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
23 #include <linux/sched.h>
24 #include <linux/latencytop.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency = 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity = 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency = 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static unsigned int get_update_sysctl_factor(void)
146 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling) {
150 case SCHED_TUNABLESCALING_NONE:
153 case SCHED_TUNABLESCALING_LINEAR:
156 case SCHED_TUNABLESCALING_LOG:
158 factor = 1 + ilog2(cpus);
165 static void update_sysctl(void)
167 unsigned int factor = get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity);
172 SET_SYSCTL(sched_latency);
173 SET_SYSCTL(sched_wakeup_granularity);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight *lw)
189 if (likely(lw->inv_weight))
192 w = scale_load_down(lw->weight);
194 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
196 else if (unlikely(!w))
197 lw->inv_weight = WMULT_CONST;
199 lw->inv_weight = WMULT_CONST / w;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
216 u64 fact = scale_load_down(weight);
217 int shift = WMULT_SHIFT;
219 __update_inv_weight(lw);
221 if (unlikely(fact >> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact = (u64)(u32)fact * lw->inv_weight;
236 return mul_u64_u32_shr(delta_exec, fact, shift);
240 const struct sched_class fair_sched_class;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct *task_of(struct sched_entity *se)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se));
262 return container_of(se, struct task_struct, se);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
310 if (cfs_rq->on_list) {
311 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq *
322 is_same_group(struct sched_entity *se, struct sched_entity *pse)
324 if (se->cfs_rq == pse->cfs_rq)
330 static inline struct sched_entity *parent_entity(struct sched_entity *se)
336 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
338 int se_depth, pse_depth;
341 * preemption test can be made between sibling entities who are in the
342 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343 * both tasks until we find their ancestors who are siblings of common
347 /* First walk up until both entities are at same depth */
348 se_depth = (*se)->depth;
349 pse_depth = (*pse)->depth;
351 while (se_depth > pse_depth) {
353 *se = parent_entity(*se);
356 while (pse_depth > se_depth) {
358 *pse = parent_entity(*pse);
361 while (!is_same_group(*se, *pse)) {
362 *se = parent_entity(*se);
363 *pse = parent_entity(*pse);
367 #else /* !CONFIG_FAIR_GROUP_SCHED */
369 static inline struct task_struct *task_of(struct sched_entity *se)
371 return container_of(se, struct task_struct, se);
374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
376 return container_of(cfs_rq, struct rq, cfs);
379 #define entity_is_task(se) 1
381 #define for_each_sched_entity(se) \
382 for (; se; se = NULL)
384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
386 return &task_rq(p)->cfs;
389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
391 struct task_struct *p = task_of(se);
392 struct rq *rq = task_rq(p);
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
414 static inline struct sched_entity *parent_entity(struct sched_entity *se)
420 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
424 #endif /* CONFIG_FAIR_GROUP_SCHED */
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
429 /**************************************************************
430 * Scheduling class tree data structure manipulation methods:
433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
435 s64 delta = (s64)(vruntime - max_vruntime);
437 max_vruntime = vruntime;
442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
444 s64 delta = (s64)(vruntime - min_vruntime);
446 min_vruntime = vruntime;
451 static inline int entity_before(struct sched_entity *a,
452 struct sched_entity *b)
454 return (s64)(a->vruntime - b->vruntime) < 0;
457 static void update_min_vruntime(struct cfs_rq *cfs_rq)
459 u64 vruntime = cfs_rq->min_vruntime;
462 vruntime = cfs_rq->curr->vruntime;
464 if (cfs_rq->rb_leftmost) {
465 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
470 vruntime = se->vruntime;
472 vruntime = min_vruntime(vruntime, se->vruntime);
475 /* ensure we never gain time by being placed backwards. */
476 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
479 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484 * Enqueue an entity into the rb-tree:
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
488 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489 struct rb_node *parent = NULL;
490 struct sched_entity *entry;
494 * Find the right place in the rbtree:
498 entry = rb_entry(parent, struct sched_entity, run_node);
500 * We dont care about collisions. Nodes with
501 * the same key stay together.
503 if (entity_before(se, entry)) {
504 link = &parent->rb_left;
506 link = &parent->rb_right;
512 * Maintain a cache of leftmost tree entries (it is frequently
516 cfs_rq->rb_leftmost = &se->run_node;
518 rb_link_node(&se->run_node, parent, link);
519 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
524 if (cfs_rq->rb_leftmost == &se->run_node) {
525 struct rb_node *next_node;
527 next_node = rb_next(&se->run_node);
528 cfs_rq->rb_leftmost = next_node;
531 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
536 struct rb_node *left = cfs_rq->rb_leftmost;
541 return rb_entry(left, struct sched_entity, run_node);
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
546 struct rb_node *next = rb_next(&se->run_node);
551 return rb_entry(next, struct sched_entity, run_node);
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
557 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
562 return rb_entry(last, struct sched_entity, run_node);
565 /**************************************************************
566 * Scheduling class statistics methods:
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570 void __user *buffer, size_t *lenp,
573 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574 unsigned int factor = get_update_sysctl_factor();
579 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580 sysctl_sched_min_granularity);
582 #define WRT_SYSCTL(name) \
583 (normalized_sysctl_##name = sysctl_##name / (factor))
584 WRT_SYSCTL(sched_min_granularity);
585 WRT_SYSCTL(sched_latency);
586 WRT_SYSCTL(sched_wakeup_granularity);
596 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
598 if (unlikely(se->load.weight != NICE_0_LOAD))
599 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
605 * The idea is to set a period in which each task runs once.
607 * When there are too many tasks (sched_nr_latency) we have to stretch
608 * this period because otherwise the slices get too small.
610 * p = (nr <= nl) ? l : l*nr/nl
612 static u64 __sched_period(unsigned long nr_running)
614 if (unlikely(nr_running > sched_nr_latency))
615 return nr_running * sysctl_sched_min_granularity;
617 return sysctl_sched_latency;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
630 for_each_sched_entity(se) {
631 struct load_weight *load;
632 struct load_weight lw;
634 cfs_rq = cfs_rq_of(se);
635 load = &cfs_rq->load;
637 if (unlikely(!se->on_rq)) {
640 update_load_add(&lw, se->load.weight);
643 slice = __calc_delta(slice, se->load.weight, load);
649 * We calculate the vruntime slice of a to-be-inserted task.
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
655 return calc_delta_fair(sched_slice(cfs_rq, se), se);
659 static int select_idle_sibling(struct task_struct *p, int cpu);
660 static unsigned long task_h_load(struct task_struct *p);
663 * We choose a half-life close to 1 scheduling period.
664 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
665 * dependent on this value.
667 #define LOAD_AVG_PERIOD 32
668 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
669 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
671 /* Give new sched_entity start runnable values to heavy its load in infant time */
672 void init_entity_runnable_average(struct sched_entity *se)
674 struct sched_avg *sa = &se->avg;
676 sa->last_update_time = 0;
678 * sched_avg's period_contrib should be strictly less then 1024, so
679 * we give it 1023 to make sure it is almost a period (1024us), and
680 * will definitely be update (after enqueue).
682 sa->period_contrib = 1023;
683 sa->load_avg = scale_load_down(se->load.weight);
684 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
685 sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
686 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
687 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
690 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
691 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
693 void init_entity_runnable_average(struct sched_entity *se)
699 * Update the current task's runtime statistics.
701 static void update_curr(struct cfs_rq *cfs_rq)
703 struct sched_entity *curr = cfs_rq->curr;
704 u64 now = rq_clock_task(rq_of(cfs_rq));
710 delta_exec = now - curr->exec_start;
711 if (unlikely((s64)delta_exec <= 0))
714 curr->exec_start = now;
716 schedstat_set(curr->statistics.exec_max,
717 max(delta_exec, curr->statistics.exec_max));
719 curr->sum_exec_runtime += delta_exec;
720 schedstat_add(cfs_rq, exec_clock, delta_exec);
722 curr->vruntime += calc_delta_fair(delta_exec, curr);
723 update_min_vruntime(cfs_rq);
725 if (entity_is_task(curr)) {
726 struct task_struct *curtask = task_of(curr);
728 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
729 cpuacct_charge(curtask, delta_exec);
730 account_group_exec_runtime(curtask, delta_exec);
733 account_cfs_rq_runtime(cfs_rq, delta_exec);
736 static void update_curr_fair(struct rq *rq)
738 update_curr(cfs_rq_of(&rq->curr->se));
741 #ifdef CONFIG_SCHEDSTATS
743 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
745 u64 wait_start = rq_clock(rq_of(cfs_rq));
747 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
748 likely(wait_start > se->statistics.wait_start))
749 wait_start -= se->statistics.wait_start;
751 se->statistics.wait_start = wait_start;
755 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
757 struct task_struct *p;
760 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start;
762 if (entity_is_task(se)) {
764 if (task_on_rq_migrating(p)) {
766 * Preserve migrating task's wait time so wait_start
767 * time stamp can be adjusted to accumulate wait time
768 * prior to migration.
770 se->statistics.wait_start = delta;
773 trace_sched_stat_wait(p, delta);
776 se->statistics.wait_max = max(se->statistics.wait_max, delta);
777 se->statistics.wait_count++;
778 se->statistics.wait_sum += delta;
779 se->statistics.wait_start = 0;
783 * Task is being enqueued - update stats:
786 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
789 * Are we enqueueing a waiting task? (for current tasks
790 * a dequeue/enqueue event is a NOP)
792 if (se != cfs_rq->curr)
793 update_stats_wait_start(cfs_rq, se);
797 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
800 * Mark the end of the wait period if dequeueing a
803 if (se != cfs_rq->curr)
804 update_stats_wait_end(cfs_rq, se);
806 if (flags & DEQUEUE_SLEEP) {
807 if (entity_is_task(se)) {
808 struct task_struct *tsk = task_of(se);
810 if (tsk->state & TASK_INTERRUPTIBLE)
811 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
812 if (tsk->state & TASK_UNINTERRUPTIBLE)
813 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
820 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
825 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
830 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
835 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
841 * We are picking a new current task - update its stats:
844 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
847 * We are starting a new run period:
849 se->exec_start = rq_clock_task(rq_of(cfs_rq));
852 /**************************************************
853 * Scheduling class queueing methods:
856 #ifdef CONFIG_NUMA_BALANCING
858 * Approximate time to scan a full NUMA task in ms. The task scan period is
859 * calculated based on the tasks virtual memory size and
860 * numa_balancing_scan_size.
862 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
863 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
865 /* Portion of address space to scan in MB */
866 unsigned int sysctl_numa_balancing_scan_size = 256;
868 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
869 unsigned int sysctl_numa_balancing_scan_delay = 1000;
871 static unsigned int task_nr_scan_windows(struct task_struct *p)
873 unsigned long rss = 0;
874 unsigned long nr_scan_pages;
877 * Calculations based on RSS as non-present and empty pages are skipped
878 * by the PTE scanner and NUMA hinting faults should be trapped based
881 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
882 rss = get_mm_rss(p->mm);
886 rss = round_up(rss, nr_scan_pages);
887 return rss / nr_scan_pages;
890 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
891 #define MAX_SCAN_WINDOW 2560
893 static unsigned int task_scan_min(struct task_struct *p)
895 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
896 unsigned int scan, floor;
897 unsigned int windows = 1;
899 if (scan_size < MAX_SCAN_WINDOW)
900 windows = MAX_SCAN_WINDOW / scan_size;
901 floor = 1000 / windows;
903 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
904 return max_t(unsigned int, floor, scan);
907 static unsigned int task_scan_max(struct task_struct *p)
909 unsigned int smin = task_scan_min(p);
912 /* Watch for min being lower than max due to floor calculations */
913 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
914 return max(smin, smax);
917 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
919 rq->nr_numa_running += (p->numa_preferred_nid != -1);
920 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
923 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
925 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
926 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
932 spinlock_t lock; /* nr_tasks, tasks */
938 unsigned long total_faults;
939 unsigned long max_faults_cpu;
941 * Faults_cpu is used to decide whether memory should move
942 * towards the CPU. As a consequence, these stats are weighted
943 * more by CPU use than by memory faults.
945 unsigned long *faults_cpu;
946 unsigned long faults[0];
949 /* Shared or private faults. */
950 #define NR_NUMA_HINT_FAULT_TYPES 2
952 /* Memory and CPU locality */
953 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
955 /* Averaged statistics, and temporary buffers. */
956 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
958 pid_t task_numa_group_id(struct task_struct *p)
960 return p->numa_group ? p->numa_group->gid : 0;
964 * The averaged statistics, shared & private, memory & cpu,
965 * occupy the first half of the array. The second half of the
966 * array is for current counters, which are averaged into the
967 * first set by task_numa_placement.
969 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
971 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
974 static inline unsigned long task_faults(struct task_struct *p, int nid)
979 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
980 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
983 static inline unsigned long group_faults(struct task_struct *p, int nid)
988 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
989 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
992 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
994 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
995 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
999 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1000 * considered part of a numa group's pseudo-interleaving set. Migrations
1001 * between these nodes are slowed down, to allow things to settle down.
1003 #define ACTIVE_NODE_FRACTION 3
1005 static bool numa_is_active_node(int nid, struct numa_group *ng)
1007 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1010 /* Handle placement on systems where not all nodes are directly connected. */
1011 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1012 int maxdist, bool task)
1014 unsigned long score = 0;
1018 * All nodes are directly connected, and the same distance
1019 * from each other. No need for fancy placement algorithms.
1021 if (sched_numa_topology_type == NUMA_DIRECT)
1025 * This code is called for each node, introducing N^2 complexity,
1026 * which should be ok given the number of nodes rarely exceeds 8.
1028 for_each_online_node(node) {
1029 unsigned long faults;
1030 int dist = node_distance(nid, node);
1033 * The furthest away nodes in the system are not interesting
1034 * for placement; nid was already counted.
1036 if (dist == sched_max_numa_distance || node == nid)
1040 * On systems with a backplane NUMA topology, compare groups
1041 * of nodes, and move tasks towards the group with the most
1042 * memory accesses. When comparing two nodes at distance
1043 * "hoplimit", only nodes closer by than "hoplimit" are part
1044 * of each group. Skip other nodes.
1046 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1050 /* Add up the faults from nearby nodes. */
1052 faults = task_faults(p, node);
1054 faults = group_faults(p, node);
1057 * On systems with a glueless mesh NUMA topology, there are
1058 * no fixed "groups of nodes". Instead, nodes that are not
1059 * directly connected bounce traffic through intermediate
1060 * nodes; a numa_group can occupy any set of nodes.
1061 * The further away a node is, the less the faults count.
1062 * This seems to result in good task placement.
1064 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1065 faults *= (sched_max_numa_distance - dist);
1066 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1076 * These return the fraction of accesses done by a particular task, or
1077 * task group, on a particular numa node. The group weight is given a
1078 * larger multiplier, in order to group tasks together that are almost
1079 * evenly spread out between numa nodes.
1081 static inline unsigned long task_weight(struct task_struct *p, int nid,
1084 unsigned long faults, total_faults;
1086 if (!p->numa_faults)
1089 total_faults = p->total_numa_faults;
1094 faults = task_faults(p, nid);
1095 faults += score_nearby_nodes(p, nid, dist, true);
1097 return 1000 * faults / total_faults;
1100 static inline unsigned long group_weight(struct task_struct *p, int nid,
1103 unsigned long faults, total_faults;
1108 total_faults = p->numa_group->total_faults;
1113 faults = group_faults(p, nid);
1114 faults += score_nearby_nodes(p, nid, dist, false);
1116 return 1000 * faults / total_faults;
1119 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1120 int src_nid, int dst_cpu)
1122 struct numa_group *ng = p->numa_group;
1123 int dst_nid = cpu_to_node(dst_cpu);
1124 int last_cpupid, this_cpupid;
1126 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1129 * Multi-stage node selection is used in conjunction with a periodic
1130 * migration fault to build a temporal task<->page relation. By using
1131 * a two-stage filter we remove short/unlikely relations.
1133 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1134 * a task's usage of a particular page (n_p) per total usage of this
1135 * page (n_t) (in a given time-span) to a probability.
1137 * Our periodic faults will sample this probability and getting the
1138 * same result twice in a row, given these samples are fully
1139 * independent, is then given by P(n)^2, provided our sample period
1140 * is sufficiently short compared to the usage pattern.
1142 * This quadric squishes small probabilities, making it less likely we
1143 * act on an unlikely task<->page relation.
1145 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1146 if (!cpupid_pid_unset(last_cpupid) &&
1147 cpupid_to_nid(last_cpupid) != dst_nid)
1150 /* Always allow migrate on private faults */
1151 if (cpupid_match_pid(p, last_cpupid))
1154 /* A shared fault, but p->numa_group has not been set up yet. */
1159 * Destination node is much more heavily used than the source
1160 * node? Allow migration.
1162 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1163 ACTIVE_NODE_FRACTION)
1167 * Distribute memory according to CPU & memory use on each node,
1168 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1170 * faults_cpu(dst) 3 faults_cpu(src)
1171 * --------------- * - > ---------------
1172 * faults_mem(dst) 4 faults_mem(src)
1174 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1175 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1178 static unsigned long weighted_cpuload(const int cpu);
1179 static unsigned long source_load(int cpu, int type);
1180 static unsigned long target_load(int cpu, int type);
1181 static unsigned long capacity_of(int cpu);
1182 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1184 /* Cached statistics for all CPUs within a node */
1186 unsigned long nr_running;
1189 /* Total compute capacity of CPUs on a node */
1190 unsigned long compute_capacity;
1192 /* Approximate capacity in terms of runnable tasks on a node */
1193 unsigned long task_capacity;
1194 int has_free_capacity;
1198 * XXX borrowed from update_sg_lb_stats
1200 static void update_numa_stats(struct numa_stats *ns, int nid)
1202 int smt, cpu, cpus = 0;
1203 unsigned long capacity;
1205 memset(ns, 0, sizeof(*ns));
1206 for_each_cpu(cpu, cpumask_of_node(nid)) {
1207 struct rq *rq = cpu_rq(cpu);
1209 ns->nr_running += rq->nr_running;
1210 ns->load += weighted_cpuload(cpu);
1211 ns->compute_capacity += capacity_of(cpu);
1217 * If we raced with hotplug and there are no CPUs left in our mask
1218 * the @ns structure is NULL'ed and task_numa_compare() will
1219 * not find this node attractive.
1221 * We'll either bail at !has_free_capacity, or we'll detect a huge
1222 * imbalance and bail there.
1227 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1228 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1229 capacity = cpus / smt; /* cores */
1231 ns->task_capacity = min_t(unsigned, capacity,
1232 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1233 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1236 struct task_numa_env {
1237 struct task_struct *p;
1239 int src_cpu, src_nid;
1240 int dst_cpu, dst_nid;
1242 struct numa_stats src_stats, dst_stats;
1247 struct task_struct *best_task;
1252 static void task_numa_assign(struct task_numa_env *env,
1253 struct task_struct *p, long imp)
1256 put_task_struct(env->best_task);
1259 env->best_imp = imp;
1260 env->best_cpu = env->dst_cpu;
1263 static bool load_too_imbalanced(long src_load, long dst_load,
1264 struct task_numa_env *env)
1267 long orig_src_load, orig_dst_load;
1268 long src_capacity, dst_capacity;
1271 * The load is corrected for the CPU capacity available on each node.
1274 * ------------ vs ---------
1275 * src_capacity dst_capacity
1277 src_capacity = env->src_stats.compute_capacity;
1278 dst_capacity = env->dst_stats.compute_capacity;
1280 /* We care about the slope of the imbalance, not the direction. */
1281 if (dst_load < src_load)
1282 swap(dst_load, src_load);
1284 /* Is the difference below the threshold? */
1285 imb = dst_load * src_capacity * 100 -
1286 src_load * dst_capacity * env->imbalance_pct;
1291 * The imbalance is above the allowed threshold.
1292 * Compare it with the old imbalance.
1294 orig_src_load = env->src_stats.load;
1295 orig_dst_load = env->dst_stats.load;
1297 if (orig_dst_load < orig_src_load)
1298 swap(orig_dst_load, orig_src_load);
1300 old_imb = orig_dst_load * src_capacity * 100 -
1301 orig_src_load * dst_capacity * env->imbalance_pct;
1303 /* Would this change make things worse? */
1304 return (imb > old_imb);
1308 * This checks if the overall compute and NUMA accesses of the system would
1309 * be improved if the source tasks was migrated to the target dst_cpu taking
1310 * into account that it might be best if task running on the dst_cpu should
1311 * be exchanged with the source task
1313 static void task_numa_compare(struct task_numa_env *env,
1314 long taskimp, long groupimp)
1316 struct rq *src_rq = cpu_rq(env->src_cpu);
1317 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1318 struct task_struct *cur;
1319 long src_load, dst_load;
1321 long imp = env->p->numa_group ? groupimp : taskimp;
1323 int dist = env->dist;
1324 bool assigned = false;
1328 raw_spin_lock_irq(&dst_rq->lock);
1331 * No need to move the exiting task or idle task.
1333 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1337 * The task_struct must be protected here to protect the
1338 * p->numa_faults access in the task_weight since the
1339 * numa_faults could already be freed in the following path:
1340 * finish_task_switch()
1341 * --> put_task_struct()
1342 * --> __put_task_struct()
1343 * --> task_numa_free()
1345 get_task_struct(cur);
1348 raw_spin_unlock_irq(&dst_rq->lock);
1351 * Because we have preemption enabled we can get migrated around and
1352 * end try selecting ourselves (current == env->p) as a swap candidate.
1358 * "imp" is the fault differential for the source task between the
1359 * source and destination node. Calculate the total differential for
1360 * the source task and potential destination task. The more negative
1361 * the value is, the more rmeote accesses that would be expected to
1362 * be incurred if the tasks were swapped.
1365 /* Skip this swap candidate if cannot move to the source cpu */
1366 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1370 * If dst and source tasks are in the same NUMA group, or not
1371 * in any group then look only at task weights.
1373 if (cur->numa_group == env->p->numa_group) {
1374 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1375 task_weight(cur, env->dst_nid, dist);
1377 * Add some hysteresis to prevent swapping the
1378 * tasks within a group over tiny differences.
1380 if (cur->numa_group)
1384 * Compare the group weights. If a task is all by
1385 * itself (not part of a group), use the task weight
1388 if (cur->numa_group)
1389 imp += group_weight(cur, env->src_nid, dist) -
1390 group_weight(cur, env->dst_nid, dist);
1392 imp += task_weight(cur, env->src_nid, dist) -
1393 task_weight(cur, env->dst_nid, dist);
1397 if (imp <= env->best_imp && moveimp <= env->best_imp)
1401 /* Is there capacity at our destination? */
1402 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1403 !env->dst_stats.has_free_capacity)
1409 /* Balance doesn't matter much if we're running a task per cpu */
1410 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1411 dst_rq->nr_running == 1)
1415 * In the overloaded case, try and keep the load balanced.
1418 load = task_h_load(env->p);
1419 dst_load = env->dst_stats.load + load;
1420 src_load = env->src_stats.load - load;
1422 if (moveimp > imp && moveimp > env->best_imp) {
1424 * If the improvement from just moving env->p direction is
1425 * better than swapping tasks around, check if a move is
1426 * possible. Store a slightly smaller score than moveimp,
1427 * so an actually idle CPU will win.
