACITY))
5778 capacity *= arch_scale_smt_capacity(sd, cpu);
5780 capacity *= default_scale_smt_capacity(sd, cpu);
5782 capacity >>= SCHED_CAPACITY_SHIFT;
5785 sdg->sgc->capacity_orig = capacity;
5787 if (sched_feat(ARCH_CAPACITY))
5788 capacity *= arch_scale_freq_capacity(sd, cpu);
5790 capacity *= default_scale_capacity(sd, cpu);
5792 capacity >>= SCHED_CAPACITY_SHIFT;
5794 capacity *= scale_rt_capacity(cpu);
5795 capacity >>= SCHED_CAPACITY_SHIFT;
5800 cpu_rq(cpu)->cpu_capacity = capacity;
5801 sdg->sgc->capacity = capacity;
5804 void update_group_capacity(struct sched_domain *sd, int cpu)
5806 struct sched_domain *child = sd->child;
5807 struct sched_group *group, *sdg = sd->groups;
5808 unsigned long capacity, capacity_orig;
5809 unsigned long interval;
5811 interval = msecs_to_jiffies(sd->balance_interval);
5812 interval = clamp(interval, 1UL, max_load_balance_interval);
5813 sdg->sgc->next_update = jiffies + interval;
5816 update_cpu_capacity(sd, cpu);
5820 capacity_orig = capacity = 0;
5822 if (child->flags & SD_OVERLAP) {
5824 * SD_OVERLAP domains cannot assume that child groups
5825 * span the current group.
5828 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5829 struct sched_group_capacity *sgc;
5830 struct rq *rq = cpu_rq(cpu);
5833 * build_sched_domains() -> init_sched_groups_capacity()
5834 * gets here before we've attached the domains to the
5837 * Use capacity_of(), which is set irrespective of domains
5838 * in update_cpu_capacity().
5840 * This avoids capacity/capacity_orig from being 0 and
5841 * causing divide-by-zero issues on boot.
5843 * Runtime updates will correct capacity_orig.
5845 if (unlikely(!rq->sd)) {
5846 capacity_orig += capacity_of(cpu);
5847 capacity += capacity_of(cpu);
5851 sgc = rq->sd->groups->sgc;
5852 capacity_orig += sgc->capacity_orig;
5853 capacity += sgc->capacity;
5857 * !SD_OVERLAP domains can assume that child groups
5858 * span the current group.
5861 group = child->groups;
5863 capacity_orig += group->sgc->capacity_orig;
5864 capacity += group->sgc->capacity;
5865 group = group->next;
5866 } while (group != child->groups);
5869 sdg->sgc->capacity_orig = capacity_orig;
5870 sdg->sgc->capacity = capacity;
5874 * Try and fix up capacity for tiny siblings, this is needed when
5875 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5876 * which on its own isn't powerful enough.
5878 * See update_sd_pick_busiest() and check_asym_packing().
5881 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5884 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5886 if (!(sd->flags & SD_SHARE_CPUCAPACITY))
5890 * If ~90% of the cpu_capacity is still there, we're good.
5892 if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29)
5899 * Group imbalance indicates (and tries to solve) the problem where balancing
5900 * groups is inadequate due to tsk_cpus_allowed() constraints.
5902 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5903 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5906 * { 0 1 2 3 } { 4 5 6 7 }
5909 * If we were to balance group-wise we'd place two tasks in the first group and
5910 * two tasks in the second group. Clearly this is undesired as it will overload
5911 * cpu 3 and leave one of the cpus in the second group unused.
5913 * The current solution to this issue is detecting the skew in the first group
5914 * by noticing the lower domain failed to reach balance and had difficulty
5915 * moving tasks due to affinity constraints.
5917 * When this is so detected; this group becomes a candidate for busiest; see
5918 * update_sd_pick_busiest(). And calculate_imbalance() and
5919 * find_busiest_group() avoid some of the usual balance conditions to allow it
5920 * to create an effective group imbalance.
5922 * This is a somewhat tricky proposition since the next run might not find the
5923 * group imbalance and decide the groups need to be balanced again. A most
5924 * subtle and fragile situation.
5927 static inline int sg_imbalanced(struct sched_group *group)
5929 return group->sgc->imbalance;
5933 * Compute the group capacity factor.
5935 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5936 * first dividing out the smt factor and computing the actual number of cores
5937 * and limit unit capacity with that.
5939 static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group)
5941 unsigned int capacity_factor, smt, cpus;
5942 unsigned int capacity, capacity_orig;
5944 capacity = group->sgc->capacity;
5945 capacity_orig = group->sgc->capacity_orig;
5946 cpus = group->group_weight;
5948 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5949 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig);
5950 capacity_factor = cpus / smt; /* cores */
5952 capacity_factor = min_t(unsigned,
5953 capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE));
5954 if (!capacity_factor)
5955 capacity_factor = fix_small_capacity(env->sd, group);
5957 return capacity_factor;
5960 static enum group_type
5961 group_classify(struct sched_group *group, struct sg_lb_stats *sgs)
5963 if (sgs->sum_nr_running > sgs->group_capacity_factor)
5964 return group_overloaded;
5966 if (sg_imbalanced(group))
5967 return group_imbalanced;
5973 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5974 * @env: The load balancing environment.
5975 * @group: sched_group whose statistics are to be updated.
5976 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5977 * @local_group: Does group contain this_cpu.
5978 * @sgs: variable to hold the statistics for this group.
5979 * @overload: Indicate more than one runnable task for any CPU.
5981 static inline void update_sg_lb_stats(struct lb_env *env,
5982 struct sched_group *group, int load_idx,
5983 int local_group, struct sg_lb_stats *sgs,
5989 memset(sgs, 0, sizeof(*sgs));
5991 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5992 struct rq *rq = cpu_rq(i);
5994 /* Bias balancing toward cpus of our domain */
5996 load = target_load(i, load_idx);
5998 load = source_load(i, load_idx);
6000 sgs->group_load += load;
6001 sgs->sum_nr_running += rq->nr_running;
6003 if (rq->nr_running > 1)
6006 #ifdef CONFIG_NUMA_BALANCING
6007 sgs->nr_numa_running += rq->nr_numa_running;
6008 sgs->nr_preferred_running += rq->nr_preferred_running;
6010 sgs->sum_weighted_load += weighted_cpuload(i);
6015 /* Adjust by relative CPU capacity of the group */
6016 sgs->group_capacity = group->sgc->capacity;
6017 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6019 if (sgs->sum_nr_running)
6020 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6022 sgs->group_weight = group->group_weight;
6023 sgs->group_capacity_factor = sg_capacity_factor(env, group);
6024 sgs->group_type = group_classify(group, sgs);
6026 if (sgs->group_capacity_factor > sgs->sum_nr_running)
6027 sgs->group_has_free_capacity = 1;
6031 * update_sd_pick_busiest - return 1 on busiest group
6032 * @env: The load balancing environment.
