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,
937 unsigned long faults, total_faults;
939 if (!p->numa_faults_memory)
942 total_faults = p->total_numa_faults;
947 faults = task_faults(p, nid);
948 return 1000 * faults / total_faults;
951 static inline unsigned long group_weight(struct task_struct *p, int nid,
954 unsigned long faults, total_faults;
959 total_faults = p->numa_group->total_faults;
964 faults = group_faults(p, nid);
965 return 1000 * faults / total_faults;
968 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
969 int src_nid, int dst_cpu)
971 struct numa_group *ng = p->numa_group;
972 int dst_nid = cpu_to_node(dst_cpu);
973 int last_cpupid, this_cpupid;
975 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
978 * Multi-stage node selection is used in conjunction with a periodic
979 * migration fault to build a temporal task<->page relation. By using
980 * a two-stage filter we remove short/unlikely relations.
982 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
983 * a task's usage of a particular page (n_p) per total usage of this
984 * page (n_t) (in a given time-span) to a probability.
986 * Our periodic faults will sample this probability and getting the
987 * same result twice in a row, given these samples are fully
988 * independent, is then given by P(n)^2, provided our sample period
989 * is sufficiently short compared to the usage pattern.
991 * This quadric squishes small probabilities, making it less likely we
992 * act on an unlikely task<->page relation.
994 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
995 if (!cpupid_pid_unset(last_cpupid) &&
996 cpupid_to_nid(last_cpupid) != dst_nid)
999 /* Always allow migrate on private faults */
1000 if (cpupid_match_pid(p, last_cpupid))
1003 /* A shared fault, but p->numa_group has not been set up yet. */
1008 * Do not migrate if the destination is not a node that
1009 * is actively used by this numa group.
1011 if (!node_isset(dst_nid, ng->active_nodes))
1015 * Source is a node that is not actively used by this
1016 * numa group, while the destination is. Migrate.
1018 if (!node_isset(src_nid, ng->active_nodes))
1022 * Both source and destination are nodes in active
1023 * use by this numa group. Maximize memory bandwidth
1024 * by migrating from more heavily used groups, to less
1025 * heavily used ones, spreading the load around.
1026 * Use a 1/4 hysteresis to avoid spurious page movement.
1028 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1031 static unsigned long weighted_cpuload(const int cpu);
1032 static unsigned long source_load(int cpu, int type);
1033 static unsigned long target_load(int cpu, int type);
1034 static unsigned long capacity_of(int cpu);
1035 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1037 /* Cached statistics for all CPUs within a node */
1039 unsigned long nr_running;
1042 /* Total compute capacity of CPUs on a node */
1043 unsigned long compute_capacity;
1045 /* Approximate capacity in terms of runnable tasks on a node */
1046 unsigned long task_capacity;
1047 int has_free_capacity;
1051 * XXX borrowed from update_sg_lb_stats
1053 static void update_numa_stats(struct numa_stats *ns, int nid)
1055 int smt, cpu, cpus = 0;
1056 unsigned long capacity;
1058 memset(ns, 0, sizeof(*ns));
1059 for_each_cpu(cpu, cpumask_of_node(nid)) {
1060 struct rq *rq = cpu_rq(cpu);
1062 ns->nr_running += rq->nr_running;
1063 ns->load += weighted_cpuload(cpu);
1064 ns->compute_capacity += capacity_of(cpu);
1070 * If we raced with hotplug and there are no CPUs left in our mask
1071 * the @ns structure is NULL'ed and task_numa_compare() will
1072 * not find this node attractive.
1074 * We'll either bail at !has_free_capacity, or we'll detect a huge
1075 * imbalance and bail there.
1080 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1081 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1082 capacity = cpus / smt; /* cores */
1084 ns->task_capacity = min_t(unsigned, capacity,
1085 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1086 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1089 struct task_numa_env {
1090 struct task_struct *p;
1092 int src_cpu, src_nid;
1093 int dst_cpu, dst_nid;
1095 struct numa_stats src_stats, dst_stats;
1100 struct task_struct *best_task;
1105 static void task_numa_assign(struct task_numa_env *env,
1106 struct task_struct *p, long imp)
1109 put_task_struct(env->best_task);
1114 env->best_imp = imp;
1115 env->best_cpu = env->dst_cpu;
1118 static bool load_too_imbalanced(long src_load, long dst_load,
1119 struct task_numa_env *env)
1122 long orig_src_load, orig_dst_load;
1123 long src_capacity, dst_capacity;
1126 * The load is corrected for the CPU capacity available on each node.
1129 * ------------ vs ---------
1130 * src_capacity dst_capacity
1132 src_capacity = env->src_stats.compute_capacity;
1133 dst_capacity = env->dst_stats.compute_capacity;
1135 /* We care about the slope of the imbalance, not the direction. */
1136 if (dst_load < src_load)
1137 swap(dst_load, src_load);
1139 /* Is the difference below the threshold? */
1140 imb = dst_load * src_capacity * 100 -
1141 src_load * dst_capacity * env->imbalance_pct;
1146 * The imbalance is above the allowed threshold.
1147 * Compare it with the old imbalance.
1149 orig_src_load = env->src_stats.load;
1150 orig_dst_load = env->dst_stats.load;
1152 if (orig_dst_load < orig_src_load)
1153 swap(orig_dst_load, orig_src_load);
1155 old_imb = orig_dst_load * src_capacity * 100 -
1156 orig_src_load * dst_capacity * env->imbalance_pct;
1158 /* Would this change make things worse? */
1159 return (imb > old_imb);
1163 * This checks if the overall compute and NUMA accesses of the system would
1164 * be improved if the source tasks was migrated to the target dst_cpu taking
1165 * into account that it might be best if task running on the dst_cpu should
1166 * be exchanged with the source task
1168 static void task_numa_compare(struct task_numa_env *env,
1169 long taskimp, long groupimp)
1171 struct rq *src_rq = cpu_rq(env->src_cpu);
1172 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1173 struct task_struct *cur;
1174 long src_load, dst_load;
1176 long imp = env->p->numa_group ? groupimp : taskimp;
1178 int dist = env->dist;
1182 raw_spin_lock_irq(&dst_rq->lock);
1185 * No need to move the exiting task, and this ensures that ->curr
1186 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1187 * is safe under RCU read lock.
1188 * Note that rcu_read_lock() itself can't protect from the final
1189 * put_task_struct() after the last schedule().
1191 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1193 raw_spin_unlock_irq(&dst_rq->lock);
1196 * "imp" is the fault differential for the source task between the
1197 * source and destination node. Calculate the total differential for
1198 * the source task and potential destination task. The more negative
1199 * the value is, the more rmeote accesses that would be expected to
1200 * be incurred if the tasks were swapped.
1203 /* Skip this swap candidate if cannot move to the source cpu */
1204 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1208 * If dst and source tasks are in the same NUMA group, or not
1209 * in any group then look only at task weights.
1211 if (cur->numa_group == env->p->numa_group) {
1212 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1213 task_weight(cur, env->dst_nid, dist);
1215 * Add some hysteresis to prevent swapping the
1216 * tasks within a group over tiny differences.
1218 if (cur->numa_group)
1222 * Compare the group weights. If a task is all by
1223 * itself (not part of a group), use the task weight
1226 if (cur->numa_group)
1227 imp += group_weight(cur, env->src_nid, dist) -
1228 group_weight(cur, env->dst_nid, dist);
1230 imp += task_weight(cur, env->src_nid, dist) -
1231 task_weight(cur, env->dst_nid, dist);
1235 if (imp <= env->best_imp && moveimp <= env->best_imp)
1239 /* Is there capacity at our destination? */
1240 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1241 !env->dst_stats.has_free_capacity)
1247 /* Balance doesn't matter much if we're running a task per cpu */
1248 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1249 dst_rq->nr_running == 1)
1253 * In the overloaded case, try and keep the load balanced.
1256 load = task_h_load(env->p);
1257 dst_load = env->dst_stats.load + load;
1258 src_load = env->src_stats.load - load;
1260 if (moveimp > imp && moveimp > env->best_imp) {
1262 * If the improvement from just moving env->p direction is
1263 * better than swapping tasks around, check if a move is
1264 * possible. Store a slightly smaller score than moveimp,
1265 * so an actually idle CPU will win.
1267 if (!load_too_imbalanced(src_load, dst_load, env)) {
1274 if (imp <= env->best_imp)
1278 load = task_h_load(cur);
1283 if (load_too_imbalanced(src_load, dst_load, env))
1287 * One idle CPU per node is evaluated for a task numa move.
1288 * Call select_idle_sibling to maybe find a better one.
1291 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1294 task_numa_assign(env, cur, imp);
1299 static void task_numa_find_cpu(struct task_numa_env *env,
1300 long taskimp, long groupimp)
1304 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1305 /* Skip this CPU if the source task cannot migrate */
1306 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1310 task_numa_compare(env, taskimp, groupimp);
1314 static int task_numa_migrate(struct task_struct *p)
1316 struct task_numa_env env = {
1319 .src_cpu = task_cpu(p),
1320 .src_nid = task_node(p),
1322 .imbalance_pct = 112,
1328 struct sched_domain *sd;
1329 unsigned long taskweight, groupweight;
1331 long taskimp, groupimp;
1334 * Pick the lowest SD_NUMA domain, as that would have the smallest
1335 * imbalance and would be the first to start moving tasks about.
1337 * And we want to avoid any moving of tasks about, as that would create
1338 * random movement of tasks -- counter the numa conditions we're trying
1342 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1344 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1348 * Cpusets can break the scheduler domain tree into smaller
1349 * balance domains, some of which do not cross NUMA boundaries.
1350 * Tasks that are "trapped" in such domains cannot be migrated
1351 * elsewhere, so there is no point in (re)trying.
1353 if (unlikely(!sd)) {
1354 p->numa_preferred_nid = task_node(p);
1358 env.dst_nid = p->numa_preferred_nid;
1359 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1360 taskweight = task_weight(p, env.src_nid, dist);
1361 groupweight = group_weight(p, env.src_nid, dist);
1362 update_numa_stats(&env.src_stats, env.src_nid);
1363 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1364 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1365 update_numa_stats(&env.dst_stats, env.dst_nid);
1367 /* Try to find a spot on the preferred nid. */
1368 task_numa_find_cpu(&env, taskimp, groupimp);
1370 /* No space available on the preferred nid. Look elsewhere. */
1371 if (env.best_cpu == -1) {
1372 for_each_online_node(nid) {
1373 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1376 dist = node_distance(env.src_nid, env.dst_nid);
1378 /* Only consider nodes where both task and groups benefit */
1379 taskimp = task_weight(p, nid, dist) - taskweight;
1380 groupimp = group_weight(p, nid, dist) - groupweight;
1381 if (taskimp < 0 && groupimp < 0)
1386 update_numa_stats(&env.dst_stats, env.dst_nid);
1387 task_numa_find_cpu(&env, taskimp, groupimp);
1392 * If the task is part of a workload that spans multiple NUMA nodes,
1393 * and is migrating into one of the workload's active nodes, remember
1394 * this node as the task's preferred numa node, so the workload can
1396 * A task that migrated to a second choice node will be better off
1397 * trying for a better one later. Do not set the preferred node here.
1399 if (p->numa_group) {
1400 if (env.best_cpu == -1)
1405 if (node_isset(nid, p->numa_group->active_nodes))
1406 sched_setnuma(p, env.dst_nid);
1409 /* No better CPU than the current one was found. */
1410 if (env.best_cpu == -1)
1414 * Reset the scan period if the task is being rescheduled on an
1415 * alternative node to recheck if the tasks is now properly placed.
1417 p->numa_scan_period = task_scan_min(p);
1419 if (env.best_task == NULL) {
1420 ret = migrate_task_to(p, env.best_cpu);
1422 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1426 ret = migrate_swap(p, env.best_task);
1428 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1429 put_task_struct(env.best_task);
1433 /* Attempt to migrate a task to a CPU on the preferred node. */
1434 static void numa_migrate_preferred(struct task_struct *p)
1436 unsigned long interval = HZ;
1438 /* This task has no NUMA fault statistics yet */
1439 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1442 /* Periodically retry migrating the task to the preferred node */
1443 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1444 p->numa_migrate_retry = jiffies + interval;
1446 /* Success if task is already running on preferred CPU */
1447 if (task_node(p) == p->numa_preferred_nid)
1450 /* Otherwise, try migrate to a CPU on the preferred node */
1451 task_numa_migrate(p);
1455 * Find the nodes on which the workload is actively running. We do this by
1456 * tracking the nodes from which NUMA hinting faults are triggered. This can
1457 * be different from the set of nodes where the workload's memory is currently
1460 * The bitmask is used to make smarter decisions on when to do NUMA page
1461 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1462 * are added when they cause over 6/16 of the maximum number of faults, but
1463 * only removed when they drop below 3/16.
1465 static void update_numa_active_node_mask(struct numa_group *numa_group)
1467 unsigned long faults, max_faults = 0;
1470 for_each_online_node(nid) {
1471 faults = group_faults_cpu(numa_group, nid);
1472 if (faults > max_faults)
1473 max_faults = faults;
1476 for_each_online_node(nid) {
1477 faults = group_faults_cpu(numa_group, nid);
1478 if (!node_isset(nid, numa_group->active_nodes)) {
1479 if (faults > max_faults * 6 / 16)
1480 node_set(nid, numa_group->active_nodes);
1481 } else if (faults < max_faults * 3 / 16)
1482 node_clear(nid, numa_group->active_nodes);
1487 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1488 * increments. The more local the fault statistics are, the higher the scan
1489 * period will be for the next scan window. If local/(local+remote) ratio is
1490 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1491 * the scan period will decrease. Aim for 70% local accesses.
1493 #define NUMA_PERIOD_SLOTS 10
1494 #define NUMA_PERIOD_THRESHOLD 7
1497 * Increase the scan period (slow down scanning) if the majority of
1498 * our memory is already on our local node, or if the majority of
1499 * the page accesses are shared with other processes.
1500 * Otherwise, decrease the scan period.
1502 static void update_task_scan_period(struct task_struct *p,
1503 unsigned long shared, unsigned long private)
1505 unsigned int period_slot;
1509 unsigned long remote = p->numa_faults_locality[0];
1510 unsigned long local = p->numa_faults_locality[1];
1513 * If there were no record hinting faults then either the task is
1514 * completely idle or all activity is areas that are not of interest
1515 * to automatic numa balancing. Scan slower
1517 if (local + shared == 0) {
1518 p->numa_scan_period = min(p->numa_scan_period_max,
1519 p->numa_scan_period << 1);
1521 p->mm->numa_next_scan = jiffies +
1522 msecs_to_jiffies(p->numa_scan_period);
1528 * Prepare to scale scan period relative to the current period.
1529 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1530 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1531 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1533 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1534 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1535 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1536 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1539 diff = slot * period_slot;
1541 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1544 * Scale scan rate increases based on sharing. There is an
1545 * inverse relationship between the degree of sharing and
1546 * the adjustment made to the scanning period. Broadly
1547 * speaking the intent is that there is little point
1548 * scanning faster if shared accesses dominate as it may
1549 * simply bounce migrations uselessly
1551 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1552 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1555 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1556 task_scan_min(p), task_scan_max(p));
1557 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1561 * Get the fraction of time the task has been running since the last
1562 * NUMA placement cycle. The scheduler keeps similar statistics, but
1563 * decays those on a 32ms period, which is orders of magnitude off
1564 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1565 * stats only if the task is so new there are no NUMA statistics yet.
1567 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1569 u64 runtime, delta, now;
1570 /* Use the start of this time slice to avoid calculations. */
1571 now = p->se.exec_start;
1572 runtime = p->se.sum_exec_runtime;
1574 if (p->last_task_numa_placement) {
1575 delta = runtime - p->last_sum_exec_runtime;
1576 *period = now - p->last_task_numa_placement;
1578 delta = p->se.avg.runnable_avg_sum;
1579 *period = p->se.avg.runnable_avg_period;
1582 p->last_sum_exec_runtime = runtime;
1583 p->last_task_numa_placement = now;
1588 static void task_numa_placement(struct task_struct *p)
1590 int seq, nid, max_nid = -1, max_group_nid = -1;
1591 unsigned long max_faults = 0, max_group_faults = 0;
1592 unsigned long fault_types[2] = { 0, 0 };
1593 unsigned long total_faults;
1594 u64 runtime, period;
1595 spinlock_t *group_lock = NULL;
1597 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1598 if (p->numa_scan_seq == seq)
1600 p->numa_scan_seq = seq;
1601 p->numa_scan_period_max = task_scan_max(p);
1603 total_faults = p->numa_faults_locality[0] +
1604 p->numa_faults_locality[1];
1605 runtime = numa_get_avg_runtime(p, &period);
1607 /* If the task is part of a group prevent parallel updates to group stats */
1608 if (p->numa_group) {
1609 group_lock = &p->numa_group->lock;
1610 spin_lock_irq(group_lock);
1613 /* Find the node with the highest number of faults */
1614 for_each_online_node(nid) {
1615 unsigned long faults = 0, group_faults = 0;
1618 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1619 long diff, f_diff, f_weight;
1621 i = task_faults_idx(nid, priv);
1623 /* Decay existing window, copy faults since last scan */
1624 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1625 fault_types[priv] += p->numa_faults_buffer_memory[i];
1626 p->numa_faults_buffer_memory[i] = 0;
1629 * Normalize the faults_from, so all tasks in a group
1630 * count according to CPU use, instead of by the raw
1631 * number of faults. Tasks with little runtime have
1632 * little over-all impact on throughput, and thus their
1633 * faults are less important.
1635 f_weight = div64_u64(runtime << 16, period + 1);
1636 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1638 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1639 p->numa_faults_buffer_cpu[i] = 0;
1641 p->numa_faults_memory[i] += diff;
1642 p->numa_faults_cpu[i] += f_diff;
1643 faults += p->numa_faults_memory[i];
1644 p->total_numa_faults += diff;
1645 if (p->numa_group) {
1646 /* safe because we can only change our own group */
1647 p->numa_group->faults[i] += diff;
1648 p->numa_group->faults_cpu[i] += f_diff;
1649 p->numa_group->total_faults += diff;
1650 group_faults += p->numa_group->faults[i];
1654 if (faults > max_faults) {
1655 max_faults = faults;
1659 if (group_faults > max_group_faults) {
1660 max_group_faults = group_faults;
1661 max_group_nid = nid;
1665 update_task_scan_period(p, fault_types[0], fault_types[1]);
1667 if (p->numa_group) {
1668 update_numa_active_node_mask(p->numa_group);
1669 spin_unlock_irq(group_lock);
1670 max_nid = max_group_nid;
1674 /* Set the new preferred node */
1675 if (max_nid != p->numa_preferred_nid)
1676 sched_setnuma(p, max_nid);
1678 if (task_node(p) != p->numa_preferred_nid)
1679 numa_migrate_preferred(p);
1683 static inline int get_numa_group(struct numa_group *grp)
1685 return atomic_inc_not_zero(&grp->refcount);
1688 static inline void put_numa_group(struct numa_group *grp)
1690 if (atomic_dec_and_test(&grp->refcount))
1691 kfree_rcu(grp, rcu);
1694 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1697 struct numa_group *grp, *my_grp;
1698 struct task_struct *tsk;
1700 int cpu = cpupid_to_cpu(cpupid);
1703 if (unlikely(!p->numa_group)) {
1704 unsigned int size = sizeof(struct numa_group) +
1705 4*nr_node_ids*sizeof(unsigned long);
1707 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1711 atomic_set(&grp->refcount, 1);
1712 spin_lock_init(&grp->lock);
1713 INIT_LIST_HEAD(&grp->task_list);
1715 /* Second half of the array tracks nids where faults happen */
1716 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1719 node_set(task_node(current), grp->active_nodes);
1721 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1722 grp->faults[i] = p->numa_faults_memory[i];
1724 grp->total_faults = p->total_numa_faults;
1726 list_add(&p->numa_entry, &grp->task_list);
1728 rcu_assign_pointer(p->numa_group, grp);
1732 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1734 if (!cpupid_match_pid(tsk, cpupid))
1737 grp = rcu_dereference(tsk->numa_group);
1741 my_grp = p->numa_group;
1746 * Only join the other group if its bigger; if we're the bigger group,
1747 * the other task will join us.
