2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency = 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity = 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency = 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static 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));
742 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
744 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
748 * Task is being enqueued - update stats:
750 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
753 * Are we enqueueing a waiting task? (for current tasks
754 * a dequeue/enqueue event is a NOP)
756 if (se != cfs_rq->curr)
757 update_stats_wait_start(cfs_rq, se);
761 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
763 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
764 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
765 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
766 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
767 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
768 #ifdef CONFIG_SCHEDSTATS
769 if (entity_is_task(se)) {
770 trace_sched_stat_wait(task_of(se),
771 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
774 schedstat_set(se->statistics.wait_start, 0);
778 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
781 * Mark the end of the wait period if dequeueing a
784 if (se != cfs_rq->curr)
785 update_stats_wait_end(cfs_rq, se);
789 * We are picking a new current task - update its stats:
792 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
795 * We are starting a new run period:
797 se->exec_start = rq_clock_task(rq_of(cfs_rq));
800 /**************************************************
801 * Scheduling class queueing methods:
804 #ifdef CONFIG_NUMA_BALANCING
806 * Approximate time to scan a full NUMA task in ms. The task scan period is
807 * calculated based on the tasks virtual memory size and
808 * numa_balancing_scan_size.
810 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
811 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
813 /* Portion of address space to scan in MB */
814 unsigned int sysctl_numa_balancing_scan_size = 256;
816 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
817 unsigned int sysctl_numa_balancing_scan_delay = 1000;
819 static unsigned int task_nr_scan_windows(struct task_struct *p)
821 unsigned long rss = 0;
822 unsigned long nr_scan_pages;
825 * Calculations based on RSS as non-present and empty pages are skipped
826 * by the PTE scanner and NUMA hinting faults should be trapped based
829 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
830 rss = get_mm_rss(p->mm);
834 rss = round_up(rss, nr_scan_pages);
835 return rss / nr_scan_pages;
838 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
839 #define MAX_SCAN_WINDOW 2560
841 static unsigned int task_scan_min(struct task_struct *p)
843 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
844 unsigned int scan, floor;
845 unsigned int windows = 1;
847 if (scan_size < MAX_SCAN_WINDOW)
848 windows = MAX_SCAN_WINDOW / scan_size;
849 floor = 1000 / windows;
851 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
852 return max_t(unsigned int, floor, scan);
855 static unsigned int task_scan_max(struct task_struct *p)
857 unsigned int smin = task_scan_min(p);
860 /* Watch for min being lower than max due to floor calculations */
861 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
862 return max(smin, smax);
865 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
867 rq->nr_numa_running += (p->numa_preferred_nid != -1);
868 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
871 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
873 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
874 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
880 spinlock_t lock; /* nr_tasks, tasks */
885 nodemask_t active_nodes;
886 unsigned long total_faults;
888 * Faults_cpu is used to decide whether memory should move
889 * towards the CPU. As a consequence, these stats are weighted
890 * more by CPU use than by memory faults.
892 unsigned long *faults_cpu;
893 unsigned long faults[0];
896 /* Shared or private faults. */
897 #define NR_NUMA_HINT_FAULT_TYPES 2
899 /* Memory and CPU locality */
900 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
902 /* Averaged statistics, and temporary buffers. */
903 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
905 pid_t task_numa_group_id(struct task_struct *p)
907 return p->numa_group ? p->numa_group->gid : 0;
911 * The averaged statistics, shared & private, memory & cpu,
912 * occupy the first half of the array. The second half of the
913 * array is for current counters, which are averaged into the
914 * first set by task_numa_placement.
916 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
918 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
921 static inline unsigned long task_faults(struct task_struct *p, int nid)
926 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
927 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
930 static inline unsigned long group_faults(struct task_struct *p, int nid)
935 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
936 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
939 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
941 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
942 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
945 /* Handle placement on systems where not all nodes are directly connected. */
946 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
947 int maxdist, bool task)
949 unsigned long score = 0;
953 * All nodes are directly connected, and the same distance
954 * from each other. No need for fancy placement algorithms.
956 if (sched_numa_topology_type == NUMA_DIRECT)
960 * This code is called for each node, introducing N^2 complexity,
961 * which should be ok given the number of nodes rarely exceeds 8.
963 for_each_online_node(node) {
964 unsigned long faults;
965 int dist = node_distance(nid, node);
968 * The furthest away nodes in the system are not interesting
969 * for placement; nid was already counted.
971 if (dist == sched_max_numa_distance || node == nid)
975 * On systems with a backplane NUMA topology, compare groups
976 * of nodes, and move tasks towards the group with the most
977 * memory accesses. When comparing two nodes at distance
978 * "hoplimit", only nodes closer by than "hoplimit" are part
979 * of each group. Skip other nodes.
981 if (sched_numa_topology_type == NUMA_BACKPLANE &&
985 /* Add up the faults from nearby nodes. */
987 faults = task_faults(p, node);
989 faults = group_faults(p, node);
992 * On systems with a glueless mesh NUMA topology, there are
993 * no fixed "groups of nodes". Instead, nodes that are not
994 * directly connected bounce traffic through intermediate
995 * nodes; a numa_group can occupy any set of nodes.
996 * The further away a node is, the less the faults count.
997 * This seems to result in good task placement.
999 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1000 faults *= (sched_max_numa_distance - dist);
1001 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1011 * These return the fraction of accesses done by a particular task, or
1012 * task group, on a particular numa node. The group weight is given a
1013 * larger multiplier, in order to group tasks together that are almost
1014 * evenly spread out between numa nodes.
1016 static inline unsigned long task_weight(struct task_struct *p, int nid,
1019 unsigned long faults, total_faults;
1021 if (!p->numa_faults)
1024 total_faults = p->total_numa_faults;
1029 faults = task_faults(p, nid);
1030 faults += score_nearby_nodes(p, nid, dist, true);
1032 return 1000 * faults / total_faults;
1035 static inline unsigned long group_weight(struct task_struct *p, int nid,
1038 unsigned long faults, total_faults;
1043 total_faults = p->numa_group->total_faults;
1048 faults = group_faults(p, nid);
1049 faults += score_nearby_nodes(p, nid, dist, false);
1051 return 1000 * faults / total_faults;
1054 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1055 int src_nid, int dst_cpu)
1057 struct numa_group *ng = p->numa_group;
1058 int dst_nid = cpu_to_node(dst_cpu);
1059 int last_cpupid, this_cpupid;
1061 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1064 * Multi-stage node selection is used in conjunction with a periodic
1065 * migration fault to build a temporal task<->page relation. By using
1066 * a two-stage filter we remove short/unlikely relations.
1068 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1069 * a task's usage of a particular page (n_p) per total usage of this
1070 * page (n_t) (in a given time-span) to a probability.
1072 * Our periodic faults will sample this probability and getting the
1073 * same result twice in a row, given these samples are fully
1074 * independent, is then given by P(n)^2, provided our sample period
1075 * is sufficiently short compared to the usage pattern.
1077 * This quadric squishes small probabilities, making it less likely we
1078 * act on an unlikely task<->page relation.
1080 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1081 if (!cpupid_pid_unset(last_cpupid) &&
1082 cpupid_to_nid(last_cpupid) != dst_nid)
1085 /* Always allow migrate on private faults */
1086 if (cpupid_match_pid(p, last_cpupid))
1089 /* A shared fault, but p->numa_group has not been set up yet. */
1094 * Do not migrate if the destination is not a node that
1095 * is actively used by this numa group.
1097 if (!node_isset(dst_nid, ng->active_nodes))
1101 * Source is a node that is not actively used by this
1102 * numa group, while the destination is. Migrate.
1104 if (!node_isset(src_nid, ng->active_nodes))
1108 * Both source and destination are nodes in active
1109 * use by this numa group. Maximize memory bandwidth
1110 * by migrating from more heavily used groups, to less
1111 * heavily used ones, spreading the load around.
1112 * Use a 1/4 hysteresis to avoid spurious page movement.
1114 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1117 static unsigned long weighted_cpuload(const int cpu);
1118 static unsigned long source_load(int cpu, int type);
1119 static unsigned long target_load(int cpu, int type);
1120 static unsigned long capacity_of(int cpu);
1121 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1123 /* Cached statistics for all CPUs within a node */
1125 unsigned long nr_running;
1128 /* Total compute capacity of CPUs on a node */
1129 unsigned long compute_capacity;
1131 /* Approximate capacity in terms of runnable tasks on a node */
1132 unsigned long task_capacity;
1133 int has_free_capacity;
1137 * XXX borrowed from update_sg_lb_stats
1139 static void update_numa_stats(struct numa_stats *ns, int nid)
1141 int smt, cpu, cpus = 0;
1142 unsigned long capacity;
1144 memset(ns, 0, sizeof(*ns));
1145 for_each_cpu(cpu, cpumask_of_node(nid)) {
1146 struct rq *rq = cpu_rq(cpu);
1148 ns->nr_running += rq->nr_running;
1149 ns->load += weighted_cpuload(cpu);
1150 ns->compute_capacity += capacity_of(cpu);
1156 * If we raced with hotplug and there are no CPUs left in our mask
1157 * the @ns structure is NULL'ed and task_numa_compare() will
1158 * not find this node attractive.
1160 * We'll either bail at !has_free_capacity, or we'll detect a huge
1161 * imbalance and bail there.
1166 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1167 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1168 capacity = cpus / smt; /* cores */
1170 ns->task_capacity = min_t(unsigned, capacity,
1171 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1172 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1175 struct task_numa_env {
1176 struct task_struct *p;
1178 int src_cpu, src_nid;
1179 int dst_cpu, dst_nid;
1181 struct numa_stats src_stats, dst_stats;
1186 struct task_struct *best_task;
1191 static void task_numa_assign(struct task_numa_env *env,
1192 struct task_struct *p, long imp)
1195 put_task_struct(env->best_task);
1200 env->best_imp = imp;
1201 env->best_cpu = env->dst_cpu;
1204 static bool load_too_imbalanced(long src_load, long dst_load,
1205 struct task_numa_env *env)
1208 long orig_src_load, orig_dst_load;
1209 long src_capacity, dst_capacity;
1212 * The load is corrected for the CPU capacity available on each node.
1215 * ------------ vs ---------
1216 * src_capacity dst_capacity
1218 src_capacity = env->src_stats.compute_capacity;
1219 dst_capacity = env->dst_stats.compute_capacity;
1221 /* We care about the slope of the imbalance, not the direction. */
1222 if (dst_load < src_load)
1223 swap(dst_load, src_load);
1225 /* Is the difference below the threshold? */
1226 imb = dst_load * src_capacity * 100 -
1227 src_load * dst_capacity * env->imbalance_pct;
1232 * The imbalance is above the allowed threshold.
1233 * Compare it with the old imbalance.
1235 orig_src_load = env->src_stats.load;
1236 orig_dst_load = env->dst_stats.load;
1238 if (orig_dst_load < orig_src_load)
1239 swap(orig_dst_load, orig_src_load);
1241 old_imb = orig_dst_load * src_capacity * 100 -
1242 orig_src_load * dst_capacity * env->imbalance_pct;
1244 /* Would this change make things worse? */
1245 return (imb > old_imb);
1249 * This checks if the overall compute and NUMA accesses of the system would
1250 * be improved if the source tasks was migrated to the target dst_cpu taking
1251 * into account that it might be best if task running on the dst_cpu should
1252 * be exchanged with the source task
1254 static void task_numa_compare(struct task_numa_env *env,
1255 long taskimp, long groupimp)
1257 struct rq *src_rq = cpu_rq(env->src_cpu);
1258 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1259 struct task_struct *cur;
1260 long src_load, dst_load;
1262 long imp = env->p->numa_group ? groupimp : taskimp;
1264 int dist = env->dist;
1268 raw_spin_lock_irq(&dst_rq->lock);
1271 * No need to move the exiting task, and this ensures that ->curr
1272 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1273 * is safe under RCU read lock.
1274 * Note that rcu_read_lock() itself can't protect from the final
1275 * put_task_struct() after the last schedule().
1277 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1279 raw_spin_unlock_irq(&dst_rq->lock);
1282 * Because we have preemption enabled we can get migrated around and
1283 * end try selecting ourselves (current == env->p) as a swap candidate.
1289 * "imp" is the fault differential for the source task between the
1290 * source and destination node. Calculate the total differential for
1291 * the source task and potential destination task. The more negative
1292 * the value is, the more rmeote accesses that would be expected to
1293 * be incurred if the tasks were swapped.
1296 /* Skip this swap candidate if cannot move to the source cpu */
1297 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1301 * If dst and source tasks are in the same NUMA group, or not
1302 * in any group then look only at task weights.
1304 if (cur->numa_group == env->p->numa_group) {
1305 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1306 task_weight(cur, env->dst_nid, dist);
1308 * Add some hysteresis to prevent swapping the
1309 * tasks within a group over tiny differences.
1311 if (cur->numa_group)
1315 * Compare the group weights. If a task is all by
1316 * itself (not part of a group), use the task weight
1319 if (cur->numa_group)
1320 imp += group_weight(cur, env->src_nid, dist) -
1321 group_weight(cur, env->dst_nid, dist);
1323 imp += task_weight(cur, env->src_nid, dist) -
1324 task_weight(cur, env->dst_nid, dist);
1328 if (imp <= env->best_imp && moveimp <= env->best_imp)
1332 /* Is there capacity at our destination? */
1333 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1334 !env->dst_stats.has_free_capacity)
1340 /* Balance doesn't matter much if we're running a task per cpu */
1341 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1342 dst_rq->nr_running == 1)
1346 * In the overloaded case, try and keep the load balanced.
1349 load = task_h_load(env->p);
1350 dst_load = env->dst_stats.load + load;
1351 src_load = env->src_stats.load - load;
1353 if (moveimp > imp && moveimp > env->best_imp) {
1355 * If the improvement from just moving env->p direction is
1356 * better than swapping tasks around, check if a move is
1357 * possible. Store a slightly smaller score than moveimp,
1358 * so an actually idle CPU will win.
1360 if (!load_too_imbalanced(src_load, dst_load, env)) {
1367 if (imp <= env->best_imp)
1371 load = task_h_load(cur);
1376 if (load_too_imbalanced(src_load, dst_load, env))
1380 * One idle CPU per node is evaluated for a task numa move.
1381 * Call select_idle_sibling to maybe find a better one.
1384 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1387 task_numa_assign(env, cur, imp);
1392 static void task_numa_find_cpu(struct task_numa_env *env,
1393 long taskimp, long groupimp)
1397 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1398 /* Skip this CPU if the source task cannot migrate */
1399 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1403 task_numa_compare(env, taskimp, groupimp);
1407 /* Only move tasks to a NUMA node less busy than the current node. */
1408 static bool numa_has_capacity(struct task_numa_env *env)
1410 struct numa_stats *src = &env->src_stats;
1411 struct numa_stats *dst = &env->dst_stats;
1413 if (src->has_free_capacity && !dst->has_free_capacity)
1417 * Only consider a task move if the source has a higher load
1418 * than the destination, corrected for CPU capacity on each node.
1420 * src->load dst->load
1421 * --------------------- vs ---------------------
1422 * src->compute_capacity dst->compute_capacity
1424 if (src->load * dst->compute_capacity * env->imbalance_pct >
1426 dst->load * src->compute_capacity * 100)
1432 static int task_numa_migrate(struct task_struct *p)
1434 struct task_numa_env env = {
1437 .src_cpu = task_cpu(p),
1438 .src_nid = task_node(p),
1440 .imbalance_pct = 112,
1446 struct sched_domain *sd;
1447 unsigned long taskweight, groupweight;
1449 long taskimp, groupimp;
1452 * Pick the lowest SD_NUMA domain, as that would have the smallest
1453 * imbalance and would be the first to start moving tasks about.
1455 * And we want to avoid any moving of tasks about, as that would create
1456 * random movement of tasks -- counter the numa conditions we're trying
1460 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1462 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1466 * Cpusets can break the scheduler domain tree into smaller
1467 * balance domains, some of which do not cross NUMA boundaries.
1468 * Tasks that are "trapped" in such domains cannot be migrated
1469 * elsewhere, so there is no point in (re)trying.
1471 if (unlikely(!sd)) {
1472 p->numa_preferred_nid = task_node(p);
1476 env.dst_nid = p->numa_preferred_nid;
1477 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1478 taskweight = task_weight(p, env.src_nid, dist);
1479 groupweight = group_weight(p, env.src_nid, dist);
1480 update_numa_stats(&env.src_stats, env.src_nid);
1481 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1482 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1483 update_numa_stats(&env.dst_stats, env.dst_nid);
1485 /* Try to find a spot on the preferred nid. */
1486 if (numa_has_capacity(&env))
1487 task_numa_find_cpu(&env, taskimp, groupimp);
1490 * Look at other nodes in these cases:
1491 * - there is no space available on the preferred_nid
1492 * - the task is part of a numa_group that is interleaved across
1493 * multiple NUMA nodes; in order to better consolidate the group,
1494 * we need to check other locations.
1496 if (env.best_cpu == -1 || (p->numa_group &&
1497 nodes_weight(p->numa_group->active_nodes) > 1)) {
1498 for_each_online_node(nid) {
1499 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1502 dist = node_distance(env.src_nid, env.dst_nid);
1503 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1505 taskweight = task_weight(p, env.src_nid, dist);
1506 groupweight = group_weight(p, env.src_nid, dist);
1509 /* Only consider nodes where both task and groups benefit */
1510 taskimp = task_weight(p, nid, dist) - taskweight;
1511 groupimp = group_weight(p, nid, dist) - groupweight;
1512 if (taskimp < 0 && groupimp < 0)
1517 update_numa_stats(&env.dst_stats, env.dst_nid);
1518 if (numa_has_capacity(&env))
1519 task_numa_find_cpu(&env, taskimp, groupimp);
1524 * If the task is part of a workload that spans multiple NUMA nodes,
1525 * and is migrating into one of the workload's active nodes, remember
1526 * this node as the task's preferred numa node, so the workload can
1528 * A task that migrated to a second choice node will be better off
1529 * trying for a better one later. Do not set the preferred node here.
1531 if (p->numa_group) {
1532 if (env.best_cpu == -1)
1537 if (node_isset(nid, p->numa_group->active_nodes))
1538 sched_setnuma(p, env.dst_nid);
1541 /* No better CPU than the current one was found. */
1542 if (env.best_cpu == -1)
1546 * Reset the scan period if the task is being rescheduled on an
1547 * alternative node to recheck if the tasks is now properly placed.
1549 p->numa_scan_period = task_scan_min(p);
1551 if (env.best_task == NULL) {
1552 ret = migrate_task_to(p, env.best_cpu);
1554 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1558 ret = migrate_swap(p, env.best_task);
1560 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1561 put_task_struct(env.best_task);
1565 /* Attempt to migrate a task to a CPU on the preferred node. */
1566 static void numa_migrate_preferred(struct task_struct *p)
1568 unsigned long interval = HZ;
1570 /* This task has no NUMA fault statistics yet */
1571 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1574 /* Periodically retry migrating the task to the preferred node */
1575 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1576 p->numa_migrate_retry = jiffies + interval;
1578 /* Success if task is already running on preferred CPU */
1579 if (task_node(p) == p->numa_preferred_nid)
1582 /* Otherwise, try migrate to a CPU on the preferred node */
1583 task_numa_migrate(p);
1587 * Find the nodes on which the workload is actively running. We do this by
1588 * tracking the nodes from which NUMA hinting faults are triggered. This can
1589 * be different from the set of nodes where the workload's memory is currently
1592 * The bitmask is used to make smarter decisions on when to do NUMA page
1593 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1594 * are added when they cause over 6/16 of the maximum number of faults, but
1595 * only removed when they drop below 3/16.
1597 static void update_numa_active_node_mask(struct numa_group *numa_group)
1599 unsigned long faults, max_faults = 0;
1602 for_each_online_node(nid) {
1603 faults = group_faults_cpu(numa_group, nid);
1604 if (faults > max_faults)
1605 max_faults = faults;
1608 for_each_online_node(nid) {
1609 faults = group_faults_cpu(numa_group, nid);
1610 if (!node_isset(nid, numa_group->active_nodes)) {
1611 if (faults > max_faults * 6 / 16)
1612 node_set(nid, numa_group->active_nodes);
1613 } else if (faults < max_faults * 3 / 16)
1614 node_clear(nid, numa_group->active_nodes);
1619 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1620 * increments. The more local the fault statistics are, the higher the scan
1621 * period will be for the next scan window. If local/(local+remote) ratio is
1622 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1623 * the scan period will decrease. Aim for 70% local accesses.
1625 #define NUMA_PERIOD_SLOTS 10
1626 #define NUMA_PERIOD_THRESHOLD 7
1629 * Increase the scan period (slow down scanning) if the majority of
1630 * our memory is already on our local node, or if the majority of
1631 * the page accesses are shared with other processes.
1632 * Otherwise, decrease the scan period.
1634 static void update_task_scan_period(struct task_struct *p,
1635 unsigned long shared, unsigned long private)
1637 unsigned int period_slot;
1641 unsigned long remote = p->numa_faults_locality[0];
1642 unsigned long local = p->numa_faults_locality[1];
1645 * If there were no record hinting faults then either the task is
1646 * completely idle or all activity is areas that are not of interest
1647 * to automatic numa balancing. Related to that, if there were failed
1648 * migration then it implies we are migrating too quickly or the local
1649 * node is overloaded. In either case, scan slower
1651 if (local + shared == 0 || p->numa_faults_locality[2]) {
1652 p->numa_scan_period = min(p->numa_scan_period_max,
1653 p->numa_scan_period << 1);
1655 p->mm->numa_next_scan = jiffies +
1656 msecs_to_jiffies(p->numa_scan_period);
1662 * Prepare to scale scan period relative to the current period.
