2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/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.weight;
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.weight;
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;
2689 int decayed, removed = 0;
2691 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2692 s64 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);
2698 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2699 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2700 sa->util_avg = max_t(long, sa->util_avg - r, 0);
2701 sa->util_sum = max_t(s32, sa->util_sum - r * LOAD_AVG_MAX, 0);
2704 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2705 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2707 #ifndef CONFIG_64BIT
2709 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2712 return decayed || removed;
2715 /* Update task and its cfs_rq load average */
2716 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2718 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2719 u64 now = cfs_rq_clock_task(cfs_rq);
2720 int cpu = cpu_of(rq_of(cfs_rq));
2723 * Track task load average for carrying it to new CPU after migrated, and
2724 * track group sched_entity load average for task_h_load calc in migration
2726 __update_load_avg(now, cpu, &se->avg,
2727 se->on_rq * scale_load_down(se->load.weight),
2728 cfs_rq->curr == se, NULL);
2730 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2731 update_tg_load_avg(cfs_rq, 0);
2734 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2736 if (!sched_feat(ATTACH_AGE_LOAD))
2740 * If we got migrated (either between CPUs or between cgroups) we'll
2741 * have aged the average right before clearing @last_update_time.
2743 if (se->avg.last_update_time) {
2744 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2745 &se->avg, 0, 0, NULL);
2748 * XXX: we could have just aged the entire load away if we've been
2749 * absent from the fair class for too long.
2754 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2755 cfs_rq->avg.load_avg += se->avg.load_avg;
2756 cfs_rq->avg.load_sum += se->avg.load_sum;
2757 cfs_rq->avg.util_avg += se->avg.util_avg;
2758 cfs_rq->avg.util_sum += se->avg.util_sum;
2761 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2763 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2764 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2765 cfs_rq->curr == se, NULL);
2767 cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
2768 cfs_rq->avg.load_sum = max_t(s64, cfs_rq->avg.load_sum - se->avg.load_sum, 0);
2769 cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
2770 cfs_rq->avg.util_sum = max_t(s32, cfs_rq->avg.util_sum - se->avg.util_sum, 0);
2773 /* Add the load generated by se into cfs_rq's load average */
2775 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2777 struct sched_avg *sa = &se->avg;
2778 u64 now = cfs_rq_clock_task(cfs_rq);
2779 int migrated, decayed;
2781 migrated = !sa->last_update_time;
2783 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2784 se->on_rq * scale_load_down(se->load.weight),
2785 cfs_rq->curr == se, NULL);
2788 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2790 cfs_rq->runnable_load_avg += sa->load_avg;
2791 cfs_rq->runnable_load_sum += sa->load_sum;
2794 attach_entity_load_avg(cfs_rq, se);
2796 if (decayed || migrated)
2797 update_tg_load_avg(cfs_rq, 0);
2800 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2802 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2804 update_load_avg(se, 1);
2806 cfs_rq->runnable_load_avg =
2807 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2808 cfs_rq->runnable_load_sum =
2809 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2812 #ifndef CONFIG_64BIT
2813 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2815 u64 last_update_time_copy;
2816 u64 last_update_time;
2819 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2821 last_update_time = cfs_rq->avg.last_update_time;
2822 } while (last_update_time != last_update_time_copy);
2824 return last_update_time;
2827 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2829 return cfs_rq->avg.last_update_time;
2834 * Task first catches up with cfs_rq, and then subtract
2835 * itself from the cfs_rq (task must be off the queue now).
2837 void remove_entity_load_avg(struct sched_entity *se)
2839 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2840 u64 last_update_time;
2843 * Newly created task or never used group entity should not be removed
2844 * from its (source) cfs_rq
2846 if (se->avg.last_update_time == 0)
2849 last_update_time = cfs_rq_last_update_time(cfs_rq);
2851 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2852 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2853 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2857 * Update the rq's load with the elapsed running time before entering
2858 * idle. if the last scheduled task is not a CFS task, idle_enter will
2859 * be the only way to update the runnable statistic.