6033 * @sds: sched_domain statistics
6034 * @sg: sched_group candidate to be checked for being the busiest
6035 * @sgs: sched_group statistics
6037 * Determine if @sg is a busier group than the previously selected
6040 * Return: %true if @sg is a busier group than the previously selected
6041 * busiest group. %false otherwise.
6043 static bool update_sd_pick_busiest(struct lb_env *env,
6044 struct sd_lb_stats *sds,
6045 struct sched_group *sg,
6046 struct sg_lb_stats *sgs)
6048 struct sg_lb_stats *busiest = &sds->busiest_stat;
6050 if (sgs->group_type > busiest->group_type)
6053 if (sgs->group_type < busiest->group_type)
6056 if (sgs->avg_load <= busiest->avg_load)
6059 /* This is the busiest node in its class. */
6060 if (!(env->sd->flags & SD_ASYM_PACKING))
6064 * ASYM_PACKING needs to move all the work to the lowest
6065 * numbered CPUs in the group, therefore mark all groups
6066 * higher than ourself as busy.
6068 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6072 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6079 #ifdef CONFIG_NUMA_BALANCING
6080 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6082 if (sgs->sum_nr_running > sgs->nr_numa_running)
6084 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6089 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6091 if (rq->nr_running > rq->nr_numa_running)
6093 if (rq->nr_running > rq->nr_preferred_running)
6098 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6103 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6107 #endif /* CONFIG_NUMA_BALANCING */
6110 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6111 * @env: The load balancing environment.
6112 * @sds: variable to hold the statistics for this sched_domain.
6114 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6116 struct sched_domain *child = env->sd->child;
6117 struct sched_group *sg = env->sd->groups;
6118 struct sg_lb_stats tmp_sgs;
6119 int load_idx, prefer_sibling = 0;
6120 bool overload = false;
6122 if (child && child->flags & SD_PREFER_SIBLING)
6125 load_idx = get_sd_load_idx(env->sd, env->idle);
6128 struct sg_lb_stats *sgs = &tmp_sgs;
6131 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6134 sgs = &sds->local_stat;
6136 if (env->idle != CPU_NEWLY_IDLE ||
6137 time_after_eq(jiffies, sg->sgc->next_update))
6138 update_group_capacity(env->sd, env->dst_cpu);
6141 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6148 * In case the child domain prefers tasks go to siblings
6149 * first, lower the sg capacity factor to one so that we'll try
6150 * and move all the excess tasks away. We lower the capacity
6151 * of a group only if the local group has the capacity to fit
6152 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6153 * extra check prevents the case where you always pull from the
6154 * heaviest group when it is already under-utilized (possible
6155 * with a large weight task outweighs the tasks on the system).
6157 if (prefer_sibling && sds->local &&
6158 sds->local_stat.group_has_free_capacity)
6159 sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U);
6161 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6163 sds->busiest_stat = *sgs;
6167 /* Now, start updating sd_lb_stats */
6168 sds->total_load += sgs->group_load;
6169 sds->total_capacity += sgs->group_capacity;
6172 } while (sg != env->sd->groups);
6174 if (env->sd->flags & SD_NUMA)
6175 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6177 if (!env->sd->parent) {
6178 /* update overload indicator if we are at root domain */
6179 if (env->dst_rq->rd->overload != overload)
6180 env->dst_rq->rd->overload = overload;
6186 * check_asym_packing - Check to see if the group is packed into the
6189 * This is primarily intended to used at the sibling level. Some
6190 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6191 * case of POWER7, it can move to lower SMT modes only when higher
6192 * threads are idle. When in lower SMT modes, the threads will
6193 * perform better since they share less core resources. Hence when we
6194 * have idle threads, we want them to be the higher ones.
6196 * This packing function is run on idle threads. It checks to see if
6197 * the busiest CPU in this domain (core in the P7 case) has a higher
6198 * CPU number than the packing function is being run on. Here we are
6199 * assuming lower CPU number will be equivalent to lower a SMT thread
6202 * Return: 1 when packing is required and a task should be moved to
6203 * this CPU. The amount of the imbalance is returned in *imbalance.
6205 * @env: The load balancing environment.
6206 * @sds: Statistics of the sched_domain which is to be packed
6208 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6212 if (!(env->sd->flags & SD_ASYM_PACKING))
6218 busiest_cpu = group_first_cpu(sds->busiest);
6219 if (env->dst_cpu > busiest_cpu)
6222 env->imbalance = DIV_ROUND_CLOSEST(
6223 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6224 SCHED_CAPACITY_SCALE);
6230 * fix_small_imbalance - Calculate the minor imbalance that exists
6231 * amongst the groups of a sched_domain, during
6233 * @env: The load balancing environment.
6234 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6237 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6239 unsigned long tmp, capa_now = 0, capa_move = 0;
6240 unsigned int imbn = 2;
6241 unsigned long scaled_busy_load_per_task;
6242 struct sg_lb_stats *local, *busiest;
6244 local = &sds->local_stat;
6245 busiest = &sds->busiest_stat;
6247 if (!local->sum_nr_running)
6248 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6249 else if (busiest->load_per_task > local->load_per_task)
6252 scaled_busy_load_per_task =
6253 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6254 busiest->group_capacity;
6256 if (busiest->avg_load + scaled_busy_load_per_task >=
6257 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6258 env->imbalance = busiest->load_per_task;
6263 * OK, we don't have enough imbalance to justify moving tasks,
6264 * however we may be able to increase total CPU capacity used by
6268 capa_now += busiest->group_capacity *
6269 min(busiest->load_per_task, busiest->avg_load);
6270 capa_now += local->group_capacity *
6271 min(local->load_per_task, local->avg_load);
6272 capa_now /= SCHED_CAPACITY_SCALE;
6274 /* Amount of load we'd subtract */
6275 if (busiest->avg_load > scaled_busy_load_per_task) {
6276 capa_move += busiest->group_capacity *
6277 min(busiest->load_per_task,
6278 busiest->avg_load - scaled_busy_load_per_task);
6281 /* Amount of load we'd add */
6282 if (busiest->avg_load * busiest->group_capacity <
6283 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6284 tmp = (busiest->avg_load * busiest->group_capacity) /
6285 local->group_capacity;
6287 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6288 local->group_capacity;
6290 capa_move += local->group_capacity *
6291 min(local->load_per_task, local->avg_load + tmp);
6292 capa_move /= SCHED_CAPACITY_SCALE;
6294 /* Move if we gain throughput */
6295 if (capa_move > capa_now)
6296 env->imbalance = busiest->load_per_task;
6300 * calculate_imbalance - Calculate the amount of imbalance present within the
6301 * groups of a given sched_domain during load balance.
