2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency = 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity = 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency = 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static int get_update_sysctl_factor(void)
146 unsigned int cpus = min_t(int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling) {
150 case SCHED_TUNABLESCALING_NONE:
153 case SCHED_TUNABLESCALING_LINEAR:
156 case SCHED_TUNABLESCALING_LOG:
158 factor = 1 + ilog2(cpus);
165 static void update_sysctl(void)
167 unsigned int factor = get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity);
172 SET_SYSCTL(sched_latency);
173 SET_SYSCTL(sched_wakeup_granularity);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight *lw)
189 if (likely(lw->inv_weight))
192 w = scale_load_down(lw->weight);
194 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
196 else if (unlikely(!w))
197 lw->inv_weight = WMULT_CONST;
199 lw->inv_weight = WMULT_CONST / w;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
216 u64 fact = scale_load_down(weight);
217 int shift = WMULT_SHIFT;
219 __update_inv_weight(lw);
221 if (unlikely(fact >> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact = (u64)(u32)fact * lw->inv_weight;
236 return mul_u64_u32_shr(delta_exec, fact, shift);
240 const struct sched_class fair_sched_class;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct *task_of(struct sched_entity *se)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se));
262 return container_of(se, struct task_struct, se);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
286 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
289 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
291 if (!cfs_rq->on_list) {
293 * Ensure we either appear before our parent (if already
294 * enqueued) or force our parent to appear after us when it is
295 * enqueued. The fact that we always enqueue bottom-up
296 * reduces this to two cases.
298 if (cfs_rq->tg->parent &&
299 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
300 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
303 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
304 &rq_of(cfs_rq)->leaf_cfs_rq_list);
308 /* We should have no load, but we need to update last_decay. */
309 update_cfs_rq_blocked_load(cfs_rq, 0);
313 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
315 if (cfs_rq->on_list) {
316 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
321 /* Iterate thr' all leaf cfs_rq's on a runqueue */
322 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
323 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
325 /* Do the two (enqueued) entities belong to the same group ? */
326 static inline struct cfs_rq *
327 is_same_group(struct sched_entity *se, struct sched_entity *pse)
329 if (se->cfs_rq == pse->cfs_rq)
335 static inline struct sched_entity *parent_entity(struct sched_entity *se)
341 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
343 int se_depth, pse_depth;
346 * preemption test can be made between sibling entities who are in the
347 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
348 * both tasks until we find their ancestors who are siblings of common
352 /* First walk up until both entities are at same depth */
353 se_depth = (*se)->depth;
354 pse_depth = (*pse)->depth;
356 while (se_depth > pse_depth) {
358 *se = parent_entity(*se);
361 while (pse_depth > se_depth) {
363 *pse = parent_entity(*pse);
366 while (!is_same_group(*se, *pse)) {
367 *se = parent_entity(*se);
368 *pse = parent_entity(*pse);
372 #else /* !CONFIG_FAIR_GROUP_SCHED */
374 static inline struct task_struct *task_of(struct sched_entity *se)
376 return container_of(se, struct task_struct, se);
379 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
381 return container_of(cfs_rq, struct rq, cfs);
384 #define entity_is_task(se) 1
386 #define for_each_sched_entity(se) \
387 for (; se; se = NULL)
389 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
391 return &task_rq(p)->cfs;
394 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
396 struct task_struct *p = task_of(se);
397 struct rq *rq = task_rq(p);
402 /* runqueue "owned" by this group */
403 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
408 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
412 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
416 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
417 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
419 static inline struct sched_entity *parent_entity(struct sched_entity *se)
425 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
429 #endif /* CONFIG_FAIR_GROUP_SCHED */
431 static __always_inline
432 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
434 /**************************************************************
435 * Scheduling class tree data structure manipulation methods:
438 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
440 s64 delta = (s64)(vruntime - max_vruntime);
442 max_vruntime = vruntime;
447 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
449 s64 delta = (s64)(vruntime - min_vruntime);
451 min_vruntime = vruntime;
456 static inline int entity_before(struct sched_entity *a,
457 struct sched_entity *b)
459 return (s64)(a->vruntime - b->vruntime) < 0;
462 static void update_min_vruntime(struct cfs_rq *cfs_rq)
464 u64 vruntime = cfs_rq->min_vruntime;
467 vruntime = cfs_rq->curr->vruntime;
469 if (cfs_rq->rb_leftmost) {
470 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
475 vruntime = se->vruntime;
477 vruntime = min_vruntime(vruntime, se->vruntime);
480 /* ensure we never gain time by being placed backwards. */
481 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
484 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
489 * Enqueue an entity into the rb-tree:
491 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
493 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
494 struct rb_node *parent = NULL;
495 struct sched_entity *entry;
499 * Find the right place in the rbtree:
503 entry = rb_entry(parent, struct sched_entity, run_node);
505 * We dont care about collisions. Nodes with
506 * the same key stay together.
508 if (entity_before(se, entry)) {
509 link = &parent->rb_left;
511 link = &parent->rb_right;
517 * Maintain a cache of leftmost tree entries (it is frequently
521 cfs_rq->rb_leftmost = &se->run_node;
523 rb_link_node(&se->run_node, parent, link);
524 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
527 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
529 if (cfs_rq->rb_leftmost == &se->run_node) {
530 struct rb_node *next_node;
532 next_node = rb_next(&se->run_node);
533 cfs_rq->rb_leftmost = next_node;
536 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
539 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
541 struct rb_node *left = cfs_rq->rb_leftmost;
546 return rb_entry(left, struct sched_entity, run_node);
549 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
551 struct rb_node *next = rb_next(&se->run_node);
556 return rb_entry(next, struct sched_entity, run_node);
559 #ifdef CONFIG_SCHED_DEBUG
560 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
562 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
567 return rb_entry(last, struct sched_entity, run_node);
570 /**************************************************************
571 * Scheduling class statistics methods:
574 int sched_proc_update_handler(struct ctl_table *table, int write,
575 void __user *buffer, size_t *lenp,
578 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
579 int factor = get_update_sysctl_factor();
584 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
585 sysctl_sched_min_granularity);
587 #define WRT_SYSCTL(name) \
588 (normalized_sysctl_##name = sysctl_##name / (factor))
589 WRT_SYSCTL(sched_min_granularity);
590 WRT_SYSCTL(sched_latency);
591 WRT_SYSCTL(sched_wakeup_granularity);
601 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
603 if (unlikely(se->load.weight != NICE_0_LOAD))
604 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
610 * The idea is to set a period in which each task runs once.
612 * When there are too many tasks (sched_nr_latency) we have to stretch
613 * this period because otherwise the slices get too small.
615 * p = (nr <= nl) ? l : l*nr/nl
617 static u64 __sched_period(unsigned long nr_running)
619 u64 period = sysctl_sched_latency;
620 unsigned long nr_latency = sched_nr_latency;
622 if (unlikely(nr_running > nr_latency)) {
623 period = sysctl_sched_min_granularity;
624 period *= nr_running;
631 * We calculate the wall-time slice from the period by taking a part
632 * proportional to the weight.
636 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
638 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
640 for_each_sched_entity(se) {
641 struct load_weight *load;
642 struct load_weight lw;
644 cfs_rq = cfs_rq_of(se);
645 load = &cfs_rq->load;
647 if (unlikely(!se->on_rq)) {
650 update_load_add(&lw, se->load.weight);
653 slice = __calc_delta(slice, se->load.weight, load);
659 * We calculate the vruntime slice of a to-be-inserted task.
663 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
665 return calc_delta_fair(sched_slice(cfs_rq, se), se);
669 static int select_idle_sibling(struct task_struct *p, int cpu);
670 static unsigned long task_h_load(struct task_struct *p);
672 static inline void __update_task_entity_contrib(struct sched_entity *se);
674 /* Give new task start runnable values to heavy its load in infant time */
675 void init_task_runnable_average(struct task_struct *p)
679 p->se.avg.decay_count = 0;
680 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
681 p->se.avg.runnable_avg_sum = slice;
682 p->se.avg.runnable_avg_period = slice;
683 __update_task_entity_contrib(&p->se);
686 void init_task_runnable_average(struct task_struct *p)
692 * Update the current task's runtime statistics.
694 static void update_curr(struct cfs_rq *cfs_rq)
696 struct sched_entity *curr = cfs_rq->curr;
697 u64 now = rq_clock_task(rq_of(cfs_rq));
703 delta_exec = now - curr->exec_start;
704 if (unlikely((s64)delta_exec <= 0))
707 curr->exec_start = now;
709 schedstat_set(curr->statistics.exec_max,
710 max(delta_exec, curr->statistics.exec_max));
712 curr->sum_exec_runtime += delta_exec;
713 schedstat_add(cfs_rq, exec_clock, delta_exec);
715 curr->vruntime += calc_delta_fair(delta_exec, curr);
716 update_min_vruntime(cfs_rq);
718 if (entity_is_task(curr)) {
719 struct task_struct *curtask = task_of(curr);
721 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
722 cpuacct_charge(curtask, delta_exec);
723 account_group_exec_runtime(curtask, delta_exec);
726 account_cfs_rq_runtime(cfs_rq, delta_exec);
730 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
732 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
736 * Task is being enqueued - update stats:
738 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
741 * Are we enqueueing a waiting task? (for current tasks
742 * a dequeue/enqueue event is a NOP)
744 if (se != cfs_rq->curr)
745 update_stats_wait_start(cfs_rq, se);
749 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
751 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
752 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
753 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
754 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
755 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
756 #ifdef CONFIG_SCHEDSTATS
757 if (entity_is_task(se)) {
758 trace_sched_stat_wait(task_of(se),
759 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
762 schedstat_set(se->statistics.wait_start, 0);
766 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
769 * Mark the end of the wait period if dequeueing a
772 if (se != cfs_rq->curr)
773 update_stats_wait_end(cfs_rq, se);
777 * We are picking a new current task - update its stats:
780 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
783 * We are starting a new run period:
785 se->exec_start = rq_clock_task(rq_of(cfs_rq));
788 /**************************************************
789 * Scheduling class queueing methods:
792 #ifdef CONFIG_NUMA_BALANCING
794 * Approximate time to scan a full NUMA task in ms. The task scan period is
795 * calculated based on the tasks virtual memory size and
796 * numa_balancing_scan_size.
798 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
799 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
801 /* Portion of address space to scan in MB */
802 unsigned int sysctl_numa_balancing_scan_size = 256;
804 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
805 unsigned int sysctl_numa_balancing_scan_delay = 1000;
807 static unsigned int task_nr_scan_windows(struct task_struct *p)
809 unsigned long rss = 0;
810 unsigned long nr_scan_pages;
813 * Calculations based on RSS as non-present and empty pages are skipped
814 * by the PTE scanner and NUMA hinting faults should be trapped based
817 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
818 rss = get_mm_rss(p->mm);
822 rss = round_up(rss, nr_scan_pages);
823 return rss / nr_scan_pages;
826 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
827 #define MAX_SCAN_WINDOW 2560
829 static unsigned int task_scan_min(struct task_struct *p)
831 unsigned int scan_size = ACCESS_ONCE(sysctl_numa_balancing_scan_size);
832 unsigned int scan, floor;
833 unsigned int windows = 1;
835 if (scan_size < MAX_SCAN_WINDOW)
836 windows = MAX_SCAN_WINDOW / scan_size;
837 floor = 1000 / windows;
839 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
840 return max_t(unsigned int, floor, scan);
843 static unsigned int task_scan_max(struct task_struct *p)
845 unsigned int smin = task_scan_min(p);
848 /* Watch for min being lower than max due to floor calculations */
849 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
850 return max(smin, smax);
853 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
855 rq->nr_numa_running += (p->numa_preferred_nid != -1);
856 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
859 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
861 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
862 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
868 spinlock_t lock; /* nr_tasks, tasks */
871 struct list_head task_list;
874 nodemask_t active_nodes;
875 unsigned long total_faults;
877 * Faults_cpu is used to decide whether memory should move
878 * towards the CPU. As a consequence, these stats are weighted
879 * more by CPU use than by memory faults.
881 unsigned long *faults_cpu;
882 unsigned long faults[0];
885 /* Shared or private faults. */
886 #define NR_NUMA_HINT_FAULT_TYPES 2
888 /* Memory and CPU locality */
889 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
891 /* Averaged statistics, and temporary buffers. */
892 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
894 pid_t task_numa_group_id(struct task_struct *p)
896 return p->numa_group ? p->numa_group->gid : 0;
899 static inline int task_faults_idx(int nid, int priv)
901 return NR_NUMA_HINT_FAULT_TYPES * nid + priv;
904 static inline unsigned long task_faults(struct task_struct *p, int nid)
906 if (!p->numa_faults_memory)
909 return p->numa_faults_memory[task_faults_idx(nid, 0)] +
910 p->numa_faults_memory[task_faults_idx(nid, 1)];
913 static inline unsigned long group_faults(struct task_struct *p, int nid)
918 return p->numa_group->faults[task_faults_idx(nid, 0)] +
919 p->numa_group->faults[task_faults_idx(nid, 1)];
922 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
924 return group->faults_cpu[task_faults_idx(nid, 0)] +
925 group->faults_cpu[task_faults_idx(nid, 1)];
929 * These return the fraction of accesses done by a particular task, or
930 * task group, on a particular numa node. The group weight is given a
931 * larger multiplier, in order to group tasks together that are almost
932 * evenly spread out between numa nodes.
934 static inline unsigned long task_weight(struct task_struct *p, int nid)
936 unsigned long total_faults;
938 if (!p->numa_faults_memory)
941 total_faults = p->total_numa_faults;
946 return 1000 * task_faults(p, nid) / total_faults;
949 static inline unsigned long group_weight(struct task_struct *p, int nid)
951 if (!p->numa_group || !p->numa_group->total_faults)
954 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
957 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
958 int src_nid, int dst_cpu)
960 struct numa_group *ng = p->numa_group;
961 int dst_nid = cpu_to_node(dst_cpu);
962 int last_cpupid, this_cpupid;
964 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
967 * Multi-stage node selection is used in conjunction with a periodic
968 * migration fault to build a temporal task<->page relation. By using
969 * a two-stage filter we remove short/unlikely relations.
971 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
972 * a task's usage of a particular page (n_p) per total usage of this
973 * page (n_t) (in a given time-span) to a probability.
975 * Our periodic faults will sample this probability and getting the
976 * same result twice in a row, given these samples are fully
977 * independent, is then given by P(n)^2, provided our sample period
978 * is sufficiently short compared to the usage pattern.
980 * This quadric squishes small probabilities, making it less likely we
981 * act on an unlikely task<->page relation.
983 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
984 if (!cpupid_pid_unset(last_cpupid) &&
985 cpupid_to_nid(last_cpupid) != dst_nid)
988 /* Always allow migrate on private faults */
989 if (cpupid_match_pid(p, last_cpupid))
992 /* A shared fault, but p->numa_group has not been set up yet. */
997 * Do not migrate if the destination is not a node that
998 * is actively used by this numa group.
1000 if (!node_isset(dst_nid, ng->active_nodes))
1004 * Source is a node that is not actively used by this
1005 * numa group, while the destination is. Migrate.
1007 if (!node_isset(src_nid, ng->active_nodes))
1011 * Both source and destination are nodes in active
1012 * use by this numa group. Maximize memory bandwidth
1013 * by migrating from more heavily used groups, to less
1014 * heavily used ones, spreading the load around.
1015 * Use a 1/4 hysteresis to avoid spurious page movement.
1017 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1020 static unsigned long weighted_cpuload(const int cpu);
1021 static unsigned long source_load(int cpu, int type);
1022 static unsigned long target_load(int cpu, int type);
1023 static unsigned long capacity_of(int cpu);
1024 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1026 /* Cached statistics for all CPUs within a node */
1028 unsigned long nr_running;
1031 /* Total compute capacity of CPUs on a node */
1032 unsigned long compute_capacity;
1034 /* Approximate capacity in terms of runnable tasks on a node */
1035 unsigned long task_capacity;
1036 int has_free_capacity;
1040 * XXX borrowed from update_sg_lb_stats
1042 static void update_numa_stats(struct numa_stats *ns, int nid)
1044 int smt, cpu, cpus = 0;
1045 unsigned long capacity;
1047 memset(ns, 0, sizeof(*ns));
1048 for_each_cpu(cpu, cpumask_of_node(nid)) {
1049 struct rq *rq = cpu_rq(cpu);
1051 ns->nr_running += rq->nr_running;
1052 ns->load += weighted_cpuload(cpu);
1053 ns->compute_capacity += capacity_of(cpu);
1059 * If we raced with hotplug and there are no CPUs left in our mask
1060 * the @ns structure is NULL'ed and task_numa_compare() will
1061 * not find this node attractive.
1063 * We'll either bail at !has_free_capacity, or we'll detect a huge
1064 * imbalance and bail there.
1069 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1070 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1071 capacity = cpus / smt; /* cores */
1073 ns->task_capacity = min_t(unsigned, capacity,
1074 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1075 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1078 struct task_numa_env {
1079 struct task_struct *p;
1081 int src_cpu, src_nid;
1082 int dst_cpu, dst_nid;
1084 struct numa_stats src_stats, dst_stats;
1088 struct task_struct *best_task;
1093 static void task_numa_assign(struct task_numa_env *env,
1094 struct task_struct *p, long imp)
1097 put_task_struct(env->best_task);
1102 env->best_imp = imp;
1103 env->best_cpu = env->dst_cpu;
1106 static bool load_too_imbalanced(long src_load, long dst_load,
1107 struct task_numa_env *env)
1110 long orig_src_load, orig_dst_load;
1111 long src_capacity, dst_capacity;
1114 * The load is corrected for the CPU capacity available on each node.
1117 * ------------ vs ---------
1118 * src_capacity dst_capacity
1120 src_capacity = env->src_stats.compute_capacity;
1121 dst_capacity = env->dst_stats.compute_capacity;
1123 /* We care about the slope of the imbalance, not the direction. */
1124 if (dst_load < src_load)
1125 swap(dst_load, src_load);
1127 /* Is the difference below the threshold? */
1128 imb = dst_load * src_capacity * 100 -
1129 src_load * dst_capacity * env->imbalance_pct;
1134 * The imbalance is above the allowed threshold.
1135 * Compare it with the old imbalance.
1137 orig_src_load = env->src_stats.load;
1138 orig_dst_load = env->dst_stats.load;
1140 if (orig_dst_load < orig_src_load)
1141 swap(orig_dst_load, orig_src_load);
1143 old_imb = orig_dst_load * src_capacity * 100 -
1144 orig_src_load * dst_capacity * env->imbalance_pct;
1146 /* Would this change make things worse? */
1147 return (imb > old_imb);
1151 * This checks if the overall compute and NUMA accesses of the system would
1152 * be improved if the source tasks was migrated to the target dst_cpu taking
1153 * into account that it might be best if task running on the dst_cpu should
1154 * be exchanged with the source task
1156 static void task_numa_compare(struct task_numa_env *env,
1157 long taskimp, long groupimp)
1159 struct rq *src_rq = cpu_rq(env->src_cpu);
1160 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1161 struct task_struct *cur;
1162 long src_load, dst_load;
1164 long imp = env->p->numa_group ? groupimp : taskimp;
1169 raw_spin_lock_irq(&dst_rq->lock);
1172 * No need to move the exiting task, and this ensures that ->curr
1173 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1174 * is safe under RCU read lock.
1175 * Note that rcu_read_lock() itself can't protect from the final
1176 * put_task_struct() after the last schedule().
1178 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1180 raw_spin_unlock_irq(&dst_rq->lock);
1183 * "imp" is the fault differential for the source task between the
1184 * source and destination node. Calculate the total differential for
1185 * the source task and potential destination task. The more negative
1186 * the value is, the more rmeote accesses that would be expected to
1187 * be incurred if the tasks were swapped.
1190 /* Skip this swap candidate if cannot move to the source cpu */
1191 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1195 * If dst and source tasks are in the same NUMA group, or not
1196 * in any group then look only at task weights.
1198 if (cur->numa_group == env->p->numa_group) {
1199 imp = taskimp + task_weight(cur, env->src_nid) -
1200 task_weight(cur, env->dst_nid);
1202 * Add some hysteresis to prevent swapping the
1203 * tasks within a group over tiny differences.
1205 if (cur->numa_group)
1209 * Compare the group weights. If a task is all by
1210 * itself (not part of a group), use the task weight
1213 if (cur->numa_group)
1214 imp += group_weight(cur, env->src_nid) -
1215 group_weight(cur, env->dst_nid);
1217 imp += task_weight(cur, env->src_nid) -
1218 task_weight(cur, env->dst_nid);
1222 if (imp <= env->best_imp && moveimp <= env->best_imp)
1226 /* Is there capacity at our destination? */
1227 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1228 !env->dst_stats.has_free_capacity)
1234 /* Balance doesn't matter much if we're running a task per cpu */
1235 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1236 dst_rq->nr_running == 1)
1240 * In the overloaded case, try and keep the load balanced.
1243 load = task_h_load(env->p);
1244 dst_load = env->dst_stats.load + load;
1245 src_load = env->src_stats.load - load;
1247 if (moveimp > imp && moveimp > env->best_imp) {
1249 * If the improvement from just moving env->p direction is
1250 * better than swapping tasks around, check if a move is
1251 * possible. Store a slightly smaller score than moveimp,
1252 * so an actually idle CPU will win.
1254 if (!load_too_imbalanced(src_load, dst_load, env)) {
1261 if (imp <= env->best_imp)
1265 load = task_h_load(cur);
1270 if (load_too_imbalanced(src_load, dst_load, env))
1274 * One idle CPU per node is evaluated for a task numa move.