2861 void idle_enter_fair(struct rq *this_rq)
2866 * Update the rq's load with the elapsed idle time before a task is
2867 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2868 * be the only way to update the runnable statistic.
2870 void idle_exit_fair(struct rq *this_rq)
2874 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2876 return cfs_rq->runnable_load_avg;
2879 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2881 return cfs_rq->avg.load_avg;
2884 static int idle_balance(struct rq *this_rq);
2886 #else /* CONFIG_SMP */
2888 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2890 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2892 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2893 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2896 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2898 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2900 static inline int idle_balance(struct rq *rq)
2905 #endif /* CONFIG_SMP */
2907 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2909 #ifdef CONFIG_SCHEDSTATS
2910 struct task_struct *tsk = NULL;
2912 if (entity_is_task(se))
2915 if (se->statistics.sleep_start) {
2916 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2921 if (unlikely(delta > se->statistics.sleep_max))
2922 se->statistics.sleep_max = delta;
2924 se->statistics.sleep_start = 0;
2925 se->statistics.sum_sleep_runtime += delta;
2928 account_scheduler_latency(tsk, delta >> 10, 1);
2929 trace_sched_stat_sleep(tsk, delta);
2932 if (se->statistics.block_start) {
2933 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2938 if (unlikely(delta > se->statistics.block_max))
2939 se->statistics.block_max = delta;
2941 se->statistics.block_start = 0;
2942 se->statistics.sum_sleep_runtime += delta;
2945 if (tsk->in_iowait) {
2946 se->statistics.iowait_sum += delta;
2947 se->statistics.iowait_count++;
2948 trace_sched_stat_iowait(tsk, delta);
2951 trace_sched_stat_blocked(tsk, delta);
2954 * Blocking time is in units of nanosecs, so shift by
2955 * 20 to get a milliseconds-range estimation of the
2956 * amount of time that the task spent sleeping:
2958 if (unlikely(prof_on == SLEEP_PROFILING)) {
2959 profile_hits(SLEEP_PROFILING,
2960 (void *)get_wchan(tsk),
2963 account_scheduler_latency(tsk, delta >> 10, 0);
2969 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2971 #ifdef CONFIG_SCHED_DEBUG
2972 s64 d = se->vruntime - cfs_rq->min_vruntime;
2977 if (d > 3*sysctl_sched_latency)
2978 schedstat_inc(cfs_rq, nr_spread_over);
2983 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2985 u64 vruntime = cfs_rq->min_vruntime;
2988 * The 'current' period is already promised to the current tasks,
2989 * however the extra weight of the new task will slow them down a
2990 * little, place the new task so that it fits in the slot that
2991 * stays open at the end.
2993 if (initial && sched_feat(START_DEBIT))
2994 vruntime += sched_vslice(cfs_rq, se);
2996 /* sleeps up to a single latency don't count. */
2998 unsigned long thresh = sysctl_sched_latency;
3001 * Halve their sleep time's effect, to allow
3002 * for a gentler effect of sleepers:
3004 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3010 /* ensure we never gain time by being placed backwards. */
3011 se->vruntime = max_vruntime(se->vruntime, vruntime);
3014 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3017 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3020 * Update the normalized vruntime before updating min_vruntime
3021 * through calling update_curr().
3023 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3024 se->vruntime += cfs_rq->min_vruntime;
3027 * Update run-time statistics of the 'current'.