6302 * @env: load balance environment
6303 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6305 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6307 unsigned long max_pull, load_above_capacity = ~0UL;
6308 struct sg_lb_stats *local, *busiest;
6310 local = &sds->local_stat;
6311 busiest = &sds->busiest_stat;
6313 if (busiest->group_type == group_imbalanced) {
6315 * In the group_imb case we cannot rely on group-wide averages
6316 * to ensure cpu-load equilibrium, look at wider averages. XXX
6318 busiest->load_per_task =
6319 min(busiest->load_per_task, sds->avg_load);
6323 * In the presence of smp nice balancing, certain scenarios can have
6324 * max load less than avg load(as we skip the groups at or below
6325 * its cpu_capacity, while calculating max_load..)
6327 if (busiest->avg_load <= sds->avg_load ||
6328 local->avg_load >= sds->avg_load) {
6330 return fix_small_imbalance(env, sds);
6334 * If there aren't any idle cpus, avoid creating some.
6336 if (busiest->group_type == group_overloaded &&
6337 local->group_type == group_overloaded) {
6338 load_above_capacity =
6339 (busiest->sum_nr_running - busiest->group_capacity_factor);
6341 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE);
6342 load_above_capacity /= busiest->group_capacity;
6346 * We're trying to get all the cpus to the average_load, so we don't
6347 * want to push ourselves above the average load, nor do we wish to
6348 * reduce the max loaded cpu below the average load. At the same time,
6349 * we also don't want to reduce the group load below the group capacity
6350 * (so that we can implement power-savings policies etc). Thus we look
6351 * for the minimum possible imbalance.
6353 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6355 /* How much load to actually move to equalise the imbalance */
6356 env->imbalance = min(
6357 max_pull * busiest->group_capacity,
6358 (sds->avg_load - local->avg_load) * local->group_capacity
6359 ) / SCHED_CAPACITY_SCALE;
6362 * if *imbalance is less than the average load per runnable task
6363 * there is no guarantee that any tasks will be moved so we'll have
6364 * a think about bumping its value to force at least one task to be
6367 if (env->imbalance < busiest->load_per_task)
6368 return fix_small_imbalance(env, sds);
6371 /******* find_busiest_group() helpers end here *********************/
6374 * find_busiest_group - Returns the busiest group within the sched_domain
6375 * if there is an imbalance. If there isn't an imbalance, and
6376 * the user has opted for power-savings, it returns a group whose
6377 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6378 * such a group exists.
6380 * Also calculates the amount of weighted load which should be moved
6381 * to restore balance.
6383 * @env: The load balancing environment.
6385 * Return: - The busiest group if imbalance exists.
6386 * - If no imbalance and user has opted for power-savings balance,
6387 * return the least loaded group whose CPUs can be
6388 * put to idle by rebalancing its tasks onto our group.
6390 static struct sched_group *find_busiest_group(struct lb_env *env)
6392 struct sg_lb_stats *local, *busiest;
6393 struct sd_lb_stats sds;
6395 init_sd_lb_stats(&sds);
6398 * Compute the various statistics relavent for load balancing at
6401 update_sd_lb_stats(env, &sds);
6402 local = &sds.local_stat;
6403 busiest = &sds.busiest_stat;
6405 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6406 check_asym_packing(env, &sds))
6409 /* There is no busy sibling group to pull tasks from */
6410 if (!sds.busiest || busiest->sum_nr_running == 0)
6413 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6414 / sds.total_capacity;
6417 * If the busiest group is imbalanced the below checks don't
6418 * work because they assume all things are equal, which typically
6419 * isn't true due to cpus_allowed constraints and the like.
6421 if (busiest->group_type == group_imbalanced)
6424 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6425 if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity &&
6426 !busiest->group_has_free_capacity)
6430 * If the local group is more busy than the selected busiest group
6431 * don't try and pull any tasks.
6433 if (local->avg_load >= busiest->avg_load)
6437 * Don't pull any tasks if this group is already above the domain
6440 if (local->avg_load >= sds.avg_load)
6443 if (env->idle == CPU_IDLE) {
6445 * This cpu is idle. If the busiest group load doesn't
6446 * have more tasks than the number of available cpu's and
6447 * there is no imbalance between this and busiest group
6448 * wrt to idle cpu's, it is balanced.
6450 if ((local->idle_cpus < busiest->idle_cpus) &&
6451 busiest->sum_nr_running <= busiest->group_weight)
6455 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6456 * imbalance_pct to be conservative.
6458 if (100 * busiest->avg_load <=
6459 env->sd->imbalance_pct * local->avg_load)
6464 /* Looks like there is an imbalance. Compute it */
6465 calculate_imbalance(env, &sds);
6474 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6476 static struct rq *find_busiest_queue(struct lb_env *env,
6477 struct sched_group *group)
6479 struct rq *busiest = NULL, *rq;
6480 unsigned long busiest_load = 0, busiest_capacity = 1;
6483 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6484 unsigned long capacity, capacity_factor, wl;
6488 rt = fbq_classify_rq(rq);
6491 * We classify groups/runqueues into three groups:
6492 * - regular: there are !numa tasks
6493 * - remote: there are numa tasks that run on the 'wrong' node
6494 * - all: there is no distinction
6496 * In order to avoid migrating ideally placed numa tasks,
6497 * ignore those when there's better options.
6499 * If we ignore the actual busiest queue to migrate another
6500 * task, the next balance pass can still reduce the busiest
6501 * queue by moving tasks around inside the node.
6503 * If we cannot move enough load due to this classification
6504 * the next pass will adjust the group classification and
6505 * allow migration of more tasks.
6507 * Both cases only affect the total convergence complexity.
6509 if (rt > env->fbq_type)
6512 capacity = capacity_of(i);
6513 capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE);
6514 if (!capacity_factor)
6515 capacity_factor = fix_small_capacity(env->sd, group);
6517 wl = weighted_cpuload(i);
6520 * When comparing with imbalance, use weighted_cpuload()
6521 * which is not scaled with the cpu capacity.
6523 if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance)
6527 * For the load comparisons with the other cpu's, consider
6528 * the weighted_cpuload() scaled with the cpu capacity, so
6529 * that the load can be moved away from the cpu that is
6530 * potentially running at a lower capacity.