1749 if (my_grp->nr_tasks > grp->nr_tasks)
1753 * Tie-break on the grp address.
1755 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1758 /* Always join threads in the same process. */
1759 if (tsk->mm == current->mm)
1762 /* Simple filter to avoid false positives due to PID collisions */
1763 if (flags & TNF_SHARED)
1766 /* Update priv based on whether false sharing was detected */
1769 if (join && !get_numa_group(grp))
1777 BUG_ON(irqs_disabled());
1778 double_lock_irq(&my_grp->lock, &grp->lock);
1780 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1781 my_grp->faults[i] -= p->numa_faults_memory[i];
1782 grp->faults[i] += p->numa_faults_memory[i];
1784 my_grp->total_faults -= p->total_numa_faults;
1785 grp->total_faults += p->total_numa_faults;
1787 list_move(&p->numa_entry, &grp->task_list);
1791 spin_unlock(&my_grp->lock);
1792 spin_unlock_irq(&grp->lock);
1794 rcu_assign_pointer(p->numa_group, grp);
1796 put_numa_group(my_grp);
1804 void task_numa_free(struct task_struct *p)
1806 struct numa_group *grp = p->numa_group;
1807 void *numa_faults = p->numa_faults_memory;
1808 unsigned long flags;
1812 spin_lock_irqsave(&grp->lock, flags);
1813 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1814 grp->faults[i] -= p->numa_faults_memory[i];
1815 grp->total_faults -= p->total_numa_faults;
1817 list_del(&p->numa_entry);
1819 spin_unlock_irqrestore(&grp->lock, flags);
1820 RCU_INIT_POINTER(p->numa_group, NULL);
1821 put_numa_group(grp);
1824 p->numa_faults_memory = NULL;
1825 p->numa_faults_buffer_memory = NULL;
1826 p->numa_faults_cpu= NULL;
1827 p->numa_faults_buffer_cpu = NULL;
1832 * Got a PROT_NONE fault for a page on @node.
1834 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1836 struct task_struct *p = current;
1837 bool migrated = flags & TNF_MIGRATED;
1838 int cpu_node = task_node(current);
1839 int local = !!(flags & TNF_FAULT_LOCAL);
1842 if (!numabalancing_enabled)
1845 /* for example, ksmd faulting in a user's mm */
1849 /* Allocate buffer to track faults on a per-node basis */
1850 if (unlikely(!p->numa_faults_memory)) {
1851 int size = sizeof(*p->numa_faults_memory) *
1852 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1854 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1855 if (!p->numa_faults_memory)
1858 BUG_ON(p->numa_faults_buffer_memory);
1860 * The averaged statistics, shared & private, memory & cpu,
1861 * occupy the first half of the array. The second half of the
1862 * array is for current counters, which are averaged into the
1863 * first set by task_numa_placement.
1865 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1866 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1867 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1868 p->total_numa_faults = 0;
1869 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1873 * First accesses are treated as private, otherwise consider accesses
1874 * to be private if the accessing pid has not changed
1876 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1879 priv = cpupid_match_pid(p, last_cpupid);
1880 if (!priv && !(flags & TNF_NO_GROUP))
1881 task_numa_group(p, last_cpupid, flags, &priv);
1885 * If a workload spans multiple NUMA nodes, a shared fault that
1886 * occurs wholly within the set of nodes that the workload is
1887 * actively using should be counted as local. This allows the
1888 * scan rate to slow down when a workload has settled down.
1890 if (!priv && !local && p->numa_group &&
1891 node_isset(cpu_node, p->numa_group->active_nodes) &&
1892 node_isset(mem_node, p->numa_group->active_nodes))
1895 task_numa_placement(p);
1898 * Retry task to preferred node migration periodically, in case it
1899 * case it previously failed, or the scheduler moved us.
1901 if (time_after(jiffies, p->numa_migrate_retry))
1902 numa_migrate_preferred(p);
1905 p->numa_pages_migrated += pages;
1907 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1908 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1909 p->numa_faults_locality[local] += pages;
1912 static void reset_ptenuma_scan(struct task_struct *p)
1914 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1915 p->mm->numa_scan_offset = 0;
1919 * The expensive part of numa migration is done from task_work context.
1920 * Triggered from task_tick_numa().
1922 void task_numa_work(struct callback_head *work)
1924 unsigned long migrate, next_scan, now = jiffies;
1925 struct task_struct *p = current;
1926 struct mm_struct *mm = p->mm;
1927 struct vm_area_struct *vma;
1928 unsigned long start, end;
1929 unsigned long nr_pte_updates = 0;
1932 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1934 work->next = work; /* protect against double add */
1936 * Who cares about NUMA placement when they're dying.
1938 * NOTE: make sure not to dereference p->mm before this check,
1939 * exit_task_work() happens _after_ exit_mm() so we could be called
1940 * without p->mm even though we still had it when we enqueued this
1943 if (p->flags & PF_EXITING)
1946 if (!mm->numa_next_scan) {
1947 mm->numa_next_scan = now +
1948 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1952 * Enforce maximal scan/migration frequency..
1954 migrate = mm->numa_next_scan;
1955 if (time_before(now, migrate))
1958 if (p->numa_scan_period == 0) {
1959 p->numa_scan_period_max = task_scan_max(p);
1960 p->numa_scan_period = task_scan_min(p);
1963 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1964 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1968 * Delay this task enough that another task of this mm will likely win
1969 * the next time around.
1971 p->node_stamp += 2 * TICK_NSEC;
1973 start = mm->numa_scan_offset;
1974 pages = sysctl_numa_balancing_scan_size;
1975 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1979 down_read(&mm->mmap_sem);
1980 vma = find_vma(mm, start);
1982 reset_ptenuma_scan(p);
1986 for (; vma; vma = vma->vm_next) {
1987 if (!vma_migratable(vma) || !vma_policy_mof(vma))
1991 * Shared library pages mapped by multiple processes are not
1992 * migrated as it is expected they are cache replicated. Avoid
1993 * hinting faults in read-only file-backed mappings or the vdso
1994 * as migrating the pages will be of marginal benefit.
1997 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2001 * Skip inaccessible VMAs to avoid any confusion between
2002 * PROT_NONE and NUMA hinting ptes
2004 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2008 start = max(start, vma->vm_start);
2009 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2010 end = min(end, vma->vm_end);
2011 nr_pte_updates += change_prot_numa(vma, start, end);
2014 * Scan sysctl_numa_balancing_scan_size but ensure that
2015 * at least one PTE is updated so that unused virtual
2016 * address space is quickly skipped.
2019 pages -= (end - start) >> PAGE_SHIFT;
2026 } while (end != vma->vm_end);
2031 * It is possible to reach the end of the VMA list but the last few
2032 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2033 * would find the !migratable VMA on the next scan but not reset the
2034 * scanner to the start so check it now.
2037 mm->numa_scan_offset = start;
2039 reset_ptenuma_scan(p);
2040 up_read(&mm->mmap_sem);
2044 * Drive the periodic memory faults..
2046 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2048 struct callback_head *work = &curr->numa_work;
2052 * We don't care about NUMA placement if we don't have memory.
2054 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2058 * Using runtime rather than walltime has the dual advantage that
2059 * we (mostly) drive the selection from busy threads and that the
2060 * task needs to have done some actual work before we bother with
2063 now = curr->se.sum_exec_runtime;
2064 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2066 if (now - curr->node_stamp > period) {
2067 if (!curr->node_stamp)
2068 curr->numa_scan_period = task_scan_min(curr);
2069 curr->node_stamp += period;
2071 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2072 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2073 task_work_add(curr, work, true);
2078 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2082 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2086 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2089 #endif /* CONFIG_NUMA_BALANCING */
2092 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2094 update_load_add(&cfs_rq->load, se->load.weight);
2095 if (!parent_entity(se))
2096 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2098 if (entity_is_task(se)) {
2099 struct rq *rq = rq_of(cfs_rq);
2101 account_numa_enqueue(rq, task_of(se));
2102 list_add(&se->group_node, &rq->cfs_tasks);
2105 cfs_rq->nr_running++;
2109 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2111 update_load_sub(&cfs_rq->load, se->load.weight);
2112 if (!parent_entity(se))
2113 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2114 if (entity_is_task(se)) {
2115 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2116 list_del_init(&se->group_node);
2118 cfs_rq->nr_running--;
2121 #ifdef CONFIG_FAIR_GROUP_SCHED
2123 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2128 * Use this CPU's actual weight instead of the last load_contribution
2129 * to gain a more accurate current total weight. See
2130 * update_cfs_rq_load_contribution().
2132 tg_weight = atomic_long_read(&tg->load_avg);
2133 tg_weight -= cfs_rq->tg_load_contrib;
2134 tg_weight += cfs_rq->load.weight;
2139 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2141 long tg_weight, load, shares;
2143 tg_weight = calc_tg_weight(tg, cfs_rq);
2144 load = cfs_rq->load.weight;
2146 shares = (tg->shares * load);
2148 shares /= tg_weight;
2150 if (shares < MIN_SHARES)
2151 shares = MIN_SHARES;
2152 if (shares > tg->shares)
2153 shares = tg->shares;
2157 # else /* CONFIG_SMP */
2158 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2162 # endif /* CONFIG_SMP */
2163 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2164 unsigned long weight)
2167 /* commit outstanding execution time */
2168 if (cfs_rq->curr == se)
2169 update_curr(cfs_rq);
2170 account_entity_dequeue(cfs_rq, se);
2173 update_load_set(&se->load, weight);
2176 account_entity_enqueue(cfs_rq, se);
2179 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2181 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2183 struct task_group *tg;
2184 struct sched_entity *se;
2188 se = tg->se[cpu_of(rq_of(cfs_rq))];
2189 if (!se || throttled_hierarchy(cfs_rq))
2192 if (likely(se->load.weight == tg->shares))
2195 shares = calc_cfs_shares(cfs_rq, tg);
2197 reweight_entity(cfs_rq_of(se), se, shares);
2199 #else /* CONFIG_FAIR_GROUP_SCHED */
2200 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2203 #endif /* CONFIG_FAIR_GROUP_SCHED */
2207 * We choose a half-life close to 1 scheduling period.
2208 * Note: The tables below are dependent on this value.
2210 #define LOAD_AVG_PERIOD 32
2211 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2212 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2214 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2215 static const u32 runnable_avg_yN_inv[] = {
2216 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2217 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2218 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2219 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2220 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2221 0x85aac367, 0x82cd8698,
2225 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2226 * over-estimates when re-combining.
2228 static const u32 runnable_avg_yN_sum[] = {
2229 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2230 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2231 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2236 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2238 static __always_inline u64 decay_load(u64 val, u64 n)
2240 unsigned int local_n;
2244 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2247 /* after bounds checking we can collapse to 32-bit */
2251 * As y^PERIOD = 1/2, we can combine
2252 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2253 * With a look-up table which covers y^n (n<PERIOD)
2255 * To achieve constant time decay_load.
2257 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2258 val >>= local_n / LOAD_AVG_PERIOD;
2259 local_n %= LOAD_AVG_PERIOD;
2262 val *= runnable_avg_yN_inv[local_n];
2263 /* We don't use SRR here since we always want to round down. */
2268 * For updates fully spanning n periods, the contribution to runnable
2269 * average will be: \Sum 1024*y^n
2271 * We can compute this reasonably efficiently by combining:
2272 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2274 static u32 __compute_runnable_contrib(u64 n)
2278 if (likely(n <= LOAD_AVG_PERIOD))
2279 return runnable_avg_yN_sum[n];
2280 else if (unlikely(n >= LOAD_AVG_MAX_N))
2281 return LOAD_AVG_MAX;
2283 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2285 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2286 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2288 n -= LOAD_AVG_PERIOD;
2289 } while (n > LOAD_AVG_PERIOD);
2291 contrib = decay_load(contrib, n);
2292 return contrib + runnable_avg_yN_sum[n];
2296 * We can represent the historical contribution to runnable average as the
2297 * coefficients of a geometric series. To do this we sub-divide our runnable
2298 * history into segments of approximately 1ms (1024us); label the segment that
2299 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2301 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2303 * (now) (~1ms ago) (~2ms ago)
2305 * Let u_i denote the fraction of p_i that the entity was runnable.
2307 * We then designate the fractions u_i as our co-efficients, yielding the
2308 * following representation of historical load:
2309 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2311 * We choose y based on the with of a reasonably scheduling period, fixing:
2314 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2315 * approximately half as much as the contribution to load within the last ms
2318 * When a period "rolls over" and we have new u_0`, multiplying the previous
2319 * sum again by y is sufficient to update:
2320 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2321 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2323 static __always_inline int __update_entity_runnable_avg(u64 now,
2324 struct sched_avg *sa,
2328 u32 runnable_contrib;
2329 int delta_w, decayed = 0;
2331 delta = now - sa->last_runnable_update;
2333 * This should only happen when time goes backwards, which it
2334 * unfortunately does during sched clock init when we swap over to TSC.
2336 if ((s64)delta < 0) {
2337 sa->last_runnable_update = now;
2342 * Use 1024ns as the unit of measurement since it's a reasonable
2343 * approximation of 1us and fast to compute.
2348 sa->last_runnable_update = now;
2350 /* delta_w is the amount already accumulated against our next period */
2351 delta_w = sa->runnable_avg_period % 1024;
2352 if (delta + delta_w >= 1024) {
2353 /* period roll-over */
2357 * Now that we know we're crossing a period boundary, figure
2358 * out how much from delta we need to complete the current
2359 * period and accrue it.
2361 delta_w = 1024 - delta_w;
2363 sa->runnable_avg_sum += delta_w;
2364 sa->runnable_avg_period += delta_w;
2368 /* Figure out how many additional periods this update spans */
2369 periods = delta / 1024;
2372 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2374 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2377 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2378 runnable_contrib = __compute_runnable_contrib(periods);
2380 sa->runnable_avg_sum += runnable_contrib;
2381 sa->runnable_avg_period += runnable_contrib;
2384 /* Remainder of delta accrued against u_0` */
2386 sa->runnable_avg_sum += delta;
2387 sa->runnable_avg_period += delta;
2392 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2393 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2395 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2396 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2398 decays -= se->avg.decay_count;
2402 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2403 se->avg.decay_count = 0;
2408 #ifdef CONFIG_FAIR_GROUP_SCHED
2409 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2412 struct task_group *tg = cfs_rq->tg;
2415 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2416 tg_contrib -= cfs_rq->tg_load_contrib;
2421 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2422 atomic_long_add(tg_contrib, &tg->load_avg);
2423 cfs_rq->tg_load_contrib += tg_contrib;
2428 * Aggregate cfs_rq runnable averages into an equivalent task_group
2429 * representation for computing load contributions.
2431 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2432 struct cfs_rq *cfs_rq)
2434 struct task_group *tg = cfs_rq->tg;
2437 /* The fraction of a cpu used by this cfs_rq */
2438 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2439 sa->runnable_avg_period + 1);
2440 contrib -= cfs_rq->tg_runnable_contrib;
2442 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2443 atomic_add(contrib, &tg->runnable_avg);
2444 cfs_rq->tg_runnable_contrib += contrib;
2448 static inline void __update_group_entity_contrib(struct sched_entity *se)
2450 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2451 struct task_group *tg = cfs_rq->tg;
2456 contrib = cfs_rq->tg_load_contrib * tg->shares;
2457 se->avg.load_avg_contrib = div_u64(contrib,
2458 atomic_long_read(&tg->load_avg) + 1);
2461 * For group entities we need to compute a correction term in the case
2462 * that they are consuming <1 cpu so that we would contribute the same
2463 * load as a task of equal weight.
2465 * Explicitly co-ordinating this measurement would be expensive, but
2466 * fortunately the sum of each cpus contribution forms a usable
2467 * lower-bound on the true value.
2469 * Consider the aggregate of 2 contributions. Either they are disjoint
2470 * (and the sum represents true value) or they are disjoint and we are
2471 * understating by the aggregate of their overlap.
2473 * Extending this to N cpus, for a given overlap, the maximum amount we
2474 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2475 * cpus that overlap for this interval and w_i is the interval width.
2477 * On a small machine; the first term is well-bounded which bounds the
2478 * total error since w_i is a subset of the period. Whereas on a
2479 * larger machine, while this first term can be larger, if w_i is the
2480 * of consequential size guaranteed to see n_i*w_i quickly converge to
2481 * our upper bound of 1-cpu.
2483 runnable_avg = atomic_read(&tg->runnable_avg);
2484 if (runnable_avg < NICE_0_LOAD) {
2485 se->avg.load_avg_contrib *= runnable_avg;
2486 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2490 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2492 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2493 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2495 #else /* CONFIG_FAIR_GROUP_SCHED */
2496 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2497 int force_update) {}
2498 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2499 struct cfs_rq *cfs_rq) {}
2500 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2501 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2502 #endif /* CONFIG_FAIR_GROUP_SCHED */
2504 static inline void __update_task_entity_contrib(struct sched_entity *se)
2508 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2509 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2510 contrib /= (se->avg.runnable_avg_period + 1);
2511 se->avg.load_avg_contrib = scale_load(contrib);
2514 /* Compute the current contribution to load_avg by se, return any delta */
2515 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2517 long old_contrib = se->avg.load_avg_contrib;
2519 if (entity_is_task(se)) {
2520 __update_task_entity_contrib(se);
2522 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2523 __update_group_entity_contrib(se);
2526 return se->avg.load_avg_contrib - old_contrib;
2529 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2532 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2533 cfs_rq->blocked_load_avg -= load_contrib;
2535 cfs_rq->blocked_load_avg = 0;
2538 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2540 /* Update a sched_entity's runnable average */
2541 static inline void update_entity_load_avg(struct sched_entity *se,
2544 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2549 * For a group entity we need to use their owned cfs_rq_clock_task() in
2550 * case they are the parent of a throttled hierarchy.
2552 if (entity_is_task(se))
2553 now = cfs_rq_clock_task(cfs_rq);
2555 now = cfs_rq_clock_task(group_cfs_rq(se));
2557 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2560 contrib_delta = __update_entity_load_avg_contrib(se);
2566 cfs_rq->runnable_load_avg += contrib_delta;
2568 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2572 * Decay the load contributed by all blocked children and account this so that
2573 * their contribution may appropriately discounted when they wake up.
2575 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2577 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2580 decays = now - cfs_rq->last_decay;
2581 if (!decays && !force_update)
2584 if (atomic_long_read(&cfs_rq->removed_load)) {
2585 unsigned long removed_load;
2586 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2587 subtract_blocked_load_contrib(cfs_rq, removed_load);
2591 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2593 atomic64_add(decays, &cfs_rq->decay_counter);
2594 cfs_rq->last_decay = now;
2597 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2600 /* Add the load generated by se into cfs_rq's child load-average */
2601 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2602 struct sched_entity *se,
2606 * We track migrations using entity decay_count <= 0, on a wake-up
2607 * migration we use a negative decay count to track the remote decays
2608 * accumulated while sleeping.
2610 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2611 * are seen by enqueue_entity_load_avg() as a migration with an already
2612 * constructed load_avg_contrib.
2614 if (unlikely(se->avg.decay_count <= 0)) {
2615 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2616 if (se->avg.decay_count) {
2618 * In a wake-up migration we have to approximate the
2619 * time sleeping. This is because we can't synchronize
2620 * clock_task between the two cpus, and it is not
2621 * guaranteed to be read-safe. Instead, we can
2622 * approximate this using our carried decays, which are
2623 * explicitly atomically readable.
2625 se->avg.last_runnable_update -= (-se->avg.decay_count)
2627 update_entity_load_avg(se, 0);
2628 /* Indicate that we're now synchronized and on-rq */
2629 se->avg.decay_count = 0;
2633 __synchronize_entity_decay(se);
2636 /* migrated tasks did not contribute to our blocked load */
2638 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2639 update_entity_load_avg(se, 0);
2642 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2643 /* we force update consideration on load-balancer moves */
2644 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2648 * Remove se's load from this cfs_rq child load-average, if the entity is
2649 * transitioning to a blocked state we track its projected decay using
2652 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2653 struct sched_entity *se,
2656 update_entity_load_avg(se, 1);
2657 /* we force update consideration on load-balancer moves */
2658 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2660 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2662 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2663 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2664 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2668 * Update the rq's load with the elapsed running time before entering
2669 * idle. if the last scheduled task is not a CFS task, idle_enter will
2670 * be the only way to update the runnable statistic.