3029 update_curr(cfs_rq);
3030 enqueue_entity_load_avg(cfs_rq, se);
3031 account_entity_enqueue(cfs_rq, se);
3032 update_cfs_shares(cfs_rq);
3034 if (flags & ENQUEUE_WAKEUP) {
3035 place_entity(cfs_rq, se, 0);
3036 enqueue_sleeper(cfs_rq, se);
3039 update_stats_enqueue(cfs_rq, se);
3040 check_spread(cfs_rq, se);
3041 if (se != cfs_rq->curr)
3042 __enqueue_entity(cfs_rq, se);
3045 if (cfs_rq->nr_running == 1) {
3046 list_add_leaf_cfs_rq(cfs_rq);
3047 check_enqueue_throttle(cfs_rq);
3051 static void __clear_buddies_last(struct sched_entity *se)
3053 for_each_sched_entity(se) {
3054 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3055 if (cfs_rq->last != se)
3058 cfs_rq->last = NULL;
3062 static void __clear_buddies_next(struct sched_entity *se)
3064 for_each_sched_entity(se) {
3065 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3066 if (cfs_rq->next != se)
3069 cfs_rq->next = NULL;
3073 static void __clear_buddies_skip(struct sched_entity *se)
3075 for_each_sched_entity(se) {
3076 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3077 if (cfs_rq->skip != se)
3080 cfs_rq->skip = NULL;
3084 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3086 if (cfs_rq->last == se)
3087 __clear_buddies_last(se);
3089 if (cfs_rq->next == se)
3090 __clear_buddies_next(se);
3092 if (cfs_rq->skip == se)
3093 __clear_buddies_skip(se);
3096 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3099 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3102 * Update run-time statistics of the 'current'.
3104 update_curr(cfs_rq);
3105 dequeue_entity_load_avg(cfs_rq, se);
3107 update_stats_dequeue(cfs_rq, se);
3108 if (flags & DEQUEUE_SLEEP) {
3109 #ifdef CONFIG_SCHEDSTATS
3110 if (entity_is_task(se)) {
3111 struct task_struct *tsk = task_of(se);
3113 if (tsk->state & TASK_INTERRUPTIBLE)
3114 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3115 if (tsk->state & TASK_UNINTERRUPTIBLE)
3116 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3121 clear_buddies(cfs_rq, se);
3123 if (se != cfs_rq->curr)
3124 __dequeue_entity(cfs_rq, se);
3126 account_entity_dequeue(cfs_rq, se);
3129 * Normalize the entity after updating the min_vruntime because the
3130 * update can refer to the ->curr item and we need to reflect this
3131 * movement in our normalized position.
3133 if (!(flags & DEQUEUE_SLEEP))
3134 se->vruntime -= cfs_rq->min_vruntime;
3136 /* return excess runtime on last dequeue */
3137 return_cfs_rq_runtime(cfs_rq);
3139 update_min_vruntime(cfs_rq);
3140 update_cfs_shares(cfs_rq);
3144 * Preempt the current task with a newly woken task if needed:
3147 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3149 unsigned long ideal_runtime, delta_exec;
3150 struct sched_entity *se;
3153 ideal_runtime = sched_slice(cfs_rq, curr);
3154 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3155 if (delta_exec > ideal_runtime) {
3156 resched_curr(rq_of(cfs_rq));
3158 * The current task ran long enough, ensure it doesn't get
3159 * re-elected due to buddy favours.
3161 clear_buddies(cfs_rq, curr);
3166 * Ensure that a task that missed wakeup preemption by a
3167 * narrow margin doesn't have to wait for a full slice.
3168 * This also mitigates buddy induced latencies under load.
3170 if (delta_exec < sysctl_sched_min_granularity)
3173 se = __pick_first_entity(cfs_rq);
3174 delta = curr->vruntime - se->vruntime;
3179 if (delta > ideal_runtime)
3180 resched_curr(rq_of(cfs_rq));
3184 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3186 /* 'current' is not kept within the tree. */
3189 * Any task has to be enqueued before it get to execute on
3190 * a CPU. So account for the time it spent waiting on the
3193 update_stats_wait_end(cfs_rq, se);
3194 __dequeue_entity(cfs_rq, se);
3195 update_load_avg(se, 1);
3198 update_stats_curr_start(cfs_rq, se);
3200 #ifdef CONFIG_SCHEDSTATS
3202 * Track our maximum slice length, if the CPU's load is at
3203 * least twice that of our own weight (i.e. dont track it
3204 * when there are only lesser-weight tasks around):
3206 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3207 se->statistics.slice_max = max(se->statistics.slice_max,
3208 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3211 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3215 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3218 * Pick the next process, keeping these things in mind, in this order:
3219 * 1) keep things fair between processes/task groups
3220 * 2) pick the "next" process, since someone really wants that to run
3221 * 3) pick the "last" process, for cache locality
3222 * 4) do not run the "skip" process, if something else is available
3224 static struct sched_entity *
3225 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3227 struct sched_entity *left = __pick_first_entity(cfs_rq);
3228 struct sched_entity *se;
3231 * If curr is set we have to see if its left of the leftmost entity
3232 * still in the tree, provided there was anything in the tree at all.