6532 * Thus we're looking for max(wl_i / capacity_i), crosswise
6533 * multiplication to rid ourselves of the division works out
6534 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6535 * our previous maximum.
6537 if (wl * busiest_capacity > busiest_load * capacity) {
6539 busiest_capacity = capacity;
6548 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6549 * so long as it is large enough.
6551 #define MAX_PINNED_INTERVAL 512
6553 /* Working cpumask for load_balance and load_balance_newidle. */
6554 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6556 static int need_active_balance(struct lb_env *env)
6558 struct sched_domain *sd = env->sd;
6560 if (env->idle == CPU_NEWLY_IDLE) {
6563 * ASYM_PACKING needs to force migrate tasks from busy but
6564 * higher numbered CPUs in order to pack all tasks in the
6565 * lowest numbered CPUs.
6567 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6571 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6574 static int active_load_balance_cpu_stop(void *data);
6576 static int should_we_balance(struct lb_env *env)
6578 struct sched_group *sg = env->sd->groups;
6579 struct cpumask *sg_cpus, *sg_mask;
6580 int cpu, balance_cpu = -1;
6583 * In the newly idle case, we will allow all the cpu's
6584 * to do the newly idle load balance.
6586 if (env->idle == CPU_NEWLY_IDLE)
6589 sg_cpus = sched_group_cpus(sg);
6590 sg_mask = sched_group_mask(sg);
6591 /* Try to find first idle cpu */
6592 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6593 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6600 if (balance_cpu == -1)
6601 balance_cpu = group_balance_cpu(sg);
6604 * First idle cpu or the first cpu(busiest) in this sched group
6605 * is eligible for doing load balancing at this and above domains.
6607 return balance_cpu == env->dst_cpu;
6611 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6612 * tasks if there is an imbalance.
6614 static int load_balance(int this_cpu, struct rq *this_rq,
6615 struct sched_domain *sd, enum cpu_idle_type idle,
6616 int *continue_balancing)
6618 int ld_moved, cur_ld_moved, active_balance = 0;
6619 struct sched_domain *sd_parent = sd->parent;
6620 struct sched_group *group;
6622 unsigned long flags;
6623 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6625 struct lb_env env = {
6627 .dst_cpu = this_cpu,
6629 .dst_grpmask = sched_group_cpus(sd->groups),
6631 .loop_break = sched_nr_migrate_break,
6634 .tasks = LIST_HEAD_INIT(env.tasks),
6638 * For NEWLY_IDLE load_balancing, we don't need to consider
6639 * other cpus in our group
6641 if (idle == CPU_NEWLY_IDLE)
6642 env.dst_grpmask = NULL;
6644 cpumask_copy(cpus, cpu_active_mask);
6646 schedstat_inc(sd, lb_count[idle]);
6649 if (!should_we_balance(&env)) {
6650 *continue_balancing = 0;
6654 group = find_busiest_group(&env);
6656 schedstat_inc(sd, lb_nobusyg[idle]);
6660 busiest = find_busiest_queue(&env, group);
6662 schedstat_inc(sd, lb_nobusyq[idle]);
6666 BUG_ON(busiest == env.dst_rq);
6668 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6671 if (busiest->nr_running > 1) {
6673 * Attempt to move tasks. If find_busiest_group has found
6674 * an imbalance but busiest->nr_running <= 1, the group is
6675 * still unbalanced. ld_moved simply stays zero, so it is
6676 * correctly treated as an imbalance.
6678 env.flags |= LBF_ALL_PINNED;
6679 env.src_cpu = busiest->cpu;
6680 env.src_rq = busiest;
6681 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6684 raw_spin_lock_irqsave(&busiest->lock, flags);
6687 * cur_ld_moved - load moved in current iteration
6688 * ld_moved - cumulative load moved across iterations
6690 cur_ld_moved = detach_tasks(&env);
6693 * We've detached some tasks from busiest_rq. Every
6694 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6695 * unlock busiest->lock, and we are able to be sure
6696 * that nobody can manipulate the tasks in parallel.
6697 * See task_rq_lock() family for the details.
6700 raw_spin_unlock(&busiest->lock);
6704 ld_moved += cur_ld_moved;
6707 local_irq_restore(flags);
6710 * some other cpu did the load balance for us.
6712 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6713 resched_cpu(env.dst_cpu);
6715 if (env.flags & LBF_NEED_BREAK) {
6716 env.flags &= ~LBF_NEED_BREAK;
6721 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6722 * us and move them to an alternate dst_cpu in our sched_group
6723 * where they can run. The upper limit on how many times we
6724 * iterate on same src_cpu is dependent on number of cpus in our
6727 * This changes load balance semantics a bit on who can move
6728 * load to a given_cpu. In addition to the given_cpu itself
6729 * (or a ilb_cpu acting on its behalf where given_cpu is
6730 * nohz-idle), we now have balance_cpu in a position to move
6731 * load to given_cpu. In rare situations, this may cause
6732 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6733 * _independently_ and at _same_ time to move some load to
6734 * given_cpu) causing exceess load to be moved to given_cpu.
6735 * This however should not happen so much in practice and
6736 * moreover subsequent load balance cycles should correct the
6737 * excess load moved.
6739 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6741 /* Prevent to re-select dst_cpu via env's cpus */
6742 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6744 env.dst_rq = cpu_rq(env.new_dst_cpu);
6745 env.dst_cpu = env.new_dst_cpu;
6746 env.flags &= ~LBF_DST_PINNED;
6748 env.loop_break = sched_nr_migrate_break;
6751 * Go back to "more_balance" rather than "redo" since we
6752 * need to continue with same src_cpu.
6758 * We failed to reach balance because of affinity.
6761 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
6763 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6764 *group_imbalance = 1;
6765 } else if (*group_imbalance)
6766 *group_imbalance = 0;
6769 /* All tasks on this runqueue were pinned by CPU affinity */
6770 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6771 cpumask_clear_cpu(cpu_of(busiest), cpus);
6772 if (!cpumask_empty(cpus)) {
6774 env.loop_break = sched_nr_migrate_break;
6782 schedstat_inc(sd, lb_failed[idle]);
6784 * Increment the failure counter only on periodic balance.
6785 * We do not want newidle balance, which can be very
6786 * frequent, pollute the failure counter causing
6787 * excessive cache_hot migrations and active balances.