2672 void idle_enter_fair(struct rq *this_rq)
2674 update_rq_runnable_avg(this_rq, 1);
2678 * Update the rq's load with the elapsed idle time before a task is
2679 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2680 * be the only way to update the runnable statistic.
2682 void idle_exit_fair(struct rq *this_rq)
2684 update_rq_runnable_avg(this_rq, 0);
2687 static int idle_balance(struct rq *this_rq);
2689 #else /* CONFIG_SMP */
2691 static inline void update_entity_load_avg(struct sched_entity *se,
2692 int update_cfs_rq) {}
2693 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2694 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2695 struct sched_entity *se,
2697 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2698 struct sched_entity *se,
2700 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2701 int force_update) {}
2703 static inline int idle_balance(struct rq *rq)
2708 #endif /* CONFIG_SMP */
2710 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2712 #ifdef CONFIG_SCHEDSTATS
2713 struct task_struct *tsk = NULL;
2715 if (entity_is_task(se))
2718 if (se->statistics.sleep_start) {
2719 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2724 if (unlikely(delta > se->statistics.sleep_max))
2725 se->statistics.sleep_max = delta;
2727 se->statistics.sleep_start = 0;
2728 se->statistics.sum_sleep_runtime += delta;
2731 account_scheduler_latency(tsk, delta >> 10, 1);
2732 trace_sched_stat_sleep(tsk, delta);
2735 if (se->statistics.block_start) {
2736 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2741 if (unlikely(delta > se->statistics.block_max))
2742 se->statistics.block_max = delta;
2744 se->statistics.block_start = 0;
2745 se->statistics.sum_sleep_runtime += delta;
2748 if (tsk->in_iowait) {
2749 se->statistics.iowait_sum += delta;
2750 se->statistics.iowait_count++;
2751 trace_sched_stat_iowait(tsk, delta);
2754 trace_sched_stat_blocked(tsk, delta);
2757 * Blocking time is in units of nanosecs, so shift by
2758 * 20 to get a milliseconds-range estimation of the
2759 * amount of time that the task spent sleeping:
2761 if (unlikely(prof_on == SLEEP_PROFILING)) {
2762 profile_hits(SLEEP_PROFILING,
2763 (void *)get_wchan(tsk),
2766 account_scheduler_latency(tsk, delta >> 10, 0);
2772 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2774 #ifdef CONFIG_SCHED_DEBUG
2775 s64 d = se->vruntime - cfs_rq->min_vruntime;
2780 if (d > 3*sysctl_sched_latency)
2781 schedstat_inc(cfs_rq, nr_spread_over);
2786 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2788 u64 vruntime = cfs_rq->min_vruntime;
2791 * The 'current' period is already promised to the current tasks,
2792 * however the extra weight of the new task will slow them down a
2793 * little, place the new task so that it fits in the slot that
2794 * stays open at the end.
2796 if (initial && sched_feat(START_DEBIT))
2797 vruntime += sched_vslice(cfs_rq, se);
2799 /* sleeps up to a single latency don't count. */
2801 unsigned long thresh = sysctl_sched_latency;
2804 * Halve their sleep time's effect, to allow
2805 * for a gentler effect of sleepers:
2807 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2813 /* ensure we never gain time by being placed backwards. */
2814 se->vruntime = max_vruntime(se->vruntime, vruntime);
2817 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2820 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2823 * Update the normalized vruntime before updating min_vruntime
2824 * through calling update_curr().
2826 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2827 se->vruntime += cfs_rq->min_vruntime;
2830 * Update run-time statistics of the 'current'.
2832 update_curr(cfs_rq);
2833 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2834 account_entity_enqueue(cfs_rq, se);
2835 update_cfs_shares(cfs_rq);
2837 if (flags & ENQUEUE_WAKEUP) {
2838 place_entity(cfs_rq, se, 0);
2839 enqueue_sleeper(cfs_rq, se);
2842 update_stats_enqueue(cfs_rq, se);
2843 check_spread(cfs_rq, se);
2844 if (se != cfs_rq->curr)
2845 __enqueue_entity(cfs_rq, se);
2848 if (cfs_rq->nr_running == 1) {
2849 list_add_leaf_cfs_rq(cfs_rq);
2850 check_enqueue_throttle(cfs_rq);
2854 static void __clear_buddies_last(struct sched_entity *se)
2856 for_each_sched_entity(se) {
2857 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2858 if (cfs_rq->last != se)
2861 cfs_rq->last = NULL;
2865 static void __clear_buddies_next(struct sched_entity *se)
2867 for_each_sched_entity(se) {
2868 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2869 if (cfs_rq->next != se)
2872 cfs_rq->next = NULL;
2876 static void __clear_buddies_skip(struct sched_entity *se)
2878 for_each_sched_entity(se) {
2879 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2880 if (cfs_rq->skip != se)
2883 cfs_rq->skip = NULL;
2887 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2889 if (cfs_rq->last == se)
2890 __clear_buddies_last(se);
2892 if (cfs_rq->next == se)
2893 __clear_buddies_next(se);
2895 if (cfs_rq->skip == se)
2896 __clear_buddies_skip(se);
2899 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2902 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2905 * Update run-time statistics of the 'current'.
2907 update_curr(cfs_rq);
2908 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2910 update_stats_dequeue(cfs_rq, se);
2911 if (flags & DEQUEUE_SLEEP) {
2912 #ifdef CONFIG_SCHEDSTATS
2913 if (entity_is_task(se)) {
2914 struct task_struct *tsk = task_of(se);
2916 if (tsk->state & TASK_INTERRUPTIBLE)
2917 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2918 if (tsk->state & TASK_UNINTERRUPTIBLE)
2919 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2924 clear_buddies(cfs_rq, se);
2926 if (se != cfs_rq->curr)
2927 __dequeue_entity(cfs_rq, se);
2929 account_entity_dequeue(cfs_rq, se);
2932 * Normalize the entity after updating the min_vruntime because the
2933 * update can refer to the ->curr item and we need to reflect this
2934 * movement in our normalized position.
2936 if (!(flags & DEQUEUE_SLEEP))
2937 se->vruntime -= cfs_rq->min_vruntime;
2939 /* return excess runtime on last dequeue */
2940 return_cfs_rq_runtime(cfs_rq);
2942 update_min_vruntime(cfs_rq);
2943 update_cfs_shares(cfs_rq);
2947 * Preempt the current task with a newly woken task if needed:
2950 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2952 unsigned long ideal_runtime, delta_exec;
2953 struct sched_entity *se;
2956 ideal_runtime = sched_slice(cfs_rq, curr);
2957 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2958 if (delta_exec > ideal_runtime) {
2959 resched_curr(rq_of(cfs_rq));
2961 * The current task ran long enough, ensure it doesn't get
2962 * re-elected due to buddy favours.
2964 clear_buddies(cfs_rq, curr);
2969 * Ensure that a task that missed wakeup preemption by a
2970 * narrow margin doesn't have to wait for a full slice.
2971 * This also mitigates buddy induced latencies under load.
2973 if (delta_exec < sysctl_sched_min_granularity)
2976 se = __pick_first_entity(cfs_rq);
2977 delta = curr->vruntime - se->vruntime;
2982 if (delta > ideal_runtime)
2983 resched_curr(rq_of(cfs_rq));
2987 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2989 /* 'current' is not kept within the tree. */
2992 * Any task has to be enqueued before it get to execute on
2993 * a CPU. So account for the time it spent waiting on the
2996 update_stats_wait_end(cfs_rq, se);
2997 __dequeue_entity(cfs_rq, se);
3000 update_stats_curr_start(cfs_rq, se);
3002 #ifdef CONFIG_SCHEDSTATS
3004 * Track our maximum slice length, if the CPU's load is at
3005 * least twice that of our own weight (i.e. dont track it
3006 * when there are only lesser-weight tasks around):
3008 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3009 se->statistics.slice_max = max(se->statistics.slice_max,
3010 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3013 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3017 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3020 * Pick the next process, keeping these things in mind, in this order:
3021 * 1) keep things fair between processes/task groups
3022 * 2) pick the "next" process, since someone really wants that to run
3023 * 3) pick the "last" process, for cache locality
3024 * 4) do not run the "skip" process, if something else is available
3026 static struct sched_entity *
3027 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3029 struct sched_entity *left = __pick_first_entity(cfs_rq);
3030 struct sched_entity *se;
3033 * If curr is set we have to see if its left of the leftmost entity
3034 * still in the tree, provided there was anything in the tree at all.
3036 if (!left || (curr && entity_before(curr, left)))
3039 se = left; /* ideally we run the leftmost entity */
3042 * Avoid running the skip buddy, if running something else can
3043 * be done without getting too unfair.
3045 if (cfs_rq->skip == se) {
3046 struct sched_entity *second;
3049 second = __pick_first_entity(cfs_rq);
3051 second = __pick_next_entity(se);
3052 if (!second || (curr && entity_before(curr, second)))
3056 if (second && wakeup_preempt_entity(second, left) < 1)
3061 * Prefer last buddy, try to return the CPU to a preempted task.
3063 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3067 * Someone really wants this to run. If it's not unfair, run it.
3069 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3072 clear_buddies(cfs_rq, se);
3077 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3079 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3082 * If still on the runqueue then deactivate_task()
3083 * was not called and update_curr() has to be done:
3086 update_curr(cfs_rq);
3088 /* throttle cfs_rqs exceeding runtime */
3089 check_cfs_rq_runtime(cfs_rq);
3091 check_spread(cfs_rq, prev);
3093 update_stats_wait_start(cfs_rq, prev);
3094 /* Put 'current' back into the tree. */
3095 __enqueue_entity(cfs_rq, prev);
3096 /* in !on_rq case, update occurred at dequeue */
3097 update_entity_load_avg(prev, 1);
3099 cfs_rq->curr = NULL;
3103 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3106 * Update run-time statistics of the 'current'.
3108 update_curr(cfs_rq);
3111 * Ensure that runnable average is periodically updated.
3113 update_entity_load_avg(curr, 1);
3114 update_cfs_rq_blocked_load(cfs_rq, 1);
3115 update_cfs_shares(cfs_rq);
3117 #ifdef CONFIG_SCHED_HRTICK
3119 * queued ticks are scheduled to match the slice, so don't bother
3120 * validating it and just reschedule.
3123 resched_curr(rq_of(cfs_rq));
3127 * don't let the period tick interfere with the hrtick preemption
3129 if (!sched_feat(DOUBLE_TICK) &&
3130 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3134 if (cfs_rq->nr_running > 1)
3135 check_preempt_tick(cfs_rq, curr);
3139 /**************************************************
3140 * CFS bandwidth control machinery
3143 #ifdef CONFIG_CFS_BANDWIDTH
3145 #ifdef HAVE_JUMP_LABEL
3146 static struct static_key __cfs_bandwidth_used;
3148 static inline bool cfs_bandwidth_used(void)
3150 return static_key_false(&__cfs_bandwidth_used);
3153 void cfs_bandwidth_usage_inc(void)
3155 static_key_slow_inc(&__cfs_bandwidth_used);
3158 void cfs_bandwidth_usage_dec(void)
3160 static_key_slow_dec(&__cfs_bandwidth_used);
3162 #else /* HAVE_JUMP_LABEL */
3163 static bool cfs_bandwidth_used(void)
3168 void cfs_bandwidth_usage_inc(void) {}
3169 void cfs_bandwidth_usage_dec(void) {}
3170 #endif /* HAVE_JUMP_LABEL */
3173 * default period for cfs group bandwidth.
3174 * default: 0.1s, units: nanoseconds
3176 static inline u64 default_cfs_period(void)
3178 return 100000000ULL;
3181 static inline u64 sched_cfs_bandwidth_slice(void)
3183 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3187 * Replenish runtime according to assigned quota and update expiration time.
3188 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3189 * additional synchronization around rq->lock.
3191 * requires cfs_b->lock
3193 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3197 if (cfs_b->quota == RUNTIME_INF)
3200 now = sched_clock_cpu(smp_processor_id());
3201 cfs_b->runtime = cfs_b->quota;
3202 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3205 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3207 return &tg->cfs_bandwidth;
3210 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3211 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3213 if (unlikely(cfs_rq->throttle_count))
3214 return cfs_rq->throttled_clock_task;
3216 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3219 /* returns 0 on failure to allocate runtime */
3220 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3222 struct task_group *tg = cfs_rq->tg;
3223 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3224 u64 amount = 0, min_amount, expires;
3226 /* note: this is a positive sum as runtime_remaining <= 0 */
3227 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3229 raw_spin_lock(&cfs_b->lock);
3230 if (cfs_b->quota == RUNTIME_INF)
3231 amount = min_amount;
3234 * If the bandwidth pool has become inactive, then at least one
3235 * period must have elapsed since the last consumption.
3236 * Refresh the global state and ensure bandwidth timer becomes
3239 if (!cfs_b->timer_active) {
3240 __refill_cfs_bandwidth_runtime(cfs_b);
3241 __start_cfs_bandwidth(cfs_b, false);
3244 if (cfs_b->runtime > 0) {
3245 amount = min(cfs_b->runtime, min_amount);
3246 cfs_b->runtime -= amount;
3250 expires = cfs_b->runtime_expires;
3251 raw_spin_unlock(&cfs_b->lock);
3253 cfs_rq->runtime_remaining += amount;
3255 * we may have advanced our local expiration to account for allowed
3256 * spread between our sched_clock and the one on which runtime was
3259 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3260 cfs_rq->runtime_expires = expires;
3262 return cfs_rq->runtime_remaining > 0;
3266 * Note: This depends on the synchronization provided by sched_clock and the
3267 * fact that rq->clock snapshots this value.
3269 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3271 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3273 /* if the deadline is ahead of our clock, nothing to do */
3274 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3277 if (cfs_rq->runtime_remaining < 0)
3281 * If the local deadline has passed we have to consider the
3282 * possibility that our sched_clock is 'fast' and the global deadline
3283 * has not truly expired.
3285 * Fortunately we can check determine whether this the case by checking
3286 * whether the global deadline has advanced. It is valid to compare
3287 * cfs_b->runtime_expires without any locks since we only care about
3288 * exact equality, so a partial write will still work.
3291 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3292 /* extend local deadline, drift is bounded above by 2 ticks */
3293 cfs_rq->runtime_expires += TICK_NSEC;
3295 /* global deadline is ahead, expiration has passed */
3296 cfs_rq->runtime_remaining = 0;
3300 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3302 /* dock delta_exec before expiring quota (as it could span periods) */
3303 cfs_rq->runtime_remaining -= delta_exec;
3304 expire_cfs_rq_runtime(cfs_rq);
3306 if (likely(cfs_rq->runtime_remaining > 0))
3310 * if we're unable to extend our runtime we resched so that the active
3311 * hierarchy can be throttled
3313 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3314 resched_curr(rq_of(cfs_rq));
3317 static __always_inline
3318 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3320 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3323 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3326 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3328 return cfs_bandwidth_used() && cfs_rq->throttled;
3331 /* check whether cfs_rq, or any parent, is throttled */
3332 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3334 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3338 * Ensure that neither of the group entities corresponding to src_cpu or
3339 * dest_cpu are members of a throttled hierarchy when performing group
3340 * load-balance operations.
3342 static inline int throttled_lb_pair(struct task_group *tg,
3343 int src_cpu, int dest_cpu)
3345 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3347 src_cfs_rq = tg->cfs_rq[src_cpu];
3348 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3350 return throttled_hierarchy(src_cfs_rq) ||
3351 throttled_hierarchy(dest_cfs_rq);
3354 /* updated child weight may affect parent so we have to do this bottom up */
3355 static int tg_unthrottle_up(struct task_group *tg, void *data)
3357 struct rq *rq = data;
3358 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3360 cfs_rq->throttle_count--;
3362 if (!cfs_rq->throttle_count) {
3363 /* adjust cfs_rq_clock_task() */
3364 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3365 cfs_rq->throttled_clock_task;
3372 static int tg_throttle_down(struct task_group *tg, void *data)
3374 struct rq *rq = data;
3375 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3377 /* group is entering throttled state, stop time */
3378 if (!cfs_rq->throttle_count)
3379 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3380 cfs_rq->throttle_count++;
3385 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3387 struct rq *rq = rq_of(cfs_rq);
3388 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3389 struct sched_entity *se;
3390 long task_delta, dequeue = 1;
3392 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3394 /* freeze hierarchy runnable averages while throttled */
3396 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3399 task_delta = cfs_rq->h_nr_running;
3400 for_each_sched_entity(se) {
3401 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3402 /* throttled entity or throttle-on-deactivate */
3407 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3408 qcfs_rq->h_nr_running -= task_delta;
3410 if (qcfs_rq->load.weight)
3415 sub_nr_running(rq, task_delta);
3417 cfs_rq->throttled = 1;
3418 cfs_rq->throttled_clock = rq_clock(rq);
3419 raw_spin_lock(&cfs_b->lock);
3421 * Add to the _head_ of the list, so that an already-started
3422 * distribute_cfs_runtime will not see us
3424 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3425 if (!cfs_b->timer_active)
3426 __start_cfs_bandwidth(cfs_b, false);
3427 raw_spin_unlock(&cfs_b->lock);
3430 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3432 struct rq *rq = rq_of(cfs_rq);
3433 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3434 struct sched_entity *se;
3438 se = cfs_rq->tg->se[cpu_of(rq)];
3440 cfs_rq->throttled = 0;
3442 update_rq_clock(rq);
3444 raw_spin_lock(&cfs_b->lock);
3445 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3446 list_del_rcu(&cfs_rq->throttled_list);
3447 raw_spin_unlock(&cfs_b->lock);
3449 /* update hierarchical throttle state */
3450 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3452 if (!cfs_rq->load.weight)
3455 task_delta = cfs_rq->h_nr_running;
3456 for_each_sched_entity(se) {
3460 cfs_rq = cfs_rq_of(se);
3462 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3463 cfs_rq->h_nr_running += task_delta;
3465 if (cfs_rq_throttled(cfs_rq))
3470 add_nr_running(rq, task_delta);
3472 /* determine whether we need to wake up potentially idle cpu */
3473 if (rq->curr == rq->idle && rq->cfs.nr_running)
3477 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3478 u64 remaining, u64 expires)
3480 struct cfs_rq *cfs_rq;
3482 u64 starting_runtime = remaining;
3485 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3487 struct rq *rq = rq_of(cfs_rq);
3489 raw_spin_lock(&rq->lock);
3490 if (!cfs_rq_throttled(cfs_rq))
3493 runtime = -cfs_rq->runtime_remaining + 1;
3494 if (runtime > remaining)
3495 runtime = remaining;
3496 remaining -= runtime;
3498 cfs_rq->runtime_remaining += runtime;
3499 cfs_rq->runtime_expires = expires;
3501 /* we check whether we're throttled above */
3502 if (cfs_rq->runtime_remaining > 0)
3503 unthrottle_cfs_rq(cfs_rq);
3506 raw_spin_unlock(&rq->lock);
3513 return starting_runtime - remaining;
3517 * Responsible for refilling a task_group's bandwidth and unthrottling its
3518 * cfs_rqs as appropriate. If there has been no activity within the last
3519 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3520 * used to track this state.
3522 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3524 u64 runtime, runtime_expires;
3527 /* no need to continue the timer with no bandwidth constraint */
3528 if (cfs_b->quota == RUNTIME_INF)
3529 goto out_deactivate;
3531 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3532 cfs_b->nr_periods += overrun;
3535 * idle depends on !throttled (for the case of a large deficit), and if
3536 * we're going inactive then everything else can be deferred
3538 if (cfs_b->idle && !throttled)
3539 goto out_deactivate;
3542 * if we have relooped after returning idle once, we need to update our
3543 * status as actually running, so that other cpus doing
3544 * __start_cfs_bandwidth will stop trying to cancel us.