3234 if (!left || (curr && entity_before(curr, left)))
3237 se = left; /* ideally we run the leftmost entity */
3240 * Avoid running the skip buddy, if running something else can
3241 * be done without getting too unfair.
3243 if (cfs_rq->skip == se) {
3244 struct sched_entity *second;
3247 second = __pick_first_entity(cfs_rq);
3249 second = __pick_next_entity(se);
3250 if (!second || (curr && entity_before(curr, second)))
3254 if (second && wakeup_preempt_entity(second, left) < 1)
3259 * Prefer last buddy, try to return the CPU to a preempted task.
3261 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3265 * Someone really wants this to run. If it's not unfair, run it.
3267 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3270 clear_buddies(cfs_rq, se);
3275 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3277 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3280 * If still on the runqueue then deactivate_task()
3281 * was not called and update_curr() has to be done:
3284 update_curr(cfs_rq);
3286 /* throttle cfs_rqs exceeding runtime */
3287 check_cfs_rq_runtime(cfs_rq);
3289 check_spread(cfs_rq, prev);
3291 update_stats_wait_start(cfs_rq, prev);
3292 /* Put 'current' back into the tree. */
3293 __enqueue_entity(cfs_rq, prev);
3294 /* in !on_rq case, update occurred at dequeue */
3295 update_load_avg(prev, 0);
3297 cfs_rq->curr = NULL;
3301 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3304 * Update run-time statistics of the 'current'.
3306 update_curr(cfs_rq);
3309 * Ensure that runnable average is periodically updated.
3311 update_load_avg(curr, 1);
3312 update_cfs_shares(cfs_rq);
3314 #ifdef CONFIG_SCHED_HRTICK
3316 * queued ticks are scheduled to match the slice, so don't bother
3317 * validating it and just reschedule.
3320 resched_curr(rq_of(cfs_rq));
3324 * don't let the period tick interfere with the hrtick preemption
3326 if (!sched_feat(DOUBLE_TICK) &&
3327 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3331 if (cfs_rq->nr_running > 1)
3332 check_preempt_tick(cfs_rq, curr);
3336 /**************************************************
3337 * CFS bandwidth control machinery
3340 #ifdef CONFIG_CFS_BANDWIDTH
3342 #ifdef HAVE_JUMP_LABEL
3343 static struct static_key __cfs_bandwidth_used;
3345 static inline bool cfs_bandwidth_used(void)
3347 return static_key_false(&__cfs_bandwidth_used);
3350 void cfs_bandwidth_usage_inc(void)
3352 static_key_slow_inc(&__cfs_bandwidth_used);
3355 void cfs_bandwidth_usage_dec(void)
3357 static_key_slow_dec(&__cfs_bandwidth_used);
3359 #else /* HAVE_JUMP_LABEL */
3360 static bool cfs_bandwidth_used(void)
3365 void cfs_bandwidth_usage_inc(void) {}
3366 void cfs_bandwidth_usage_dec(void) {}
3367 #endif /* HAVE_JUMP_LABEL */
3370 * default period for cfs group bandwidth.
3371 * default: 0.1s, units: nanoseconds
3373 static inline u64 default_cfs_period(void)
3375 return 100000000ULL;
3378 static inline u64 sched_cfs_bandwidth_slice(void)
3380 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3384 * Replenish runtime according to assigned quota and update expiration time.
3385 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3386 * additional synchronization around rq->lock.