1663 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1664 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1665 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1667 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1668 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1669 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1670 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1673 diff = slot * period_slot;
1675 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1678 * Scale scan rate increases based on sharing. There is an
1679 * inverse relationship between the degree of sharing and
1680 * the adjustment made to the scanning period. Broadly
1681 * speaking the intent is that there is little point
1682 * scanning faster if shared accesses dominate as it may
1683 * simply bounce migrations uselessly
1685 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1686 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1689 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1690 task_scan_min(p), task_scan_max(p));
1691 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1695 * Get the fraction of time the task has been running since the last
1696 * NUMA placement cycle. The scheduler keeps similar statistics, but
1697 * decays those on a 32ms period, which is orders of magnitude off
1698 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1699 * stats only if the task is so new there are no NUMA statistics yet.
1701 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1703 u64 runtime, delta, now;
1704 /* Use the start of this time slice to avoid calculations. */
1705 now = p->se.exec_start;
1706 runtime = p->se.sum_exec_runtime;
1708 if (p->last_task_numa_placement) {
1709 delta = runtime - p->last_sum_exec_runtime;
1710 *period = now - p->last_task_numa_placement;
1712 delta = p->se.avg.load_sum / p->se.load.weight;
1713 *period = LOAD_AVG_MAX;
1716 p->last_sum_exec_runtime = runtime;
1717 p->last_task_numa_placement = now;
1723 * Determine the preferred nid for a task in a numa_group. This needs to
1724 * be done in a way that produces consistent results with group_weight,
1725 * otherwise workloads might not converge.
1727 static int preferred_group_nid(struct task_struct *p, int nid)
1732 /* Direct connections between all NUMA nodes. */
1733 if (sched_numa_topology_type == NUMA_DIRECT)
1737 * On a system with glueless mesh NUMA topology, group_weight
1738 * scores nodes according to the number of NUMA hinting faults on
1739 * both the node itself, and on nearby nodes.
1741 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1742 unsigned long score, max_score = 0;
1743 int node, max_node = nid;
1745 dist = sched_max_numa_distance;
1747 for_each_online_node(node) {
1748 score = group_weight(p, node, dist);
1749 if (score > max_score) {
1758 * Finding the preferred nid in a system with NUMA backplane
1759 * interconnect topology is more involved. The goal is to locate
1760 * tasks from numa_groups near each other in the system, and
1761 * untangle workloads from different sides of the system. This requires
1762 * searching down the hierarchy of node groups, recursively searching
1763 * inside the highest scoring group of nodes. The nodemask tricks
1764 * keep the complexity of the search down.
1766 nodes = node_online_map;
1767 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1768 unsigned long max_faults = 0;
1769 nodemask_t max_group = NODE_MASK_NONE;
1772 /* Are there nodes at this distance from each other? */
1773 if (!find_numa_distance(dist))
1776 for_each_node_mask(a, nodes) {
1777 unsigned long faults = 0;
1778 nodemask_t this_group;
1779 nodes_clear(this_group);
1781 /* Sum group's NUMA faults; includes a==b case. */
1782 for_each_node_mask(b, nodes) {
1783 if (node_distance(a, b) < dist) {
1784 faults += group_faults(p, b);
1785 node_set(b, this_group);
1786 node_clear(b, nodes);
1790 /* Remember the top group. */
1791 if (faults > max_faults) {
1792 max_faults = faults;
1793 max_group = this_group;
1795 * subtle: at the smallest distance there is
1796 * just one node left in each "group", the
1797 * winner is the preferred nid.
1802 /* Next round, evaluate the nodes within max_group. */
1810 static void task_numa_placement(struct task_struct *p)
1812 int seq, nid, max_nid = -1, max_group_nid = -1;
1813 unsigned long max_faults = 0, max_group_faults = 0;
1814 unsigned long fault_types[2] = { 0, 0 };
1815 unsigned long total_faults;
1816 u64 runtime, period;
1817 spinlock_t *group_lock = NULL;
1820 * The p->mm->numa_scan_seq field gets updated without
1821 * exclusive access. Use READ_ONCE() here to ensure
1822 * that the field is read in a single access:
1824 seq = READ_ONCE(p->mm->numa_scan_seq);
1825 if (p->numa_scan_seq == seq)
1827 p->numa_scan_seq = seq;
1828 p->numa_scan_period_max = task_scan_max(p);
1830 total_faults = p->numa_faults_locality[0] +
1831 p->numa_faults_locality[1];
1832 runtime = numa_get_avg_runtime(p, &period);
1834 /* If the task is part of a group prevent parallel updates to group stats */
1835 if (p->numa_group) {
1836 group_lock = &p->numa_group->lock;
1837 spin_lock_irq(group_lock);
1840 /* Find the node with the highest number of faults */
1841 for_each_online_node(nid) {
1842 /* Keep track of the offsets in numa_faults array */
1843 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1844 unsigned long faults = 0, group_faults = 0;
1847 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1848 long diff, f_diff, f_weight;
1850 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1851 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1852 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1853 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1855 /* Decay existing window, copy faults since last scan */
1856 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1857 fault_types[priv] += p->numa_faults[membuf_idx];
1858 p->numa_faults[membuf_idx] = 0;
1861 * Normalize the faults_from, so all tasks in a group
1862 * count according to CPU use, instead of by the raw
1863 * number of faults. Tasks with little runtime have
1864 * little over-all impact on throughput, and thus their
1865 * faults are less important.
1867 f_weight = div64_u64(runtime << 16, period + 1);
1868 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1870 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1871 p->numa_faults[cpubuf_idx] = 0;
1873 p->numa_faults[mem_idx] += diff;
1874 p->numa_faults[cpu_idx] += f_diff;
1875 faults += p->numa_faults[mem_idx];
1876 p->total_numa_faults += diff;
1877 if (p->numa_group) {
1879 * safe because we can only change our own group
1881 * mem_idx represents the offset for a given
1882 * nid and priv in a specific region because it
1883 * is at the beginning of the numa_faults array.
1885 p->numa_group->faults[mem_idx] += diff;
1886 p->numa_group->faults_cpu[mem_idx] += f_diff;
1887 p->numa_group->total_faults += diff;
1888 group_faults += p->numa_group->faults[mem_idx];
1892 if (faults > max_faults) {
1893 max_faults = faults;
1897 if (group_faults > max_group_faults) {
1898 max_group_faults = group_faults;
1899 max_group_nid = nid;
1903 update_task_scan_period(p, fault_types[0], fault_types[1]);
1905 if (p->numa_group) {
1906 update_numa_active_node_mask(p->numa_group);
1907 spin_unlock_irq(group_lock);
1908 max_nid = preferred_group_nid(p, max_group_nid);
1912 /* Set the new preferred node */
1913 if (max_nid != p->numa_preferred_nid)
1914 sched_setnuma(p, max_nid);
1916 if (task_node(p) != p->numa_preferred_nid)
1917 numa_migrate_preferred(p);
1921 static inline int get_numa_group(struct numa_group *grp)
1923 return atomic_inc_not_zero(&grp->refcount);
1926 static inline void put_numa_group(struct numa_group *grp)
1928 if (atomic_dec_and_test(&grp->refcount))
1929 kfree_rcu(grp, rcu);
1932 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1935 struct numa_group *grp, *my_grp;
1936 struct task_struct *tsk;
1938 int cpu = cpupid_to_cpu(cpupid);
1941 if (unlikely(!p->numa_group)) {
1942 unsigned int size = sizeof(struct numa_group) +
1943 4*nr_node_ids*sizeof(unsigned long);
1945 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1949 atomic_set(&grp->refcount, 1);
1950 spin_lock_init(&grp->lock);
1952 /* Second half of the array tracks nids where faults happen */
1953 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1956 node_set(task_node(current), grp->active_nodes);
1958 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1959 grp->faults[i] = p->numa_faults[i];
1961 grp->total_faults = p->total_numa_faults;
1964 rcu_assign_pointer(p->numa_group, grp);
1968 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1970 if (!cpupid_match_pid(tsk, cpupid))
1973 grp = rcu_dereference(tsk->numa_group);
1977 my_grp = p->numa_group;
1982 * Only join the other group if its bigger; if we're the bigger group,
1983 * the other task will join us.
1985 if (my_grp->nr_tasks > grp->nr_tasks)
1989 * Tie-break on the grp address.
1991 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1994 /* Always join threads in the same process. */
1995 if (tsk->mm == current->mm)
1998 /* Simple filter to avoid false positives due to PID collisions */
1999 if (flags & TNF_SHARED)
2002 /* Update priv based on whether false sharing was detected */
2005 if (join && !get_numa_group(grp))
2013 BUG_ON(irqs_disabled());
2014 double_lock_irq(&my_grp->lock, &grp->lock);
2016 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2017 my_grp->faults[i] -= p->numa_faults[i];
2018 grp->faults[i] += p->numa_faults[i];
2020 my_grp->total_faults -= p->total_numa_faults;
2021 grp->total_faults += p->total_numa_faults;
2026 spin_unlock(&my_grp->lock);
2027 spin_unlock_irq(&grp->lock);
2029 rcu_assign_pointer(p->numa_group, grp);
2031 put_numa_group(my_grp);
2039 void task_numa_free(struct task_struct *p)
2041 struct numa_group *grp = p->numa_group;
2042 void *numa_faults = p->numa_faults;
2043 unsigned long flags;
2047 spin_lock_irqsave(&grp->lock, flags);
2048 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2049 grp->faults[i] -= p->numa_faults[i];
2050 grp->total_faults -= p->total_numa_faults;
2053 spin_unlock_irqrestore(&grp->lock, flags);
2054 RCU_INIT_POINTER(p->numa_group, NULL);
2055 put_numa_group(grp);
2058 p->numa_faults = NULL;
2063 * Got a PROT_NONE fault for a page on @node.
2065 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2067 struct task_struct *p = current;
2068 bool migrated = flags & TNF_MIGRATED;
2069 int cpu_node = task_node(current);
2070 int local = !!(flags & TNF_FAULT_LOCAL);
2073 if (!static_branch_likely(&sched_numa_balancing))
2076 /* for example, ksmd faulting in a user's mm */
2080 /* Allocate buffer to track faults on a per-node basis */
2081 if (unlikely(!p->numa_faults)) {
2082 int size = sizeof(*p->numa_faults) *
2083 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2085 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2086 if (!p->numa_faults)
2089 p->total_numa_faults = 0;
2090 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2094 * First accesses are treated as private, otherwise consider accesses
2095 * to be private if the accessing pid has not changed
2097 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2100 priv = cpupid_match_pid(p, last_cpupid);
2101 if (!priv && !(flags & TNF_NO_GROUP))
2102 task_numa_group(p, last_cpupid, flags, &priv);
2106 * If a workload spans multiple NUMA nodes, a shared fault that
2107 * occurs wholly within the set of nodes that the workload is
2108 * actively using should be counted as local. This allows the
2109 * scan rate to slow down when a workload has settled down.
2111 if (!priv && !local && p->numa_group &&
2112 node_isset(cpu_node, p->numa_group->active_nodes) &&
2113 node_isset(mem_node, p->numa_group->active_nodes))
2116 task_numa_placement(p);
2119 * Retry task to preferred node migration periodically, in case it
2120 * case it previously failed, or the scheduler moved us.
2122 if (time_after(jiffies, p->numa_migrate_retry))
2123 numa_migrate_preferred(p);
2126 p->numa_pages_migrated += pages;
2127 if (flags & TNF_MIGRATE_FAIL)
2128 p->numa_faults_locality[2] += pages;
2130 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2131 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2132 p->numa_faults_locality[local] += pages;
2135 static void reset_ptenuma_scan(struct task_struct *p)
2138 * We only did a read acquisition of the mmap sem, so
2139 * p->mm->numa_scan_seq is written to without exclusive access
2140 * and the update is not guaranteed to be atomic. That's not
2141 * much of an issue though, since this is just used for
2142 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2143 * expensive, to avoid any form of compiler optimizations:
2145 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2146 p->mm->numa_scan_offset = 0;
2150 * The expensive part of numa migration is done from task_work context.
2151 * Triggered from task_tick_numa().
2153 void task_numa_work(struct callback_head *work)
2155 unsigned long migrate, next_scan, now = jiffies;
2156 struct task_struct *p = current;
2157 struct mm_struct *mm = p->mm;
2158 struct vm_area_struct *vma;
2159 unsigned long start, end;
2160 unsigned long nr_pte_updates = 0;
2161 long pages, virtpages;
2163 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2165 work->next = work; /* protect against double add */
2167 * Who cares about NUMA placement when they're dying.
2169 * NOTE: make sure not to dereference p->mm before this check,
2170 * exit_task_work() happens _after_ exit_mm() so we could be called
2171 * without p->mm even though we still had it when we enqueued this
2174 if (p->flags & PF_EXITING)
2177 if (!mm->numa_next_scan) {
2178 mm->numa_next_scan = now +
2179 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2183 * Enforce maximal scan/migration frequency..
2185 migrate = mm->numa_next_scan;
2186 if (time_before(now, migrate))
2189 if (p->numa_scan_period == 0) {
2190 p->numa_scan_period_max = task_scan_max(p);
2191 p->numa_scan_period = task_scan_min(p);
2194 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2195 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2199 * Delay this task enough that another task of this mm will likely win
2200 * the next time around.
2202 p->node_stamp += 2 * TICK_NSEC;
2204 start = mm->numa_scan_offset;
2205 pages = sysctl_numa_balancing_scan_size;
2206 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2207 virtpages = pages * 8; /* Scan up to this much virtual space */
2212 down_read(&mm->mmap_sem);
2213 vma = find_vma(mm, start);
2215 reset_ptenuma_scan(p);
2219 for (; vma; vma = vma->vm_next) {
2220 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2221 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2226 * Shared library pages mapped by multiple processes are not
2227 * migrated as it is expected they are cache replicated. Avoid
2228 * hinting faults in read-only file-backed mappings or the vdso
2229 * as migrating the pages will be of marginal benefit.
2232 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2236 * Skip inaccessible VMAs to avoid any confusion between
2237 * PROT_NONE and NUMA hinting ptes
2239 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2243 start = max(start, vma->vm_start);
2244 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2245 end = min(end, vma->vm_end);
2246 nr_pte_updates = change_prot_numa(vma, start, end);
2249 * Try to scan sysctl_numa_balancing_size worth of
2250 * hpages that have at least one present PTE that
2251 * is not already pte-numa. If the VMA contains
2252 * areas that are unused or already full of prot_numa
2253 * PTEs, scan up to virtpages, to skip through those
2257 pages -= (end - start) >> PAGE_SHIFT;
2258 virtpages -= (end - start) >> PAGE_SHIFT;
2261 if (pages <= 0 || virtpages <= 0)
2265 } while (end != vma->vm_end);
2270 * It is possible to reach the end of the VMA list but the last few
2271 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2272 * would find the !migratable VMA on the next scan but not reset the
2273 * scanner to the start so check it now.
2276 mm->numa_scan_offset = start;
2278 reset_ptenuma_scan(p);
2279 up_read(&mm->mmap_sem);
2283 * Drive the periodic memory faults..
2285 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2287 struct callback_head *work = &curr->numa_work;
2291 * We don't care about NUMA placement if we don't have memory.
2293 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2297 * Using runtime rather than walltime has the dual advantage that
2298 * we (mostly) drive the selection from busy threads and that the
2299 * task needs to have done some actual work before we bother with
2302 now = curr->se.sum_exec_runtime;
2303 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2305 if (now - curr->node_stamp > period) {
2306 if (!curr->node_stamp)
2307 curr->numa_scan_period = task_scan_min(curr);
2308 curr->node_stamp += period;
2310 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2311 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2312 task_work_add(curr, work, true);
2317 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2321 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2325 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2328 #endif /* CONFIG_NUMA_BALANCING */
2331 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2333 update_load_add(&cfs_rq->load, se->load.weight);
2334 if (!parent_entity(se))
2335 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2337 if (entity_is_task(se)) {
2338 struct rq *rq = rq_of(cfs_rq);
2340 account_numa_enqueue(rq, task_of(se));
2341 list_add(&se->group_node, &rq->cfs_tasks);
2344 cfs_rq->nr_running++;
2348 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2350 update_load_sub(&cfs_rq->load, se->load.weight);
2351 if (!parent_entity(se))
2352 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2353 if (entity_is_task(se)) {
2354 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2355 list_del_init(&se->group_node);
2357 cfs_rq->nr_running--;
2360 #ifdef CONFIG_FAIR_GROUP_SCHED
2362 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2367 * Use this CPU's real-time load instead of the last load contribution
2368 * as the updating of the contribution is delayed, and we will use the
2369 * the real-time load to calc the share. See update_tg_load_avg().
2371 tg_weight = atomic_long_read(&tg->load_avg);
2372 tg_weight -= cfs_rq->tg_load_avg_contrib;
2373 tg_weight += cfs_rq_load_avg(cfs_rq);
2378 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2380 long tg_weight, load, shares;
2382 tg_weight = calc_tg_weight(tg, cfs_rq);
2383 load = cfs_rq_load_avg(cfs_rq);
2385 shares = (tg->shares * load);
2387 shares /= tg_weight;
2389 if (shares < MIN_SHARES)
2390 shares = MIN_SHARES;
2391 if (shares > tg->shares)
2392 shares = tg->shares;
2396 # else /* CONFIG_SMP */
2397 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2401 # endif /* CONFIG_SMP */
2402 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2403 unsigned long weight)
2406 /* commit outstanding execution time */
2407 if (cfs_rq->curr == se)
2408 update_curr(cfs_rq);
2409 account_entity_dequeue(cfs_rq, se);
2412 update_load_set(&se->load, weight);
2415 account_entity_enqueue(cfs_rq, se);
2418 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2420 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2422 struct task_group *tg;
2423 struct sched_entity *se;
2427 se = tg->se[cpu_of(rq_of(cfs_rq))];
2428 if (!se || throttled_hierarchy(cfs_rq))
2431 if (likely(se->load.weight == tg->shares))
2434 shares = calc_cfs_shares(cfs_rq, tg);
2436 reweight_entity(cfs_rq_of(se), se, shares);
2438 #else /* CONFIG_FAIR_GROUP_SCHED */
2439 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2442 #endif /* CONFIG_FAIR_GROUP_SCHED */
2445 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2446 static const u32 runnable_avg_yN_inv[] = {
2447 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2448 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2449 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2450 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2451 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2452 0x85aac367, 0x82cd8698,
2456 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2457 * over-estimates when re-combining.
2459 static const u32 runnable_avg_yN_sum[] = {
2460 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2461 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2462 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2467 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2469 static __always_inline u64 decay_load(u64 val, u64 n)
2471 unsigned int local_n;
2475 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2478 /* after bounds checking we can collapse to 32-bit */
2482 * As y^PERIOD = 1/2, we can combine
2483 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2484 * With a look-up table which covers y^n (n<PERIOD)
2486 * To achieve constant time decay_load.
2488 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2489 val >>= local_n / LOAD_AVG_PERIOD;
2490 local_n %= LOAD_AVG_PERIOD;
2493 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2498 * For updates fully spanning n periods, the contribution to runnable
2499 * average will be: \Sum 1024*y^n
2501 * We can compute this reasonably efficiently by combining:
2502 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2504 static u32 __compute_runnable_contrib(u64 n)
2508 if (likely(n <= LOAD_AVG_PERIOD))
2509 return runnable_avg_yN_sum[n];
2510 else if (unlikely(n >= LOAD_AVG_MAX_N))
2511 return LOAD_AVG_MAX;
2513 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2515 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2516 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2518 n -= LOAD_AVG_PERIOD;
2519 } while (n > LOAD_AVG_PERIOD);
2521 contrib = decay_load(contrib, n);
2522 return contrib + runnable_avg_yN_sum[n];
2525 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2526 #error "load tracking assumes 2^10 as unit"
2529 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2532 * We can represent the historical contribution to runnable average as the
2533 * coefficients of a geometric series. To do this we sub-divide our runnable
2534 * history into segments of approximately 1ms (1024us); label the segment that
2535 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2537 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2539 * (now) (~1ms ago) (~2ms ago)
2541 * Let u_i denote the fraction of p_i that the entity was runnable.
2543 * We then designate the fractions u_i as our co-efficients, yielding the
2544 * following representation of historical load:
2545 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2547 * We choose y based on the with of a reasonably scheduling period, fixing:
2550 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2551 * approximately half as much as the contribution to load within the last ms
2554 * When a period "rolls over" and we have new u_0`, multiplying the previous
2555 * sum again by y is sufficient to update:
2556 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2557 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2559 static __always_inline int
2560 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2561 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2563 u64 delta, scaled_delta, periods;
2565 unsigned int delta_w, scaled_delta_w, decayed = 0;
2566 unsigned long scale_freq, scale_cpu;
2568 delta = now - sa->last_update_time;
2570 * This should only happen when time goes backwards, which it
2571 * unfortunately does during sched clock init when we swap over to TSC.
2573 if ((s64)delta < 0) {
2574 sa->last_update_time = now;
2579 * Use 1024ns as the unit of measurement since it's a reasonable
2580 * approximation of 1us and fast to compute.
2585 sa->last_update_time = now;
2587 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2588 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2590 /* delta_w is the amount already accumulated against our next period */
2591 delta_w = sa->period_contrib;
2592 if (delta + delta_w >= 1024) {
2595 /* how much left for next period will start over, we don't know yet */
2596 sa->period_contrib = 0;
2599 * Now that we know we're crossing a period boundary, figure
2600 * out how much from delta we need to complete the current
2601 * period and accrue it.