1275 * Call select_idle_sibling to maybe find a better one.
1278 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1281 task_numa_assign(env, cur, imp);
1286 static void task_numa_find_cpu(struct task_numa_env *env,
1287 long taskimp, long groupimp)
1291 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1292 /* Skip this CPU if the source task cannot migrate */
1293 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1297 task_numa_compare(env, taskimp, groupimp);
1301 static int task_numa_migrate(struct task_struct *p)
1303 struct task_numa_env env = {
1306 .src_cpu = task_cpu(p),
1307 .src_nid = task_node(p),
1309 .imbalance_pct = 112,
1315 struct sched_domain *sd;
1316 unsigned long taskweight, groupweight;
1318 long taskimp, groupimp;
1321 * Pick the lowest SD_NUMA domain, as that would have the smallest
1322 * imbalance and would be the first to start moving tasks about.
1324 * And we want to avoid any moving of tasks about, as that would create
1325 * random movement of tasks -- counter the numa conditions we're trying
1329 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1331 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1335 * Cpusets can break the scheduler domain tree into smaller
1336 * balance domains, some of which do not cross NUMA boundaries.
1337 * Tasks that are "trapped" in such domains cannot be migrated
1338 * elsewhere, so there is no point in (re)trying.
1340 if (unlikely(!sd)) {
1341 p->numa_preferred_nid = task_node(p);
1345 taskweight = task_weight(p, env.src_nid);
1346 groupweight = group_weight(p, env.src_nid);
1347 update_numa_stats(&env.src_stats, env.src_nid);
1348 env.dst_nid = p->numa_preferred_nid;
1349 taskimp = task_weight(p, env.dst_nid) - taskweight;
1350 groupimp = group_weight(p, env.dst_nid) - groupweight;
1351 update_numa_stats(&env.dst_stats, env.dst_nid);
1353 /* Try to find a spot on the preferred nid. */
1354 task_numa_find_cpu(&env, taskimp, groupimp);
1356 /* No space available on the preferred nid. Look elsewhere. */
1357 if (env.best_cpu == -1) {
1358 for_each_online_node(nid) {
1359 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1362 /* Only consider nodes where both task and groups benefit */
1363 taskimp = task_weight(p, nid) - taskweight;
1364 groupimp = group_weight(p, nid) - groupweight;
1365 if (taskimp < 0 && groupimp < 0)
1369 update_numa_stats(&env.dst_stats, env.dst_nid);
1370 task_numa_find_cpu(&env, taskimp, groupimp);
1375 * If the task is part of a workload that spans multiple NUMA nodes,
1376 * and is migrating into one of the workload's active nodes, remember
1377 * this node as the task's preferred numa node, so the workload can
1379 * A task that migrated to a second choice node will be better off
1380 * trying for a better one later. Do not set the preferred node here.
1382 if (p->numa_group) {
1383 if (env.best_cpu == -1)
1388 if (node_isset(nid, p->numa_group->active_nodes))
1389 sched_setnuma(p, env.dst_nid);
1392 /* No better CPU than the current one was found. */
1393 if (env.best_cpu == -1)
1397 * Reset the scan period if the task is being rescheduled on an
1398 * alternative node to recheck if the tasks is now properly placed.
1400 p->numa_scan_period = task_scan_min(p);
1402 if (env.best_task == NULL) {
1403 ret = migrate_task_to(p, env.best_cpu);
1405 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1409 ret = migrate_swap(p, env.best_task);
1411 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1412 put_task_struct(env.best_task);
1416 /* Attempt to migrate a task to a CPU on the preferred node. */
1417 static void numa_migrate_preferred(struct task_struct *p)
1419 unsigned long interval = HZ;
1421 /* This task has no NUMA fault statistics yet */
1422 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1425 /* Periodically retry migrating the task to the preferred node */
1426 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1427 p->numa_migrate_retry = jiffies + interval;
1429 /* Success if task is already running on preferred CPU */
1430 if (task_node(p) == p->numa_preferred_nid)
1433 /* Otherwise, try migrate to a CPU on the preferred node */
1434 task_numa_migrate(p);
1438 * Find the nodes on which the workload is actively running. We do this by
1439 * tracking the nodes from which NUMA hinting faults are triggered. This can
1440 * be different from the set of nodes where the workload's memory is currently
1443 * The bitmask is used to make smarter decisions on when to do NUMA page
1444 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1445 * are added when they cause over 6/16 of the maximum number of faults, but
1446 * only removed when they drop below 3/16.
1448 static void update_numa_active_node_mask(struct numa_group *numa_group)
1450 unsigned long faults, max_faults = 0;
1453 for_each_online_node(nid) {
1454 faults = group_faults_cpu(numa_group, nid);
1455 if (faults > max_faults)
1456 max_faults = faults;
1459 for_each_online_node(nid) {
1460 faults = group_faults_cpu(numa_group, nid);
1461 if (!node_isset(nid, numa_group->active_nodes)) {
1462 if (faults > max_faults * 6 / 16)
1463 node_set(nid, numa_group->active_nodes);
1464 } else if (faults < max_faults * 3 / 16)
1465 node_clear(nid, numa_group->active_nodes);
1470 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1471 * increments. The more local the fault statistics are, the higher the scan
1472 * period will be for the next scan window. If local/(local+remote) ratio is
1473 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1474 * the scan period will decrease. Aim for 70% local accesses.
1476 #define NUMA_PERIOD_SLOTS 10
1477 #define NUMA_PERIOD_THRESHOLD 7
1480 * Increase the scan period (slow down scanning) if the majority of
1481 * our memory is already on our local node, or if the majority of
1482 * the page accesses are shared with other processes.
1483 * Otherwise, decrease the scan period.
1485 static void update_task_scan_period(struct task_struct *p,
1486 unsigned long shared, unsigned long private)
1488 unsigned int period_slot;
1492 unsigned long remote = p->numa_faults_locality[0];
1493 unsigned long local = p->numa_faults_locality[1];
1496 * If there were no record hinting faults then either the task is
1497 * completely idle or all activity is areas that are not of interest
1498 * to automatic numa balancing. Scan slower
1500 if (local + shared == 0) {
1501 p->numa_scan_period = min(p->numa_scan_period_max,
1502 p->numa_scan_period << 1);
1504 p->mm->numa_next_scan = jiffies +
1505 msecs_to_jiffies(p->numa_scan_period);
1511 * Prepare to scale scan period relative to the current period.
1512 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1513 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1514 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1516 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1517 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1518 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1519 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1522 diff = slot * period_slot;
1524 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1527 * Scale scan rate increases based on sharing. There is an
1528 * inverse relationship between the degree of sharing and
1529 * the adjustment made to the scanning period. Broadly
1530 * speaking the intent is that there is little point
1531 * scanning faster if shared accesses dominate as it may
1532 * simply bounce migrations uselessly
1534 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1535 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1538 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1539 task_scan_min(p), task_scan_max(p));
1540 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1544 * Get the fraction of time the task has been running since the last
1545 * NUMA placement cycle. The scheduler keeps similar statistics, but
1546 * decays those on a 32ms period, which is orders of magnitude off
1547 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1548 * stats only if the task is so new there are no NUMA statistics yet.
1550 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1552 u64 runtime, delta, now;
1553 /* Use the start of this time slice to avoid calculations. */
1554 now = p->se.exec_start;
1555 runtime = p->se.sum_exec_runtime;
1557 if (p->last_task_numa_placement) {
1558 delta = runtime - p->last_sum_exec_runtime;
1559 *period = now - p->last_task_numa_placement;
1561 delta = p->se.avg.runnable_avg_sum;
1562 *period = p->se.avg.runnable_avg_period;
1565 p->last_sum_exec_runtime = runtime;
1566 p->last_task_numa_placement = now;
1571 static void task_numa_placement(struct task_struct *p)
1573 int seq, nid, max_nid = -1, max_group_nid = -1;
1574 unsigned long max_faults = 0, max_group_faults = 0;
1575 unsigned long fault_types[2] = { 0, 0 };
1576 unsigned long total_faults;
1577 u64 runtime, period;
1578 spinlock_t *group_lock = NULL;
1580 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1581 if (p->numa_scan_seq == seq)
1583 p->numa_scan_seq = seq;
1584 p->numa_scan_period_max = task_scan_max(p);
1586 total_faults = p->numa_faults_locality[0] +
1587 p->numa_faults_locality[1];
1588 runtime = numa_get_avg_runtime(p, &period);
1590 /* If the task is part of a group prevent parallel updates to group stats */
1591 if (p->numa_group) {
1592 group_lock = &p->numa_group->lock;
1593 spin_lock_irq(group_lock);
1596 /* Find the node with the highest number of faults */
1597 for_each_online_node(nid) {
1598 unsigned long faults = 0, group_faults = 0;
1601 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1602 long diff, f_diff, f_weight;
1604 i = task_faults_idx(nid, priv);
1606 /* Decay existing window, copy faults since last scan */
1607 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1608 fault_types[priv] += p->numa_faults_buffer_memory[i];
1609 p->numa_faults_buffer_memory[i] = 0;
1612 * Normalize the faults_from, so all tasks in a group
1613 * count according to CPU use, instead of by the raw
1614 * number of faults. Tasks with little runtime have
1615 * little over-all impact on throughput, and thus their
1616 * faults are less important.
1618 f_weight = div64_u64(runtime << 16, period + 1);
1619 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1621 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1622 p->numa_faults_buffer_cpu[i] = 0;
1624 p->numa_faults_memory[i] += diff;
1625 p->numa_faults_cpu[i] += f_diff;
1626 faults += p->numa_faults_memory[i];
1627 p->total_numa_faults += diff;
1628 if (p->numa_group) {
1629 /* safe because we can only change our own group */
1630 p->numa_group->faults[i] += diff;
1631 p->numa_group->faults_cpu[i] += f_diff;
1632 p->numa_group->total_faults += diff;
1633 group_faults += p->numa_group->faults[i];
1637 if (faults > max_faults) {
1638 max_faults = faults;
1642 if (group_faults > max_group_faults) {
1643 max_group_faults = group_faults;
1644 max_group_nid = nid;
1648 update_task_scan_period(p, fault_types[0], fault_types[1]);
1650 if (p->numa_group) {
1651 update_numa_active_node_mask(p->numa_group);
1652 spin_unlock_irq(group_lock);
1653 max_nid = max_group_nid;
1657 /* Set the new preferred node */
1658 if (max_nid != p->numa_preferred_nid)
1659 sched_setnuma(p, max_nid);
1661 if (task_node(p) != p->numa_preferred_nid)
1662 numa_migrate_preferred(p);
1666 static inline int get_numa_group(struct numa_group *grp)
1668 return atomic_inc_not_zero(&grp->refcount);
1671 static inline void put_numa_group(struct numa_group *grp)
1673 if (atomic_dec_and_test(&grp->refcount))
1674 kfree_rcu(grp, rcu);
1677 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1680 struct numa_group *grp, *my_grp;
1681 struct task_struct *tsk;
1683 int cpu = cpupid_to_cpu(cpupid);
1686 if (unlikely(!p->numa_group)) {
1687 unsigned int size = sizeof(struct numa_group) +
1688 4*nr_node_ids*sizeof(unsigned long);
1690 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1694 atomic_set(&grp->refcount, 1);
1695 spin_lock_init(&grp->lock);
1696 INIT_LIST_HEAD(&grp->task_list);
1698 /* Second half of the array tracks nids where faults happen */
1699 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1702 node_set(task_node(current), grp->active_nodes);
1704 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1705 grp->faults[i] = p->numa_faults_memory[i];
1707 grp->total_faults = p->total_numa_faults;
1709 list_add(&p->numa_entry, &grp->task_list);
1711 rcu_assign_pointer(p->numa_group, grp);
1715 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1717 if (!cpupid_match_pid(tsk, cpupid))
1720 grp = rcu_dereference(tsk->numa_group);
1724 my_grp = p->numa_group;
1729 * Only join the other group if its bigger; if we're the bigger group,
1730 * the other task will join us.
1732 if (my_grp->nr_tasks > grp->nr_tasks)
1736 * Tie-break on the grp address.
1738 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1741 /* Always join threads in the same process. */
1742 if (tsk->mm == current->mm)
1745 /* Simple filter to avoid false positives due to PID collisions */
1746 if (flags & TNF_SHARED)
1749 /* Update priv based on whether false sharing was detected */
1752 if (join && !get_numa_group(grp))
1760 BUG_ON(irqs_disabled());
1761 double_lock_irq(&my_grp->lock, &grp->lock);
1763 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1764 my_grp->faults[i] -= p->numa_faults_memory[i];
1765 grp->faults[i] += p->numa_faults_memory[i];
1767 my_grp->total_faults -= p->total_numa_faults;
1768 grp->total_faults += p->total_numa_faults;
1770 list_move(&p->numa_entry, &grp->task_list);
1774 spin_unlock(&my_grp->lock);
1775 spin_unlock_irq(&grp->lock);
1777 rcu_assign_pointer(p->numa_group, grp);
1779 put_numa_group(my_grp);
1787 void task_numa_free(struct task_struct *p)
1789 struct numa_group *grp = p->numa_group;
1790 void *numa_faults = p->numa_faults_memory;
1791 unsigned long flags;
1795 spin_lock_irqsave(&grp->lock, flags);
1796 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1797 grp->faults[i] -= p->numa_faults_memory[i];
1798 grp->total_faults -= p->total_numa_faults;
1800 list_del(&p->numa_entry);
1802 spin_unlock_irqrestore(&grp->lock, flags);
1803 RCU_INIT_POINTER(p->numa_group, NULL);
1804 put_numa_group(grp);
1807 p->numa_faults_memory = NULL;
1808 p->numa_faults_buffer_memory = NULL;
1809 p->numa_faults_cpu= NULL;
1810 p->numa_faults_buffer_cpu = NULL;
1815 * Got a PROT_NONE fault for a page on @node.
1817 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1819 struct task_struct *p = current;
1820 bool migrated = flags & TNF_MIGRATED;
1821 int cpu_node = task_node(current);
1822 int local = !!(flags & TNF_FAULT_LOCAL);
1825 if (!numabalancing_enabled)
1828 /* for example, ksmd faulting in a user's mm */
1832 /* Allocate buffer to track faults on a per-node basis */
1833 if (unlikely(!p->numa_faults_memory)) {
1834 int size = sizeof(*p->numa_faults_memory) *
1835 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1837 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1838 if (!p->numa_faults_memory)
1841 BUG_ON(p->numa_faults_buffer_memory);
1843 * The averaged statistics, shared & private, memory & cpu,
1844 * occupy the first half of the array. The second half of the
1845 * array is for current counters, which are averaged into the
1846 * first set by task_numa_placement.
1848 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1849 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1850 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1851 p->total_numa_faults = 0;
1852 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1856 * First accesses are treated as private, otherwise consider accesses
1857 * to be private if the accessing pid has not changed
1859 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1862 priv = cpupid_match_pid(p, last_cpupid);
1863 if (!priv && !(flags & TNF_NO_GROUP))
1864 task_numa_group(p, last_cpupid, flags, &priv);
1868 * If a workload spans multiple NUMA nodes, a shared fault that
1869 * occurs wholly within the set of nodes that the workload is
1870 * actively using should be counted as local. This allows the
1871 * scan rate to slow down when a workload has settled down.
1873 if (!priv && !local && p->numa_group &&
1874 node_isset(cpu_node, p->numa_group->active_nodes) &&
1875 node_isset(mem_node, p->numa_group->active_nodes))
1878 task_numa_placement(p);
1881 * Retry task to preferred node migration periodically, in case it
1882 * case it previously failed, or the scheduler moved us.
1884 if (time_after(jiffies, p->numa_migrate_retry))
1885 numa_migrate_preferred(p);
1888 p->numa_pages_migrated += pages;
1890 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1891 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1892 p->numa_faults_locality[local] += pages;
1895 static void reset_ptenuma_scan(struct task_struct *p)
1897 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1898 p->mm->numa_scan_offset = 0;
1902 * The expensive part of numa migration is done from task_work context.
1903 * Triggered from task_tick_numa().
1905 void task_numa_work(struct callback_head *work)
1907 unsigned long migrate, next_scan, now = jiffies;
1908 struct task_struct *p = current;
1909 struct mm_struct *mm = p->mm;
1910 struct vm_area_struct *vma;
1911 unsigned long start, end;
1912 unsigned long nr_pte_updates = 0;
1915 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1917 work->next = work; /* protect against double add */
1919 * Who cares about NUMA placement when they're dying.
1921 * NOTE: make sure not to dereference p->mm before this check,
1922 * exit_task_work() happens _after_ exit_mm() so we could be called
1923 * without p->mm even though we still had it when we enqueued this
1926 if (p->flags & PF_EXITING)
1929 if (!mm->numa_next_scan) {
1930 mm->numa_next_scan = now +
1931 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1935 * Enforce maximal scan/migration frequency..
1937 migrate = mm->numa_next_scan;
1938 if (time_before(now, migrate))
1941 if (p->numa_scan_period == 0) {
1942 p->numa_scan_period_max = task_scan_max(p);
1943 p->numa_scan_period = task_scan_min(p);
1946 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1947 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1951 * Delay this task enough that another task of this mm will likely win
1952 * the next time around.
1954 p->node_stamp += 2 * TICK_NSEC;
1956 start = mm->numa_scan_offset;
1957 pages = sysctl_numa_balancing_scan_size;
1958 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1962 down_read(&mm->mmap_sem);
1963 vma = find_vma(mm, start);
1965 reset_ptenuma_scan(p);
1969 for (; vma; vma = vma->vm_next) {
1970 if (!vma_migratable(vma) || !vma_policy_mof(vma))
1974 * Shared library pages mapped by multiple processes are not
1975 * migrated as it is expected they are cache replicated. Avoid
1976 * hinting faults in read-only file-backed mappings or the vdso
1977 * as migrating the pages will be of marginal benefit.
1980 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1984 * Skip inaccessible VMAs to avoid any confusion between
1985 * PROT_NONE and NUMA hinting ptes
1987 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1991 start = max(start, vma->vm_start);
1992 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1993 end = min(end, vma->vm_end);
1994 nr_pte_updates += change_prot_numa(vma, start, end);
1997 * Scan sysctl_numa_balancing_scan_size but ensure that
1998 * at least one PTE is updated so that unused virtual
1999 * address space is quickly skipped.
2002 pages -= (end - start) >> PAGE_SHIFT;
2009 } while (end != vma->vm_end);
2014 * It is possible to reach the end of the VMA list but the last few
2015 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2016 * would find the !migratable VMA on the next scan but not reset the
2017 * scanner to the start so check it now.
2020 mm->numa_scan_offset = start;
2022 reset_ptenuma_scan(p);
2023 up_read(&mm->mmap_sem);
2027 * Drive the periodic memory faults..
2029 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2031 struct callback_head *work = &curr->numa_work;
2035 * We don't care about NUMA placement if we don't have memory.
2037 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2041 * Using runtime rather than walltime has the dual advantage that
2042 * we (mostly) drive the selection from busy threads and that the
2043 * task needs to have done some actual work before we bother with
2046 now = curr->se.sum_exec_runtime;
2047 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2049 if (now - curr->node_stamp > period) {
2050 if (!curr->node_stamp)
2051 curr->numa_scan_period = task_scan_min(curr);
2052 curr->node_stamp += period;
2054 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2055 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2056 task_work_add(curr, work, true);
2061 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2065 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2069 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2072 #endif /* CONFIG_NUMA_BALANCING */
2075 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2077 update_load_add(&cfs_rq->load, se->load.weight);
2078 if (!parent_entity(se))
2079 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2081 if (entity_is_task(se)) {
2082 struct rq *rq = rq_of(cfs_rq);
2084 account_numa_enqueue(rq, task_of(se));
2085 list_add(&se->group_node, &rq->cfs_tasks);
2088 cfs_rq->nr_running++;
2092 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2094 update_load_sub(&cfs_rq->load, se->load.weight);
2095 if (!parent_entity(se))
2096 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2097 if (entity_is_task(se)) {
2098 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2099 list_del_init(&se->group_node);
2101 cfs_rq->nr_running--;
2104 #ifdef CONFIG_FAIR_GROUP_SCHED
2106 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2111 * Use this CPU's actual weight instead of the last load_contribution
2112 * to gain a more accurate current total weight. See
2113 * update_cfs_rq_load_contribution().
2115 tg_weight = atomic_long_read(&tg->load_avg);
2116 tg_weight -= cfs_rq->tg_load_contrib;
2117 tg_weight += cfs_rq->load.weight;
2122 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2124 long tg_weight, load, shares;
2126 tg_weight = calc_tg_weight(tg, cfs_rq);
2127 load = cfs_rq->load.weight;
2129 shares = (tg->shares * load);
2131 shares /= tg_weight;
2133 if (shares < MIN_SHARES)
2134 shares = MIN_SHARES;
2135 if (shares > tg->shares)
2136 shares = tg->shares;
2140 # else /* CONFIG_SMP */
2141 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2145 # endif /* CONFIG_SMP */
2146 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2147 unsigned long weight)
2150 /* commit outstanding execution time */
2151 if (cfs_rq->curr == se)
2152 update_curr(cfs_rq);
2153 account_entity_dequeue(cfs_rq, se);
2156 update_load_set(&se->load, weight);
2159 account_entity_enqueue(cfs_rq, se);
2162 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2164 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2166 struct task_group *tg;
2167 struct sched_entity *se;
2171 se = tg->se[cpu_of(rq_of(cfs_rq))];
2172 if (!se || throttled_hierarchy(cfs_rq))
2175 if (likely(se->load.weight == tg->shares))
2178 shares = calc_cfs_shares(cfs_rq, tg);
2180 reweight_entity(cfs_rq_of(se), se, shares);
2182 #else /* CONFIG_FAIR_GROUP_SCHED */
2183 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2186 #endif /* CONFIG_FAIR_GROUP_SCHED */
2190 * We choose a half-life close to 1 scheduling period.