3546 cfs_b->timer_active = 1;
3548 __refill_cfs_bandwidth_runtime(cfs_b);
3551 /* mark as potentially idle for the upcoming period */
3556 /* account preceding periods in which throttling occurred */
3557 cfs_b->nr_throttled += overrun;
3559 runtime_expires = cfs_b->runtime_expires;
3562 * This check is repeated as we are holding onto the new bandwidth while
3563 * we unthrottle. This can potentially race with an unthrottled group
3564 * trying to acquire new bandwidth from the global pool. This can result
3565 * in us over-using our runtime if it is all used during this loop, but
3566 * only by limited amounts in that extreme case.
3568 while (throttled && cfs_b->runtime > 0) {
3569 runtime = cfs_b->runtime;
3570 raw_spin_unlock(&cfs_b->lock);
3571 /* we can't nest cfs_b->lock while distributing bandwidth */
3572 runtime = distribute_cfs_runtime(cfs_b, runtime,
3574 raw_spin_lock(&cfs_b->lock);
3576 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3578 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3582 * While we are ensured activity in the period following an
3583 * unthrottle, this also covers the case in which the new bandwidth is
3584 * insufficient to cover the existing bandwidth deficit. (Forcing the
3585 * timer to remain active while there are any throttled entities.)
3592 cfs_b->timer_active = 0;
3596 /* a cfs_rq won't donate quota below this amount */
3597 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3598 /* minimum remaining period time to redistribute slack quota */
3599 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3600 /* how long we wait to gather additional slack before distributing */
3601 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3604 * Are we near the end of the current quota period?
3606 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3607 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3608 * migrate_hrtimers, base is never cleared, so we are fine.
3610 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3612 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3615 /* if the call-back is running a quota refresh is already occurring */
3616 if (hrtimer_callback_running(refresh_timer))
3619 /* is a quota refresh about to occur? */
3620 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3621 if (remaining < min_expire)
3627 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3629 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3631 /* if there's a quota refresh soon don't bother with slack */
3632 if (runtime_refresh_within(cfs_b, min_left))
3635 start_bandwidth_timer(&cfs_b->slack_timer,
3636 ns_to_ktime(cfs_bandwidth_slack_period));
3639 /* we know any runtime found here is valid as update_curr() precedes return */
3640 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3642 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3643 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3645 if (slack_runtime <= 0)
3648 raw_spin_lock(&cfs_b->lock);
3649 if (cfs_b->quota != RUNTIME_INF &&
3650 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3651 cfs_b->runtime += slack_runtime;
3653 /* we are under rq->lock, defer unthrottling using a timer */
3654 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3655 !list_empty(&cfs_b->throttled_cfs_rq))
3656 start_cfs_slack_bandwidth(cfs_b);
3658 raw_spin_unlock(&cfs_b->lock);
3660 /* even if it's not valid for return we don't want to try again */
3661 cfs_rq->runtime_remaining -= slack_runtime;
3664 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3666 if (!cfs_bandwidth_used())
3669 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3672 __return_cfs_rq_runtime(cfs_rq);
3676 * This is done with a timer (instead of inline with bandwidth return) since
3677 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3679 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3681 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3684 /* confirm we're still not at a refresh boundary */
3685 raw_spin_lock(&cfs_b->lock);
3686 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3687 raw_spin_unlock(&cfs_b->lock);
3691 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3692 runtime = cfs_b->runtime;
3694 expires = cfs_b->runtime_expires;
3695 raw_spin_unlock(&cfs_b->lock);
3700 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3702 raw_spin_lock(&cfs_b->lock);
3703 if (expires == cfs_b->runtime_expires)
3704 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3705 raw_spin_unlock(&cfs_b->lock);
3709 * When a group wakes up we want to make sure that its quota is not already
3710 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3711 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3713 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3715 if (!cfs_bandwidth_used())
3718 /* an active group must be handled by the update_curr()->put() path */
3719 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3722 /* ensure the group is not already throttled */
3723 if (cfs_rq_throttled(cfs_rq))
3726 /* update runtime allocation */
3727 account_cfs_rq_runtime(cfs_rq, 0);
3728 if (cfs_rq->runtime_remaining <= 0)
3729 throttle_cfs_rq(cfs_rq);
3732 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3733 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3735 if (!cfs_bandwidth_used())
3738 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3742 * it's possible for a throttled entity to be forced into a running
3743 * state (e.g. set_curr_task), in this case we're finished.
3745 if (cfs_rq_throttled(cfs_rq))
3748 throttle_cfs_rq(cfs_rq);
3752 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3754 struct cfs_bandwidth *cfs_b =
3755 container_of(timer, struct cfs_bandwidth, slack_timer);
3756 do_sched_cfs_slack_timer(cfs_b);
3758 return HRTIMER_NORESTART;
3761 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3763 struct cfs_bandwidth *cfs_b =
3764 container_of(timer, struct cfs_bandwidth, period_timer);
3769 raw_spin_lock(&cfs_b->lock);
3771 now = hrtimer_cb_get_time(timer);
3772 overrun = hrtimer_forward(timer, now, cfs_b->period);
3777 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3779 raw_spin_unlock(&cfs_b->lock);
3781 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3784 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3786 raw_spin_lock_init(&cfs_b->lock);
3788 cfs_b->quota = RUNTIME_INF;
3789 cfs_b->period = ns_to_ktime(default_cfs_period());
3791 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3792 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3793 cfs_b->period_timer.function = sched_cfs_period_timer;
3794 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3795 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3798 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3800 cfs_rq->runtime_enabled = 0;
3801 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3804 /* requires cfs_b->lock, may release to reprogram timer */
3805 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force)
3808 * The timer may be active because we're trying to set a new bandwidth
3809 * period or because we're racing with the tear-down path
3810 * (timer_active==0 becomes visible before the hrtimer call-back
3811 * terminates). In either case we ensure that it's re-programmed
3813 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3814 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3815 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3816 raw_spin_unlock(&cfs_b->lock);
3818 raw_spin_lock(&cfs_b->lock);
3819 /* if someone else restarted the timer then we're done */
3820 if (!force && cfs_b->timer_active)
3824 cfs_b->timer_active = 1;
3825 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3828 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3830 hrtimer_cancel(&cfs_b->period_timer);
3831 hrtimer_cancel(&cfs_b->slack_timer);
3834 static void __maybe_unused update_runtime_enabled(struct rq *rq)
3836 struct cfs_rq *cfs_rq;
3838 for_each_leaf_cfs_rq(rq, cfs_rq) {
3839 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
3841 raw_spin_lock(&cfs_b->lock);
3842 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
3843 raw_spin_unlock(&cfs_b->lock);
3847 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3849 struct cfs_rq *cfs_rq;
3851 for_each_leaf_cfs_rq(rq, cfs_rq) {
3852 if (!cfs_rq->runtime_enabled)
3856 * clock_task is not advancing so we just need to make sure
3857 * there's some valid quota amount
3859 cfs_rq->runtime_remaining = 1;
3861 * Offline rq is schedulable till cpu is completely disabled
3862 * in take_cpu_down(), so we prevent new cfs throttling here.
3864 cfs_rq->runtime_enabled = 0;
3866 if (cfs_rq_throttled(cfs_rq))
3867 unthrottle_cfs_rq(cfs_rq);
3871 #else /* CONFIG_CFS_BANDWIDTH */
3872 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3874 return rq_clock_task(rq_of(cfs_rq));
3877 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3878 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3879 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3880 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3882 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3887 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3892 static inline int throttled_lb_pair(struct task_group *tg,
3893 int src_cpu, int dest_cpu)
3898 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3900 #ifdef CONFIG_FAIR_GROUP_SCHED
3901 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3904 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3908 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3909 static inline void update_runtime_enabled(struct rq *rq) {}
3910 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3912 #endif /* CONFIG_CFS_BANDWIDTH */
3914 /**************************************************
3915 * CFS operations on tasks:
3918 #ifdef CONFIG_SCHED_HRTICK
3919 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3921 struct sched_entity *se = &p->se;
3922 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3924 WARN_ON(task_rq(p) != rq);
3926 if (cfs_rq->nr_running > 1) {
3927 u64 slice = sched_slice(cfs_rq, se);
3928 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3929 s64 delta = slice - ran;
3936 hrtick_start(rq, delta);
3941 * called from enqueue/dequeue and updates the hrtick when the
3942 * current task is from our class and nr_running is low enough
3945 static void hrtick_update(struct rq *rq)
3947 struct task_struct *curr = rq->curr;
3949 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3952 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3953 hrtick_start_fair(rq, curr);
3955 #else /* !CONFIG_SCHED_HRTICK */
3957 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3961 static inline void hrtick_update(struct rq *rq)
3967 * The enqueue_task method is called before nr_running is
3968 * increased. Here we update the fair scheduling stats and
3969 * then put the task into the rbtree:
3972 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3974 struct cfs_rq *cfs_rq;
3975 struct sched_entity *se = &p->se;
3977 for_each_sched_entity(se) {
3980 cfs_rq = cfs_rq_of(se);
3981 enqueue_entity(cfs_rq, se, flags);
3984 * end evaluation on encountering a throttled cfs_rq
3986 * note: in the case of encountering a throttled cfs_rq we will
3987 * post the final h_nr_running increment below.
3989 if (cfs_rq_throttled(cfs_rq))
3991 cfs_rq->h_nr_running++;
3993 flags = ENQUEUE_WAKEUP;
3996 for_each_sched_entity(se) {
3997 cfs_rq = cfs_rq_of(se);
3998 cfs_rq->h_nr_running++;
4000 if (cfs_rq_throttled(cfs_rq))
4003 update_cfs_shares(cfs_rq);
4004 update_entity_load_avg(se, 1);
4008 update_rq_runnable_avg(rq, rq->nr_running);
4009 add_nr_running(rq, 1);
4014 static void set_next_buddy(struct sched_entity *se);
4017 * The dequeue_task method is called before nr_running is
4018 * decreased. We remove the task from the rbtree and
4019 * update the fair scheduling stats:
4021 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4023 struct cfs_rq *cfs_rq;
4024 struct sched_entity *se = &p->se;
4025 int task_sleep = flags & DEQUEUE_SLEEP;
4027 for_each_sched_entity(se) {
4028 cfs_rq = cfs_rq_of(se);
4029 dequeue_entity(cfs_rq, se, flags);
4032 * end evaluation on encountering a throttled cfs_rq
4034 * note: in the case of encountering a throttled cfs_rq we will
4035 * post the final h_nr_running decrement below.
4037 if (cfs_rq_throttled(cfs_rq))
4039 cfs_rq->h_nr_running--;
4041 /* Don't dequeue parent if it has other entities besides us */
4042 if (cfs_rq->load.weight) {
4044 * Bias pick_next to pick a task from this cfs_rq, as
4045 * p is sleeping when it is within its sched_slice.
4047 if (task_sleep && parent_entity(se))
4048 set_next_buddy(parent_entity(se));
4050 /* avoid re-evaluating load for this entity */
4051 se = parent_entity(se);
4054 flags |= DEQUEUE_SLEEP;
4057 for_each_sched_entity(se) {
4058 cfs_rq = cfs_rq_of(se);
4059 cfs_rq->h_nr_running--;
4061 if (cfs_rq_throttled(cfs_rq))
4064 update_cfs_shares(cfs_rq);
4065 update_entity_load_avg(se, 1);
4069 sub_nr_running(rq, 1);
4070 update_rq_runnable_avg(rq, 1);
4076 /* Used instead of source_load when we know the type == 0 */
4077 static unsigned long weighted_cpuload(const int cpu)
4079 return cpu_rq(cpu)->cfs.runnable_load_avg;
4083 * Return a low guess at the load of a migration-source cpu weighted
4084 * according to the scheduling class and "nice" value.
4086 * We want to under-estimate the load of migration sources, to
4087 * balance conservatively.
4089 static unsigned long source_load(int cpu, int type)
4091 struct rq *rq = cpu_rq(cpu);
4092 unsigned long total = weighted_cpuload(cpu);
4094 if (type == 0 || !sched_feat(LB_BIAS))
4097 return min(rq->cpu_load[type-1], total);
4101 * Return a high guess at the load of a migration-target cpu weighted
4102 * according to the scheduling class and "nice" value.
4104 static unsigned long target_load(int cpu, int type)
4106 struct rq *rq = cpu_rq(cpu);
4107 unsigned long total = weighted_cpuload(cpu);
4109 if (type == 0 || !sched_feat(LB_BIAS))
4112 return max(rq->cpu_load[type-1], total);
4115 static unsigned long capacity_of(int cpu)
4117 return cpu_rq(cpu)->cpu_capacity;
4120 static unsigned long cpu_avg_load_per_task(int cpu)
4122 struct rq *rq = cpu_rq(cpu);
4123 unsigned long nr_running = ACCESS_ONCE(rq->cfs.h_nr_running);
4124 unsigned long load_avg = rq->cfs.runnable_load_avg;
4127 return load_avg / nr_running;
4132 static void record_wakee(struct task_struct *p)
4135 * Rough decay (wiping) for cost saving, don't worry
4136 * about the boundary, really active task won't care
4139 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4140 current->wakee_flips >>= 1;
4141 current->wakee_flip_decay_ts = jiffies;
4144 if (current->last_wakee != p) {
4145 current->last_wakee = p;
4146 current->wakee_flips++;
4150 static void task_waking_fair(struct task_struct *p)
4152 struct sched_entity *se = &p->se;
4153 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4156 #ifndef CONFIG_64BIT
4157 u64 min_vruntime_copy;
4160 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4162 min_vruntime = cfs_rq->min_vruntime;
4163 } while (min_vruntime != min_vruntime_copy);
4165 min_vruntime = cfs_rq->min_vruntime;
4168 se->vruntime -= min_vruntime;
4172 #ifdef CONFIG_FAIR_GROUP_SCHED
4174 * effective_load() calculates the load change as seen from the root_task_group
4176 * Adding load to a group doesn't make a group heavier, but can cause movement
4177 * of group shares between cpus. Assuming the shares were perfectly aligned one
4178 * can calculate the shift in shares.
4180 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4181 * on this @cpu and results in a total addition (subtraction) of @wg to the
4182 * total group weight.
4184 * Given a runqueue weight distribution (rw_i) we can compute a shares
4185 * distribution (s_i) using:
4187 * s_i = rw_i / \Sum rw_j (1)
4189 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4190 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4191 * shares distribution (s_i):
4193 * rw_i = { 2, 4, 1, 0 }
4194 * s_i = { 2/7, 4/7, 1/7, 0 }
4196 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4197 * task used to run on and the CPU the waker is running on), we need to
4198 * compute the effect of waking a task on either CPU and, in case of a sync
4199 * wakeup, compute the effect of the current task going to sleep.
4201 * So for a change of @wl to the local @cpu with an overall group weight change
4202 * of @wl we can compute the new shares distribution (s'_i) using:
4204 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4206 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4207 * differences in waking a task to CPU 0. The additional task changes the
4208 * weight and shares distributions like:
4210 * rw'_i = { 3, 4, 1, 0 }
4211 * s'_i = { 3/8, 4/8, 1/8, 0 }
4213 * We can then compute the difference in effective weight by using:
4215 * dw_i = S * (s'_i - s_i) (3)
4217 * Where 'S' is the group weight as seen by its parent.
4219 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4220 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4221 * 4/7) times the weight of the group.
4223 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4225 struct sched_entity *se = tg->se[cpu];
4227 if (!tg->parent) /* the trivial, non-cgroup case */
4230 for_each_sched_entity(se) {
4236 * W = @wg + \Sum rw_j
4238 W = wg + calc_tg_weight(tg, se->my_q);
4243 w = se->my_q->load.weight + wl;
4246 * wl = S * s'_i; see (2)
4249 wl = (w * tg->shares) / W;
4254 * Per the above, wl is the new se->load.weight value; since
4255 * those are clipped to [MIN_SHARES, ...) do so now. See
4256 * calc_cfs_shares().
4258 if (wl < MIN_SHARES)
4262 * wl = dw_i = S * (s'_i - s_i); see (3)
4264 wl -= se->load.weight;
4267 * Recursively apply this logic to all parent groups to compute
4268 * the final effective load change on the root group. Since
4269 * only the @tg group gets extra weight, all parent groups can
4270 * only redistribute existing shares. @wl is the shift in shares
4271 * resulting from this level per the above.
4280 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4287 static int wake_wide(struct task_struct *p)
4289 int factor = this_cpu_read(sd_llc_size);
4292 * Yeah, it's the switching-frequency, could means many wakee or
4293 * rapidly switch, use factor here will just help to automatically
4294 * adjust the loose-degree, so bigger node will lead to more pull.
4296 if (p->wakee_flips > factor) {
4298 * wakee is somewhat hot, it needs certain amount of cpu
4299 * resource, so if waker is far more hot, prefer to leave
4302 if (current->wakee_flips > (factor * p->wakee_flips))
4309 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4311 s64 this_load, load;
4312 s64 this_eff_load, prev_eff_load;
4313 int idx, this_cpu, prev_cpu;
4314 struct task_group *tg;
4315 unsigned long weight;
4319 * If we wake multiple tasks be careful to not bounce
4320 * ourselves around too much.
4326 this_cpu = smp_processor_id();
4327 prev_cpu = task_cpu(p);
4328 load = source_load(prev_cpu, idx);
4329 this_load = target_load(this_cpu, idx);
4332 * If sync wakeup then subtract the (maximum possible)
4333 * effect of the currently running task from the load
4334 * of the current CPU:
4337 tg = task_group(current);
4338 weight = current->se.load.weight;
4340 this_load += effective_load(tg, this_cpu, -weight, -weight);
4341 load += effective_load(tg, prev_cpu, 0, -weight);
4345 weight = p->se.load.weight;
4348 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4349 * due to the sync cause above having dropped this_load to 0, we'll
4350 * always have an imbalance, but there's really nothing you can do
4351 * about that, so that's good too.
4353 * Otherwise check if either cpus are near enough in load to allow this
4354 * task to be woken on this_cpu.
4356 this_eff_load = 100;
4357 this_eff_load *= capacity_of(prev_cpu);
4359 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4360 prev_eff_load *= capacity_of(this_cpu);
4362 if (this_load > 0) {
4363 this_eff_load *= this_load +
4364 effective_load(tg, this_cpu, weight, weight);
4366 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4369 balanced = this_eff_load <= prev_eff_load;
4371 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4376 schedstat_inc(sd, ttwu_move_affine);
4377 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4383 * find_idlest_group finds and returns the least busy CPU group within the
4386 static struct sched_group *
4387 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4388 int this_cpu, int sd_flag)
4390 struct sched_group *idlest = NULL, *group = sd->groups;
4391 unsigned long min_load = ULONG_MAX, this_load = 0;
4392 int load_idx = sd->forkexec_idx;
4393 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4395 if (sd_flag & SD_BALANCE_WAKE)
4396 load_idx = sd->wake_idx;
4399 unsigned long load, avg_load;
4403 /* Skip over this group if it has no CPUs allowed */
4404 if (!cpumask_intersects(sched_group_cpus(group),
4405 tsk_cpus_allowed(p)))
4408 local_group = cpumask_test_cpu(this_cpu,
4409 sched_group_cpus(group));
4411 /* Tally up the load of all CPUs in the group */
4414 for_each_cpu(i, sched_group_cpus(group)) {
4415 /* Bias balancing toward cpus of our domain */
4417 load = source_load(i, load_idx);
4419 load = target_load(i, load_idx);
4424 /* Adjust by relative CPU capacity of the group */
4425 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4428 this_load = avg_load;
4429 } else if (avg_load < min_load) {
4430 min_load = avg_load;
4433 } while (group = group->next, group != sd->groups);
4435 if (!idlest || 100*this_load < imbalance*min_load)
4441 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4444 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4446 unsigned long load, min_load = ULONG_MAX;
4447 unsigned int min_exit_latency = UINT_MAX;
4448 u64 latest_idle_timestamp = 0;
4449 int least_loaded_cpu = this_cpu;
4450 int shallowest_idle_cpu = -1;
4453 /* Traverse only the allowed CPUs */
4454 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4456 struct rq *rq = cpu_rq(i);
4457 struct cpuidle_state *idle = idle_get_state(rq);
4458 if (idle && idle->exit_latency < min_exit_latency) {
4460 * We give priority to a CPU whose idle state
4461 * has the smallest exit latency irrespective
4462 * of any idle timestamp.