2603 delta_w = 1024 - delta_w;
2604 scaled_delta_w = cap_scale(delta_w, scale_freq);
2606 sa->load_sum += weight * scaled_delta_w;
2608 cfs_rq->runnable_load_sum +=
2609 weight * scaled_delta_w;
2613 sa->util_sum += scaled_delta_w * scale_cpu;
2617 /* Figure out how many additional periods this update spans */
2618 periods = delta / 1024;
2621 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2623 cfs_rq->runnable_load_sum =
2624 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2626 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2628 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2629 contrib = __compute_runnable_contrib(periods);
2630 contrib = cap_scale(contrib, scale_freq);
2632 sa->load_sum += weight * contrib;
2634 cfs_rq->runnable_load_sum += weight * contrib;
2637 sa->util_sum += contrib * scale_cpu;
2640 /* Remainder of delta accrued against u_0` */
2641 scaled_delta = cap_scale(delta, scale_freq);
2643 sa->load_sum += weight * scaled_delta;
2645 cfs_rq->runnable_load_sum += weight * scaled_delta;
2648 sa->util_sum += scaled_delta * scale_cpu;
2650 sa->period_contrib += delta;
2653 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2655 cfs_rq->runnable_load_avg =
2656 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2658 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2664 #ifdef CONFIG_FAIR_GROUP_SCHED
2666 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2667 * and effective_load (which is not done because it is too costly).
2669 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2671 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2673 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2674 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2675 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2679 #else /* CONFIG_FAIR_GROUP_SCHED */
2680 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2681 #endif /* CONFIG_FAIR_GROUP_SCHED */
2683 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2685 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2686 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2688 struct sched_avg *sa = &cfs_rq->avg;
2691 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2692 long r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2693 sa->load_avg = max_t(long, sa->load_avg - r, 0);
2694 sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
2697 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2698 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2699 sa->util_avg = max_t(long, sa->util_avg - r, 0);
2700 sa->util_sum = max_t(s32, sa->util_sum - r * LOAD_AVG_MAX, 0);
2703 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2704 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2706 #ifndef CONFIG_64BIT
2708 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2714 /* Update task and its cfs_rq load average */
2715 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2717 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2718 u64 now = cfs_rq_clock_task(cfs_rq);
2719 int cpu = cpu_of(rq_of(cfs_rq));
2722 * Track task load average for carrying it to new CPU after migrated, and
2723 * track group sched_entity load average for task_h_load calc in migration
2725 __update_load_avg(now, cpu, &se->avg,
2726 se->on_rq * scale_load_down(se->load.weight),
2727 cfs_rq->curr == se, NULL);
2729 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2730 update_tg_load_avg(cfs_rq, 0);
2733 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2735 if (!sched_feat(ATTACH_AGE_LOAD))
2739 * If we got migrated (either between CPUs or between cgroups) we'll
2740 * have aged the average right before clearing @last_update_time.
2742 if (se->avg.last_update_time) {
2743 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2744 &se->avg, 0, 0, NULL);
2747 * XXX: we could have just aged the entire load away if we've been
2748 * absent from the fair class for too long.
2753 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2754 cfs_rq->avg.load_avg += se->avg.load_avg;
2755 cfs_rq->avg.load_sum += se->avg.load_sum;
2756 cfs_rq->avg.util_avg += se->avg.util_avg;
2757 cfs_rq->avg.util_sum += se->avg.util_sum;
2760 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2762 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2763 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2764 cfs_rq->curr == se, NULL);
2766 cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
2767 cfs_rq->avg.load_sum = max_t(s64, cfs_rq->avg.load_sum - se->avg.load_sum, 0);
2768 cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
2769 cfs_rq->avg.util_sum = max_t(s32, cfs_rq->avg.util_sum - se->avg.util_sum, 0);
2772 /* Add the load generated by se into cfs_rq's load average */
2774 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2776 struct sched_avg *sa = &se->avg;
2777 u64 now = cfs_rq_clock_task(cfs_rq);
2778 int migrated, decayed;
2780 migrated = !sa->last_update_time;
2782 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2783 se->on_rq * scale_load_down(se->load.weight),
2784 cfs_rq->curr == se, NULL);
2787 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2789 cfs_rq->runnable_load_avg += sa->load_avg;
2790 cfs_rq->runnable_load_sum += sa->load_sum;
2793 attach_entity_load_avg(cfs_rq, se);
2795 if (decayed || migrated)
2796 update_tg_load_avg(cfs_rq, 0);
2799 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2801 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2803 update_load_avg(se, 1);
2805 cfs_rq->runnable_load_avg =
2806 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2807 cfs_rq->runnable_load_sum =
2808 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2812 * Task first catches up with cfs_rq, and then subtract
2813 * itself from the cfs_rq (task must be off the queue now).
2815 void remove_entity_load_avg(struct sched_entity *se)
2817 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2818 u64 last_update_time;
2820 #ifndef CONFIG_64BIT
2821 u64 last_update_time_copy;
2824 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2826 last_update_time = cfs_rq->avg.last_update_time;
2827 } while (last_update_time != last_update_time_copy);
2829 last_update_time = cfs_rq->avg.last_update_time;
2832 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2833 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2834 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2838 * Update the rq's load with the elapsed running time before entering
2839 * idle. if the last scheduled task is not a CFS task, idle_enter will
2840 * be the only way to update the runnable statistic.
2842 void idle_enter_fair(struct rq *this_rq)
2847 * Update the rq's load with the elapsed idle time before a task is
2848 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2849 * be the only way to update the runnable statistic.
2851 void idle_exit_fair(struct rq *this_rq)
2855 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2857 return cfs_rq->runnable_load_avg;
2860 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2862 return cfs_rq->avg.load_avg;
2865 static int idle_balance(struct rq *this_rq);
2867 #else /* CONFIG_SMP */
2869 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2871 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2873 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2874 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2877 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2879 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2881 static inline int idle_balance(struct rq *rq)
2886 #endif /* CONFIG_SMP */
2888 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2890 #ifdef CONFIG_SCHEDSTATS
2891 struct task_struct *tsk = NULL;
2893 if (entity_is_task(se))
2896 if (se->statistics.sleep_start) {
2897 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2902 if (unlikely(delta > se->statistics.sleep_max))
2903 se->statistics.sleep_max = delta;
2905 se->statistics.sleep_start = 0;
2906 se->statistics.sum_sleep_runtime += delta;
2909 account_scheduler_latency(tsk, delta >> 10, 1);
2910 trace_sched_stat_sleep(tsk, delta);
2913 if (se->statistics.block_start) {
2914 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2919 if (unlikely(delta > se->statistics.block_max))
2920 se->statistics.block_max = delta;
2922 se->statistics.block_start = 0;
2923 se->statistics.sum_sleep_runtime += delta;
2926 if (tsk->in_iowait) {
2927 se->statistics.iowait_sum += delta;
2928 se->statistics.iowait_count++;
2929 trace_sched_stat_iowait(tsk, delta);
2932 trace_sched_stat_blocked(tsk, delta);
2935 * Blocking time is in units of nanosecs, so shift by
2936 * 20 to get a milliseconds-range estimation of the
2937 * amount of time that the task spent sleeping:
2939 if (unlikely(prof_on == SLEEP_PROFILING)) {
2940 profile_hits(SLEEP_PROFILING,
2941 (void *)get_wchan(tsk),
2944 account_scheduler_latency(tsk, delta >> 10, 0);
2950 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2952 #ifdef CONFIG_SCHED_DEBUG
2953 s64 d = se->vruntime - cfs_rq->min_vruntime;
2958 if (d > 3*sysctl_sched_latency)
2959 schedstat_inc(cfs_rq, nr_spread_over);
2964 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2966 u64 vruntime = cfs_rq->min_vruntime;
2969 * The 'current' period is already promised to the current tasks,
2970 * however the extra weight of the new task will slow them down a
2971 * little, place the new task so that it fits in the slot that
2972 * stays open at the end.
2974 if (initial && sched_feat(START_DEBIT))
2975 vruntime += sched_vslice(cfs_rq, se);
2977 /* sleeps up to a single latency don't count. */
2979 unsigned long thresh = sysctl_sched_latency;
2982 * Halve their sleep time's effect, to allow
2983 * for a gentler effect of sleepers:
2985 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2991 /* ensure we never gain time by being placed backwards. */
2992 se->vruntime = max_vruntime(se->vruntime, vruntime);
2995 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2998 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3001 * Update the normalized vruntime before updating min_vruntime
3002 * through calling update_curr().
3004 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3005 se->vruntime += cfs_rq->min_vruntime;
3008 * Update run-time statistics of the 'current'.
3010 update_curr(cfs_rq);
3011 enqueue_entity_load_avg(cfs_rq, se);
3012 account_entity_enqueue(cfs_rq, se);
3013 update_cfs_shares(cfs_rq);
3015 if (flags & ENQUEUE_WAKEUP) {
3016 place_entity(cfs_rq, se, 0);
3017 enqueue_sleeper(cfs_rq, se);
3020 update_stats_enqueue(cfs_rq, se);
3021 check_spread(cfs_rq, se);
3022 if (se != cfs_rq->curr)
3023 __enqueue_entity(cfs_rq, se);
3026 if (cfs_rq->nr_running == 1) {
3027 list_add_leaf_cfs_rq(cfs_rq);
3028 check_enqueue_throttle(cfs_rq);
3032 static void __clear_buddies_last(struct sched_entity *se)
3034 for_each_sched_entity(se) {
3035 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3036 if (cfs_rq->last != se)
3039 cfs_rq->last = NULL;
3043 static void __clear_buddies_next(struct sched_entity *se)
3045 for_each_sched_entity(se) {
3046 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3047 if (cfs_rq->next != se)
3050 cfs_rq->next = NULL;
3054 static void __clear_buddies_skip(struct sched_entity *se)
3056 for_each_sched_entity(se) {
3057 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3058 if (cfs_rq->skip != se)
3061 cfs_rq->skip = NULL;
3065 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3067 if (cfs_rq->last == se)
3068 __clear_buddies_last(se);
3070 if (cfs_rq->next == se)
3071 __clear_buddies_next(se);
3073 if (cfs_rq->skip == se)
3074 __clear_buddies_skip(se);
3077 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3080 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3083 * Update run-time statistics of the 'current'.
3085 update_curr(cfs_rq);
3086 dequeue_entity_load_avg(cfs_rq, se);
3088 update_stats_dequeue(cfs_rq, se);
3089 if (flags & DEQUEUE_SLEEP) {
3090 #ifdef CONFIG_SCHEDSTATS
3091 if (entity_is_task(se)) {
3092 struct task_struct *tsk = task_of(se);
3094 if (tsk->state & TASK_INTERRUPTIBLE)
3095 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3096 if (tsk->state & TASK_UNINTERRUPTIBLE)
3097 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3102 clear_buddies(cfs_rq, se);
3104 if (se != cfs_rq->curr)
3105 __dequeue_entity(cfs_rq, se);
3107 account_entity_dequeue(cfs_rq, se);
3110 * Normalize the entity after updating the min_vruntime because the
3111 * update can refer to the ->curr item and we need to reflect this
3112 * movement in our normalized position.
3114 if (!(flags & DEQUEUE_SLEEP))
3115 se->vruntime -= cfs_rq->min_vruntime;
3117 /* return excess runtime on last dequeue */
3118 return_cfs_rq_runtime(cfs_rq);
3120 update_min_vruntime(cfs_rq);
3121 update_cfs_shares(cfs_rq);
3125 * Preempt the current task with a newly woken task if needed:
3128 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3130 unsigned long ideal_runtime, delta_exec;
3131 struct sched_entity *se;
3134 ideal_runtime = sched_slice(cfs_rq, curr);
3135 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3136 if (delta_exec > ideal_runtime) {
3137 resched_curr(rq_of(cfs_rq));
3139 * The current task ran long enough, ensure it doesn't get
3140 * re-elected due to buddy favours.
3142 clear_buddies(cfs_rq, curr);
3147 * Ensure that a task that missed wakeup preemption by a
3148 * narrow margin doesn't have to wait for a full slice.
3149 * This also mitigates buddy induced latencies under load.
3151 if (delta_exec < sysctl_sched_min_granularity)
3154 se = __pick_first_entity(cfs_rq);
3155 delta = curr->vruntime - se->vruntime;
3160 if (delta > ideal_runtime)
3161 resched_curr(rq_of(cfs_rq));
3165 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3167 /* 'current' is not kept within the tree. */
3170 * Any task has to be enqueued before it get to execute on
3171 * a CPU. So account for the time it spent waiting on the
3174 update_stats_wait_end(cfs_rq, se);
3175 __dequeue_entity(cfs_rq, se);
3176 update_load_avg(se, 1);
3179 update_stats_curr_start(cfs_rq, se);
3181 #ifdef CONFIG_SCHEDSTATS
3183 * Track our maximum slice length, if the CPU's load is at
3184 * least twice that of our own weight (i.e. dont track it
3185 * when there are only lesser-weight tasks around):
3187 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3188 se->statistics.slice_max = max(se->statistics.slice_max,
3189 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3192 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3196 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3199 * Pick the next process, keeping these things in mind, in this order:
3200 * 1) keep things fair between processes/task groups
3201 * 2) pick the "next" process, since someone really wants that to run
3202 * 3) pick the "last" process, for cache locality
3203 * 4) do not run the "skip" process, if something else is available
3205 static struct sched_entity *
3206 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3208 struct sched_entity *left = __pick_first_entity(cfs_rq);
3209 struct sched_entity *se;
3212 * If curr is set we have to see if its left of the leftmost entity
3213 * still in the tree, provided there was anything in the tree at all.
3215 if (!left || (curr && entity_before(curr, left)))
3218 se = left; /* ideally we run the leftmost entity */
3221 * Avoid running the skip buddy, if running something else can
3222 * be done without getting too unfair.
3224 if (cfs_rq->skip == se) {
3225 struct sched_entity *second;
3228 second = __pick_first_entity(cfs_rq);
3230 second = __pick_next_entity(se);
3231 if (!second || (curr && entity_before(curr, second)))
3235 if (second && wakeup_preempt_entity(second, left) < 1)
3240 * Prefer last buddy, try to return the CPU to a preempted task.
3242 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3246 * Someone really wants this to run. If it's not unfair, run it.
3248 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3251 clear_buddies(cfs_rq, se);
3256 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3258 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3261 * If still on the runqueue then deactivate_task()
3262 * was not called and update_curr() has to be done:
3265 update_curr(cfs_rq);
3267 /* throttle cfs_rqs exceeding runtime */
3268 check_cfs_rq_runtime(cfs_rq);
3270 check_spread(cfs_rq, prev);
3272 update_stats_wait_start(cfs_rq, prev);
3273 /* Put 'current' back into the tree. */
3274 __enqueue_entity(cfs_rq, prev);
3275 /* in !on_rq case, update occurred at dequeue */
3276 update_load_avg(prev, 0);
3278 cfs_rq->curr = NULL;
3282 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3285 * Update run-time statistics of the 'current'.
3287 update_curr(cfs_rq);
3290 * Ensure that runnable average is periodically updated.
3292 update_load_avg(curr, 1);
3293 update_cfs_shares(cfs_rq);
3295 #ifdef CONFIG_SCHED_HRTICK
3297 * queued ticks are scheduled to match the slice, so don't bother
3298 * validating it and just reschedule.
3301 resched_curr(rq_of(cfs_rq));
3305 * don't let the period tick interfere with the hrtick preemption
3307 if (!sched_feat(DOUBLE_TICK) &&
3308 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3312 if (cfs_rq->nr_running > 1)
3313 check_preempt_tick(cfs_rq, curr);
3317 /**************************************************
3318 * CFS bandwidth control machinery
3321 #ifdef CONFIG_CFS_BANDWIDTH
3323 #ifdef HAVE_JUMP_LABEL
3324 static struct static_key __cfs_bandwidth_used;
3326 static inline bool cfs_bandwidth_used(void)
3328 return static_key_false(&__cfs_bandwidth_used);
3331 void cfs_bandwidth_usage_inc(void)
3333 static_key_slow_inc(&__cfs_bandwidth_used);
3336 void cfs_bandwidth_usage_dec(void)
3338 static_key_slow_dec(&__cfs_bandwidth_used);
3340 #else /* HAVE_JUMP_LABEL */
3341 static bool cfs_bandwidth_used(void)
3346 void cfs_bandwidth_usage_inc(void) {}
3347 void cfs_bandwidth_usage_dec(void) {}
3348 #endif /* HAVE_JUMP_LABEL */
3351 * default period for cfs group bandwidth.
3352 * default: 0.1s, units: nanoseconds
3354 static inline u64 default_cfs_period(void)
3356 return 100000000ULL;
3359 static inline u64 sched_cfs_bandwidth_slice(void)
3361 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3365 * Replenish runtime according to assigned quota and update expiration time.
3366 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3367 * additional synchronization around rq->lock.
3369 * requires cfs_b->lock
3371 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3375 if (cfs_b->quota == RUNTIME_INF)
3378 now = sched_clock_cpu(smp_processor_id());
3379 cfs_b->runtime = cfs_b->quota;
3380 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3383 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3385 return &tg->cfs_bandwidth;
3388 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3389 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3391 if (unlikely(cfs_rq->throttle_count))
3392 return cfs_rq->throttled_clock_task;
3394 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3397 /* returns 0 on failure to allocate runtime */
3398 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3400 struct task_group *tg = cfs_rq->tg;
3401 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3402 u64 amount = 0, min_amount, expires;
3404 /* note: this is a positive sum as runtime_remaining <= 0 */
3405 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3407 raw_spin_lock(&cfs_b->lock);
3408 if (cfs_b->quota == RUNTIME_INF)
3409 amount = min_amount;
3411 start_cfs_bandwidth(cfs_b);
3413 if (cfs_b->runtime > 0) {
3414 amount = min(cfs_b->runtime, min_amount);
3415 cfs_b->runtime -= amount;
3419 expires = cfs_b->runtime_expires;
3420 raw_spin_unlock(&cfs_b->lock);
3422 cfs_rq->runtime_remaining += amount;
3424 * we may have advanced our local expiration to account for allowed
3425 * spread between our sched_clock and the one on which runtime was
3428 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3429 cfs_rq->runtime_expires = expires;
3431 return cfs_rq->runtime_remaining > 0;
3435 * Note: This depends on the synchronization provided by sched_clock and the
3436 * fact that rq->clock snapshots this value.
3438 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3440 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3442 /* if the deadline is ahead of our clock, nothing to do */
3443 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3446 if (cfs_rq->runtime_remaining < 0)
3450 * If the local deadline has passed we have to consider the
3451 * possibility that our sched_clock is 'fast' and the global deadline
3452 * has not truly expired.
3454 * Fortunately we can check determine whether this the case by checking
3455 * whether the global deadline has advanced. It is valid to compare
3456 * cfs_b->runtime_expires without any locks since we only care about
3457 * exact equality, so a partial write will still work.
3460 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3461 /* extend local deadline, drift is bounded above by 2 ticks */
3462 cfs_rq->runtime_expires += TICK_NSEC;
3464 /* global deadline is ahead, expiration has passed */
3465 cfs_rq->runtime_remaining = 0;
3469 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3471 /* dock delta_exec before expiring quota (as it could span periods) */
3472 cfs_rq->runtime_remaining -= delta_exec;
3473 expire_cfs_rq_runtime(cfs_rq);
3475 if (likely(cfs_rq->runtime_remaining > 0))
3479 * if we're unable to extend our runtime we resched so that the active
3480 * hierarchy can be throttled
3482 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3483 resched_curr(rq_of(cfs_rq));
3486 static __always_inline
3487 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3489 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3492 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3495 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3497 return cfs_bandwidth_used() && cfs_rq->throttled;
3500 /* check whether cfs_rq, or any parent, is throttled */
3501 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3503 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3507 * Ensure that neither of the group entities corresponding to src_cpu or
3508 * dest_cpu are members of a throttled hierarchy when performing group
3509 * load-balance operations.