2191 * Note: The tables below are dependent on this value.
2193 #define LOAD_AVG_PERIOD 32
2194 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2195 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2197 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2198 static const u32 runnable_avg_yN_inv[] = {
2199 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2200 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2201 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2202 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2203 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2204 0x85aac367, 0x82cd8698,
2208 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2209 * over-estimates when re-combining.
2211 static const u32 runnable_avg_yN_sum[] = {
2212 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2213 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2214 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2219 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2221 static __always_inline u64 decay_load(u64 val, u64 n)
2223 unsigned int local_n;
2227 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2230 /* after bounds checking we can collapse to 32-bit */
2234 * As y^PERIOD = 1/2, we can combine
2235 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2236 * With a look-up table which covers y^n (n<PERIOD)
2238 * To achieve constant time decay_load.
2240 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2241 val >>= local_n / LOAD_AVG_PERIOD;
2242 local_n %= LOAD_AVG_PERIOD;
2245 val *= runnable_avg_yN_inv[local_n];
2246 /* We don't use SRR here since we always want to round down. */
2251 * For updates fully spanning n periods, the contribution to runnable
2252 * average will be: \Sum 1024*y^n
2254 * We can compute this reasonably efficiently by combining:
2255 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2257 static u32 __compute_runnable_contrib(u64 n)
2261 if (likely(n <= LOAD_AVG_PERIOD))
2262 return runnable_avg_yN_sum[n];
2263 else if (unlikely(n >= LOAD_AVG_MAX_N))
2264 return LOAD_AVG_MAX;
2266 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2268 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2269 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2271 n -= LOAD_AVG_PERIOD;
2272 } while (n > LOAD_AVG_PERIOD);
2274 contrib = decay_load(contrib, n);
2275 return contrib + runnable_avg_yN_sum[n];
2279 * We can represent the historical contribution to runnable average as the
2280 * coefficients of a geometric series. To do this we sub-divide our runnable
2281 * history into segments of approximately 1ms (1024us); label the segment that
2282 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2284 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2286 * (now) (~1ms ago) (~2ms ago)
2288 * Let u_i denote the fraction of p_i that the entity was runnable.
2290 * We then designate the fractions u_i as our co-efficients, yielding the
2291 * following representation of historical load:
2292 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2294 * We choose y based on the with of a reasonably scheduling period, fixing:
2297 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2298 * approximately half as much as the contribution to load within the last ms
2301 * When a period "rolls over" and we have new u_0`, multiplying the previous
2302 * sum again by y is sufficient to update:
2303 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2304 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2306 static __always_inline int __update_entity_runnable_avg(u64 now,
2307 struct sched_avg *sa,
2311 u32 runnable_contrib;
2312 int delta_w, decayed = 0;
2314 delta = now - sa->last_runnable_update;
2316 * This should only happen when time goes backwards, which it
2317 * unfortunately does during sched clock init when we swap over to TSC.
2319 if ((s64)delta < 0) {
2320 sa->last_runnable_update = now;
2325 * Use 1024ns as the unit of measurement since it's a reasonable
2326 * approximation of 1us and fast to compute.
2331 sa->last_runnable_update = now;
2333 /* delta_w is the amount already accumulated against our next period */
2334 delta_w = sa->runnable_avg_period % 1024;
2335 if (delta + delta_w >= 1024) {
2336 /* period roll-over */
2340 * Now that we know we're crossing a period boundary, figure
2341 * out how much from delta we need to complete the current
2342 * period and accrue it.
2344 delta_w = 1024 - delta_w;
2346 sa->runnable_avg_sum += delta_w;
2347 sa->runnable_avg_period += delta_w;
2351 /* Figure out how many additional periods this update spans */
2352 periods = delta / 1024;
2355 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2357 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2360 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2361 runnable_contrib = __compute_runnable_contrib(periods);
2363 sa->runnable_avg_sum += runnable_contrib;
2364 sa->runnable_avg_period += runnable_contrib;
2367 /* Remainder of delta accrued against u_0` */
2369 sa->runnable_avg_sum += delta;
2370 sa->runnable_avg_period += delta;
2375 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2376 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2378 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2379 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2381 decays -= se->avg.decay_count;
2385 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2386 se->avg.decay_count = 0;
2391 #ifdef CONFIG_FAIR_GROUP_SCHED
2392 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2395 struct task_group *tg = cfs_rq->tg;
2398 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2399 tg_contrib -= cfs_rq->tg_load_contrib;
2404 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2405 atomic_long_add(tg_contrib, &tg->load_avg);
2406 cfs_rq->tg_load_contrib += tg_contrib;
2411 * Aggregate cfs_rq runnable averages into an equivalent task_group
2412 * representation for computing load contributions.
2414 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2415 struct cfs_rq *cfs_rq)
2417 struct task_group *tg = cfs_rq->tg;
2420 /* The fraction of a cpu used by this cfs_rq */
2421 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2422 sa->runnable_avg_period + 1);
2423 contrib -= cfs_rq->tg_runnable_contrib;
2425 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2426 atomic_add(contrib, &tg->runnable_avg);
2427 cfs_rq->tg_runnable_contrib += contrib;
2431 static inline void __update_group_entity_contrib(struct sched_entity *se)
2433 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2434 struct task_group *tg = cfs_rq->tg;
2439 contrib = cfs_rq->tg_load_contrib * tg->shares;
2440 se->avg.load_avg_contrib = div_u64(contrib,
2441 atomic_long_read(&tg->load_avg) + 1);
2444 * For group entities we need to compute a correction term in the case
2445 * that they are consuming <1 cpu so that we would contribute the same
2446 * load as a task of equal weight.
2448 * Explicitly co-ordinating this measurement would be expensive, but
2449 * fortunately the sum of each cpus contribution forms a usable
2450 * lower-bound on the true value.
2452 * Consider the aggregate of 2 contributions. Either they are disjoint
2453 * (and the sum represents true value) or they are disjoint and we are
2454 * understating by the aggregate of their overlap.
2456 * Extending this to N cpus, for a given overlap, the maximum amount we
2457 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2458 * cpus that overlap for this interval and w_i is the interval width.
2460 * On a small machine; the first term is well-bounded which bounds the
2461 * total error since w_i is a subset of the period. Whereas on a
2462 * larger machine, while this first term can be larger, if w_i is the
2463 * of consequential size guaranteed to see n_i*w_i quickly converge to
2464 * our upper bound of 1-cpu.
2466 runnable_avg = atomic_read(&tg->runnable_avg);
2467 if (runnable_avg < NICE_0_LOAD) {
2468 se->avg.load_avg_contrib *= runnable_avg;
2469 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2473 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2475 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2476 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2478 #else /* CONFIG_FAIR_GROUP_SCHED */
2479 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2480 int force_update) {}
2481 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2482 struct cfs_rq *cfs_rq) {}
2483 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2484 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2485 #endif /* CONFIG_FAIR_GROUP_SCHED */
2487 static inline void __update_task_entity_contrib(struct sched_entity *se)
2491 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2492 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2493 contrib /= (se->avg.runnable_avg_period + 1);
2494 se->avg.load_avg_contrib = scale_load(contrib);
2497 /* Compute the current contribution to load_avg by se, return any delta */
2498 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2500 long old_contrib = se->avg.load_avg_contrib;
2502 if (entity_is_task(se)) {
2503 __update_task_entity_contrib(se);
2505 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2506 __update_group_entity_contrib(se);
2509 return se->avg.load_avg_contrib - old_contrib;
2512 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2515 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2516 cfs_rq->blocked_load_avg -= load_contrib;
2518 cfs_rq->blocked_load_avg = 0;
2521 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2523 /* Update a sched_entity's runnable average */
2524 static inline void update_entity_load_avg(struct sched_entity *se,
2527 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2532 * For a group entity we need to use their owned cfs_rq_clock_task() in
2533 * case they are the parent of a throttled hierarchy.
2535 if (entity_is_task(se))
2536 now = cfs_rq_clock_task(cfs_rq);
2538 now = cfs_rq_clock_task(group_cfs_rq(se));
2540 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2543 contrib_delta = __update_entity_load_avg_contrib(se);
2549 cfs_rq->runnable_load_avg += contrib_delta;
2551 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2555 * Decay the load contributed by all blocked children and account this so that
2556 * their contribution may appropriately discounted when they wake up.
2558 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2560 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2563 decays = now - cfs_rq->last_decay;
2564 if (!decays && !force_update)
2567 if (atomic_long_read(&cfs_rq->removed_load)) {
2568 unsigned long removed_load;
2569 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2570 subtract_blocked_load_contrib(cfs_rq, removed_load);
2574 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2576 atomic64_add(decays, &cfs_rq->decay_counter);
2577 cfs_rq->last_decay = now;
2580 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2583 /* Add the load generated by se into cfs_rq's child load-average */
2584 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2585 struct sched_entity *se,
2589 * We track migrations using entity decay_count <= 0, on a wake-up
2590 * migration we use a negative decay count to track the remote decays
2591 * accumulated while sleeping.
2593 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2594 * are seen by enqueue_entity_load_avg() as a migration with an already
2595 * constructed load_avg_contrib.
2597 if (unlikely(se->avg.decay_count <= 0)) {
2598 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2599 if (se->avg.decay_count) {
2601 * In a wake-up migration we have to approximate the
2602 * time sleeping. This is because we can't synchronize
2603 * clock_task between the two cpus, and it is not
2604 * guaranteed to be read-safe. Instead, we can
2605 * approximate this using our carried decays, which are
2606 * explicitly atomically readable.
2608 se->avg.last_runnable_update -= (-se->avg.decay_count)
2610 update_entity_load_avg(se, 0);
2611 /* Indicate that we're now synchronized and on-rq */
2612 se->avg.decay_count = 0;
2616 __synchronize_entity_decay(se);
2619 /* migrated tasks did not contribute to our blocked load */
2621 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2622 update_entity_load_avg(se, 0);
2625 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2626 /* we force update consideration on load-balancer moves */
2627 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2631 * Remove se's load from this cfs_rq child load-average, if the entity is
2632 * transitioning to a blocked state we track its projected decay using
2635 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2636 struct sched_entity *se,
2639 update_entity_load_avg(se, 1);
2640 /* we force update consideration on load-balancer moves */
2641 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2643 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2645 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2646 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2647 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2651 * Update the rq's load with the elapsed running time before entering
2652 * idle. if the last scheduled task is not a CFS task, idle_enter will
2653 * be the only way to update the runnable statistic.
2655 void idle_enter_fair(struct rq *this_rq)
2657 update_rq_runnable_avg(this_rq, 1);
2661 * Update the rq's load with the elapsed idle time before a task is
2662 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2663 * be the only way to update the runnable statistic.
2665 void idle_exit_fair(struct rq *this_rq)
2667 update_rq_runnable_avg(this_rq, 0);
2670 static int idle_balance(struct rq *this_rq);
2672 #else /* CONFIG_SMP */
2674 static inline void update_entity_load_avg(struct sched_entity *se,
2675 int update_cfs_rq) {}
2676 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2677 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2678 struct sched_entity *se,
2680 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2681 struct sched_entity *se,
2683 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2684 int force_update) {}
2686 static inline int idle_balance(struct rq *rq)
2691 #endif /* CONFIG_SMP */
2693 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2695 #ifdef CONFIG_SCHEDSTATS
2696 struct task_struct *tsk = NULL;
2698 if (entity_is_task(se))
2701 if (se->statistics.sleep_start) {
2702 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2707 if (unlikely(delta > se->statistics.sleep_max))
2708 se->statistics.sleep_max = delta;
2710 se->statistics.sleep_start = 0;
2711 se->statistics.sum_sleep_runtime += delta;
2714 account_scheduler_latency(tsk, delta >> 10, 1);
2715 trace_sched_stat_sleep(tsk, delta);
2718 if (se->statistics.block_start) {
2719 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2724 if (unlikely(delta > se->statistics.block_max))
2725 se->statistics.block_max = delta;
2727 se->statistics.block_start = 0;
2728 se->statistics.sum_sleep_runtime += delta;
2731 if (tsk->in_iowait) {
2732 se->statistics.iowait_sum += delta;
2733 se->statistics.iowait_count++;
2734 trace_sched_stat_iowait(tsk, delta);
2737 trace_sched_stat_blocked(tsk, delta);
2740 * Blocking time is in units of nanosecs, so shift by
2741 * 20 to get a milliseconds-range estimation of the
2742 * amount of time that the task spent sleeping:
2744 if (unlikely(prof_on == SLEEP_PROFILING)) {
2745 profile_hits(SLEEP_PROFILING,
2746 (void *)get_wchan(tsk),
2749 account_scheduler_latency(tsk, delta >> 10, 0);
2755 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2757 #ifdef CONFIG_SCHED_DEBUG
2758 s64 d = se->vruntime - cfs_rq->min_vruntime;
2763 if (d > 3*sysctl_sched_latency)
2764 schedstat_inc(cfs_rq, nr_spread_over);
2769 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2771 u64 vruntime = cfs_rq->min_vruntime;
2774 * The 'current' period is already promised to the current tasks,
2775 * however the extra weight of the new task will slow them down a
2776 * little, place the new task so that it fits in the slot that
2777 * stays open at the end.
2779 if (initial && sched_feat(START_DEBIT))
2780 vruntime += sched_vslice(cfs_rq, se);
2782 /* sleeps up to a single latency don't count. */
2784 unsigned long thresh = sysctl_sched_latency;
2787 * Halve their sleep time's effect, to allow
2788 * for a gentler effect of sleepers:
2790 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2796 /* ensure we never gain time by being placed backwards. */
2797 se->vruntime = max_vruntime(se->vruntime, vruntime);
2800 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2803 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2806 * Update the normalized vruntime before updating min_vruntime
2807 * through calling update_curr().
2809 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2810 se->vruntime += cfs_rq->min_vruntime;
2813 * Update run-time statistics of the 'current'.
2815 update_curr(cfs_rq);
2816 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2817 account_entity_enqueue(cfs_rq, se);
2818 update_cfs_shares(cfs_rq);
2820 if (flags & ENQUEUE_WAKEUP) {
2821 place_entity(cfs_rq, se, 0);
2822 enqueue_sleeper(cfs_rq, se);
2825 update_stats_enqueue(cfs_rq, se);
2826 check_spread(cfs_rq, se);
2827 if (se != cfs_rq->curr)
2828 __enqueue_entity(cfs_rq, se);
2831 if (cfs_rq->nr_running == 1) {
2832 list_add_leaf_cfs_rq(cfs_rq);
2833 check_enqueue_throttle(cfs_rq);
2837 static void __clear_buddies_last(struct sched_entity *se)
2839 for_each_sched_entity(se) {
2840 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2841 if (cfs_rq->last != se)
2844 cfs_rq->last = NULL;
2848 static void __clear_buddies_next(struct sched_entity *se)
2850 for_each_sched_entity(se) {
2851 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2852 if (cfs_rq->next != se)
2855 cfs_rq->next = NULL;
2859 static void __clear_buddies_skip(struct sched_entity *se)
2861 for_each_sched_entity(se) {
2862 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2863 if (cfs_rq->skip != se)
2866 cfs_rq->skip = NULL;
2870 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2872 if (cfs_rq->last == se)
2873 __clear_buddies_last(se);
2875 if (cfs_rq->next == se)
2876 __clear_buddies_next(se);
2878 if (cfs_rq->skip == se)
2879 __clear_buddies_skip(se);
2882 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2885 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2888 * Update run-time statistics of the 'current'.
2890 update_curr(cfs_rq);
2891 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2893 update_stats_dequeue(cfs_rq, se);
2894 if (flags & DEQUEUE_SLEEP) {
2895 #ifdef CONFIG_SCHEDSTATS
2896 if (entity_is_task(se)) {
2897 struct task_struct *tsk = task_of(se);
2899 if (tsk->state & TASK_INTERRUPTIBLE)
2900 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2901 if (tsk->state & TASK_UNINTERRUPTIBLE)
2902 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2907 clear_buddies(cfs_rq, se);
2909 if (se != cfs_rq->curr)
2910 __dequeue_entity(cfs_rq, se);
2912 account_entity_dequeue(cfs_rq, se);
2915 * Normalize the entity after updating the min_vruntime because the
2916 * update can refer to the ->curr item and we need to reflect this
2917 * movement in our normalized position.
2919 if (!(flags & DEQUEUE_SLEEP))
2920 se->vruntime -= cfs_rq->min_vruntime;
2922 /* return excess runtime on last dequeue */
2923 return_cfs_rq_runtime(cfs_rq);
2925 update_min_vruntime(cfs_rq);
2926 update_cfs_shares(cfs_rq);
2930 * Preempt the current task with a newly woken task if needed:
2933 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2935 unsigned long ideal_runtime, delta_exec;
2936 struct sched_entity *se;
2939 ideal_runtime = sched_slice(cfs_rq, curr);
2940 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2941 if (delta_exec > ideal_runtime) {
2942 resched_curr(rq_of(cfs_rq));
2944 * The current task ran long enough, ensure it doesn't get
2945 * re-elected due to buddy favours.
2947 clear_buddies(cfs_rq, curr);
2952 * Ensure that a task that missed wakeup preemption by a
2953 * narrow margin doesn't have to wait for a full slice.
2954 * This also mitigates buddy induced latencies under load.
2956 if (delta_exec < sysctl_sched_min_granularity)
2959 se = __pick_first_entity(cfs_rq);
2960 delta = curr->vruntime - se->vruntime;
2965 if (delta > ideal_runtime)
2966 resched_curr(rq_of(cfs_rq));
2970 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2972 /* 'current' is not kept within the tree. */
2975 * Any task has to be enqueued before it get to execute on
2976 * a CPU. So account for the time it spent waiting on the
2979 update_stats_wait_end(cfs_rq, se);
2980 __dequeue_entity(cfs_rq, se);
2983 update_stats_curr_start(cfs_rq, se);
2985 #ifdef CONFIG_SCHEDSTATS
2987 * Track our maximum slice length, if the CPU's load is at
2988 * least twice that of our own weight (i.e. dont track it
2989 * when there are only lesser-weight tasks around):
2991 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2992 se->statistics.slice_max = max(se->statistics.slice_max,
2993 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2996 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3000 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3003 * Pick the next process, keeping these things in mind, in this order:
3004 * 1) keep things fair between processes/task groups
3005 * 2) pick the "next" process, since someone really wants that to run
3006 * 3) pick the "last" process, for cache locality
3007 * 4) do not run the "skip" process, if something else is available
3009 static struct sched_entity *
3010 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3012 struct sched_entity *left = __pick_first_entity(cfs_rq);
3013 struct sched_entity *se;
3016 * If curr is set we have to see if its left of the leftmost entity
3017 * still in the tree, provided there was anything in the tree at all.
3019 if (!left || (curr && entity_before(curr, left)))
3022 se = left; /* ideally we run the leftmost entity */
3025 * Avoid running the skip buddy, if running something else can
3026 * be done without getting too unfair.
3028 if (cfs_rq->skip == se) {
3029 struct sched_entity *second;
3032 second = __pick_first_entity(cfs_rq);
3034 second = __pick_next_entity(se);
3035 if (!second || (curr && entity_before(curr, second)))
3039 if (second && wakeup_preempt_entity(second, left) < 1)
3044 * Prefer last buddy, try to return the CPU to a preempted task.
3046 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3050 * Someone really wants this to run. If it's not unfair, run it.
3052 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3055 clear_buddies(cfs_rq, se);
3060 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3062 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3065 * If still on the runqueue then deactivate_task()
3066 * was not called and update_curr() has to be done:
3069 update_curr(cfs_rq);
3071 /* throttle cfs_rqs exceeding runtime */
3072 check_cfs_rq_runtime(cfs_rq);
3074 check_spread(cfs_rq, prev);
3076 update_stats_wait_start(cfs_rq, prev);
3077 /* Put 'current' back into the tree. */
3078 __enqueue_entity(cfs_rq, prev);
3079 /* in !on_rq case, update occurred at dequeue */
3080 update_entity_load_avg(prev, 1);
3082 cfs_rq->curr = NULL;
3086 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3089 * Update run-time statistics of the 'current'.
3091 update_curr(cfs_rq);
3094 * Ensure that runnable average is periodically updated.
3096 update_entity_load_avg(curr, 1);
3097 update_cfs_rq_blocked_load(cfs_rq, 1);
3098 update_cfs_shares(cfs_rq);
3100 #ifdef CONFIG_SCHED_HRTICK
3102 * queued ticks are scheduled to match the slice, so don't bother
3103 * validating it and just reschedule.
3106 resched_curr(rq_of(cfs_rq));
3110 * don't let the period tick interfere with the hrtick preemption
3112 if (!sched_feat(DOUBLE_TICK) &&
3113 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3117 if (cfs_rq->nr_running > 1)
3118 check_preempt_tick(cfs_rq, curr);
3122 /**************************************************
3123 * CFS bandwidth control machinery
3126 #ifdef CONFIG_CFS_BANDWIDTH
3128 #ifdef HAVE_JUMP_LABEL
3129 static struct static_key __cfs_bandwidth_used;
3131 static inline bool cfs_bandwidth_used(void)
3133 return static_key_false(&__cfs_bandwidth_used);
3136 void cfs_bandwidth_usage_inc(void)
3138 static_key_slow_inc(&__cfs_bandwidth_used);
3141 void cfs_bandwidth_usage_dec(void)
3143 static_key_slow_dec(&__cfs_bandwidth_used);
3145 #else /* HAVE_JUMP_LABEL */
3146 static bool cfs_bandwidth_used(void)
3151 void cfs_bandwidth_usage_inc(void) {}
3152 void cfs_bandwidth_usage_dec(void) {}
3153 #endif /* HAVE_JUMP_LABEL */
3156 * default period for cfs group bandwidth.