4464 min_exit_latency = idle->exit_latency;
4465 latest_idle_timestamp = rq->idle_stamp;
4466 shallowest_idle_cpu = i;
4467 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4468 rq->idle_stamp > latest_idle_timestamp) {
4470 * If equal or no active idle state, then
4471 * the most recently idled CPU might have
4474 latest_idle_timestamp = rq->idle_stamp;
4475 shallowest_idle_cpu = i;
4478 load = weighted_cpuload(i);
4479 if (load < min_load || (load == min_load && i == this_cpu)) {
4481 least_loaded_cpu = i;
4486 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4490 * Try and locate an idle CPU in the sched_domain.
4492 static int select_idle_sibling(struct task_struct *p, int target)
4494 struct sched_domain *sd;
4495 struct sched_group *sg;
4496 int i = task_cpu(p);
4498 if (idle_cpu(target))
4502 * If the prevous cpu is cache affine and idle, don't be stupid.
4504 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4508 * Otherwise, iterate the domains and find an elegible idle cpu.
4510 sd = rcu_dereference(per_cpu(sd_llc, target));
4511 for_each_lower_domain(sd) {
4514 if (!cpumask_intersects(sched_group_cpus(sg),
4515 tsk_cpus_allowed(p)))
4518 for_each_cpu(i, sched_group_cpus(sg)) {
4519 if (i == target || !idle_cpu(i))
4523 target = cpumask_first_and(sched_group_cpus(sg),
4524 tsk_cpus_allowed(p));
4528 } while (sg != sd->groups);
4535 * select_task_rq_fair: Select target runqueue for the waking task in domains
4536 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4537 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4539 * Balances load by selecting the idlest cpu in the idlest group, or under
4540 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4542 * Returns the target cpu number.
4544 * preempt must be disabled.
4547 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4549 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4550 int cpu = smp_processor_id();
4552 int want_affine = 0;
4553 int sync = wake_flags & WF_SYNC;
4555 if (p->nr_cpus_allowed == 1)
4558 if (sd_flag & SD_BALANCE_WAKE)
4559 want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4562 for_each_domain(cpu, tmp) {
4563 if (!(tmp->flags & SD_LOAD_BALANCE))
4567 * If both cpu and prev_cpu are part of this domain,
4568 * cpu is a valid SD_WAKE_AFFINE target.
4570 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4571 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4576 if (tmp->flags & sd_flag)
4580 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4583 if (sd_flag & SD_BALANCE_WAKE) {
4584 new_cpu = select_idle_sibling(p, prev_cpu);
4589 struct sched_group *group;
4592 if (!(sd->flags & sd_flag)) {
4597 group = find_idlest_group(sd, p, cpu, sd_flag);
4603 new_cpu = find_idlest_cpu(group, p, cpu);
4604 if (new_cpu == -1 || new_cpu == cpu) {
4605 /* Now try balancing at a lower domain level of cpu */
4610 /* Now try balancing at a lower domain level of new_cpu */
4612 weight = sd->span_weight;
4614 for_each_domain(cpu, tmp) {
4615 if (weight <= tmp->span_weight)
4617 if (tmp->flags & sd_flag)
4620 /* while loop will break here if sd == NULL */
4629 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4630 * cfs_rq_of(p) references at time of call are still valid and identify the
4631 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4632 * other assumptions, including the state of rq->lock, should be made.
4635 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4637 struct sched_entity *se = &p->se;
4638 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4641 * Load tracking: accumulate removed load so that it can be processed
4642 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4643 * to blocked load iff they have a positive decay-count. It can never
4644 * be negative here since on-rq tasks have decay-count == 0.
4646 if (se->avg.decay_count) {
4647 se->avg.decay_count = -__synchronize_entity_decay(se);
4648 atomic_long_add(se->avg.load_avg_contrib,
4649 &cfs_rq->removed_load);
4652 /* We have migrated, no longer consider this task hot */
4655 #endif /* CONFIG_SMP */
4657 static unsigned long
4658 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4660 unsigned long gran = sysctl_sched_wakeup_granularity;
4663 * Since its curr running now, convert the gran from real-time
4664 * to virtual-time in his units.
4666 * By using 'se' instead of 'curr' we penalize light tasks, so
4667 * they get preempted easier. That is, if 'se' < 'curr' then
4668 * the resulting gran will be larger, therefore penalizing the
4669 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4670 * be smaller, again penalizing the lighter task.
4672 * This is especially important for buddies when the leftmost
4673 * task is higher priority than the buddy.
4675 return calc_delta_fair(gran, se);
4679 * Should 'se' preempt 'curr'.
4693 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4695 s64 gran, vdiff = curr->vruntime - se->vruntime;
4700 gran = wakeup_gran(curr, se);
4707 static void set_last_buddy(struct sched_entity *se)
4709 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4712 for_each_sched_entity(se)
4713 cfs_rq_of(se)->last = se;
4716 static void set_next_buddy(struct sched_entity *se)
4718 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4721 for_each_sched_entity(se)
4722 cfs_rq_of(se)->next = se;
4725 static void set_skip_buddy(struct sched_entity *se)
4727 for_each_sched_entity(se)
4728 cfs_rq_of(se)->skip = se;
4732 * Preempt the current task with a newly woken task if needed:
4734 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4736 struct task_struct *curr = rq->curr;
4737 struct sched_entity *se = &curr->se, *pse = &p->se;
4738 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4739 int scale = cfs_rq->nr_running >= sched_nr_latency;
4740 int next_buddy_marked = 0;
4742 if (unlikely(se == pse))
4746 * This is possible from callers such as attach_tasks(), in which we
4747 * unconditionally check_prempt_curr() after an enqueue (which may have
4748 * lead to a throttle). This both saves work and prevents false
4749 * next-buddy nomination below.
4751 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4754 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4755 set_next_buddy(pse);
4756 next_buddy_marked = 1;
4760 * We can come here with TIF_NEED_RESCHED already set from new task
4763 * Note: this also catches the edge-case of curr being in a throttled
4764 * group (e.g. via set_curr_task), since update_curr() (in the
4765 * enqueue of curr) will have resulted in resched being set. This
4766 * prevents us from potentially nominating it as a false LAST_BUDDY
4769 if (test_tsk_need_resched(curr))
4772 /* Idle tasks are by definition preempted by non-idle tasks. */
4773 if (unlikely(curr->policy == SCHED_IDLE) &&
4774 likely(p->policy != SCHED_IDLE))
4778 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4779 * is driven by the tick):
4781 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4784 find_matching_se(&se, &pse);
4785 update_curr(cfs_rq_of(se));
4787 if (wakeup_preempt_entity(se, pse) == 1) {
4789 * Bias pick_next to pick the sched entity that is
4790 * triggering this preemption.
4792 if (!next_buddy_marked)
4793 set_next_buddy(pse);
4802 * Only set the backward buddy when the current task is still
4803 * on the rq. This can happen when a wakeup gets interleaved
4804 * with schedule on the ->pre_schedule() or idle_balance()
4805 * point, either of which can * drop the rq lock.
4807 * Also, during early boot the idle thread is in the fair class,
4808 * for obvious reasons its a bad idea to schedule back to it.
4810 if (unlikely(!se->on_rq || curr == rq->idle))
4813 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4817 static struct task_struct *
4818 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4820 struct cfs_rq *cfs_rq = &rq->cfs;
4821 struct sched_entity *se;
4822 struct task_struct *p;
4826 #ifdef CONFIG_FAIR_GROUP_SCHED
4827 if (!cfs_rq->nr_running)
4830 if (prev->sched_class != &fair_sched_class)
4834 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4835 * likely that a next task is from the same cgroup as the current.
4837 * Therefore attempt to avoid putting and setting the entire cgroup
4838 * hierarchy, only change the part that actually changes.
4842 struct sched_entity *curr = cfs_rq->curr;
4845 * Since we got here without doing put_prev_entity() we also
4846 * have to consider cfs_rq->curr. If it is still a runnable
4847 * entity, update_curr() will update its vruntime, otherwise
4848 * forget we've ever seen it.
4850 if (curr && curr->on_rq)
4851 update_curr(cfs_rq);
4856 * This call to check_cfs_rq_runtime() will do the throttle and
4857 * dequeue its entity in the parent(s). Therefore the 'simple'
4858 * nr_running test will indeed be correct.
4860 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4863 se = pick_next_entity(cfs_rq, curr);
4864 cfs_rq = group_cfs_rq(se);
4870 * Since we haven't yet done put_prev_entity and if the selected task
4871 * is a different task than we started out with, try and touch the
4872 * least amount of cfs_rqs.
4875 struct sched_entity *pse = &prev->se;
4877 while (!(cfs_rq = is_same_group(se, pse))) {
4878 int se_depth = se->depth;
4879 int pse_depth = pse->depth;
4881 if (se_depth <= pse_depth) {
4882 put_prev_entity(cfs_rq_of(pse), pse);
4883 pse = parent_entity(pse);
4885 if (se_depth >= pse_depth) {
4886 set_next_entity(cfs_rq_of(se), se);
4887 se = parent_entity(se);
4891 put_prev_entity(cfs_rq, pse);
4892 set_next_entity(cfs_rq, se);
4895 if (hrtick_enabled(rq))
4896 hrtick_start_fair(rq, p);
4903 if (!cfs_rq->nr_running)
4906 put_prev_task(rq, prev);
4909 se = pick_next_entity(cfs_rq, NULL);
4910 set_next_entity(cfs_rq, se);
4911 cfs_rq = group_cfs_rq(se);
4916 if (hrtick_enabled(rq))
4917 hrtick_start_fair(rq, p);
4922 new_tasks = idle_balance(rq);
4924 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4925 * possible for any higher priority task to appear. In that case we
4926 * must re-start the pick_next_entity() loop.
4938 * Account for a descheduled task:
4940 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4942 struct sched_entity *se = &prev->se;
4943 struct cfs_rq *cfs_rq;
4945 for_each_sched_entity(se) {
4946 cfs_rq = cfs_rq_of(se);
4947 put_prev_entity(cfs_rq, se);
4952 * sched_yield() is very simple
4954 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4956 static void yield_task_fair(struct rq *rq)
4958 struct task_struct *curr = rq->curr;
4959 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4960 struct sched_entity *se = &curr->se;
4963 * Are we the only task in the tree?
4965 if (unlikely(rq->nr_running == 1))
4968 clear_buddies(cfs_rq, se);
4970 if (curr->policy != SCHED_BATCH) {
4971 update_rq_clock(rq);
4973 * Update run-time statistics of the 'current'.
4975 update_curr(cfs_rq);
4977 * Tell update_rq_clock() that we've just updated,
4978 * so we don't do microscopic update in schedule()
4979 * and double the fastpath cost.
4981 rq->skip_clock_update = 1;
4987 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4989 struct sched_entity *se = &p->se;
4991 /* throttled hierarchies are not runnable */
4992 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4995 /* Tell the scheduler that we'd really like pse to run next. */
4998 yield_task_fair(rq);
5004 /**************************************************
5005 * Fair scheduling class load-balancing methods.
5009 * The purpose of load-balancing is to achieve the same basic fairness the
5010 * per-cpu scheduler provides, namely provide a proportional amount of compute
5011 * time to each task. This is expressed in the following equation:
5013 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5015 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5016 * W_i,0 is defined as:
5018 * W_i,0 = \Sum_j w_i,j (2)
5020 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5021 * is derived from the nice value as per prio_to_weight[].
5023 * The weight average is an exponential decay average of the instantaneous
5026 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5028 * C_i is the compute capacity of cpu i, typically it is the
5029 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5030 * can also include other factors [XXX].
5032 * To achieve this balance we define a measure of imbalance which follows
5033 * directly from (1):
5035 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5037 * We them move tasks around to minimize the imbalance. In the continuous
5038 * function space it is obvious this converges, in the discrete case we get
5039 * a few fun cases generally called infeasible weight scenarios.
5042 * - infeasible weights;
5043 * - local vs global optima in the discrete case. ]
5048 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5049 * for all i,j solution, we create a tree of cpus that follows the hardware
5050 * topology where each level pairs two lower groups (or better). This results
5051 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5052 * tree to only the first of the previous level and we decrease the frequency
5053 * of load-balance at each level inv. proportional to the number of cpus in
5059 * \Sum { --- * --- * 2^i } = O(n) (5)
5061 * `- size of each group
5062 * | | `- number of cpus doing load-balance
5064 * `- sum over all levels
5066 * Coupled with a limit on how many tasks we can migrate every balance pass,
5067 * this makes (5) the runtime complexity of the balancer.
5069 * An important property here is that each CPU is still (indirectly) connected
5070 * to every other cpu in at most O(log n) steps:
5072 * The adjacency matrix of the resulting graph is given by:
5075 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5078 * And you'll find that:
5080 * A^(log_2 n)_i,j != 0 for all i,j (7)
5082 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5083 * The task movement gives a factor of O(m), giving a convergence complexity
5086 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5091 * In order to avoid CPUs going idle while there's still work to do, new idle
5092 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5093 * tree itself instead of relying on other CPUs to bring it work.
5095 * This adds some complexity to both (5) and (8) but it reduces the total idle
5103 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5106 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5111 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5113 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5115 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5118 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5119 * rewrite all of this once again.]
5122 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5124 enum fbq_type { regular, remote, all };
5126 #define LBF_ALL_PINNED 0x01
5127 #define LBF_NEED_BREAK 0x02
5128 #define LBF_DST_PINNED 0x04
5129 #define LBF_SOME_PINNED 0x08
5132 struct sched_domain *sd;
5140 struct cpumask *dst_grpmask;
5142 enum cpu_idle_type idle;
5144 /* The set of CPUs under consideration for load-balancing */
5145 struct cpumask *cpus;
5150 unsigned int loop_break;
5151 unsigned int loop_max;
5153 enum fbq_type fbq_type;
5154 struct list_head tasks;
5158 * Is this task likely cache-hot:
5160 static int task_hot(struct task_struct *p, struct lb_env *env)
5164 lockdep_assert_held(&env->src_rq->lock);
5166 if (p->sched_class != &fair_sched_class)
5169 if (unlikely(p->policy == SCHED_IDLE))
5173 * Buddy candidates are cache hot:
5175 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5176 (&p->se == cfs_rq_of(&p->se)->next ||
5177 &p->se == cfs_rq_of(&p->se)->last))
5180 if (sysctl_sched_migration_cost == -1)
5182 if (sysctl_sched_migration_cost == 0)
5185 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5187 return delta < (s64)sysctl_sched_migration_cost;
5190 #ifdef CONFIG_NUMA_BALANCING
5191 /* Returns true if the destination node has incurred more faults */
5192 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5194 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5195 int src_nid, dst_nid;
5197 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5198 !(env->sd->flags & SD_NUMA)) {
5202 src_nid = cpu_to_node(env->src_cpu);
5203 dst_nid = cpu_to_node(env->dst_cpu);
5205 if (src_nid == dst_nid)
5209 /* Task is already in the group's interleave set. */
5210 if (node_isset(src_nid, numa_group->active_nodes))
5213 /* Task is moving into the group's interleave set. */
5214 if (node_isset(dst_nid, numa_group->active_nodes))
5217 return group_faults(p, dst_nid) > group_faults(p, src_nid);
5220 /* Encourage migration to the preferred node. */
5221 if (dst_nid == p->numa_preferred_nid)
5224 return task_faults(p, dst_nid) > task_faults(p, src_nid);
5228 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5230 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5231 int src_nid, dst_nid;
5233 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5236 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5239 src_nid = cpu_to_node(env->src_cpu);
5240 dst_nid = cpu_to_node(env->dst_cpu);
5242 if (src_nid == dst_nid)
5246 /* Task is moving within/into the group's interleave set. */
5247 if (node_isset(dst_nid, numa_group->active_nodes))
5250 /* Task is moving out of the group's interleave set. */
5251 if (node_isset(src_nid, numa_group->active_nodes))
5254 return group_faults(p, dst_nid) < group_faults(p, src_nid);
5257 /* Migrating away from the preferred node is always bad. */
5258 if (src_nid == p->numa_preferred_nid)
5261 return task_faults(p, dst_nid) < task_faults(p, src_nid);
5265 static inline bool migrate_improves_locality(struct task_struct *p,
5271 static inline bool migrate_degrades_locality(struct task_struct *p,
5279 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5282 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5284 int tsk_cache_hot = 0;
5286 lockdep_assert_held(&env->src_rq->lock);
5289 * We do not migrate tasks that are:
5290 * 1) throttled_lb_pair, or
5291 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5292 * 3) running (obviously), or
5293 * 4) are cache-hot on their current CPU.
5295 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5298 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5301 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5303 env->flags |= LBF_SOME_PINNED;
5306 * Remember if this task can be migrated to any other cpu in
5307 * our sched_group. We may want to revisit it if we couldn't
5308 * meet load balance goals by pulling other tasks on src_cpu.
5310 * Also avoid computing new_dst_cpu if we have already computed
5311 * one in current iteration.
5313 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5316 /* Prevent to re-select dst_cpu via env's cpus */
5317 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5318 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5319 env->flags |= LBF_DST_PINNED;
5320 env->new_dst_cpu = cpu;
5328 /* Record that we found atleast one task that could run on dst_cpu */
5329 env->flags &= ~LBF_ALL_PINNED;
5331 if (task_running(env->src_rq, p)) {
5332 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5337 * Aggressive migration if:
5338 * 1) destination numa is preferred
5339 * 2) task is cache cold, or
5340 * 3) too many balance attempts have failed.
5342 tsk_cache_hot = task_hot(p, env);
5344 tsk_cache_hot = migrate_degrades_locality(p, env);
5346 if (migrate_improves_locality(p, env) || !tsk_cache_hot ||
5347 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5348 if (tsk_cache_hot) {
5349 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5350 schedstat_inc(p, se.statistics.nr_forced_migrations);
5355 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5360 * detach_task() -- detach the task for the migration specified in env
5362 static void detach_task(struct task_struct *p, struct lb_env *env)
5364 lockdep_assert_held(&env->src_rq->lock);
5366 deactivate_task(env->src_rq, p, 0);
5367 p->on_rq = TASK_ON_RQ_MIGRATING;
5368 set_task_cpu(p, env->dst_cpu);
5372 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5373 * part of active balancing operations within "domain".
5375 * Returns a task if successful and NULL otherwise.
5377 static struct task_struct *detach_one_task(struct lb_env *env)
5379 struct task_struct *p, *n;
5381 lockdep_assert_held(&env->src_rq->lock);
5383 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5384 if (!can_migrate_task(p, env))
5387 detach_task(p, env);
5390 * Right now, this is only the second place where
5391 * lb_gained[env->idle] is updated (other is detach_tasks)
5392 * so we can safely collect stats here rather than
5393 * inside detach_tasks().
5395 schedstat_inc(env->sd, lb_gained[env->idle]);
5401 static const unsigned int sched_nr_migrate_break = 32;
5404 * detach_tasks() -- tries to detach up to imbalance weighted load from
5405 * busiest_rq, as part of a balancing operation within domain "sd".
5407 * Returns number of detached tasks if successful and 0 otherwise.
5409 static int detach_tasks(struct lb_env *env)
5411 struct list_head *tasks = &env->src_rq->cfs_tasks;
5412 struct task_struct *p;
5416 lockdep_assert_held(&env->src_rq->lock);
5418 if (env->imbalance <= 0)
5421 while (!list_empty(tasks)) {
5422 p = list_first_entry(tasks, struct task_struct, se.group_node);
5425 /* We've more or less seen every task there is, call it quits */
5426 if (env->loop > env->loop_max)
5429 /* take a breather every nr_migrate tasks */
5430 if (env->loop > env->loop_break) {
5431 env->loop_break += sched_nr_migrate_break;
5432 env->flags |= LBF_NEED_BREAK;
5436 if (!can_migrate_task(p, env))
5439 load = task_h_load(p);
5441 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5444 if ((load / 2) > env->imbalance)
5447 detach_task(p, env);
5448 list_add(&p->se.group_node, &env->tasks);
5451 env->imbalance -= load;
5453 #ifdef CONFIG_PREEMPT
5455 * NEWIDLE balancing is a source of latency, so preemptible
5456 * kernels will stop after the first task is detached to minimize
5457 * the critical section.