6789 if (idle != CPU_NEWLY_IDLE)
6790 sd->nr_balance_failed++;
6792 if (need_active_balance(&env)) {
6793 raw_spin_lock_irqsave(&busiest->lock, flags);
6795 /* don't kick the active_load_balance_cpu_stop,
6796 * if the curr task on busiest cpu can't be
6799 if (!cpumask_test_cpu(this_cpu,
6800 tsk_cpus_allowed(busiest->curr))) {
6801 raw_spin_unlock_irqrestore(&busiest->lock,
6803 env.flags |= LBF_ALL_PINNED;
6804 goto out_one_pinned;
6808 * ->active_balance synchronizes accesses to
6809 * ->active_balance_work. Once set, it's cleared
6810 * only after active load balance is finished.
6812 if (!busiest->active_balance) {
6813 busiest->active_balance = 1;
6814 busiest->push_cpu = this_cpu;
6817 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6819 if (active_balance) {
6820 stop_one_cpu_nowait(cpu_of(busiest),
6821 active_load_balance_cpu_stop, busiest,
6822 &busiest->active_balance_work);
6826 * We've kicked active balancing, reset the failure
6829 sd->nr_balance_failed = sd->cache_nice_tries+1;
6832 sd->nr_balance_failed = 0;
6834 if (likely(!active_balance)) {
6835 /* We were unbalanced, so reset the balancing interval */
6836 sd->balance_interval = sd->min_interval;
6839 * If we've begun active balancing, start to back off. This
6840 * case may not be covered by the all_pinned logic if there
6841 * is only 1 task on the busy runqueue (because we don't call
6844 if (sd->balance_interval < sd->max_interval)
6845 sd->balance_interval *= 2;
6851 schedstat_inc(sd, lb_balanced[idle]);
6853 sd->nr_balance_failed = 0;
6856 /* tune up the balancing interval */
6857 if (((env.flags & LBF_ALL_PINNED) &&
6858 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6859 (sd->balance_interval < sd->max_interval))
6860 sd->balance_interval *= 2;
6867 static inline unsigned long
6868 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
6870 unsigned long interval = sd->balance_interval;
6873 interval *= sd->busy_factor;
6875 /* scale ms to jiffies */
6876 interval = msecs_to_jiffies(interval);
6877 interval = clamp(interval, 1UL, max_load_balance_interval);
6883 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
6885 unsigned long interval, next;
6887 interval = get_sd_balance_interval(sd, cpu_busy);
6888 next = sd->last_balance + interval;
6890 if (time_after(*next_balance, next))
6891 *next_balance = next;
6895 * idle_balance is called by schedule() if this_cpu is about to become
6896 * idle. Attempts to pull tasks from other CPUs.
6898 static int idle_balance(struct rq *this_rq)
6900 unsigned long next_balance = jiffies + HZ;
6901 int this_cpu = this_rq->cpu;
6902 struct sched_domain *sd;
6903 int pulled_task = 0;
6906 idle_enter_fair(this_rq);
6909 * We must set idle_stamp _before_ calling idle_balance(), such that we
6910 * measure the duration of idle_balance() as idle time.
6912 this_rq->idle_stamp = rq_clock(this_rq);
6914 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
6915 !this_rq->rd->overload) {
6917 sd = rcu_dereference_check_sched_domain(this_rq->sd);
6919 update_next_balance(sd, 0, &next_balance);
6926 * Drop the rq->lock, but keep IRQ/preempt disabled.
6928 raw_spin_unlock(&this_rq->lock);
6930 update_blocked_averages(this_cpu);
6932 for_each_domain(this_cpu, sd) {
6933 int continue_balancing = 1;
6934 u64 t0, domain_cost;
6936 if (!(sd->flags & SD_LOAD_BALANCE))
6939 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
6940 update_next_balance(sd, 0, &next_balance);
6944 if (sd->flags & SD_BALANCE_NEWIDLE) {
6945 t0 = sched_clock_cpu(this_cpu);
6947 pulled_task = load_balance(this_cpu, this_rq,
6949 &continue_balancing);
6951 domain_cost = sched_clock_cpu(this_cpu) - t0;
6952 if (domain_cost > sd->max_newidle_lb_cost)
6953 sd->max_newidle_lb_cost = domain_cost;
6955 curr_cost += domain_cost;
6958 update_next_balance(sd, 0, &next_balance);
6961 * Stop searching for tasks to pull if there are
6962 * now runnable tasks on this rq.
6964 if (pulled_task || this_rq->nr_running > 0)
6969 raw_spin_lock(&this_rq->lock);
6971 if (curr_cost > this_rq->max_idle_balance_cost)
6972 this_rq->max_idle_balance_cost = curr_cost;
6975 * While browsing the domains, we released the rq lock, a task could
6976 * have been enqueued in the meantime. Since we're not going idle,
6977 * pretend we pulled a task.
6979 if (this_rq->cfs.h_nr_running && !pulled_task)
6983 /* Move the next balance forward */
6984 if (time_after(this_rq->next_balance, next_balance))
6985 this_rq->next_balance = next_balance;
6987 /* Is there a task of a high priority class? */
6988 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
6992 idle_exit_fair(this_rq);
6993 this_rq->idle_stamp = 0;
7000 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7001 * running tasks off the busiest CPU onto idle CPUs. It requires at
7002 * least 1 task to be running on each physical CPU where possible, and
7003 * avoids physical / logical imbalances.
7005 static int active_load_balance_cpu_stop(void *data)
7007 struct rq *busiest_rq = data;
7008 int busiest_cpu = cpu_of(busiest_rq);
7009 int target_cpu = busiest_rq->push_cpu;
7010 struct rq *target_rq = cpu_rq(target_cpu);
7011 struct sched_domain *sd;
7012 struct task_struct *p = NULL;
7014 raw_spin_lock_irq(&busiest_rq->lock);
7016 /* make sure the requested cpu hasn't gone down in the meantime */
7017 if (unlikely(busiest_cpu != smp_processor_id() ||
7018 !busiest_rq->active_balance))
7021 /* Is there any task to move? */
7022 if (busiest_rq->nr_running <= 1)
7026 * This condition is "impossible", if it occurs
7027 * we need to fix it. Originally reported by
7028 * Bjorn Helgaas on a 128-cpu setup.