3157 * default: 0.1s, units: nanoseconds
3159 static inline u64 default_cfs_period(void)
3161 return 100000000ULL;
3164 static inline u64 sched_cfs_bandwidth_slice(void)
3166 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3170 * Replenish runtime according to assigned quota and update expiration time.
3171 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3172 * additional synchronization around rq->lock.
3174 * requires cfs_b->lock
3176 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3180 if (cfs_b->quota == RUNTIME_INF)
3183 now = sched_clock_cpu(smp_processor_id());
3184 cfs_b->runtime = cfs_b->quota;
3185 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3188 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3190 return &tg->cfs_bandwidth;
3193 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3194 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3196 if (unlikely(cfs_rq->throttle_count))
3197 return cfs_rq->throttled_clock_task;
3199 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3202 /* returns 0 on failure to allocate runtime */
3203 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3205 struct task_group *tg = cfs_rq->tg;
3206 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3207 u64 amount = 0, min_amount, expires;
3209 /* note: this is a positive sum as runtime_remaining <= 0 */
3210 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3212 raw_spin_lock(&cfs_b->lock);
3213 if (cfs_b->quota == RUNTIME_INF)
3214 amount = min_amount;
3217 * If the bandwidth pool has become inactive, then at least one
3218 * period must have elapsed since the last consumption.
3219 * Refresh the global state and ensure bandwidth timer becomes
3222 if (!cfs_b->timer_active) {
3223 __refill_cfs_bandwidth_runtime(cfs_b);
3224 __start_cfs_bandwidth(cfs_b, false);
3227 if (cfs_b->runtime > 0) {
3228 amount = min(cfs_b->runtime, min_amount);
3229 cfs_b->runtime -= amount;
3233 expires = cfs_b->runtime_expires;
3234 raw_spin_unlock(&cfs_b->lock);
3236 cfs_rq->runtime_remaining += amount;
3238 * we may have advanced our local expiration to account for allowed
3239 * spread between our sched_clock and the one on which runtime was
3242 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3243 cfs_rq->runtime_expires = expires;
3245 return cfs_rq->runtime_remaining > 0;
3249 * Note: This depends on the synchronization provided by sched_clock and the
3250 * fact that rq->clock snapshots this value.
3252 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3254 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3256 /* if the deadline is ahead of our clock, nothing to do */
3257 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3260 if (cfs_rq->runtime_remaining < 0)
3264 * If the local deadline has passed we have to consider the
3265 * possibility that our sched_clock is 'fast' and the global deadline
3266 * has not truly expired.
3268 * Fortunately we can check determine whether this the case by checking
3269 * whether the global deadline has advanced. It is valid to compare
3270 * cfs_b->runtime_expires without any locks since we only care about
3271 * exact equality, so a partial write will still work.
3274 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3275 /* extend local deadline, drift is bounded above by 2 ticks */
3276 cfs_rq->runtime_expires += TICK_NSEC;
3278 /* global deadline is ahead, expiration has passed */
3279 cfs_rq->runtime_remaining = 0;
3283 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3285 /* dock delta_exec before expiring quota (as it could span periods) */
3286 cfs_rq->runtime_remaining -= delta_exec;
3287 expire_cfs_rq_runtime(cfs_rq);
3289 if (likely(cfs_rq->runtime_remaining > 0))
3293 * if we're unable to extend our runtime we resched so that the active
3294 * hierarchy can be throttled
3296 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3297 resched_curr(rq_of(cfs_rq));
3300 static __always_inline
3301 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3303 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3306 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3309 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3311 return cfs_bandwidth_used() && cfs_rq->throttled;
3314 /* check whether cfs_rq, or any parent, is throttled */
3315 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3317 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3321 * Ensure that neither of the group entities corresponding to src_cpu or
3322 * dest_cpu are members of a throttled hierarchy when performing group
3323 * load-balance operations.
3325 static inline int throttled_lb_pair(struct task_group *tg,
3326 int src_cpu, int dest_cpu)
3328 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3330 src_cfs_rq = tg->cfs_rq[src_cpu];
3331 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3333 return throttled_hierarchy(src_cfs_rq) ||
3334 throttled_hierarchy(dest_cfs_rq);
3337 /* updated child weight may affect parent so we have to do this bottom up */
3338 static int tg_unthrottle_up(struct task_group *tg, void *data)
3340 struct rq *rq = data;
3341 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3343 cfs_rq->throttle_count--;
3345 if (!cfs_rq->throttle_count) {
3346 /* adjust cfs_rq_clock_task() */
3347 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3348 cfs_rq->throttled_clock_task;
3355 static int tg_throttle_down(struct task_group *tg, void *data)
3357 struct rq *rq = data;
3358 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3360 /* group is entering throttled state, stop time */
3361 if (!cfs_rq->throttle_count)
3362 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3363 cfs_rq->throttle_count++;
3368 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3370 struct rq *rq = rq_of(cfs_rq);
3371 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3372 struct sched_entity *se;
3373 long task_delta, dequeue = 1;
3375 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3377 /* freeze hierarchy runnable averages while throttled */
3379 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3382 task_delta = cfs_rq->h_nr_running;
3383 for_each_sched_entity(se) {
3384 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3385 /* throttled entity or throttle-on-deactivate */
3390 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3391 qcfs_rq->h_nr_running -= task_delta;
3393 if (qcfs_rq->load.weight)
3398 sub_nr_running(rq, task_delta);
3400 cfs_rq->throttled = 1;
3401 cfs_rq->throttled_clock = rq_clock(rq);
3402 raw_spin_lock(&cfs_b->lock);
3404 * Add to the _head_ of the list, so that an already-started
3405 * distribute_cfs_runtime will not see us
3407 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3408 if (!cfs_b->timer_active)
3409 __start_cfs_bandwidth(cfs_b, false);
3410 raw_spin_unlock(&cfs_b->lock);
3413 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3415 struct rq *rq = rq_of(cfs_rq);
3416 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3417 struct sched_entity *se;
3421 se = cfs_rq->tg->se[cpu_of(rq)];
3423 cfs_rq->throttled = 0;
3425 update_rq_clock(rq);
3427 raw_spin_lock(&cfs_b->lock);
3428 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3429 list_del_rcu(&cfs_rq->throttled_list);
3430 raw_spin_unlock(&cfs_b->lock);
3432 /* update hierarchical throttle state */
3433 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3435 if (!cfs_rq->load.weight)
3438 task_delta = cfs_rq->h_nr_running;
3439 for_each_sched_entity(se) {
3443 cfs_rq = cfs_rq_of(se);
3445 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3446 cfs_rq->h_nr_running += task_delta;
3448 if (cfs_rq_throttled(cfs_rq))
3453 add_nr_running(rq, task_delta);
3455 /* determine whether we need to wake up potentially idle cpu */
3456 if (rq->curr == rq->idle && rq->cfs.nr_running)
3460 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3461 u64 remaining, u64 expires)
3463 struct cfs_rq *cfs_rq;
3465 u64 starting_runtime = remaining;
3468 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3470 struct rq *rq = rq_of(cfs_rq);
3472 raw_spin_lock(&rq->lock);
3473 if (!cfs_rq_throttled(cfs_rq))
3476 runtime = -cfs_rq->runtime_remaining + 1;
3477 if (runtime > remaining)
3478 runtime = remaining;
3479 remaining -= runtime;
3481 cfs_rq->runtime_remaining += runtime;
3482 cfs_rq->runtime_expires = expires;
3484 /* we check whether we're throttled above */
3485 if (cfs_rq->runtime_remaining > 0)
3486 unthrottle_cfs_rq(cfs_rq);
3489 raw_spin_unlock(&rq->lock);
3496 return starting_runtime - remaining;
3500 * Responsible for refilling a task_group's bandwidth and unthrottling its
3501 * cfs_rqs as appropriate. If there has been no activity within the last
3502 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3503 * used to track this state.
3505 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3507 u64 runtime, runtime_expires;
3510 /* no need to continue the timer with no bandwidth constraint */
3511 if (cfs_b->quota == RUNTIME_INF)
3512 goto out_deactivate;
3514 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3515 cfs_b->nr_periods += overrun;
3518 * idle depends on !throttled (for the case of a large deficit), and if
3519 * we're going inactive then everything else can be deferred
3521 if (cfs_b->idle && !throttled)
3522 goto out_deactivate;
3525 * if we have relooped after returning idle once, we need to update our
3526 * status as actually running, so that other cpus doing
3527 * __start_cfs_bandwidth will stop trying to cancel us.
3529 cfs_b->timer_active = 1;
3531 __refill_cfs_bandwidth_runtime(cfs_b);
3534 /* mark as potentially idle for the upcoming period */
3539 /* account preceding periods in which throttling occurred */
3540 cfs_b->nr_throttled += overrun;
3542 runtime_expires = cfs_b->runtime_expires;
3545 * This check is repeated as we are holding onto the new bandwidth while
3546 * we unthrottle. This can potentially race with an unthrottled group
3547 * trying to acquire new bandwidth from the global pool. This can result
3548 * in us over-using our runtime if it is all used during this loop, but
3549 * only by limited amounts in that extreme case.
3551 while (throttled && cfs_b->runtime > 0) {
3552 runtime = cfs_b->runtime;
3553 raw_spin_unlock(&cfs_b->lock);
3554 /* we can't nest cfs_b->lock while distributing bandwidth */
3555 runtime = distribute_cfs_runtime(cfs_b, runtime,
3557 raw_spin_lock(&cfs_b->lock);
3559 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3561 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3565 * While we are ensured activity in the period following an
3566 * unthrottle, this also covers the case in which the new bandwidth is
3567 * insufficient to cover the existing bandwidth deficit. (Forcing the
3568 * timer to remain active while there are any throttled entities.)
3575 cfs_b->timer_active = 0;
3579 /* a cfs_rq won't donate quota below this amount */
3580 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3581 /* minimum remaining period time to redistribute slack quota */
3582 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3583 /* how long we wait to gather additional slack before distributing */
3584 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3587 * Are we near the end of the current quota period?
3589 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3590 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3591 * migrate_hrtimers, base is never cleared, so we are fine.
3593 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3595 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3598 /* if the call-back is running a quota refresh is already occurring */
3599 if (hrtimer_callback_running(refresh_timer))
3602 /* is a quota refresh about to occur? */
3603 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3604 if (remaining < min_expire)
3610 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3612 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3614 /* if there's a quota refresh soon don't bother with slack */
3615 if (runtime_refresh_within(cfs_b, min_left))
3618 start_bandwidth_timer(&cfs_b->slack_timer,
3619 ns_to_ktime(cfs_bandwidth_slack_period));
3622 /* we know any runtime found here is valid as update_curr() precedes return */
3623 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3625 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3626 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3628 if (slack_runtime <= 0)
3631 raw_spin_lock(&cfs_b->lock);
3632 if (cfs_b->quota != RUNTIME_INF &&
3633 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3634 cfs_b->runtime += slack_runtime;
3636 /* we are under rq->lock, defer unthrottling using a timer */
3637 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3638 !list_empty(&cfs_b->throttled_cfs_rq))
3639 start_cfs_slack_bandwidth(cfs_b);
3641 raw_spin_unlock(&cfs_b->lock);
3643 /* even if it's not valid for return we don't want to try again */
3644 cfs_rq->runtime_remaining -= slack_runtime;
3647 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3649 if (!cfs_bandwidth_used())
3652 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3655 __return_cfs_rq_runtime(cfs_rq);
3659 * This is done with a timer (instead of inline with bandwidth return) since
3660 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3662 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3664 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3667 /* confirm we're still not at a refresh boundary */
3668 raw_spin_lock(&cfs_b->lock);
3669 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3670 raw_spin_unlock(&cfs_b->lock);
3674 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3675 runtime = cfs_b->runtime;
3677 expires = cfs_b->runtime_expires;
3678 raw_spin_unlock(&cfs_b->lock);
3683 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3685 raw_spin_lock(&cfs_b->lock);
3686 if (expires == cfs_b->runtime_expires)
3687 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3688 raw_spin_unlock(&cfs_b->lock);
3692 * When a group wakes up we want to make sure that its quota is not already
3693 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3694 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3696 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3698 if (!cfs_bandwidth_used())
3701 /* an active group must be handled by the update_curr()->put() path */
3702 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3705 /* ensure the group is not already throttled */
3706 if (cfs_rq_throttled(cfs_rq))
3709 /* update runtime allocation */
3710 account_cfs_rq_runtime(cfs_rq, 0);
3711 if (cfs_rq->runtime_remaining <= 0)
3712 throttle_cfs_rq(cfs_rq);
3715 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3716 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3718 if (!cfs_bandwidth_used())
3721 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3725 * it's possible for a throttled entity to be forced into a running
3726 * state (e.g. set_curr_task), in this case we're finished.
3728 if (cfs_rq_throttled(cfs_rq))
3731 throttle_cfs_rq(cfs_rq);
3735 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3737 struct cfs_bandwidth *cfs_b =
3738 container_of(timer, struct cfs_bandwidth, slack_timer);
3739 do_sched_cfs_slack_timer(cfs_b);
3741 return HRTIMER_NORESTART;
3744 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3746 struct cfs_bandwidth *cfs_b =
3747 container_of(timer, struct cfs_bandwidth, period_timer);
3752 raw_spin_lock(&cfs_b->lock);
3754 now = hrtimer_cb_get_time(timer);
3755 overrun = hrtimer_forward(timer, now, cfs_b->period);
3760 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3762 raw_spin_unlock(&cfs_b->lock);
3764 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3767 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3769 raw_spin_lock_init(&cfs_b->lock);
3771 cfs_b->quota = RUNTIME_INF;
3772 cfs_b->period = ns_to_ktime(default_cfs_period());
3774 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3775 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3776 cfs_b->period_timer.function = sched_cfs_period_timer;
3777 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3778 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3781 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3783 cfs_rq->runtime_enabled = 0;
3784 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3787 /* requires cfs_b->lock, may release to reprogram timer */
3788 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force)
3791 * The timer may be active because we're trying to set a new bandwidth
3792 * period or because we're racing with the tear-down path
3793 * (timer_active==0 becomes visible before the hrtimer call-back
3794 * terminates). In either case we ensure that it's re-programmed
3796 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3797 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3798 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3799 raw_spin_unlock(&cfs_b->lock);
3801 raw_spin_lock(&cfs_b->lock);
3802 /* if someone else restarted the timer then we're done */
3803 if (!force && cfs_b->timer_active)
3807 cfs_b->timer_active = 1;
3808 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3811 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3813 hrtimer_cancel(&cfs_b->period_timer);
3814 hrtimer_cancel(&cfs_b->slack_timer);
3817 static void __maybe_unused update_runtime_enabled(struct rq *rq)
3819 struct cfs_rq *cfs_rq;
3821 for_each_leaf_cfs_rq(rq, cfs_rq) {
3822 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
3824 raw_spin_lock(&cfs_b->lock);
3825 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
3826 raw_spin_unlock(&cfs_b->lock);
3830 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3832 struct cfs_rq *cfs_rq;
3834 for_each_leaf_cfs_rq(rq, cfs_rq) {
3835 if (!cfs_rq->runtime_enabled)
3839 * clock_task is not advancing so we just need to make sure
3840 * there's some valid quota amount
3842 cfs_rq->runtime_remaining = 1;
3844 * Offline rq is schedulable till cpu is completely disabled
3845 * in take_cpu_down(), so we prevent new cfs throttling here.
3847 cfs_rq->runtime_enabled = 0;
3849 if (cfs_rq_throttled(cfs_rq))
3850 unthrottle_cfs_rq(cfs_rq);
3854 #else /* CONFIG_CFS_BANDWIDTH */
3855 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3857 return rq_clock_task(rq_of(cfs_rq));
3860 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3861 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3862 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3863 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3865 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3870 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3875 static inline int throttled_lb_pair(struct task_group *tg,
3876 int src_cpu, int dest_cpu)
3881 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3883 #ifdef CONFIG_FAIR_GROUP_SCHED
3884 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3887 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3891 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3892 static inline void update_runtime_enabled(struct rq *rq) {}
3893 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3895 #endif /* CONFIG_CFS_BANDWIDTH */
3897 /**************************************************
3898 * CFS operations on tasks:
3901 #ifdef CONFIG_SCHED_HRTICK
3902 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3904 struct sched_entity *se = &p->se;
3905 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3907 WARN_ON(task_rq(p) != rq);
3909 if (cfs_rq->nr_running > 1) {
3910 u64 slice = sched_slice(cfs_rq, se);
3911 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3912 s64 delta = slice - ran;
3919 hrtick_start(rq, delta);
3924 * called from enqueue/dequeue and updates the hrtick when the
3925 * current task is from our class and nr_running is low enough
3928 static void hrtick_update(struct rq *rq)
3930 struct task_struct *curr = rq->curr;
3932 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3935 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3936 hrtick_start_fair(rq, curr);
3938 #else /* !CONFIG_SCHED_HRTICK */
3940 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3944 static inline void hrtick_update(struct rq *rq)
3950 * The enqueue_task method is called before nr_running is
3951 * increased. Here we update the fair scheduling stats and
3952 * then put the task into the rbtree:
3955 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3957 struct cfs_rq *cfs_rq;
3958 struct sched_entity *se = &p->se;
3960 for_each_sched_entity(se) {
3963 cfs_rq = cfs_rq_of(se);
3964 enqueue_entity(cfs_rq, se, flags);
3967 * end evaluation on encountering a throttled cfs_rq
3969 * note: in the case of encountering a throttled cfs_rq we will
3970 * post the final h_nr_running increment below.
3972 if (cfs_rq_throttled(cfs_rq))
3974 cfs_rq->h_nr_running++;
3976 flags = ENQUEUE_WAKEUP;
3979 for_each_sched_entity(se) {
3980 cfs_rq = cfs_rq_of(se);
3981 cfs_rq->h_nr_running++;
3983 if (cfs_rq_throttled(cfs_rq))
3986 update_cfs_shares(cfs_rq);
3987 update_entity_load_avg(se, 1);
3991 update_rq_runnable_avg(rq, rq->nr_running);
3992 add_nr_running(rq, 1);
3997 static void set_next_buddy(struct sched_entity *se);
4000 * The dequeue_task method is called before nr_running is
4001 * decreased. We remove the task from the rbtree and
4002 * update the fair scheduling stats:
4004 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4006 struct cfs_rq *cfs_rq;
4007 struct sched_entity *se = &p->se;
4008 int task_sleep = flags & DEQUEUE_SLEEP;
4010 for_each_sched_entity(se) {
4011 cfs_rq = cfs_rq_of(se);
4012 dequeue_entity(cfs_rq, se, flags);
4015 * end evaluation on encountering a throttled cfs_rq
4017 * note: in the case of encountering a throttled cfs_rq we will
4018 * post the final h_nr_running decrement below.
4020 if (cfs_rq_throttled(cfs_rq))
4022 cfs_rq->h_nr_running--;
4024 /* Don't dequeue parent if it has other entities besides us */
4025 if (cfs_rq->load.weight) {
4027 * Bias pick_next to pick a task from this cfs_rq, as
4028 * p is sleeping when it is within its sched_slice.
4030 if (task_sleep && parent_entity(se))
4031 set_next_buddy(parent_entity(se));
4033 /* avoid re-evaluating load for this entity */
4034 se = parent_entity(se);
4037 flags |= DEQUEUE_SLEEP;
4040 for_each_sched_entity(se) {
4041 cfs_rq = cfs_rq_of(se);
4042 cfs_rq->h_nr_running--;
4044 if (cfs_rq_throttled(cfs_rq))
4047 update_cfs_shares(cfs_rq);
4048 update_entity_load_avg(se, 1);
4052 sub_nr_running(rq, 1);
4053 update_rq_runnable_avg(rq, 1);
4059 /* Used instead of source_load when we know the type == 0 */
4060 static unsigned long weighted_cpuload(const int cpu)
4062 return cpu_rq(cpu)->cfs.runnable_load_avg;
4066 * Return a low guess at the load of a migration-source cpu weighted
4067 * according to the scheduling class and "nice" value.
4069 * We want to under-estimate the load of migration sources, to
4070 * balance conservatively.
4072 static unsigned long source_load(int cpu, int type)
4074 struct rq *rq = cpu_rq(cpu);
4075 unsigned long total = weighted_cpuload(cpu);
4077 if (type == 0 || !sched_feat(LB_BIAS))
4080 return min(rq->cpu_load[type-1], total);
4084 * Return a high guess at the load of a migration-target cpu weighted
4085 * according to the scheduling class and "nice" value.