5459 if (env->idle == CPU_NEWLY_IDLE)
5464 * We only want to steal up to the prescribed amount of
5467 if (env->imbalance <= 0)
5472 list_move_tail(&p->se.group_node, tasks);
5476 * Right now, this is one of only two places we collect this stat
5477 * so we can safely collect detach_one_task() stats here rather
5478 * than inside detach_one_task().
5480 schedstat_add(env->sd, lb_gained[env->idle], detached);
5486 * attach_task() -- attach the task detached by detach_task() to its new rq.
5488 static void attach_task(struct rq *rq, struct task_struct *p)
5490 lockdep_assert_held(&rq->lock);
5492 BUG_ON(task_rq(p) != rq);
5493 p->on_rq = TASK_ON_RQ_QUEUED;
5494 activate_task(rq, p, 0);
5495 check_preempt_curr(rq, p, 0);
5499 * attach_one_task() -- attaches the task returned from detach_one_task() to
5502 static void attach_one_task(struct rq *rq, struct task_struct *p)
5504 raw_spin_lock(&rq->lock);
5506 raw_spin_unlock(&rq->lock);
5510 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5513 static void attach_tasks(struct lb_env *env)
5515 struct list_head *tasks = &env->tasks;
5516 struct task_struct *p;
5518 raw_spin_lock(&env->dst_rq->lock);
5520 while (!list_empty(tasks)) {
5521 p = list_first_entry(tasks, struct task_struct, se.group_node);
5522 list_del_init(&p->se.group_node);
5524 attach_task(env->dst_rq, p);
5527 raw_spin_unlock(&env->dst_rq->lock);
5530 #ifdef CONFIG_FAIR_GROUP_SCHED
5532 * update tg->load_weight by folding this cpu's load_avg
5534 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5536 struct sched_entity *se = tg->se[cpu];
5537 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5539 /* throttled entities do not contribute to load */
5540 if (throttled_hierarchy(cfs_rq))
5543 update_cfs_rq_blocked_load(cfs_rq, 1);
5546 update_entity_load_avg(se, 1);
5548 * We pivot on our runnable average having decayed to zero for
5549 * list removal. This generally implies that all our children
5550 * have also been removed (modulo rounding error or bandwidth
5551 * control); however, such cases are rare and we can fix these
5554 * TODO: fix up out-of-order children on enqueue.
5556 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5557 list_del_leaf_cfs_rq(cfs_rq);
5559 struct rq *rq = rq_of(cfs_rq);
5560 update_rq_runnable_avg(rq, rq->nr_running);
5564 static void update_blocked_averages(int cpu)
5566 struct rq *rq = cpu_rq(cpu);
5567 struct cfs_rq *cfs_rq;
5568 unsigned long flags;
5570 raw_spin_lock_irqsave(&rq->lock, flags);
5571 update_rq_clock(rq);
5573 * Iterates the task_group tree in a bottom up fashion, see
5574 * list_add_leaf_cfs_rq() for details.
5576 for_each_leaf_cfs_rq(rq, cfs_rq) {
5578 * Note: We may want to consider periodically releasing
5579 * rq->lock about these updates so that creating many task
5580 * groups does not result in continually extending hold time.
5582 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5585 raw_spin_unlock_irqrestore(&rq->lock, flags);
5589 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5590 * This needs to be done in a top-down fashion because the load of a child
5591 * group is a fraction of its parents load.
5593 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5595 struct rq *rq = rq_of(cfs_rq);
5596 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5597 unsigned long now = jiffies;
5600 if (cfs_rq->last_h_load_update == now)
5603 cfs_rq->h_load_next = NULL;
5604 for_each_sched_entity(se) {
5605 cfs_rq = cfs_rq_of(se);
5606 cfs_rq->h_load_next = se;
5607 if (cfs_rq->last_h_load_update == now)
5612 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5613 cfs_rq->last_h_load_update = now;
5616 while ((se = cfs_rq->h_load_next) != NULL) {
5617 load = cfs_rq->h_load;
5618 load = div64_ul(load * se->avg.load_avg_contrib,
5619 cfs_rq->runnable_load_avg + 1);
5620 cfs_rq = group_cfs_rq(se);
5621 cfs_rq->h_load = load;
5622 cfs_rq->last_h_load_update = now;
5626 static unsigned long task_h_load(struct task_struct *p)
5628 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5630 update_cfs_rq_h_load(cfs_rq);
5631 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5632 cfs_rq->runnable_load_avg + 1);
5635 static inline void update_blocked_averages(int cpu)
5639 static unsigned long task_h_load(struct task_struct *p)
5641 return p->se.avg.load_avg_contrib;
5645 /********** Helpers for find_busiest_group ************************/
5654 * sg_lb_stats - stats of a sched_group required for load_balancing
5656 struct sg_lb_stats {
5657 unsigned long avg_load; /*Avg load across the CPUs of the group */
5658 unsigned long group_load; /* Total load over the CPUs of the group */
5659 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5660 unsigned long load_per_task;
5661 unsigned long group_capacity;
5662 unsigned int sum_nr_running; /* Nr tasks running in the group */
5663 unsigned int group_capacity_factor;
5664 unsigned int idle_cpus;
5665 unsigned int group_weight;
5666 enum group_type group_type;
5667 int group_has_free_capacity;
5668 #ifdef CONFIG_NUMA_BALANCING
5669 unsigned int nr_numa_running;
5670 unsigned int nr_preferred_running;
5675 * sd_lb_stats - Structure to store the statistics of a sched_domain
5676 * during load balancing.
5678 struct sd_lb_stats {
5679 struct sched_group *busiest; /* Busiest group in this sd */
5680 struct sched_group *local; /* Local group in this sd */
5681 unsigned long total_load; /* Total load of all groups in sd */
5682 unsigned long total_capacity; /* Total capacity of all groups in sd */
5683 unsigned long avg_load; /* Average load across all groups in sd */
5685 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5686 struct sg_lb_stats local_stat; /* Statistics of the local group */
5689 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5692 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5693 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5694 * We must however clear busiest_stat::avg_load because
5695 * update_sd_pick_busiest() reads this before assignment.
5697 *sds = (struct sd_lb_stats){
5701 .total_capacity = 0UL,
5704 .sum_nr_running = 0,
5705 .group_type = group_other,
5711 * get_sd_load_idx - Obtain the load index for a given sched domain.
5712 * @sd: The sched_domain whose load_idx is to be obtained.
5713 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5715 * Return: The load index.
5717 static inline int get_sd_load_idx(struct sched_domain *sd,
5718 enum cpu_idle_type idle)
5724 load_idx = sd->busy_idx;
5727 case CPU_NEWLY_IDLE:
5728 load_idx = sd->newidle_idx;
5731 load_idx = sd->idle_idx;
5738 static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu)
5740 return SCHED_CAPACITY_SCALE;
5743 unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu)
5745 return default_scale_capacity(sd, cpu);
5748 static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu)
5750 if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1))
5751 return sd->smt_gain / sd->span_weight;
5753 return SCHED_CAPACITY_SCALE;
5756 unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
5758 return default_scale_cpu_capacity(sd, cpu);
5761 static unsigned long scale_rt_capacity(int cpu)
5763 struct rq *rq = cpu_rq(cpu);
5764 u64 total, available, age_stamp, avg;
5768 * Since we're reading these variables without serialization make sure
5769 * we read them once before doing sanity checks on them.
5771 age_stamp = ACCESS_ONCE(rq->age_stamp);
5772 avg = ACCESS_ONCE(rq->rt_avg);
5774 delta = rq_clock(rq) - age_stamp;
5775 if (unlikely(delta < 0))
5778 total = sched_avg_period() + delta;
5780 if (unlikely(total < avg)) {
5781 /* Ensures that capacity won't end up being negative */
5784 available = total - avg;
5787 if (unlikely((s64)total < SCHED_CAPACITY_SCALE))
5788 total = SCHED_CAPACITY_SCALE;
5790 total >>= SCHED_CAPACITY_SHIFT;
5792 return div_u64(available, total);
5795 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
5797 unsigned long capacity = SCHED_CAPACITY_SCALE;
5798 struct sched_group *sdg = sd->groups;
5800 if (sched_feat(ARCH_CAPACITY))
5801 capacity *= arch_scale_cpu_capacity(sd, cpu);
5803 capacity *= default_scale_cpu_capacity(sd, cpu);
5805 capacity >>= SCHED_CAPACITY_SHIFT;
5807 sdg->sgc->capacity_orig = capacity;
5809 if (sched_feat(ARCH_CAPACITY))
5810 capacity *= arch_scale_freq_capacity(sd, cpu);
5812 capacity *= default_scale_capacity(sd, cpu);
5814 capacity >>= SCHED_CAPACITY_SHIFT;
5816 capacity *= scale_rt_capacity(cpu);
5817 capacity >>= SCHED_CAPACITY_SHIFT;
5822 cpu_rq(cpu)->cpu_capacity = capacity;
5823 sdg->sgc->capacity = capacity;
5826 void update_group_capacity(struct sched_domain *sd, int cpu)
5828 struct sched_domain *child = sd->child;
5829 struct sched_group *group, *sdg = sd->groups;
5830 unsigned long capacity, capacity_orig;
5831 unsigned long interval;
5833 interval = msecs_to_jiffies(sd->balance_interval);
5834 interval = clamp(interval, 1UL, max_load_balance_interval);
5835 sdg->sgc->next_update = jiffies + interval;
5838 update_cpu_capacity(sd, cpu);
5842 capacity_orig = capacity = 0;
5844 if (child->flags & SD_OVERLAP) {
5846 * SD_OVERLAP domains cannot assume that child groups
5847 * span the current group.
5850 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5851 struct sched_group_capacity *sgc;
5852 struct rq *rq = cpu_rq(cpu);
5855 * build_sched_domains() -> init_sched_groups_capacity()
5856 * gets here before we've attached the domains to the
5859 * Use capacity_of(), which is set irrespective of domains
5860 * in update_cpu_capacity().
5862 * This avoids capacity/capacity_orig from being 0 and
5863 * causing divide-by-zero issues on boot.
5865 * Runtime updates will correct capacity_orig.
5867 if (unlikely(!rq->sd)) {
5868 capacity_orig += capacity_of(cpu);
5869 capacity += capacity_of(cpu);
5873 sgc = rq->sd->groups->sgc;
5874 capacity_orig += sgc->capacity_orig;
5875 capacity += sgc->capacity;
5879 * !SD_OVERLAP domains can assume that child groups
5880 * span the current group.
5883 group = child->groups;
5885 capacity_orig += group->sgc->capacity_orig;
5886 capacity += group->sgc->capacity;
5887 group = group->next;
5888 } while (group != child->groups);
5891 sdg->sgc->capacity_orig = capacity_orig;
5892 sdg->sgc->capacity = capacity;
5896 * Try and fix up capacity for tiny siblings, this is needed when
5897 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5898 * which on its own isn't powerful enough.
5900 * See update_sd_pick_busiest() and check_asym_packing().
5903 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5906 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5908 if (!(sd->flags & SD_SHARE_CPUCAPACITY))
5912 * If ~90% of the cpu_capacity is still there, we're good.
5914 if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29)
5921 * Group imbalance indicates (and tries to solve) the problem where balancing
5922 * groups is inadequate due to tsk_cpus_allowed() constraints.
5924 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5925 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5928 * { 0 1 2 3 } { 4 5 6 7 }
5931 * If we were to balance group-wise we'd place two tasks in the first group and
5932 * two tasks in the second group. Clearly this is undesired as it will overload
5933 * cpu 3 and leave one of the cpus in the second group unused.
5935 * The current solution to this issue is detecting the skew in the first group
5936 * by noticing the lower domain failed to reach balance and had difficulty
5937 * moving tasks due to affinity constraints.
5939 * When this is so detected; this group becomes a candidate for busiest; see
5940 * update_sd_pick_busiest(). And calculate_imbalance() and
5941 * find_busiest_group() avoid some of the usual balance conditions to allow it
5942 * to create an effective group imbalance.
5944 * This is a somewhat tricky proposition since the next run might not find the
5945 * group imbalance and decide the groups need to be balanced again. A most
5946 * subtle and fragile situation.
5949 static inline int sg_imbalanced(struct sched_group *group)
5951 return group->sgc->imbalance;
5955 * Compute the group capacity factor.
5957 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5958 * first dividing out the smt factor and computing the actual number of cores
5959 * and limit unit capacity with that.
5961 static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group)
5963 unsigned int capacity_factor, smt, cpus;
5964 unsigned int capacity, capacity_orig;
5966 capacity = group->sgc->capacity;
5967 capacity_orig = group->sgc->capacity_orig;
5968 cpus = group->group_weight;
5970 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5971 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig);
5972 capacity_factor = cpus / smt; /* cores */
5974 capacity_factor = min_t(unsigned,
5975 capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE));
5976 if (!capacity_factor)
5977 capacity_factor = fix_small_capacity(env->sd, group);
5979 return capacity_factor;
5982 static enum group_type
5983 group_classify(struct sched_group *group, struct sg_lb_stats *sgs)
5985 if (sgs->sum_nr_running > sgs->group_capacity_factor)
5986 return group_overloaded;
5988 if (sg_imbalanced(group))
5989 return group_imbalanced;
5995 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5996 * @env: The load balancing environment.
5997 * @group: sched_group whose statistics are to be updated.
5998 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5999 * @local_group: Does group contain this_cpu.
6000 * @sgs: variable to hold the statistics for this group.
6001 * @overload: Indicate more than one runnable task for any CPU.
6003 static inline void update_sg_lb_stats(struct lb_env *env,
6004 struct sched_group *group, int load_idx,
6005 int local_group, struct sg_lb_stats *sgs,
6011 memset(sgs, 0, sizeof(*sgs));
6013 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6014 struct rq *rq = cpu_rq(i);
6016 /* Bias balancing toward cpus of our domain */
6018 load = target_load(i, load_idx);
6020 load = source_load(i, load_idx);
6022 sgs->group_load += load;
6023 sgs->sum_nr_running += rq->cfs.h_nr_running;
6025 if (rq->nr_running > 1)
6028 #ifdef CONFIG_NUMA_BALANCING
6029 sgs->nr_numa_running += rq->nr_numa_running;
6030 sgs->nr_preferred_running += rq->nr_preferred_running;
6032 sgs->sum_weighted_load += weighted_cpuload(i);
6037 /* Adjust by relative CPU capacity of the group */
6038 sgs->group_capacity = group->sgc->capacity;
6039 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6041 if (sgs->sum_nr_running)
6042 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6044 sgs->group_weight = group->group_weight;
6045 sgs->group_capacity_factor = sg_capacity_factor(env, group);
6046 sgs->group_type = group_classify(group, sgs);
6048 if (sgs->group_capacity_factor > sgs->sum_nr_running)
6049 sgs->group_has_free_capacity = 1;
6053 * update_sd_pick_busiest - return 1 on busiest group
6054 * @env: The load balancing environment.
6055 * @sds: sched_domain statistics
6056 * @sg: sched_group candidate to be checked for being the busiest
6057 * @sgs: sched_group statistics
6059 * Determine if @sg is a busier group than the previously selected
6062 * Return: %true if @sg is a busier group than the previously selected
6063 * busiest group. %false otherwise.
6065 static bool update_sd_pick_busiest(struct lb_env *env,
6066 struct sd_lb_stats *sds,
6067 struct sched_group *sg,
6068 struct sg_lb_stats *sgs)
6070 struct sg_lb_stats *busiest = &sds->busiest_stat;
6072 if (sgs->group_type > busiest->group_type)
6075 if (sgs->group_type < busiest->group_type)
6078 if (sgs->avg_load <= busiest->avg_load)
6081 /* This is the busiest node in its class. */
6082 if (!(env->sd->flags & SD_ASYM_PACKING))
6086 * ASYM_PACKING needs to move all the work to the lowest
6087 * numbered CPUs in the group, therefore mark all groups
6088 * higher than ourself as busy.
6090 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6094 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6101 #ifdef CONFIG_NUMA_BALANCING
6102 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6104 if (sgs->sum_nr_running > sgs->nr_numa_running)
6106 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6111 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6113 if (rq->nr_running > rq->nr_numa_running)
6115 if (rq->nr_running > rq->nr_preferred_running)
6120 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6125 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6129 #endif /* CONFIG_NUMA_BALANCING */
6132 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6133 * @env: The load balancing environment.
6134 * @sds: variable to hold the statistics for this sched_domain.
6136 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6138 struct sched_domain *child = env->sd->child;
6139 struct sched_group *sg = env->sd->groups;
6140 struct sg_lb_stats tmp_sgs;
6141 int load_idx, prefer_sibling = 0;
6142 bool overload = false;
6144 if (child && child->flags & SD_PREFER_SIBLING)
6147 load_idx = get_sd_load_idx(env->sd, env->idle);
6150 struct sg_lb_stats *sgs = &tmp_sgs;
6153 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6156 sgs = &sds->local_stat;
6158 if (env->idle != CPU_NEWLY_IDLE ||
6159 time_after_eq(jiffies, sg->sgc->next_update))
6160 update_group_capacity(env->sd, env->dst_cpu);
6163 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6170 * In case the child domain prefers tasks go to siblings
6171 * first, lower the sg capacity factor to one so that we'll try
6172 * and move all the excess tasks away. We lower the capacity
6173 * of a group only if the local group has the capacity to fit
6174 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6175 * extra check prevents the case where you always pull from the
6176 * heaviest group when it is already under-utilized (possible
6177 * with a large weight task outweighs the tasks on the system).
6179 if (prefer_sibling && sds->local &&
6180 sds->local_stat.group_has_free_capacity)
6181 sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U);
6183 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6185 sds->busiest_stat = *sgs;
6189 /* Now, start updating sd_lb_stats */
6190 sds->total_load += sgs->group_load;
6191 sds->total_capacity += sgs->group_capacity;
6194 } while (sg != env->sd->groups);
6196 if (env->sd->flags & SD_NUMA)
6197 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6199 if (!env->sd->parent) {
6200 /* update overload indicator if we are at root domain */
6201 if (env->dst_rq->rd->overload != overload)
6202 env->dst_rq->rd->overload = overload;
6208 * check_asym_packing - Check to see if the group is packed into the
6211 * This is primarily intended to used at the sibling level. Some
6212 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6213 * case of POWER7, it can move to lower SMT modes only when higher
6214 * threads are idle. When in lower SMT modes, the threads will
6215 * perform better since they share less core resources. Hence when we
6216 * have idle threads, we want them to be the higher ones.
6218 * This packing function is run on idle threads. It checks to see if
6219 * the busiest CPU in this domain (core in the P7 case) has a higher
6220 * CPU number than the packing function is being run on. Here we are
6221 * assuming lower CPU number will be equivalent to lower a SMT thread
6224 * Return: 1 when packing is required and a task should be moved to
6225 * this CPU. The amount of the imbalance is returned in *imbalance.
6227 * @env: The load balancing environment.
6228 * @sds: Statistics of the sched_domain which is to be packed
6230 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6234 if (!(env->sd->flags & SD_ASYM_PACKING))
6240 busiest_cpu = group_first_cpu(sds->busiest);
6241 if (env->dst_cpu > busiest_cpu)
6244 env->imbalance = DIV_ROUND_CLOSEST(
6245 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6246 SCHED_CAPACITY_SCALE);
6252 * fix_small_imbalance - Calculate the minor imbalance that exists
6253 * amongst the groups of a sched_domain, during
6255 * @env: The load balancing environment.