3511 static inline int throttled_lb_pair(struct task_group *tg,
3512 int src_cpu, int dest_cpu)
3514 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3516 src_cfs_rq = tg->cfs_rq[src_cpu];
3517 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3519 return throttled_hierarchy(src_cfs_rq) ||
3520 throttled_hierarchy(dest_cfs_rq);
3523 /* updated child weight may affect parent so we have to do this bottom up */
3524 static int tg_unthrottle_up(struct task_group *tg, void *data)
3526 struct rq *rq = data;
3527 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3529 cfs_rq->throttle_count--;
3531 if (!cfs_rq->throttle_count) {
3532 /* adjust cfs_rq_clock_task() */
3533 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3534 cfs_rq->throttled_clock_task;
3541 static int tg_throttle_down(struct task_group *tg, void *data)
3543 struct rq *rq = data;
3544 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3546 /* group is entering throttled state, stop time */
3547 if (!cfs_rq->throttle_count)
3548 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3549 cfs_rq->throttle_count++;
3554 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3556 struct rq *rq = rq_of(cfs_rq);
3557 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3558 struct sched_entity *se;
3559 long task_delta, dequeue = 1;
3562 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3564 /* freeze hierarchy runnable averages while throttled */
3566 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3569 task_delta = cfs_rq->h_nr_running;
3570 for_each_sched_entity(se) {
3571 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3572 /* throttled entity or throttle-on-deactivate */
3577 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3578 qcfs_rq->h_nr_running -= task_delta;
3580 if (qcfs_rq->load.weight)
3585 sub_nr_running(rq, task_delta);
3587 cfs_rq->throttled = 1;
3588 cfs_rq->throttled_clock = rq_clock(rq);
3589 raw_spin_lock(&cfs_b->lock);
3590 empty = list_empty(&cfs_b->throttled_cfs_rq);
3593 * Add to the _head_ of the list, so that an already-started
3594 * distribute_cfs_runtime will not see us
3596 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3599 * If we're the first throttled task, make sure the bandwidth
3603 start_cfs_bandwidth(cfs_b);
3605 raw_spin_unlock(&cfs_b->lock);
3608 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3610 struct rq *rq = rq_of(cfs_rq);
3611 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3612 struct sched_entity *se;
3616 se = cfs_rq->tg->se[cpu_of(rq)];
3618 cfs_rq->throttled = 0;
3620 update_rq_clock(rq);
3622 raw_spin_lock(&cfs_b->lock);
3623 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3624 list_del_rcu(&cfs_rq->throttled_list);
3625 raw_spin_unlock(&cfs_b->lock);
3627 /* update hierarchical throttle state */
3628 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3630 if (!cfs_rq->load.weight)
3633 task_delta = cfs_rq->h_nr_running;
3634 for_each_sched_entity(se) {
3638 cfs_rq = cfs_rq_of(se);
3640 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3641 cfs_rq->h_nr_running += task_delta;
3643 if (cfs_rq_throttled(cfs_rq))
3648 add_nr_running(rq, task_delta);
3650 /* determine whether we need to wake up potentially idle cpu */
3651 if (rq->curr == rq->idle && rq->cfs.nr_running)
3655 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3656 u64 remaining, u64 expires)
3658 struct cfs_rq *cfs_rq;
3660 u64 starting_runtime = remaining;
3663 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3665 struct rq *rq = rq_of(cfs_rq);
3667 raw_spin_lock(&rq->lock);
3668 if (!cfs_rq_throttled(cfs_rq))
3671 runtime = -cfs_rq->runtime_remaining + 1;
3672 if (runtime > remaining)
3673 runtime = remaining;
3674 remaining -= runtime;
3676 cfs_rq->runtime_remaining += runtime;
3677 cfs_rq->runtime_expires = expires;
3679 /* we check whether we're throttled above */
3680 if (cfs_rq->runtime_remaining > 0)
3681 unthrottle_cfs_rq(cfs_rq);
3684 raw_spin_unlock(&rq->lock);
3691 return starting_runtime - remaining;
3695 * Responsible for refilling a task_group's bandwidth and unthrottling its
3696 * cfs_rqs as appropriate. If there has been no activity within the last
3697 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3698 * used to track this state.
3700 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3702 u64 runtime, runtime_expires;
3705 /* no need to continue the timer with no bandwidth constraint */
3706 if (cfs_b->quota == RUNTIME_INF)
3707 goto out_deactivate;
3709 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3710 cfs_b->nr_periods += overrun;
3713 * idle depends on !throttled (for the case of a large deficit), and if
3714 * we're going inactive then everything else can be deferred
3716 if (cfs_b->idle && !throttled)
3717 goto out_deactivate;
3719 __refill_cfs_bandwidth_runtime(cfs_b);
3722 /* mark as potentially idle for the upcoming period */
3727 /* account preceding periods in which throttling occurred */
3728 cfs_b->nr_throttled += overrun;
3730 runtime_expires = cfs_b->runtime_expires;
3733 * This check is repeated as we are holding onto the new bandwidth while
3734 * we unthrottle. This can potentially race with an unthrottled group
3735 * trying to acquire new bandwidth from the global pool. This can result
3736 * in us over-using our runtime if it is all used during this loop, but
3737 * only by limited amounts in that extreme case.
3739 while (throttled && cfs_b->runtime > 0) {
3740 runtime = cfs_b->runtime;
3741 raw_spin_unlock(&cfs_b->lock);
3742 /* we can't nest cfs_b->lock while distributing bandwidth */
3743 runtime = distribute_cfs_runtime(cfs_b, runtime,
3745 raw_spin_lock(&cfs_b->lock);
3747 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3749 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3753 * While we are ensured activity in the period following an
3754 * unthrottle, this also covers the case in which the new bandwidth is
3755 * insufficient to cover the existing bandwidth deficit. (Forcing the
3756 * timer to remain active while there are any throttled entities.)
3766 /* a cfs_rq won't donate quota below this amount */
3767 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3768 /* minimum remaining period time to redistribute slack quota */
3769 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3770 /* how long we wait to gather additional slack before distributing */
3771 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3774 * Are we near the end of the current quota period?
3776 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3777 * hrtimer base being cleared by hrtimer_start. In the case of
3778 * migrate_hrtimers, base is never cleared, so we are fine.
3780 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3782 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3785 /* if the call-back is running a quota refresh is already occurring */
3786 if (hrtimer_callback_running(refresh_timer))
3789 /* is a quota refresh about to occur? */
3790 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3791 if (remaining < min_expire)
3797 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3799 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3801 /* if there's a quota refresh soon don't bother with slack */
3802 if (runtime_refresh_within(cfs_b, min_left))
3805 hrtimer_start(&cfs_b->slack_timer,
3806 ns_to_ktime(cfs_bandwidth_slack_period),
3810 /* we know any runtime found here is valid as update_curr() precedes return */
3811 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3813 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3814 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3816 if (slack_runtime <= 0)
3819 raw_spin_lock(&cfs_b->lock);
3820 if (cfs_b->quota != RUNTIME_INF &&
3821 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3822 cfs_b->runtime += slack_runtime;
3824 /* we are under rq->lock, defer unthrottling using a timer */
3825 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3826 !list_empty(&cfs_b->throttled_cfs_rq))
3827 start_cfs_slack_bandwidth(cfs_b);
3829 raw_spin_unlock(&cfs_b->lock);
3831 /* even if it's not valid for return we don't want to try again */
3832 cfs_rq->runtime_remaining -= slack_runtime;
3835 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3837 if (!cfs_bandwidth_used())
3840 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3843 __return_cfs_rq_runtime(cfs_rq);
3847 * This is done with a timer (instead of inline with bandwidth return) since
3848 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3850 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3852 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3855 /* confirm we're still not at a refresh boundary */
3856 raw_spin_lock(&cfs_b->lock);
3857 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3858 raw_spin_unlock(&cfs_b->lock);
3862 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3863 runtime = cfs_b->runtime;
3865 expires = cfs_b->runtime_expires;
3866 raw_spin_unlock(&cfs_b->lock);
3871 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3873 raw_spin_lock(&cfs_b->lock);
3874 if (expires == cfs_b->runtime_expires)
3875 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3876 raw_spin_unlock(&cfs_b->lock);
3880 * When a group wakes up we want to make sure that its quota is not already
3881 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3882 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3884 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3886 if (!cfs_bandwidth_used())
3889 /* an active group must be handled by the update_curr()->put() path */
3890 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3893 /* ensure the group is not already throttled */
3894 if (cfs_rq_throttled(cfs_rq))
3897 /* update runtime allocation */
3898 account_cfs_rq_runtime(cfs_rq, 0);
3899 if (cfs_rq->runtime_remaining <= 0)
3900 throttle_cfs_rq(cfs_rq);
3903 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3904 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3906 if (!cfs_bandwidth_used())
3909 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3913 * it's possible for a throttled entity to be forced into a running
3914 * state (e.g. set_curr_task), in this case we're finished.
3916 if (cfs_rq_throttled(cfs_rq))
3919 throttle_cfs_rq(cfs_rq);
3923 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3925 struct cfs_bandwidth *cfs_b =
3926 container_of(timer, struct cfs_bandwidth, slack_timer);
3928 do_sched_cfs_slack_timer(cfs_b);
3930 return HRTIMER_NORESTART;
3933 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3935 struct cfs_bandwidth *cfs_b =
3936 container_of(timer, struct cfs_bandwidth, period_timer);
3940 raw_spin_lock(&cfs_b->lock);
3942 overrun = hrtimer_forward_now(timer, cfs_b->period);
3946 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3949 cfs_b->period_active = 0;
3950 raw_spin_unlock(&cfs_b->lock);
3952 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3955 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3957 raw_spin_lock_init(&cfs_b->lock);
3959 cfs_b->quota = RUNTIME_INF;
3960 cfs_b->period = ns_to_ktime(default_cfs_period());
3962 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3963 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
3964 cfs_b->period_timer.function = sched_cfs_period_timer;
3965 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3966 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3969 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3971 cfs_rq->runtime_enabled = 0;
3972 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3975 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3977 lockdep_assert_held(&cfs_b->lock);
3979 if (!cfs_b->period_active) {
3980 cfs_b->period_active = 1;
3981 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
3982 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
3986 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3988 /* init_cfs_bandwidth() was not called */
3989 if (!cfs_b->throttled_cfs_rq.next)
3992 hrtimer_cancel(&cfs_b->period_timer);
3993 hrtimer_cancel(&cfs_b->slack_timer);
3996 static void __maybe_unused update_runtime_enabled(struct rq *rq)
3998 struct cfs_rq *cfs_rq;
4000 for_each_leaf_cfs_rq(rq, cfs_rq) {
4001 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4003 raw_spin_lock(&cfs_b->lock);
4004 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4005 raw_spin_unlock(&cfs_b->lock);
4009 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4011 struct cfs_rq *cfs_rq;
4013 for_each_leaf_cfs_rq(rq, cfs_rq) {
4014 if (!cfs_rq->runtime_enabled)
4018 * clock_task is not advancing so we just need to make sure
4019 * there's some valid quota amount
4021 cfs_rq->runtime_remaining = 1;
4023 * Offline rq is schedulable till cpu is completely disabled
4024 * in take_cpu_down(), so we prevent new cfs throttling here.
4026 cfs_rq->runtime_enabled = 0;
4028 if (cfs_rq_throttled(cfs_rq))
4029 unthrottle_cfs_rq(cfs_rq);
4033 #else /* CONFIG_CFS_BANDWIDTH */
4034 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4036 return rq_clock_task(rq_of(cfs_rq));
4039 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4040 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4041 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4042 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4044 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4049 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4054 static inline int throttled_lb_pair(struct task_group *tg,
4055 int src_cpu, int dest_cpu)
4060 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4062 #ifdef CONFIG_FAIR_GROUP_SCHED
4063 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4066 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4070 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4071 static inline void update_runtime_enabled(struct rq *rq) {}
4072 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4074 #endif /* CONFIG_CFS_BANDWIDTH */
4076 /**************************************************
4077 * CFS operations on tasks:
4080 #ifdef CONFIG_SCHED_HRTICK
4081 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4083 struct sched_entity *se = &p->se;
4084 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4086 WARN_ON(task_rq(p) != rq);
4088 if (cfs_rq->nr_running > 1) {
4089 u64 slice = sched_slice(cfs_rq, se);
4090 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4091 s64 delta = slice - ran;
4098 hrtick_start(rq, delta);
4103 * called from enqueue/dequeue and updates the hrtick when the
4104 * current task is from our class and nr_running is low enough
4107 static void hrtick_update(struct rq *rq)
4109 struct task_struct *curr = rq->curr;
4111 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4114 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4115 hrtick_start_fair(rq, curr);
4117 #else /* !CONFIG_SCHED_HRTICK */
4119 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4123 static inline void hrtick_update(struct rq *rq)
4129 * The enqueue_task method is called before nr_running is
4130 * increased. Here we update the fair scheduling stats and
4131 * then put the task into the rbtree:
4134 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4136 struct cfs_rq *cfs_rq;
4137 struct sched_entity *se = &p->se;
4139 for_each_sched_entity(se) {
4142 cfs_rq = cfs_rq_of(se);
4143 enqueue_entity(cfs_rq, se, flags);
4146 * end evaluation on encountering a throttled cfs_rq
4148 * note: in the case of encountering a throttled cfs_rq we will
4149 * post the final h_nr_running increment below.
4151 if (cfs_rq_throttled(cfs_rq))
4153 cfs_rq->h_nr_running++;
4155 flags = ENQUEUE_WAKEUP;
4158 for_each_sched_entity(se) {
4159 cfs_rq = cfs_rq_of(se);
4160 cfs_rq->h_nr_running++;
4162 if (cfs_rq_throttled(cfs_rq))
4165 update_load_avg(se, 1);
4166 update_cfs_shares(cfs_rq);
4170 add_nr_running(rq, 1);
4175 static void set_next_buddy(struct sched_entity *se);
4178 * The dequeue_task method is called before nr_running is
4179 * decreased. We remove the task from the rbtree and
4180 * update the fair scheduling stats:
4182 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4184 struct cfs_rq *cfs_rq;
4185 struct sched_entity *se = &p->se;
4186 int task_sleep = flags & DEQUEUE_SLEEP;
4188 for_each_sched_entity(se) {
4189 cfs_rq = cfs_rq_of(se);
4190 dequeue_entity(cfs_rq, se, flags);
4193 * end evaluation on encountering a throttled cfs_rq
4195 * note: in the case of encountering a throttled cfs_rq we will
4196 * post the final h_nr_running decrement below.
4198 if (cfs_rq_throttled(cfs_rq))
4200 cfs_rq->h_nr_running--;
4202 /* Don't dequeue parent if it has other entities besides us */
4203 if (cfs_rq->load.weight) {
4205 * Bias pick_next to pick a task from this cfs_rq, as
4206 * p is sleeping when it is within its sched_slice.
4208 if (task_sleep && parent_entity(se))
4209 set_next_buddy(parent_entity(se));
4211 /* avoid re-evaluating load for this entity */
4212 se = parent_entity(se);
4215 flags |= DEQUEUE_SLEEP;
4218 for_each_sched_entity(se) {
4219 cfs_rq = cfs_rq_of(se);
4220 cfs_rq->h_nr_running--;
4222 if (cfs_rq_throttled(cfs_rq))
4225 update_load_avg(se, 1);
4226 update_cfs_shares(cfs_rq);
4230 sub_nr_running(rq, 1);
4238 * per rq 'load' arrray crap; XXX kill this.
4242 * The exact cpuload at various idx values, calculated at every tick would be
4243 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4245 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4246 * on nth tick when cpu may be busy, then we have:
4247 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4248 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4250 * decay_load_missed() below does efficient calculation of
4251 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4252 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4254 * The calculation is approximated on a 128 point scale.
4255 * degrade_zero_ticks is the number of ticks after which load at any
4256 * particular idx is approximated to be zero.
4257 * degrade_factor is a precomputed table, a row for each load idx.
4258 * Each column corresponds to degradation factor for a power of two ticks,
4259 * based on 128 point scale.
4261 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4262 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4264 * With this power of 2 load factors, we can degrade the load n times
4265 * by looking at 1 bits in n and doing as many mult/shift instead of
4266 * n mult/shifts needed by the exact degradation.
4268 #define DEGRADE_SHIFT 7
4269 static const unsigned char
4270 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4271 static const unsigned char
4272 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4273 {0, 0, 0, 0, 0, 0, 0, 0},
4274 {64, 32, 8, 0, 0, 0, 0, 0},
4275 {96, 72, 40, 12, 1, 0, 0},
4276 {112, 98, 75, 43, 15, 1, 0},
4277 {120, 112, 98, 76, 45, 16, 2} };
4280 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4281 * would be when CPU is idle and so we just decay the old load without
4282 * adding any new load.
4284 static unsigned long
4285 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4289 if (!missed_updates)
4292 if (missed_updates >= degrade_zero_ticks[idx])
4296 return load >> missed_updates;
4298 while (missed_updates) {
4299 if (missed_updates % 2)
4300 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4302 missed_updates >>= 1;
4309 * Update rq->cpu_load[] statistics. This function is usually called every
4310 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4311 * every tick. We fix it up based on jiffies.
4313 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4314 unsigned long pending_updates)
4318 this_rq->nr_load_updates++;
4320 /* Update our load: */
4321 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4322 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4323 unsigned long old_load, new_load;
4325 /* scale is effectively 1 << i now, and >> i divides by scale */
4327 old_load = this_rq->cpu_load[i];
4328 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4329 new_load = this_load;
4331 * Round up the averaging division if load is increasing. This
4332 * prevents us from getting stuck on 9 if the load is 10, for
4335 if (new_load > old_load)
4336 new_load += scale - 1;
4338 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4341 sched_avg_update(this_rq);
4344 /* Used instead of source_load when we know the type == 0 */
4345 static unsigned long weighted_cpuload(const int cpu)
4347 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4350 #ifdef CONFIG_NO_HZ_COMMON
4352 * There is no sane way to deal with nohz on smp when using jiffies because the
4353 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4354 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4356 * Therefore we cannot use the delta approach from the regular tick since that
4357 * would seriously skew the load calculation. However we'll make do for those
4358 * updates happening while idle (nohz_idle_balance) or coming out of idle
4359 * (tick_nohz_idle_exit).
4361 * This means we might still be one tick off for nohz periods.
4365 * Called from nohz_idle_balance() to update the load ratings before doing the
4368 static void update_idle_cpu_load(struct rq *this_rq)
4370 unsigned long curr_jiffies = READ_ONCE(jiffies);
4371 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4372 unsigned long pending_updates;
4375 * bail if there's load or we're actually up-to-date.
4377 if (load || curr_jiffies == this_rq->last_load_update_tick)
4380 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4381 this_rq->last_load_update_tick = curr_jiffies;
4383 __update_cpu_load(this_rq, load, pending_updates);
4387 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4389 void update_cpu_load_nohz(void)
4391 struct rq *this_rq = this_rq();
4392 unsigned long curr_jiffies = READ_ONCE(jiffies);
4393 unsigned long pending_updates;
4395 if (curr_jiffies == this_rq->last_load_update_tick)
4398 raw_spin_lock(&this_rq->lock);
4399 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4400 if (pending_updates) {
4401 this_rq->last_load_update_tick = curr_jiffies;
4403 * We were idle, this means load 0, the current load might be
4404 * !0 due to remote wakeups and the sort.
4406 __update_cpu_load(this_rq, 0, pending_updates);
4408 raw_spin_unlock(&this_rq->lock);
4410 #endif /* CONFIG_NO_HZ */
4413 * Called from scheduler_tick()
4415 void update_cpu_load_active(struct rq *this_rq)
4417 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4419 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4421 this_rq->last_load_update_tick = jiffies;
4422 __update_cpu_load(this_rq, load, 1);
4426 * Return a low guess at the load of a migration-source cpu weighted
4427 * according to the scheduling class and "nice" value.
4429 * We want to under-estimate the load of migration sources, to
4430 * balance conservatively.
4432 static unsigned long source_load(int cpu, int type)
4434 struct rq *rq = cpu_rq(cpu);
4435 unsigned long total = weighted_cpuload(cpu);
4437 if (type == 0 || !sched_feat(LB_BIAS))
4440 return min(rq->cpu_load[type-1], total);
4444 * Return a high guess at the load of a migration-target cpu weighted
4445 * according to the scheduling class and "nice" value.
4447 static unsigned long target_load(int cpu, int type)
4449 struct rq *rq = cpu_rq(cpu);
4450 unsigned long total = weighted_cpuload(cpu);
4452 if (type == 0 || !sched_feat(LB_BIAS))
4455 return max(rq->cpu_load[type-1], total);
4458 static unsigned long capacity_of(int cpu)
4460 return cpu_rq(cpu)->cpu_capacity;
4463 static unsigned long capacity_orig_of(int cpu)
4465 return cpu_rq(cpu)->cpu_capacity_orig;
4468 static unsigned long cpu_avg_load_per_task(int cpu)
4470 struct rq *rq = cpu_rq(cpu);
4471 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4472 unsigned long load_avg = weighted_cpuload(cpu);
4475 return load_avg / nr_running;
4480 static void record_wakee(struct task_struct *p)
4483 * Rough decay (wiping) for cost saving, don't worry
4484 * about the boundary, really active task won't care
4487 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4488 current->wakee_flips >>= 1;
4489 current->wakee_flip_decay_ts = jiffies;
4492 if (current->last_wakee != p) {
4493 current->last_wakee = p;
4494 current->wakee_flips++;
4498 static void task_waking_fair(struct task_struct *p)
4500 struct sched_entity *se = &p->se;
4501 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4504 #ifndef CONFIG_64BIT
4505 u64 min_vruntime_copy;
4508 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4510 min_vruntime = cfs_rq->min_vruntime;
4511 } while (min_vruntime != min_vruntime_copy);
4513 min_vruntime = cfs_rq->min_vruntime;
4516 se->vruntime -= min_vruntime;
4520 #ifdef CONFIG_FAIR_GROUP_SCHED
4522 * effective_load() calculates the load change as seen from the root_task_group
4524 * Adding load to a group doesn't make a group heavier, but can cause movement
4525 * of group shares between cpus. Assuming the shares were perfectly aligned one
4526 * can calculate the shift in shares.
4528 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4529 * on this @cpu and results in a total addition (subtraction) of @wg to the
4530 * total group weight.
4532 * Given a runqueue weight distribution (rw_i) we can compute a shares
4533 * distribution (s_i) using:
4535 * s_i = rw_i / \Sum rw_j (1)
4537 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4538 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4539 * shares distribution (s_i):
4541 * rw_i = { 2, 4, 1, 0 }
4542 * s_i = { 2/7, 4/7, 1/7, 0 }
4544 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4545 * task used to run on and the CPU the waker is running on), we need to
4546 * compute the effect of waking a task on either CPU and, in case of a sync
4547 * wakeup, compute the effect of the current task going to sleep.
4549 * So for a change of @wl to the local @cpu with an overall group weight change
4550 * of @wl we can compute the new shares distribution (s'_i) using:
4552 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4554 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4555 * differences in waking a task to CPU 0. The additional task changes the
4556 * weight and shares distributions like:
4558 * rw'_i = { 3, 4, 1, 0 }
4559 * s'_i = { 3/8, 4/8, 1/8, 0 }
4561 * We can then compute the difference in effective weight by using:
4563 * dw_i = S * (s'_i - s_i) (3)
4565 * Where 'S' is the group weight as seen by its parent.