1429 if (!load_too_imbalanced(src_load, dst_load, env)) {
1431 put_task_struct(cur);
1437 if (imp <= env->best_imp)
1441 load = task_h_load(cur);
1446 if (load_too_imbalanced(src_load, dst_load, env))
1450 * One idle CPU per node is evaluated for a task numa move.
1451 * Call select_idle_sibling to maybe find a better one.
1454 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1458 task_numa_assign(env, cur, imp);
1462 * The dst_rq->curr isn't assigned. The protection for task_struct is
1465 if (cur && !assigned)
1466 put_task_struct(cur);
1469 static void task_numa_find_cpu(struct task_numa_env *env,
1470 long taskimp, long groupimp)
1474 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1475 /* Skip this CPU if the source task cannot migrate */
1476 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1480 task_numa_compare(env, taskimp, groupimp);
1484 /* Only move tasks to a NUMA node less busy than the current node. */
1485 static bool numa_has_capacity(struct task_numa_env *env)
1487 struct numa_stats *src = &env->src_stats;
1488 struct numa_stats *dst = &env->dst_stats;
1490 if (src->has_free_capacity && !dst->has_free_capacity)
1494 * Only consider a task move if the source has a higher load
1495 * than the destination, corrected for CPU capacity on each node.
1497 * src->load dst->load
1498 * --------------------- vs ---------------------
1499 * src->compute_capacity dst->compute_capacity
1501 if (src->load * dst->compute_capacity * env->imbalance_pct >
1503 dst->load * src->compute_capacity * 100)
1509 static int task_numa_migrate(struct task_struct *p)
1511 struct task_numa_env env = {
1514 .src_cpu = task_cpu(p),
1515 .src_nid = task_node(p),
1517 .imbalance_pct = 112,
1523 struct sched_domain *sd;
1524 unsigned long taskweight, groupweight;
1526 long taskimp, groupimp;
1529 * Pick the lowest SD_NUMA domain, as that would have the smallest
1530 * imbalance and would be the first to start moving tasks about.
1532 * And we want to avoid any moving of tasks about, as that would create
1533 * random movement of tasks -- counter the numa conditions we're trying
1537 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1539 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1543 * Cpusets can break the scheduler domain tree into smaller
1544 * balance domains, some of which do not cross NUMA boundaries.
1545 * Tasks that are "trapped" in such domains cannot be migrated
1546 * elsewhere, so there is no point in (re)trying.
1548 if (unlikely(!sd)) {
1549 p->numa_preferred_nid = task_node(p);
1553 env.dst_nid = p->numa_preferred_nid;
1554 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1555 taskweight = task_weight(p, env.src_nid, dist);
1556 groupweight = group_weight(p, env.src_nid, dist);
1557 update_numa_stats(&env.src_stats, env.src_nid);
1558 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1559 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1560 update_numa_stats(&env.dst_stats, env.dst_nid);
1562 /* Try to find a spot on the preferred nid. */
1563 if (numa_has_capacity(&env))
1564 task_numa_find_cpu(&env, taskimp, groupimp);
1567 * Look at other nodes in these cases:
1568 * - there is no space available on the preferred_nid
1569 * - the task is part of a numa_group that is interleaved across
1570 * multiple NUMA nodes; in order to better consolidate the group,
1571 * we need to check other locations.
1573 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1574 for_each_online_node(nid) {
1575 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1578 dist = node_distance(env.src_nid, env.dst_nid);
1579 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1581 taskweight = task_weight(p, env.src_nid, dist);
1582 groupweight = group_weight(p, env.src_nid, dist);
1585 /* Only consider nodes where both task and groups benefit */
1586 taskimp = task_weight(p, nid, dist) - taskweight;
1587 groupimp = group_weight(p, nid, dist) - groupweight;
1588 if (taskimp < 0 && groupimp < 0)
1593 update_numa_stats(&env.dst_stats, env.dst_nid);
1594 if (numa_has_capacity(&env))
1595 task_numa_find_cpu(&env, taskimp, groupimp);
1600 * If the task is part of a workload that spans multiple NUMA nodes,
1601 * and is migrating into one of the workload's active nodes, remember
1602 * this node as the task's preferred numa node, so the workload can
1604 * A task that migrated to a second choice node will be better off
1605 * trying for a better one later. Do not set the preferred node here.
1607 if (p->numa_group) {
1608 struct numa_group *ng = p->numa_group;
1610 if (env.best_cpu == -1)
1615 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1616 sched_setnuma(p, env.dst_nid);
1619 /* No better CPU than the current one was found. */
1620 if (env.best_cpu == -1)
1624 * Reset the scan period if the task is being rescheduled on an
1625 * alternative node to recheck if the tasks is now properly placed.
1627 p->numa_scan_period = task_scan_min(p);
1629 if (env.best_task == NULL) {
1630 ret = migrate_task_to(p, env.best_cpu);
1632 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1636 ret = migrate_swap(p, env.best_task);
1638 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1639 put_task_struct(env.best_task);
1643 /* Attempt to migrate a task to a CPU on the preferred node. */
1644 static void numa_migrate_preferred(struct task_struct *p)
1646 unsigned long interval = HZ;
1648 /* This task has no NUMA fault statistics yet */
1649 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1652 /* Periodically retry migrating the task to the preferred node */
1653 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1654 p->numa_migrate_retry = jiffies + interval;
1656 /* Success if task is already running on preferred CPU */
1657 if (task_node(p) == p->numa_preferred_nid)
1660 /* Otherwise, try migrate to a CPU on the preferred node */
1661 task_numa_migrate(p);
1665 * Find out how many nodes on the workload is actively running on. Do this by
1666 * tracking the nodes from which NUMA hinting faults are triggered. This can
1667 * be different from the set of nodes where the workload's memory is currently
1670 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1672 unsigned long faults, max_faults = 0;
1673 int nid, active_nodes = 0;
1675 for_each_online_node(nid) {
1676 faults = group_faults_cpu(numa_group, nid);
1677 if (faults > max_faults)
1678 max_faults = faults;
1681 for_each_online_node(nid) {
1682 faults = group_faults_cpu(numa_group, nid);
1683 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1687 numa_group->max_faults_cpu = max_faults;
1688 numa_group->active_nodes = active_nodes;
1692 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1693 * increments. The more local the fault statistics are, the higher the scan
1694 * period will be for the next scan window. If local/(local+remote) ratio is
1695 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1696 * the scan period will decrease. Aim for 70% local accesses.
1698 #define NUMA_PERIOD_SLOTS 10
1699 #define NUMA_PERIOD_THRESHOLD 7
1702 * Increase the scan period (slow down scanning) if the majority of
1703 * our memory is already on our local node, or if the majority of
1704 * the page accesses are shared with other processes.
1705 * Otherwise, decrease the scan period.
1707 static void update_task_scan_period(struct task_struct *p,
1708 unsigned long shared, unsigned long private)
1710 unsigned int period_slot;
1714 unsigned long remote = p->numa_faults_locality[0];
1715 unsigned long local = p->numa_faults_locality[1];
1718 * If there were no record hinting faults then either the task is
1719 * completely idle or all activity is areas that are not of interest
1720 * to automatic numa balancing. Related to that, if there were failed
1721 * migration then it implies we are migrating too quickly or the local
1722 * node is overloaded. In either case, scan slower
1724 if (local + shared == 0 || p->numa_faults_locality[2]) {
1725 p->numa_scan_period = min(p->numa_scan_period_max,
1726 p->numa_scan_period << 1);
1728 p->mm->numa_next_scan = jiffies +
1729 msecs_to_jiffies(p->numa_scan_period);
1735 * Prepare to scale scan period relative to the current period.
1736 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1737 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1738 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1740 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1741 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1742 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1743 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1746 diff = slot * period_slot;
1748 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1751 * Scale scan rate increases based on sharing. There is an
1752 * inverse relationship between the degree of sharing and
1753 * the adjustment made to the scanning period. Broadly
1754 * speaking the intent is that there is little point
1755 * scanning faster if shared accesses dominate as it may
1756 * simply bounce migrations uselessly
1758 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1759 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1762 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1763 task_scan_min(p), task_scan_max(p));
1764 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1768 * Get the fraction of time the task has been running since the last
1769 * NUMA placement cycle. The scheduler keeps similar statistics, but
1770 * decays those on a 32ms period, which is orders of magnitude off
1771 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1772 * stats only if the task is so new there are no NUMA statistics yet.
1774 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1776 u64 runtime, delta, now;
1777 /* Use the start of this time slice to avoid calculations. */
1778 now = p->se.exec_start;
1779 runtime = p->se.sum_exec_runtime;
1781 if (p->last_task_numa_placement) {
1782 delta = runtime - p->last_sum_exec_runtime;
1783 *period = now - p->last_task_numa_placement;
1785 delta = p->se.avg.load_sum / p->se.load.weight;
1786 *period = LOAD_AVG_MAX;
1789 p->last_sum_exec_runtime = runtime;
1790 p->last_task_numa_placement = now;
1796 * Determine the preferred nid for a task in a numa_group. This needs to
1797 * be done in a way that produces consistent results with group_weight,
1798 * otherwise workloads might not converge.
1800 static int preferred_group_nid(struct task_struct *p, int nid)
1805 /* Direct connections between all NUMA nodes. */
1806 if (sched_numa_topology_type == NUMA_DIRECT)
1810 * On a system with glueless mesh NUMA topology, group_weight
1811 * scores nodes according to the number of NUMA hinting faults on
1812 * both the node itself, and on nearby nodes.
1814 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1815 unsigned long score, max_score = 0;
1816 int node, max_node = nid;
1818 dist = sched_max_numa_distance;
1820 for_each_online_node(node) {
1821 score = group_weight(p, node, dist);
1822 if (score > max_score) {
1831 * Finding the preferred nid in a system with NUMA backplane
1832 * interconnect topology is more involved. The goal is to locate
1833 * tasks from numa_groups near each other in the system, and
1834 * untangle workloads from different sides of the system. This requires
1835 * searching down the hierarchy of node groups, recursively searching
1836 * inside the highest scoring group of nodes. The nodemask tricks
1837 * keep the complexity of the search down.
1839 nodes = node_online_map;
1840 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1841 unsigned long max_faults = 0;
1842 nodemask_t max_group = NODE_MASK_NONE;
1845 /* Are there nodes at this distance from each other? */
1846 if (!find_numa_distance(dist))
1849 for_each_node_mask(a, nodes) {
1850 unsigned long faults = 0;
1851 nodemask_t this_group;
1852 nodes_clear(this_group);
1854 /* Sum group's NUMA faults; includes a==b case. */
1855 for_each_node_mask(b, nodes) {
1856 if (node_distance(a, b) < dist) {
1857 faults += group_faults(p, b);
1858 node_set(b, this_group);
1859 node_clear(b, nodes);
1863 /* Remember the top group. */
1864 if (faults > max_faults) {
1865 max_faults = faults;
1866 max_group = this_group;
1868 * subtle: at the smallest distance there is
1869 * just one node left in each "group", the
1870 * winner is the preferred nid.
1875 /* Next round, evaluate the nodes within max_group. */
1883 static void task_numa_placement(struct task_struct *p)
1885 int seq, nid, max_nid = -1, max_group_nid = -1;
1886 unsigned long max_faults = 0, max_group_faults = 0;
1887 unsigned long fault_types[2] = { 0, 0 };
1888 unsigned long total_faults;
1889 u64 runtime, period;
1890 spinlock_t *group_lock = NULL;
1893 * The p->mm->numa_scan_seq field gets updated without
1894 * exclusive access. Use READ_ONCE() here to ensure
1895 * that the field is read in a single access:
1897 seq = READ_ONCE(p->mm->numa_scan_seq);
1898 if (p->numa_scan_seq == seq)
1900 p->numa_scan_seq = seq;
1901 p->numa_scan_period_max = task_scan_max(p);
1903 total_faults = p->numa_faults_locality[0] +
1904 p->numa_faults_locality[1];
1905 runtime = numa_get_avg_runtime(p, &period);
1907 /* If the task is part of a group prevent parallel updates to group stats */
1908 if (p->numa_group) {
1909 group_lock = &p->numa_group->lock;
1910 spin_lock_irq(group_lock);
1913 /* Find the node with the highest number of faults */
1914 for_each_online_node(nid) {
1915 /* Keep track of the offsets in numa_faults array */
1916 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1917 unsigned long faults = 0, group_faults = 0;
1920 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1921 long diff, f_diff, f_weight;
1923 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1924 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1925 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1926 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1928 /* Decay existing window, copy faults since last scan */
1929 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1930 fault_types[priv] += p->numa_faults[membuf_idx];
1931 p->numa_faults[membuf_idx] = 0;
1934 * Normalize the faults_from, so all tasks in a group
1935 * count according to CPU use, instead of by the raw
1936 * number of faults. Tasks with little runtime have
1937 * little over-all impact on throughput, and thus their
1938 * faults are less important.
1940 f_weight = div64_u64(runtime << 16, period + 1);
1941 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1943 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1944 p->numa_faults[cpubuf_idx] = 0;
1946 p->numa_faults[mem_idx] += diff;
1947 p->numa_faults[cpu_idx] += f_diff;
1948 faults += p->numa_faults[mem_idx];
1949 p->total_numa_faults += diff;
1950 if (p->numa_group) {
1952 * safe because we can only change our own group
1954 * mem_idx represents the offset for a given
1955 * nid and priv in a specific region because it
1956 * is at the beginning of the numa_faults array.
1958 p->numa_group->faults[mem_idx] += diff;
1959 p->numa_group->faults_cpu[mem_idx] += f_diff;
1960 p->numa_group->total_faults += diff;
1961 group_faults += p->numa_group->faults[mem_idx];
1965 if (faults > max_faults) {
1966 max_faults = faults;
1970 if (group_faults > max_group_faults) {
1971 max_group_faults = group_faults;
1972 max_group_nid = nid;
1976 update_task_scan_period(p, fault_types[0], fault_types[1]);
1978 if (p->numa_group) {
1979 numa_group_count_active_nodes(p->numa_group);
1980 spin_unlock_irq(group_lock);
1981 max_nid = preferred_group_nid(p, max_group_nid);
1985 /* Set the new preferred node */
1986 if (max_nid != p->numa_preferred_nid)
1987 sched_setnuma(p, max_nid);
1989 if (task_node(p) != p->numa_preferred_nid)
1990 numa_migrate_preferred(p);
1994 static inline int get_numa_group(struct numa_group *grp)
1996 return atomic_inc_not_zero(&grp->refcount);
1999 static inline void put_numa_group(struct numa_group *grp)
2001 if (atomic_dec_and_test(&grp->refcount))
2002 kfree_rcu(grp, rcu);
2005 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2008 struct numa_group *grp, *my_grp;
2009 struct task_struct *tsk;
2011 int cpu = cpupid_to_cpu(cpupid);
2014 if (unlikely(!p->numa_group)) {
2015 unsigned int size = sizeof(struct numa_group) +
2016 4*nr_node_ids*sizeof(unsigned long);
2018 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2022 atomic_set(&grp->refcount, 1);
2023 grp->active_nodes = 1;
2024 grp->max_faults_cpu = 0;
2025 spin_lock_init(&grp->lock);
2027 /* Second half of the array tracks nids where faults happen */
2028 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2031 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2032 grp->faults[i] = p->numa_faults[i];
2034 grp->total_faults = p->total_numa_faults;
2037 rcu_assign_pointer(p->numa_group, grp);
2041 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2043 if (!cpupid_match_pid(tsk, cpupid))
2046 grp = rcu_dereference(tsk->numa_group);
2050 my_grp = p->numa_group;
2055 * Only join the other group if its bigger; if we're the bigger group,
2056 * the other task will join us.
2058 if (my_grp->nr_tasks > grp->nr_tasks)
2062 * Tie-break on the grp address.
2064 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2067 /* Always join threads in the same process. */
2068 if (tsk->mm == current->mm)
2071 /* Simple filter to avoid false positives due to PID collisions */
2072 if (flags & TNF_SHARED)
2075 /* Update priv based on whether false sharing was detected */
2078 if (join && !get_numa_group(grp))
2086 BUG_ON(irqs_disabled());
2087 double_lock_irq(&my_grp->lock, &grp->lock);
2089 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2090 my_grp->faults[i] -= p->numa_faults[i];
2091 grp->faults[i] += p->numa_faults[i];
2093 my_grp->total_faults -= p->total_numa_faults;
2094 grp->total_faults += p->total_numa_faults;
2099 spin_unlock(&my_grp->lock);
2100 spin_unlock_irq(&grp->lock);
2102 rcu_assign_pointer(p->numa_group, grp);
2104 put_numa_group(my_grp);
2112 void task_numa_free(struct task_struct *p)
2114 struct numa_group *grp = p->numa_group;
2115 void *numa_faults = p->numa_faults;
2116 unsigned long flags;
2120 spin_lock_irqsave(&grp->lock, flags);
2121 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2122 grp->faults[i] -= p->numa_faults[i];
2123 grp->total_faults -= p->total_numa_faults;
2126 spin_unlock_irqrestore(&grp->lock, flags);
2127 RCU_INIT_POINTER(p->numa_group, NULL);
2128 put_numa_group(grp);
2131 p->numa_faults = NULL;
2136 * Got a PROT_NONE fault for a page on @node.
2138 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2140 struct task_struct *p = current;
2141 bool migrated = flags & TNF_MIGRATED;
2142 int cpu_node = task_node(current);
2143 int local = !!(flags & TNF_FAULT_LOCAL);
2144 struct numa_group *ng;
2147 if (!static_branch_likely(&sched_numa_balancing))
2150 /* for example, ksmd faulting in a user's mm */
2154 /* Allocate buffer to track faults on a per-node basis */
2155 if (unlikely(!p->numa_faults)) {
2156 int size = sizeof(*p->numa_faults) *
2157 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2159 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2160 if (!p->numa_faults)
2163 p->total_numa_faults = 0;
2164 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2168 * First accesses are treated as private, otherwise consider accesses
2169 * to be private if the accessing pid has not changed
2171 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2174 priv = cpupid_match_pid(p, last_cpupid);
2175 if (!priv && !(flags & TNF_NO_GROUP))
2176 task_numa_group(p, last_cpupid, flags, &priv);
2180 * If a workload spans multiple NUMA nodes, a shared fault that
2181 * occurs wholly within the set of nodes that the workload is
2182 * actively using should be counted as local. This allows the
2183 * scan rate to slow down when a workload has settled down.
2186 if (!priv && !local && ng && ng->active_nodes > 1 &&
2187 numa_is_active_node(cpu_node, ng) &&
2188 numa_is_active_node(mem_node, ng))
2191 task_numa_placement(p);
2194 * Retry task to preferred node migration periodically, in case it
2195 * case it previously failed, or the scheduler moved us.
2197 if (time_after(jiffies, p->numa_migrate_retry))
2198 numa_migrate_preferred(p);
2201 p->numa_pages_migrated += pages;
2202 if (flags & TNF_MIGRATE_FAIL)
2203 p->numa_faults_locality[2] += pages;
2205 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2206 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2207 p->numa_faults_locality[local] += pages;
2210 static void reset_ptenuma_scan(struct task_struct *p)
2213 * We only did a read acquisition of the mmap sem, so
2214 * p->mm->numa_scan_seq is written to without exclusive access
2215 * and the update is not guaranteed to be atomic. That's not
2216 * much of an issue though, since this is just used for
2217 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2218 * expensive, to avoid any form of compiler optimizations:
2220 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2221 p->mm->numa_scan_offset = 0;
2225 * The expensive part of numa migration is done from task_work context.
2226 * Triggered from task_tick_numa().
2228 void task_numa_work(struct callback_head *work)
2230 unsigned long migrate, next_scan, now = jiffies;
2231 struct task_struct *p = current;
2232 struct mm_struct *mm = p->mm;
2233 u64 runtime = p->se.sum_exec_runtime;
2234 struct vm_area_struct *vma;
2235 unsigned long start, end;
2236 unsigned long nr_pte_updates = 0;
2237 long pages, virtpages;
2239 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2241 work->next = work; /* protect against double add */
2243 * Who cares about NUMA placement when they're dying.
2245 * NOTE: make sure not to dereference p->mm before this check,
2246 * exit_task_work() happens _after_ exit_mm() so we could be called
2247 * without p->mm even though we still had it when we enqueued this
2250 if (p->flags & PF_EXITING)
2253 if (!mm->numa_next_scan) {
2254 mm->numa_next_scan = now +
2255 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2259 * Enforce maximal scan/migration frequency..
2261 migrate = mm->numa_next_scan;
2262 if (time_before(now, migrate))
2265 if (p->numa_scan_period == 0) {
2266 p->numa_scan_period_max = task_scan_max(p);
2267 p->numa_scan_period = task_scan_min(p);
2270 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2271 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2275 * Delay this task enough that another task of this mm will likely win
2276 * the next time around.
2278 p->node_stamp += 2 * TICK_NSEC;
2280 start = mm->numa_scan_offset;
2281 pages = sysctl_numa_balancing_scan_size;
2282 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2283 virtpages = pages * 8; /* Scan up to this much virtual space */
2288 down_read(&mm->mmap_sem);
2289 vma = find_vma(mm, start);
2291 reset_ptenuma_scan(p);
2295 for (; vma; vma = vma->vm_next) {
2296 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2297 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2302 * Shared library pages mapped by multiple processes are not
2303 * migrated as it is expected they are cache replicated. Avoid
2304 * hinting faults in read-only file-backed mappings or the vdso
2305 * as migrating the pages will be of marginal benefit.
2308 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2312 * Skip inaccessible VMAs to avoid any confusion between
2313 * PROT_NONE and NUMA hinting ptes
2315 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2319 start = max(start, vma->vm_start);
2320 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2321 end = min(end, vma->vm_end);
2322 nr_pte_updates = change_prot_numa(vma, start, end);
2325 * Try to scan sysctl_numa_balancing_size worth of
2326 * hpages that have at least one present PTE that
2327 * is not already pte-numa. If the VMA contains
2328 * areas that are unused or already full of prot_numa
2329 * PTEs, scan up to virtpages, to skip through those
2333 pages -= (end - start) >> PAGE_SHIFT;
2334 virtpages -= (end - start) >> PAGE_SHIFT;
2337 if (pages <= 0 || virtpages <= 0)
2341 } while (end != vma->vm_end);
2346 * It is possible to reach the end of the VMA list but the last few
2347 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2348 * would find the !migratable VMA on the next scan but not reset the
2349 * scanner to the start so check it now.
2352 mm->numa_scan_offset = start;
2354 reset_ptenuma_scan(p);
2355 up_read(&mm->mmap_sem);
2358 * Make sure tasks use at least 32x as much time to run other code
2359 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2360 * Usually update_task_scan_period slows down scanning enough; on an
2361 * overloaded system we need to limit overhead on a per task basis.
2363 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2364 u64 diff = p->se.sum_exec_runtime - runtime;
2365 p->node_stamp += 32 * diff;
2370 * Drive the periodic memory faults..
2372 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2374 struct callback_head *work = &curr->numa_work;
2378 * We don't care about NUMA placement if we don't have memory.