4087 static unsigned long target_load(int cpu, int type)
4089 struct rq *rq = cpu_rq(cpu);
4090 unsigned long total = weighted_cpuload(cpu);
4092 if (type == 0 || !sched_feat(LB_BIAS))
4095 return max(rq->cpu_load[type-1], total);
4098 static unsigned long capacity_of(int cpu)
4100 return cpu_rq(cpu)->cpu_capacity;
4103 static unsigned long cpu_avg_load_per_task(int cpu)
4105 struct rq *rq = cpu_rq(cpu);
4106 unsigned long nr_running = ACCESS_ONCE(rq->cfs.h_nr_running);
4107 unsigned long load_avg = rq->cfs.runnable_load_avg;
4110 return load_avg / nr_running;
4115 static void record_wakee(struct task_struct *p)
4118 * Rough decay (wiping) for cost saving, don't worry
4119 * about the boundary, really active task won't care
4122 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4123 current->wakee_flips >>= 1;
4124 current->wakee_flip_decay_ts = jiffies;
4127 if (current->last_wakee != p) {
4128 current->last_wakee = p;
4129 current->wakee_flips++;
4133 static void task_waking_fair(struct task_struct *p)
4135 struct sched_entity *se = &p->se;
4136 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4139 #ifndef CONFIG_64BIT
4140 u64 min_vruntime_copy;
4143 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4145 min_vruntime = cfs_rq->min_vruntime;
4146 } while (min_vruntime != min_vruntime_copy);
4148 min_vruntime = cfs_rq->min_vruntime;
4151 se->vruntime -= min_vruntime;
4155 #ifdef CONFIG_FAIR_GROUP_SCHED
4157 * effective_load() calculates the load change as seen from the root_task_group
4159 * Adding load to a group doesn't make a group heavier, but can cause movement
4160 * of group shares between cpus. Assuming the shares were perfectly aligned one
4161 * can calculate the shift in shares.
4163 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4164 * on this @cpu and results in a total addition (subtraction) of @wg to the
4165 * total group weight.
4167 * Given a runqueue weight distribution (rw_i) we can compute a shares
4168 * distribution (s_i) using:
4170 * s_i = rw_i / \Sum rw_j (1)
4172 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4173 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4174 * shares distribution (s_i):
4176 * rw_i = { 2, 4, 1, 0 }
4177 * s_i = { 2/7, 4/7, 1/7, 0 }
4179 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4180 * task used to run on and the CPU the waker is running on), we need to
4181 * compute the effect of waking a task on either CPU and, in case of a sync
4182 * wakeup, compute the effect of the current task going to sleep.
4184 * So for a change of @wl to the local @cpu with an overall group weight change
4185 * of @wl we can compute the new shares distribution (s'_i) using:
4187 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4189 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4190 * differences in waking a task to CPU 0. The additional task changes the
4191 * weight and shares distributions like:
4193 * rw'_i = { 3, 4, 1, 0 }
4194 * s'_i = { 3/8, 4/8, 1/8, 0 }
4196 * We can then compute the difference in effective weight by using:
4198 * dw_i = S * (s'_i - s_i) (3)
4200 * Where 'S' is the group weight as seen by its parent.
4202 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4203 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4204 * 4/7) times the weight of the group.
4206 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4208 struct sched_entity *se = tg->se[cpu];
4210 if (!tg->parent) /* the trivial, non-cgroup case */
4213 for_each_sched_entity(se) {
4219 * W = @wg + \Sum rw_j
4221 W = wg + calc_tg_weight(tg, se->my_q);
4226 w = se->my_q->load.weight + wl;
4229 * wl = S * s'_i; see (2)
4232 wl = (w * tg->shares) / W;
4237 * Per the above, wl is the new se->load.weight value; since
4238 * those are clipped to [MIN_SHARES, ...) do so now. See
4239 * calc_cfs_shares().
4241 if (wl < MIN_SHARES)
4245 * wl = dw_i = S * (s'_i - s_i); see (3)
4247 wl -= se->load.weight;
4250 * Recursively apply this logic to all parent groups to compute
4251 * the final effective load change on the root group. Since
4252 * only the @tg group gets extra weight, all parent groups can
4253 * only redistribute existing shares. @wl is the shift in shares
4254 * resulting from this level per the above.
4263 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4270 static int wake_wide(struct task_struct *p)
4272 int factor = this_cpu_read(sd_llc_size);
4275 * Yeah, it's the switching-frequency, could means many wakee or
4276 * rapidly switch, use factor here will just help to automatically
4277 * adjust the loose-degree, so bigger node will lead to more pull.
4279 if (p->wakee_flips > factor) {
4281 * wakee is somewhat hot, it needs certain amount of cpu
4282 * resource, so if waker is far more hot, prefer to leave
4285 if (current->wakee_flips > (factor * p->wakee_flips))
4292 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4294 s64 this_load, load;
4295 s64 this_eff_load, prev_eff_load;
4296 int idx, this_cpu, prev_cpu;
4297 struct task_group *tg;
4298 unsigned long weight;
4302 * If we wake multiple tasks be careful to not bounce
4303 * ourselves around too much.
4309 this_cpu = smp_processor_id();
4310 prev_cpu = task_cpu(p);
4311 load = source_load(prev_cpu, idx);
4312 this_load = target_load(this_cpu, idx);
4315 * If sync wakeup then subtract the (maximum possible)
4316 * effect of the currently running task from the load
4317 * of the current CPU:
4320 tg = task_group(current);
4321 weight = current->se.load.weight;
4323 this_load += effective_load(tg, this_cpu, -weight, -weight);
4324 load += effective_load(tg, prev_cpu, 0, -weight);
4328 weight = p->se.load.weight;
4331 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4332 * due to the sync cause above having dropped this_load to 0, we'll
4333 * always have an imbalance, but there's really nothing you can do
4334 * about that, so that's good too.
4336 * Otherwise check if either cpus are near enough in load to allow this
4337 * task to be woken on this_cpu.
4339 this_eff_load = 100;
4340 this_eff_load *= capacity_of(prev_cpu);
4342 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4343 prev_eff_load *= capacity_of(this_cpu);
4345 if (this_load > 0) {
4346 this_eff_load *= this_load +
4347 effective_load(tg, this_cpu, weight, weight);
4349 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4352 balanced = this_eff_load <= prev_eff_load;
4354 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4359 schedstat_inc(sd, ttwu_move_affine);
4360 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4366 * find_idlest_group finds and returns the least busy CPU group within the
4369 static struct sched_group *
4370 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4371 int this_cpu, int sd_flag)
4373 struct sched_group *idlest = NULL, *group = sd->groups;
4374 unsigned long min_load = ULONG_MAX, this_load = 0;
4375 int load_idx = sd->forkexec_idx;
4376 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4378 if (sd_flag & SD_BALANCE_WAKE)
4379 load_idx = sd->wake_idx;
4382 unsigned long load, avg_load;
4386 /* Skip over this group if it has no CPUs allowed */
4387 if (!cpumask_intersects(sched_group_cpus(group),
4388 tsk_cpus_allowed(p)))
4391 local_group = cpumask_test_cpu(this_cpu,
4392 sched_group_cpus(group));
4394 /* Tally up the load of all CPUs in the group */
4397 for_each_cpu(i, sched_group_cpus(group)) {
4398 /* Bias balancing toward cpus of our domain */
4400 load = source_load(i, load_idx);
4402 load = target_load(i, load_idx);
4407 /* Adjust by relative CPU capacity of the group */
4408 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4411 this_load = avg_load;
4412 } else if (avg_load < min_load) {
4413 min_load = avg_load;
4416 } while (group = group->next, group != sd->groups);
4418 if (!idlest || 100*this_load < imbalance*min_load)
4424 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4427 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4429 unsigned long load, min_load = ULONG_MAX;
4430 unsigned int min_exit_latency = UINT_MAX;
4431 u64 latest_idle_timestamp = 0;
4432 int least_loaded_cpu = this_cpu;
4433 int shallowest_idle_cpu = -1;
4436 /* Traverse only the allowed CPUs */
4437 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4439 struct rq *rq = cpu_rq(i);
4440 struct cpuidle_state *idle = idle_get_state(rq);
4441 if (idle && idle->exit_latency < min_exit_latency) {
4443 * We give priority to a CPU whose idle state
4444 * has the smallest exit latency irrespective
4445 * of any idle timestamp.
4447 min_exit_latency = idle->exit_latency;
4448 latest_idle_timestamp = rq->idle_stamp;
4449 shallowest_idle_cpu = i;
4450 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4451 rq->idle_stamp > latest_idle_timestamp) {
4453 * If equal or no active idle state, then
4454 * the most recently idled CPU might have
4457 latest_idle_timestamp = rq->idle_stamp;
4458 shallowest_idle_cpu = i;
4461 load = weighted_cpuload(i);
4462 if (load < min_load || (load == min_load && i == this_cpu)) {
4464 least_loaded_cpu = i;
4469 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4473 * Try and locate an idle CPU in the sched_domain.
4475 static int select_idle_sibling(struct task_struct *p, int target)
4477 struct sched_domain *sd;
4478 struct sched_group *sg;
4479 int i = task_cpu(p);
4481 if (idle_cpu(target))
4485 * If the prevous cpu is cache affine and idle, don't be stupid.
4487 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4491 * Otherwise, iterate the domains and find an elegible idle cpu.
4493 sd = rcu_dereference(per_cpu(sd_llc, target));
4494 for_each_lower_domain(sd) {
4497 if (!cpumask_intersects(sched_group_cpus(sg),
4498 tsk_cpus_allowed(p)))
4501 for_each_cpu(i, sched_group_cpus(sg)) {
4502 if (i == target || !idle_cpu(i))
4506 target = cpumask_first_and(sched_group_cpus(sg),
4507 tsk_cpus_allowed(p));
4511 } while (sg != sd->groups);
4518 * select_task_rq_fair: Select target runqueue for the waking task in domains
4519 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4520 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4522 * Balances load by selecting the idlest cpu in the idlest group, or under
4523 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4525 * Returns the target cpu number.
4527 * preempt must be disabled.
4530 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4532 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4533 int cpu = smp_processor_id();
4535 int want_affine = 0;
4536 int sync = wake_flags & WF_SYNC;
4538 if (p->nr_cpus_allowed == 1)
4541 if (sd_flag & SD_BALANCE_WAKE)
4542 want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4545 for_each_domain(cpu, tmp) {
4546 if (!(tmp->flags & SD_LOAD_BALANCE))
4550 * If both cpu and prev_cpu are part of this domain,
4551 * cpu is a valid SD_WAKE_AFFINE target.
4553 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4554 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4559 if (tmp->flags & sd_flag)
4563 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4566 if (sd_flag & SD_BALANCE_WAKE) {
4567 new_cpu = select_idle_sibling(p, prev_cpu);
4572 struct sched_group *group;
4575 if (!(sd->flags & sd_flag)) {
4580 group = find_idlest_group(sd, p, cpu, sd_flag);
4586 new_cpu = find_idlest_cpu(group, p, cpu);
4587 if (new_cpu == -1 || new_cpu == cpu) {
4588 /* Now try balancing at a lower domain level of cpu */
4593 /* Now try balancing at a lower domain level of new_cpu */
4595 weight = sd->span_weight;
4597 for_each_domain(cpu, tmp) {
4598 if (weight <= tmp->span_weight)
4600 if (tmp->flags & sd_flag)
4603 /* while loop will break here if sd == NULL */
4612 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4613 * cfs_rq_of(p) references at time of call are still valid and identify the
4614 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4615 * other assumptions, including the state of rq->lock, should be made.
4618 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4620 struct sched_entity *se = &p->se;
4621 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4624 * Load tracking: accumulate removed load so that it can be processed
4625 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4626 * to blocked load iff they have a positive decay-count. It can never
4627 * be negative here since on-rq tasks have decay-count == 0.
4629 if (se->avg.decay_count) {
4630 se->avg.decay_count = -__synchronize_entity_decay(se);
4631 atomic_long_add(se->avg.load_avg_contrib,
4632 &cfs_rq->removed_load);
4635 /* We have migrated, no longer consider this task hot */
4638 #endif /* CONFIG_SMP */
4640 static unsigned long
4641 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4643 unsigned long gran = sysctl_sched_wakeup_granularity;
4646 * Since its curr running now, convert the gran from real-time
4647 * to virtual-time in his units.
4649 * By using 'se' instead of 'curr' we penalize light tasks, so
4650 * they get preempted easier. That is, if 'se' < 'curr' then
4651 * the resulting gran will be larger, therefore penalizing the
4652 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4653 * be smaller, again penalizing the lighter task.
4655 * This is especially important for buddies when the leftmost
4656 * task is higher priority than the buddy.
4658 return calc_delta_fair(gran, se);
4662 * Should 'se' preempt 'curr'.
4676 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4678 s64 gran, vdiff = curr->vruntime - se->vruntime;
4683 gran = wakeup_gran(curr, se);
4690 static void set_last_buddy(struct sched_entity *se)
4692 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4695 for_each_sched_entity(se)
4696 cfs_rq_of(se)->last = se;
4699 static void set_next_buddy(struct sched_entity *se)
4701 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4704 for_each_sched_entity(se)
4705 cfs_rq_of(se)->next = se;
4708 static void set_skip_buddy(struct sched_entity *se)
4710 for_each_sched_entity(se)
4711 cfs_rq_of(se)->skip = se;
4715 * Preempt the current task with a newly woken task if needed:
4717 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4719 struct task_struct *curr = rq->curr;
4720 struct sched_entity *se = &curr->se, *pse = &p->se;
4721 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4722 int scale = cfs_rq->nr_running >= sched_nr_latency;
4723 int next_buddy_marked = 0;
4725 if (unlikely(se == pse))
4729 * This is possible from callers such as attach_tasks(), in which we
4730 * unconditionally check_prempt_curr() after an enqueue (which may have
4731 * lead to a throttle). This both saves work and prevents false
4732 * next-buddy nomination below.
4734 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4737 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4738 set_next_buddy(pse);
4739 next_buddy_marked = 1;
4743 * We can come here with TIF_NEED_RESCHED already set from new task
4746 * Note: this also catches the edge-case of curr being in a throttled
4747 * group (e.g. via set_curr_task), since update_curr() (in the
4748 * enqueue of curr) will have resulted in resched being set. This
4749 * prevents us from potentially nominating it as a false LAST_BUDDY
4752 if (test_tsk_need_resched(curr))
4755 /* Idle tasks are by definition preempted by non-idle tasks. */
4756 if (unlikely(curr->policy == SCHED_IDLE) &&
4757 likely(p->policy != SCHED_IDLE))
4761 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4762 * is driven by the tick):
4764 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4767 find_matching_se(&se, &pse);
4768 update_curr(cfs_rq_of(se));
4770 if (wakeup_preempt_entity(se, pse) == 1) {
4772 * Bias pick_next to pick the sched entity that is
4773 * triggering this preemption.
4775 if (!next_buddy_marked)
4776 set_next_buddy(pse);
4785 * Only set the backward buddy when the current task is still
4786 * on the rq. This can happen when a wakeup gets interleaved
4787 * with schedule on the ->pre_schedule() or idle_balance()
4788 * point, either of which can * drop the rq lock.
4790 * Also, during early boot the idle thread is in the fair class,
4791 * for obvious reasons its a bad idea to schedule back to it.
4793 if (unlikely(!se->on_rq || curr == rq->idle))
4796 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4800 static struct task_struct *
4801 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4803 struct cfs_rq *cfs_rq = &rq->cfs;
4804 struct sched_entity *se;
4805 struct task_struct *p;
4809 #ifdef CONFIG_FAIR_GROUP_SCHED
4810 if (!cfs_rq->nr_running)
4813 if (prev->sched_class != &fair_sched_class)
4817 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4818 * likely that a next task is from the same cgroup as the current.
4820 * Therefore attempt to avoid putting and setting the entire cgroup
4821 * hierarchy, only change the part that actually changes.
4825 struct sched_entity *curr = cfs_rq->curr;
4828 * Since we got here without doing put_prev_entity() we also
4829 * have to consider cfs_rq->curr. If it is still a runnable
4830 * entity, update_curr() will update its vruntime, otherwise
4831 * forget we've ever seen it.
4833 if (curr && curr->on_rq)
4834 update_curr(cfs_rq);
4839 * This call to check_cfs_rq_runtime() will do the throttle and
4840 * dequeue its entity in the parent(s). Therefore the 'simple'
4841 * nr_running test will indeed be correct.
4843 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4846 se = pick_next_entity(cfs_rq, curr);
4847 cfs_rq = group_cfs_rq(se);
4853 * Since we haven't yet done put_prev_entity and if the selected task
4854 * is a different task than we started out with, try and touch the
4855 * least amount of cfs_rqs.
4858 struct sched_entity *pse = &prev->se;
4860 while (!(cfs_rq = is_same_group(se, pse))) {
4861 int se_depth = se->depth;
4862 int pse_depth = pse->depth;
4864 if (se_depth <= pse_depth) {
4865 put_prev_entity(cfs_rq_of(pse), pse);
4866 pse = parent_entity(pse);
4868 if (se_depth >= pse_depth) {
4869 set_next_entity(cfs_rq_of(se), se);
4870 se = parent_entity(se);
4874 put_prev_entity(cfs_rq, pse);
4875 set_next_entity(cfs_rq, se);
4878 if (hrtick_enabled(rq))
4879 hrtick_start_fair(rq, p);
4886 if (!cfs_rq->nr_running)
4889 put_prev_task(rq, prev);
4892 se = pick_next_entity(cfs_rq, NULL);
4893 set_next_entity(cfs_rq, se);
4894 cfs_rq = group_cfs_rq(se);
4899 if (hrtick_enabled(rq))
4900 hrtick_start_fair(rq, p);
4905 new_tasks = idle_balance(rq);
4907 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4908 * possible for any higher priority task to appear. In that case we
4909 * must re-start the pick_next_entity() loop.
4921 * Account for a descheduled task:
4923 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4925 struct sched_entity *se = &prev->se;
4926 struct cfs_rq *cfs_rq;
4928 for_each_sched_entity(se) {
4929 cfs_rq = cfs_rq_of(se);
4930 put_prev_entity(cfs_rq, se);
4935 * sched_yield() is very simple
4937 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4939 static void yield_task_fair(struct rq *rq)
4941 struct task_struct *curr = rq->curr;
4942 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4943 struct sched_entity *se = &curr->se;
4946 * Are we the only task in the tree?
4948 if (unlikely(rq->nr_running == 1))
4951 clear_buddies(cfs_rq, se);
4953 if (curr->policy != SCHED_BATCH) {
4954 update_rq_clock(rq);
4956 * Update run-time statistics of the 'current'.
4958 update_curr(cfs_rq);
4960 * Tell update_rq_clock() that we've just updated,
4961 * so we don't do microscopic update in schedule()
4962 * and double the fastpath cost.
4964 rq->skip_clock_update = 1;
4970 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4972 struct sched_entity *se = &p->se;
4974 /* throttled hierarchies are not runnable */
4975 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4978 /* Tell the scheduler that we'd really like pse to run next. */
4981 yield_task_fair(rq);
4987 /**************************************************
4988 * Fair scheduling class load-balancing methods.
4992 * The purpose of load-balancing is to achieve the same basic fairness the
4993 * per-cpu scheduler provides, namely provide a proportional amount of compute
4994 * time to each task. This is expressed in the following equation:
4996 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4998 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4999 * W_i,0 is defined as:
5001 * W_i,0 = \Sum_j w_i,j (2)
5003 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5004 * is derived from the nice value as per prio_to_weight[].
5006 * The weight average is an exponential decay average of the instantaneous
5009 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5011 * C_i is the compute capacity of cpu i, typically it is the
5012 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5013 * can also include other factors [XXX].
5015 * To achieve this balance we define a measure of imbalance which follows
5016 * directly from (1):
5018 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5020 * We them move tasks around to minimize the imbalance. In the continuous
5021 * function space it is obvious this converges, in the discrete case we get
5022 * a few fun cases generally called infeasible weight scenarios.
5025 * - infeasible weights;
5026 * - local vs global optima in the discrete case. ]
5031 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5032 * for all i,j solution, we create a tree of cpus that follows the hardware
5033 * topology where each level pairs two lower groups (or better). This results
5034 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5035 * tree to only the first of the previous level and we decrease the frequency
5036 * of load-balance at each level inv. proportional to the number of cpus in
5042 * \Sum { --- * --- * 2^i } = O(n) (5)
5044 * `- size of each group
5045 * | | `- number of cpus doing load-balance
5047 * `- sum over all levels
5049 * Coupled with a limit on how many tasks we can migrate every balance pass,
5050 * this makes (5) the runtime complexity of the balancer.
5052 * An important property here is that each CPU is still (indirectly) connected
5053 * to every other cpu in at most O(log n) steps:
5055 * The adjacency matrix of the resulting graph is given by:
5058 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5061 * And you'll find that:
5063 * A^(log_2 n)_i,j != 0 for all i,j (7)
5065 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5066 * The task movement gives a factor of O(m), giving a convergence complexity
5069 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5074 * In order to avoid CPUs going idle while there's still work to do, new idle
5075 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5076 * tree itself instead of relying on other CPUs to bring it work.
5078 * This adds some complexity to both (5) and (8) but it reduces the total idle
5086 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5089 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5094 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5096 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5098 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5101 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5102 * rewrite all of this once again.]