4567 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4568 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4569 * 4/7) times the weight of the group.
4571 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4573 struct sched_entity *se = tg->se[cpu];
4575 if (!tg->parent) /* the trivial, non-cgroup case */
4578 for_each_sched_entity(se) {
4584 * W = @wg + \Sum rw_j
4586 W = wg + calc_tg_weight(tg, se->my_q);
4591 w = cfs_rq_load_avg(se->my_q) + wl;
4594 * wl = S * s'_i; see (2)
4597 wl = (w * (long)tg->shares) / W;
4602 * Per the above, wl is the new se->load.weight value; since
4603 * those are clipped to [MIN_SHARES, ...) do so now. See
4604 * calc_cfs_shares().
4606 if (wl < MIN_SHARES)
4610 * wl = dw_i = S * (s'_i - s_i); see (3)
4612 wl -= se->avg.load_avg;
4615 * Recursively apply this logic to all parent groups to compute
4616 * the final effective load change on the root group. Since
4617 * only the @tg group gets extra weight, all parent groups can
4618 * only redistribute existing shares. @wl is the shift in shares
4619 * resulting from this level per the above.
4628 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4636 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4637 * A waker of many should wake a different task than the one last awakened
4638 * at a frequency roughly N times higher than one of its wakees. In order
4639 * to determine whether we should let the load spread vs consolodating to
4640 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4641 * partner, and a factor of lls_size higher frequency in the other. With
4642 * both conditions met, we can be relatively sure that the relationship is
4643 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4644 * being client/server, worker/dispatcher, interrupt source or whatever is
4645 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4647 static int wake_wide(struct task_struct *p)
4649 unsigned int master = current->wakee_flips;
4650 unsigned int slave = p->wakee_flips;
4651 int factor = this_cpu_read(sd_llc_size);
4654 swap(master, slave);
4655 if (slave < factor || master < slave * factor)
4660 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4662 s64 this_load, load;
4663 s64 this_eff_load, prev_eff_load;
4664 int idx, this_cpu, prev_cpu;
4665 struct task_group *tg;
4666 unsigned long weight;
4670 this_cpu = smp_processor_id();
4671 prev_cpu = task_cpu(p);
4672 load = source_load(prev_cpu, idx);
4673 this_load = target_load(this_cpu, idx);
4676 * If sync wakeup then subtract the (maximum possible)
4677 * effect of the currently running task from the load
4678 * of the current CPU:
4681 tg = task_group(current);
4682 weight = current->se.avg.load_avg;
4684 this_load += effective_load(tg, this_cpu, -weight, -weight);
4685 load += effective_load(tg, prev_cpu, 0, -weight);
4689 weight = p->se.avg.load_avg;
4692 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4693 * due to the sync cause above having dropped this_load to 0, we'll
4694 * always have an imbalance, but there's really nothing you can do
4695 * about that, so that's good too.
4697 * Otherwise check if either cpus are near enough in load to allow this
4698 * task to be woken on this_cpu.
4700 this_eff_load = 100;
4701 this_eff_load *= capacity_of(prev_cpu);
4703 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4704 prev_eff_load *= capacity_of(this_cpu);
4706 if (this_load > 0) {
4707 this_eff_load *= this_load +
4708 effective_load(tg, this_cpu, weight, weight);
4710 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4713 balanced = this_eff_load <= prev_eff_load;
4715 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4720 schedstat_inc(sd, ttwu_move_affine);
4721 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4727 * find_idlest_group finds and returns the least busy CPU group within the
4730 static struct sched_group *
4731 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4732 int this_cpu, int sd_flag)
4734 struct sched_group *idlest = NULL, *group = sd->groups;
4735 unsigned long min_load = ULONG_MAX, this_load = 0;
4736 int load_idx = sd->forkexec_idx;
4737 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4739 if (sd_flag & SD_BALANCE_WAKE)
4740 load_idx = sd->wake_idx;
4743 unsigned long load, avg_load;
4747 /* Skip over this group if it has no CPUs allowed */
4748 if (!cpumask_intersects(sched_group_cpus(group),
4749 tsk_cpus_allowed(p)))
4752 local_group = cpumask_test_cpu(this_cpu,
4753 sched_group_cpus(group));
4755 /* Tally up the load of all CPUs in the group */
4758 for_each_cpu(i, sched_group_cpus(group)) {
4759 /* Bias balancing toward cpus of our domain */
4761 load = source_load(i, load_idx);
4763 load = target_load(i, load_idx);
4768 /* Adjust by relative CPU capacity of the group */
4769 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4772 this_load = avg_load;
4773 } else if (avg_load < min_load) {
4774 min_load = avg_load;
4777 } while (group = group->next, group != sd->groups);
4779 if (!idlest || 100*this_load < imbalance*min_load)
4785 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4788 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4790 unsigned long load, min_load = ULONG_MAX;
4791 unsigned int min_exit_latency = UINT_MAX;
4792 u64 latest_idle_timestamp = 0;
4793 int least_loaded_cpu = this_cpu;
4794 int shallowest_idle_cpu = -1;
4797 /* Traverse only the allowed CPUs */
4798 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4800 struct rq *rq = cpu_rq(i);
4801 struct cpuidle_state *idle = idle_get_state(rq);
4802 if (idle && idle->exit_latency < min_exit_latency) {
4804 * We give priority to a CPU whose idle state
4805 * has the smallest exit latency irrespective
4806 * of any idle timestamp.
4808 min_exit_latency = idle->exit_latency;
4809 latest_idle_timestamp = rq->idle_stamp;
4810 shallowest_idle_cpu = i;
4811 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4812 rq->idle_stamp > latest_idle_timestamp) {
4814 * If equal or no active idle state, then
4815 * the most recently idled CPU might have
4818 latest_idle_timestamp = rq->idle_stamp;
4819 shallowest_idle_cpu = i;
4821 } else if (shallowest_idle_cpu == -1) {
4822 load = weighted_cpuload(i);
4823 if (load < min_load || (load == min_load && i == this_cpu)) {
4825 least_loaded_cpu = i;
4830 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4834 * Try and locate an idle CPU in the sched_domain.
4836 static int select_idle_sibling(struct task_struct *p, int target)
4838 struct sched_domain *sd;
4839 struct sched_group *sg;
4840 int i = task_cpu(p);
4842 if (idle_cpu(target))
4846 * If the prevous cpu is cache affine and idle, don't be stupid.
4848 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4852 * Otherwise, iterate the domains and find an elegible idle cpu.
4854 sd = rcu_dereference(per_cpu(sd_llc, target));
4855 for_each_lower_domain(sd) {
4858 if (!cpumask_intersects(sched_group_cpus(sg),
4859 tsk_cpus_allowed(p)))
4862 for_each_cpu(i, sched_group_cpus(sg)) {
4863 if (i == target || !idle_cpu(i))
4867 target = cpumask_first_and(sched_group_cpus(sg),
4868 tsk_cpus_allowed(p));
4872 } while (sg != sd->groups);
4879 * cpu_util returns the amount of capacity of a CPU that is used by CFS
4880 * tasks. The unit of the return value must be the one of capacity so we can
4881 * compare the utilization with the capacity of the CPU that is available for
4882 * CFS task (ie cpu_capacity).
4884 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
4885 * recent utilization of currently non-runnable tasks on a CPU. It represents
4886 * the amount of utilization of a CPU in the range [0..capacity_orig] where
4887 * capacity_orig is the cpu_capacity available at the highest frequency
4888 * (arch_scale_freq_capacity()).
4889 * The utilization of a CPU converges towards a sum equal to or less than the
4890 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
4891 * the running time on this CPU scaled by capacity_curr.
4893 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
4894 * higher than capacity_orig because of unfortunate rounding in
4895 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
4896 * the average stabilizes with the new running time. We need to check that the
4897 * utilization stays within the range of [0..capacity_orig] and cap it if
4898 * necessary. Without utilization capping, a group could be seen as overloaded
4899 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
4900 * available capacity. We allow utilization to overshoot capacity_curr (but not
4901 * capacity_orig) as it useful for predicting the capacity required after task
4902 * migrations (scheduler-driven DVFS).
4904 static int cpu_util(int cpu)
4906 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
4907 unsigned long capacity = capacity_orig_of(cpu);
4909 return (util >= capacity) ? capacity : util;
4913 * select_task_rq_fair: Select target runqueue for the waking task in domains
4914 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4915 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4917 * Balances load by selecting the idlest cpu in the idlest group, or under
4918 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4920 * Returns the target cpu number.
4922 * preempt must be disabled.
4925 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4927 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4928 int cpu = smp_processor_id();
4929 int new_cpu = prev_cpu;
4930 int want_affine = 0;
4931 int sync = wake_flags & WF_SYNC;
4933 if (sd_flag & SD_BALANCE_WAKE)
4934 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4937 for_each_domain(cpu, tmp) {
4938 if (!(tmp->flags & SD_LOAD_BALANCE))
4942 * If both cpu and prev_cpu are part of this domain,
4943 * cpu is a valid SD_WAKE_AFFINE target.
4945 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4946 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4951 if (tmp->flags & sd_flag)
4953 else if (!want_affine)
4958 sd = NULL; /* Prefer wake_affine over balance flags */
4959 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4964 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
4965 new_cpu = select_idle_sibling(p, new_cpu);
4968 struct sched_group *group;
4971 if (!(sd->flags & sd_flag)) {
4976 group = find_idlest_group(sd, p, cpu, sd_flag);
4982 new_cpu = find_idlest_cpu(group, p, cpu);
4983 if (new_cpu == -1 || new_cpu == cpu) {
4984 /* Now try balancing at a lower domain level of cpu */
4989 /* Now try balancing at a lower domain level of new_cpu */
4991 weight = sd->span_weight;
4993 for_each_domain(cpu, tmp) {
4994 if (weight <= tmp->span_weight)
4996 if (tmp->flags & sd_flag)
4999 /* while loop will break here if sd == NULL */
5007 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5008 * cfs_rq_of(p) references at time of call are still valid and identify the
5009 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5010 * other assumptions, including the state of rq->lock, should be made.
5012 static void migrate_task_rq_fair(struct task_struct *p, int next_cpu)
5015 * We are supposed to update the task to "current" time, then its up to date
5016 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5017 * what current time is, so simply throw away the out-of-date time. This
5018 * will result in the wakee task is less decayed, but giving the wakee more
5019 * load sounds not bad.
5021 remove_entity_load_avg(&p->se);
5023 /* Tell new CPU we are migrated */
5024 p->se.avg.last_update_time = 0;
5026 /* We have migrated, no longer consider this task hot */
5027 p->se.exec_start = 0;
5030 static void task_dead_fair(struct task_struct *p)
5032 remove_entity_load_avg(&p->se);
5034 #endif /* CONFIG_SMP */
5036 static unsigned long
5037 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5039 unsigned long gran = sysctl_sched_wakeup_granularity;
5042 * Since its curr running now, convert the gran from real-time
5043 * to virtual-time in his units.
5045 * By using 'se' instead of 'curr' we penalize light tasks, so
5046 * they get preempted easier. That is, if 'se' < 'curr' then
5047 * the resulting gran will be larger, therefore penalizing the
5048 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5049 * be smaller, again penalizing the lighter task.
5051 * This is especially important for buddies when the leftmost
5052 * task is higher priority than the buddy.
5054 return calc_delta_fair(gran, se);
5058 * Should 'se' preempt 'curr'.
5072 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5074 s64 gran, vdiff = curr->vruntime - se->vruntime;
5079 gran = wakeup_gran(curr, se);
5086 static void set_last_buddy(struct sched_entity *se)
5088 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5091 for_each_sched_entity(se)
5092 cfs_rq_of(se)->last = se;
5095 static void set_next_buddy(struct sched_entity *se)
5097 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5100 for_each_sched_entity(se)
5101 cfs_rq_of(se)->next = se;
5104 static void set_skip_buddy(struct sched_entity *se)
5106 for_each_sched_entity(se)
5107 cfs_rq_of(se)->skip = se;
5111 * Preempt the current task with a newly woken task if needed:
5113 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5115 struct task_struct *curr = rq->curr;
5116 struct sched_entity *se = &curr->se, *pse = &p->se;
5117 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5118 int scale = cfs_rq->nr_running >= sched_nr_latency;
5119 int next_buddy_marked = 0;
5121 if (unlikely(se == pse))
5125 * This is possible from callers such as attach_tasks(), in which we
5126 * unconditionally check_prempt_curr() after an enqueue (which may have
5127 * lead to a throttle). This both saves work and prevents false
5128 * next-buddy nomination below.
5130 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5133 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5134 set_next_buddy(pse);
5135 next_buddy_marked = 1;
5139 * We can come here with TIF_NEED_RESCHED already set from new task
5142 * Note: this also catches the edge-case of curr being in a throttled
5143 * group (e.g. via set_curr_task), since update_curr() (in the
5144 * enqueue of curr) will have resulted in resched being set. This
5145 * prevents us from potentially nominating it as a false LAST_BUDDY
5148 if (test_tsk_need_resched(curr))
5151 /* Idle tasks are by definition preempted by non-idle tasks. */
5152 if (unlikely(curr->policy == SCHED_IDLE) &&
5153 likely(p->policy != SCHED_IDLE))
5157 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5158 * is driven by the tick):
5160 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5163 find_matching_se(&se, &pse);
5164 update_curr(cfs_rq_of(se));
5166 if (wakeup_preempt_entity(se, pse) == 1) {
5168 * Bias pick_next to pick the sched entity that is
5169 * triggering this preemption.
5171 if (!next_buddy_marked)
5172 set_next_buddy(pse);
5181 * Only set the backward buddy when the current task is still
5182 * on the rq. This can happen when a wakeup gets interleaved
5183 * with schedule on the ->pre_schedule() or idle_balance()
5184 * point, either of which can * drop the rq lock.
5186 * Also, during early boot the idle thread is in the fair class,
5187 * for obvious reasons its a bad idea to schedule back to it.
5189 if (unlikely(!se->on_rq || curr == rq->idle))
5192 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5196 static struct task_struct *
5197 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5199 struct cfs_rq *cfs_rq = &rq->cfs;
5200 struct sched_entity *se;
5201 struct task_struct *p;
5205 #ifdef CONFIG_FAIR_GROUP_SCHED
5206 if (!cfs_rq->nr_running)
5209 if (prev->sched_class != &fair_sched_class)
5213 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5214 * likely that a next task is from the same cgroup as the current.
5216 * Therefore attempt to avoid putting and setting the entire cgroup
5217 * hierarchy, only change the part that actually changes.
5221 struct sched_entity *curr = cfs_rq->curr;
5224 * Since we got here without doing put_prev_entity() we also
5225 * have to consider cfs_rq->curr. If it is still a runnable
5226 * entity, update_curr() will update its vruntime, otherwise
5227 * forget we've ever seen it.
5231 update_curr(cfs_rq);
5236 * This call to check_cfs_rq_runtime() will do the
5237 * throttle and dequeue its entity in the parent(s).
5238 * Therefore the 'simple' nr_running test will indeed
5241 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5245 se = pick_next_entity(cfs_rq, curr);
5246 cfs_rq = group_cfs_rq(se);
5252 * Since we haven't yet done put_prev_entity and if the selected task
5253 * is a different task than we started out with, try and touch the
5254 * least amount of cfs_rqs.
5257 struct sched_entity *pse = &prev->se;
5259 while (!(cfs_rq = is_same_group(se, pse))) {
5260 int se_depth = se->depth;
5261 int pse_depth = pse->depth;
5263 if (se_depth <= pse_depth) {
5264 put_prev_entity(cfs_rq_of(pse), pse);
5265 pse = parent_entity(pse);
5267 if (se_depth >= pse_depth) {
5268 set_next_entity(cfs_rq_of(se), se);
5269 se = parent_entity(se);
5273 put_prev_entity(cfs_rq, pse);
5274 set_next_entity(cfs_rq, se);
5277 if (hrtick_enabled(rq))
5278 hrtick_start_fair(rq, p);
5285 if (!cfs_rq->nr_running)
5288 put_prev_task(rq, prev);
5291 se = pick_next_entity(cfs_rq, NULL);
5292 set_next_entity(cfs_rq, se);
5293 cfs_rq = group_cfs_rq(se);
5298 if (hrtick_enabled(rq))
5299 hrtick_start_fair(rq, p);
5305 * This is OK, because current is on_cpu, which avoids it being picked
5306 * for load-balance and preemption/IRQs are still disabled avoiding
5307 * further scheduler activity on it and we're being very careful to
5308 * re-start the picking loop.
5310 lockdep_unpin_lock(&rq->lock);
5311 new_tasks = idle_balance(rq);
5312 lockdep_pin_lock(&rq->lock);
5314 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5315 * possible for any higher priority task to appear. In that case we
5316 * must re-start the pick_next_entity() loop.
5328 * Account for a descheduled task:
5330 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5332 struct sched_entity *se = &prev->se;
5333 struct cfs_rq *cfs_rq;
5335 for_each_sched_entity(se) {
5336 cfs_rq = cfs_rq_of(se);
5337 put_prev_entity(cfs_rq, se);
5342 * sched_yield() is very simple
5344 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5346 static void yield_task_fair(struct rq *rq)
5348 struct task_struct *curr = rq->curr;
5349 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5350 struct sched_entity *se = &curr->se;
5353 * Are we the only task in the tree?
5355 if (unlikely(rq->nr_running == 1))
5358 clear_buddies(cfs_rq, se);
5360 if (curr->policy != SCHED_BATCH) {
5361 update_rq_clock(rq);
5363 * Update run-time statistics of the 'current'.
5365 update_curr(cfs_rq);
5367 * Tell update_rq_clock() that we've just updated,
5368 * so we don't do microscopic update in schedule()
5369 * and double the fastpath cost.
5371 rq_clock_skip_update(rq, true);
5377 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5379 struct sched_entity *se = &p->se;
5381 /* throttled hierarchies are not runnable */
5382 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5385 /* Tell the scheduler that we'd really like pse to run next. */
5388 yield_task_fair(rq);
5394 /**************************************************
5395 * Fair scheduling class load-balancing methods.
5399 * The purpose of load-balancing is to achieve the same basic fairness the
5400 * per-cpu scheduler provides, namely provide a proportional amount of compute
5401 * time to each task. This is expressed in the following equation:
5403 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5405 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5406 * W_i,0 is defined as:
5408 * W_i,0 = \Sum_j w_i,j (2)
5410 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5411 * is derived from the nice value as per prio_to_weight[].
5413 * The weight average is an exponential decay average of the instantaneous
5416 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5418 * C_i is the compute capacity of cpu i, typically it is the
5419 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5420 * can also include other factors [XXX].
5422 * To achieve this balance we define a measure of imbalance which follows
5423 * directly from (1):
5425 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5427 * We them move tasks around to minimize the imbalance. In the continuous
5428 * function space it is obvious this converges, in the discrete case we get
5429 * a few fun cases generally called infeasible weight scenarios.
5432 * - infeasible weights;
5433 * - local vs global optima in the discrete case. ]
5438 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5439 * for all i,j solution, we create a tree of cpus that follows the hardware
5440 * topology where each level pairs two lower groups (or better). This results
5441 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5442 * tree to only the first of the previous level and we decrease the frequency
5443 * of load-balance at each level inv. proportional to the number of cpus in
5449 * \Sum { --- * --- * 2^i } = O(n) (5)
5451 * `- size of each group
5452 * | | `- number of cpus doing load-balance
5454 * `- sum over all levels
5456 * Coupled with a limit on how many tasks we can migrate every balance pass,
5457 * this makes (5) the runtime complexity of the balancer.
5459 * An important property here is that each CPU is still (indirectly) connected
5460 * to every other cpu in at most O(log n) steps:
5462 * The adjacency matrix of the resulting graph is given by:
5465 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5468 * And you'll find that:
5470 * A^(log_2 n)_i,j != 0 for all i,j (7)
5472 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5473 * The task movement gives a factor of O(m), giving a convergence complexity
5476 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5481 * In order to avoid CPUs going idle while there's still work to do, new idle
5482 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5483 * tree itself instead of relying on other CPUs to bring it work.
5485 * This adds some complexity to both (5) and (8) but it reduces the total idle
5493 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5496 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5501 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5503 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5505 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5508 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5509 * rewrite all of this once again.]
5512 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5514 enum fbq_type { regular, remote, all };
5516 #define LBF_ALL_PINNED 0x01
5517 #define LBF_NEED_BREAK 0x02
5518 #define LBF_DST_PINNED 0x04
5519 #define LBF_SOME_PINNED 0x08
5522 struct sched_domain *sd;
5530 struct cpumask *dst_grpmask;
5532 enum cpu_idle_type idle;
5534 /* The set of CPUs under consideration for load-balancing */
5535 struct cpumask *cpus;
5540 unsigned int loop_break;
5541 unsigned int loop_max;
5543 enum fbq_type fbq_type;
5544 struct list_head tasks;
5548 * Is this task likely cache-hot:
5550 static int task_hot(struct task_struct *p, struct lb_env *env)
5554 lockdep_assert_held(&env->src_rq->lock);
5556 if (p->sched_class != &fair_sched_class)
5559 if (unlikely(p->policy == SCHED_IDLE))
5563 * Buddy candidates are cache hot:
5565 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5566 (&p->se == cfs_rq_of(&p->se)->next ||
5567 &p->se == cfs_rq_of(&p->se)->last))
5570 if (sysctl_sched_migration_cost == -1)
5572 if (sysctl_sched_migration_cost == 0)
5575 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5577 return delta < (s64)sysctl_sched_migration_cost;
5580 #ifdef CONFIG_NUMA_BALANCING
5582 * Returns 1, if task migration degrades locality
5583 * Returns 0, if task migration improves locality i.e migration preferred.