2380 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2384 * Using runtime rather than walltime has the dual advantage that
2385 * we (mostly) drive the selection from busy threads and that the
2386 * task needs to have done some actual work before we bother with
2389 now = curr->se.sum_exec_runtime;
2390 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2392 if (now > curr->node_stamp + period) {
2393 if (!curr->node_stamp)
2394 curr->numa_scan_period = task_scan_min(curr);
2395 curr->node_stamp += period;
2397 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2398 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2399 task_work_add(curr, work, true);
2404 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2408 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2412 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2415 #endif /* CONFIG_NUMA_BALANCING */
2418 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2420 update_load_add(&cfs_rq->load, se->load.weight);
2421 if (!parent_entity(se))
2422 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2424 if (entity_is_task(se)) {
2425 struct rq *rq = rq_of(cfs_rq);
2427 account_numa_enqueue(rq, task_of(se));
2428 list_add(&se->group_node, &rq->cfs_tasks);
2431 cfs_rq->nr_running++;
2435 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2437 update_load_sub(&cfs_rq->load, se->load.weight);
2438 if (!parent_entity(se))
2439 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2440 if (entity_is_task(se)) {
2441 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2442 list_del_init(&se->group_node);
2444 cfs_rq->nr_running--;
2447 #ifdef CONFIG_FAIR_GROUP_SCHED
2449 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2454 * Use this CPU's real-time load instead of the last load contribution
2455 * as the updating of the contribution is delayed, and we will use the
2456 * the real-time load to calc the share. See update_tg_load_avg().
2458 tg_weight = atomic_long_read(&tg->load_avg);
2459 tg_weight -= cfs_rq->tg_load_avg_contrib;
2460 tg_weight += cfs_rq->load.weight;
2465 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2467 long tg_weight, load, shares;
2469 tg_weight = calc_tg_weight(tg, cfs_rq);
2470 load = cfs_rq->load.weight;
2472 shares = (tg->shares * load);
2474 shares /= tg_weight;
2476 if (shares < MIN_SHARES)
2477 shares = MIN_SHARES;
2478 if (shares > tg->shares)
2479 shares = tg->shares;
2483 # else /* CONFIG_SMP */
2484 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2488 # endif /* CONFIG_SMP */
2489 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2490 unsigned long weight)
2493 /* commit outstanding execution time */
2494 if (cfs_rq->curr == se)
2495 update_curr(cfs_rq);
2496 account_entity_dequeue(cfs_rq, se);
2499 update_load_set(&se->load, weight);
2502 account_entity_enqueue(cfs_rq, se);
2505 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2507 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2509 struct task_group *tg;
2510 struct sched_entity *se;
2514 se = tg->se[cpu_of(rq_of(cfs_rq))];
2515 if (!se || throttled_hierarchy(cfs_rq))
2518 if (likely(se->load.weight == tg->shares))
2521 shares = calc_cfs_shares(cfs_rq, tg);
2523 reweight_entity(cfs_rq_of(se), se, shares);
2525 #else /* CONFIG_FAIR_GROUP_SCHED */
2526 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2529 #endif /* CONFIG_FAIR_GROUP_SCHED */
2532 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2533 static const u32 runnable_avg_yN_inv[] = {
2534 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2535 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2536 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2537 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2538 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2539 0x85aac367, 0x82cd8698,
2543 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2544 * over-estimates when re-combining.
2546 static const u32 runnable_avg_yN_sum[] = {
2547 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2548 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2549 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2554 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2556 static __always_inline u64 decay_load(u64 val, u64 n)
2558 unsigned int local_n;
2562 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2565 /* after bounds checking we can collapse to 32-bit */
2569 * As y^PERIOD = 1/2, we can combine
2570 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2571 * With a look-up table which covers y^n (n<PERIOD)
2573 * To achieve constant time decay_load.
2575 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2576 val >>= local_n / LOAD_AVG_PERIOD;
2577 local_n %= LOAD_AVG_PERIOD;
2580 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2585 * For updates fully spanning n periods, the contribution to runnable
2586 * average will be: \Sum 1024*y^n
2588 * We can compute this reasonably efficiently by combining:
2589 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2591 static u32 __compute_runnable_contrib(u64 n)
2595 if (likely(n <= LOAD_AVG_PERIOD))
2596 return runnable_avg_yN_sum[n];
2597 else if (unlikely(n >= LOAD_AVG_MAX_N))
2598 return LOAD_AVG_MAX;
2600 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2602 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2603 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2605 n -= LOAD_AVG_PERIOD;
2606 } while (n > LOAD_AVG_PERIOD);
2608 contrib = decay_load(contrib, n);
2609 return contrib + runnable_avg_yN_sum[n];
2612 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2613 #error "load tracking assumes 2^10 as unit"
2616 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2619 * We can represent the historical contribution to runnable average as the
2620 * coefficients of a geometric series. To do this we sub-divide our runnable
2621 * history into segments of approximately 1ms (1024us); label the segment that
2622 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2624 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2626 * (now) (~1ms ago) (~2ms ago)
2628 * Let u_i denote the fraction of p_i that the entity was runnable.
2630 * We then designate the fractions u_i as our co-efficients, yielding the
2631 * following representation of historical load:
2632 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2634 * We choose y based on the with of a reasonably scheduling period, fixing:
2637 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2638 * approximately half as much as the contribution to load within the last ms
2641 * When a period "rolls over" and we have new u_0`, multiplying the previous
2642 * sum again by y is sufficient to update:
2643 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2644 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2646 static __always_inline int
2647 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2648 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2650 u64 delta, scaled_delta, periods;
2652 unsigned int delta_w, scaled_delta_w, decayed = 0;
2653 unsigned long scale_freq, scale_cpu;
2655 delta = now - sa->last_update_time;
2657 * This should only happen when time goes backwards, which it
2658 * unfortunately does during sched clock init when we swap over to TSC.
2660 if ((s64)delta < 0) {
2661 sa->last_update_time = now;
2666 * Use 1024ns as the unit of measurement since it's a reasonable
2667 * approximation of 1us and fast to compute.
2672 sa->last_update_time = now;
2674 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2675 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2677 /* delta_w is the amount already accumulated against our next period */
2678 delta_w = sa->period_contrib;
2679 if (delta + delta_w >= 1024) {
2682 /* how much left for next period will start over, we don't know yet */
2683 sa->period_contrib = 0;
2686 * Now that we know we're crossing a period boundary, figure
2687 * out how much from delta we need to complete the current
2688 * period and accrue it.
2690 delta_w = 1024 - delta_w;
2691 scaled_delta_w = cap_scale(delta_w, scale_freq);
2693 sa->load_sum += weight * scaled_delta_w;
2695 cfs_rq->runnable_load_sum +=
2696 weight * scaled_delta_w;
2700 sa->util_sum += scaled_delta_w * scale_cpu;
2704 /* Figure out how many additional periods this update spans */
2705 periods = delta / 1024;
2708 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2710 cfs_rq->runnable_load_sum =
2711 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2713 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2715 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2716 contrib = __compute_runnable_contrib(periods);
2717 contrib = cap_scale(contrib, scale_freq);
2719 sa->load_sum += weight * contrib;
2721 cfs_rq->runnable_load_sum += weight * contrib;
2724 sa->util_sum += contrib * scale_cpu;
2727 /* Remainder of delta accrued against u_0` */
2728 scaled_delta = cap_scale(delta, scale_freq);
2730 sa->load_sum += weight * scaled_delta;
2732 cfs_rq->runnable_load_sum += weight * scaled_delta;
2735 sa->util_sum += scaled_delta * scale_cpu;
2737 sa->period_contrib += delta;
2740 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2742 cfs_rq->runnable_load_avg =
2743 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2745 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2751 #ifdef CONFIG_FAIR_GROUP_SCHED
2753 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2754 * and effective_load (which is not done because it is too costly).
2756 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2758 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2761 * No need to update load_avg for root_task_group as it is not used.
2763 if (cfs_rq->tg == &root_task_group)
2766 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2767 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2768 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2773 * Called within set_task_rq() right before setting a task's cpu. The
2774 * caller only guarantees p->pi_lock is held; no other assumptions,
2775 * including the state of rq->lock, should be made.
2777 void set_task_rq_fair(struct sched_entity *se,
2778 struct cfs_rq *prev, struct cfs_rq *next)
2780 if (!sched_feat(ATTACH_AGE_LOAD))
2784 * We are supposed to update the task to "current" time, then its up to
2785 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2786 * getting what current time is, so simply throw away the out-of-date
2787 * time. This will result in the wakee task is less decayed, but giving
2788 * the wakee more load sounds not bad.
2790 if (se->avg.last_update_time && prev) {
2791 u64 p_last_update_time;
2792 u64 n_last_update_time;
2794 #ifndef CONFIG_64BIT
2795 u64 p_last_update_time_copy;
2796 u64 n_last_update_time_copy;
2799 p_last_update_time_copy = prev->load_last_update_time_copy;
2800 n_last_update_time_copy = next->load_last_update_time_copy;
2804 p_last_update_time = prev->avg.last_update_time;
2805 n_last_update_time = next->avg.last_update_time;
2807 } while (p_last_update_time != p_last_update_time_copy ||
2808 n_last_update_time != n_last_update_time_copy);
2810 p_last_update_time = prev->avg.last_update_time;
2811 n_last_update_time = next->avg.last_update_time;
2813 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2814 &se->avg, 0, 0, NULL);
2815 se->avg.last_update_time = n_last_update_time;
2818 #else /* CONFIG_FAIR_GROUP_SCHED */
2819 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2820 #endif /* CONFIG_FAIR_GROUP_SCHED */
2822 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2824 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2825 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2827 struct sched_avg *sa = &cfs_rq->avg;
2828 int decayed, removed = 0;
2830 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2831 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2832 sa->load_avg = max_t(long, sa->load_avg - r, 0);
2833 sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
2837 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2838 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2839 sa->util_avg = max_t(long, sa->util_avg - r, 0);
2840 sa->util_sum = max_t(s32, sa->util_sum - r * LOAD_AVG_MAX, 0);
2843 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2844 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2846 #ifndef CONFIG_64BIT
2848 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2851 return decayed || removed;
2854 /* Update task and its cfs_rq load average */
2855 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2857 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2858 u64 now = cfs_rq_clock_task(cfs_rq);
2859 struct rq *rq = rq_of(cfs_rq);
2860 int cpu = cpu_of(rq);
2863 * Track task load average for carrying it to new CPU after migrated, and
2864 * track group sched_entity load average for task_h_load calc in migration
2866 __update_load_avg(now, cpu, &se->avg,
2867 se->on_rq * scale_load_down(se->load.weight),
2868 cfs_rq->curr == se, NULL);
2870 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2871 update_tg_load_avg(cfs_rq, 0);
2873 if (cpu == smp_processor_id() && &rq->cfs == cfs_rq) {
2874 unsigned long max = rq->cpu_capacity_orig;
2877 * There are a few boundary cases this might miss but it should
2878 * get called often enough that that should (hopefully) not be
2879 * a real problem -- added to that it only calls on the local
2880 * CPU, so if we enqueue remotely we'll miss an update, but
2881 * the next tick/schedule should update.
2883 * It will not get called when we go idle, because the idle
2884 * thread is a different class (!fair), nor will the utilization
2885 * number include things like RT tasks.
2887 * As is, the util number is not freq-invariant (we'd have to
2888 * implement arch_scale_freq_capacity() for that).
2892 cpufreq_update_util(rq_clock(rq),
2893 min(cfs_rq->avg.util_avg, max), max);
2897 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2899 if (!sched_feat(ATTACH_AGE_LOAD))
2903 * If we got migrated (either between CPUs or between cgroups) we'll
2904 * have aged the average right before clearing @last_update_time.
2906 if (se->avg.last_update_time) {
2907 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2908 &se->avg, 0, 0, NULL);
2911 * XXX: we could have just aged the entire load away if we've been
2912 * absent from the fair class for too long.
2917 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2918 cfs_rq->avg.load_avg += se->avg.load_avg;
2919 cfs_rq->avg.load_sum += se->avg.load_sum;
2920 cfs_rq->avg.util_avg += se->avg.util_avg;
2921 cfs_rq->avg.util_sum += se->avg.util_sum;
2924 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2926 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2927 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2928 cfs_rq->curr == se, NULL);
2930 cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
2931 cfs_rq->avg.load_sum = max_t(s64, cfs_rq->avg.load_sum - se->avg.load_sum, 0);
2932 cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
2933 cfs_rq->avg.util_sum = max_t(s32, cfs_rq->avg.util_sum - se->avg.util_sum, 0);
2936 /* Add the load generated by se into cfs_rq's load average */
2938 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2940 struct sched_avg *sa = &se->avg;
2941 u64 now = cfs_rq_clock_task(cfs_rq);
2942 int migrated, decayed;
2944 migrated = !sa->last_update_time;
2946 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2947 se->on_rq * scale_load_down(se->load.weight),
2948 cfs_rq->curr == se, NULL);
2951 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2953 cfs_rq->runnable_load_avg += sa->load_avg;
2954 cfs_rq->runnable_load_sum += sa->load_sum;
2957 attach_entity_load_avg(cfs_rq, se);
2959 if (decayed || migrated)
2960 update_tg_load_avg(cfs_rq, 0);
2963 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2965 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2967 update_load_avg(se, 1);
2969 cfs_rq->runnable_load_avg =
2970 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2971 cfs_rq->runnable_load_sum =
2972 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2975 #ifndef CONFIG_64BIT
2976 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2978 u64 last_update_time_copy;
2979 u64 last_update_time;
2982 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2984 last_update_time = cfs_rq->avg.last_update_time;
2985 } while (last_update_time != last_update_time_copy);
2987 return last_update_time;
2990 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2992 return cfs_rq->avg.last_update_time;
2997 * Task first catches up with cfs_rq, and then subtract
2998 * itself from the cfs_rq (task must be off the queue now).
3000 void remove_entity_load_avg(struct sched_entity *se)
3002 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3003 u64 last_update_time;
3006 * Newly created task or never used group entity should not be removed
3007 * from its (source) cfs_rq
3009 if (se->avg.last_update_time == 0)
3012 last_update_time = cfs_rq_last_update_time(cfs_rq);
3014 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3015 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3016 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3019 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3021 return cfs_rq->runnable_load_avg;
3024 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3026 return cfs_rq->avg.load_avg;
3029 static int idle_balance(struct rq *this_rq);
3031 #else /* CONFIG_SMP */
3033 static inline void update_load_avg(struct sched_entity *se, int not_used)
3035 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3036 struct rq *rq = rq_of(cfs_rq);
3038 cpufreq_trigger_update(rq_clock(rq));
3042 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3044 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3045 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3048 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3050 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3052 static inline int idle_balance(struct rq *rq)
3057 #endif /* CONFIG_SMP */
3059 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3061 #ifdef CONFIG_SCHEDSTATS
3062 struct task_struct *tsk = NULL;
3064 if (entity_is_task(se))
3067 if (se->statistics.sleep_start) {
3068 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3073 if (unlikely(delta > se->statistics.sleep_max))
3074 se->statistics.sleep_max = delta;
3076 se->statistics.sleep_start = 0;
3077 se->statistics.sum_sleep_runtime += delta;
3080 account_scheduler_latency(tsk, delta >> 10, 1);
3081 trace_sched_stat_sleep(tsk, delta);
3084 if (se->statistics.block_start) {
3085 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3090 if (unlikely(delta > se->statistics.block_max))
3091 se->statistics.block_max = delta;
3093 se->statistics.block_start = 0;
3094 se->statistics.sum_sleep_runtime += delta;
3097 if (tsk->in_iowait) {
3098 se->statistics.iowait_sum += delta;
3099 se->statistics.iowait_count++;
3100 trace_sched_stat_iowait(tsk, delta);
3103 trace_sched_stat_blocked(tsk, delta);
3106 * Blocking time is in units of nanosecs, so shift by
3107 * 20 to get a milliseconds-range estimation of the
3108 * amount of time that the task spent sleeping:
3110 if (unlikely(prof_on == SLEEP_PROFILING)) {
3111 profile_hits(SLEEP_PROFILING,
3112 (void *)get_wchan(tsk),
3115 account_scheduler_latency(tsk, delta >> 10, 0);
3121 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3123 #ifdef CONFIG_SCHED_DEBUG
3124 s64 d = se->vruntime - cfs_rq->min_vruntime;
3129 if (d > 3*sysctl_sched_latency)
3130 schedstat_inc(cfs_rq, nr_spread_over);
3135 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3137 u64 vruntime = cfs_rq->min_vruntime;
3140 * The 'current' period is already promised to the current tasks,
3141 * however the extra weight of the new task will slow them down a
3142 * little, place the new task so that it fits in the slot that
3143 * stays open at the end.
3145 if (initial && sched_feat(START_DEBIT))
3146 vruntime += sched_vslice(cfs_rq, se);
3148 /* sleeps up to a single latency don't count. */
3150 unsigned long thresh = sysctl_sched_latency;
3153 * Halve their sleep time's effect, to allow
3154 * for a gentler effect of sleepers:
3156 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3162 /* ensure we never gain time by being placed backwards. */
3163 se->vruntime = max_vruntime(se->vruntime, vruntime);
3166 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3168 static inline void check_schedstat_required(void)
3170 #ifdef CONFIG_SCHEDSTATS
3171 if (schedstat_enabled())
3174 /* Force schedstat enabled if a dependent tracepoint is active */
3175 if (trace_sched_stat_wait_enabled() ||
3176 trace_sched_stat_sleep_enabled() ||
3177 trace_sched_stat_iowait_enabled() ||
3178 trace_sched_stat_blocked_enabled() ||
3179 trace_sched_stat_runtime_enabled()) {
3180 pr_warn_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3181 "stat_blocked and stat_runtime require the "
3182 "kernel parameter schedstats=enabled or "
3183 "kernel.sched_schedstats=1\n");
3189 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3192 * Update the normalized vruntime before updating min_vruntime
3193 * through calling update_curr().
3195 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3196 se->vruntime += cfs_rq->min_vruntime;
3199 * Update run-time statistics of the 'current'.
3201 update_curr(cfs_rq);
3202 enqueue_entity_load_avg(cfs_rq, se);
3203 account_entity_enqueue(cfs_rq, se);
3204 update_cfs_shares(cfs_rq);
3206 if (flags & ENQUEUE_WAKEUP) {
3207 place_entity(cfs_rq, se, 0);
3208 if (schedstat_enabled())
3209 enqueue_sleeper(cfs_rq, se);
3212 check_schedstat_required();
3213 if (schedstat_enabled()) {
3214 update_stats_enqueue(cfs_rq, se);
3215 check_spread(cfs_rq, se);
3217 if (se != cfs_rq->curr)
3218 __enqueue_entity(cfs_rq, se);
3221 if (cfs_rq->nr_running == 1) {
3222 list_add_leaf_cfs_rq(cfs_rq);
3223 check_enqueue_throttle(cfs_rq);
3227 static void __clear_buddies_last(struct sched_entity *se)
3229 for_each_sched_entity(se) {
3230 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3231 if (cfs_rq->last != se)
3234 cfs_rq->last = NULL;
3238 static void __clear_buddies_next(struct sched_entity *se)
3240 for_each_sched_entity(se) {
3241 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3242 if (cfs_rq->next != se)
3245 cfs_rq->next = NULL;
3249 static void __clear_buddies_skip(struct sched_entity *se)
3251 for_each_sched_entity(se) {
3252 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3253 if (cfs_rq->skip != se)
3256 cfs_rq->skip = NULL;
3260 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3262 if (cfs_rq->last == se)
3263 __clear_buddies_last(se);
3265 if (cfs_rq->next == se)
3266 __clear_buddies_next(se);
3268 if (cfs_rq->skip == se)
3269 __clear_buddies_skip(se);
3272 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3275 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3278 * Update run-time statistics of the 'current'.
3280 update_curr(cfs_rq);
3281 dequeue_entity_load_avg(cfs_rq, se);
3283 if (schedstat_enabled())
3284 update_stats_dequeue(cfs_rq, se, flags);
3286 clear_buddies(cfs_rq, se);
3288 if (se != cfs_rq->curr)
3289 __dequeue_entity(cfs_rq, se);
3291 account_entity_dequeue(cfs_rq, se);
3294 * Normalize the entity after updating the min_vruntime because the
3295 * update can refer to the ->curr item and we need to reflect this
3296 * movement in our normalized position.
3298 if (!(flags & DEQUEUE_SLEEP))
3299 se->vruntime -= cfs_rq->min_vruntime;
3301 /* return excess runtime on last dequeue */
3302 return_cfs_rq_runtime(cfs_rq);
3304 update_min_vruntime(cfs_rq);
3305 update_cfs_shares(cfs_rq);
3309 * Preempt the current task with a newly woken task if needed:
3312 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3314 unsigned long ideal_runtime, delta_exec;
3315 struct sched_entity *se;
3318 ideal_runtime = sched_slice(cfs_rq, curr);
3319 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3320 if (delta_exec > ideal_runtime) {
3321 resched_curr(rq_of(cfs_rq));
3323 * The current task ran long enough, ensure it doesn't get
3324 * re-elected due to buddy favours.
3326 clear_buddies(cfs_rq, curr);
3331 * Ensure that a task that missed wakeup preemption by a
3332 * narrow margin doesn't have to wait for a full slice.
3333 * This also mitigates buddy induced latencies under load.
3335 if (delta_exec < sysctl_sched_min_granularity)
3338 se = __pick_first_entity(cfs_rq);
3339 delta = curr->vruntime - se->vruntime;
3344 if (delta > ideal_runtime)
3345 resched_curr(rq_of(cfs_rq));
3349 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3351 /* 'current' is not kept within the tree. */
3354 * Any task has to be enqueued before it get to execute on
3355 * a CPU. So account for the time it spent waiting on the
3358 if (schedstat_enabled())
3359 update_stats_wait_end(cfs_rq, se);
3360 __dequeue_entity(cfs_rq, se);
3361 update_load_avg(se, 1);
3364 update_stats_curr_start(cfs_rq, se);
3366 #ifdef CONFIG_SCHEDSTATS
3368 * Track our maximum slice length, if the CPU's load is at
3369 * least twice that of our own weight (i.e. dont track it
3370 * when there are only lesser-weight tasks around):
3372 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3373 se->statistics.slice_max = max(se->statistics.slice_max,
3374 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3377 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3381 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3384 * Pick the next process, keeping these things in mind, in this order:
3385 * 1) keep things fair between processes/task groups
3386 * 2) pick the "next" process, since someone really wants that to run
3387 * 3) pick the "last" process, for cache locality
3388 * 4) do not run the "skip" process, if something else is available
3390 static struct sched_entity *
3391 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3393 struct sched_entity *left = __pick_first_entity(cfs_rq);
3394 struct sched_entity *se;
3397 * If curr is set we have to see if its left of the leftmost entity
3398 * still in the tree, provided there was anything in the tree at all.
3400 if (!left || (curr && entity_before(curr, left)))
3403 se = left; /* ideally we run the leftmost entity */
3406 * Avoid running the skip buddy, if running something else can
3407 * be done without getting too unfair.
3409 if (cfs_rq->skip == se) {
3410 struct sched_entity *second;
3413 second = __pick_first_entity(cfs_rq);
3415 second = __pick_next_entity(se);
3416 if (!second || (curr && entity_before(curr, second)))
3420 if (second && wakeup_preempt_entity(second, left) < 1)
3425 * Prefer last buddy, try to return the CPU to a preempted task.
3427 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3431 * Someone really wants this to run. If it's not unfair, run it.