5105 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5107 enum fbq_type { regular, remote, all };
5109 #define LBF_ALL_PINNED 0x01
5110 #define LBF_NEED_BREAK 0x02
5111 #define LBF_DST_PINNED 0x04
5112 #define LBF_SOME_PINNED 0x08
5115 struct sched_domain *sd;
5123 struct cpumask *dst_grpmask;
5125 enum cpu_idle_type idle;
5127 /* The set of CPUs under consideration for load-balancing */
5128 struct cpumask *cpus;
5133 unsigned int loop_break;
5134 unsigned int loop_max;
5136 enum fbq_type fbq_type;
5137 struct list_head tasks;
5141 * Is this task likely cache-hot:
5143 static int task_hot(struct task_struct *p, struct lb_env *env)
5147 lockdep_assert_held(&env->src_rq->lock);
5149 if (p->sched_class != &fair_sched_class)
5152 if (unlikely(p->policy == SCHED_IDLE))
5156 * Buddy candidates are cache hot:
5158 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5159 (&p->se == cfs_rq_of(&p->se)->next ||
5160 &p->se == cfs_rq_of(&p->se)->last))
5163 if (sysctl_sched_migration_cost == -1)
5165 if (sysctl_sched_migration_cost == 0)
5168 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5170 return delta < (s64)sysctl_sched_migration_cost;
5173 #ifdef CONFIG_NUMA_BALANCING
5174 /* Returns true if the destination node has incurred more faults */
5175 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5177 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5178 int src_nid, dst_nid;
5180 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5181 !(env->sd->flags & SD_NUMA)) {
5185 src_nid = cpu_to_node(env->src_cpu);
5186 dst_nid = cpu_to_node(env->dst_cpu);
5188 if (src_nid == dst_nid)
5192 /* Task is already in the group's interleave set. */
5193 if (node_isset(src_nid, numa_group->active_nodes))
5196 /* Task is moving into the group's interleave set. */
5197 if (node_isset(dst_nid, numa_group->active_nodes))
5200 return group_faults(p, dst_nid) > group_faults(p, src_nid);
5203 /* Encourage migration to the preferred node. */
5204 if (dst_nid == p->numa_preferred_nid)
5207 return task_faults(p, dst_nid) > task_faults(p, src_nid);
5211 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5213 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5214 int src_nid, dst_nid;
5216 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5219 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5222 src_nid = cpu_to_node(env->src_cpu);
5223 dst_nid = cpu_to_node(env->dst_cpu);
5225 if (src_nid == dst_nid)
5229 /* Task is moving within/into the group's interleave set. */
5230 if (node_isset(dst_nid, numa_group->active_nodes))
5233 /* Task is moving out of the group's interleave set. */
5234 if (node_isset(src_nid, numa_group->active_nodes))
5237 return group_faults(p, dst_nid) < group_faults(p, src_nid);
5240 /* Migrating away from the preferred node is always bad. */
5241 if (src_nid == p->numa_preferred_nid)
5244 return task_faults(p, dst_nid) < task_faults(p, src_nid);
5248 static inline bool migrate_improves_locality(struct task_struct *p,
5254 static inline bool migrate_degrades_locality(struct task_struct *p,
5262 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5265 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5267 int tsk_cache_hot = 0;
5269 lockdep_assert_held(&env->src_rq->lock);
5272 * We do not migrate tasks that are:
5273 * 1) throttled_lb_pair, or
5274 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5275 * 3) running (obviously), or
5276 * 4) are cache-hot on their current CPU.
5278 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5281 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5284 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5286 env->flags |= LBF_SOME_PINNED;
5289 * Remember if this task can be migrated to any other cpu in
5290 * our sched_group. We may want to revisit it if we couldn't
5291 * meet load balance goals by pulling other tasks on src_cpu.
5293 * Also avoid computing new_dst_cpu if we have already computed
5294 * one in current iteration.
5296 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5299 /* Prevent to re-select dst_cpu via env's cpus */
5300 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5301 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5302 env->flags |= LBF_DST_PINNED;
5303 env->new_dst_cpu = cpu;
5311 /* Record that we found atleast one task that could run on dst_cpu */
5312 env->flags &= ~LBF_ALL_PINNED;
5314 if (task_running(env->src_rq, p)) {
5315 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5320 * Aggressive migration if:
5321 * 1) destination numa is preferred
5322 * 2) task is cache cold, or
5323 * 3) too many balance attempts have failed.
5325 tsk_cache_hot = task_hot(p, env);
5327 tsk_cache_hot = migrate_degrades_locality(p, env);
5329 if (migrate_improves_locality(p, env) || !tsk_cache_hot ||
5330 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5331 if (tsk_cache_hot) {
5332 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5333 schedstat_inc(p, se.statistics.nr_forced_migrations);
5338 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5343 * detach_task() -- detach the task for the migration specified in env
5345 static void detach_task(struct task_struct *p, struct lb_env *env)
5347 lockdep_assert_held(&env->src_rq->lock);
5349 deactivate_task(env->src_rq, p, 0);
5350 p->on_rq = TASK_ON_RQ_MIGRATING;
5351 set_task_cpu(p, env->dst_cpu);
5355 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5356 * part of active balancing operations within "domain".
5358 * Returns a task if successful and NULL otherwise.
5360 static struct task_struct *detach_one_task(struct lb_env *env)
5362 struct task_struct *p, *n;
5364 lockdep_assert_held(&env->src_rq->lock);
5366 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5367 if (!can_migrate_task(p, env))
5370 detach_task(p, env);
5373 * Right now, this is only the second place where
5374 * lb_gained[env->idle] is updated (other is detach_tasks)
5375 * so we can safely collect stats here rather than
5376 * inside detach_tasks().
5378 schedstat_inc(env->sd, lb_gained[env->idle]);
5384 static const unsigned int sched_nr_migrate_break = 32;
5387 * detach_tasks() -- tries to detach up to imbalance weighted load from
5388 * busiest_rq, as part of a balancing operation within domain "sd".
5390 * Returns number of detached tasks if successful and 0 otherwise.
5392 static int detach_tasks(struct lb_env *env)
5394 struct list_head *tasks = &env->src_rq->cfs_tasks;
5395 struct task_struct *p;
5399 lockdep_assert_held(&env->src_rq->lock);
5401 if (env->imbalance <= 0)
5404 while (!list_empty(tasks)) {
5405 p = list_first_entry(tasks, struct task_struct, se.group_node);
5408 /* We've more or less seen every task there is, call it quits */
5409 if (env->loop > env->loop_max)
5412 /* take a breather every nr_migrate tasks */
5413 if (env->loop > env->loop_break) {
5414 env->loop_break += sched_nr_migrate_break;
5415 env->flags |= LBF_NEED_BREAK;
5419 if (!can_migrate_task(p, env))
5422 load = task_h_load(p);
5424 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5427 if ((load / 2) > env->imbalance)
5430 detach_task(p, env);
5431 list_add(&p->se.group_node, &env->tasks);
5434 env->imbalance -= load;
5436 #ifdef CONFIG_PREEMPT
5438 * NEWIDLE balancing is a source of latency, so preemptible
5439 * kernels will stop after the first task is detached to minimize
5440 * the critical section.
5442 if (env->idle == CPU_NEWLY_IDLE)
5447 * We only want to steal up to the prescribed amount of
5450 if (env->imbalance <= 0)
5455 list_move_tail(&p->se.group_node, tasks);
5459 * Right now, this is one of only two places we collect this stat
5460 * so we can safely collect detach_one_task() stats here rather
5461 * than inside detach_one_task().
5463 schedstat_add(env->sd, lb_gained[env->idle], detached);
5469 * attach_task() -- attach the task detached by detach_task() to its new rq.
5471 static void attach_task(struct rq *rq, struct task_struct *p)
5473 lockdep_assert_held(&rq->lock);
5475 BUG_ON(task_rq(p) != rq);
5476 p->on_rq = TASK_ON_RQ_QUEUED;
5477 activate_task(rq, p, 0);
5478 check_preempt_curr(rq, p, 0);
5482 * attach_one_task() -- attaches the task returned from detach_one_task() to
5485 static void attach_one_task(struct rq *rq, struct task_struct *p)
5487 raw_spin_lock(&rq->lock);
5489 raw_spin_unlock(&rq->lock);
5493 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5496 static void attach_tasks(struct lb_env *env)
5498 struct list_head *tasks = &env->tasks;
5499 struct task_struct *p;
5501 raw_spin_lock(&env->dst_rq->lock);
5503 while (!list_empty(tasks)) {
5504 p = list_first_entry(tasks, struct task_struct, se.group_node);
5505 list_del_init(&p->se.group_node);
5507 attach_task(env->dst_rq, p);
5510 raw_spin_unlock(&env->dst_rq->lock);
5513 #ifdef CONFIG_FAIR_GROUP_SCHED
5515 * update tg->load_weight by folding this cpu's load_avg
5517 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5519 struct sched_entity *se = tg->se[cpu];
5520 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5522 /* throttled entities do not contribute to load */
5523 if (throttled_hierarchy(cfs_rq))
5526 update_cfs_rq_blocked_load(cfs_rq, 1);
5529 update_entity_load_avg(se, 1);
5531 * We pivot on our runnable average having decayed to zero for
5532 * list removal. This generally implies that all our children
5533 * have also been removed (modulo rounding error or bandwidth
5534 * control); however, such cases are rare and we can fix these
5537 * TODO: fix up out-of-order children on enqueue.
5539 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5540 list_del_leaf_cfs_rq(cfs_rq);
5542 struct rq *rq = rq_of(cfs_rq);
5543 update_rq_runnable_avg(rq, rq->nr_running);
5547 static void update_blocked_averages(int cpu)
5549 struct rq *rq = cpu_rq(cpu);
5550 struct cfs_rq *cfs_rq;
5551 unsigned long flags;
5553 raw_spin_lock_irqsave(&rq->lock, flags);
5554 update_rq_clock(rq);
5556 * Iterates the task_group tree in a bottom up fashion, see
5557 * list_add_leaf_cfs_rq() for details.
5559 for_each_leaf_cfs_rq(rq, cfs_rq) {
5561 * Note: We may want to consider periodically releasing
5562 * rq->lock about these updates so that creating many task
5563 * groups does not result in continually extending hold time.
5565 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5568 raw_spin_unlock_irqrestore(&rq->lock, flags);
5572 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5573 * This needs to be done in a top-down fashion because the load of a child
5574 * group is a fraction of its parents load.
5576 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5578 struct rq *rq = rq_of(cfs_rq);
5579 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5580 unsigned long now = jiffies;
5583 if (cfs_rq->last_h_load_update == now)
5586 cfs_rq->h_load_next = NULL;
5587 for_each_sched_entity(se) {
5588 cfs_rq = cfs_rq_of(se);
5589 cfs_rq->h_load_next = se;
5590 if (cfs_rq->last_h_load_update == now)
5595 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5596 cfs_rq->last_h_load_update = now;
5599 while ((se = cfs_rq->h_load_next) != NULL) {
5600 load = cfs_rq->h_load;
5601 load = div64_ul(load * se->avg.load_avg_contrib,
5602 cfs_rq->runnable_load_avg + 1);
5603 cfs_rq = group_cfs_rq(se);
5604 cfs_rq->h_load = load;
5605 cfs_rq->last_h_load_update = now;
5609 static unsigned long task_h_load(struct task_struct *p)
5611 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5613 update_cfs_rq_h_load(cfs_rq);
5614 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5615 cfs_rq->runnable_load_avg + 1);
5618 static inline void update_blocked_averages(int cpu)
5622 static unsigned long task_h_load(struct task_struct *p)
5624 return p->se.avg.load_avg_contrib;
5628 /********** Helpers for find_busiest_group ************************/
5637 * sg_lb_stats - stats of a sched_group required for load_balancing
5639 struct sg_lb_stats {
5640 unsigned long avg_load; /*Avg load across the CPUs of the group */
5641 unsigned long group_load; /* Total load over the CPUs of the group */
5642 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5643 unsigned long load_per_task;
5644 unsigned long group_capacity;
5645 unsigned int sum_nr_running; /* Nr tasks running in the group */
5646 unsigned int group_capacity_factor;
5647 unsigned int idle_cpus;
5648 unsigned int group_weight;
5649 enum group_type group_type;
5650 int group_has_free_capacity;
5651 #ifdef CONFIG_NUMA_BALANCING
5652 unsigned int nr_numa_running;
5653 unsigned int nr_preferred_running;
5658 * sd_lb_stats - Structure to store the statistics of a sched_domain
5659 * during load balancing.
5661 struct sd_lb_stats {
5662 struct sched_group *busiest; /* Busiest group in this sd */
5663 struct sched_group *local; /* Local group in this sd */
5664 unsigned long total_load; /* Total load of all groups in sd */
5665 unsigned long total_capacity; /* Total capacity of all groups in sd */
5666 unsigned long avg_load; /* Average load across all groups in sd */
5668 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5669 struct sg_lb_stats local_stat; /* Statistics of the local group */
5672 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5675 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5676 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5677 * We must however clear busiest_stat::avg_load because
5678 * update_sd_pick_busiest() reads this before assignment.
5680 *sds = (struct sd_lb_stats){
5684 .total_capacity = 0UL,
5687 .sum_nr_running = 0,
5688 .group_type = group_other,
5694 * get_sd_load_idx - Obtain the load index for a given sched domain.
5695 * @sd: The sched_domain whose load_idx is to be obtained.
5696 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5698 * Return: The load index.
5700 static inline int get_sd_load_idx(struct sched_domain *sd,
5701 enum cpu_idle_type idle)
5707 load_idx = sd->busy_idx;
5710 case CPU_NEWLY_IDLE:
5711 load_idx = sd->newidle_idx;
5714 load_idx = sd->idle_idx;
5721 static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu)
5723 return SCHED_CAPACITY_SCALE;
5726 unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu)
5728 return default_scale_capacity(sd, cpu);
5731 static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu)
5733 if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1))
5734 return sd->smt_gain / sd->span_weight;
5736 return SCHED_CAPACITY_SCALE;
5739 unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
5741 return default_scale_cpu_capacity(sd, cpu);
5744 static unsigned long scale_rt_capacity(int cpu)
5746 struct rq *rq = cpu_rq(cpu);
5747 u64 total, available, age_stamp, avg;
5751 * Since we're reading these variables without serialization make sure
5752 * we read them once before doing sanity checks on them.
5754 age_stamp = ACCESS_ONCE(rq->age_stamp);
5755 avg = ACCESS_ONCE(rq->rt_avg);
5757 delta = rq_clock(rq) - age_stamp;
5758 if (unlikely(delta < 0))
5761 total = sched_avg_period() + delta;
5763 if (unlikely(total < avg)) {
5764 /* Ensures that capacity won't end up being negative */
5767 available = total - avg;
5770 if (unlikely((s64)total < SCHED_CAPACITY_SCALE))
5771 total = SCHED_CAPACITY_SCALE;
5773 total >>= SCHED_CAPACITY_SHIFT;
5775 return div_u64(available, total);
5778 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
5780 unsigned long capacity = SCHED_CAPACITY_SCALE;
5781 struct sched_group *sdg = sd->groups;
5783 if (sched_feat(ARCH_CAPACITY))
5784 capacity *= arch_scale_cpu_capacity(sd, cpu);
5786 capacity *= default_scale_cpu_capacity(sd, cpu);
5788 capacity >>= SCHED_CAPACITY_SHIFT;
5790 sdg->sgc->capacity_orig = capacity;
5792 if (sched_feat(ARCH_CAPACITY))
5793 capacity *= arch_scale_freq_capacity(sd, cpu);
5795 capacity *= default_scale_capacity(sd, cpu);
5797 capacity >>= SCHED_CAPACITY_SHIFT;
5799 capacity *= scale_rt_capacity(cpu);
5800 capacity >>= SCHED_CAPACITY_SHIFT;
5805 cpu_rq(cpu)->cpu_capacity = capacity;
5806 sdg->sgc->capacity = capacity;
5809 void update_group_capacity(struct sched_domain *sd, int cpu)
5811 struct sched_domain *child = sd->child;
5812 struct sched_group *group, *sdg = sd->groups;
5813 unsigned long capacity, capacity_orig;
5814 unsigned long interval;
5816 interval = msecs_to_jiffies(sd->balance_interval);
5817 interval = clamp(interval, 1UL, max_load_balance_interval);
5818 sdg->sgc->next_update = jiffies + interval;
5821 update_cpu_capacity(sd, cpu);
5825 capacity_orig = capacity = 0;
5827 if (child->flags & SD_OVERLAP) {
5829 * SD_OVERLAP domains cannot assume that child groups
5830 * span the current group.
5833 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5834 struct sched_group_capacity *sgc;
5835 struct rq *rq = cpu_rq(cpu);
5838 * build_sched_domains() -> init_sched_groups_capacity()
5839 * gets here before we've attached the domains to the
5842 * Use capacity_of(), which is set irrespective of domains
5843 * in update_cpu_capacity().
5845 * This avoids capacity/capacity_orig from being 0 and
5846 * causing divide-by-zero issues on boot.
5848 * Runtime updates will correct capacity_orig.
5850 if (unlikely(!rq->sd)) {
5851 capacity_orig += capacity_of(cpu);
5852 capacity += capacity_of(cpu);
5856 sgc = rq->sd->groups->sgc;
5857 capacity_orig += sgc->capacity_orig;
5858 capacity += sgc->capacity;
5862 * !SD_OVERLAP domains can assume that child groups
5863 * span the current group.
5866 group = child->groups;
5868 capacity_orig += group->sgc->capacity_orig;
5869 capacity += group->sgc->capacity;
5870 group = group->next;
5871 } while (group != child->groups);
5874 sdg->sgc->capacity_orig = capacity_orig;
5875 sdg->sgc->capacity = capacity;
5879 * Try and fix up capacity for tiny siblings, this is needed when
5880 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5881 * which on its own isn't powerful enough.
5883 * See update_sd_pick_busiest() and check_asym_packing().
5886 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5889 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5891 if (!(sd->flags & SD_SHARE_CPUCAPACITY))
5895 * If ~90% of the cpu_capacity is still there, we're good.
5897 if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29)
5904 * Group imbalance indicates (and tries to solve) the problem where balancing
5905 * groups is inadequate due to tsk_cpus_allowed() constraints.
5907 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5908 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5911 * { 0 1 2 3 } { 4 5 6 7 }
5914 * If we were to balance group-wise we'd place two tasks in the first group and
5915 * two tasks in the second group. Clearly this is undesired as it will overload
5916 * cpu 3 and leave one of the cpus in the second group unused.
5918 * The current solution to this issue is detecting the skew in the first group
5919 * by noticing the lower domain failed to reach balance and had difficulty
5920 * moving tasks due to affinity constraints.
5922 * When this is so detected; this group becomes a candidate for busiest; see
5923 * update_sd_pick_busiest(). And calculate_imbalance() and
5924 * find_busiest_group() avoid some of the usual balance conditions to allow it
5925 * to create an effective group imbalance.
5927 * This is a somewhat tricky proposition since the next run might not find the
5928 * group imbalance and decide the groups need to be balanced again. A most
5929 * subtle and fragile situation.
5932 static inline int sg_imbalanced(struct sched_group *group)
5934 return group->sgc->imbalance;
5938 * Compute the group capacity factor.
5940 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5941 * first dividing out the smt factor and computing the actual number of cores
5942 * and limit unit capacity with that.
5944 static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group)
5946 unsigned int capacity_factor, smt, cpus;
5947 unsigned int capacity, capacity_orig;
5949 capacity = group->sgc->capacity;
5950 capacity_orig = group->sgc->capacity_orig;
5951 cpus = group->group_weight;
5953 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5954 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig);
5955 capacity_factor = cpus / smt; /* cores */
5957 capacity_factor = min_t(unsigned,
5958 capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE));
5959 if (!capacity_factor)
5960 capacity_factor = fix_small_capacity(env->sd, group);
5962 return capacity_factor;
5965 static enum group_type
5966 group_classify(struct sched_group *group, struct sg_lb_stats *sgs)
5968 if (sgs->sum_nr_running > sgs->group_capacity_factor)
5969 return group_overloaded;
5971 if (sg_imbalanced(group))
5972 return group_imbalanced;
5978 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5979 * @env: The load balancing environment.
5980 * @group: sched_group whose statistics are to be updated.
5981 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5982 * @local_group: Does group contain this_cpu.
5983 * @sgs: variable to hold the statistics for this group.
5984 * @overload: Indicate more than one runnable task for any CPU.
5986 static inline void update_sg_lb_stats(struct lb_env *env,
5987 struct sched_group *group, int load_idx,
5988 int local_group, struct sg_lb_stats *sgs,
5994 memset(sgs, 0, sizeof(*sgs));
5996 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5997 struct rq *rq = cpu_rq(i);
5999 /* Bias balancing toward cpus of our domain */
6001 load = target_load(i, load_idx);
6003 load = source_load(i, load_idx);
6005 sgs->group_load += load;
6006 sgs->sum_nr_running += rq->cfs.h_nr_running;
6008 if (rq->nr_running > 1)
6011 #ifdef CONFIG_NUMA_BALANCING
6012 sgs->nr_numa_running += rq->nr_numa_running;
6013 sgs->nr_preferred_running += rq->nr_preferred_running;
6015 sgs->sum_weighted_load += weighted_cpuload(i);
6020 /* Adjust by relative CPU capacity of the group */
6021 sgs->group_capacity = group->sgc->capacity;
6022 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6024 if (sgs->sum_nr_running)
6025 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6027 sgs->group_weight = group->group_weight;
6028 sgs->group_capacity_factor = sg_capacity_factor(env, group);
6029 sgs->group_type = group_classify(group, sgs);
6031 if (sgs->group_capacity_factor > sgs->sum_nr_running)
6032 sgs->group_has_free_capacity = 1;
6036 * update_sd_pick_busiest - return 1 on busiest group
6037 * @env: The load balancing environment.
6038 * @sds: sched_domain statistics
6039 * @sg: sched_group candidate to be checked for being the busiest
6040 * @sgs: sched_group statistics
6042 * Determine if @sg is a busier group than the previously selected
6045 * Return: %true if @sg is a busier group than the previously selected
6046 * busiest group. %false otherwise.