7030 BUG_ON(busiest_rq == target_rq);
7032 /* Search for an sd spanning us and the target CPU. */
7034 for_each_domain(target_cpu, sd) {
7035 if ((sd->flags & SD_LOAD_BALANCE) &&
7036 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7041 struct lb_env env = {
7043 .dst_cpu = target_cpu,
7044 .dst_rq = target_rq,
7045 .src_cpu = busiest_rq->cpu,
7046 .src_rq = busiest_rq,
7050 schedstat_inc(sd, alb_count);
7052 p = detach_one_task(&env);
7054 schedstat_inc(sd, alb_pushed);
7056 schedstat_inc(sd, alb_failed);
7060 busiest_rq->active_balance = 0;
7061 raw_spin_unlock(&busiest_rq->lock);
7064 attach_one_task(target_rq, p);
7071 static inline int on_null_domain(struct rq *rq)
7073 return unlikely(!rcu_dereference_sched(rq->sd));
7076 #ifdef CONFIG_NO_HZ_COMMON
7078 * idle load balancing details
7079 * - When one of the busy CPUs notice that there may be an idle rebalancing
7080 * needed, they will kick the idle load balancer, which then does idle
7081 * load balancing for all the idle CPUs.
7084 cpumask_var_t idle_cpus_mask;
7086 unsigned long next_balance; /* in jiffy units */
7087 } nohz ____cacheline_aligned;
7089 static inline int find_new_ilb(void)
7091 int ilb = cpumask_first(nohz.idle_cpus_mask);
7093 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7100 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7101 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7102 * CPU (if there is one).
7104 static void nohz_balancer_kick(void)
7108 nohz.next_balance++;
7110 ilb_cpu = find_new_ilb();
7112 if (ilb_cpu >= nr_cpu_ids)
7115 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7118 * Use smp_send_reschedule() instead of resched_cpu().
7119 * This way we generate a sched IPI on the target cpu which
7120 * is idle. And the softirq performing nohz idle load balance
7121 * will be run before returning from the IPI.
7123 smp_send_reschedule(ilb_cpu);
7127 static inline void nohz_balance_exit_idle(int cpu)
7129 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7131 * Completely isolated CPUs don't ever set, so we must test.
7133 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7134 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7135 atomic_dec(&nohz.nr_cpus);
7137 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7141 static inline void set_cpu_sd_state_busy(void)
7143 struct sched_domain *sd;
7144 int cpu = smp_processor_id();
7147 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7149 if (!sd || !sd->nohz_idle)
7153 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7158 void set_cpu_sd_state_idle(void)
7160 struct sched_domain *sd;
7161 int cpu = smp_processor_id();
7164 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7166 if (!sd || sd->nohz_idle)
7170 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7176 * This routine will record that the cpu is going idle with tick stopped.
7177 * This info will be used in performing idle load balancing in the future.
7179 void nohz_balance_enter_idle(int cpu)
7182 * If this cpu is going down, then nothing needs to be done.
7184 if (!cpu_active(cpu))
7187 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7191 * If we're a completely isolated CPU, we don't play.
7193 if (on_null_domain(cpu_rq(cpu)))
7196 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7197 atomic_inc(&nohz.nr_cpus);
7198 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7201 static int sched_ilb_notifier(struct notifier_block *nfb,
7202 unsigned long action, void *hcpu)
7204 switch (action & ~CPU_TASKS_FROZEN) {
7206 nohz_balance_exit_idle(smp_processor_id());
7214 static DEFINE_SPINLOCK(balancing);
7217 * Scale the max load_balance interval with the number of CPUs in the system.
7218 * This trades load-balance latency on larger machines for less cross talk.
7220 void update_max_interval(void)
7222 max_load_balance_interval = HZ*num_online_cpus()/10;
7226 * It checks each scheduling domain to see if it is due to be balanced,
7227 * and initiates a balancing operation if so.
7229 * Balancing parameters are set up in init_sched_domains.
7231 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7233 int continue_balancing = 1;
7235 unsigned long interval;
7236 struct sched_domain *sd;
7237 /* Earliest time when we have to do rebalance again */
7238 unsigned long next_balance = jiffies + 60*HZ;
7239 int update_next_balance = 0;
7240 int need_serialize, need_decay = 0;
7243 update_blocked_averages(cpu);
7246 for_each_domain(cpu, sd) {
7248 * Decay the newidle max times here because this is a regular
7249 * visit to all the domains. Decay ~1% per second.
7251 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7252 sd->max_newidle_lb_cost =
7253 (sd->max_newidle_lb_cost * 253) / 256;
7254 sd->next_decay_max_lb_cost = jiffies + HZ;
7257 max_cost += sd->max_newidle_lb_cost;
7259 if (!(sd->flags & SD_LOAD_BALANCE))
7263 * Stop the load balance at this level. There is another
7264 * CPU in our sched group which is doing load balancing more
7267 if (!continue_balancing) {
7273 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7275 need_serialize = sd->flags & SD_SERIALIZE;
7276 if (need_serialize) {
7277 if (!spin_trylock(&balancing))
7281 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7282 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7284 * The LBF_DST_PINNED logic could have changed
7285 * env->dst_cpu, so we can't know our idle
7286 * state even if we migrated tasks. Update it.
7288 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7290 sd->last_balance = jiffies;
7291 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7294 spin_unlock(&balancing);
7296 if (time_after(next_balance, sd->last_balance + interval)) {
7297 next_balance = sd->last_balance + interval;
7298 update_next_balance = 1;
7303 * Ensure the rq-wide value also decays but keep it at a
7304 * reasonable floor to avoid funnies with rq->avg_idle.
7306 rq->max_idle_balance_cost =
7307 max((u64)sysctl_sched_migration_cost, max_cost);
7312 * next_balance will be updated only when there is a need.
7313 * When the cpu is attached to null domain for ex, it will not be
7316 if (likely(update_next_balance))
7317 rq->next_balance = next_balance;
7320 #ifdef CONFIG_NO_HZ_COMMON
7322 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7323 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7325 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7327 int this_cpu = this_rq->cpu;
7331 if (idle != CPU_IDLE ||
7332 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7335 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7336 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7340 * If this cpu gets work to do, stop the load balancing
7341 * work being done for other cpus. Next load
7342 * balancing owner will pick it up.
7347 rq = cpu_rq(balance_cpu);
7350 * If time for next balance is due,
7353 if (time_after_eq(jiffies, rq->next_balance)) {
7354 raw_spin_lock_irq(&rq->lock);
7355 update_rq_clock(rq);
7356 update_idle_cpu_load(rq);
7357 raw_spin_unlock_irq(&rq->lock);
7358 rebalance_domains(rq, CPU_IDLE);
7361 if (time_after(this_rq->next_balance, rq->next_balance))
7362 this_rq->next_balance = rq->next_balance;
7364 nohz.next_balance = this_rq->next_balance;
7366 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7370 * Current heuristic for kicking the idle load balancer in the presence
7371 * of an idle cpu is the system.
7372 * - This rq has more than one task.
7373 * - At any scheduler domain level, this cpu's scheduler group has multiple
7374 * busy cpu's exceeding the group's capacity.