3433 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3436 clear_buddies(cfs_rq, se);
3441 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3443 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3446 * If still on the runqueue then deactivate_task()
3447 * was not called and update_curr() has to be done:
3450 update_curr(cfs_rq);
3452 /* throttle cfs_rqs exceeding runtime */
3453 check_cfs_rq_runtime(cfs_rq);
3455 if (schedstat_enabled()) {
3456 check_spread(cfs_rq, prev);
3458 update_stats_wait_start(cfs_rq, prev);
3462 /* Put 'current' back into the tree. */
3463 __enqueue_entity(cfs_rq, prev);
3464 /* in !on_rq case, update occurred at dequeue */
3465 update_load_avg(prev, 0);
3467 cfs_rq->curr = NULL;
3471 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3474 * Update run-time statistics of the 'current'.
3476 update_curr(cfs_rq);
3479 * Ensure that runnable average is periodically updated.
3481 update_load_avg(curr, 1);
3482 update_cfs_shares(cfs_rq);
3484 #ifdef CONFIG_SCHED_HRTICK
3486 * queued ticks are scheduled to match the slice, so don't bother
3487 * validating it and just reschedule.
3490 resched_curr(rq_of(cfs_rq));
3494 * don't let the period tick interfere with the hrtick preemption
3496 if (!sched_feat(DOUBLE_TICK) &&
3497 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3501 if (cfs_rq->nr_running > 1)
3502 check_preempt_tick(cfs_rq, curr);
3506 /**************************************************
3507 * CFS bandwidth control machinery
3510 #ifdef CONFIG_CFS_BANDWIDTH
3512 #ifdef HAVE_JUMP_LABEL
3513 static struct static_key __cfs_bandwidth_used;
3515 static inline bool cfs_bandwidth_used(void)
3517 return static_key_false(&__cfs_bandwidth_used);
3520 void cfs_bandwidth_usage_inc(void)
3522 static_key_slow_inc(&__cfs_bandwidth_used);
3525 void cfs_bandwidth_usage_dec(void)
3527 static_key_slow_dec(&__cfs_bandwidth_used);
3529 #else /* HAVE_JUMP_LABEL */
3530 static bool cfs_bandwidth_used(void)
3535 void cfs_bandwidth_usage_inc(void) {}
3536 void cfs_bandwidth_usage_dec(void) {}
3537 #endif /* HAVE_JUMP_LABEL */
3540 * default period for cfs group bandwidth.
3541 * default: 0.1s, units: nanoseconds
3543 static inline u64 default_cfs_period(void)
3545 return 100000000ULL;
3548 static inline u64 sched_cfs_bandwidth_slice(void)
3550 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3554 * Replenish runtime according to assigned quota and update expiration time.
3555 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3556 * additional synchronization around rq->lock.
3558 * requires cfs_b->lock
3560 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3564 if (cfs_b->quota == RUNTIME_INF)
3567 now = sched_clock_cpu(smp_processor_id());
3568 cfs_b->runtime = cfs_b->quota;
3569 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3572 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3574 return &tg->cfs_bandwidth;
3577 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3578 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3580 if (unlikely(cfs_rq->throttle_count))
3581 return cfs_rq->throttled_clock_task;
3583 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3586 /* returns 0 on failure to allocate runtime */
3587 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3589 struct task_group *tg = cfs_rq->tg;
3590 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3591 u64 amount = 0, min_amount, expires;
3593 /* note: this is a positive sum as runtime_remaining <= 0 */
3594 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3596 raw_spin_lock(&cfs_b->lock);
3597 if (cfs_b->quota == RUNTIME_INF)
3598 amount = min_amount;
3600 start_cfs_bandwidth(cfs_b);
3602 if (cfs_b->runtime > 0) {
3603 amount = min(cfs_b->runtime, min_amount);
3604 cfs_b->runtime -= amount;
3608 expires = cfs_b->runtime_expires;
3609 raw_spin_unlock(&cfs_b->lock);
3611 cfs_rq->runtime_remaining += amount;
3613 * we may have advanced our local expiration to account for allowed
3614 * spread between our sched_clock and the one on which runtime was
3617 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3618 cfs_rq->runtime_expires = expires;
3620 return cfs_rq->runtime_remaining > 0;
3624 * Note: This depends on the synchronization provided by sched_clock and the
3625 * fact that rq->clock snapshots this value.
3627 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3629 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3631 /* if the deadline is ahead of our clock, nothing to do */
3632 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3635 if (cfs_rq->runtime_remaining < 0)
3639 * If the local deadline has passed we have to consider the
3640 * possibility that our sched_clock is 'fast' and the global deadline
3641 * has not truly expired.
3643 * Fortunately we can check determine whether this the case by checking
3644 * whether the global deadline has advanced. It is valid to compare
3645 * cfs_b->runtime_expires without any locks since we only care about
3646 * exact equality, so a partial write will still work.
3649 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3650 /* extend local deadline, drift is bounded above by 2 ticks */
3651 cfs_rq->runtime_expires += TICK_NSEC;
3653 /* global deadline is ahead, expiration has passed */
3654 cfs_rq->runtime_remaining = 0;
3658 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3660 /* dock delta_exec before expiring quota (as it could span periods) */
3661 cfs_rq->runtime_remaining -= delta_exec;
3662 expire_cfs_rq_runtime(cfs_rq);
3664 if (likely(cfs_rq->runtime_remaining > 0))
3668 * if we're unable to extend our runtime we resched so that the active
3669 * hierarchy can be throttled
3671 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3672 resched_curr(rq_of(cfs_rq));
3675 static __always_inline
3676 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3678 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3681 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3684 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3686 return cfs_bandwidth_used() && cfs_rq->throttled;
3689 /* check whether cfs_rq, or any parent, is throttled */
3690 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3692 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3696 * Ensure that neither of the group entities corresponding to src_cpu or
3697 * dest_cpu are members of a throttled hierarchy when performing group
3698 * load-balance operations.
3700 static inline int throttled_lb_pair(struct task_group *tg,
3701 int src_cpu, int dest_cpu)
3703 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3705 src_cfs_rq = tg->cfs_rq[src_cpu];
3706 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3708 return throttled_hierarchy(src_cfs_rq) ||
3709 throttled_hierarchy(dest_cfs_rq);
3712 /* updated child weight may affect parent so we have to do this bottom up */
3713 static int tg_unthrottle_up(struct task_group *tg, void *data)
3715 struct rq *rq = data;
3716 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3718 cfs_rq->throttle_count--;
3720 if (!cfs_rq->throttle_count) {
3721 /* adjust cfs_rq_clock_task() */
3722 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3723 cfs_rq->throttled_clock_task;
3730 static int tg_throttle_down(struct task_group *tg, void *data)
3732 struct rq *rq = data;
3733 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3735 /* group is entering throttled state, stop time */
3736 if (!cfs_rq->throttle_count)
3737 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3738 cfs_rq->throttle_count++;
3743 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3745 struct rq *rq = rq_of(cfs_rq);
3746 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3747 struct sched_entity *se;
3748 long task_delta, dequeue = 1;
3751 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3753 /* freeze hierarchy runnable averages while throttled */
3755 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3758 task_delta = cfs_rq->h_nr_running;
3759 for_each_sched_entity(se) {
3760 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3761 /* throttled entity or throttle-on-deactivate */
3766 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3767 qcfs_rq->h_nr_running -= task_delta;
3769 if (qcfs_rq->load.weight)
3774 sub_nr_running(rq, task_delta);
3776 cfs_rq->throttled = 1;
3777 cfs_rq->throttled_clock = rq_clock(rq);
3778 raw_spin_lock(&cfs_b->lock);
3779 empty = list_empty(&cfs_b->throttled_cfs_rq);
3782 * Add to the _head_ of the list, so that an already-started
3783 * distribute_cfs_runtime will not see us
3785 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3788 * If we're the first throttled task, make sure the bandwidth
3792 start_cfs_bandwidth(cfs_b);
3794 raw_spin_unlock(&cfs_b->lock);
3797 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3799 struct rq *rq = rq_of(cfs_rq);
3800 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3801 struct sched_entity *se;
3805 se = cfs_rq->tg->se[cpu_of(rq)];
3807 cfs_rq->throttled = 0;
3809 update_rq_clock(rq);
3811 raw_spin_lock(&cfs_b->lock);
3812 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3813 list_del_rcu(&cfs_rq->throttled_list);
3814 raw_spin_unlock(&cfs_b->lock);
3816 /* update hierarchical throttle state */
3817 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3819 if (!cfs_rq->load.weight)
3822 task_delta = cfs_rq->h_nr_running;
3823 for_each_sched_entity(se) {
3827 cfs_rq = cfs_rq_of(se);
3829 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3830 cfs_rq->h_nr_running += task_delta;
3832 if (cfs_rq_throttled(cfs_rq))
3837 add_nr_running(rq, task_delta);
3839 /* determine whether we need to wake up potentially idle cpu */
3840 if (rq->curr == rq->idle && rq->cfs.nr_running)
3844 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3845 u64 remaining, u64 expires)
3847 struct cfs_rq *cfs_rq;
3849 u64 starting_runtime = remaining;
3852 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3854 struct rq *rq = rq_of(cfs_rq);
3856 raw_spin_lock(&rq->lock);
3857 if (!cfs_rq_throttled(cfs_rq))
3860 runtime = -cfs_rq->runtime_remaining + 1;
3861 if (runtime > remaining)
3862 runtime = remaining;
3863 remaining -= runtime;
3865 cfs_rq->runtime_remaining += runtime;
3866 cfs_rq->runtime_expires = expires;
3868 /* we check whether we're throttled above */
3869 if (cfs_rq->runtime_remaining > 0)
3870 unthrottle_cfs_rq(cfs_rq);
3873 raw_spin_unlock(&rq->lock);
3880 return starting_runtime - remaining;
3884 * Responsible for refilling a task_group's bandwidth and unthrottling its
3885 * cfs_rqs as appropriate. If there has been no activity within the last
3886 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3887 * used to track this state.
3889 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3891 u64 runtime, runtime_expires;
3894 /* no need to continue the timer with no bandwidth constraint */
3895 if (cfs_b->quota == RUNTIME_INF)
3896 goto out_deactivate;
3898 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3899 cfs_b->nr_periods += overrun;
3902 * idle depends on !throttled (for the case of a large deficit), and if
3903 * we're going inactive then everything else can be deferred
3905 if (cfs_b->idle && !throttled)
3906 goto out_deactivate;
3908 __refill_cfs_bandwidth_runtime(cfs_b);
3911 /* mark as potentially idle for the upcoming period */
3916 /* account preceding periods in which throttling occurred */
3917 cfs_b->nr_throttled += overrun;
3919 runtime_expires = cfs_b->runtime_expires;
3922 * This check is repeated as we are holding onto the new bandwidth while
3923 * we unthrottle. This can potentially race with an unthrottled group
3924 * trying to acquire new bandwidth from the global pool. This can result
3925 * in us over-using our runtime if it is all used during this loop, but
3926 * only by limited amounts in that extreme case.
3928 while (throttled && cfs_b->runtime > 0) {
3929 runtime = cfs_b->runtime;
3930 raw_spin_unlock(&cfs_b->lock);
3931 /* we can't nest cfs_b->lock while distributing bandwidth */
3932 runtime = distribute_cfs_runtime(cfs_b, runtime,
3934 raw_spin_lock(&cfs_b->lock);
3936 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3938 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3942 * While we are ensured activity in the period following an
3943 * unthrottle, this also covers the case in which the new bandwidth is
3944 * insufficient to cover the existing bandwidth deficit. (Forcing the
3945 * timer to remain active while there are any throttled entities.)
3955 /* a cfs_rq won't donate quota below this amount */
3956 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3957 /* minimum remaining period time to redistribute slack quota */
3958 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3959 /* how long we wait to gather additional slack before distributing */
3960 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3963 * Are we near the end of the current quota period?
3965 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3966 * hrtimer base being cleared by hrtimer_start. In the case of
3967 * migrate_hrtimers, base is never cleared, so we are fine.
3969 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3971 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3974 /* if the call-back is running a quota refresh is already occurring */
3975 if (hrtimer_callback_running(refresh_timer))
3978 /* is a quota refresh about to occur? */
3979 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3980 if (remaining < min_expire)
3986 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3988 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3990 /* if there's a quota refresh soon don't bother with slack */
3991 if (runtime_refresh_within(cfs_b, min_left))
3994 hrtimer_start(&cfs_b->slack_timer,
3995 ns_to_ktime(cfs_bandwidth_slack_period),
3999 /* we know any runtime found here is valid as update_curr() precedes return */
4000 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4002 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4003 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4005 if (slack_runtime <= 0)
4008 raw_spin_lock(&cfs_b->lock);
4009 if (cfs_b->quota != RUNTIME_INF &&
4010 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4011 cfs_b->runtime += slack_runtime;
4013 /* we are under rq->lock, defer unthrottling using a timer */
4014 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4015 !list_empty(&cfs_b->throttled_cfs_rq))
4016 start_cfs_slack_bandwidth(cfs_b);
4018 raw_spin_unlock(&cfs_b->lock);
4020 /* even if it's not valid for return we don't want to try again */
4021 cfs_rq->runtime_remaining -= slack_runtime;
4024 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4026 if (!cfs_bandwidth_used())
4029 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4032 __return_cfs_rq_runtime(cfs_rq);
4036 * This is done with a timer (instead of inline with bandwidth return) since
4037 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4039 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4041 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4044 /* confirm we're still not at a refresh boundary */
4045 raw_spin_lock(&cfs_b->lock);
4046 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4047 raw_spin_unlock(&cfs_b->lock);
4051 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4052 runtime = cfs_b->runtime;
4054 expires = cfs_b->runtime_expires;
4055 raw_spin_unlock(&cfs_b->lock);
4060 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4062 raw_spin_lock(&cfs_b->lock);
4063 if (expires == cfs_b->runtime_expires)
4064 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4065 raw_spin_unlock(&cfs_b->lock);
4069 * When a group wakes up we want to make sure that its quota is not already
4070 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4071 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4073 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4075 if (!cfs_bandwidth_used())
4078 /* an active group must be handled by the update_curr()->put() path */
4079 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4082 /* ensure the group is not already throttled */
4083 if (cfs_rq_throttled(cfs_rq))
4086 /* update runtime allocation */
4087 account_cfs_rq_runtime(cfs_rq, 0);
4088 if (cfs_rq->runtime_remaining <= 0)
4089 throttle_cfs_rq(cfs_rq);
4092 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4093 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4095 if (!cfs_bandwidth_used())
4098 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4102 * it's possible for a throttled entity to be forced into a running
4103 * state (e.g. set_curr_task), in this case we're finished.
4105 if (cfs_rq_throttled(cfs_rq))
4108 throttle_cfs_rq(cfs_rq);
4112 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4114 struct cfs_bandwidth *cfs_b =
4115 container_of(timer, struct cfs_bandwidth, slack_timer);
4117 do_sched_cfs_slack_timer(cfs_b);
4119 return HRTIMER_NORESTART;
4122 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4124 struct cfs_bandwidth *cfs_b =
4125 container_of(timer, struct cfs_bandwidth, period_timer);
4129 raw_spin_lock(&cfs_b->lock);
4131 overrun = hrtimer_forward_now(timer, cfs_b->period);
4135 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4138 cfs_b->period_active = 0;
4139 raw_spin_unlock(&cfs_b->lock);
4141 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4144 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4146 raw_spin_lock_init(&cfs_b->lock);
4148 cfs_b->quota = RUNTIME_INF;
4149 cfs_b->period = ns_to_ktime(default_cfs_period());
4151 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4152 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4153 cfs_b->period_timer.function = sched_cfs_period_timer;
4154 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4155 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4158 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4160 cfs_rq->runtime_enabled = 0;
4161 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4164 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4166 lockdep_assert_held(&cfs_b->lock);
4168 if (!cfs_b->period_active) {
4169 cfs_b->period_active = 1;
4170 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4171 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4175 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4177 /* init_cfs_bandwidth() was not called */
4178 if (!cfs_b->throttled_cfs_rq.next)
4181 hrtimer_cancel(&cfs_b->period_timer);
4182 hrtimer_cancel(&cfs_b->slack_timer);
4185 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4187 struct cfs_rq *cfs_rq;
4189 for_each_leaf_cfs_rq(rq, cfs_rq) {
4190 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4192 raw_spin_lock(&cfs_b->lock);
4193 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4194 raw_spin_unlock(&cfs_b->lock);
4198 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4200 struct cfs_rq *cfs_rq;
4202 for_each_leaf_cfs_rq(rq, cfs_rq) {
4203 if (!cfs_rq->runtime_enabled)
4207 * clock_task is not advancing so we just need to make sure
4208 * there's some valid quota amount
4210 cfs_rq->runtime_remaining = 1;
4212 * Offline rq is schedulable till cpu is completely disabled
4213 * in take_cpu_down(), so we prevent new cfs throttling here.
4215 cfs_rq->runtime_enabled = 0;
4217 if (cfs_rq_throttled(cfs_rq))
4218 unthrottle_cfs_rq(cfs_rq);
4222 #else /* CONFIG_CFS_BANDWIDTH */
4223 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4225 return rq_clock_task(rq_of(cfs_rq));
4228 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4229 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4230 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4231 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4233 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4238 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4243 static inline int throttled_lb_pair(struct task_group *tg,
4244 int src_cpu, int dest_cpu)
4249 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4251 #ifdef CONFIG_FAIR_GROUP_SCHED
4252 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4255 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4259 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4260 static inline void update_runtime_enabled(struct rq *rq) {}
4261 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4263 #endif /* CONFIG_CFS_BANDWIDTH */
4265 /**************************************************
4266 * CFS operations on tasks:
4269 #ifdef CONFIG_SCHED_HRTICK
4270 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4272 struct sched_entity *se = &p->se;
4273 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4275 WARN_ON(task_rq(p) != rq);
4277 if (cfs_rq->nr_running > 1) {
4278 u64 slice = sched_slice(cfs_rq, se);
4279 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4280 s64 delta = slice - ran;
4287 hrtick_start(rq, delta);
4292 * called from enqueue/dequeue and updates the hrtick when the
4293 * current task is from our class and nr_running is low enough
4296 static void hrtick_update(struct rq *rq)
4298 struct task_struct *curr = rq->curr;
4300 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4303 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4304 hrtick_start_fair(rq, curr);
4306 #else /* !CONFIG_SCHED_HRTICK */
4308 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4312 static inline void hrtick_update(struct rq *rq)
4318 * The enqueue_task method is called before nr_running is
4319 * increased. Here we update the fair scheduling stats and
4320 * then put the task into the rbtree:
4323 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4325 struct cfs_rq *cfs_rq;
4326 struct sched_entity *se = &p->se;
4328 for_each_sched_entity(se) {
4331 cfs_rq = cfs_rq_of(se);
4332 enqueue_entity(cfs_rq, se, flags);
4335 * end evaluation on encountering a throttled cfs_rq
4337 * note: in the case of encountering a throttled cfs_rq we will
4338 * post the final h_nr_running increment below.
4340 if (cfs_rq_throttled(cfs_rq))
4342 cfs_rq->h_nr_running++;
4344 flags = ENQUEUE_WAKEUP;
4347 for_each_sched_entity(se) {
4348 cfs_rq = cfs_rq_of(se);
4349 cfs_rq->h_nr_running++;
4351 if (cfs_rq_throttled(cfs_rq))
4354 update_load_avg(se, 1);
4355 update_cfs_shares(cfs_rq);
4359 add_nr_running(rq, 1);
4364 static void set_next_buddy(struct sched_entity *se);
4367 * The dequeue_task method is called before nr_running is
4368 * decreased. We remove the task from the rbtree and
4369 * update the fair scheduling stats:
4371 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4373 struct cfs_rq *cfs_rq;
4374 struct sched_entity *se = &p->se;
4375 int task_sleep = flags & DEQUEUE_SLEEP;
4377 for_each_sched_entity(se) {
4378 cfs_rq = cfs_rq_of(se);
4379 dequeue_entity(cfs_rq, se, flags);
4382 * end evaluation on encountering a throttled cfs_rq
4384 * note: in the case of encountering a throttled cfs_rq we will
4385 * post the final h_nr_running decrement below.
4387 if (cfs_rq_throttled(cfs_rq))
4389 cfs_rq->h_nr_running--;
4391 /* Don't dequeue parent if it has other entities besides us */
4392 if (cfs_rq->load.weight) {
4394 * Bias pick_next to pick a task from this cfs_rq, as
4395 * p is sleeping when it is within its sched_slice.
4397 if (task_sleep && parent_entity(se))
4398 set_next_buddy(parent_entity(se));
4400 /* avoid re-evaluating load for this entity */
4401 se = parent_entity(se);
4404 flags |= DEQUEUE_SLEEP;
4407 for_each_sched_entity(se) {
4408 cfs_rq = cfs_rq_of(se);
4409 cfs_rq->h_nr_running--;
4411 if (cfs_rq_throttled(cfs_rq))
4414 update_load_avg(se, 1);
4415 update_cfs_shares(cfs_rq);
4419 sub_nr_running(rq, 1);
4427 * per rq 'load' arrray crap; XXX kill this.
4431 * The exact cpuload calculated at every tick would be:
4433 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4435 * If a cpu misses updates for n ticks (as it was idle) and update gets
4436 * called on the n+1-th tick when cpu may be busy, then we have:
4438 * load_n = (1 - 1/2^i)^n * load_0
4439 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4441 * decay_load_missed() below does efficient calculation of
4443 * load' = (1 - 1/2^i)^n * load
4445 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4446 * This allows us to precompute the above in said factors, thereby allowing the
4447 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4448 * fixed_power_int())
4450 * The calculation is approximated on a 128 point scale.
4452 #define DEGRADE_SHIFT 7
4454 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4455 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4456 { 0, 0, 0, 0, 0, 0, 0, 0 },
4457 { 64, 32, 8, 0, 0, 0, 0, 0 },
4458 { 96, 72, 40, 12, 1, 0, 0, 0 },
4459 { 112, 98, 75, 43, 15, 1, 0, 0 },
4460 { 120, 112, 98, 76, 45, 16, 2, 0 }
4464 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4465 * would be when CPU is idle and so we just decay the old load without
4466 * adding any new load.
4468 static unsigned long
4469 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4473 if (!missed_updates)
4476 if (missed_updates >= degrade_zero_ticks[idx])
4480 return load >> missed_updates;
4482 while (missed_updates) {
4483 if (missed_updates % 2)
4484 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4486 missed_updates >>= 1;
4493 * __update_cpu_load - update the rq->cpu_load[] statistics
4494 * @this_rq: The rq to update statistics for
4495 * @this_load: The current load
4496 * @pending_updates: The number of missed updates
4497 * @active: !0 for NOHZ_FULL
4499 * Update rq->cpu_load[] statistics. This function is usually called every
4500 * scheduler tick (TICK_NSEC).
4502 * This function computes a decaying average:
4504 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4506 * Because of NOHZ it might not get called on every tick which gives need for
4507 * the @pending_updates argument.
4509 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4510 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4511 * = A * (A * load[i]_n-2 + B) + B
4512 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4513 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4514 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4515 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4516 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4518 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4519 * any change in load would have resulted in the tick being turned back on.
4521 * For regular NOHZ, this reduces to:
4523 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4525 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4526 * term. See the @active paramter.