6048 static bool update_sd_pick_busiest(struct lb_env *env,
6049 struct sd_lb_stats *sds,
6050 struct sched_group *sg,
6051 struct sg_lb_stats *sgs)
6053 struct sg_lb_stats *busiest = &sds->busiest_stat;
6055 if (sgs->group_type > busiest->group_type)
6058 if (sgs->group_type < busiest->group_type)
6061 if (sgs->avg_load <= busiest->avg_load)
6064 /* This is the busiest node in its class. */
6065 if (!(env->sd->flags & SD_ASYM_PACKING))
6069 * ASYM_PACKING needs to move all the work to the lowest
6070 * numbered CPUs in the group, therefore mark all groups
6071 * higher than ourself as busy.
6073 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6077 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6084 #ifdef CONFIG_NUMA_BALANCING
6085 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6087 if (sgs->sum_nr_running > sgs->nr_numa_running)
6089 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6094 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6096 if (rq->nr_running > rq->nr_numa_running)
6098 if (rq->nr_running > rq->nr_preferred_running)
6103 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6108 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6112 #endif /* CONFIG_NUMA_BALANCING */
6115 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6116 * @env: The load balancing environment.
6117 * @sds: variable to hold the statistics for this sched_domain.
6119 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6121 struct sched_domain *child = env->sd->child;
6122 struct sched_group *sg = env->sd->groups;
6123 struct sg_lb_stats tmp_sgs;
6124 int load_idx, prefer_sibling = 0;
6125 bool overload = false;
6127 if (child && child->flags & SD_PREFER_SIBLING)
6130 load_idx = get_sd_load_idx(env->sd, env->idle);
6133 struct sg_lb_stats *sgs = &tmp_sgs;
6136 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6139 sgs = &sds->local_stat;
6141 if (env->idle != CPU_NEWLY_IDLE ||
6142 time_after_eq(jiffies, sg->sgc->next_update))
6143 update_group_capacity(env->sd, env->dst_cpu);
6146 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6153 * In case the child domain prefers tasks go to siblings
6154 * first, lower the sg capacity factor to one so that we'll try
6155 * and move all the excess tasks away. We lower the capacity
6156 * of a group only if the local group has the capacity to fit
6157 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6158 * extra check prevents the case where you always pull from the
6159 * heaviest group when it is already under-utilized (possible
6160 * with a large weight task outweighs the tasks on the system).
6162 if (prefer_sibling && sds->local &&
6163 sds->local_stat.group_has_free_capacity)
6164 sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U);
6166 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6168 sds->busiest_stat = *sgs;
6172 /* Now, start updating sd_lb_stats */
6173 sds->total_load += sgs->group_load;
6174 sds->total_capacity += sgs->group_capacity;
6177 } while (sg != env->sd->groups);
6179 if (env->sd->flags & SD_NUMA)
6180 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6182 if (!env->sd->parent) {
6183 /* update overload indicator if we are at root domain */
6184 if (env->dst_rq->rd->overload != overload)
6185 env->dst_rq->rd->overload = overload;
6191 * check_asym_packing - Check to see if the group is packed into the
6194 * This is primarily intended to used at the sibling level. Some
6195 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6196 * case of POWER7, it can move to lower SMT modes only when higher
6197 * threads are idle. When in lower SMT modes, the threads will
6198 * perform better since they share less core resources. Hence when we
6199 * have idle threads, we want them to be the higher ones.
6201 * This packing function is run on idle threads. It checks to see if
6202 * the busiest CPU in this domain (core in the P7 case) has a higher
6203 * CPU number than the packing function is being run on. Here we are
6204 * assuming lower CPU number will be equivalent to lower a SMT thread
6207 * Return: 1 when packing is required and a task should be moved to
6208 * this CPU. The amount of the imbalance is returned in *imbalance.
6210 * @env: The load balancing environment.
6211 * @sds: Statistics of the sched_domain which is to be packed
6213 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6217 if (!(env->sd->flags & SD_ASYM_PACKING))
6223 busiest_cpu = group_first_cpu(sds->busiest);
6224 if (env->dst_cpu > busiest_cpu)
6227 env->imbalance = DIV_ROUND_CLOSEST(
6228 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6229 SCHED_CAPACITY_SCALE);
6235 * fix_small_imbalance - Calculate the minor imbalance that exists
6236 * amongst the groups of a sched_domain, during
6238 * @env: The load balancing environment.
6239 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6242 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6244 unsigned long tmp, capa_now = 0, capa_move = 0;
6245 unsigned int imbn = 2;
6246 unsigned long scaled_busy_load_per_task;
6247 struct sg_lb_stats *local, *busiest;
6249 local = &sds->local_stat;
6250 busiest = &sds->busiest_stat;
6252 if (!local->sum_nr_running)
6253 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6254 else if (busiest->load_per_task > local->load_per_task)
6257 scaled_busy_load_per_task =
6258 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6259 busiest->group_capacity;
6261 if (busiest->avg_load + scaled_busy_load_per_task >=
6262 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6263 env->imbalance = busiest->load_per_task;
6268 * OK, we don't have enough imbalance to justify moving tasks,
6269 * however we may be able to increase total CPU capacity used by
6273 capa_now += busiest->group_capacity *
6274 min(busiest->load_per_task, busiest->avg_load);
6275 capa_now += local->group_capacity *
6276 min(local->load_per_task, local->avg_load);
6277 capa_now /= SCHED_CAPACITY_SCALE;
6279 /* Amount of load we'd subtract */
6280 if (busiest->avg_load > scaled_busy_load_per_task) {
6281 capa_move += busiest->group_capacity *
6282 min(busiest->load_per_task,
6283 busiest->avg_load - scaled_busy_load_per_task);
6286 /* Amount of load we'd add */
6287 if (busiest->avg_load * busiest->group_capacity <
6288 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6289 tmp = (busiest->avg_load * busiest->group_capacity) /
6290 local->group_capacity;
6292 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6293 local->group_capacity;
6295 capa_move += local->group_capacity *
6296 min(local->load_per_task, local->avg_load + tmp);
6297 capa_move /= SCHED_CAPACITY_SCALE;
6299 /* Move if we gain throughput */
6300 if (capa_move > capa_now)
6301 env->imbalance = busiest->load_per_task;
6305 * calculate_imbalance - Calculate the amount of imbalance present within the
6306 * groups of a given sched_domain during load balance.
6307 * @env: load balance environment
6308 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6310 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6312 unsigned long max_pull, load_above_capacity = ~0UL;
6313 struct sg_lb_stats *local, *busiest;
6315 local = &sds->local_stat;
6316 busiest = &sds->busiest_stat;
6318 if (busiest->group_type == group_imbalanced) {
6320 * In the group_imb case we cannot rely on group-wide averages
6321 * to ensure cpu-load equilibrium, look at wider averages. XXX
6323 busiest->load_per_task =
6324 min(busiest->load_per_task, sds->avg_load);
6328 * In the presence of smp nice balancing, certain scenarios can have
6329 * max load less than avg load(as we skip the groups at or below
6330 * its cpu_capacity, while calculating max_load..)
6332 if (busiest->avg_load <= sds->avg_load ||
6333 local->avg_load >= sds->avg_load) {
6335 return fix_small_imbalance(env, sds);
6339 * If there aren't any idle cpus, avoid creating some.
6341 if (busiest->group_type == group_overloaded &&
6342 local->group_type == group_overloaded) {
6343 load_above_capacity =
6344 (busiest->sum_nr_running - busiest->group_capacity_factor);
6346 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE);
6347 load_above_capacity /= busiest->group_capacity;
6351 * We're trying to get all the cpus to the average_load, so we don't
6352 * want to push ourselves above the average load, nor do we wish to
6353 * reduce the max loaded cpu below the average load. At the same time,
6354 * we also don't want to reduce the group load below the group capacity
6355 * (so that we can implement power-savings policies etc). Thus we look
6356 * for the minimum possible imbalance.
6358 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6360 /* How much load to actually move to equalise the imbalance */
6361 env->imbalance = min(
6362 max_pull * busiest->group_capacity,
6363 (sds->avg_load - local->avg_load) * local->group_capacity
6364 ) / SCHED_CAPACITY_SCALE;
6367 * if *imbalance is less than the average load per runnable task
6368 * there is no guarantee that any tasks will be moved so we'll have
6369 * a think about bumping its value to force at least one task to be
6372 if (env->imbalance < busiest->load_per_task)
6373 return fix_small_imbalance(env, sds);
6376 /******* find_busiest_group() helpers end here *********************/
6379 * find_busiest_group - Returns the busiest group within the sched_domain
6380 * if there is an imbalance. If there isn't an imbalance, and
6381 * the user has opted for power-savings, it returns a group whose
6382 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6383 * such a group exists.
6385 * Also calculates the amount of weighted load which should be moved
6386 * to restore balance.
6388 * @env: The load balancing environment.
6390 * Return: - The busiest group if imbalance exists.
6391 * - If no imbalance and user has opted for power-savings balance,
6392 * return the least loaded group whose CPUs can be
6393 * put to idle by rebalancing its tasks onto our group.
6395 static struct sched_group *find_busiest_group(struct lb_env *env)
6397 struct sg_lb_stats *local, *busiest;
6398 struct sd_lb_stats sds;
6400 init_sd_lb_stats(&sds);
6403 * Compute the various statistics relavent for load balancing at
6406 update_sd_lb_stats(env, &sds);
6407 local = &sds.local_stat;
6408 busiest = &sds.busiest_stat;
6410 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6411 check_asym_packing(env, &sds))
6414 /* There is no busy sibling group to pull tasks from */
6415 if (!sds.busiest || busiest->sum_nr_running == 0)
6418 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6419 / sds.total_capacity;
6422 * If the busiest group is imbalanced the below checks don't
6423 * work because they assume all things are equal, which typically
6424 * isn't true due to cpus_allowed constraints and the like.
6426 if (busiest->group_type == group_imbalanced)
6429 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6430 if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity &&
6431 !busiest->group_has_free_capacity)
6435 * If the local group is busier than the selected busiest group
6436 * don't try and pull any tasks.
6438 if (local->avg_load >= busiest->avg_load)
6442 * Don't pull any tasks if this group is already above the domain
6445 if (local->avg_load >= sds.avg_load)
6448 if (env->idle == CPU_IDLE) {
6450 * This cpu is idle. If the busiest group is not overloaded
6451 * and there is no imbalance between this and busiest group
6452 * wrt idle cpus, it is balanced. The imbalance becomes
6453 * significant if the diff is greater than 1 otherwise we
6454 * might end up to just move the imbalance on another group
6456 if ((busiest->group_type != group_overloaded) &&
6457 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6461 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6462 * imbalance_pct to be conservative.
6464 if (100 * busiest->avg_load <=
6465 env->sd->imbalance_pct * local->avg_load)
6470 /* Looks like there is an imbalance. Compute it */
6471 calculate_imbalance(env, &sds);
6480 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6482 static struct rq *find_busiest_queue(struct lb_env *env,
6483 struct sched_group *group)
6485 struct rq *busiest = NULL, *rq;
6486 unsigned long busiest_load = 0, busiest_capacity = 1;
6489 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6490 unsigned long capacity, capacity_factor, wl;
6494 rt = fbq_classify_rq(rq);
6497 * We classify groups/runqueues into three groups:
6498 * - regular: there are !numa tasks
6499 * - remote: there are numa tasks that run on the 'wrong' node
6500 * - all: there is no distinction
6502 * In order to avoid migrating ideally placed numa tasks,
6503 * ignore those when there's better options.
6505 * If we ignore the actual busiest queue to migrate another
6506 * task, the next balance pass can still reduce the busiest
6507 * queue by moving tasks around inside the node.
6509 * If we cannot move enough load due to this classification
6510 * the next pass will adjust the group classification and
6511 * allow migration of more tasks.
6513 * Both cases only affect the total convergence complexity.
6515 if (rt > env->fbq_type)
6518 capacity = capacity_of(i);
6519 capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE);
6520 if (!capacity_factor)
6521 capacity_factor = fix_small_capacity(env->sd, group);
6523 wl = weighted_cpuload(i);
6526 * When comparing with imbalance, use weighted_cpuload()
6527 * which is not scaled with the cpu capacity.
6529 if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance)
6533 * For the load comparisons with the other cpu's, consider
6534 * the weighted_cpuload() scaled with the cpu capacity, so
6535 * that the load can be moved away from the cpu that is
6536 * potentially running at a lower capacity.
6538 * Thus we're looking for max(wl_i / capacity_i), crosswise
6539 * multiplication to rid ourselves of the division works out
6540 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6541 * our previous maximum.
6543 if (wl * busiest_capacity > busiest_load * capacity) {
6545 busiest_capacity = capacity;
6554 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6555 * so long as it is large enough.
6557 #define MAX_PINNED_INTERVAL 512
6559 /* Working cpumask for load_balance and load_balance_newidle. */
6560 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6562 static int need_active_balance(struct lb_env *env)
6564 struct sched_domain *sd = env->sd;
6566 if (env->idle == CPU_NEWLY_IDLE) {
6569 * ASYM_PACKING needs to force migrate tasks from busy but
6570 * higher numbered CPUs in order to pack all tasks in the
6571 * lowest numbered CPUs.
6573 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6577 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6580 static int active_load_balance_cpu_stop(void *data);
6582 static int should_we_balance(struct lb_env *env)
6584 struct sched_group *sg = env->sd->groups;
6585 struct cpumask *sg_cpus, *sg_mask;
6586 int cpu, balance_cpu = -1;
6589 * In the newly idle case, we will allow all the cpu's
6590 * to do the newly idle load balance.
6592 if (env->idle == CPU_NEWLY_IDLE)
6595 sg_cpus = sched_group_cpus(sg);
6596 sg_mask = sched_group_mask(sg);
6597 /* Try to find first idle cpu */
6598 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6599 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6606 if (balance_cpu == -1)
6607 balance_cpu = group_balance_cpu(sg);
6610 * First idle cpu or the first cpu(busiest) in this sched group
6611 * is eligible for doing load balancing at this and above domains.
6613 return balance_cpu == env->dst_cpu;
6617 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6618 * tasks if there is an imbalance.
6620 static int load_balance(int this_cpu, struct rq *this_rq,
6621 struct sched_domain *sd, enum cpu_idle_type idle,
6622 int *continue_balancing)
6624 int ld_moved, cur_ld_moved, active_balance = 0;
6625 struct sched_domain *sd_parent = sd->parent;
6626 struct sched_group *group;
6628 unsigned long flags;
6629 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
6631 struct lb_env env = {
6633 .dst_cpu = this_cpu,
6635 .dst_grpmask = sched_group_cpus(sd->groups),
6637 .loop_break = sched_nr_migrate_break,
6640 .tasks = LIST_HEAD_INIT(env.tasks),
6644 * For NEWLY_IDLE load_balancing, we don't need to consider
6645 * other cpus in our group
6647 if (idle == CPU_NEWLY_IDLE)
6648 env.dst_grpmask = NULL;
6650 cpumask_copy(cpus, cpu_active_mask);
6652 schedstat_inc(sd, lb_count[idle]);
6655 if (!should_we_balance(&env)) {
6656 *continue_balancing = 0;
6660 group = find_busiest_group(&env);
6662 schedstat_inc(sd, lb_nobusyg[idle]);
6666 busiest = find_busiest_queue(&env, group);
6668 schedstat_inc(sd, lb_nobusyq[idle]);
6672 BUG_ON(busiest == env.dst_rq);
6674 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6677 if (busiest->nr_running > 1) {
6679 * Attempt to move tasks. If find_busiest_group has found
6680 * an imbalance but busiest->nr_running <= 1, the group is
6681 * still unbalanced. ld_moved simply stays zero, so it is
6682 * correctly treated as an imbalance.
6684 env.flags |= LBF_ALL_PINNED;
6685 env.src_cpu = busiest->cpu;
6686 env.src_rq = busiest;
6687 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6690 raw_spin_lock_irqsave(&busiest->lock, flags);
6693 * cur_ld_moved - load moved in current iteration
6694 * ld_moved - cumulative load moved across iterations
6696 cur_ld_moved = detach_tasks(&env);
6699 * We've detached some tasks from busiest_rq. Every
6700 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6701 * unlock busiest->lock, and we are able to be sure
6702 * that nobody can manipulate the tasks in parallel.
6703 * See task_rq_lock() family for the details.
6706 raw_spin_unlock(&busiest->lock);
6710 ld_moved += cur_ld_moved;
6713 local_irq_restore(flags);
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;
6767 /* All tasks on this runqueue were pinned by CPU affinity */
6768 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6769 cpumask_clear_cpu(cpu_of(busiest), cpus);
6770 if (!cpumask_empty(cpus)) {
6772 env.loop_break = sched_nr_migrate_break;
6775 goto out_all_pinned;
6780 schedstat_inc(sd, lb_failed[idle]);
6782 * Increment the failure counter only on periodic balance.
6783 * We do not want newidle balance, which can be very
6784 * frequent, pollute the failure counter causing
6785 * excessive cache_hot migrations and active balances.
6787 if (idle != CPU_NEWLY_IDLE)
6788 sd->nr_balance_failed++;
6790 if (need_active_balance(&env)) {
6791 raw_spin_lock_irqsave(&busiest->lock, flags);
6793 /* don't kick the active_load_balance_cpu_stop,
6794 * if the curr task on busiest cpu can't be
6797 if (!cpumask_test_cpu(this_cpu,
6798 tsk_cpus_allowed(busiest->curr))) {
6799 raw_spin_unlock_irqrestore(&busiest->lock,
6801 env.flags |= LBF_ALL_PINNED;
6802 goto out_one_pinned;
6806 * ->active_balance synchronizes accesses to
6807 * ->active_balance_work. Once set, it's cleared
6808 * only after active load balance is finished.
6810 if (!busiest->active_balance) {
6811 busiest->active_balance = 1;
6812 busiest->push_cpu = this_cpu;
6815 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6817 if (active_balance) {
6818 stop_one_cpu_nowait(cpu_of(busiest),
6819 active_load_balance_cpu_stop, busiest,
6820 &busiest->active_balance_work);
6824 * We've kicked active balancing, reset the failure
6827 sd->nr_balance_failed = sd->cache_nice_tries+1;
6830 sd->nr_balance_failed = 0;
6832 if (likely(!active_balance)) {
6833 /* We were unbalanced, so reset the balancing interval */
6834 sd->balance_interval = sd->min_interval;
6837 * If we've begun active balancing, start to back off. This
6838 * case may not be covered by the all_pinned logic if there
6839 * is only 1 task on the busy runqueue (because we don't call
6842 if (sd->balance_interval < sd->max_interval)
6843 sd->balance_interval *= 2;
6850 * We reach balance although we may have faced some affinity
6851 * constraints. Clear the imbalance flag if it was set.
6854 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
6856 if (*group_imbalance)
6857 *group_imbalance = 0;
6862 * We reach balance because all tasks are pinned at this level so
6863 * we can't migrate them. Let the imbalance flag set so parent level
6864 * can try to migrate them.
6866 schedstat_inc(sd, lb_balanced[idle]);
6868 sd->nr_balance_failed = 0;
6871 /* tune up the balancing interval */
6872 if (((env.flags & LBF_ALL_PINNED) &&
6873 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6874 (sd->balance_interval < sd->max_interval))
6875 sd->balance_interval *= 2;
6882 static inline unsigned long
6883 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
6885 unsigned long interval = sd->balance_interval;
6888 interval *= sd->busy_factor;
6890 /* scale ms to jiffies */
6891 interval = msecs_to_jiffies(interval);
6892 interval = clamp(interval, 1UL, max_load_balance_interval);
6898 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
6900 unsigned long interval, next;
6902 interval = get_sd_balance_interval(sd, cpu_busy);
6903 next = sd->last_balance + interval;
6905 if (time_after(*next_balance, next))
6906 *next_balance = next;
6910 * idle_balance is called by schedule() if this_cpu is about to become
6911 * idle. Attempts to pull tasks from other CPUs.
6913 static int idle_balance(struct rq *this_rq)
6915 unsigned long next_balance = jiffies + HZ;
6916 int this_cpu = this_rq->cpu;
6917 struct sched_domain *sd;
6918 int pulled_task = 0;
6921 idle_enter_fair(this_rq);
6924 * We must set idle_stamp _before_ calling idle_balance(), such that we
6925 * measure the duration of idle_balance() as idle time.
6927 this_rq->idle_stamp = rq_clock(this_rq);
6929 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
6930 !this_rq->rd->overload) {
6932 sd = rcu_dereference_check_sched_domain(this_rq->sd);
6934 update_next_balance(sd, 0, &next_balance);
6941 * Drop the rq->lock, but keep IRQ/preempt disabled.
6943 raw_spin_unlock(&this_rq->lock);
6945 update_blocked_averages(this_cpu);
6947 for_each_domain(this_cpu, sd) {
6948 int continue_balancing = 1;
6949 u64 t0, domain_cost;
6951 if (!(sd->flags & SD_LOAD_BALANCE))
6954 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
6955 update_next_balance(sd, 0, &next_balance);
6959 if (sd->flags & SD_BALANCE_NEWIDLE) {
6960 t0 = sched_clock_cpu(this_cpu);
6962 pulled_task = load_balance(this_cpu, this_rq,
6964 &continue_balancing);
6966 domain_cost = sched_clock_cpu(this_cpu) - t0;
6967 if (domain_cost > sd->max_newidle_lb_cost)
6968 sd->max_newidle_lb_cost = domain_cost;
6970 curr_cost += domain_cost;
6973 update_next_balance(sd, 0, &next_balance);
6976 * Stop searching for tasks to pull if there are
6977 * now runnable tasks on this rq.