5584 * Returns -1, if task migration is not affected by locality.
5586 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5588 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5589 unsigned long src_faults, dst_faults;
5590 int src_nid, dst_nid;
5592 if (!static_branch_likely(&sched_numa_balancing))
5595 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5598 src_nid = cpu_to_node(env->src_cpu);
5599 dst_nid = cpu_to_node(env->dst_cpu);
5601 if (src_nid == dst_nid)
5604 /* Migrating away from the preferred node is always bad. */
5605 if (src_nid == p->numa_preferred_nid) {
5606 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5612 /* Encourage migration to the preferred node. */
5613 if (dst_nid == p->numa_preferred_nid)
5617 src_faults = group_faults(p, src_nid);
5618 dst_faults = group_faults(p, dst_nid);
5620 src_faults = task_faults(p, src_nid);
5621 dst_faults = task_faults(p, dst_nid);
5624 return dst_faults < src_faults;
5628 static inline int migrate_degrades_locality(struct task_struct *p,
5636 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5639 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5643 lockdep_assert_held(&env->src_rq->lock);
5646 * We do not migrate tasks that are:
5647 * 1) throttled_lb_pair, or
5648 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5649 * 3) running (obviously), or
5650 * 4) are cache-hot on their current CPU.
5652 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5655 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5658 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5660 env->flags |= LBF_SOME_PINNED;
5663 * Remember if this task can be migrated to any other cpu in
5664 * our sched_group. We may want to revisit it if we couldn't
5665 * meet load balance goals by pulling other tasks on src_cpu.
5667 * Also avoid computing new_dst_cpu if we have already computed
5668 * one in current iteration.
5670 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5673 /* Prevent to re-select dst_cpu via env's cpus */
5674 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5675 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5676 env->flags |= LBF_DST_PINNED;
5677 env->new_dst_cpu = cpu;
5685 /* Record that we found atleast one task that could run on dst_cpu */
5686 env->flags &= ~LBF_ALL_PINNED;
5688 if (task_running(env->src_rq, p)) {
5689 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5694 * Aggressive migration if:
5695 * 1) destination numa is preferred
5696 * 2) task is cache cold, or
5697 * 3) too many balance attempts have failed.
5699 tsk_cache_hot = migrate_degrades_locality(p, env);
5700 if (tsk_cache_hot == -1)
5701 tsk_cache_hot = task_hot(p, env);
5703 if (tsk_cache_hot <= 0 ||
5704 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5705 if (tsk_cache_hot == 1) {
5706 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5707 schedstat_inc(p, se.statistics.nr_forced_migrations);
5712 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5717 * detach_task() -- detach the task for the migration specified in env
5719 static void detach_task(struct task_struct *p, struct lb_env *env)
5721 lockdep_assert_held(&env->src_rq->lock);
5723 deactivate_task(env->src_rq, p, 0);
5724 p->on_rq = TASK_ON_RQ_MIGRATING;
5725 set_task_cpu(p, env->dst_cpu);
5729 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5730 * part of active balancing operations within "domain".
5732 * Returns a task if successful and NULL otherwise.
5734 static struct task_struct *detach_one_task(struct lb_env *env)
5736 struct task_struct *p, *n;
5738 lockdep_assert_held(&env->src_rq->lock);
5740 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5741 if (!can_migrate_task(p, env))
5744 detach_task(p, env);
5747 * Right now, this is only the second place where
5748 * lb_gained[env->idle] is updated (other is detach_tasks)
5749 * so we can safely collect stats here rather than
5750 * inside detach_tasks().
5752 schedstat_inc(env->sd, lb_gained[env->idle]);
5758 static const unsigned int sched_nr_migrate_break = 32;
5761 * detach_tasks() -- tries to detach up to imbalance weighted load from
5762 * busiest_rq, as part of a balancing operation within domain "sd".
5764 * Returns number of detached tasks if successful and 0 otherwise.
5766 static int detach_tasks(struct lb_env *env)
5768 struct list_head *tasks = &env->src_rq->cfs_tasks;
5769 struct task_struct *p;
5773 lockdep_assert_held(&env->src_rq->lock);
5775 if (env->imbalance <= 0)
5778 while (!list_empty(tasks)) {
5780 * We don't want to steal all, otherwise we may be treated likewise,
5781 * which could at worst lead to a livelock crash.
5783 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5786 p = list_first_entry(tasks, struct task_struct, se.group_node);
5789 /* We've more or less seen every task there is, call it quits */
5790 if (env->loop > env->loop_max)
5793 /* take a breather every nr_migrate tasks */
5794 if (env->loop > env->loop_break) {
5795 env->loop_break += sched_nr_migrate_break;
5796 env->flags |= LBF_NEED_BREAK;
5800 if (!can_migrate_task(p, env))
5803 load = task_h_load(p);
5805 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5808 if ((load / 2) > env->imbalance)
5811 detach_task(p, env);
5812 list_add(&p->se.group_node, &env->tasks);
5815 env->imbalance -= load;
5817 #ifdef CONFIG_PREEMPT
5819 * NEWIDLE balancing is a source of latency, so preemptible
5820 * kernels will stop after the first task is detached to minimize
5821 * the critical section.
5823 if (env->idle == CPU_NEWLY_IDLE)
5828 * We only want to steal up to the prescribed amount of
5831 if (env->imbalance <= 0)
5836 list_move_tail(&p->se.group_node, tasks);
5840 * Right now, this is one of only two places we collect this stat
5841 * so we can safely collect detach_one_task() stats here rather
5842 * than inside detach_one_task().
5844 schedstat_add(env->sd, lb_gained[env->idle], detached);
5850 * attach_task() -- attach the task detached by detach_task() to its new rq.
5852 static void attach_task(struct rq *rq, struct task_struct *p)
5854 lockdep_assert_held(&rq->lock);
5856 BUG_ON(task_rq(p) != rq);
5857 p->on_rq = TASK_ON_RQ_QUEUED;
5858 activate_task(rq, p, 0);
5859 check_preempt_curr(rq, p, 0);
5863 * attach_one_task() -- attaches the task returned from detach_one_task() to
5866 static void attach_one_task(struct rq *rq, struct task_struct *p)
5868 raw_spin_lock(&rq->lock);
5870 raw_spin_unlock(&rq->lock);
5874 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5877 static void attach_tasks(struct lb_env *env)
5879 struct list_head *tasks = &env->tasks;
5880 struct task_struct *p;
5882 raw_spin_lock(&env->dst_rq->lock);
5884 while (!list_empty(tasks)) {
5885 p = list_first_entry(tasks, struct task_struct, se.group_node);
5886 list_del_init(&p->se.group_node);
5888 attach_task(env->dst_rq, p);
5891 raw_spin_unlock(&env->dst_rq->lock);
5894 #ifdef CONFIG_FAIR_GROUP_SCHED
5895 static void update_blocked_averages(int cpu)
5897 struct rq *rq = cpu_rq(cpu);
5898 struct cfs_rq *cfs_rq;
5899 unsigned long flags;
5901 raw_spin_lock_irqsave(&rq->lock, flags);
5902 update_rq_clock(rq);
5905 * Iterates the task_group tree in a bottom up fashion, see
5906 * list_add_leaf_cfs_rq() for details.
5908 for_each_leaf_cfs_rq(rq, cfs_rq) {
5909 /* throttled entities do not contribute to load */
5910 if (throttled_hierarchy(cfs_rq))
5913 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
5914 update_tg_load_avg(cfs_rq, 0);
5916 raw_spin_unlock_irqrestore(&rq->lock, flags);
5920 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5921 * This needs to be done in a top-down fashion because the load of a child
5922 * group is a fraction of its parents load.
5924 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5926 struct rq *rq = rq_of(cfs_rq);
5927 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5928 unsigned long now = jiffies;
5931 if (cfs_rq->last_h_load_update == now)
5934 cfs_rq->h_load_next = NULL;
5935 for_each_sched_entity(se) {
5936 cfs_rq = cfs_rq_of(se);
5937 cfs_rq->h_load_next = se;
5938 if (cfs_rq->last_h_load_update == now)
5943 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
5944 cfs_rq->last_h_load_update = now;
5947 while ((se = cfs_rq->h_load_next) != NULL) {
5948 load = cfs_rq->h_load;
5949 load = div64_ul(load * se->avg.load_avg,
5950 cfs_rq_load_avg(cfs_rq) + 1);
5951 cfs_rq = group_cfs_rq(se);
5952 cfs_rq->h_load = load;
5953 cfs_rq->last_h_load_update = now;
5957 static unsigned long task_h_load(struct task_struct *p)
5959 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5961 update_cfs_rq_h_load(cfs_rq);
5962 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
5963 cfs_rq_load_avg(cfs_rq) + 1);
5966 static inline void update_blocked_averages(int cpu)
5968 struct rq *rq = cpu_rq(cpu);
5969 struct cfs_rq *cfs_rq = &rq->cfs;
5970 unsigned long flags;
5972 raw_spin_lock_irqsave(&rq->lock, flags);
5973 update_rq_clock(rq);
5974 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
5975 raw_spin_unlock_irqrestore(&rq->lock, flags);
5978 static unsigned long task_h_load(struct task_struct *p)
5980 return p->se.avg.load_avg;
5984 /********** Helpers for find_busiest_group ************************/
5993 * sg_lb_stats - stats of a sched_group required for load_balancing
5995 struct sg_lb_stats {
5996 unsigned long avg_load; /*Avg load across the CPUs of the group */
5997 unsigned long group_load; /* Total load over the CPUs of the group */
5998 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5999 unsigned long load_per_task;
6000 unsigned long group_capacity;
6001 unsigned long group_util; /* Total utilization of the group */
6002 unsigned int sum_nr_running; /* Nr tasks running in the group */
6003 unsigned int idle_cpus;
6004 unsigned int group_weight;
6005 enum group_type group_type;
6006 int group_no_capacity;
6007 #ifdef CONFIG_NUMA_BALANCING
6008 unsigned int nr_numa_running;
6009 unsigned int nr_preferred_running;
6014 * sd_lb_stats - Structure to store the statistics of a sched_domain
6015 * during load balancing.
6017 struct sd_lb_stats {
6018 struct sched_group *busiest; /* Busiest group in this sd */
6019 struct sched_group *local; /* Local group in this sd */
6020 unsigned long total_load; /* Total load of all groups in sd */
6021 unsigned long total_capacity; /* Total capacity of all groups in sd */
6022 unsigned long avg_load; /* Average load across all groups in sd */
6024 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6025 struct sg_lb_stats local_stat; /* Statistics of the local group */
6028 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6031 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6032 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6033 * We must however clear busiest_stat::avg_load because
6034 * update_sd_pick_busiest() reads this before assignment.
6036 *sds = (struct sd_lb_stats){
6040 .total_capacity = 0UL,
6043 .sum_nr_running = 0,
6044 .group_type = group_other,
6050 * get_sd_load_idx - Obtain the load index for a given sched domain.
6051 * @sd: The sched_domain whose load_idx is to be obtained.
6052 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6054 * Return: The load index.
6056 static inline int get_sd_load_idx(struct sched_domain *sd,
6057 enum cpu_idle_type idle)
6063 load_idx = sd->busy_idx;
6066 case CPU_NEWLY_IDLE:
6067 load_idx = sd->newidle_idx;
6070 load_idx = sd->idle_idx;
6077 static unsigned long scale_rt_capacity(int cpu)
6079 struct rq *rq = cpu_rq(cpu);
6080 u64 total, used, age_stamp, avg;
6084 * Since we're reading these variables without serialization make sure
6085 * we read them once before doing sanity checks on them.
6087 age_stamp = READ_ONCE(rq->age_stamp);
6088 avg = READ_ONCE(rq->rt_avg);
6089 delta = __rq_clock_broken(rq) - age_stamp;
6091 if (unlikely(delta < 0))
6094 total = sched_avg_period() + delta;
6096 used = div_u64(avg, total);
6098 if (likely(used < SCHED_CAPACITY_SCALE))
6099 return SCHED_CAPACITY_SCALE - used;
6104 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6106 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6107 struct sched_group *sdg = sd->groups;
6109 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6111 capacity *= scale_rt_capacity(cpu);
6112 capacity >>= SCHED_CAPACITY_SHIFT;
6117 cpu_rq(cpu)->cpu_capacity = capacity;
6118 sdg->sgc->capacity = capacity;
6121 void update_group_capacity(struct sched_domain *sd, int cpu)
6123 struct sched_domain *child = sd->child;
6124 struct sched_group *group, *sdg = sd->groups;
6125 unsigned long capacity;
6126 unsigned long interval;
6128 interval = msecs_to_jiffies(sd->balance_interval);
6129 interval = clamp(interval, 1UL, max_load_balance_interval);
6130 sdg->sgc->next_update = jiffies + interval;
6133 update_cpu_capacity(sd, cpu);
6139 if (child->flags & SD_OVERLAP) {
6141 * SD_OVERLAP domains cannot assume that child groups
6142 * span the current group.
6145 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6146 struct sched_group_capacity *sgc;
6147 struct rq *rq = cpu_rq(cpu);
6150 * build_sched_domains() -> init_sched_groups_capacity()
6151 * gets here before we've attached the domains to the
6154 * Use capacity_of(), which is set irrespective of domains
6155 * in update_cpu_capacity().
6157 * This avoids capacity from being 0 and
6158 * causing divide-by-zero issues on boot.
6160 if (unlikely(!rq->sd)) {
6161 capacity += capacity_of(cpu);
6165 sgc = rq->sd->groups->sgc;
6166 capacity += sgc->capacity;
6170 * !SD_OVERLAP domains can assume that child groups
6171 * span the current group.
6174 group = child->groups;
6176 capacity += group->sgc->capacity;
6177 group = group->next;
6178 } while (group != child->groups);
6181 sdg->sgc->capacity = capacity;
6185 * Check whether the capacity of the rq has been noticeably reduced by side
6186 * activity. The imbalance_pct is used for the threshold.
6187 * Return true is the capacity is reduced
6190 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6192 return ((rq->cpu_capacity * sd->imbalance_pct) <
6193 (rq->cpu_capacity_orig * 100));
6197 * Group imbalance indicates (and tries to solve) the problem where balancing
6198 * groups is inadequate due to tsk_cpus_allowed() constraints.
6200 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6201 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6204 * { 0 1 2 3 } { 4 5 6 7 }
6207 * If we were to balance group-wise we'd place two tasks in the first group and
6208 * two tasks in the second group. Clearly this is undesired as it will overload
6209 * cpu 3 and leave one of the cpus in the second group unused.
6211 * The current solution to this issue is detecting the skew in the first group
6212 * by noticing the lower domain failed to reach balance and had difficulty
6213 * moving tasks due to affinity constraints.
6215 * When this is so detected; this group becomes a candidate for busiest; see
6216 * update_sd_pick_busiest(). And calculate_imbalance() and
6217 * find_busiest_group() avoid some of the usual balance conditions to allow it
6218 * to create an effective group imbalance.
6220 * This is a somewhat tricky proposition since the next run might not find the
6221 * group imbalance and decide the groups need to be balanced again. A most
6222 * subtle and fragile situation.
6225 static inline int sg_imbalanced(struct sched_group *group)
6227 return group->sgc->imbalance;
6231 * group_has_capacity returns true if the group has spare capacity that could
6232 * be used by some tasks.
6233 * We consider that a group has spare capacity if the * number of task is
6234 * smaller than the number of CPUs or if the utilization is lower than the
6235 * available capacity for CFS tasks.
6236 * For the latter, we use a threshold to stabilize the state, to take into
6237 * account the variance of the tasks' load and to return true if the available
6238 * capacity in meaningful for the load balancer.
6239 * As an example, an available capacity of 1% can appear but it doesn't make
6240 * any benefit for the load balance.
6243 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6245 if (sgs->sum_nr_running < sgs->group_weight)
6248 if ((sgs->group_capacity * 100) >
6249 (sgs->group_util * env->sd->imbalance_pct))
6256 * group_is_overloaded returns true if the group has more tasks than it can
6258 * group_is_overloaded is not equals to !group_has_capacity because a group
6259 * with the exact right number of tasks, has no more spare capacity but is not
6260 * overloaded so both group_has_capacity and group_is_overloaded return
6264 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6266 if (sgs->sum_nr_running <= sgs->group_weight)
6269 if ((sgs->group_capacity * 100) <
6270 (sgs->group_util * env->sd->imbalance_pct))
6277 group_type group_classify(struct sched_group *group,
6278 struct sg_lb_stats *sgs)
6280 if (sgs->group_no_capacity)
6281 return group_overloaded;
6283 if (sg_imbalanced(group))
6284 return group_imbalanced;
6290 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6291 * @env: The load balancing environment.
6292 * @group: sched_group whose statistics are to be updated.
6293 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6294 * @local_group: Does group contain this_cpu.
6295 * @sgs: variable to hold the statistics for this group.
6296 * @overload: Indicate more than one runnable task for any CPU.
6298 static inline void update_sg_lb_stats(struct lb_env *env,
6299 struct sched_group *group, int load_idx,
6300 int local_group, struct sg_lb_stats *sgs,
6306 memset(sgs, 0, sizeof(*sgs));
6308 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6309 struct rq *rq = cpu_rq(i);
6311 /* Bias balancing toward cpus of our domain */
6313 load = target_load(i, load_idx);
6315 load = source_load(i, load_idx);
6317 sgs->group_load += load;
6318 sgs->group_util += cpu_util(i);
6319 sgs->sum_nr_running += rq->cfs.h_nr_running;
6321 if (rq->nr_running > 1)
6324 #ifdef CONFIG_NUMA_BALANCING
6325 sgs->nr_numa_running += rq->nr_numa_running;
6326 sgs->nr_preferred_running += rq->nr_preferred_running;
6328 sgs->sum_weighted_load += weighted_cpuload(i);
6333 /* Adjust by relative CPU capacity of the group */
6334 sgs->group_capacity = group->sgc->capacity;
6335 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6337 if (sgs->sum_nr_running)
6338 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6340 sgs->group_weight = group->group_weight;
6342 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6343 sgs->group_type = group_classify(group, sgs);
6347 * update_sd_pick_busiest - return 1 on busiest group
6348 * @env: The load balancing environment.
6349 * @sds: sched_domain statistics
6350 * @sg: sched_group candidate to be checked for being the busiest
6351 * @sgs: sched_group statistics
6353 * Determine if @sg is a busier group than the previously selected
6356 * Return: %true if @sg is a busier group than the previously selected
6357 * busiest group. %false otherwise.
6359 static bool update_sd_pick_busiest(struct lb_env *env,
6360 struct sd_lb_stats *sds,
6361 struct sched_group *sg,
6362 struct sg_lb_stats *sgs)
6364 struct sg_lb_stats *busiest = &sds->busiest_stat;
6366 if (sgs->group_type > busiest->group_type)
6369 if (sgs->group_type < busiest->group_type)
6372 if (sgs->avg_load <= busiest->avg_load)
6375 /* This is the busiest node in its class. */
6376 if (!(env->sd->flags & SD_ASYM_PACKING))
6380 * ASYM_PACKING needs to move all the work to the lowest
6381 * numbered CPUs in the group, therefore mark all groups
6382 * higher than ourself as busy.
6384 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6388 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6395 #ifdef CONFIG_NUMA_BALANCING
6396 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6398 if (sgs->sum_nr_running > sgs->nr_numa_running)
6400 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6405 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6407 if (rq->nr_running > rq->nr_numa_running)
6409 if (rq->nr_running > rq->nr_preferred_running)
6414 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6419 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6423 #endif /* CONFIG_NUMA_BALANCING */
6426 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6427 * @env: The load balancing environment.
6428 * @sds: variable to hold the statistics for this sched_domain.
6430 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6432 struct sched_domain *child = env->sd->child;
6433 struct sched_group *sg = env->sd->groups;
6434 struct sg_lb_stats tmp_sgs;
6435 int load_idx, prefer_sibling = 0;
6436 bool overload = false;
6438 if (child && child->flags & SD_PREFER_SIBLING)
6441 load_idx = get_sd_load_idx(env->sd, env->idle);
6444 struct sg_lb_stats *sgs = &tmp_sgs;
6447 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6450 sgs = &sds->local_stat;
6452 if (env->idle != CPU_NEWLY_IDLE ||
6453 time_after_eq(jiffies, sg->sgc->next_update))
6454 update_group_capacity(env->sd, env->dst_cpu);
6457 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6464 * In case the child domain prefers tasks go to siblings
6465 * first, lower the sg capacity so that we'll try
6466 * and move all the excess tasks away. We lower the capacity
6467 * of a group only if the local group has the capacity to fit
6468 * these excess tasks. The extra check prevents the case where
6469 * you always pull from the heaviest group when it is already
6470 * under-utilized (possible with a large weight task outweighs
6471 * the tasks on the system).
6473 if (prefer_sibling && sds->local &&
6474 group_has_capacity(env, &sds->local_stat) &&
6475 (sgs->sum_nr_running > 1)) {
6476 sgs->group_no_capacity = 1;
6477 sgs->group_type = group_classify(sg, sgs);
6480 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6482 sds->busiest_stat = *sgs;
6486 /* Now, start updating sd_lb_stats */
6487 sds->total_load += sgs->group_load;
6488 sds->total_capacity += sgs->group_capacity;
6491 } while (sg != env->sd->groups);
6493 if (env->sd->flags & SD_NUMA)
6494 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6496 if (!env->sd->parent) {
6497 /* update overload indicator if we are at root domain */
6498 if (env->dst_rq->rd->overload != overload)
6499 env->dst_rq->rd->overload = overload;
6505 * check_asym_packing - Check to see if the group is packed into the
6508 * This is primarily intended to used at the sibling level. Some
6509 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6510 * case of POWER7, it can move to lower SMT modes only when higher
6511 * threads are idle. When in lower SMT modes, the threads will
6512 * perform better since they share less core resources. Hence when we
6513 * have idle threads, we want them to be the higher ones.