4528 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4529 unsigned long pending_updates, int active)
4531 unsigned long tickless_load = active ? this_rq->cpu_load[0] : 0;
4534 this_rq->nr_load_updates++;
4536 /* Update our load: */
4537 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4538 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4539 unsigned long old_load, new_load;
4541 /* scale is effectively 1 << i now, and >> i divides by scale */
4543 old_load = this_rq->cpu_load[i];
4544 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4545 if (tickless_load) {
4546 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4548 * old_load can never be a negative value because a
4549 * decayed tickless_load cannot be greater than the
4550 * original tickless_load.
4552 old_load += tickless_load;
4554 new_load = this_load;
4556 * Round up the averaging division if load is increasing. This
4557 * prevents us from getting stuck on 9 if the load is 10, for
4560 if (new_load > old_load)
4561 new_load += scale - 1;
4563 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4566 sched_avg_update(this_rq);
4569 /* Used instead of source_load when we know the type == 0 */
4570 static unsigned long weighted_cpuload(const int cpu)
4572 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4575 #ifdef CONFIG_NO_HZ_COMMON
4576 static void __update_cpu_load_nohz(struct rq *this_rq,
4577 unsigned long curr_jiffies,
4581 unsigned long pending_updates;
4583 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4584 if (pending_updates) {
4585 this_rq->last_load_update_tick = curr_jiffies;
4587 * In the regular NOHZ case, we were idle, this means load 0.
4588 * In the NOHZ_FULL case, we were non-idle, we should consider
4589 * its weighted load.
4591 __update_cpu_load(this_rq, load, pending_updates, active);
4596 * There is no sane way to deal with nohz on smp when using jiffies because the
4597 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4598 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4600 * Therefore we cannot use the delta approach from the regular tick since that
4601 * would seriously skew the load calculation. However we'll make do for those
4602 * updates happening while idle (nohz_idle_balance) or coming out of idle
4603 * (tick_nohz_idle_exit).
4605 * This means we might still be one tick off for nohz periods.
4609 * Called from nohz_idle_balance() to update the load ratings before doing the
4612 static void update_cpu_load_idle(struct rq *this_rq)
4615 * bail if there's load or we're actually up-to-date.
4617 if (weighted_cpuload(cpu_of(this_rq)))
4620 __update_cpu_load_nohz(this_rq, READ_ONCE(jiffies), 0, 0);
4624 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4626 void update_cpu_load_nohz(int active)
4628 struct rq *this_rq = this_rq();
4629 unsigned long curr_jiffies = READ_ONCE(jiffies);
4630 unsigned long load = active ? weighted_cpuload(cpu_of(this_rq)) : 0;
4632 if (curr_jiffies == this_rq->last_load_update_tick)
4635 raw_spin_lock(&this_rq->lock);
4636 __update_cpu_load_nohz(this_rq, curr_jiffies, load, active);
4637 raw_spin_unlock(&this_rq->lock);
4639 #endif /* CONFIG_NO_HZ */
4642 * Called from scheduler_tick()
4644 void update_cpu_load_active(struct rq *this_rq)
4646 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4648 * See the mess around update_cpu_load_idle() / update_cpu_load_nohz().
4650 this_rq->last_load_update_tick = jiffies;
4651 __update_cpu_load(this_rq, load, 1, 1);
4655 * Return a low guess at the load of a migration-source cpu weighted
4656 * according to the scheduling class and "nice" value.
4658 * We want to under-estimate the load of migration sources, to
4659 * balance conservatively.
4661 static unsigned long source_load(int cpu, int type)
4663 struct rq *rq = cpu_rq(cpu);
4664 unsigned long total = weighted_cpuload(cpu);
4666 if (type == 0 || !sched_feat(LB_BIAS))
4669 return min(rq->cpu_load[type-1], total);
4673 * Return a high guess at the load of a migration-target cpu weighted
4674 * according to the scheduling class and "nice" value.
4676 static unsigned long target_load(int cpu, int type)
4678 struct rq *rq = cpu_rq(cpu);
4679 unsigned long total = weighted_cpuload(cpu);
4681 if (type == 0 || !sched_feat(LB_BIAS))
4684 return max(rq->cpu_load[type-1], total);
4687 static unsigned long capacity_of(int cpu)
4689 return cpu_rq(cpu)->cpu_capacity;
4692 static unsigned long capacity_orig_of(int cpu)
4694 return cpu_rq(cpu)->cpu_capacity_orig;
4697 static unsigned long cpu_avg_load_per_task(int cpu)
4699 struct rq *rq = cpu_rq(cpu);
4700 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4701 unsigned long load_avg = weighted_cpuload(cpu);
4704 return load_avg / nr_running;
4709 static void record_wakee(struct task_struct *p)
4712 * Rough decay (wiping) for cost saving, don't worry
4713 * about the boundary, really active task won't care
4716 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4717 current->wakee_flips >>= 1;
4718 current->wakee_flip_decay_ts = jiffies;
4721 if (current->last_wakee != p) {
4722 current->last_wakee = p;
4723 current->wakee_flips++;
4727 static void task_waking_fair(struct task_struct *p)
4729 struct sched_entity *se = &p->se;
4730 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4733 #ifndef CONFIG_64BIT
4734 u64 min_vruntime_copy;
4737 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4739 min_vruntime = cfs_rq->min_vruntime;
4740 } while (min_vruntime != min_vruntime_copy);
4742 min_vruntime = cfs_rq->min_vruntime;
4745 se->vruntime -= min_vruntime;
4749 #ifdef CONFIG_FAIR_GROUP_SCHED
4751 * effective_load() calculates the load change as seen from the root_task_group
4753 * Adding load to a group doesn't make a group heavier, but can cause movement
4754 * of group shares between cpus. Assuming the shares were perfectly aligned one
4755 * can calculate the shift in shares.
4757 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4758 * on this @cpu and results in a total addition (subtraction) of @wg to the
4759 * total group weight.
4761 * Given a runqueue weight distribution (rw_i) we can compute a shares
4762 * distribution (s_i) using:
4764 * s_i = rw_i / \Sum rw_j (1)
4766 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4767 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4768 * shares distribution (s_i):
4770 * rw_i = { 2, 4, 1, 0 }
4771 * s_i = { 2/7, 4/7, 1/7, 0 }
4773 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4774 * task used to run on and the CPU the waker is running on), we need to
4775 * compute the effect of waking a task on either CPU and, in case of a sync
4776 * wakeup, compute the effect of the current task going to sleep.
4778 * So for a change of @wl to the local @cpu with an overall group weight change
4779 * of @wl we can compute the new shares distribution (s'_i) using:
4781 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4783 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4784 * differences in waking a task to CPU 0. The additional task changes the
4785 * weight and shares distributions like:
4787 * rw'_i = { 3, 4, 1, 0 }
4788 * s'_i = { 3/8, 4/8, 1/8, 0 }
4790 * We can then compute the difference in effective weight by using:
4792 * dw_i = S * (s'_i - s_i) (3)
4794 * Where 'S' is the group weight as seen by its parent.
4796 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4797 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4798 * 4/7) times the weight of the group.
4800 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4802 struct sched_entity *se = tg->se[cpu];
4804 if (!tg->parent) /* the trivial, non-cgroup case */
4807 for_each_sched_entity(se) {
4813 * W = @wg + \Sum rw_j
4815 W = wg + calc_tg_weight(tg, se->my_q);
4820 w = cfs_rq_load_avg(se->my_q) + wl;
4823 * wl = S * s'_i; see (2)
4826 wl = (w * (long)tg->shares) / W;
4831 * Per the above, wl is the new se->load.weight value; since
4832 * those are clipped to [MIN_SHARES, ...) do so now. See
4833 * calc_cfs_shares().
4835 if (wl < MIN_SHARES)
4839 * wl = dw_i = S * (s'_i - s_i); see (3)
4841 wl -= se->avg.load_avg;
4844 * Recursively apply this logic to all parent groups to compute
4845 * the final effective load change on the root group. Since
4846 * only the @tg group gets extra weight, all parent groups can
4847 * only redistribute existing shares. @wl is the shift in shares
4848 * resulting from this level per the above.
4857 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4865 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4866 * A waker of many should wake a different task than the one last awakened
4867 * at a frequency roughly N times higher than one of its wakees. In order
4868 * to determine whether we should let the load spread vs consolodating to
4869 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4870 * partner, and a factor of lls_size higher frequency in the other. With
4871 * both conditions met, we can be relatively sure that the relationship is
4872 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4873 * being client/server, worker/dispatcher, interrupt source or whatever is
4874 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4876 static int wake_wide(struct task_struct *p)
4878 unsigned int master = current->wakee_flips;
4879 unsigned int slave = p->wakee_flips;
4880 int factor = this_cpu_read(sd_llc_size);
4883 swap(master, slave);
4884 if (slave < factor || master < slave * factor)
4889 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4891 s64 this_load, load;
4892 s64 this_eff_load, prev_eff_load;
4893 int idx, this_cpu, prev_cpu;
4894 struct task_group *tg;
4895 unsigned long weight;
4899 this_cpu = smp_processor_id();
4900 prev_cpu = task_cpu(p);
4901 load = source_load(prev_cpu, idx);
4902 this_load = target_load(this_cpu, idx);
4905 * If sync wakeup then subtract the (maximum possible)
4906 * effect of the currently running task from the load
4907 * of the current CPU:
4910 tg = task_group(current);
4911 weight = current->se.avg.load_avg;
4913 this_load += effective_load(tg, this_cpu, -weight, -weight);
4914 load += effective_load(tg, prev_cpu, 0, -weight);
4918 weight = p->se.avg.load_avg;
4921 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4922 * due to the sync cause above having dropped this_load to 0, we'll
4923 * always have an imbalance, but there's really nothing you can do
4924 * about that, so that's good too.
4926 * Otherwise check if either cpus are near enough in load to allow this
4927 * task to be woken on this_cpu.
4929 this_eff_load = 100;
4930 this_eff_load *= capacity_of(prev_cpu);
4932 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4933 prev_eff_load *= capacity_of(this_cpu);
4935 if (this_load > 0) {
4936 this_eff_load *= this_load +
4937 effective_load(tg, this_cpu, weight, weight);
4939 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4942 balanced = this_eff_load <= prev_eff_load;
4944 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4949 schedstat_inc(sd, ttwu_move_affine);
4950 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4956 * find_idlest_group finds and returns the least busy CPU group within the
4959 static struct sched_group *
4960 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4961 int this_cpu, int sd_flag)
4963 struct sched_group *idlest = NULL, *group = sd->groups;
4964 unsigned long min_load = ULONG_MAX, this_load = 0;
4965 int load_idx = sd->forkexec_idx;
4966 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4968 if (sd_flag & SD_BALANCE_WAKE)
4969 load_idx = sd->wake_idx;
4972 unsigned long load, avg_load;
4976 /* Skip over this group if it has no CPUs allowed */
4977 if (!cpumask_intersects(sched_group_cpus(group),
4978 tsk_cpus_allowed(p)))
4981 local_group = cpumask_test_cpu(this_cpu,
4982 sched_group_cpus(group));
4984 /* Tally up the load of all CPUs in the group */
4987 for_each_cpu(i, sched_group_cpus(group)) {
4988 /* Bias balancing toward cpus of our domain */
4990 load = source_load(i, load_idx);
4992 load = target_load(i, load_idx);
4997 /* Adjust by relative CPU capacity of the group */
4998 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5001 this_load = avg_load;
5002 } else if (avg_load < min_load) {
5003 min_load = avg_load;
5006 } while (group = group->next, group != sd->groups);
5008 if (!idlest || 100*this_load < imbalance*min_load)
5014 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5017 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5019 unsigned long load, min_load = ULONG_MAX;
5020 unsigned int min_exit_latency = UINT_MAX;
5021 u64 latest_idle_timestamp = 0;
5022 int least_loaded_cpu = this_cpu;
5023 int shallowest_idle_cpu = -1;
5026 /* Traverse only the allowed CPUs */
5027 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5029 struct rq *rq = cpu_rq(i);
5030 struct cpuidle_state *idle = idle_get_state(rq);
5031 if (idle && idle->exit_latency < min_exit_latency) {
5033 * We give priority to a CPU whose idle state
5034 * has the smallest exit latency irrespective
5035 * of any idle timestamp.
5037 min_exit_latency = idle->exit_latency;
5038 latest_idle_timestamp = rq->idle_stamp;
5039 shallowest_idle_cpu = i;
5040 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5041 rq->idle_stamp > latest_idle_timestamp) {
5043 * If equal or no active idle state, then
5044 * the most recently idled CPU might have
5047 latest_idle_timestamp = rq->idle_stamp;
5048 shallowest_idle_cpu = i;
5050 } else if (shallowest_idle_cpu == -1) {
5051 load = weighted_cpuload(i);
5052 if (load < min_load || (load == min_load && i == this_cpu)) {
5054 least_loaded_cpu = i;
5059 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5063 * Try and locate an idle CPU in the sched_domain.
5065 static int select_idle_sibling(struct task_struct *p, int target)
5067 struct sched_domain *sd;
5068 struct sched_group *sg;
5069 int i = task_cpu(p);
5071 if (idle_cpu(target))
5075 * If the prevous cpu is cache affine and idle, don't be stupid.
5077 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5081 * Otherwise, iterate the domains and find an eligible idle cpu.
5083 * A completely idle sched group at higher domains is more
5084 * desirable than an idle group at a lower level, because lower
5085 * domains have smaller groups and usually share hardware
5086 * resources which causes tasks to contend on them, e.g. x86
5087 * hyperthread siblings in the lowest domain (SMT) can contend
5088 * on the shared cpu pipeline.
5090 * However, while we prefer idle groups at higher domains
5091 * finding an idle cpu at the lowest domain is still better than
5092 * returning 'target', which we've already established, isn't
5095 sd = rcu_dereference(per_cpu(sd_llc, target));
5096 for_each_lower_domain(sd) {
5099 if (!cpumask_intersects(sched_group_cpus(sg),
5100 tsk_cpus_allowed(p)))
5103 /* Ensure the entire group is idle */
5104 for_each_cpu(i, sched_group_cpus(sg)) {
5105 if (i == target || !idle_cpu(i))
5110 * It doesn't matter which cpu we pick, the
5111 * whole group is idle.
5113 target = cpumask_first_and(sched_group_cpus(sg),
5114 tsk_cpus_allowed(p));
5118 } while (sg != sd->groups);
5125 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5126 * tasks. The unit of the return value must be the one of capacity so we can
5127 * compare the utilization with the capacity of the CPU that is available for
5128 * CFS task (ie cpu_capacity).
5130 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5131 * recent utilization of currently non-runnable tasks on a CPU. It represents
5132 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5133 * capacity_orig is the cpu_capacity available at the highest frequency
5134 * (arch_scale_freq_capacity()).
5135 * The utilization of a CPU converges towards a sum equal to or less than the
5136 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5137 * the running time on this CPU scaled by capacity_curr.
5139 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5140 * higher than capacity_orig because of unfortunate rounding in
5141 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5142 * the average stabilizes with the new running time. We need to check that the
5143 * utilization stays within the range of [0..capacity_orig] and cap it if
5144 * necessary. Without utilization capping, a group could be seen as overloaded
5145 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5146 * available capacity. We allow utilization to overshoot capacity_curr (but not
5147 * capacity_orig) as it useful for predicting the capacity required after task
5148 * migrations (scheduler-driven DVFS).
5150 static int cpu_util(int cpu)
5152 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5153 unsigned long capacity = capacity_orig_of(cpu);
5155 return (util >= capacity) ? capacity : util;
5159 * select_task_rq_fair: Select target runqueue for the waking task in domains
5160 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5161 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5163 * Balances load by selecting the idlest cpu in the idlest group, or under
5164 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5166 * Returns the target cpu number.
5168 * preempt must be disabled.
5171 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5173 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5174 int cpu = smp_processor_id();
5175 int new_cpu = prev_cpu;
5176 int want_affine = 0;
5177 int sync = wake_flags & WF_SYNC;
5179 if (sd_flag & SD_BALANCE_WAKE)
5180 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5183 for_each_domain(cpu, tmp) {
5184 if (!(tmp->flags & SD_LOAD_BALANCE))
5188 * If both cpu and prev_cpu are part of this domain,
5189 * cpu is a valid SD_WAKE_AFFINE target.
5191 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5192 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5197 if (tmp->flags & sd_flag)
5199 else if (!want_affine)
5204 sd = NULL; /* Prefer wake_affine over balance flags */
5205 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5210 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5211 new_cpu = select_idle_sibling(p, new_cpu);
5214 struct sched_group *group;
5217 if (!(sd->flags & sd_flag)) {
5222 group = find_idlest_group(sd, p, cpu, sd_flag);
5228 new_cpu = find_idlest_cpu(group, p, cpu);
5229 if (new_cpu == -1 || new_cpu == cpu) {
5230 /* Now try balancing at a lower domain level of cpu */
5235 /* Now try balancing at a lower domain level of new_cpu */
5237 weight = sd->span_weight;
5239 for_each_domain(cpu, tmp) {
5240 if (weight <= tmp->span_weight)
5242 if (tmp->flags & sd_flag)
5245 /* while loop will break here if sd == NULL */
5253 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5254 * cfs_rq_of(p) references at time of call are still valid and identify the
5255 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5257 static void migrate_task_rq_fair(struct task_struct *p)
5260 * We are supposed to update the task to "current" time, then its up to date
5261 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5262 * what current time is, so simply throw away the out-of-date time. This
5263 * will result in the wakee task is less decayed, but giving the wakee more
5264 * load sounds not bad.
5266 remove_entity_load_avg(&p->se);
5268 /* Tell new CPU we are migrated */
5269 p->se.avg.last_update_time = 0;
5271 /* We have migrated, no longer consider this task hot */
5272 p->se.exec_start = 0;
5275 static void task_dead_fair(struct task_struct *p)
5277 remove_entity_load_avg(&p->se);
5279 #endif /* CONFIG_SMP */
5281 static unsigned long
5282 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5284 unsigned long gran = sysctl_sched_wakeup_granularity;
5287 * Since its curr running now, convert the gran from real-time
5288 * to virtual-time in his units.
5290 * By using 'se' instead of 'curr' we penalize light tasks, so
5291 * they get preempted easier. That is, if 'se' < 'curr' then
5292 * the resulting gran will be larger, therefore penalizing the
5293 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5294 * be smaller, again penalizing the lighter task.
5296 * This is especially important for buddies when the leftmost
5297 * task is higher priority than the buddy.
5299 return calc_delta_fair(gran, se);
5303 * Should 'se' preempt 'curr'.
5317 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5319 s64 gran, vdiff = curr->vruntime - se->vruntime;
5324 gran = wakeup_gran(curr, se);
5331 static void set_last_buddy(struct sched_entity *se)
5333 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5336 for_each_sched_entity(se)
5337 cfs_rq_of(se)->last = se;
5340 static void set_next_buddy(struct sched_entity *se)
5342 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5345 for_each_sched_entity(se)
5346 cfs_rq_of(se)->next = se;
5349 static void set_skip_buddy(struct sched_entity *se)
5351 for_each_sched_entity(se)
5352 cfs_rq_of(se)->skip = se;
5356 * Preempt the current task with a newly woken task if needed:
5358 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5360 struct task_struct *curr = rq->curr;
5361 struct sched_entity *se = &curr->se, *pse = &p->se;
5362 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5363 int scale = cfs_rq->nr_running >= sched_nr_latency;
5364 int next_buddy_marked = 0;
5366 if (unlikely(se == pse))
5370 * This is possible from callers such as attach_tasks(), in which we
5371 * unconditionally check_prempt_curr() after an enqueue (which may have
5372 * lead to a throttle). This both saves work and prevents false
5373 * next-buddy nomination below.
5375 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5378 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5379 set_next_buddy(pse);
5380 next_buddy_marked = 1;
5384 * We can come here with TIF_NEED_RESCHED already set from new task
5387 * Note: this also catches the edge-case of curr being in a throttled
5388 * group (e.g. via set_curr_task), since update_curr() (in the
5389 * enqueue of curr) will have resulted in resched being set. This
5390 * prevents us from potentially nominating it as a false LAST_BUDDY
5393 if (test_tsk_need_resched(curr))
5396 /* Idle tasks are by definition preempted by non-idle tasks. */
5397 if (unlikely(curr->policy == SCHED_IDLE) &&
5398 likely(p->policy != SCHED_IDLE))
5402 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5403 * is driven by the tick):
5405 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5408 find_matching_se(&se, &pse);
5409 update_curr(cfs_rq_of(se));
5411 if (wakeup_preempt_entity(se, pse) == 1) {
5413 * Bias pick_next to pick the sched entity that is
5414 * triggering this preemption.
5416 if (!next_buddy_marked)
5417 set_next_buddy(pse);
5426 * Only set the backward buddy when the current task is still
5427 * on the rq. This can happen when a wakeup gets interleaved
5428 * with schedule on the ->pre_schedule() or idle_balance()
5429 * point, either of which can * drop the rq lock.
5431 * Also, during early boot the idle thread is in the fair class,
5432 * for obvious reasons its a bad idea to schedule back to it.
5434 if (unlikely(!se->on_rq || curr == rq->idle))
5437 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5441 static struct task_struct *
5442 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5444 struct cfs_rq *cfs_rq = &rq->cfs;
5445 struct sched_entity *se;
5446 struct task_struct *p;
5450 #ifdef CONFIG_FAIR_GROUP_SCHED
5451 if (!cfs_rq->nr_running)
5454 if (prev->sched_class != &fair_sched_class)
5458 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5459 * likely that a next task is from the same cgroup as the current.
5461 * Therefore attempt to avoid putting and setting the entire cgroup
5462 * hierarchy, only change the part that actually changes.
5466 struct sched_entity *curr = cfs_rq->curr;
5469 * Since we got here without doing put_prev_entity() we also
5470 * have to consider cfs_rq->curr. If it is still a runnable
5471 * entity, update_curr() will update its vruntime, otherwise
5472 * forget we've ever seen it.
5476 update_curr(cfs_rq);
5481 * This call to check_cfs_rq_runtime() will do the
5482 * throttle and dequeue its entity in the parent(s).
5483 * Therefore the 'simple' nr_running test will indeed
5486 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5490 se = pick_next_entity(cfs_rq, curr);
5491 cfs_rq = group_cfs_rq(se);
5497 * Since we haven't yet done put_prev_entity and if the selected task
5498 * is a different task than we started out with, try and touch the
5499 * least amount of cfs_rqs.
5502 struct sched_entity *pse = &prev->se;
5504 while (!(cfs_rq = is_same_group(se, pse))) {
5505 int se_depth = se->depth;
5506 int pse_depth = pse->depth;
5508 if (se_depth <= pse_depth) {
5509 put_prev_entity(cfs_rq_of(pse), pse);
5510 pse = parent_entity(pse);
5512 if (se_depth >= pse_depth) {
5513 set_next_entity(cfs_rq_of(se), se);
5514 se = parent_entity(se);
5518 put_prev_entity(cfs_rq, pse);
5519 set_next_entity(cfs_rq, se);
5522 if (hrtick_enabled(rq))
5523 hrtick_start_fair(rq, p);
5530 if (!cfs_rq->nr_running)
5533 put_prev_task(rq, prev);
5536 se = pick_next_entity(cfs_rq, NULL);
5537 set_next_entity(cfs_rq, se);
5538 cfs_rq = group_cfs_rq(se);
5543 if (hrtick_enabled(rq))
5544 hrtick_start_fair(rq, p);
5550 * This is OK, because current is on_cpu, which avoids it being picked
5551 * for load-balance and preemption/IRQs are still disabled avoiding
5552 * further scheduler activity on it and we're being very careful to
5553 * re-start the picking loop.