6979 if (pulled_task || this_rq->nr_running > 0)
6984 raw_spin_lock(&this_rq->lock);
6986 if (curr_cost > this_rq->max_idle_balance_cost)
6987 this_rq->max_idle_balance_cost = curr_cost;
6990 * While browsing the domains, we released the rq lock, a task could
6991 * have been enqueued in the meantime. Since we're not going idle,
6992 * pretend we pulled a task.
6994 if (this_rq->cfs.h_nr_running && !pulled_task)
6998 /* Move the next balance forward */
6999 if (time_after(this_rq->next_balance, next_balance))
7000 this_rq->next_balance = next_balance;
7002 /* Is there a task of a high priority class? */
7003 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7007 idle_exit_fair(this_rq);
7008 this_rq->idle_stamp = 0;
7015 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7016 * running tasks off the busiest CPU onto idle CPUs. It requires at
7017 * least 1 task to be running on each physical CPU where possible, and
7018 * avoids physical / logical imbalances.
7020 static int active_load_balance_cpu_stop(void *data)
7022 struct rq *busiest_rq = data;
7023 int busiest_cpu = cpu_of(busiest_rq);
7024 int target_cpu = busiest_rq->push_cpu;
7025 struct rq *target_rq = cpu_rq(target_cpu);
7026 struct sched_domain *sd;
7027 struct task_struct *p = NULL;
7029 raw_spin_lock_irq(&busiest_rq->lock);
7031 /* make sure the requested cpu hasn't gone down in the meantime */
7032 if (unlikely(busiest_cpu != smp_processor_id() ||
7033 !busiest_rq->active_balance))
7036 /* Is there any task to move? */
7037 if (busiest_rq->nr_running <= 1)
7041 * This condition is "impossible", if it occurs
7042 * we need to fix it. Originally reported by
7043 * Bjorn Helgaas on a 128-cpu setup.
7045 BUG_ON(busiest_rq == target_rq);
7047 /* Search for an sd spanning us and the target CPU. */
7049 for_each_domain(target_cpu, sd) {
7050 if ((sd->flags & SD_LOAD_BALANCE) &&
7051 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7056 struct lb_env env = {
7058 .dst_cpu = target_cpu,
7059 .dst_rq = target_rq,
7060 .src_cpu = busiest_rq->cpu,
7061 .src_rq = busiest_rq,
7065 schedstat_inc(sd, alb_count);
7067 p = detach_one_task(&env);
7069 schedstat_inc(sd, alb_pushed);
7071 schedstat_inc(sd, alb_failed);
7075 busiest_rq->active_balance = 0;
7076 raw_spin_unlock(&busiest_rq->lock);
7079 attach_one_task(target_rq, p);
7086 static inline int on_null_domain(struct rq *rq)
7088 return unlikely(!rcu_dereference_sched(rq->sd));
7091 #ifdef CONFIG_NO_HZ_COMMON
7093 * idle load balancing details
7094 * - When one of the busy CPUs notice that there may be an idle rebalancing
7095 * needed, they will kick the idle load balancer, which then does idle
7096 * load balancing for all the idle CPUs.
7099 cpumask_var_t idle_cpus_mask;
7101 unsigned long next_balance; /* in jiffy units */
7102 } nohz ____cacheline_aligned;
7104 static inline int find_new_ilb(void)
7106 int ilb = cpumask_first(nohz.idle_cpus_mask);
7108 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7115 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7116 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7117 * CPU (if there is one).
7119 static void nohz_balancer_kick(void)
7123 nohz.next_balance++;
7125 ilb_cpu = find_new_ilb();
7127 if (ilb_cpu >= nr_cpu_ids)
7130 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7133 * Use smp_send_reschedule() instead of resched_cpu().
7134 * This way we generate a sched IPI on the target cpu which
7135 * is idle. And the softirq performing nohz idle load balance
7136 * will be run before returning from the IPI.
7138 smp_send_reschedule(ilb_cpu);
7142 static inline void nohz_balance_exit_idle(int cpu)
7144 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7146 * Completely isolated CPUs don't ever set, so we must test.
7148 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7149 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7150 atomic_dec(&nohz.nr_cpus);
7152 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7156 static inline void set_cpu_sd_state_busy(void)
7158 struct sched_domain *sd;
7159 int cpu = smp_processor_id();
7162 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7164 if (!sd || !sd->nohz_idle)
7168 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7173 void set_cpu_sd_state_idle(void)
7175 struct sched_domain *sd;
7176 int cpu = smp_processor_id();
7179 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7181 if (!sd || sd->nohz_idle)
7185 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7191 * This routine will record that the cpu is going idle with tick stopped.
7192 * This info will be used in performing idle load balancing in the future.
7194 void nohz_balance_enter_idle(int cpu)
7197 * If this cpu is going down, then nothing needs to be done.
7199 if (!cpu_active(cpu))
7202 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7206 * If we're a completely isolated CPU, we don't play.
7208 if (on_null_domain(cpu_rq(cpu)))
7211 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7212 atomic_inc(&nohz.nr_cpus);
7213 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7216 static int sched_ilb_notifier(struct notifier_block *nfb,
7217 unsigned long action, void *hcpu)
7219 switch (action & ~CPU_TASKS_FROZEN) {
7221 nohz_balance_exit_idle(smp_processor_id());
7229 static DEFINE_SPINLOCK(balancing);
7232 * Scale the max load_balance interval with the number of CPUs in the system.
7233 * This trades load-balance latency on larger machines for less cross talk.
7235 void update_max_interval(void)
7237 max_load_balance_interval = HZ*num_online_cpus()/10;
7241 * It checks each scheduling domain to see if it is due to be balanced,
7242 * and initiates a balancing operation if so.
7244 * Balancing parameters are set up in init_sched_domains.
7246 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7248 int continue_balancing = 1;
7250 unsigned long interval;
7251 struct sched_domain *sd;
7252 /* Earliest time when we have to do rebalance again */
7253 unsigned long next_balance = jiffies + 60*HZ;
7254 int update_next_balance = 0;
7255 int need_serialize, need_decay = 0;
7258 update_blocked_averages(cpu);
7261 for_each_domain(cpu, sd) {
7263 * Decay the newidle max times here because this is a regular
7264 * visit to all the domains. Decay ~1% per second.
7266 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7267 sd->max_newidle_lb_cost =
7268 (sd->max_newidle_lb_cost * 253) / 256;
7269 sd->next_decay_max_lb_cost = jiffies + HZ;
7272 max_cost += sd->max_newidle_lb_cost;
7274 if (!(sd->flags & SD_LOAD_BALANCE))
7278 * Stop the load balance at this level. There is another
7279 * CPU in our sched group which is doing load balancing more
7282 if (!continue_balancing) {
7288 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7290 need_serialize = sd->flags & SD_SERIALIZE;
7291 if (need_serialize) {
7292 if (!spin_trylock(&balancing))
7296 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7297 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7299 * The LBF_DST_PINNED logic could have changed
7300 * env->dst_cpu, so we can't know our idle
7301 * state even if we migrated tasks. Update it.
7303 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7305 sd->last_balance = jiffies;
7306 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7309 spin_unlock(&balancing);
7311 if (time_after(next_balance, sd->last_balance + interval)) {
7312 next_balance = sd->last_balance + interval;
7313 update_next_balance = 1;
7318 * Ensure the rq-wide value also decays but keep it at a
7319 * reasonable floor to avoid funnies with rq->avg_idle.
7321 rq->max_idle_balance_cost =
7322 max((u64)sysctl_sched_migration_cost, max_cost);
7327 * next_balance will be updated only when there is a need.
7328 * When the cpu is attached to null domain for ex, it will not be
7331 if (likely(update_next_balance))
7332 rq->next_balance = next_balance;
7335 #ifdef CONFIG_NO_HZ_COMMON
7337 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7338 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7340 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7342 int this_cpu = this_rq->cpu;
7346 if (idle != CPU_IDLE ||
7347 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7350 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7351 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7355 * If this cpu gets work to do, stop the load balancing
7356 * work being done for other cpus. Next load
7357 * balancing owner will pick it up.
7362 rq = cpu_rq(balance_cpu);
7365 * If time for next balance is due,
7368 if (time_after_eq(jiffies, rq->next_balance)) {
7369 raw_spin_lock_irq(&rq->lock);
7370 update_rq_clock(rq);
7371 update_idle_cpu_load(rq);
7372 raw_spin_unlock_irq(&rq->lock);
7373 rebalance_domains(rq, CPU_IDLE);
7376 if (time_after(this_rq->next_balance, rq->next_balance))
7377 this_rq->next_balance = rq->next_balance;
7379 nohz.next_balance = this_rq->next_balance;
7381 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7385 * Current heuristic for kicking the idle load balancer in the presence
7386 * of an idle cpu is the system.
7387 * - This rq has more than one task.
7388 * - At any scheduler domain level, this cpu's scheduler group has multiple
7389 * busy cpu's exceeding the group's capacity.
7390 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7391 * domain span are idle.
7393 static inline int nohz_kick_needed(struct rq *rq)
7395 unsigned long now = jiffies;
7396 struct sched_domain *sd;
7397 struct sched_group_capacity *sgc;
7398 int nr_busy, cpu = rq->cpu;
7400 if (unlikely(rq->idle_balance))
7404 * We may be recently in ticked or tickless idle mode. At the first
7405 * busy tick after returning from idle, we will update the busy stats.
7407 set_cpu_sd_state_busy();
7408 nohz_balance_exit_idle(cpu);
7411 * None are in tickless mode and hence no need for NOHZ idle load
7414 if (likely(!atomic_read(&nohz.nr_cpus)))
7417 if (time_before(now, nohz.next_balance))
7420 if (rq->nr_running >= 2)
7424 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7427 sgc = sd->groups->sgc;
7428 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7431 goto need_kick_unlock;
7434 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7436 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7437 sched_domain_span(sd)) < cpu))
7438 goto need_kick_unlock;
7449 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7453 * run_rebalance_domains is triggered when needed from the scheduler tick.
7454 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7456 static void run_rebalance_domains(struct softirq_action *h)
7458 struct rq *this_rq = this_rq();
7459 enum cpu_idle_type idle = this_rq->idle_balance ?
7460 CPU_IDLE : CPU_NOT_IDLE;
7462 rebalance_domains(this_rq, idle);
7465 * If this cpu has a pending nohz_balance_kick, then do the
7466 * balancing on behalf of the other idle cpus whose ticks are
7469 nohz_idle_balance(this_rq, idle);
7473 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7475 void trigger_load_balance(struct rq *rq)
7477 /* Don't need to rebalance while attached to NULL domain */
7478 if (unlikely(on_null_domain(rq)))
7481 if (time_after_eq(jiffies, rq->next_balance))
7482 raise_softirq(SCHED_SOFTIRQ);
7483 #ifdef CONFIG_NO_HZ_COMMON
7484 if (nohz_kick_needed(rq))
7485 nohz_balancer_kick();
7489 static void rq_online_fair(struct rq *rq)
7493 update_runtime_enabled(rq);
7496 static void rq_offline_fair(struct rq *rq)
7500 /* Ensure any throttled groups are reachable by pick_next_task */
7501 unthrottle_offline_cfs_rqs(rq);
7504 #endif /* CONFIG_SMP */
7507 * scheduler tick hitting a task of our scheduling class:
7509 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7511 struct cfs_rq *cfs_rq;
7512 struct sched_entity *se = &curr->se;
7514 for_each_sched_entity(se) {
7515 cfs_rq = cfs_rq_of(se);
7516 entity_tick(cfs_rq, se, queued);
7519 if (numabalancing_enabled)
7520 task_tick_numa(rq, curr);
7522 update_rq_runnable_avg(rq, 1);
7526 * called on fork with the child task as argument from the parent's context
7527 * - child not yet on the tasklist
7528 * - preemption disabled
7530 static void task_fork_fair(struct task_struct *p)
7532 struct cfs_rq *cfs_rq;
7533 struct sched_entity *se = &p->se, *curr;
7534 int this_cpu = smp_processor_id();
7535 struct rq *rq = this_rq();
7536 unsigned long flags;
7538 raw_spin_lock_irqsave(&rq->lock, flags);
7540 update_rq_clock(rq);
7542 cfs_rq = task_cfs_rq(current);
7543 curr = cfs_rq->curr;
7546 * Not only the cpu but also the task_group of the parent might have
7547 * been changed after parent->se.parent,cfs_rq were copied to
7548 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7549 * of child point to valid ones.
7552 __set_task_cpu(p, this_cpu);
7555 update_curr(cfs_rq);
7558 se->vruntime = curr->vruntime;
7559 place_entity(cfs_rq, se, 1);
7561 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7563 * Upon rescheduling, sched_class::put_prev_task() will place
7564 * 'current' within the tree based on its new key value.
7566 swap(curr->vruntime, se->vruntime);
7570 se->vruntime -= cfs_rq->min_vruntime;
7572 raw_spin_unlock_irqrestore(&rq->lock, flags);
7576 * Priority of the task has changed. Check to see if we preempt
7580 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7582 if (!task_on_rq_queued(p))
7586 * Reschedule if we are currently running on this runqueue and
7587 * our priority decreased, or if we are not currently running on
7588 * this runqueue and our priority is higher than the current's
7590 if (rq->curr == p) {
7591 if (p->prio > oldprio)
7594 check_preempt_curr(rq, p, 0);
7597 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7599 struct sched_entity *se = &p->se;
7600 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7603 * Ensure the task's vruntime is normalized, so that when it's
7604 * switched back to the fair class the enqueue_entity(.flags=0) will
7605 * do the right thing.
7607 * If it's queued, then the dequeue_entity(.flags=0) will already
7608 * have normalized the vruntime, if it's !queued, then only when
7609 * the task is sleeping will it still have non-normalized vruntime.
7611 if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) {
7613 * Fix up our vruntime so that the current sleep doesn't
7614 * cause 'unlimited' sleep bonus.
7616 place_entity(cfs_rq, se, 0);
7617 se->vruntime -= cfs_rq->min_vruntime;
7622 * Remove our load from contribution when we leave sched_fair
7623 * and ensure we don't carry in an old decay_count if we
7626 if (se->avg.decay_count) {
7627 __synchronize_entity_decay(se);
7628 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7634 * We switched to the sched_fair class.
7636 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7638 #ifdef CONFIG_FAIR_GROUP_SCHED
7639 struct sched_entity *se = &p->se;
7641 * Since the real-depth could have been changed (only FAIR
7642 * class maintain depth value), reset depth properly.
7644 se->depth = se->parent ? se->parent->depth + 1 : 0;
7646 if (!task_on_rq_queued(p))
7650 * We were most likely switched from sched_rt, so
7651 * kick off the schedule if running, otherwise just see
7652 * if we can still preempt the current task.
7657 check_preempt_curr(rq, p, 0);
7660 /* Account for a task changing its policy or group.
7662 * This routine is mostly called to set cfs_rq->curr field when a task
7663 * migrates between groups/classes.
7665 static void set_curr_task_fair(struct rq *rq)
7667 struct sched_entity *se = &rq->curr->se;
7669 for_each_sched_entity(se) {
7670 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7672 set_next_entity(cfs_rq, se);
7673 /* ensure bandwidth has been allocated on our new cfs_rq */
7674 account_cfs_rq_runtime(cfs_rq, 0);
7678 void init_cfs_rq(struct cfs_rq *cfs_rq)
7680 cfs_rq->tasks_timeline = RB_ROOT;
7681 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7682 #ifndef CONFIG_64BIT
7683 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7686 atomic64_set(&cfs_rq->decay_counter, 1);
7687 atomic_long_set(&cfs_rq->removed_load, 0);
7691 #ifdef CONFIG_FAIR_GROUP_SCHED
7692 static void task_move_group_fair(struct task_struct *p, int queued)
7694 struct sched_entity *se = &p->se;
7695 struct cfs_rq *cfs_rq;
7698 * If the task was not on the rq at the time of this cgroup movement
7699 * it must have been asleep, sleeping tasks keep their ->vruntime
7700 * absolute on their old rq until wakeup (needed for the fair sleeper
7701 * bonus in place_entity()).
7703 * If it was on the rq, we've just 'preempted' it, which does convert
7704 * ->vruntime to a relative base.
7706 * Make sure both cases convert their relative position when migrating
7707 * to another cgroup's rq. This does somewhat interfere with the
7708 * fair sleeper stuff for the first placement, but who cares.
7711 * When !queued, vruntime of the task has usually NOT been normalized.
7712 * But there are some cases where it has already been normalized:
7714 * - Moving a forked child which is waiting for being woken up by
7715 * wake_up_new_task().
7716 * - Moving a task which has been woken up by try_to_wake_up() and
7717 * waiting for actually being woken up by sched_ttwu_pending().
7719 * To prevent boost or penalty in the new cfs_rq caused by delta
7720 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7722 if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7726 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7727 set_task_rq(p, task_cpu(p));
7728 se->depth = se->parent ? se->parent->depth + 1 : 0;
7730 cfs_rq = cfs_rq_of(se);
7731 se->vruntime += cfs_rq->min_vruntime;
7734 * migrate_task_rq_fair() will have removed our previous
7735 * contribution, but we must synchronize for ongoing future
7738 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7739 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7744 void free_fair_sched_group(struct task_group *tg)
7748 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7750 for_each_possible_cpu(i) {
7752 kfree(tg->cfs_rq[i]);
7761 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7763 struct cfs_rq *cfs_rq;
7764 struct sched_entity *se;
7767 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7770 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7774 tg->shares = NICE_0_LOAD;
7776 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7778 for_each_possible_cpu(i) {
7779 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7780 GFP_KERNEL, cpu_to_node(i));
7784 se = kzalloc_node(sizeof(struct sched_entity),
7785 GFP_KERNEL, cpu_to_node(i));
7789 init_cfs_rq(cfs_rq);
7790 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7801 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7803 struct rq *rq = cpu_rq(cpu);
7804 unsigned long flags;
7807 * Only empty task groups can be destroyed; so we can speculatively
7808 * check on_list without danger of it being re-added.
7810 if (!tg->cfs_rq[cpu]->on_list)
7813 raw_spin_lock_irqsave(&rq->lock, flags);
7814 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7815 raw_spin_unlock_irqrestore(&rq->lock, flags);
7818 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7819 struct sched_entity *se, int cpu,
7820 struct sched_entity *parent)
7822 struct rq *rq = cpu_rq(cpu);
7826 init_cfs_rq_runtime(cfs_rq);
7828 tg->cfs_rq[cpu] = cfs_rq;
7831 /* se could be NULL for root_task_group */
7836 se->cfs_rq = &rq->cfs;
7839 se->cfs_rq = parent->my_q;
7840 se->depth = parent->depth + 1;
7844 /* guarantee group entities always have weight */
7845 update_load_set(&se->load, NICE_0_LOAD);
7846 se->parent = parent;
7849 static DEFINE_MUTEX(shares_mutex);
7851 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7854 unsigned long flags;
7857 * We can't change the weight of the root cgroup.
7862 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7864 mutex_lock(&shares_mutex);
7865 if (tg->shares == shares)
7868 tg->shares = shares;
7869 for_each_possible_cpu(i) {
7870 struct rq *rq = cpu_rq(i);
7871 struct sched_entity *se;
7874 /* Propagate contribution to hierarchy */
7875 raw_spin_lock_irqsave(&rq->lock, flags);
7877 /* Possible calls to update_curr() need rq clock */
7878 update_rq_clock(rq);
7879 for_each_sched_entity(se)
7880 update_cfs_shares(group_cfs_rq(se));
7881 raw_spin_unlock_irqrestore(&rq->lock, flags);
7885 mutex_unlock(&shares_mutex);
7888 #else /* CONFIG_FAIR_GROUP_SCHED */
7890 void free_fair_sched_group(struct task_group *tg) { }
7892 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7897 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7899 #endif /* CONFIG_FAIR_GROUP_SCHED */
7902 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7904 struct sched_entity *se = &task->se;
7905 unsigned int rr_interval = 0;
7908 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7911 if (rq->cfs.load.weight)
7912 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7918 * All the scheduling class methods:
7920 const struct sched_class fair_sched_class = {
7921 .next = &idle_sched_class,
7922 .enqueue_task = enqueue_task_fair,
7923 .dequeue_task = dequeue_task_fair,
7924 .yield_task = yield_task_fair,
7925 .yield_to_task = yield_to_task_fair,
7927 .check_preempt_curr = check_preempt_wakeup,
7929 .pick_next_task = pick_next_task_fair,
7930 .put_prev_task = put_prev_task_fair,
7933 .select_task_rq = select_task_rq_fair,
7934 .migrate_task_rq = migrate_task_rq_fair,
7936 .rq_online = rq_online_fair,
7937 .rq_offline = rq_offline_fair,
7939 .task_waking = task_waking_fair,
7942 .set_curr_task = set_curr_task_fair,
7943 .task_tick = task_tick_fair,
7944 .task_fork = task_fork_fair,
7946 .prio_changed = prio_changed_fair,
7947 .switched_from = switched_from_fair,
7948 .switched_to = switched_to_fair,
7950 .get_rr_interval = get_rr_interval_fair,
7952 #ifdef CONFIG_FAIR_GROUP_SCHED
7953 .task_move_group = task_move_group_fair,
7957 #ifdef CONFIG_SCHED_DEBUG
7958 void print_cfs_stats(struct seq_file *m, int cpu)
7960 struct cfs_rq *cfs_rq;
7963 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7964 print_cfs_rq(m, cpu, cfs_rq);
7969 __init void init_sched_fair_class(void)
7972 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7974 #ifdef CONFIG_NO_HZ_COMMON
7975 nohz.next_balance = jiffies;
7976 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7977 cpu_notifier(sched_ilb_notifier, 0);