6256 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6259 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6261 unsigned long tmp, capa_now = 0, capa_move = 0;
6262 unsigned int imbn = 2;
6263 unsigned long scaled_busy_load_per_task;
6264 struct sg_lb_stats *local, *busiest;
6266 local = &sds->local_stat;
6267 busiest = &sds->busiest_stat;
6269 if (!local->sum_nr_running)
6270 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6271 else if (busiest->load_per_task > local->load_per_task)
6274 scaled_busy_load_per_task =
6275 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6276 busiest->group_capacity;
6278 if (busiest->avg_load + scaled_busy_load_per_task >=
6279 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6280 env->imbalance = busiest->load_per_task;
6285 * OK, we don't have enough imbalance to justify moving tasks,
6286 * however we may be able to increase total CPU capacity used by
6290 capa_now += busiest->group_capacity *
6291 min(busiest->load_per_task, busiest->avg_load);
6292 capa_now += local->group_capacity *
6293 min(local->load_per_task, local->avg_load);
6294 capa_now /= SCHED_CAPACITY_SCALE;
6296 /* Amount of load we'd subtract */
6297 if (busiest->avg_load > scaled_busy_load_per_task) {
6298 capa_move += busiest->group_capacity *
6299 min(busiest->load_per_task,
6300 busiest->avg_load - scaled_busy_load_per_task);
6303 /* Amount of load we'd add */
6304 if (busiest->avg_load * busiest->group_capacity <
6305 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6306 tmp = (busiest->avg_load * busiest->group_capacity) /
6307 local->group_capacity;
6309 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6310 local->group_capacity;
6312 capa_move += local->group_capacity *
6313 min(local->load_per_task, local->avg_load + tmp);
6314 capa_move /= SCHED_CAPACITY_SCALE;
6316 /* Move if we gain throughput */
6317 if (capa_move > capa_now)
6318 env->imbalance = busiest->load_per_task;
6322 * calculate_imbalance - Calculate the amount of imbalance present within the
6323 * groups of a given sched_domain during load balance.
6324 * @env: load balance environment
6325 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6327 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6329 unsigned long max_pull, load_above_capacity = ~0UL;
6330 struct sg_lb_stats *local, *busiest;
6332 local = &sds->local_stat;
6333 busiest = &sds->busiest_stat;
6335 if (busiest->group_type == group_imbalanced) {
6337 * In the group_imb case we cannot rely on group-wide averages
6338 * to ensure cpu-load equilibrium, look at wider averages. XXX
6340 busiest->load_per_task =
6341 min(busiest->load_per_task, sds->avg_load);
6345 * In the presence of smp nice balancing, certain scenarios can have
6346 * max load less than avg load(as we skip the groups at or below
6347 * its cpu_capacity, while calculating max_load..)
6349 if (busiest->avg_load <= sds->avg_load ||
6350 local->avg_load >= sds->avg_load) {
6352 return fix_small_imbalance(env, sds);
6356 * If there aren't any idle cpus, avoid creating some.
6358 if (busiest->group_type == group_overloaded &&
6359 local->group_type == group_overloaded) {
6360 load_above_capacity =
6361 (busiest->sum_nr_running - busiest->group_capacity_factor);
6363 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE);
6364 load_above_capacity /= busiest->group_capacity;
6368 * We're trying to get all the cpus to the average_load, so we don't
6369 * want to push ourselves above the average load, nor do we wish to
6370 * reduce the max loaded cpu below the average load. At the same time,
6371 * we also don't want to reduce the group load below the group capacity
6372 * (so that we can implement power-savings policies etc). Thus we look
6373 * for the minimum possible imbalance.
6375 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6377 /* How much load to actually move to equalise the imbalance */
6378 env->imbalance = min(
6379 max_pull * busiest->group_capacity,
6380 (sds->avg_load - local->avg_load) * local->group_capacity
6381 ) / SCHED_CAPACITY_SCALE;
6384 * if *imbalance is less than the average load per runnable task
6385 * there is no guarantee that any tasks will be moved so we'll have
6386 * a think about bumping its value to force at least one task to be
6389 if (env->imbalance < busiest->load_per_task)
6390 return fix_small_imbalance(env, sds);
6393 /******* find_busiest_group() helpers end here *********************/
6396 * find_busiest_group - Returns the busiest group within the sched_domain
6397 * if there is an imbalance. If there isn't an imbalance, and
6398 * the user has opted for power-savings, it returns a group whose
6399 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6400 * such a group exists.
6402 * Also calculates the amount of weighted load which should be moved
6403 * to restore balance.
6405 * @env: The load balancing environment.
6407 * Return: - The busiest group if imbalance exists.
6408 * - If no imbalance and user has opted for power-savings balance,
6409 * return the least loaded group whose CPUs can be
6410 * put to idle by rebalancing its tasks onto our group.
6412 static struct sched_group *find_busiest_group(struct lb_env *env)
6414 struct sg_lb_stats *local, *busiest;
6415 struct sd_lb_stats sds;
6417 init_sd_lb_stats(&sds);
6420 * Compute the various statistics relavent for load balancing at
6423 update_sd_lb_stats(env, &sds);
6424 local = &sds.local_stat;
6425 busiest = &sds.busiest_stat;
6427 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6428 check_asym_packing(env, &sds))
6431 /* There is no busy sibling group to pull tasks from */
6432 if (!sds.busiest || busiest->sum_nr_running == 0)
6435 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6436 / sds.total_capacity;
6439 * If the busiest group is imbalanced the below checks don't
6440 * work because they assume all things are equal, which typically
6441 * isn't true due to cpus_allowed constraints and the like.
6443 if (busiest->group_type == group_imbalanced)
6446 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6447 if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity &&
6448 !busiest->group_has_free_capacity)
6452 * If the local group is busier than the selected busiest group
6453 * don't try and pull any tasks.
6455 if (local->avg_load >= busiest->avg_load)
6459 * Don't pull any tasks if this group is already above the domain
6462 if (local->avg_load >= sds.avg_load)
6465 if (env->idle == CPU_IDLE) {
6467 * This cpu is idle. If the busiest group is not overloaded
6468 * and there is no imbalance between this and busiest group
6469 * wrt idle cpus, it is balanced. The imbalance becomes
6470 * significant if the diff is greater than 1 otherwise we
6471 * might end up to just move the imbalance on another group
6473 if ((busiest->group_type != group_overloaded) &&
6474 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6478 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6479 * imbalance_pct to be conservative.
6481 if (100 * busiest->avg_load <=
6482 env->sd->imbalance_pct * local->avg_load)
6487 /* Looks like there is an imbalance. Compute it */
6488 calculate_imbalance(env, &sds);
6497 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6499 static struct rq *find_busiest_queue(struct lb_env *env,
6500 struct sched_group *group)
6502 struct rq *busiest = NULL, *rq;
6503 unsigned long busiest_load = 0, busiest_capacity = 1;
6506 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6507 unsigned long capacity, capacity_factor, wl;
6511 rt = fbq_classify_rq(rq);
6514 * We classify groups/runqueues into three groups:
6515 * - regular: there are !numa tasks
6516 * - remote: there are numa tasks that run on the 'wrong' node
6517 * - all: there is no distinction
6519 * In order to avoid migrating ideally placed numa tasks,
6520 * ignore those when there's better options.
6522 * If we ignore the actual busiest queue to migrate another
6523 * task, the next balance pass can still reduce the busiest
6524 * queue by moving tasks around inside the node.
6526 * If we cannot move enough load due to this classification
6527 * the next pass will adjust the group classification and
6528 * allow migration of more tasks.
6530 * Both cases only affect the total convergence complexity.
6532 if (rt > env->fbq_type)
6535 capacity = capacity_of(i);
6536 capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE);
6537 if (!capacity_factor)
6538 capacity_factor = fix_small_capacity(env->sd, group);
6540 wl = weighted_cpuload(i);
6543 * When comparing with imbalance, use weighted_cpuload()
6544 * which is not scaled with the cpu capacity.
6546 if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance)
6550 * For the load comparisons with the other cpu's, consider
6551 * the weighted_cpuload() scaled with the cpu capacity, so
6552 * that the load can be moved away from the cpu that is
6553 * potentially running at a lower capacity.
6555 * Thus we're looking for max(wl_i / capacity_i), crosswise
6556 * multiplication to rid ourselves of the division works out
6557 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6558 * our previous maximum.
6560 if (wl * busiest_capacity > busiest_load * capacity) {
6562 busiest_capacity = capacity;
6571 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6572 * so long as it is large enough.
6574 #define MAX_PINNED_INTERVAL 512
6576 /* Working cpumask for load_balance and load_balance_newidle. */
6577 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6579 static int need_active_balance(struct lb_env *env)
6581 struct sched_domain *sd = env->sd;
6583 if (env->idle == CPU_NEWLY_IDLE) {
6586 * ASYM_PACKING needs to force migrate tasks from busy but
6587 * higher numbered CPUs in order to pack all tasks in the
6588 * lowest numbered CPUs.
6590 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6594 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6597 static int active_load_balance_cpu_stop(void *data);
6599 static int should_we_balance(struct lb_env *env)
6601 struct sched_group *sg = env->sd->groups;
6602 struct cpumask *sg_cpus, *sg_mask;
6603 int cpu, balance_cpu = -1;
6606 * In the newly idle case, we will allow all the cpu's
6607 * to do the newly idle load balance.
6609 if (env->idle == CPU_NEWLY_IDLE)
6612 sg_cpus = sched_group_cpus(sg);
6613 sg_mask = sched_group_mask(sg);
6614 /* Try to find first idle cpu */
6615 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6616 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6623 if (balance_cpu == -1)
6624 balance_cpu = group_balance_cpu(sg);
6627 * First idle cpu or the first cpu(busiest) in this sched group
6628 * is eligible for doing load balancing at this and above domains.
6630 return balance_cpu == env->dst_cpu;
6634 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6635 * tasks if there is an imbalance.
6637 static int load_balance(int this_cpu, struct rq *this_rq,
6638 struct sched_domain *sd, enum cpu_idle_type idle,
6639 int *continue_balancing)
6641 int ld_moved, cur_ld_moved, active_balance = 0;
6642 struct sched_domain *sd_parent = sd->parent;
6643 struct sched_group *group;
6645 unsigned long flags;
6646 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
6648 struct lb_env env = {
6650 .dst_cpu = this_cpu,
6652 .dst_grpmask = sched_group_cpus(sd->groups),
6654 .loop_break = sched_nr_migrate_break,
6657 .tasks = LIST_HEAD_INIT(env.tasks),
6661 * For NEWLY_IDLE load_balancing, we don't need to consider
6662 * other cpus in our group
6664 if (idle == CPU_NEWLY_IDLE)
6665 env.dst_grpmask = NULL;
6667 cpumask_copy(cpus, cpu_active_mask);
6669 schedstat_inc(sd, lb_count[idle]);
6672 if (!should_we_balance(&env)) {
6673 *continue_balancing = 0;
6677 group = find_busiest_group(&env);
6679 schedstat_inc(sd, lb_nobusyg[idle]);
6683 busiest = find_busiest_queue(&env, group);
6685 schedstat_inc(sd, lb_nobusyq[idle]);
6689 BUG_ON(busiest == env.dst_rq);
6691 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6694 if (busiest->nr_running > 1) {
6696 * Attempt to move tasks. If find_busiest_group has found
6697 * an imbalance but busiest->nr_running <= 1, the group is
6698 * still unbalanced. ld_moved simply stays zero, so it is
6699 * correctly treated as an imbalance.
6701 env.flags |= LBF_ALL_PINNED;
6702 env.src_cpu = busiest->cpu;
6703 env.src_rq = busiest;
6704 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6707 raw_spin_lock_irqsave(&busiest->lock, flags);
6710 * cur_ld_moved - load moved in current iteration
6711 * ld_moved - cumulative load moved across iterations
6713 cur_ld_moved = detach_tasks(&env);
6716 * We've detached some tasks from busiest_rq. Every
6717 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6718 * unlock busiest->lock, and we are able to be sure
6719 * that nobody can manipulate the tasks in parallel.
6720 * See task_rq_lock() family for the details.
6723 raw_spin_unlock(&busiest->lock);
6727 ld_moved += cur_ld_moved;
6730 local_irq_restore(flags);
6732 if (env.flags & LBF_NEED_BREAK) {
6733 env.flags &= ~LBF_NEED_BREAK;
6738 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6739 * us and move them to an alternate dst_cpu in our sched_group
6740 * where they can run. The upper limit on how many times we
6741 * iterate on same src_cpu is dependent on number of cpus in our
6744 * This changes load balance semantics a bit on who can move
6745 * load to a given_cpu. In addition to the given_cpu itself
6746 * (or a ilb_cpu acting on its behalf where given_cpu is
6747 * nohz-idle), we now have balance_cpu in a position to move
6748 * load to given_cpu. In rare situations, this may cause
6749 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6750 * _independently_ and at _same_ time to move some load to
6751 * given_cpu) causing exceess load to be moved to given_cpu.
6752 * This however should not happen so much in practice and
6753 * moreover subsequent load balance cycles should correct the
6754 * excess load moved.
6756 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6758 /* Prevent to re-select dst_cpu via env's cpus */
6759 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6761 env.dst_rq = cpu_rq(env.new_dst_cpu);
6762 env.dst_cpu = env.new_dst_cpu;
6763 env.flags &= ~LBF_DST_PINNED;
6765 env.loop_break = sched_nr_migrate_break;
6768 * Go back to "more_balance" rather than "redo" since we
6769 * need to continue with same src_cpu.
6775 * We failed to reach balance because of affinity.
6778 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
6780 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
6781 *group_imbalance = 1;
6784 /* All tasks on this runqueue were pinned by CPU affinity */
6785 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6786 cpumask_clear_cpu(cpu_of(busiest), cpus);
6787 if (!cpumask_empty(cpus)) {
6789 env.loop_break = sched_nr_migrate_break;
6792 goto out_all_pinned;
6797 schedstat_inc(sd, lb_failed[idle]);
6799 * Increment the failure counter only on periodic balance.
6800 * We do not want newidle balance, which can be very
6801 * frequent, pollute the failure counter causing
6802 * excessive cache_hot migrations and active balances.
6804 if (idle != CPU_NEWLY_IDLE)
6805 sd->nr_balance_failed++;
6807 if (need_active_balance(&env)) {
6808 raw_spin_lock_irqsave(&busiest->lock, flags);
6810 /* don't kick the active_load_balance_cpu_stop,
6811 * if the curr task on busiest cpu can't be
6814 if (!cpumask_test_cpu(this_cpu,
6815 tsk_cpus_allowed(busiest->curr))) {
6816 raw_spin_unlock_irqrestore(&busiest->lock,
6818 env.flags |= LBF_ALL_PINNED;
6819 goto out_one_pinned;
6823 * ->active_balance synchronizes accesses to
6824 * ->active_balance_work. Once set, it's cleared
6825 * only after active load balance is finished.
6827 if (!busiest->active_balance) {
6828 busiest->active_balance = 1;
6829 busiest->push_cpu = this_cpu;
6832 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6834 if (active_balance) {
6835 stop_one_cpu_nowait(cpu_of(busiest),
6836 active_load_balance_cpu_stop, busiest,
6837 &busiest->active_balance_work);
6841 * We've kicked active balancing, reset the failure
6844 sd->nr_balance_failed = sd->cache_nice_tries+1;
6847 sd->nr_balance_failed = 0;
6849 if (likely(!active_balance)) {
6850 /* We were unbalanced, so reset the balancing interval */
6851 sd->balance_interval = sd->min_interval;
6854 * If we've begun active balancing, start to back off. This
6855 * case may not be covered by the all_pinned logic if there
6856 * is only 1 task on the busy runqueue (because we don't call
6859 if (sd->balance_interval < sd->max_interval)
6860 sd->balance_interval *= 2;
6867 * We reach balance although we may have faced some affinity
6868 * constraints. Clear the imbalance flag if it was set.
6871 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
6873 if (*group_imbalance)
6874 *group_imbalance = 0;
6879 * We reach balance because all tasks are pinned at this level so
6880 * we can't migrate them. Let the imbalance flag set so parent level
6881 * can try to migrate them.
6883 schedstat_inc(sd, lb_balanced[idle]);
6885 sd->nr_balance_failed = 0;
6888 /* tune up the balancing interval */
6889 if (((env.flags & LBF_ALL_PINNED) &&
6890 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6891 (sd->balance_interval < sd->max_interval))
6892 sd->balance_interval *= 2;
6899 static inline unsigned long
6900 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
6902 unsigned long interval = sd->balance_interval;
6905 interval *= sd->busy_factor;
6907 /* scale ms to jiffies */
6908 interval = msecs_to_jiffies(interval);
6909 interval = clamp(interval, 1UL, max_load_balance_interval);
6915 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
6917 unsigned long interval, next;
6919 interval = get_sd_balance_interval(sd, cpu_busy);
6920 next = sd->last_balance + interval;
6922 if (time_after(*next_balance, next))
6923 *next_balance = next;
6927 * idle_balance is called by schedule() if this_cpu is about to become
6928 * idle. Attempts to pull tasks from other CPUs.
6930 static int idle_balance(struct rq *this_rq)
6932 unsigned long next_balance = jiffies + HZ;
6933 int this_cpu = this_rq->cpu;
6934 struct sched_domain *sd;
6935 int pulled_task = 0;
6938 idle_enter_fair(this_rq);
6941 * We must set idle_stamp _before_ calling idle_balance(), such that we
6942 * measure the duration of idle_balance() as idle time.
6944 this_rq->idle_stamp = rq_clock(this_rq);
6946 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
6947 !this_rq->rd->overload) {
6949 sd = rcu_dereference_check_sched_domain(this_rq->sd);
6951 update_next_balance(sd, 0, &next_balance);
6958 * Drop the rq->lock, but keep IRQ/preempt disabled.
6960 raw_spin_unlock(&this_rq->lock);
6962 update_blocked_averages(this_cpu);
6964 for_each_domain(this_cpu, sd) {
6965 int continue_balancing = 1;
6966 u64 t0, domain_cost;
6968 if (!(sd->flags & SD_LOAD_BALANCE))
6971 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
6972 update_next_balance(sd, 0, &next_balance);
6976 if (sd->flags & SD_BALANCE_NEWIDLE) {
6977 t0 = sched_clock_cpu(this_cpu);
6979 pulled_task = load_balance(this_cpu, this_rq,
6981 &continue_balancing);
6983 domain_cost = sched_clock_cpu(this_cpu) - t0;
6984 if (domain_cost > sd->max_newidle_lb_cost)
6985 sd->max_newidle_lb_cost = domain_cost;
6987 curr_cost += domain_cost;
6990 update_next_balance(sd, 0, &next_balance);
6993 * Stop searching for tasks to pull if there are
6994 * now runnable tasks on this rq.
6996 if (pulled_task || this_rq->nr_running > 0)
7001 raw_spin_lock(&this_rq->lock);
7003 if (curr_cost > this_rq->max_idle_balance_cost)
7004 this_rq->max_idle_balance_cost = curr_cost;
7007 * While browsing the domains, we released the rq lock, a task could
7008 * have been enqueued in the meantime. Since we're not going idle,
7009 * pretend we pulled a task.
7011 if (this_rq->cfs.h_nr_running && !pulled_task)
7015 /* Move the next balance forward */
7016 if (time_after(this_rq->next_balance, next_balance))
7017 this_rq->next_balance = next_balance;
7019 /* Is there a task of a high priority class? */
7020 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7024 idle_exit_fair(this_rq);
7025 this_rq->idle_stamp = 0;
7032 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7033 * running tasks off the busiest CPU onto idle CPUs. It requires at
7034 * least 1 task to be running on each physical CPU where possible, and
7035 * avoids physical / logical imbalances.
7037 static int active_load_balance_cpu_stop(void *data)
7039 struct rq *busiest_rq = data;
7040 int busiest_cpu = cpu_of(busiest_rq);
7041 int target_cpu = busiest_rq->push_cpu;
7042 struct rq *target_rq = cpu_rq(target_cpu);
7043 struct sched_domain *sd;
7044 struct task_struct *p = NULL;
7046 raw_spin_lock_irq(&busiest_rq->lock);
7048 /* make sure the requested cpu hasn't gone down in the meantime */
7049 if (unlikely(busiest_cpu != smp_processor_id() ||
7050 !busiest_rq->active_balance))
7053 /* Is there any task to move? */
7054 if (busiest_rq->nr_running <= 1)
7058 * This condition is "impossible", if it occurs
7059 * we need to fix it. Originally reported by
7060 * Bjorn Helgaas on a 128-cpu setup.