7375 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7376 * domain span are idle.
7378 static inline int nohz_kick_needed(struct rq *rq)
7380 unsigned long now = jiffies;
7381 struct sched_domain *sd;
7382 struct sched_group_capacity *sgc;
7383 int nr_busy, cpu = rq->cpu;
7385 if (unlikely(rq->idle_balance))
7389 * We may be recently in ticked or tickless idle mode. At the first
7390 * busy tick after returning from idle, we will update the busy stats.
7392 set_cpu_sd_state_busy();
7393 nohz_balance_exit_idle(cpu);
7396 * None are in tickless mode and hence no need for NOHZ idle load
7399 if (likely(!atomic_read(&nohz.nr_cpus)))
7402 if (time_before(now, nohz.next_balance))
7405 if (rq->nr_running >= 2)
7409 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7412 sgc = sd->groups->sgc;
7413 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7416 goto need_kick_unlock;
7419 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7421 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7422 sched_domain_span(sd)) < cpu))
7423 goto need_kick_unlock;
7434 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7438 * run_rebalance_domains is triggered when needed from the scheduler tick.
7439 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7441 static void run_rebalance_domains(struct softirq_action *h)
7443 struct rq *this_rq = this_rq();
7444 enum cpu_idle_type idle = this_rq->idle_balance ?
7445 CPU_IDLE : CPU_NOT_IDLE;
7447 rebalance_domains(this_rq, idle);
7450 * If this cpu has a pending nohz_balance_kick, then do the
7451 * balancing on behalf of the other idle cpus whose ticks are
7454 nohz_idle_balance(this_rq, idle);
7458 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7460 void trigger_load_balance(struct rq *rq)
7462 /* Don't need to rebalance while attached to NULL domain */
7463 if (unlikely(on_null_domain(rq)))
7466 if (time_after_eq(jiffies, rq->next_balance))
7467 raise_softirq(SCHED_SOFTIRQ);
7468 #ifdef CONFIG_NO_HZ_COMMON
7469 if (nohz_kick_needed(rq))
7470 nohz_balancer_kick();
7474 static void rq_online_fair(struct rq *rq)
7478 update_runtime_enabled(rq);
7481 static void rq_offline_fair(struct rq *rq)
7485 /* Ensure any throttled groups are reachable by pick_next_task */
7486 unthrottle_offline_cfs_rqs(rq);
7489 #endif /* CONFIG_SMP */
7492 * scheduler tick hitting a task of our scheduling class:
7494 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7496 struct cfs_rq *cfs_rq;
7497 struct sched_entity *se = &curr->se;
7499 for_each_sched_entity(se) {
7500 cfs_rq = cfs_rq_of(se);
7501 entity_tick(cfs_rq, se, queued);
7504 if (numabalancing_enabled)
7505 task_tick_numa(rq, curr);
7507 update_rq_runnable_avg(rq, 1);
7511 * called on fork with the child task as argument from the parent's context
7512 * - child not yet on the tasklist
7513 * - preemption disabled
7515 static void task_fork_fair(struct task_struct *p)
7517 struct cfs_rq *cfs_rq;
7518 struct sched_entity *se = &p->se, *curr;
7519 int this_cpu = smp_processor_id();
7520 struct rq *rq = this_rq();
7521 unsigned long flags;
7523 raw_spin_lock_irqsave(&rq->lock, flags);
7525 update_rq_clock(rq);
7527 cfs_rq = task_cfs_rq(current);
7528 curr = cfs_rq->curr;
7531 * Not only the cpu but also the task_group of the parent might have
7532 * been changed after parent->se.parent,cfs_rq were copied to
7533 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7534 * of child point to valid ones.
7537 __set_task_cpu(p, this_cpu);
7540 update_curr(cfs_rq);
7543 se->vruntime = curr->vruntime;
7544 place_entity(cfs_rq, se, 1);
7546 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7548 * Upon rescheduling, sched_class::put_prev_task() will place
7549 * 'current' within the tree based on its new key value.
7551 swap(curr->vruntime, se->vruntime);
7555 se->vruntime -= cfs_rq->min_vruntime;
7557 raw_spin_unlock_irqrestore(&rq->lock, flags);
7561 * Priority of the task has changed. Check to see if we preempt
7565 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7567 if (!task_on_rq_queued(p))
7571 * Reschedule if we are currently running on this runqueue and
7572 * our priority decreased, or if we are not currently running on
7573 * this runqueue and our priority is higher than the current's
7575 if (rq->curr == p) {
7576 if (p->prio > oldprio)
7579 check_preempt_curr(rq, p, 0);
7582 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7584 struct sched_entity *se = &p->se;
7585 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7588 * Ensure the task's vruntime is normalized, so that when it's
7589 * switched back to the fair class the enqueue_entity(.flags=0) will
7590 * do the right thing.
7592 * If it's queued, then the dequeue_entity(.flags=0) will already
7593 * have normalized the vruntime, if it's !queued, then only when
7594 * the task is sleeping will it still have non-normalized vruntime.
7596 if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) {
7598 * Fix up our vruntime so that the current sleep doesn't
7599 * cause 'unlimited' sleep bonus.
7601 place_entity(cfs_rq, se, 0);
7602 se->vruntime -= cfs_rq->min_vruntime;
7607 * Remove our load from contribution when we leave sched_fair
7608 * and ensure we don't carry in an old decay_count if we
7611 if (se->avg.decay_count) {
7612 __synchronize_entity_decay(se);
7613 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7619 * We switched to the sched_fair class.
7621 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7623 #ifdef CONFIG_FAIR_GROUP_SCHED
7624 struct sched_entity *se = &p->se;
7626 * Since the real-depth could have been changed (only FAIR
7627 * class maintain depth value), reset depth properly.
7629 se->depth = se->parent ? se->parent->depth + 1 : 0;
7631 if (!task_on_rq_queued(p))
7635 * We were most likely switched from sched_rt, so
7636 * kick off the schedule if running, otherwise just see
7637 * if we can still preempt the current task.
7642 check_preempt_curr(rq, p, 0);
7645 /* Account for a task changing its policy or group.
7647 * This routine is mostly called to set cfs_rq->curr field when a task
7648 * migrates between groups/classes.