5555 lockdep_unpin_lock(&rq->lock);
5556 new_tasks = idle_balance(rq);
5557 lockdep_pin_lock(&rq->lock);
5559 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5560 * possible for any higher priority task to appear. In that case we
5561 * must re-start the pick_next_entity() loop.
5573 * Account for a descheduled task:
5575 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5577 struct sched_entity *se = &prev->se;
5578 struct cfs_rq *cfs_rq;
5580 for_each_sched_entity(se) {
5581 cfs_rq = cfs_rq_of(se);
5582 put_prev_entity(cfs_rq, se);
5587 * sched_yield() is very simple
5589 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5591 static void yield_task_fair(struct rq *rq)
5593 struct task_struct *curr = rq->curr;
5594 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5595 struct sched_entity *se = &curr->se;
5598 * Are we the only task in the tree?
5600 if (unlikely(rq->nr_running == 1))
5603 clear_buddies(cfs_rq, se);
5605 if (curr->policy != SCHED_BATCH) {
5606 update_rq_clock(rq);
5608 * Update run-time statistics of the 'current'.
5610 update_curr(cfs_rq);
5612 * Tell update_rq_clock() that we've just updated,
5613 * so we don't do microscopic update in schedule()
5614 * and double the fastpath cost.
5616 rq_clock_skip_update(rq, true);
5622 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5624 struct sched_entity *se = &p->se;
5626 /* throttled hierarchies are not runnable */
5627 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5630 /* Tell the scheduler that we'd really like pse to run next. */
5633 yield_task_fair(rq);
5639 /**************************************************
5640 * Fair scheduling class load-balancing methods.
5644 * The purpose of load-balancing is to achieve the same basic fairness the
5645 * per-cpu scheduler provides, namely provide a proportional amount of compute
5646 * time to each task. This is expressed in the following equation:
5648 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5650 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5651 * W_i,0 is defined as:
5653 * W_i,0 = \Sum_j w_i,j (2)
5655 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5656 * is derived from the nice value as per prio_to_weight[].
5658 * The weight average is an exponential decay average of the instantaneous
5661 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5663 * C_i is the compute capacity of cpu i, typically it is the
5664 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5665 * can also include other factors [XXX].
5667 * To achieve this balance we define a measure of imbalance which follows
5668 * directly from (1):
5670 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5672 * We them move tasks around to minimize the imbalance. In the continuous
5673 * function space it is obvious this converges, in the discrete case we get
5674 * a few fun cases generally called infeasible weight scenarios.
5677 * - infeasible weights;
5678 * - local vs global optima in the discrete case. ]
5683 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5684 * for all i,j solution, we create a tree of cpus that follows the hardware
5685 * topology where each level pairs two lower groups (or better). This results
5686 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5687 * tree to only the first of the previous level and we decrease the frequency
5688 * of load-balance at each level inv. proportional to the number of cpus in
5694 * \Sum { --- * --- * 2^i } = O(n) (5)
5696 * `- size of each group
5697 * | | `- number of cpus doing load-balance
5699 * `- sum over all levels
5701 * Coupled with a limit on how many tasks we can migrate every balance pass,
5702 * this makes (5) the runtime complexity of the balancer.
5704 * An important property here is that each CPU is still (indirectly) connected
5705 * to every other cpu in at most O(log n) steps:
5707 * The adjacency matrix of the resulting graph is given by:
5710 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5713 * And you'll find that:
5715 * A^(log_2 n)_i,j != 0 for all i,j (7)
5717 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5718 * The task movement gives a factor of O(m), giving a convergence complexity
5721 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5726 * In order to avoid CPUs going idle while there's still work to do, new idle
5727 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5728 * tree itself instead of relying on other CPUs to bring it work.
5730 * This adds some complexity to both (5) and (8) but it reduces the total idle
5738 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5741 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5746 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5748 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5750 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5753 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5754 * rewrite all of this once again.]
5757 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5759 enum fbq_type { regular, remote, all };
5761 #define LBF_ALL_PINNED 0x01
5762 #define LBF_NEED_BREAK 0x02
5763 #define LBF_DST_PINNED 0x04
5764 #define LBF_SOME_PINNED 0x08
5767 struct sched_domain *sd;
5775 struct cpumask *dst_grpmask;
5777 enum cpu_idle_type idle;
5779 /* The set of CPUs under consideration for load-balancing */
5780 struct cpumask *cpus;
5785 unsigned int loop_break;
5786 unsigned int loop_max;
5788 enum fbq_type fbq_type;
5789 struct list_head tasks;
5793 * Is this task likely cache-hot:
5795 static int task_hot(struct task_struct *p, struct lb_env *env)
5799 lockdep_assert_held(&env->src_rq->lock);
5801 if (p->sched_class != &fair_sched_class)
5804 if (unlikely(p->policy == SCHED_IDLE))
5808 * Buddy candidates are cache hot:
5810 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5811 (&p->se == cfs_rq_of(&p->se)->next ||
5812 &p->se == cfs_rq_of(&p->se)->last))
5815 if (sysctl_sched_migration_cost == -1)
5817 if (sysctl_sched_migration_cost == 0)
5820 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5822 return delta < (s64)sysctl_sched_migration_cost;
5825 #ifdef CONFIG_NUMA_BALANCING
5827 * Returns 1, if task migration degrades locality
5828 * Returns 0, if task migration improves locality i.e migration preferred.
5829 * Returns -1, if task migration is not affected by locality.
5831 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5833 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5834 unsigned long src_faults, dst_faults;
5835 int src_nid, dst_nid;
5837 if (!static_branch_likely(&sched_numa_balancing))
5840 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5843 src_nid = cpu_to_node(env->src_cpu);
5844 dst_nid = cpu_to_node(env->dst_cpu);
5846 if (src_nid == dst_nid)
5849 /* Migrating away from the preferred node is always bad. */
5850 if (src_nid == p->numa_preferred_nid) {
5851 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5857 /* Encourage migration to the preferred node. */
5858 if (dst_nid == p->numa_preferred_nid)
5862 src_faults = group_faults(p, src_nid);
5863 dst_faults = group_faults(p, dst_nid);
5865 src_faults = task_faults(p, src_nid);
5866 dst_faults = task_faults(p, dst_nid);
5869 return dst_faults < src_faults;
5873 static inline int migrate_degrades_locality(struct task_struct *p,
5881 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5884 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5888 lockdep_assert_held(&env->src_rq->lock);
5891 * We do not migrate tasks that are:
5892 * 1) throttled_lb_pair, or
5893 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5894 * 3) running (obviously), or
5895 * 4) are cache-hot on their current CPU.
5897 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5900 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5903 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5905 env->flags |= LBF_SOME_PINNED;
5908 * Remember if this task can be migrated to any other cpu in
5909 * our sched_group. We may want to revisit it if we couldn't
5910 * meet load balance goals by pulling other tasks on src_cpu.
5912 * Also avoid computing new_dst_cpu if we have already computed
5913 * one in current iteration.
5915 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5918 /* Prevent to re-select dst_cpu via env's cpus */
5919 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5920 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5921 env->flags |= LBF_DST_PINNED;
5922 env->new_dst_cpu = cpu;
5930 /* Record that we found atleast one task that could run on dst_cpu */
5931 env->flags &= ~LBF_ALL_PINNED;
5933 if (task_running(env->src_rq, p)) {
5934 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5939 * Aggressive migration if:
5940 * 1) destination numa is preferred
5941 * 2) task is cache cold, or
5942 * 3) too many balance attempts have failed.
5944 tsk_cache_hot = migrate_degrades_locality(p, env);
5945 if (tsk_cache_hot == -1)
5946 tsk_cache_hot = task_hot(p, env);
5948 if (tsk_cache_hot <= 0 ||
5949 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5950 if (tsk_cache_hot == 1) {
5951 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5952 schedstat_inc(p, se.statistics.nr_forced_migrations);
5957 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5962 * detach_task() -- detach the task for the migration specified in env
5964 static void detach_task(struct task_struct *p, struct lb_env *env)
5966 lockdep_assert_held(&env->src_rq->lock);
5968 p->on_rq = TASK_ON_RQ_MIGRATING;
5969 deactivate_task(env->src_rq, p, 0);
5970 set_task_cpu(p, env->dst_cpu);
5974 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5975 * part of active balancing operations within "domain".
5977 * Returns a task if successful and NULL otherwise.
5979 static struct task_struct *detach_one_task(struct lb_env *env)
5981 struct task_struct *p, *n;
5983 lockdep_assert_held(&env->src_rq->lock);
5985 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5986 if (!can_migrate_task(p, env))
5989 detach_task(p, env);
5992 * Right now, this is only the second place where
5993 * lb_gained[env->idle] is updated (other is detach_tasks)
5994 * so we can safely collect stats here rather than
5995 * inside detach_tasks().
5997 schedstat_inc(env->sd, lb_gained[env->idle]);
6003 static const unsigned int sched_nr_migrate_break = 32;
6006 * detach_tasks() -- tries to detach up to imbalance weighted load from
6007 * busiest_rq, as part of a balancing operation within domain "sd".
6009 * Returns number of detached tasks if successful and 0 otherwise.
6011 static int detach_tasks(struct lb_env *env)
6013 struct list_head *tasks = &env->src_rq->cfs_tasks;
6014 struct task_struct *p;
6018 lockdep_assert_held(&env->src_rq->lock);
6020 if (env->imbalance <= 0)
6023 while (!list_empty(tasks)) {
6025 * We don't want to steal all, otherwise we may be treated likewise,
6026 * which could at worst lead to a livelock crash.
6028 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6031 p = list_first_entry(tasks, struct task_struct, se.group_node);
6034 /* We've more or less seen every task there is, call it quits */
6035 if (env->loop > env->loop_max)
6038 /* take a breather every nr_migrate tasks */
6039 if (env->loop > env->loop_break) {
6040 env->loop_break += sched_nr_migrate_break;
6041 env->flags |= LBF_NEED_BREAK;
6045 if (!can_migrate_task(p, env))
6048 load = task_h_load(p);
6050 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6053 if ((load / 2) > env->imbalance)
6056 detach_task(p, env);
6057 list_add(&p->se.group_node, &env->tasks);
6060 env->imbalance -= load;
6062 #ifdef CONFIG_PREEMPT
6064 * NEWIDLE balancing is a source of latency, so preemptible
6065 * kernels will stop after the first task is detached to minimize
6066 * the critical section.
6068 if (env->idle == CPU_NEWLY_IDLE)
6073 * We only want to steal up to the prescribed amount of
6076 if (env->imbalance <= 0)
6081 list_move_tail(&p->se.group_node, tasks);
6085 * Right now, this is one of only two places we collect this stat
6086 * so we can safely collect detach_one_task() stats here rather
6087 * than inside detach_one_task().
6089 schedstat_add(env->sd, lb_gained[env->idle], detached);
6095 * attach_task() -- attach the task detached by detach_task() to its new rq.
6097 static void attach_task(struct rq *rq, struct task_struct *p)
6099 lockdep_assert_held(&rq->lock);
6101 BUG_ON(task_rq(p) != rq);
6102 activate_task(rq, p, 0);
6103 p->on_rq = TASK_ON_RQ_QUEUED;
6104 check_preempt_curr(rq, p, 0);
6108 * attach_one_task() -- attaches the task returned from detach_one_task() to
6111 static void attach_one_task(struct rq *rq, struct task_struct *p)
6113 raw_spin_lock(&rq->lock);
6115 raw_spin_unlock(&rq->lock);
6119 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6122 static void attach_tasks(struct lb_env *env)
6124 struct list_head *tasks = &env->tasks;
6125 struct task_struct *p;
6127 raw_spin_lock(&env->dst_rq->lock);
6129 while (!list_empty(tasks)) {
6130 p = list_first_entry(tasks, struct task_struct, se.group_node);
6131 list_del_init(&p->se.group_node);
6133 attach_task(env->dst_rq, p);
6136 raw_spin_unlock(&env->dst_rq->lock);
6139 #ifdef CONFIG_FAIR_GROUP_SCHED
6140 static void update_blocked_averages(int cpu)
6142 struct rq *rq = cpu_rq(cpu);
6143 struct cfs_rq *cfs_rq;
6144 unsigned long flags;
6146 raw_spin_lock_irqsave(&rq->lock, flags);
6147 update_rq_clock(rq);
6150 * Iterates the task_group tree in a bottom up fashion, see
6151 * list_add_leaf_cfs_rq() for details.
6153 for_each_leaf_cfs_rq(rq, cfs_rq) {
6154 /* throttled entities do not contribute to load */
6155 if (throttled_hierarchy(cfs_rq))
6158 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6159 update_tg_load_avg(cfs_rq, 0);
6161 raw_spin_unlock_irqrestore(&rq->lock, flags);
6165 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6166 * This needs to be done in a top-down fashion because the load of a child
6167 * group is a fraction of its parents load.
6169 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6171 struct rq *rq = rq_of(cfs_rq);
6172 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6173 unsigned long now = jiffies;
6176 if (cfs_rq->last_h_load_update == now)
6179 cfs_rq->h_load_next = NULL;
6180 for_each_sched_entity(se) {
6181 cfs_rq = cfs_rq_of(se);
6182 cfs_rq->h_load_next = se;
6183 if (cfs_rq->last_h_load_update == now)
6188 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6189 cfs_rq->last_h_load_update = now;
6192 while ((se = cfs_rq->h_load_next) != NULL) {
6193 load = cfs_rq->h_load;
6194 load = div64_ul(load * se->avg.load_avg,
6195 cfs_rq_load_avg(cfs_rq) + 1);
6196 cfs_rq = group_cfs_rq(se);
6197 cfs_rq->h_load = load;
6198 cfs_rq->last_h_load_update = now;
6202 static unsigned long task_h_load(struct task_struct *p)
6204 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6206 update_cfs_rq_h_load(cfs_rq);
6207 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6208 cfs_rq_load_avg(cfs_rq) + 1);
6211 static inline void update_blocked_averages(int cpu)
6213 struct rq *rq = cpu_rq(cpu);
6214 struct cfs_rq *cfs_rq = &rq->cfs;
6215 unsigned long flags;
6217 raw_spin_lock_irqsave(&rq->lock, flags);
6218 update_rq_clock(rq);
6219 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6220 raw_spin_unlock_irqrestore(&rq->lock, flags);
6223 static unsigned long task_h_load(struct task_struct *p)
6225 return p->se.avg.load_avg;
6229 /********** Helpers for find_busiest_group ************************/
6238 * sg_lb_stats - stats of a sched_group required for load_balancing
6240 struct sg_lb_stats {
6241 unsigned long avg_load; /*Avg load across the CPUs of the group */
6242 unsigned long group_load; /* Total load over the CPUs of the group */
6243 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6244 unsigned long load_per_task;
6245 unsigned long group_capacity;
6246 unsigned long group_util; /* Total utilization of the group */
6247 unsigned int sum_nr_running; /* Nr tasks running in the group */
6248 unsigned int idle_cpus;
6249 unsigned int group_weight;
6250 enum group_type group_type;
6251 int group_no_capacity;
6252 #ifdef CONFIG_NUMA_BALANCING
6253 unsigned int nr_numa_running;
6254 unsigned int nr_preferred_running;
6259 * sd_lb_stats - Structure to store the statistics of a sched_domain
6260 * during load balancing.
6262 struct sd_lb_stats {
6263 struct sched_group *busiest; /* Busiest group in this sd */
6264 struct sched_group *local; /* Local group in this sd */
6265 unsigned long total_load; /* Total load of all groups in sd */
6266 unsigned long total_capacity; /* Total capacity of all groups in sd */
6267 unsigned long avg_load; /* Average load across all groups in sd */
6269 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6270 struct sg_lb_stats local_stat; /* Statistics of the local group */
6273 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6276 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6277 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6278 * We must however clear busiest_stat::avg_load because
6279 * update_sd_pick_busiest() reads this before assignment.
6281 *sds = (struct sd_lb_stats){
6285 .total_capacity = 0UL,
6288 .sum_nr_running = 0,
6289 .group_type = group_other,
6295 * get_sd_load_idx - Obtain the load index for a given sched domain.
6296 * @sd: The sched_domain whose load_idx is to be obtained.
6297 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6299 * Return: The load index.
6301 static inline int get_sd_load_idx(struct sched_domain *sd,
6302 enum cpu_idle_type idle)
6308 load_idx = sd->busy_idx;
6311 case CPU_NEWLY_IDLE:
6312 load_idx = sd->newidle_idx;
6315 load_idx = sd->idle_idx;
6322 static unsigned long scale_rt_capacity(int cpu)
6324 struct rq *rq = cpu_rq(cpu);
6325 u64 total, used, age_stamp, avg;
6329 * Since we're reading these variables without serialization make sure
6330 * we read them once before doing sanity checks on them.
6332 age_stamp = READ_ONCE(rq->age_stamp);
6333 avg = READ_ONCE(rq->rt_avg);
6334 delta = __rq_clock_broken(rq) - age_stamp;
6336 if (unlikely(delta < 0))
6339 total = sched_avg_period() + delta;
6341 used = div_u64(avg, total);
6343 if (likely(used < SCHED_CAPACITY_SCALE))
6344 return SCHED_CAPACITY_SCALE - used;
6349 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6351 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6352 struct sched_group *sdg = sd->groups;
6354 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6356 capacity *= scale_rt_capacity(cpu);
6357 capacity >>= SCHED_CAPACITY_SHIFT;
6362 cpu_rq(cpu)->cpu_capacity = capacity;
6363 sdg->sgc->capacity = capacity;
6366 void update_group_capacity(struct sched_domain *sd, int cpu)
6368 struct sched_domain *child = sd->child;
6369 struct sched_group *group, *sdg = sd->groups;
6370 unsigned long capacity;
6371 unsigned long interval;
6373 interval = msecs_to_jiffies(sd->balance_interval);
6374 interval = clamp(interval, 1UL, max_load_balance_interval);
6375 sdg->sgc->next_update = jiffies + interval;
6378 update_cpu_capacity(sd, cpu);
6384 if (child->flags & SD_OVERLAP) {
6386 * SD_OVERLAP domains cannot assume that child groups
6387 * span the current group.
6390 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6391 struct sched_group_capacity *sgc;
6392 struct rq *rq = cpu_rq(cpu);
6395 * build_sched_domains() -> init_sched_groups_capacity()
6396 * gets here before we've attached the domains to the
6399 * Use capacity_of(), which is set irrespective of domains
6400 * in update_cpu_capacity().
6402 * This avoids capacity from being 0 and
6403 * causing divide-by-zero issues on boot.
6405 if (unlikely(!rq->sd)) {
6406 capacity += capacity_of(cpu);
6410 sgc = rq->sd->groups->sgc;
6411 capacity += sgc->capacity;
6415 * !SD_OVERLAP domains can assume that child groups
6416 * span the current group.
6419 group = child->groups;
6421 capacity += group->sgc->capacity;
6422 group = group->next;
6423 } while (group != child->groups);
6426 sdg->sgc->capacity = capacity;
6430 * Check whether the capacity of the rq has been noticeably reduced by side
6431 * activity. The imbalance_pct is used for the threshold.
6432 * Return true is the capacity is reduced
6435 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6437 return ((rq->cpu_capacity * sd->imbalance_pct) <
6438 (rq->cpu_capacity_orig * 100));
6442 * Group imbalance indicates (and tries to solve) the problem where balancing
6443 * groups is inadequate due to tsk_cpus_allowed() constraints.
6445 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6446 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6449 * { 0 1 2 3 } { 4 5 6 7 }
6452 * If we were to balance group-wise we'd place two tasks in the first group and
6453 * two tasks in the second group. Clearly this is undesired as it will overload
6454 * cpu 3 and leave one of the cpus in the second group unused.
6456 * The current solution to this issue is detecting the skew in the first group
6457 * by noticing the lower domain failed to reach balance and had difficulty
6458 * moving tasks due to affinity constraints.
6460 * When this is so detected; this group becomes a candidate for busiest; see
6461 * update_sd_pick_busiest(). And calculate_imbalance() and
6462 * find_busiest_group() avoid some of the usual balance conditions to allow it
6463 * to create an effective group imbalance.
6465 * This is a somewhat tricky proposition since the next run might not find the
6466 * group imbalance and decide the groups need to be balanced again. A most
6467 * subtle and fragile situation.
6470 static inline int sg_imbalanced(struct sched_group *group)
6472 return group->sgc->imbalance;
6476 * group_has_capacity returns true if the group has spare capacity that could
6477 * be used by some tasks.
6478 * We consider that a group has spare capacity if the * number of task is
6479 * smaller than the number of CPUs or if the utilization is lower than the
6480 * available capacity for CFS tasks.
6481 * For the latter, we use a threshold to stabilize the state, to take into
6482 * account the variance of the tasks' load and to return true if the available
6483 * capacity in meaningful for the load balancer.
6484 * As an example, an available capacity of 1% can appear but it doesn't make
6485 * any benefit for the load balance.
6488 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6490 if (sgs->sum_nr_running < sgs->group_weight)
6493 if ((sgs->group_capacity * 100) >
6494 (sgs->group_util * env->sd->imbalance_pct))
6501 * group_is_overloaded returns true if the group has more tasks than it can
6503 * group_is_overloaded is not equals to !group_has_capacity because a group
6504 * with the exact right number of tasks, has no more spare capacity but is not
6505 * overloaded so both group_has_capacity and group_is_overloaded return
6509 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6511 if (sgs->sum_nr_running <= sgs->group_weight)
6514 if ((sgs->group_capacity * 100) <
6515 (sgs->group_util * env->sd->imbalance_pct))
6522 group_type group_classify(struct sched_group *group,
6523 struct sg_lb_stats *sgs)
6525 if (sgs->group_no_capacity)
6526 return group_overloaded;
6528 if (sg_imbalanced(group))
6529 return group_imbalanced;
6535 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6536 * @env: The load balancing environment.
6537 * @group: sched_group whose statistics are to be updated.
6538 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6539 * @local_group: Does group contain this_cpu.
6540 * @sgs: variable to hold the statistics for this group.
6541 * @overload: Indicate more than one runnable task for any CPU.
6543 static inline void update_sg_lb_stats(struct lb_env *env,
6544 struct sched_group *group, int load_idx,
6545 int local_group, struct sg_lb_stats *sgs,
6551 memset(sgs, 0, sizeof(*sgs));
6553 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6554 struct rq *rq = cpu_rq(i);
6556 /* Bias balancing toward cpus of our domain */
6558 load = target_load(i, load_idx);
6560 load = source_load(i, load_idx);
6562 sgs->group_load += load;
6563 sgs->group_util += cpu_util(i);
6564 sgs->sum_nr_running += rq->cfs.h_nr_running;
6566 nr_running = rq->nr_running;
6570 #ifdef CONFIG_NUMA_BALANCING
6571 sgs->nr_numa_running += rq->nr_numa_running;
6572 sgs->nr_preferred_running += rq->nr_preferred_running;
6574 sgs->sum_weighted_load += weighted_cpuload(i);
6576 * No need to call idle_cpu() if nr_running is not 0
6578 if (!nr_running && idle_cpu(i))
6582 /* Adjust by relative CPU capacity of the group */
6583 sgs->group_capacity = group->sgc->capacity;
6584 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6586 if (sgs->sum_nr_running)
6587 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6589 sgs->group_weight = group->group_weight;
6591 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6592 sgs->group_type = group_classify(group, sgs);
6596 * update_sd_pick_busiest - return 1 on busiest group
6597 * @env: The load balancing environment.