6515 * This packing function is run on idle threads. It checks to see if
6516 * the busiest CPU in this domain (core in the P7 case) has a higher
6517 * CPU number than the packing function is being run on. Here we are
6518 * assuming lower CPU number will be equivalent to lower a SMT thread
6521 * Return: 1 when packing is required and a task should be moved to
6522 * this CPU. The amount of the imbalance is returned in *imbalance.
6524 * @env: The load balancing environment.
6525 * @sds: Statistics of the sched_domain which is to be packed
6527 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6531 if (!(env->sd->flags & SD_ASYM_PACKING))
6537 busiest_cpu = group_first_cpu(sds->busiest);
6538 if (env->dst_cpu > busiest_cpu)
6541 env->imbalance = DIV_ROUND_CLOSEST(
6542 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6543 SCHED_CAPACITY_SCALE);
6549 * fix_small_imbalance - Calculate the minor imbalance that exists
6550 * amongst the groups of a sched_domain, during
6552 * @env: The load balancing environment.
6553 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6556 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6558 unsigned long tmp, capa_now = 0, capa_move = 0;
6559 unsigned int imbn = 2;
6560 unsigned long scaled_busy_load_per_task;
6561 struct sg_lb_stats *local, *busiest;
6563 local = &sds->local_stat;
6564 busiest = &sds->busiest_stat;
6566 if (!local->sum_nr_running)
6567 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6568 else if (busiest->load_per_task > local->load_per_task)
6571 scaled_busy_load_per_task =
6572 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6573 busiest->group_capacity;
6575 if (busiest->avg_load + scaled_busy_load_per_task >=
6576 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6577 env->imbalance = busiest->load_per_task;
6582 * OK, we don't have enough imbalance to justify moving tasks,
6583 * however we may be able to increase total CPU capacity used by
6587 capa_now += busiest->group_capacity *
6588 min(busiest->load_per_task, busiest->avg_load);
6589 capa_now += local->group_capacity *
6590 min(local->load_per_task, local->avg_load);
6591 capa_now /= SCHED_CAPACITY_SCALE;
6593 /* Amount of load we'd subtract */
6594 if (busiest->avg_load > scaled_busy_load_per_task) {
6595 capa_move += busiest->group_capacity *
6596 min(busiest->load_per_task,
6597 busiest->avg_load - scaled_busy_load_per_task);
6600 /* Amount of load we'd add */
6601 if (busiest->avg_load * busiest->group_capacity <
6602 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6603 tmp = (busiest->avg_load * busiest->group_capacity) /
6604 local->group_capacity;
6606 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6607 local->group_capacity;
6609 capa_move += local->group_capacity *
6610 min(local->load_per_task, local->avg_load + tmp);
6611 capa_move /= SCHED_CAPACITY_SCALE;
6613 /* Move if we gain throughput */
6614 if (capa_move > capa_now)
6615 env->imbalance = busiest->load_per_task;
6619 * calculate_imbalance - Calculate the amount of imbalance present within the
6620 * groups of a given sched_domain during load balance.
6621 * @env: load balance environment
6622 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6624 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6626 unsigned long max_pull, load_above_capacity = ~0UL;
6627 struct sg_lb_stats *local, *busiest;
6629 local = &sds->local_stat;
6630 busiest = &sds->busiest_stat;
6632 if (busiest->group_type == group_imbalanced) {
6634 * In the group_imb case we cannot rely on group-wide averages
6635 * to ensure cpu-load equilibrium, look at wider averages. XXX
6637 busiest->load_per_task =
6638 min(busiest->load_per_task, sds->avg_load);
6642 * In the presence of smp nice balancing, certain scenarios can have
6643 * max load less than avg load(as we skip the groups at or below
6644 * its cpu_capacity, while calculating max_load..)
6646 if (busiest->avg_load <= sds->avg_load ||
6647 local->avg_load >= sds->avg_load) {
6649 return fix_small_imbalance(env, sds);
6653 * If there aren't any idle cpus, avoid creating some.
6655 if (busiest->group_type == group_overloaded &&
6656 local->group_type == group_overloaded) {
6657 load_above_capacity = busiest->sum_nr_running *
6659 if (load_above_capacity > busiest->group_capacity)
6660 load_above_capacity -= busiest->group_capacity;
6662 load_above_capacity = ~0UL;
6666 * We're trying to get all the cpus to the average_load, so we don't
6667 * want to push ourselves above the average load, nor do we wish to
6668 * reduce the max loaded cpu below the average load. At the same time,
6669 * we also don't want to reduce the group load below the group capacity
6670 * (so that we can implement power-savings policies etc). Thus we look
6671 * for the minimum possible imbalance.
6673 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6675 /* How much load to actually move to equalise the imbalance */
6676 env->imbalance = min(
6677 max_pull * busiest->group_capacity,
6678 (sds->avg_load - local->avg_load) * local->group_capacity
6679 ) / SCHED_CAPACITY_SCALE;
6682 * if *imbalance is less than the average load per runnable task
6683 * there is no guarantee that any tasks will be moved so we'll have
6684 * a think about bumping its value to force at least one task to be
6687 if (env->imbalance < busiest->load_per_task)
6688 return fix_small_imbalance(env, sds);
6691 /******* find_busiest_group() helpers end here *********************/
6694 * find_busiest_group - Returns the busiest group within the sched_domain
6695 * if there is an imbalance. If there isn't an imbalance, and
6696 * the user has opted for power-savings, it returns a group whose
6697 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6698 * such a group exists.
6700 * Also calculates the amount of weighted load which should be moved
6701 * to restore balance.
6703 * @env: The load balancing environment.
6705 * Return: - The busiest group if imbalance exists.
6706 * - If no imbalance and user has opted for power-savings balance,
6707 * return the least loaded group whose CPUs can be
6708 * put to idle by rebalancing its tasks onto our group.
6710 static struct sched_group *find_busiest_group(struct lb_env *env)
6712 struct sg_lb_stats *local, *busiest;
6713 struct sd_lb_stats sds;
6715 init_sd_lb_stats(&sds);
6718 * Compute the various statistics relavent for load balancing at
6721 update_sd_lb_stats(env, &sds);
6722 local = &sds.local_stat;
6723 busiest = &sds.busiest_stat;
6725 /* ASYM feature bypasses nice load balance check */
6726 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6727 check_asym_packing(env, &sds))
6730 /* There is no busy sibling group to pull tasks from */
6731 if (!sds.busiest || busiest->sum_nr_running == 0)
6734 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6735 / sds.total_capacity;
6738 * If the busiest group is imbalanced the below checks don't
6739 * work because they assume all things are equal, which typically
6740 * isn't true due to cpus_allowed constraints and the like.
6742 if (busiest->group_type == group_imbalanced)
6745 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6746 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6747 busiest->group_no_capacity)
6751 * If the local group is busier than the selected busiest group
6752 * don't try and pull any tasks.
6754 if (local->avg_load >= busiest->avg_load)
6758 * Don't pull any tasks if this group is already above the domain
6761 if (local->avg_load >= sds.avg_load)
6764 if (env->idle == CPU_IDLE) {
6766 * This cpu is idle. If the busiest group is not overloaded
6767 * and there is no imbalance between this and busiest group
6768 * wrt idle cpus, it is balanced. The imbalance becomes
6769 * significant if the diff is greater than 1 otherwise we
6770 * might end up to just move the imbalance on another group
6772 if ((busiest->group_type != group_overloaded) &&
6773 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6777 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6778 * imbalance_pct to be conservative.
6780 if (100 * busiest->avg_load <=
6781 env->sd->imbalance_pct * local->avg_load)
6786 /* Looks like there is an imbalance. Compute it */
6787 calculate_imbalance(env, &sds);
6796 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6798 static struct rq *find_busiest_queue(struct lb_env *env,
6799 struct sched_group *group)
6801 struct rq *busiest = NULL, *rq;
6802 unsigned long busiest_load = 0, busiest_capacity = 1;
6805 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6806 unsigned long capacity, wl;
6810 rt = fbq_classify_rq(rq);
6813 * We classify groups/runqueues into three groups:
6814 * - regular: there are !numa tasks
6815 * - remote: there are numa tasks that run on the 'wrong' node
6816 * - all: there is no distinction
6818 * In order to avoid migrating ideally placed numa tasks,
6819 * ignore those when there's better options.
6821 * If we ignore the actual busiest queue to migrate another
6822 * task, the next balance pass can still reduce the busiest
6823 * queue by moving tasks around inside the node.
6825 * If we cannot move enough load due to this classification
6826 * the next pass will adjust the group classification and
6827 * allow migration of more tasks.
6829 * Both cases only affect the total convergence complexity.
6831 if (rt > env->fbq_type)
6834 capacity = capacity_of(i);
6836 wl = weighted_cpuload(i);
6839 * When comparing with imbalance, use weighted_cpuload()
6840 * which is not scaled with the cpu capacity.
6843 if (rq->nr_running == 1 && wl > env->imbalance &&
6844 !check_cpu_capacity(rq, env->sd))
6848 * For the load comparisons with the other cpu's, consider
6849 * the weighted_cpuload() scaled with the cpu capacity, so
6850 * that the load can be moved away from the cpu that is
6851 * potentially running at a lower capacity.
6853 * Thus we're looking for max(wl_i / capacity_i), crosswise
6854 * multiplication to rid ourselves of the division works out
6855 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6856 * our previous maximum.
6858 if (wl * busiest_capacity > busiest_load * capacity) {
6860 busiest_capacity = capacity;
6869 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6870 * so long as it is large enough.
6872 #define MAX_PINNED_INTERVAL 512
6874 /* Working cpumask for load_balance and load_balance_newidle. */
6875 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6877 static int need_active_balance(struct lb_env *env)
6879 struct sched_domain *sd = env->sd;
6881 if (env->idle == CPU_NEWLY_IDLE) {
6884 * ASYM_PACKING needs to force migrate tasks from busy but
6885 * higher numbered CPUs in order to pack all tasks in the
6886 * lowest numbered CPUs.
6888 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6893 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
6894 * It's worth migrating the task if the src_cpu's capacity is reduced
6895 * because of other sched_class or IRQs if more capacity stays
6896 * available on dst_cpu.
6898 if ((env->idle != CPU_NOT_IDLE) &&
6899 (env->src_rq->cfs.h_nr_running == 1)) {
6900 if ((check_cpu_capacity(env->src_rq, sd)) &&
6901 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
6905 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6908 static int active_load_balance_cpu_stop(void *data);
6910 static int should_we_balance(struct lb_env *env)
6912 struct sched_group *sg = env->sd->groups;
6913 struct cpumask *sg_cpus, *sg_mask;
6914 int cpu, balance_cpu = -1;
6917 * In the newly idle case, we will allow all the cpu's
6918 * to do the newly idle load balance.
6920 if (env->idle == CPU_NEWLY_IDLE)
6923 sg_cpus = sched_group_cpus(sg);
6924 sg_mask = sched_group_mask(sg);
6925 /* Try to find first idle cpu */
6926 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6927 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6934 if (balance_cpu == -1)
6935 balance_cpu = group_balance_cpu(sg);
6938 * First idle cpu or the first cpu(busiest) in this sched group
6939 * is eligible for doing load balancing at this and above domains.
6941 return balance_cpu == env->dst_cpu;
6945 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6946 * tasks if there is an imbalance.
6948 static int load_balance(int this_cpu, struct rq *this_rq,
6949 struct sched_domain *sd, enum cpu_idle_type idle,
6950 int *continue_balancing)
6952 int ld_moved, cur_ld_moved, active_balance = 0;
6953 struct sched_domain *sd_parent = sd->parent;
6954 struct sched_group *group;
6956 unsigned long flags;
6957 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
6959 struct lb_env env = {
6961 .dst_cpu = this_cpu,
6963 .dst_grpmask = sched_group_cpus(sd->groups),
6965 .loop_break = sched_nr_migrate_break,
6968 .tasks = LIST_HEAD_INIT(env.tasks),
6972 * For NEWLY_IDLE load_balancing, we don't need to consider
6973 * other cpus in our group
6975 if (idle == CPU_NEWLY_IDLE)
6976 env.dst_grpmask = NULL;
6978 cpumask_copy(cpus, cpu_active_mask);
6980 schedstat_inc(sd, lb_count[idle]);
6983 if (!should_we_balance(&env)) {
6984 *continue_balancing = 0;
6988 group = find_busiest_group(&env);
6990 schedstat_inc(sd, lb_nobusyg[idle]);
6994 busiest = find_busiest_queue(&env, group);
6996 schedstat_inc(sd, lb_nobusyq[idle]);
7000 BUG_ON(busiest == env.dst_rq);
7002 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7004 env.src_cpu = busiest->cpu;
7005 env.src_rq = busiest;
7008 if (busiest->nr_running > 1) {
7010 * Attempt to move tasks. If find_busiest_group has found
7011 * an imbalance but busiest->nr_running <= 1, the group is
7012 * still unbalanced. ld_moved simply stays zero, so it is
7013 * correctly treated as an imbalance.
7015 env.flags |= LBF_ALL_PINNED;
7016 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7019 raw_spin_lock_irqsave(&busiest->lock, flags);
7022 * cur_ld_moved - load moved in current iteration
7023 * ld_moved - cumulative load moved across iterations
7025 cur_ld_moved = detach_tasks(&env);
7028 * We've detached some tasks from busiest_rq. Every
7029 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7030 * unlock busiest->lock, and we are able to be sure
7031 * that nobody can manipulate the tasks in parallel.
7032 * See task_rq_lock() family for the details.
7035 raw_spin_unlock(&busiest->lock);
7039 ld_moved += cur_ld_moved;
7042 local_irq_restore(flags);
7044 if (env.flags & LBF_NEED_BREAK) {
7045 env.flags &= ~LBF_NEED_BREAK;
7050 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7051 * us and move them to an alternate dst_cpu in our sched_group
7052 * where they can run. The upper limit on how many times we
7053 * iterate on same src_cpu is dependent on number of cpus in our
7056 * This changes load balance semantics a bit on who can move
7057 * load to a given_cpu. In addition to the given_cpu itself
7058 * (or a ilb_cpu acting on its behalf where given_cpu is
7059 * nohz-idle), we now have balance_cpu in a position to move
7060 * load to given_cpu. In rare situations, this may cause
7061 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7062 * _independently_ and at _same_ time to move some load to
7063 * given_cpu) causing exceess load to be moved to given_cpu.
7064 * This however should not happen so much in practice and
7065 * moreover subsequent load balance cycles should correct the
7066 * excess load moved.
7068 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7070 /* Prevent to re-select dst_cpu via env's cpus */
7071 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7073 env.dst_rq = cpu_rq(env.new_dst_cpu);
7074 env.dst_cpu = env.new_dst_cpu;
7075 env.flags &= ~LBF_DST_PINNED;
7077 env.loop_break = sched_nr_migrate_break;
7080 * Go back to "more_balance" rather than "redo" since we
7081 * need to continue with same src_cpu.
7087 * We failed to reach balance because of affinity.
7090 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7092 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7093 *group_imbalance = 1;
7096 /* All tasks on this runqueue were pinned by CPU affinity */
7097 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7098 cpumask_clear_cpu(cpu_of(busiest), cpus);
7099 if (!cpumask_empty(cpus)) {
7101 env.loop_break = sched_nr_migrate_break;
7104 goto out_all_pinned;
7109 schedstat_inc(sd, lb_failed[idle]);
7111 * Increment the failure counter only on periodic balance.
7112 * We do not want newidle balance, which can be very
7113 * frequent, pollute the failure counter causing
7114 * excessive cache_hot migrations and active balances.
7116 if (idle != CPU_NEWLY_IDLE)
7117 sd->nr_balance_failed++;
7119 if (need_active_balance(&env)) {
7120 raw_spin_lock_irqsave(&busiest->lock, flags);
7122 /* don't kick the active_load_balance_cpu_stop,
7123 * if the curr task on busiest cpu can't be
7126 if (!cpumask_test_cpu(this_cpu,
7127 tsk_cpus_allowed(busiest->curr))) {
7128 raw_spin_unlock_irqrestore(&busiest->lock,
7130 env.flags |= LBF_ALL_PINNED;
7131 goto out_one_pinned;
7135 * ->active_balance synchronizes accesses to
7136 * ->active_balance_work. Once set, it's cleared
7137 * only after active load balance is finished.
7139 if (!busiest->active_balance) {
7140 busiest->active_balance = 1;
7141 busiest->push_cpu = this_cpu;
7144 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7146 if (active_balance) {
7147 stop_one_cpu_nowait(cpu_of(busiest),
7148 active_load_balance_cpu_stop, busiest,
7149 &busiest->active_balance_work);
7153 * We've kicked active balancing, reset the failure
7156 sd->nr_balance_failed = sd->cache_nice_tries+1;
7159 sd->nr_balance_failed = 0;
7161 if (likely(!active_balance)) {
7162 /* We were unbalanced, so reset the balancing interval */
7163 sd->balance_interval = sd->min_interval;
7166 * If we've begun active balancing, start to back off. This
7167 * case may not be covered by the all_pinned logic if there
7168 * is only 1 task on the busy runqueue (because we don't call
7171 if (sd->balance_interval < sd->max_interval)
7172 sd->balance_interval *= 2;
7179 * We reach balance although we may have faced some affinity
7180 * constraints. Clear the imbalance flag if it was set.
7183 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7185 if (*group_imbalance)
7186 *group_imbalance = 0;
7191 * We reach balance because all tasks are pinned at this level so
7192 * we can't migrate them. Let the imbalance flag set so parent level
7193 * can try to migrate them.
7195 schedstat_inc(sd, lb_balanced[idle]);
7197 sd->nr_balance_failed = 0;
7200 /* tune up the balancing interval */
7201 if (((env.flags & LBF_ALL_PINNED) &&
7202 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7203 (sd->balance_interval < sd->max_interval))
7204 sd->balance_interval *= 2;
7211 static inline unsigned long
7212 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7214 unsigned long interval = sd->balance_interval;
7217 interval *= sd->busy_factor;
7219 /* scale ms to jiffies */
7220 interval = msecs_to_jiffies(interval);
7221 interval = clamp(interval, 1UL, max_load_balance_interval);
7227 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7229 unsigned long interval, next;
7231 interval = get_sd_balance_interval(sd, cpu_busy);
7232 next = sd->last_balance + interval;
7234 if (time_after(*next_balance, next))
7235 *next_balance = next;
7239 * idle_balance is called by schedule() if this_cpu is about to become
7240 * idle. Attempts to pull tasks from other CPUs.
7242 static int idle_balance(struct rq *this_rq)
7244 unsigned long next_balance = jiffies + HZ;
7245 int this_cpu = this_rq->cpu;
7246 struct sched_domain *sd;
7247 int pulled_task = 0;
7250 idle_enter_fair(this_rq);
7253 * We must set idle_stamp _before_ calling idle_balance(), such that we
7254 * measure the duration of idle_balance() as idle time.
7256 this_rq->idle_stamp = rq_clock(this_rq);
7258 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7259 !this_rq->rd->overload) {
7261 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7263 update_next_balance(sd, 0, &next_balance);
7269 raw_spin_unlock(&this_rq->lock);
7271 update_blocked_averages(this_cpu);
7273 for_each_domain(this_cpu, sd) {
7274 int continue_balancing = 1;
7275 u64 t0, domain_cost;
7277 if (!(sd->flags & SD_LOAD_BALANCE))
7280 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7281 update_next_balance(sd, 0, &next_balance);
7285 if (sd->flags & SD_BALANCE_NEWIDLE) {
7286 t0 = sched_clock_cpu(this_cpu);
7288 pulled_task = load_balance(this_cpu, this_rq,
7290 &continue_balancing);
7292 domain_cost = sched_clock_cpu(this_cpu) - t0;
7293 if (domain_cost > sd->max_newidle_lb_cost)
7294 sd->max_newidle_lb_cost = domain_cost;
7296 curr_cost += domain_cost;
7299 update_next_balance(sd, 0, &next_balance);
7302 * Stop searching for tasks to pull if there are
7303 * now runnable tasks on this rq.
7305 if (pulled_task || this_rq->nr_running > 0)
7310 raw_spin_lock(&this_rq->lock);
7312 if (curr_cost > this_rq->max_idle_balance_cost)
7313 this_rq->max_idle_balance_cost = curr_cost;
7316 * While browsing the domains, we released the rq lock, a task could
7317 * have been enqueued in the meantime. Since we're not going idle,
7318 * pretend we pulled a task.
7320 if (this_rq->cfs.h_nr_running && !pulled_task)
7324 /* Move the next balance forward */
7325 if (time_after(this_rq->next_balance, next_balance))
7326 this_rq->next_balance = next_balance;
7328 /* Is there a task of a high priority class? */
7329 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7333 idle_exit_fair(this_rq);
7334 this_rq->idle_stamp = 0;
7341 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7342 * running tasks off the busiest CPU onto idle CPUs. It requires at
7343 * least 1 task to be running on each physical CPU where possible, and
7344 * avoids physical / logical imbalances.