7650 static void set_curr_task_fair(struct rq *rq)
7652 struct sched_entity *se = &rq->curr->se;
7654 for_each_sched_entity(se) {
7655 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7657 set_next_entity(cfs_rq, se);
7658 /* ensure bandwidth has been allocated on our new cfs_rq */
7659 account_cfs_rq_runtime(cfs_rq, 0);
7663 void init_cfs_rq(struct cfs_rq *cfs_rq)
7665 cfs_rq->tasks_timeline = RB_ROOT;
7666 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7667 #ifndef CONFIG_64BIT
7668 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7671 atomic64_set(&cfs_rq->decay_counter, 1);
7672 atomic_long_set(&cfs_rq->removed_load, 0);
7676 #ifdef CONFIG_FAIR_GROUP_SCHED
7677 static void task_move_group_fair(struct task_struct *p, int queued)
7679 struct sched_entity *se = &p->se;
7680 struct cfs_rq *cfs_rq;
7683 * If the task was not on the rq at the time of this cgroup movement
7684 * it must have been asleep, sleeping tasks keep their ->vruntime
7685 * absolute on their old rq until wakeup (needed for the fair sleeper
7686 * bonus in place_entity()).
7688 * If it was on the rq, we've just 'preempted' it, which does convert
7689 * ->vruntime to a relative base.
7691 * Make sure both cases convert their relative position when migrating
7692 * to another cgroup's rq. This does somewhat interfere with the
7693 * fair sleeper stuff for the first placement, but who cares.
7696 * When !queued, vruntime of the task has usually NOT been normalized.
7697 * But there are some cases where it has already been normalized:
7699 * - Moving a forked child which is waiting for being woken up by
7700 * wake_up_new_task().
7701 * - Moving a task which has been woken up by try_to_wake_up() and
7702 * waiting for actually being woken up by sched_ttwu_pending().
7704 * To prevent boost or penalty in the new cfs_rq caused by delta
7705 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7707 if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7711 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7712 set_task_rq(p, task_cpu(p));
7713 se->depth = se->parent ? se->parent->depth + 1 : 0;
7715 cfs_rq = cfs_rq_of(se);
7716 se->vruntime += cfs_rq->min_vruntime;
7719 * migrate_task_rq_fair() will have removed our previous
7720 * contribution, but we must synchronize for ongoing future
7723 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7724 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7729 void free_fair_sched_group(struct task_group *tg)
7733 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7735 for_each_possible_cpu(i) {
7737 kfree(tg->cfs_rq[i]);
7746 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7748 struct cfs_rq *cfs_rq;
7749 struct sched_entity *se;
7752 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7755 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7759 tg->shares = NICE_0_LOAD;
7761 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7763 for_each_possible_cpu(i) {
7764 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7765 GFP_KERNEL, cpu_to_node(i));
7769 se = kzalloc_node(sizeof(struct sched_entity),
7770 GFP_KERNEL, cpu_to_node(i));
7774 init_cfs_rq(cfs_rq);
7775 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7786 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7788 struct rq *rq = cpu_rq(cpu);
7789 unsigned long flags;
7792 * Only empty task groups can be destroyed; so we can speculatively
7793 * check on_list without danger of it being re-added.
7795 if (!tg->cfs_rq[cpu]->on_list)
7798 raw_spin_lock_irqsave(&rq->lock, flags);
7799 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7800 raw_spin_unlock_irqrestore(&rq->lock, flags);
7803 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7804 struct sched_entity *se, int cpu,
7805 struct sched_entity *parent)
7807 struct rq *rq = cpu_rq(cpu);
7811 init_cfs_rq_runtime(cfs_rq);
7813 tg->cfs_rq[cpu] = cfs_rq;
7816 /* se could be NULL for root_task_group */
7821 se->cfs_rq = &rq->cfs;
7824 se->cfs_rq = parent->my_q;
7825 se->depth = parent->depth + 1;
7829 /* guarantee group entities always have weight */
7830 update_load_set(&se->load, NICE_0_LOAD);
7831 se->parent = parent;
7834 static DEFINE_MUTEX(shares_mutex);
7836 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7839 unsigned long flags;
7842 * We can't change the weight of the root cgroup.
7847 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7849 mutex_lock(&shares_mutex);
7850 if (tg->shares == shares)
7853 tg->shares = shares;
7854 for_each_possible_cpu(i) {
7855 struct rq *rq = cpu_rq(i);
7856 struct sched_entity *se;
7859 /* Propagate contribution to hierarchy */
7860 raw_spin_lock_irqsave(&rq->lock, flags);
7862 /* Possible calls to update_curr() need rq clock */
7863 update_rq_clock(rq);
7864 for_each_sched_entity(se)
7865 update_cfs_shares(group_cfs_rq(se));
7866 raw_spin_unlock_irqrestore(&rq->lock, flags);
7870 mutex_unlock(&shares_mutex);
7873 #else /* CONFIG_FAIR_GROUP_SCHED */
7875 void free_fair_sched_group(struct task_group *tg) { }
7877 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7882 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7884 #endif /* CONFIG_FAIR_GROUP_SCHED */
7887 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7889 struct sched_entity *se = &task->se;
7890 unsigned int rr_interval = 0;
7893 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7896 if (rq->cfs.load.weight)
7897 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7903 * All the scheduling class methods:
7905 const struct sched_class fair_sched_class = {
7906 .next = &idle_sched_class,
7907 .enqueue_task = enqueue_task_fair,
7908 .dequeue_task = dequeue_task_fair,
7909 .yield_task = yield_task_fair,
7910 .yield_to_task = yield_to_task_fair,
7912 .check_preempt_curr = check_preempt_wakeup,
7914 .pick_next_task = pick_next_task_fair,
7915 .put_prev_task = put_prev_task_fair,
7918 .select_task_rq = select_task_rq_fair,
7919 .migrate_task_rq = migrate_task_rq_fair,
7921 .rq_online = rq_online_fair,
7922 .rq_offline = rq_offline_fair,
7924 .task_waking = task_waking_fair,
7927 .set_curr_task = set_curr_task_fair,
7928 .task_tick = task_tick_fair,
7929 .task_fork = task_fork_fair,
7931 .prio_changed = prio_changed_fair,
7932 .switched_from = switched_from_fair,
7933 .switched_to = switched_to_fair,
7935 .get_rr_interval = get_rr_interval_fair,
7937 #ifdef CONFIG_FAIR_GROUP_SCHED
7938 .task_move_group = task_move_group_fair,
7942 #ifdef CONFIG_SCHED_DEBUG
7943 void print_cfs_stats(struct seq_file *m, int cpu)
7945 struct cfs_rq *cfs_rq;
7948 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7949 print_cfs_rq(m, cpu, cfs_rq);
7954 __init void init_sched_fair_class(void)
7957 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7959 #ifdef CONFIG_NO_HZ_COMMON
7960 nohz.next_balance = jiffies;
7961 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7962 cpu_notifier(sched_ilb_notifier, 0);