6598 * @sds: sched_domain statistics
6599 * @sg: sched_group candidate to be checked for being the busiest
6600 * @sgs: sched_group statistics
6602 * Determine if @sg is a busier group than the previously selected
6605 * Return: %true if @sg is a busier group than the previously selected
6606 * busiest group. %false otherwise.
6608 static bool update_sd_pick_busiest(struct lb_env *env,
6609 struct sd_lb_stats *sds,
6610 struct sched_group *sg,
6611 struct sg_lb_stats *sgs)
6613 struct sg_lb_stats *busiest = &sds->busiest_stat;
6615 if (sgs->group_type > busiest->group_type)
6618 if (sgs->group_type < busiest->group_type)
6621 if (sgs->avg_load <= busiest->avg_load)
6624 /* This is the busiest node in its class. */
6625 if (!(env->sd->flags & SD_ASYM_PACKING))
6629 * ASYM_PACKING needs to move all the work to the lowest
6630 * numbered CPUs in the group, therefore mark all groups
6631 * higher than ourself as busy.
6633 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6637 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6644 #ifdef CONFIG_NUMA_BALANCING
6645 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6647 if (sgs->sum_nr_running > sgs->nr_numa_running)
6649 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6654 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6656 if (rq->nr_running > rq->nr_numa_running)
6658 if (rq->nr_running > rq->nr_preferred_running)
6663 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6668 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6672 #endif /* CONFIG_NUMA_BALANCING */
6675 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6676 * @env: The load balancing environment.
6677 * @sds: variable to hold the statistics for this sched_domain.
6679 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6681 struct sched_domain *child = env->sd->child;
6682 struct sched_group *sg = env->sd->groups;
6683 struct sg_lb_stats tmp_sgs;
6684 int load_idx, prefer_sibling = 0;
6685 bool overload = false;
6687 if (child && child->flags & SD_PREFER_SIBLING)
6690 load_idx = get_sd_load_idx(env->sd, env->idle);
6693 struct sg_lb_stats *sgs = &tmp_sgs;
6696 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6699 sgs = &sds->local_stat;
6701 if (env->idle != CPU_NEWLY_IDLE ||
6702 time_after_eq(jiffies, sg->sgc->next_update))
6703 update_group_capacity(env->sd, env->dst_cpu);
6706 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6713 * In case the child domain prefers tasks go to siblings
6714 * first, lower the sg capacity so that we'll try
6715 * and move all the excess tasks away. We lower the capacity
6716 * of a group only if the local group has the capacity to fit
6717 * these excess tasks. The extra check prevents the case where
6718 * you always pull from the heaviest group when it is already
6719 * under-utilized (possible with a large weight task outweighs
6720 * the tasks on the system).
6722 if (prefer_sibling && sds->local &&
6723 group_has_capacity(env, &sds->local_stat) &&
6724 (sgs->sum_nr_running > 1)) {
6725 sgs->group_no_capacity = 1;
6726 sgs->group_type = group_classify(sg, sgs);
6729 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6731 sds->busiest_stat = *sgs;
6735 /* Now, start updating sd_lb_stats */
6736 sds->total_load += sgs->group_load;
6737 sds->total_capacity += sgs->group_capacity;
6740 } while (sg != env->sd->groups);
6742 if (env->sd->flags & SD_NUMA)
6743 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6745 if (!env->sd->parent) {
6746 /* update overload indicator if we are at root domain */
6747 if (env->dst_rq->rd->overload != overload)
6748 env->dst_rq->rd->overload = overload;
6754 * check_asym_packing - Check to see if the group is packed into the
6757 * This is primarily intended to used at the sibling level. Some
6758 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6759 * case of POWER7, it can move to lower SMT modes only when higher
6760 * threads are idle. When in lower SMT modes, the threads will
6761 * perform better since they share less core resources. Hence when we
6762 * have idle threads, we want them to be the higher ones.
6764 * This packing function is run on idle threads. It checks to see if
6765 * the busiest CPU in this domain (core in the P7 case) has a higher
6766 * CPU number than the packing function is being run on. Here we are
6767 * assuming lower CPU number will be equivalent to lower a SMT thread
6770 * Return: 1 when packing is required and a task should be moved to
6771 * this CPU. The amount of the imbalance is returned in *imbalance.
6773 * @env: The load balancing environment.
6774 * @sds: Statistics of the sched_domain which is to be packed
6776 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6780 if (!(env->sd->flags & SD_ASYM_PACKING))
6786 busiest_cpu = group_first_cpu(sds->busiest);
6787 if (env->dst_cpu > busiest_cpu)
6790 env->imbalance = DIV_ROUND_CLOSEST(
6791 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6792 SCHED_CAPACITY_SCALE);
6798 * fix_small_imbalance - Calculate the minor imbalance that exists
6799 * amongst the groups of a sched_domain, during
6801 * @env: The load balancing environment.
6802 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6805 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6807 unsigned long tmp, capa_now = 0, capa_move = 0;
6808 unsigned int imbn = 2;
6809 unsigned long scaled_busy_load_per_task;
6810 struct sg_lb_stats *local, *busiest;
6812 local = &sds->local_stat;
6813 busiest = &sds->busiest_stat;
6815 if (!local->sum_nr_running)
6816 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6817 else if (busiest->load_per_task > local->load_per_task)
6820 scaled_busy_load_per_task =
6821 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6822 busiest->group_capacity;
6824 if (busiest->avg_load + scaled_busy_load_per_task >=
6825 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6826 env->imbalance = busiest->load_per_task;
6831 * OK, we don't have enough imbalance to justify moving tasks,
6832 * however we may be able to increase total CPU capacity used by
6836 capa_now += busiest->group_capacity *
6837 min(busiest->load_per_task, busiest->avg_load);
6838 capa_now += local->group_capacity *
6839 min(local->load_per_task, local->avg_load);
6840 capa_now /= SCHED_CAPACITY_SCALE;
6842 /* Amount of load we'd subtract */
6843 if (busiest->avg_load > scaled_busy_load_per_task) {
6844 capa_move += busiest->group_capacity *
6845 min(busiest->load_per_task,
6846 busiest->avg_load - scaled_busy_load_per_task);
6849 /* Amount of load we'd add */
6850 if (busiest->avg_load * busiest->group_capacity <
6851 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6852 tmp = (busiest->avg_load * busiest->group_capacity) /
6853 local->group_capacity;
6855 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6856 local->group_capacity;
6858 capa_move += local->group_capacity *
6859 min(local->load_per_task, local->avg_load + tmp);
6860 capa_move /= SCHED_CAPACITY_SCALE;
6862 /* Move if we gain throughput */
6863 if (capa_move > capa_now)
6864 env->imbalance = busiest->load_per_task;
6868 * calculate_imbalance - Calculate the amount of imbalance present within the
6869 * groups of a given sched_domain during load balance.
6870 * @env: load balance environment
6871 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6873 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6875 unsigned long max_pull, load_above_capacity = ~0UL;
6876 struct sg_lb_stats *local, *busiest;
6878 local = &sds->local_stat;
6879 busiest = &sds->busiest_stat;
6881 if (busiest->group_type == group_imbalanced) {
6883 * In the group_imb case we cannot rely on group-wide averages
6884 * to ensure cpu-load equilibrium, look at wider averages. XXX
6886 busiest->load_per_task =
6887 min(busiest->load_per_task, sds->avg_load);
6891 * In the presence of smp nice balancing, certain scenarios can have
6892 * max load less than avg load(as we skip the groups at or below
6893 * its cpu_capacity, while calculating max_load..)
6895 if (busiest->avg_load <= sds->avg_load ||
6896 local->avg_load >= sds->avg_load) {
6898 return fix_small_imbalance(env, sds);
6902 * If there aren't any idle cpus, avoid creating some.
6904 if (busiest->group_type == group_overloaded &&
6905 local->group_type == group_overloaded) {
6906 load_above_capacity = busiest->sum_nr_running *
6908 if (load_above_capacity > busiest->group_capacity)
6909 load_above_capacity -= busiest->group_capacity;
6911 load_above_capacity = ~0UL;
6915 * We're trying to get all the cpus to the average_load, so we don't
6916 * want to push ourselves above the average load, nor do we wish to
6917 * reduce the max loaded cpu below the average load. At the same time,
6918 * we also don't want to reduce the group load below the group capacity
6919 * (so that we can implement power-savings policies etc). Thus we look
6920 * for the minimum possible imbalance.
6922 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6924 /* How much load to actually move to equalise the imbalance */
6925 env->imbalance = min(
6926 max_pull * busiest->group_capacity,
6927 (sds->avg_load - local->avg_load) * local->group_capacity
6928 ) / SCHED_CAPACITY_SCALE;
6931 * if *imbalance is less than the average load per runnable task
6932 * there is no guarantee that any tasks will be moved so we'll have
6933 * a think about bumping its value to force at least one task to be
6936 if (env->imbalance < busiest->load_per_task)
6937 return fix_small_imbalance(env, sds);
6940 /******* find_busiest_group() helpers end here *********************/
6943 * find_busiest_group - Returns the busiest group within the sched_domain
6944 * if there is an imbalance. If there isn't an imbalance, and
6945 * the user has opted for power-savings, it returns a group whose
6946 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6947 * such a group exists.
6949 * Also calculates the amount of weighted load which should be moved
6950 * to restore balance.
6952 * @env: The load balancing environment.
6954 * Return: - The busiest group if imbalance exists.
6955 * - If no imbalance and user has opted for power-savings balance,
6956 * return the least loaded group whose CPUs can be
6957 * put to idle by rebalancing its tasks onto our group.
6959 static struct sched_group *find_busiest_group(struct lb_env *env)
6961 struct sg_lb_stats *local, *busiest;
6962 struct sd_lb_stats sds;
6964 init_sd_lb_stats(&sds);
6967 * Compute the various statistics relavent for load balancing at
6970 update_sd_lb_stats(env, &sds);
6971 local = &sds.local_stat;
6972 busiest = &sds.busiest_stat;
6974 /* ASYM feature bypasses nice load balance check */
6975 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6976 check_asym_packing(env, &sds))
6979 /* There is no busy sibling group to pull tasks from */
6980 if (!sds.busiest || busiest->sum_nr_running == 0)
6983 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6984 / sds.total_capacity;
6987 * If the busiest group is imbalanced the below checks don't
6988 * work because they assume all things are equal, which typically
6989 * isn't true due to cpus_allowed constraints and the like.
6991 if (busiest->group_type == group_imbalanced)
6994 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6995 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6996 busiest->group_no_capacity)
7000 * If the local group is busier than the selected busiest group
7001 * don't try and pull any tasks.
7003 if (local->avg_load >= busiest->avg_load)
7007 * Don't pull any tasks if this group is already above the domain
7010 if (local->avg_load >= sds.avg_load)
7013 if (env->idle == CPU_IDLE) {
7015 * This cpu is idle. If the busiest group is not overloaded
7016 * and there is no imbalance between this and busiest group
7017 * wrt idle cpus, it is balanced. The imbalance becomes
7018 * significant if the diff is greater than 1 otherwise we
7019 * might end up to just move the imbalance on another group
7021 if ((busiest->group_type != group_overloaded) &&
7022 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7026 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7027 * imbalance_pct to be conservative.
7029 if (100 * busiest->avg_load <=
7030 env->sd->imbalance_pct * local->avg_load)
7035 /* Looks like there is an imbalance. Compute it */
7036 calculate_imbalance(env, &sds);
7045 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7047 static struct rq *find_busiest_queue(struct lb_env *env,
7048 struct sched_group *group)
7050 struct rq *busiest = NULL, *rq;
7051 unsigned long busiest_load = 0, busiest_capacity = 1;
7054 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7055 unsigned long capacity, wl;
7059 rt = fbq_classify_rq(rq);
7062 * We classify groups/runqueues into three groups:
7063 * - regular: there are !numa tasks
7064 * - remote: there are numa tasks that run on the 'wrong' node
7065 * - all: there is no distinction
7067 * In order to avoid migrating ideally placed numa tasks,
7068 * ignore those when there's better options.
7070 * If we ignore the actual busiest queue to migrate another
7071 * task, the next balance pass can still reduce the busiest
7072 * queue by moving tasks around inside the node.
7074 * If we cannot move enough load due to this classification
7075 * the next pass will adjust the group classification and
7076 * allow migration of more tasks.
7078 * Both cases only affect the total convergence complexity.
7080 if (rt > env->fbq_type)
7083 capacity = capacity_of(i);
7085 wl = weighted_cpuload(i);
7088 * When comparing with imbalance, use weighted_cpuload()
7089 * which is not scaled with the cpu capacity.
7092 if (rq->nr_running == 1 && wl > env->imbalance &&
7093 !check_cpu_capacity(rq, env->sd))
7097 * For the load comparisons with the other cpu's, consider
7098 * the weighted_cpuload() scaled with the cpu capacity, so
7099 * that the load can be moved away from the cpu that is
7100 * potentially running at a lower capacity.
7102 * Thus we're looking for max(wl_i / capacity_i), crosswise
7103 * multiplication to rid ourselves of the division works out
7104 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7105 * our previous maximum.
7107 if (wl * busiest_capacity > busiest_load * capacity) {
7109 busiest_capacity = capacity;
7118 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7119 * so long as it is large enough.
7121 #define MAX_PINNED_INTERVAL 512
7123 /* Working cpumask for load_balance and load_balance_newidle. */
7124 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7126 static int need_active_balance(struct lb_env *env)
7128 struct sched_domain *sd = env->sd;
7130 if (env->idle == CPU_NEWLY_IDLE) {
7133 * ASYM_PACKING needs to force migrate tasks from busy but
7134 * higher numbered CPUs in order to pack all tasks in the
7135 * lowest numbered CPUs.
7137 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7142 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7143 * It's worth migrating the task if the src_cpu's capacity is reduced
7144 * because of other sched_class or IRQs if more capacity stays
7145 * available on dst_cpu.
7147 if ((env->idle != CPU_NOT_IDLE) &&
7148 (env->src_rq->cfs.h_nr_running == 1)) {
7149 if ((check_cpu_capacity(env->src_rq, sd)) &&
7150 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7154 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7157 static int active_load_balance_cpu_stop(void *data);
7159 static int should_we_balance(struct lb_env *env)
7161 struct sched_group *sg = env->sd->groups;
7162 struct cpumask *sg_cpus, *sg_mask;
7163 int cpu, balance_cpu = -1;
7166 * In the newly idle case, we will allow all the cpu's
7167 * to do the newly idle load balance.
7169 if (env->idle == CPU_NEWLY_IDLE)
7172 sg_cpus = sched_group_cpus(sg);
7173 sg_mask = sched_group_mask(sg);
7174 /* Try to find first idle cpu */
7175 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7176 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7183 if (balance_cpu == -1)
7184 balance_cpu = group_balance_cpu(sg);
7187 * First idle cpu or the first cpu(busiest) in this sched group
7188 * is eligible for doing load balancing at this and above domains.
7190 return balance_cpu == env->dst_cpu;
7194 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7195 * tasks if there is an imbalance.
7197 static int load_balance(int this_cpu, struct rq *this_rq,
7198 struct sched_domain *sd, enum cpu_idle_type idle,
7199 int *continue_balancing)
7201 int ld_moved, cur_ld_moved, active_balance = 0;
7202 struct sched_domain *sd_parent = sd->parent;
7203 struct sched_group *group;
7205 unsigned long flags;
7206 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7208 struct lb_env env = {
7210 .dst_cpu = this_cpu,
7212 .dst_grpmask = sched_group_cpus(sd->groups),
7214 .loop_break = sched_nr_migrate_break,
7217 .tasks = LIST_HEAD_INIT(env.tasks),
7221 * For NEWLY_IDLE load_balancing, we don't need to consider
7222 * other cpus in our group
7224 if (idle == CPU_NEWLY_IDLE)
7225 env.dst_grpmask = NULL;
7227 cpumask_copy(cpus, cpu_active_mask);
7229 schedstat_inc(sd, lb_count[idle]);
7232 if (!should_we_balance(&env)) {
7233 *continue_balancing = 0;
7237 group = find_busiest_group(&env);
7239 schedstat_inc(sd, lb_nobusyg[idle]);
7243 busiest = find_busiest_queue(&env, group);
7245 schedstat_inc(sd, lb_nobusyq[idle]);
7249 BUG_ON(busiest == env.dst_rq);
7251 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7253 env.src_cpu = busiest->cpu;
7254 env.src_rq = busiest;
7257 if (busiest->nr_running > 1) {
7259 * Attempt to move tasks. If find_busiest_group has found
7260 * an imbalance but busiest->nr_running <= 1, the group is
7261 * still unbalanced. ld_moved simply stays zero, so it is
7262 * correctly treated as an imbalance.
7264 env.flags |= LBF_ALL_PINNED;
7265 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7268 raw_spin_lock_irqsave(&busiest->lock, flags);
7271 * cur_ld_moved - load moved in current iteration
7272 * ld_moved - cumulative load moved across iterations
7274 cur_ld_moved = detach_tasks(&env);
7277 * We've detached some tasks from busiest_rq. Every
7278 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7279 * unlock busiest->lock, and we are able to be sure
7280 * that nobody can manipulate the tasks in parallel.
7281 * See task_rq_lock() family for the details.
7284 raw_spin_unlock(&busiest->lock);
7288 ld_moved += cur_ld_moved;
7291 local_irq_restore(flags);
7293 if (env.flags & LBF_NEED_BREAK) {
7294 env.flags &= ~LBF_NEED_BREAK;
7299 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7300 * us and move them to an alternate dst_cpu in our sched_group
7301 * where they can run. The upper limit on how many times we
7302 * iterate on same src_cpu is dependent on number of cpus in our
7305 * This changes load balance semantics a bit on who can move
7306 * load to a given_cpu. In addition to the given_cpu itself
7307 * (or a ilb_cpu acting on its behalf where given_cpu is
7308 * nohz-idle), we now have balance_cpu in a position to move
7309 * load to given_cpu. In rare situations, this may cause
7310 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7311 * _independently_ and at _same_ time to move some load to
7312 * given_cpu) causing exceess load to be moved to given_cpu.
7313 * This however should not happen so much in practice and
7314 * moreover subsequent load balance cycles should correct the
7315 * excess load moved.
7317 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7319 /* Prevent to re-select dst_cpu via env's cpus */
7320 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7322 env.dst_rq = cpu_rq(env.new_dst_cpu);
7323 env.dst_cpu = env.new_dst_cpu;
7324 env.flags &= ~LBF_DST_PINNED;
7326 env.loop_break = sched_nr_migrate_break;
7329 * Go back to "more_balance" rather than "redo" since we
7330 * need to continue with same src_cpu.
7336 * We failed to reach balance because of affinity.
7339 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7341 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7342 *group_imbalance = 1;
7345 /* All tasks on this runqueue were pinned by CPU affinity */
7346 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7347 cpumask_clear_cpu(cpu_of(busiest), cpus);
7348 if (!cpumask_empty(cpus)) {
7350 env.loop_break = sched_nr_migrate_break;
7353 goto out_all_pinned;
7358 schedstat_inc(sd, lb_failed[idle]);
7360 * Increment the failure counter only on periodic balance.
7361 * We do not want newidle balance, which can be very
7362 * frequent, pollute the failure counter causing
7363 * excessive cache_hot migrations and active balances.
7365 if (idle != CPU_NEWLY_IDLE)
7366 sd->nr_balance_failed++;
7368 if (need_active_balance(&env)) {
7369 raw_spin_lock_irqsave(&busiest->lock, flags);
7371 /* don't kick the active_load_balance_cpu_stop,
7372 * if the curr task on busiest cpu can't be
7375 if (!cpumask_test_cpu(this_cpu,
7376 tsk_cpus_allowed(busiest->curr))) {
7377 raw_spin_unlock_irqrestore(&busiest->lock,
7379 env.flags |= LBF_ALL_PINNED;
7380 goto out_one_pinned;
7384 * ->active_balance synchronizes accesses to
7385 * ->active_balance_work. Once set, it's cleared
7386 * only after active load balance is finished.
7388 if (!busiest->active_balance) {
7389 busiest->active_balance = 1;
7390 busiest->push_cpu = this_cpu;
7393 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7395 if (active_balance) {
7396 stop_one_cpu_nowait(cpu_of(busiest),
7397 active_load_balance_cpu_stop, busiest,
7398 &busiest->active_balance_work);
7402 * We've kicked active balancing, reset the failure
7405 sd->nr_balance_failed = sd->cache_nice_tries+1;
7408 sd->nr_balance_failed = 0;
7410 if (likely(!active_balance)) {
7411 /* We were unbalanced, so reset the balancing interval */
7412 sd->balance_interval = sd->min_interval;
7415 * If we've begun active balancing, start to back off. This
7416 * case may not be covered by the all_pinned logic if there
7417 * is only 1 task on the busy runqueue (because we don't call
7420 if (sd->balance_interval < sd->max_interval)
7421 sd->balance_interval *= 2;
7428 * We reach balance although we may have faced some affinity
7429 * constraints. Clear the imbalance flag if it was set.
7432 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7434 if (*group_imbalance)
7435 *group_imbalance = 0;
7440 * We reach balance because all tasks are pinned at this level so
7441 * we can't migrate them. Let the imbalance flag set so parent level
7442 * can try to migrate them.
7444 schedstat_inc(sd, lb_balanced[idle]);
7446 sd->nr_balance_failed = 0;
7449 /* tune up the balancing interval */
7450 if (((env.flags & LBF_ALL_PINNED) &&
7451 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7452 (sd->balance_interval < sd->max_interval))
7453 sd->balance_interval *= 2;
7460 static inline unsigned long
7461 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7463 unsigned long interval = sd->balance_interval;
7466 interval *= sd->busy_factor;
7468 /* scale ms to jiffies */
7469 interval = msecs_to_jiffies(interval);
7470 interval = clamp(interval, 1UL, max_load_balance_interval);
7476 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7478 unsigned long interval, next;
7480 interval = get_sd_balance_interval(sd, cpu_busy);
7481 next = sd->last_balance + interval;
7483 if (time_after(*next_balance, next))
7484 *next_balance = next;
7488 * idle_balance is called by schedule() if this_cpu is about to become
7489 * idle. Attempts to pull tasks from other CPUs.
7491 static int idle_balance(struct rq *this_rq)
7493 unsigned long next_balance = jiffies + HZ;
7494 int this_cpu = this_rq->cpu;
7495 struct sched_domain *sd;
7496 int pulled_task = 0;
7500 * We must set idle_stamp _before_ calling idle_balance(), such that we
7501 * measure the duration of idle_balance() as idle time.