7062 BUG_ON(busiest_rq == target_rq);
7064 /* Search for an sd spanning us and the target CPU. */
7066 for_each_domain(target_cpu, sd) {
7067 if ((sd->flags & SD_LOAD_BALANCE) &&
7068 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7073 struct lb_env env = {
7075 .dst_cpu = target_cpu,
7076 .dst_rq = target_rq,
7077 .src_cpu = busiest_rq->cpu,
7078 .src_rq = busiest_rq,
7082 schedstat_inc(sd, alb_count);
7084 p = detach_one_task(&env);
7086 schedstat_inc(sd, alb_pushed);
7088 schedstat_inc(sd, alb_failed);
7092 busiest_rq->active_balance = 0;
7093 raw_spin_unlock(&busiest_rq->lock);
7096 attach_one_task(target_rq, p);
7103 static inline int on_null_domain(struct rq *rq)
7105 return unlikely(!rcu_dereference_sched(rq->sd));
7108 #ifdef CONFIG_NO_HZ_COMMON
7110 * idle load balancing details
7111 * - When one of the busy CPUs notice that there may be an idle rebalancing
7112 * needed, they will kick the idle load balancer, which then does idle
7113 * load balancing for all the idle CPUs.
7116 cpumask_var_t idle_cpus_mask;
7118 unsigned long next_balance; /* in jiffy units */
7119 } nohz ____cacheline_aligned;
7121 static inline int find_new_ilb(void)
7123 int ilb = cpumask_first(nohz.idle_cpus_mask);
7125 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7132 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7133 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7134 * CPU (if there is one).
7136 static void nohz_balancer_kick(void)
7140 nohz.next_balance++;
7142 ilb_cpu = find_new_ilb();
7144 if (ilb_cpu >= nr_cpu_ids)
7147 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7150 * Use smp_send_reschedule() instead of resched_cpu().
7151 * This way we generate a sched IPI on the target cpu which
7152 * is idle. And the softirq performing nohz idle load balance
7153 * will be run before returning from the IPI.
7155 smp_send_reschedule(ilb_cpu);
7159 static inline void nohz_balance_exit_idle(int cpu)
7161 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7163 * Completely isolated CPUs don't ever set, so we must test.
7165 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7166 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7167 atomic_dec(&nohz.nr_cpus);
7169 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7173 static inline void set_cpu_sd_state_busy(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_inc(&sd->groups->sgc->nr_busy_cpus);
7190 void set_cpu_sd_state_idle(void)
7192 struct sched_domain *sd;
7193 int cpu = smp_processor_id();
7196 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7198 if (!sd || sd->nohz_idle)
7202 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7208 * This routine will record that the cpu is going idle with tick stopped.
7209 * This info will be used in performing idle load balancing in the future.
7211 void nohz_balance_enter_idle(int cpu)
7214 * If this cpu is going down, then nothing needs to be done.
7216 if (!cpu_active(cpu))
7219 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7223 * If we're a completely isolated CPU, we don't play.
7225 if (on_null_domain(cpu_rq(cpu)))
7228 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7229 atomic_inc(&nohz.nr_cpus);
7230 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7233 static int sched_ilb_notifier(struct notifier_block *nfb,
7234 unsigned long action, void *hcpu)
7236 switch (action & ~CPU_TASKS_FROZEN) {
7238 nohz_balance_exit_idle(smp_processor_id());
7246 static DEFINE_SPINLOCK(balancing);
7249 * Scale the max load_balance interval with the number of CPUs in the system.
7250 * This trades load-balance latency on larger machines for less cross talk.
7252 void update_max_interval(void)
7254 max_load_balance_interval = HZ*num_online_cpus()/10;
7258 * It checks each scheduling domain to see if it is due to be balanced,
7259 * and initiates a balancing operation if so.
7261 * Balancing parameters are set up in init_sched_domains.
7263 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7265 int continue_balancing = 1;
7267 unsigned long interval;
7268 struct sched_domain *sd;
7269 /* Earliest time when we have to do rebalance again */
7270 unsigned long next_balance = jiffies + 60*HZ;
7271 int update_next_balance = 0;
7272 int need_serialize, need_decay = 0;
7275 update_blocked_averages(cpu);
7278 for_each_domain(cpu, sd) {
7280 * Decay the newidle max times here because this is a regular
7281 * visit to all the domains. Decay ~1% per second.
7283 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7284 sd->max_newidle_lb_cost =
7285 (sd->max_newidle_lb_cost * 253) / 256;
7286 sd->next_decay_max_lb_cost = jiffies + HZ;
7289 max_cost += sd->max_newidle_lb_cost;
7291 if (!(sd->flags & SD_LOAD_BALANCE))
7295 * Stop the load balance at this level. There is another
7296 * CPU in our sched group which is doing load balancing more
7299 if (!continue_balancing) {
7305 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7307 need_serialize = sd->flags & SD_SERIALIZE;
7308 if (need_serialize) {
7309 if (!spin_trylock(&balancing))
7313 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7314 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7316 * The LBF_DST_PINNED logic could have changed
7317 * env->dst_cpu, so we can't know our idle
7318 * state even if we migrated tasks. Update it.
7320 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7322 sd->last_balance = jiffies;
7323 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7326 spin_unlock(&balancing);
7328 if (time_after(next_balance, sd->last_balance + interval)) {
7329 next_balance = sd->last_balance + interval;
7330 update_next_balance = 1;
7335 * Ensure the rq-wide value also decays but keep it at a
7336 * reasonable floor to avoid funnies with rq->avg_idle.
7338 rq->max_idle_balance_cost =
7339 max((u64)sysctl_sched_migration_cost, max_cost);
7344 * next_balance will be updated only when there is a need.
7345 * When the cpu is attached to null domain for ex, it will not be
7348 if (likely(update_next_balance))
7349 rq->next_balance = next_balance;
7352 #ifdef CONFIG_NO_HZ_COMMON
7354 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7355 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7357 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7359 int this_cpu = this_rq->cpu;
7363 if (idle != CPU_IDLE ||
7364 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7367 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7368 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7372 * If this cpu gets work to do, stop the load balancing
7373 * work being done for other cpus. Next load
7374 * balancing owner will pick it up.
7379 rq = cpu_rq(balance_cpu);
7382 * If time for next balance is due,
7385 if (time_after_eq(jiffies, rq->next_balance)) {
7386 raw_spin_lock_irq(&rq->lock);
7387 update_rq_clock(rq);
7388 update_idle_cpu_load(rq);
7389 raw_spin_unlock_irq(&rq->lock);
7390 rebalance_domains(rq, CPU_IDLE);
7393 if (time_after(this_rq->next_balance, rq->next_balance))
7394 this_rq->next_balance = rq->next_balance;
7396 nohz.next_balance = this_rq->next_balance;
7398 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7402 * Current heuristic for kicking the idle load balancer in the presence
7403 * of an idle cpu is the system.
7404 * - This rq has more than one task.
7405 * - At any scheduler domain level, this cpu's scheduler group has multiple
7406 * busy cpu's exceeding the group's capacity.
7407 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7408 * domain span are idle.
7410 static inline int nohz_kick_needed(struct rq *rq)
7412 unsigned long now = jiffies;
7413 struct sched_domain *sd;
7414 struct sched_group_capacity *sgc;
7415 int nr_busy, cpu = rq->cpu;
7417 if (unlikely(rq->idle_balance))
7421 * We may be recently in ticked or tickless idle mode. At the first
7422 * busy tick after returning from idle, we will update the busy stats.
7424 set_cpu_sd_state_busy();
7425 nohz_balance_exit_idle(cpu);
7428 * None are in tickless mode and hence no need for NOHZ idle load
7431 if (likely(!atomic_read(&nohz.nr_cpus)))
7434 if (time_before(now, nohz.next_balance))
7437 if (rq->nr_running >= 2)
7441 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7444 sgc = sd->groups->sgc;
7445 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7448 goto need_kick_unlock;
7451 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7453 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7454 sched_domain_span(sd)) < cpu))
7455 goto need_kick_unlock;
7466 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7470 * run_rebalance_domains is triggered when needed from the scheduler tick.
7471 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7473 static void run_rebalance_domains(struct softirq_action *h)
7475 struct rq *this_rq = this_rq();
7476 enum cpu_idle_type idle = this_rq->idle_balance ?
7477 CPU_IDLE : CPU_NOT_IDLE;
7479 rebalance_domains(this_rq, idle);
7482 * If this cpu has a pending nohz_balance_kick, then do the
7483 * balancing on behalf of the other idle cpus whose ticks are
7486 nohz_idle_balance(this_rq, idle);
7490 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7492 void trigger_load_balance(struct rq *rq)
7494 /* Don't need to rebalance while attached to NULL domain */
7495 if (unlikely(on_null_domain(rq)))
7498 if (time_after_eq(jiffies, rq->next_balance))
7499 raise_softirq(SCHED_SOFTIRQ);
7500 #ifdef CONFIG_NO_HZ_COMMON
7501 if (nohz_kick_needed(rq))
7502 nohz_balancer_kick();
7506 static void rq_online_fair(struct rq *rq)
7510 update_runtime_enabled(rq);
7513 static void rq_offline_fair(struct rq *rq)
7517 /* Ensure any throttled groups are reachable by pick_next_task */
7518 unthrottle_offline_cfs_rqs(rq);
7521 #endif /* CONFIG_SMP */
7524 * scheduler tick hitting a task of our scheduling class:
7526 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7528 struct cfs_rq *cfs_rq;
7529 struct sched_entity *se = &curr->se;
7531 for_each_sched_entity(se) {
7532 cfs_rq = cfs_rq_of(se);
7533 entity_tick(cfs_rq, se, queued);
7536 if (numabalancing_enabled)
7537 task_tick_numa(rq, curr);
7539 update_rq_runnable_avg(rq, 1);
7543 * called on fork with the child task as argument from the parent's context
7544 * - child not yet on the tasklist
7545 * - preemption disabled
7547 static void task_fork_fair(struct task_struct *p)
7549 struct cfs_rq *cfs_rq;
7550 struct sched_entity *se = &p->se, *curr;
7551 int this_cpu = smp_processor_id();
7552 struct rq *rq = this_rq();
7553 unsigned long flags;
7555 raw_spin_lock_irqsave(&rq->lock, flags);
7557 update_rq_clock(rq);
7559 cfs_rq = task_cfs_rq(current);
7560 curr = cfs_rq->curr;
7563 * Not only the cpu but also the task_group of the parent might have
7564 * been changed after parent->se.parent,cfs_rq were copied to
7565 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7566 * of child point to valid ones.
7569 __set_task_cpu(p, this_cpu);
7572 update_curr(cfs_rq);
7575 se->vruntime = curr->vruntime;
7576 place_entity(cfs_rq, se, 1);
7578 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7580 * Upon rescheduling, sched_class::put_prev_task() will place
7581 * 'current' within the tree based on its new key value.
7583 swap(curr->vruntime, se->vruntime);
7587 se->vruntime -= cfs_rq->min_vruntime;
7589 raw_spin_unlock_irqrestore(&rq->lock, flags);
7593 * Priority of the task has changed. Check to see if we preempt
7597 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7599 if (!task_on_rq_queued(p))
7603 * Reschedule if we are currently running on this runqueue and
7604 * our priority decreased, or if we are not currently running on
7605 * this runqueue and our priority is higher than the current's
7607 if (rq->curr == p) {
7608 if (p->prio > oldprio)
7611 check_preempt_curr(rq, p, 0);
7614 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7616 struct sched_entity *se = &p->se;
7617 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7620 * Ensure the task's vruntime is normalized, so that when it's
7621 * switched back to the fair class the enqueue_entity(.flags=0) will
7622 * do the right thing.
7624 * If it's queued, then the dequeue_entity(.flags=0) will already
7625 * have normalized the vruntime, if it's !queued, then only when
7626 * the task is sleeping will it still have non-normalized vruntime.
7628 if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) {
7630 * Fix up our vruntime so that the current sleep doesn't
7631 * cause 'unlimited' sleep bonus.
7633 place_entity(cfs_rq, se, 0);
7634 se->vruntime -= cfs_rq->min_vruntime;
7639 * Remove our load from contribution when we leave sched_fair
7640 * and ensure we don't carry in an old decay_count if we
7643 if (se->avg.decay_count) {
7644 __synchronize_entity_decay(se);
7645 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7651 * We switched to the sched_fair class.
7653 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7655 #ifdef CONFIG_FAIR_GROUP_SCHED
7656 struct sched_entity *se = &p->se;
7658 * Since the real-depth could have been changed (only FAIR
7659 * class maintain depth value), reset depth properly.
7661 se->depth = se->parent ? se->parent->depth + 1 : 0;
7663 if (!task_on_rq_queued(p))
7667 * We were most likely switched from sched_rt, so
7668 * kick off the schedule if running, otherwise just see
7669 * if we can still preempt the current task.
7674 check_preempt_curr(rq, p, 0);
7677 /* Account for a task changing its policy or group.
7679 * This routine is mostly called to set cfs_rq->curr field when a task
7680 * migrates between groups/classes.
7682 static void set_curr_task_fair(struct rq *rq)
7684 struct sched_entity *se = &rq->curr->se;
7686 for_each_sched_entity(se) {
7687 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7689 set_next_entity(cfs_rq, se);
7690 /* ensure bandwidth has been allocated on our new cfs_rq */
7691 account_cfs_rq_runtime(cfs_rq, 0);
7695 void init_cfs_rq(struct cfs_rq *cfs_rq)
7697 cfs_rq->tasks_timeline = RB_ROOT;
7698 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7699 #ifndef CONFIG_64BIT
7700 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7703 atomic64_set(&cfs_rq->decay_counter, 1);
7704 atomic_long_set(&cfs_rq->removed_load, 0);
7708 #ifdef CONFIG_FAIR_GROUP_SCHED
7709 static void task_move_group_fair(struct task_struct *p, int queued)
7711 struct sched_entity *se = &p->se;
7712 struct cfs_rq *cfs_rq;
7715 * If the task was not on the rq at the time of this cgroup movement
7716 * it must have been asleep, sleeping tasks keep their ->vruntime
7717 * absolute on their old rq until wakeup (needed for the fair sleeper
7718 * bonus in place_entity()).
7720 * If it was on the rq, we've just 'preempted' it, which does convert
7721 * ->vruntime to a relative base.
7723 * Make sure both cases convert their relative position when migrating
7724 * to another cgroup's rq. This does somewhat interfere with the
7725 * fair sleeper stuff for the first placement, but who cares.
7728 * When !queued, vruntime of the task has usually NOT been normalized.
7729 * But there are some cases where it has already been normalized:
7731 * - Moving a forked child which is waiting for being woken up by
7732 * wake_up_new_task().
7733 * - Moving a task which has been woken up by try_to_wake_up() and
7734 * waiting for actually being woken up by sched_ttwu_pending().
7736 * To prevent boost or penalty in the new cfs_rq caused by delta
7737 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7739 if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7743 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7744 set_task_rq(p, task_cpu(p));
7745 se->depth = se->parent ? se->parent->depth + 1 : 0;
7747 cfs_rq = cfs_rq_of(se);
7748 se->vruntime += cfs_rq->min_vruntime;
7751 * migrate_task_rq_fair() will have removed our previous
7752 * contribution, but we must synchronize for ongoing future
7755 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7756 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7761 void free_fair_sched_group(struct task_group *tg)
7765 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7767 for_each_possible_cpu(i) {
7769 kfree(tg->cfs_rq[i]);
7778 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7780 struct cfs_rq *cfs_rq;
7781 struct sched_entity *se;
7784 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7787 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7791 tg->shares = NICE_0_LOAD;
7793 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7795 for_each_possible_cpu(i) {
7796 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7797 GFP_KERNEL, cpu_to_node(i));
7801 se = kzalloc_node(sizeof(struct sched_entity),
7802 GFP_KERNEL, cpu_to_node(i));
7806 init_cfs_rq(cfs_rq);
7807 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7818 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7820 struct rq *rq = cpu_rq(cpu);
7821 unsigned long flags;
7824 * Only empty task groups can be destroyed; so we can speculatively
7825 * check on_list without danger of it being re-added.
7827 if (!tg->cfs_rq[cpu]->on_list)
7830 raw_spin_lock_irqsave(&rq->lock, flags);
7831 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7832 raw_spin_unlock_irqrestore(&rq->lock, flags);
7835 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7836 struct sched_entity *se, int cpu,
7837 struct sched_entity *parent)
7839 struct rq *rq = cpu_rq(cpu);
7843 init_cfs_rq_runtime(cfs_rq);
7845 tg->cfs_rq[cpu] = cfs_rq;
7848 /* se could be NULL for root_task_group */
7853 se->cfs_rq = &rq->cfs;
7856 se->cfs_rq = parent->my_q;
7857 se->depth = parent->depth + 1;
7861 /* guarantee group entities always have weight */
7862 update_load_set(&se->load, NICE_0_LOAD);
7863 se->parent = parent;
7866 static DEFINE_MUTEX(shares_mutex);
7868 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7871 unsigned long flags;
7874 * We can't change the weight of the root cgroup.
7879 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7881 mutex_lock(&shares_mutex);
7882 if (tg->shares == shares)
7885 tg->shares = shares;
7886 for_each_possible_cpu(i) {
7887 struct rq *rq = cpu_rq(i);
7888 struct sched_entity *se;
7891 /* Propagate contribution to hierarchy */
7892 raw_spin_lock_irqsave(&rq->lock, flags);
7894 /* Possible calls to update_curr() need rq clock */
7895 update_rq_clock(rq);
7896 for_each_sched_entity(se)
7897 update_cfs_shares(group_cfs_rq(se));
7898 raw_spin_unlock_irqrestore(&rq->lock, flags);
7902 mutex_unlock(&shares_mutex);
7905 #else /* CONFIG_FAIR_GROUP_SCHED */
7907 void free_fair_sched_group(struct task_group *tg) { }
7909 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7914 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7916 #endif /* CONFIG_FAIR_GROUP_SCHED */
7919 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7921 struct sched_entity *se = &task->se;
7922 unsigned int rr_interval = 0;
7925 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7928 if (rq->cfs.load.weight)
7929 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7935 * All the scheduling class methods:
7937 const struct sched_class fair_sched_class = {
7938 .next = &idle_sched_class,
7939 .enqueue_task = enqueue_task_fair,
7940 .dequeue_task = dequeue_task_fair,
7941 .yield_task = yield_task_fair,
7942 .yield_to_task = yield_to_task_fair,
7944 .check_preempt_curr = check_preempt_wakeup,
7946 .pick_next_task = pick_next_task_fair,
7947 .put_prev_task = put_prev_task_fair,
7950 .select_task_rq = select_task_rq_fair,
7951 .migrate_task_rq = migrate_task_rq_fair,
7953 .rq_online = rq_online_fair,
7954 .rq_offline = rq_offline_fair,
7956 .task_waking = task_waking_fair,
7959 .set_curr_task = set_curr_task_fair,
7960 .task_tick = task_tick_fair,
7961 .task_fork = task_fork_fair,
7963 .prio_changed = prio_changed_fair,
7964 .switched_from = switched_from_fair,
7965 .switched_to = switched_to_fair,
7967 .get_rr_interval = get_rr_interval_fair,
7969 #ifdef CONFIG_FAIR_GROUP_SCHED
7970 .task_move_group = task_move_group_fair,
7974 #ifdef CONFIG_SCHED_DEBUG
7975 void print_cfs_stats(struct seq_file *m, int cpu)
7977 struct cfs_rq *cfs_rq;
7980 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7981 print_cfs_rq(m, cpu, cfs_rq);
7986 __init void init_sched_fair_class(void)
7989 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7991 #ifdef CONFIG_NO_HZ_COMMON
7992 nohz.next_balance = jiffies;
7993 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7994 cpu_notifier(sched_ilb_notifier, 0);