7346 static int active_load_balance_cpu_stop(void *data)
7348 struct rq *busiest_rq = data;
7349 int busiest_cpu = cpu_of(busiest_rq);
7350 int target_cpu = busiest_rq->push_cpu;
7351 struct rq *target_rq = cpu_rq(target_cpu);
7352 struct sched_domain *sd;
7353 struct task_struct *p = NULL;
7355 raw_spin_lock_irq(&busiest_rq->lock);
7357 /* make sure the requested cpu hasn't gone down in the meantime */
7358 if (unlikely(busiest_cpu != smp_processor_id() ||
7359 !busiest_rq->active_balance))
7362 /* Is there any task to move? */
7363 if (busiest_rq->nr_running <= 1)
7367 * This condition is "impossible", if it occurs
7368 * we need to fix it. Originally reported by
7369 * Bjorn Helgaas on a 128-cpu setup.
7371 BUG_ON(busiest_rq == target_rq);
7373 /* Search for an sd spanning us and the target CPU. */
7375 for_each_domain(target_cpu, sd) {
7376 if ((sd->flags & SD_LOAD_BALANCE) &&
7377 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7382 struct lb_env env = {
7384 .dst_cpu = target_cpu,
7385 .dst_rq = target_rq,
7386 .src_cpu = busiest_rq->cpu,
7387 .src_rq = busiest_rq,
7391 schedstat_inc(sd, alb_count);
7393 p = detach_one_task(&env);
7395 schedstat_inc(sd, alb_pushed);
7397 schedstat_inc(sd, alb_failed);
7401 busiest_rq->active_balance = 0;
7402 raw_spin_unlock(&busiest_rq->lock);
7405 attach_one_task(target_rq, p);
7412 static inline int on_null_domain(struct rq *rq)
7414 return unlikely(!rcu_dereference_sched(rq->sd));
7417 #ifdef CONFIG_NO_HZ_COMMON
7419 * idle load balancing details
7420 * - When one of the busy CPUs notice that there may be an idle rebalancing
7421 * needed, they will kick the idle load balancer, which then does idle
7422 * load balancing for all the idle CPUs.
7425 cpumask_var_t idle_cpus_mask;
7427 unsigned long next_balance; /* in jiffy units */
7428 } nohz ____cacheline_aligned;
7430 static inline int find_new_ilb(void)
7432 int ilb = cpumask_first(nohz.idle_cpus_mask);
7434 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7441 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7442 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7443 * CPU (if there is one).
7445 static void nohz_balancer_kick(void)
7449 nohz.next_balance++;
7451 ilb_cpu = find_new_ilb();
7453 if (ilb_cpu >= nr_cpu_ids)
7456 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7459 * Use smp_send_reschedule() instead of resched_cpu().
7460 * This way we generate a sched IPI on the target cpu which
7461 * is idle. And the softirq performing nohz idle load balance
7462 * will be run before returning from the IPI.
7464 smp_send_reschedule(ilb_cpu);
7468 static inline void nohz_balance_exit_idle(int cpu)
7470 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7472 * Completely isolated CPUs don't ever set, so we must test.
7474 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7475 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7476 atomic_dec(&nohz.nr_cpus);
7478 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7482 static inline void set_cpu_sd_state_busy(void)
7484 struct sched_domain *sd;
7485 int cpu = smp_processor_id();
7488 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7490 if (!sd || !sd->nohz_idle)
7494 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7499 void set_cpu_sd_state_idle(void)
7501 struct sched_domain *sd;
7502 int cpu = smp_processor_id();
7505 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7507 if (!sd || sd->nohz_idle)
7511 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7517 * This routine will record that the cpu is going idle with tick stopped.
7518 * This info will be used in performing idle load balancing in the future.
7520 void nohz_balance_enter_idle(int cpu)
7523 * If this cpu is going down, then nothing needs to be done.
7525 if (!cpu_active(cpu))
7528 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7532 * If we're a completely isolated CPU, we don't play.
7534 if (on_null_domain(cpu_rq(cpu)))
7537 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7538 atomic_inc(&nohz.nr_cpus);
7539 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7542 static int sched_ilb_notifier(struct notifier_block *nfb,
7543 unsigned long action, void *hcpu)
7545 switch (action & ~CPU_TASKS_FROZEN) {
7547 nohz_balance_exit_idle(smp_processor_id());
7555 static DEFINE_SPINLOCK(balancing);
7558 * Scale the max load_balance interval with the number of CPUs in the system.
7559 * This trades load-balance latency on larger machines for less cross talk.
7561 void update_max_interval(void)
7563 max_load_balance_interval = HZ*num_online_cpus()/10;
7567 * It checks each scheduling domain to see if it is due to be balanced,
7568 * and initiates a balancing operation if so.
7570 * Balancing parameters are set up in init_sched_domains.
7572 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7574 int continue_balancing = 1;
7576 unsigned long interval;
7577 struct sched_domain *sd;
7578 /* Earliest time when we have to do rebalance again */
7579 unsigned long next_balance = jiffies + 60*HZ;
7580 int update_next_balance = 0;
7581 int need_serialize, need_decay = 0;
7584 update_blocked_averages(cpu);
7587 for_each_domain(cpu, sd) {
7589 * Decay the newidle max times here because this is a regular
7590 * visit to all the domains. Decay ~1% per second.
7592 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7593 sd->max_newidle_lb_cost =
7594 (sd->max_newidle_lb_cost * 253) / 256;
7595 sd->next_decay_max_lb_cost = jiffies + HZ;
7598 max_cost += sd->max_newidle_lb_cost;
7600 if (!(sd->flags & SD_LOAD_BALANCE))
7604 * Stop the load balance at this level. There is another
7605 * CPU in our sched group which is doing load balancing more
7608 if (!continue_balancing) {
7614 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7616 need_serialize = sd->flags & SD_SERIALIZE;
7617 if (need_serialize) {
7618 if (!spin_trylock(&balancing))
7622 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7623 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7625 * The LBF_DST_PINNED logic could have changed
7626 * env->dst_cpu, so we can't know our idle
7627 * state even if we migrated tasks. Update it.
7629 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7631 sd->last_balance = jiffies;
7632 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7635 spin_unlock(&balancing);
7637 if (time_after(next_balance, sd->last_balance + interval)) {
7638 next_balance = sd->last_balance + interval;
7639 update_next_balance = 1;
7644 * Ensure the rq-wide value also decays but keep it at a
7645 * reasonable floor to avoid funnies with rq->avg_idle.
7647 rq->max_idle_balance_cost =
7648 max((u64)sysctl_sched_migration_cost, max_cost);
7653 * next_balance will be updated only when there is a need.
7654 * When the cpu is attached to null domain for ex, it will not be
7657 if (likely(update_next_balance)) {
7658 rq->next_balance = next_balance;
7660 #ifdef CONFIG_NO_HZ_COMMON
7662 * If this CPU has been elected to perform the nohz idle
7663 * balance. Other idle CPUs have already rebalanced with
7664 * nohz_idle_balance() and nohz.next_balance has been
7665 * updated accordingly. This CPU is now running the idle load
7666 * balance for itself and we need to update the
7667 * nohz.next_balance accordingly.
7669 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
7670 nohz.next_balance = rq->next_balance;
7675 #ifdef CONFIG_NO_HZ_COMMON
7677 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7678 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7680 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7682 int this_cpu = this_rq->cpu;
7685 /* Earliest time when we have to do rebalance again */
7686 unsigned long next_balance = jiffies + 60*HZ;
7687 int update_next_balance = 0;
7689 if (idle != CPU_IDLE ||
7690 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7693 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7694 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7698 * If this cpu gets work to do, stop the load balancing
7699 * work being done for other cpus. Next load
7700 * balancing owner will pick it up.
7705 rq = cpu_rq(balance_cpu);
7708 * If time for next balance is due,
7711 if (time_after_eq(jiffies, rq->next_balance)) {
7712 raw_spin_lock_irq(&rq->lock);
7713 update_rq_clock(rq);
7714 update_idle_cpu_load(rq);
7715 raw_spin_unlock_irq(&rq->lock);
7716 rebalance_domains(rq, CPU_IDLE);
7719 if (time_after(next_balance, rq->next_balance)) {
7720 next_balance = rq->next_balance;
7721 update_next_balance = 1;
7726 * next_balance will be updated only when there is a need.
7727 * When the CPU is attached to null domain for ex, it will not be
7730 if (likely(update_next_balance))
7731 nohz.next_balance = next_balance;
7733 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7737 * Current heuristic for kicking the idle load balancer in the presence
7738 * of an idle cpu in the system.
7739 * - This rq has more than one task.
7740 * - This rq has at least one CFS task and the capacity of the CPU is
7741 * significantly reduced because of RT tasks or IRQs.
7742 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7743 * multiple busy cpu.
7744 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7745 * domain span are idle.
7747 static inline bool nohz_kick_needed(struct rq *rq)
7749 unsigned long now = jiffies;
7750 struct sched_domain *sd;
7751 struct sched_group_capacity *sgc;
7752 int nr_busy, cpu = rq->cpu;
7755 if (unlikely(rq->idle_balance))
7759 * We may be recently in ticked or tickless idle mode. At the first
7760 * busy tick after returning from idle, we will update the busy stats.
7762 set_cpu_sd_state_busy();
7763 nohz_balance_exit_idle(cpu);
7766 * None are in tickless mode and hence no need for NOHZ idle load
7769 if (likely(!atomic_read(&nohz.nr_cpus)))
7772 if (time_before(now, nohz.next_balance))
7775 if (rq->nr_running >= 2)
7779 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7781 sgc = sd->groups->sgc;
7782 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7791 sd = rcu_dereference(rq->sd);
7793 if ((rq->cfs.h_nr_running >= 1) &&
7794 check_cpu_capacity(rq, sd)) {
7800 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7801 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7802 sched_domain_span(sd)) < cpu)) {
7812 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7816 * run_rebalance_domains is triggered when needed from the scheduler tick.
7817 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7819 static void run_rebalance_domains(struct softirq_action *h)
7821 struct rq *this_rq = this_rq();
7822 enum cpu_idle_type idle = this_rq->idle_balance ?
7823 CPU_IDLE : CPU_NOT_IDLE;
7826 * If this cpu has a pending nohz_balance_kick, then do the
7827 * balancing on behalf of the other idle cpus whose ticks are
7828 * stopped. Do nohz_idle_balance *before* rebalance_domains to
7829 * give the idle cpus a chance to load balance. Else we may
7830 * load balance only within the local sched_domain hierarchy
7831 * and abort nohz_idle_balance altogether if we pull some load.
7833 nohz_idle_balance(this_rq, idle);
7834 rebalance_domains(this_rq, idle);
7838 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7840 void trigger_load_balance(struct rq *rq)
7842 /* Don't need to rebalance while attached to NULL domain */
7843 if (unlikely(on_null_domain(rq)))
7846 if (time_after_eq(jiffies, rq->next_balance))
7847 raise_softirq(SCHED_SOFTIRQ);
7848 #ifdef CONFIG_NO_HZ_COMMON
7849 if (nohz_kick_needed(rq))
7850 nohz_balancer_kick();
7854 static void rq_online_fair(struct rq *rq)
7858 update_runtime_enabled(rq);
7861 static void rq_offline_fair(struct rq *rq)
7865 /* Ensure any throttled groups are reachable by pick_next_task */
7866 unthrottle_offline_cfs_rqs(rq);
7869 #endif /* CONFIG_SMP */
7872 * scheduler tick hitting a task of our scheduling class:
7874 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7876 struct cfs_rq *cfs_rq;
7877 struct sched_entity *se = &curr->se;
7879 for_each_sched_entity(se) {
7880 cfs_rq = cfs_rq_of(se);
7881 entity_tick(cfs_rq, se, queued);
7884 if (!static_branch_unlikely(&sched_numa_balancing))
7885 task_tick_numa(rq, curr);
7889 * called on fork with the child task as argument from the parent's context
7890 * - child not yet on the tasklist
7891 * - preemption disabled
7893 static void task_fork_fair(struct task_struct *p)
7895 struct cfs_rq *cfs_rq;
7896 struct sched_entity *se = &p->se, *curr;
7897 int this_cpu = smp_processor_id();
7898 struct rq *rq = this_rq();
7899 unsigned long flags;
7901 raw_spin_lock_irqsave(&rq->lock, flags);
7903 update_rq_clock(rq);
7905 cfs_rq = task_cfs_rq(current);
7906 curr = cfs_rq->curr;
7909 * Not only the cpu but also the task_group of the parent might have
7910 * been changed after parent->se.parent,cfs_rq were copied to
7911 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7912 * of child point to valid ones.
7915 __set_task_cpu(p, this_cpu);
7918 update_curr(cfs_rq);
7921 se->vruntime = curr->vruntime;
7922 place_entity(cfs_rq, se, 1);
7924 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7926 * Upon rescheduling, sched_class::put_prev_task() will place
7927 * 'current' within the tree based on its new key value.
7929 swap(curr->vruntime, se->vruntime);
7933 se->vruntime -= cfs_rq->min_vruntime;
7935 raw_spin_unlock_irqrestore(&rq->lock, flags);
7939 * Priority of the task has changed. Check to see if we preempt
7943 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7945 if (!task_on_rq_queued(p))
7949 * Reschedule if we are currently running on this runqueue and
7950 * our priority decreased, or if we are not currently running on
7951 * this runqueue and our priority is higher than the current's
7953 if (rq->curr == p) {
7954 if (p->prio > oldprio)
7957 check_preempt_curr(rq, p, 0);
7960 static inline bool vruntime_normalized(struct task_struct *p)
7962 struct sched_entity *se = &p->se;
7965 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
7966 * the dequeue_entity(.flags=0) will already have normalized the
7973 * When !on_rq, vruntime of the task has usually NOT been normalized.
7974 * But there are some cases where it has already been normalized:
7976 * - A forked child which is waiting for being woken up by
7977 * wake_up_new_task().
7978 * - A task which has been woken up by try_to_wake_up() and
7979 * waiting for actually being woken up by sched_ttwu_pending().
7981 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
7987 static void detach_task_cfs_rq(struct task_struct *p)
7989 struct sched_entity *se = &p->se;
7990 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7992 if (!vruntime_normalized(p)) {
7994 * Fix up our vruntime so that the current sleep doesn't
7995 * cause 'unlimited' sleep bonus.
7997 place_entity(cfs_rq, se, 0);
7998 se->vruntime -= cfs_rq->min_vruntime;
8001 /* Catch up with the cfs_rq and remove our load when we leave */
8002 detach_entity_load_avg(cfs_rq, se);
8005 static void attach_task_cfs_rq(struct task_struct *p)
8007 struct sched_entity *se = &p->se;
8008 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8010 #ifdef CONFIG_FAIR_GROUP_SCHED
8012 * Since the real-depth could have been changed (only FAIR
8013 * class maintain depth value), reset depth properly.
8015 se->depth = se->parent ? se->parent->depth + 1 : 0;
8018 /* Synchronize task with its cfs_rq */
8019 attach_entity_load_avg(cfs_rq, se);
8021 if (!vruntime_normalized(p))
8022 se->vruntime += cfs_rq->min_vruntime;
8025 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8027 detach_task_cfs_rq(p);
8030 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8032 attach_task_cfs_rq(p);
8034 if (task_on_rq_queued(p)) {
8036 * We were most likely switched from sched_rt, so
8037 * kick off the schedule if running, otherwise just see
8038 * if we can still preempt the current task.
8043 check_preempt_curr(rq, p, 0);
8047 /* Account for a task changing its policy or group.
8049 * This routine is mostly called to set cfs_rq->curr field when a task
8050 * migrates between groups/classes.
8052 static void set_curr_task_fair(struct rq *rq)
8054 struct sched_entity *se = &rq->curr->se;
8056 for_each_sched_entity(se) {
8057 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8059 set_next_entity(cfs_rq, se);
8060 /* ensure bandwidth has been allocated on our new cfs_rq */
8061 account_cfs_rq_runtime(cfs_rq, 0);
8065 void init_cfs_rq(struct cfs_rq *cfs_rq)
8067 cfs_rq->tasks_timeline = RB_ROOT;
8068 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8069 #ifndef CONFIG_64BIT
8070 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8073 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8074 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8078 #ifdef CONFIG_FAIR_GROUP_SCHED
8079 static void task_move_group_fair(struct task_struct *p)
8081 detach_task_cfs_rq(p);
8082 set_task_rq(p, task_cpu(p));
8085 /* Tell se's cfs_rq has been changed -- migrated */
8086 p->se.avg.last_update_time = 0;
8088 attach_task_cfs_rq(p);
8091 void free_fair_sched_group(struct task_group *tg)
8095 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8097 for_each_possible_cpu(i) {
8099 kfree(tg->cfs_rq[i]);
8102 remove_entity_load_avg(tg->se[i]);
8111 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8113 struct cfs_rq *cfs_rq;
8114 struct sched_entity *se;
8117 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8120 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8124 tg->shares = NICE_0_LOAD;
8126 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8128 for_each_possible_cpu(i) {
8129 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8130 GFP_KERNEL, cpu_to_node(i));
8134 se = kzalloc_node(sizeof(struct sched_entity),
8135 GFP_KERNEL, cpu_to_node(i));
8139 init_cfs_rq(cfs_rq);
8140 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8141 init_entity_runnable_average(se);
8152 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8154 struct rq *rq = cpu_rq(cpu);
8155 unsigned long flags;
8158 * Only empty task groups can be destroyed; so we can speculatively
8159 * check on_list without danger of it being re-added.
8161 if (!tg->cfs_rq[cpu]->on_list)
8164 raw_spin_lock_irqsave(&rq->lock, flags);
8165 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8166 raw_spin_unlock_irqrestore(&rq->lock, flags);
8169 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8170 struct sched_entity *se, int cpu,
8171 struct sched_entity *parent)
8173 struct rq *rq = cpu_rq(cpu);
8177 init_cfs_rq_runtime(cfs_rq);
8179 tg->cfs_rq[cpu] = cfs_rq;
8182 /* se could be NULL for root_task_group */
8187 se->cfs_rq = &rq->cfs;
8190 se->cfs_rq = parent->my_q;
8191 se->depth = parent->depth + 1;
8195 /* guarantee group entities always have weight */
8196 update_load_set(&se->load, NICE_0_LOAD);
8197 se->parent = parent;
8200 static DEFINE_MUTEX(shares_mutex);
8202 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8205 unsigned long flags;
8208 * We can't change the weight of the root cgroup.
8213 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8215 mutex_lock(&shares_mutex);
8216 if (tg->shares == shares)
8219 tg->shares = shares;
8220 for_each_possible_cpu(i) {
8221 struct rq *rq = cpu_rq(i);
8222 struct sched_entity *se;
8225 /* Propagate contribution to hierarchy */
8226 raw_spin_lock_irqsave(&rq->lock, flags);
8228 /* Possible calls to update_curr() need rq clock */
8229 update_rq_clock(rq);
8230 for_each_sched_entity(se)
8231 update_cfs_shares(group_cfs_rq(se));
8232 raw_spin_unlock_irqrestore(&rq->lock, flags);
8236 mutex_unlock(&shares_mutex);
8239 #else /* CONFIG_FAIR_GROUP_SCHED */
8241 void free_fair_sched_group(struct task_group *tg) { }
8243 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8248 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8250 #endif /* CONFIG_FAIR_GROUP_SCHED */
8253 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8255 struct sched_entity *se = &task->se;
8256 unsigned int rr_interval = 0;
8259 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8262 if (rq->cfs.load.weight)
8263 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8269 * All the scheduling class methods:
8271 const struct sched_class fair_sched_class = {
8272 .next = &idle_sched_class,
8273 .enqueue_task = enqueue_task_fair,
8274 .dequeue_task = dequeue_task_fair,
8275 .yield_task = yield_task_fair,
8276 .yield_to_task = yield_to_task_fair,
8278 .check_preempt_curr = check_preempt_wakeup,
8280 .pick_next_task = pick_next_task_fair,
8281 .put_prev_task = put_prev_task_fair,
8284 .select_task_rq = select_task_rq_fair,
8285 .migrate_task_rq = migrate_task_rq_fair,
8287 .rq_online = rq_online_fair,
8288 .rq_offline = rq_offline_fair,
8290 .task_waking = task_waking_fair,
8291 .task_dead = task_dead_fair,
8292 .set_cpus_allowed = set_cpus_allowed_common,
8295 .set_curr_task = set_curr_task_fair,
8296 .task_tick = task_tick_fair,
8297 .task_fork = task_fork_fair,
8299 .prio_changed = prio_changed_fair,
8300 .switched_from = switched_from_fair,
8301 .switched_to = switched_to_fair,
8303 .get_rr_interval = get_rr_interval_fair,
8305 .update_curr = update_curr_fair,
8307 #ifdef CONFIG_FAIR_GROUP_SCHED
8308 .task_move_group = task_move_group_fair,
8312 #ifdef CONFIG_SCHED_DEBUG
8313 void print_cfs_stats(struct seq_file *m, int cpu)
8315 struct cfs_rq *cfs_rq;
8318 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8319 print_cfs_rq(m, cpu, cfs_rq);
8323 #ifdef CONFIG_NUMA_BALANCING
8324 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8327 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8329 for_each_online_node(node) {
8330 if (p->numa_faults) {
8331 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8332 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8334 if (p->numa_group) {
8335 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8336 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8338 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8341 #endif /* CONFIG_NUMA_BALANCING */
8342 #endif /* CONFIG_SCHED_DEBUG */
8344 __init void init_sched_fair_class(void)
8347 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8349 #ifdef CONFIG_NO_HZ_COMMON
8350 nohz.next_balance = jiffies;
8351 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8352 cpu_notifier(sched_ilb_notifier, 0);