7503 this_rq->idle_stamp = rq_clock(this_rq);
7505 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7506 !this_rq->rd->overload) {
7508 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7510 update_next_balance(sd, 0, &next_balance);
7516 raw_spin_unlock(&this_rq->lock);
7518 update_blocked_averages(this_cpu);
7520 for_each_domain(this_cpu, sd) {
7521 int continue_balancing = 1;
7522 u64 t0, domain_cost;
7524 if (!(sd->flags & SD_LOAD_BALANCE))
7527 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7528 update_next_balance(sd, 0, &next_balance);
7532 if (sd->flags & SD_BALANCE_NEWIDLE) {
7533 t0 = sched_clock_cpu(this_cpu);
7535 pulled_task = load_balance(this_cpu, this_rq,
7537 &continue_balancing);
7539 domain_cost = sched_clock_cpu(this_cpu) - t0;
7540 if (domain_cost > sd->max_newidle_lb_cost)
7541 sd->max_newidle_lb_cost = domain_cost;
7543 curr_cost += domain_cost;
7546 update_next_balance(sd, 0, &next_balance);
7549 * Stop searching for tasks to pull if there are
7550 * now runnable tasks on this rq.
7552 if (pulled_task || this_rq->nr_running > 0)
7557 raw_spin_lock(&this_rq->lock);
7559 if (curr_cost > this_rq->max_idle_balance_cost)
7560 this_rq->max_idle_balance_cost = curr_cost;
7563 * While browsing the domains, we released the rq lock, a task could
7564 * have been enqueued in the meantime. Since we're not going idle,
7565 * pretend we pulled a task.
7567 if (this_rq->cfs.h_nr_running && !pulled_task)
7571 /* Move the next balance forward */
7572 if (time_after(this_rq->next_balance, next_balance))
7573 this_rq->next_balance = next_balance;
7575 /* Is there a task of a high priority class? */
7576 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7580 this_rq->idle_stamp = 0;
7586 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7587 * running tasks off the busiest CPU onto idle CPUs. It requires at
7588 * least 1 task to be running on each physical CPU where possible, and
7589 * avoids physical / logical imbalances.
7591 static int active_load_balance_cpu_stop(void *data)
7593 struct rq *busiest_rq = data;
7594 int busiest_cpu = cpu_of(busiest_rq);
7595 int target_cpu = busiest_rq->push_cpu;
7596 struct rq *target_rq = cpu_rq(target_cpu);
7597 struct sched_domain *sd;
7598 struct task_struct *p = NULL;
7600 raw_spin_lock_irq(&busiest_rq->lock);
7602 /* make sure the requested cpu hasn't gone down in the meantime */
7603 if (unlikely(busiest_cpu != smp_processor_id() ||
7604 !busiest_rq->active_balance))
7607 /* Is there any task to move? */
7608 if (busiest_rq->nr_running <= 1)
7612 * This condition is "impossible", if it occurs
7613 * we need to fix it. Originally reported by
7614 * Bjorn Helgaas on a 128-cpu setup.
7616 BUG_ON(busiest_rq == target_rq);
7618 /* Search for an sd spanning us and the target CPU. */
7620 for_each_domain(target_cpu, sd) {
7621 if ((sd->flags & SD_LOAD_BALANCE) &&
7622 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7627 struct lb_env env = {
7629 .dst_cpu = target_cpu,
7630 .dst_rq = target_rq,
7631 .src_cpu = busiest_rq->cpu,
7632 .src_rq = busiest_rq,
7636 schedstat_inc(sd, alb_count);
7638 p = detach_one_task(&env);
7640 schedstat_inc(sd, alb_pushed);
7642 schedstat_inc(sd, alb_failed);
7646 busiest_rq->active_balance = 0;
7647 raw_spin_unlock(&busiest_rq->lock);
7650 attach_one_task(target_rq, p);
7657 static inline int on_null_domain(struct rq *rq)
7659 return unlikely(!rcu_dereference_sched(rq->sd));
7662 #ifdef CONFIG_NO_HZ_COMMON
7664 * idle load balancing details
7665 * - When one of the busy CPUs notice that there may be an idle rebalancing
7666 * needed, they will kick the idle load balancer, which then does idle
7667 * load balancing for all the idle CPUs.
7670 cpumask_var_t idle_cpus_mask;
7672 unsigned long next_balance; /* in jiffy units */
7673 } nohz ____cacheline_aligned;
7675 static inline int find_new_ilb(void)
7677 int ilb = cpumask_first(nohz.idle_cpus_mask);
7679 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7686 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7687 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7688 * CPU (if there is one).
7690 static void nohz_balancer_kick(void)
7694 nohz.next_balance++;
7696 ilb_cpu = find_new_ilb();
7698 if (ilb_cpu >= nr_cpu_ids)
7701 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7704 * Use smp_send_reschedule() instead of resched_cpu().
7705 * This way we generate a sched IPI on the target cpu which
7706 * is idle. And the softirq performing nohz idle load balance
7707 * will be run before returning from the IPI.
7709 smp_send_reschedule(ilb_cpu);
7713 static inline void nohz_balance_exit_idle(int cpu)
7715 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7717 * Completely isolated CPUs don't ever set, so we must test.
7719 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7720 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7721 atomic_dec(&nohz.nr_cpus);
7723 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7727 static inline void set_cpu_sd_state_busy(void)
7729 struct sched_domain *sd;
7730 int cpu = smp_processor_id();
7733 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7735 if (!sd || !sd->nohz_idle)
7739 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7744 void set_cpu_sd_state_idle(void)
7746 struct sched_domain *sd;
7747 int cpu = smp_processor_id();
7750 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7752 if (!sd || sd->nohz_idle)
7756 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7762 * This routine will record that the cpu is going idle with tick stopped.
7763 * This info will be used in performing idle load balancing in the future.
7765 void nohz_balance_enter_idle(int cpu)
7768 * If this cpu is going down, then nothing needs to be done.
7770 if (!cpu_active(cpu))
7773 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7777 * If we're a completely isolated CPU, we don't play.
7779 if (on_null_domain(cpu_rq(cpu)))
7782 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7783 atomic_inc(&nohz.nr_cpus);
7784 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7787 static int sched_ilb_notifier(struct notifier_block *nfb,
7788 unsigned long action, void *hcpu)
7790 switch (action & ~CPU_TASKS_FROZEN) {
7792 nohz_balance_exit_idle(smp_processor_id());
7800 static DEFINE_SPINLOCK(balancing);
7803 * Scale the max load_balance interval with the number of CPUs in the system.
7804 * This trades load-balance latency on larger machines for less cross talk.
7806 void update_max_interval(void)
7808 max_load_balance_interval = HZ*num_online_cpus()/10;
7812 * It checks each scheduling domain to see if it is due to be balanced,
7813 * and initiates a balancing operation if so.
7815 * Balancing parameters are set up in init_sched_domains.
7817 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7819 int continue_balancing = 1;
7821 unsigned long interval;
7822 struct sched_domain *sd;
7823 /* Earliest time when we have to do rebalance again */
7824 unsigned long next_balance = jiffies + 60*HZ;
7825 int update_next_balance = 0;
7826 int need_serialize, need_decay = 0;
7829 update_blocked_averages(cpu);
7832 for_each_domain(cpu, sd) {
7834 * Decay the newidle max times here because this is a regular
7835 * visit to all the domains. Decay ~1% per second.
7837 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7838 sd->max_newidle_lb_cost =
7839 (sd->max_newidle_lb_cost * 253) / 256;
7840 sd->next_decay_max_lb_cost = jiffies + HZ;
7843 max_cost += sd->max_newidle_lb_cost;
7845 if (!(sd->flags & SD_LOAD_BALANCE))
7849 * Stop the load balance at this level. There is another
7850 * CPU in our sched group which is doing load balancing more
7853 if (!continue_balancing) {
7859 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7861 need_serialize = sd->flags & SD_SERIALIZE;
7862 if (need_serialize) {
7863 if (!spin_trylock(&balancing))
7867 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7868 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7870 * The LBF_DST_PINNED logic could have changed
7871 * env->dst_cpu, so we can't know our idle
7872 * state even if we migrated tasks. Update it.
7874 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7876 sd->last_balance = jiffies;
7877 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7880 spin_unlock(&balancing);
7882 if (time_after(next_balance, sd->last_balance + interval)) {
7883 next_balance = sd->last_balance + interval;
7884 update_next_balance = 1;
7889 * Ensure the rq-wide value also decays but keep it at a
7890 * reasonable floor to avoid funnies with rq->avg_idle.
7892 rq->max_idle_balance_cost =
7893 max((u64)sysctl_sched_migration_cost, max_cost);
7898 * next_balance will be updated only when there is a need.
7899 * When the cpu is attached to null domain for ex, it will not be
7902 if (likely(update_next_balance)) {
7903 rq->next_balance = next_balance;
7905 #ifdef CONFIG_NO_HZ_COMMON
7907 * If this CPU has been elected to perform the nohz idle
7908 * balance. Other idle CPUs have already rebalanced with
7909 * nohz_idle_balance() and nohz.next_balance has been
7910 * updated accordingly. This CPU is now running the idle load
7911 * balance for itself and we need to update the
7912 * nohz.next_balance accordingly.
7914 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
7915 nohz.next_balance = rq->next_balance;
7920 #ifdef CONFIG_NO_HZ_COMMON
7922 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7923 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7925 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7927 int this_cpu = this_rq->cpu;
7930 /* Earliest time when we have to do rebalance again */
7931 unsigned long next_balance = jiffies + 60*HZ;
7932 int update_next_balance = 0;
7934 if (idle != CPU_IDLE ||
7935 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7938 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7939 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7943 * If this cpu gets work to do, stop the load balancing
7944 * work being done for other cpus. Next load
7945 * balancing owner will pick it up.
7950 rq = cpu_rq(balance_cpu);
7953 * If time for next balance is due,
7956 if (time_after_eq(jiffies, rq->next_balance)) {
7957 raw_spin_lock_irq(&rq->lock);
7958 update_rq_clock(rq);
7959 update_cpu_load_idle(rq);
7960 raw_spin_unlock_irq(&rq->lock);
7961 rebalance_domains(rq, CPU_IDLE);
7964 if (time_after(next_balance, rq->next_balance)) {
7965 next_balance = rq->next_balance;
7966 update_next_balance = 1;
7971 * next_balance will be updated only when there is a need.
7972 * When the CPU is attached to null domain for ex, it will not be
7975 if (likely(update_next_balance))
7976 nohz.next_balance = next_balance;
7978 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7982 * Current heuristic for kicking the idle load balancer in the presence
7983 * of an idle cpu in the system.
7984 * - This rq has more than one task.
7985 * - This rq has at least one CFS task and the capacity of the CPU is
7986 * significantly reduced because of RT tasks or IRQs.
7987 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7988 * multiple busy cpu.
7989 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7990 * domain span are idle.
7992 static inline bool nohz_kick_needed(struct rq *rq)
7994 unsigned long now = jiffies;
7995 struct sched_domain *sd;
7996 struct sched_group_capacity *sgc;
7997 int nr_busy, cpu = rq->cpu;
8000 if (unlikely(rq->idle_balance))
8004 * We may be recently in ticked or tickless idle mode. At the first
8005 * busy tick after returning from idle, we will update the busy stats.
8007 set_cpu_sd_state_busy();
8008 nohz_balance_exit_idle(cpu);
8011 * None are in tickless mode and hence no need for NOHZ idle load
8014 if (likely(!atomic_read(&nohz.nr_cpus)))
8017 if (time_before(now, nohz.next_balance))
8020 if (rq->nr_running >= 2)
8024 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8026 sgc = sd->groups->sgc;
8027 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8036 sd = rcu_dereference(rq->sd);
8038 if ((rq->cfs.h_nr_running >= 1) &&
8039 check_cpu_capacity(rq, sd)) {
8045 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8046 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8047 sched_domain_span(sd)) < cpu)) {
8057 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8061 * run_rebalance_domains is triggered when needed from the scheduler tick.
8062 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8064 static void run_rebalance_domains(struct softirq_action *h)
8066 struct rq *this_rq = this_rq();
8067 enum cpu_idle_type idle = this_rq->idle_balance ?
8068 CPU_IDLE : CPU_NOT_IDLE;
8071 * If this cpu has a pending nohz_balance_kick, then do the
8072 * balancing on behalf of the other idle cpus whose ticks are
8073 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8074 * give the idle cpus a chance to load balance. Else we may
8075 * load balance only within the local sched_domain hierarchy
8076 * and abort nohz_idle_balance altogether if we pull some load.
8078 nohz_idle_balance(this_rq, idle);
8079 rebalance_domains(this_rq, idle);
8083 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8085 void trigger_load_balance(struct rq *rq)
8087 /* Don't need to rebalance while attached to NULL domain */
8088 if (unlikely(on_null_domain(rq)))
8091 if (time_after_eq(jiffies, rq->next_balance))
8092 raise_softirq(SCHED_SOFTIRQ);
8093 #ifdef CONFIG_NO_HZ_COMMON
8094 if (nohz_kick_needed(rq))
8095 nohz_balancer_kick();
8099 static void rq_online_fair(struct rq *rq)
8103 update_runtime_enabled(rq);
8106 static void rq_offline_fair(struct rq *rq)
8110 /* Ensure any throttled groups are reachable by pick_next_task */
8111 unthrottle_offline_cfs_rqs(rq);
8114 #endif /* CONFIG_SMP */
8117 * scheduler tick hitting a task of our scheduling class:
8119 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8121 struct cfs_rq *cfs_rq;
8122 struct sched_entity *se = &curr->se;
8124 for_each_sched_entity(se) {
8125 cfs_rq = cfs_rq_of(se);
8126 entity_tick(cfs_rq, se, queued);
8129 if (static_branch_unlikely(&sched_numa_balancing))
8130 task_tick_numa(rq, curr);
8134 * called on fork with the child task as argument from the parent's context
8135 * - child not yet on the tasklist
8136 * - preemption disabled
8138 static void task_fork_fair(struct task_struct *p)
8140 struct cfs_rq *cfs_rq;
8141 struct sched_entity *se = &p->se, *curr;
8142 int this_cpu = smp_processor_id();
8143 struct rq *rq = this_rq();
8144 unsigned long flags;
8146 raw_spin_lock_irqsave(&rq->lock, flags);
8148 update_rq_clock(rq);
8150 cfs_rq = task_cfs_rq(current);
8151 curr = cfs_rq->curr;
8154 * Not only the cpu but also the task_group of the parent might have
8155 * been changed after parent->se.parent,cfs_rq were copied to
8156 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8157 * of child point to valid ones.
8160 __set_task_cpu(p, this_cpu);
8163 update_curr(cfs_rq);
8166 se->vruntime = curr->vruntime;
8167 place_entity(cfs_rq, se, 1);
8169 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8171 * Upon rescheduling, sched_class::put_prev_task() will place
8172 * 'current' within the tree based on its new key value.
8174 swap(curr->vruntime, se->vruntime);
8178 se->vruntime -= cfs_rq->min_vruntime;
8180 raw_spin_unlock_irqrestore(&rq->lock, flags);
8184 * Priority of the task has changed. Check to see if we preempt
8188 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8190 if (!task_on_rq_queued(p))
8194 * Reschedule if we are currently running on this runqueue and
8195 * our priority decreased, or if we are not currently running on
8196 * this runqueue and our priority is higher than the current's
8198 if (rq->curr == p) {
8199 if (p->prio > oldprio)
8202 check_preempt_curr(rq, p, 0);
8205 static inline bool vruntime_normalized(struct task_struct *p)
8207 struct sched_entity *se = &p->se;
8210 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8211 * the dequeue_entity(.flags=0) will already have normalized the
8218 * When !on_rq, vruntime of the task has usually NOT been normalized.
8219 * But there are some cases where it has already been normalized:
8221 * - A forked child which is waiting for being woken up by
8222 * wake_up_new_task().
8223 * - A task which has been woken up by try_to_wake_up() and
8224 * waiting for actually being woken up by sched_ttwu_pending().
8226 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8232 static void detach_task_cfs_rq(struct task_struct *p)
8234 struct sched_entity *se = &p->se;
8235 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8237 if (!vruntime_normalized(p)) {
8239 * Fix up our vruntime so that the current sleep doesn't
8240 * cause 'unlimited' sleep bonus.
8242 place_entity(cfs_rq, se, 0);
8243 se->vruntime -= cfs_rq->min_vruntime;
8246 /* Catch up with the cfs_rq and remove our load when we leave */
8247 detach_entity_load_avg(cfs_rq, se);
8250 static void attach_task_cfs_rq(struct task_struct *p)
8252 struct sched_entity *se = &p->se;
8253 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8255 #ifdef CONFIG_FAIR_GROUP_SCHED
8257 * Since the real-depth could have been changed (only FAIR
8258 * class maintain depth value), reset depth properly.
8260 se->depth = se->parent ? se->parent->depth + 1 : 0;
8263 /* Synchronize task with its cfs_rq */
8264 attach_entity_load_avg(cfs_rq, se);
8266 if (!vruntime_normalized(p))
8267 se->vruntime += cfs_rq->min_vruntime;
8270 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8272 detach_task_cfs_rq(p);
8275 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8277 attach_task_cfs_rq(p);
8279 if (task_on_rq_queued(p)) {
8281 * We were most likely switched from sched_rt, so
8282 * kick off the schedule if running, otherwise just see
8283 * if we can still preempt the current task.
8288 check_preempt_curr(rq, p, 0);
8292 /* Account for a task changing its policy or group.
8294 * This routine is mostly called to set cfs_rq->curr field when a task
8295 * migrates between groups/classes.
8297 static void set_curr_task_fair(struct rq *rq)
8299 struct sched_entity *se = &rq->curr->se;
8301 for_each_sched_entity(se) {
8302 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8304 set_next_entity(cfs_rq, se);
8305 /* ensure bandwidth has been allocated on our new cfs_rq */
8306 account_cfs_rq_runtime(cfs_rq, 0);
8310 void init_cfs_rq(struct cfs_rq *cfs_rq)
8312 cfs_rq->tasks_timeline = RB_ROOT;
8313 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8314 #ifndef CONFIG_64BIT
8315 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8318 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8319 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8323 #ifdef CONFIG_FAIR_GROUP_SCHED
8324 static void task_move_group_fair(struct task_struct *p)
8326 detach_task_cfs_rq(p);
8327 set_task_rq(p, task_cpu(p));
8330 /* Tell se's cfs_rq has been changed -- migrated */
8331 p->se.avg.last_update_time = 0;
8333 attach_task_cfs_rq(p);
8336 void free_fair_sched_group(struct task_group *tg)
8340 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8342 for_each_possible_cpu(i) {
8344 kfree(tg->cfs_rq[i]);
8353 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8355 struct cfs_rq *cfs_rq;
8356 struct sched_entity *se;
8359 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8362 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8366 tg->shares = NICE_0_LOAD;
8368 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8370 for_each_possible_cpu(i) {
8371 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8372 GFP_KERNEL, cpu_to_node(i));
8376 se = kzalloc_node(sizeof(struct sched_entity),
8377 GFP_KERNEL, cpu_to_node(i));
8381 init_cfs_rq(cfs_rq);
8382 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8383 init_entity_runnable_average(se);
8394 void unregister_fair_sched_group(struct task_group *tg)
8396 unsigned long flags;
8400 for_each_possible_cpu(cpu) {
8402 remove_entity_load_avg(tg->se[cpu]);
8405 * Only empty task groups can be destroyed; so we can speculatively
8406 * check on_list without danger of it being re-added.
8408 if (!tg->cfs_rq[cpu]->on_list)
8413 raw_spin_lock_irqsave(&rq->lock, flags);
8414 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8415 raw_spin_unlock_irqrestore(&rq->lock, flags);
8419 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8420 struct sched_entity *se, int cpu,
8421 struct sched_entity *parent)
8423 struct rq *rq = cpu_rq(cpu);
8427 init_cfs_rq_runtime(cfs_rq);
8429 tg->cfs_rq[cpu] = cfs_rq;
8432 /* se could be NULL for root_task_group */
8437 se->cfs_rq = &rq->cfs;
8440 se->cfs_rq = parent->my_q;
8441 se->depth = parent->depth + 1;
8445 /* guarantee group entities always have weight */
8446 update_load_set(&se->load, NICE_0_LOAD);
8447 se->parent = parent;
8450 static DEFINE_MUTEX(shares_mutex);
8452 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8455 unsigned long flags;
8458 * We can't change the weight of the root cgroup.
8463 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8465 mutex_lock(&shares_mutex);
8466 if (tg->shares == shares)
8469 tg->shares = shares;
8470 for_each_possible_cpu(i) {
8471 struct rq *rq = cpu_rq(i);
8472 struct sched_entity *se;
8475 /* Propagate contribution to hierarchy */
8476 raw_spin_lock_irqsave(&rq->lock, flags);
8478 /* Possible calls to update_curr() need rq clock */
8479 update_rq_clock(rq);
8480 for_each_sched_entity(se)
8481 update_cfs_shares(group_cfs_rq(se));
8482 raw_spin_unlock_irqrestore(&rq->lock, flags);
8486 mutex_unlock(&shares_mutex);
8489 #else /* CONFIG_FAIR_GROUP_SCHED */
8491 void free_fair_sched_group(struct task_group *tg) { }
8493 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8498 void unregister_fair_sched_group(struct task_group *tg) { }
8500 #endif /* CONFIG_FAIR_GROUP_SCHED */
8503 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8505 struct sched_entity *se = &task->se;
8506 unsigned int rr_interval = 0;
8509 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8512 if (rq->cfs.load.weight)
8513 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8519 * All the scheduling class methods:
8521 const struct sched_class fair_sched_class = {
8522 .next = &idle_sched_class,
8523 .enqueue_task = enqueue_task_fair,
8524 .dequeue_task = dequeue_task_fair,
8525 .yield_task = yield_task_fair,
8526 .yield_to_task = yield_to_task_fair,
8528 .check_preempt_curr = check_preempt_wakeup,
8530 .pick_next_task = pick_next_task_fair,
8531 .put_prev_task = put_prev_task_fair,
8534 .select_task_rq = select_task_rq_fair,
8535 .migrate_task_rq = migrate_task_rq_fair,
8537 .rq_online = rq_online_fair,
8538 .rq_offline = rq_offline_fair,
8540 .task_waking = task_waking_fair,
8541 .task_dead = task_dead_fair,
8542 .set_cpus_allowed = set_cpus_allowed_common,
8545 .set_curr_task = set_curr_task_fair,
8546 .task_tick = task_tick_fair,
8547 .task_fork = task_fork_fair,
8549 .prio_changed = prio_changed_fair,
8550 .switched_from = switched_from_fair,
8551 .switched_to = switched_to_fair,
8553 .get_rr_interval = get_rr_interval_fair,
8555 .update_curr = update_curr_fair,
8557 #ifdef CONFIG_FAIR_GROUP_SCHED
8558 .task_move_group = task_move_group_fair,
8562 #ifdef CONFIG_SCHED_DEBUG
8563 void print_cfs_stats(struct seq_file *m, int cpu)
8565 struct cfs_rq *cfs_rq;
8568 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8569 print_cfs_rq(m, cpu, cfs_rq);
8573 #ifdef CONFIG_NUMA_BALANCING
8574 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8577 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8579 for_each_online_node(node) {
8580 if (p->numa_faults) {
8581 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8582 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8584 if (p->numa_group) {
8585 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8586 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8588 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8591 #endif /* CONFIG_NUMA_BALANCING */
8592 #endif /* CONFIG_SCHED_DEBUG */
8594 __init void init_sched_fair_class(void)
8597 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8599 #ifdef CONFIG_NO_HZ_COMMON
8600 nohz.next_balance = jiffies;
8601 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8602 cpu_notifier(sched_ilb_notifier, 0);