3388 * requires cfs_b->lock
3390 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3394 if (cfs_b->quota == RUNTIME_INF)
3397 now = sched_clock_cpu(smp_processor_id());
3398 cfs_b->runtime = cfs_b->quota;
3399 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3402 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3404 return &tg->cfs_bandwidth;
3407 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3408 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3410 if (unlikely(cfs_rq->throttle_count))
3411 return cfs_rq->throttled_clock_task;
3413 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3416 /* returns 0 on failure to allocate runtime */
3417 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3419 struct task_group *tg = cfs_rq->tg;
3420 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3421 u64 amount = 0, min_amount, expires;
3423 /* note: this is a positive sum as runtime_remaining <= 0 */
3424 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3426 raw_spin_lock(&cfs_b->lock);
3427 if (cfs_b->quota == RUNTIME_INF)
3428 amount = min_amount;
3430 start_cfs_bandwidth(cfs_b);
3432 if (cfs_b->runtime > 0) {
3433 amount = min(cfs_b->runtime, min_amount);
3434 cfs_b->runtime -= amount;
3438 expires = cfs_b->runtime_expires;
3439 raw_spin_unlock(&cfs_b->lock);
3441 cfs_rq->runtime_remaining += amount;
3443 * we may have advanced our local expiration to account for allowed
3444 * spread between our sched_clock and the one on which runtime was
3447 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3448 cfs_rq->runtime_expires = expires;
3450 return cfs_rq->runtime_remaining > 0;
3454 * Note: This depends on the synchronization provided by sched_clock and the
3455 * fact that rq->clock snapshots this value.
3457 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3459 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3461 /* if the deadline is ahead of our clock, nothing to do */
3462 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3465 if (cfs_rq->runtime_remaining < 0)
3469 * If the local deadline has passed we have to consider the
3470 * possibility that our sched_clock is 'fast' and the global deadline
3471 * has not truly expired.
3473 * Fortunately we can check determine whether this the case by checking
3474 * whether the global deadline has advanced. It is valid to compare
3475 * cfs_b->runtime_expires without any locks since we only care about
3476 * exact equality, so a partial write will still work.
3479 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3480 /* extend local deadline, drift is bounded above by 2 ticks */
3481 cfs_rq->runtime_expires += TICK_NSEC;
3483 /* global deadline is ahead, expiration has passed */
3484 cfs_rq->runtime_remaining = 0;
3488 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3490 /* dock delta_exec before expiring quota (as it could span periods) */
3491 cfs_rq->runtime_remaining -= delta_exec;
3492 expire_cfs_rq_runtime(cfs_rq);
3494 if (likely(cfs_rq->runtime_remaining > 0))
3498 * if we're unable to extend our runtime we resched so that the active
3499 * hierarchy can be throttled
3501 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3502 resched_curr(rq_of(cfs_rq));
3505 static __always_inline
3506 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3508 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3511 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3514 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3516 return cfs_bandwidth_used() && cfs_rq->throttled;
3519 /* check whether cfs_rq, or any parent, is throttled */
3520 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3522 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3526 * Ensure that neither of the group entities corresponding to src_cpu or
3527 * dest_cpu are members of a throttled hierarchy when performing group
3528 * load-balance operations.
3530 static inline int throttled_lb_pair(struct task_group *tg,
3531 int src_cpu, int dest_cpu)
3533 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3535 src_cfs_rq = tg->cfs_rq[src_cpu];
3536 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3538 return throttled_hierarchy(src_cfs_rq) ||
3539 throttled_hierarchy(dest_cfs_rq);
3542 /* updated child weight may affect parent so we have to do this bottom up */
3543 static int tg_unthrottle_up(struct task_group *tg, void *data)
3545 struct rq *rq = data;
3546 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3548 cfs_rq->throttle_count--;
3550 if (!cfs_rq->throttle_count) {
3551 /* adjust cfs_rq_clock_task() */
3552 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3553 cfs_rq->throttled_clock_task;
3560 static int tg_throttle_down(struct task_group *tg, void *data)
3562 struct rq *rq = data;
3563 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3565 /* group is entering throttled state, stop time */
3566 if (!cfs_rq->throttle_count)
3567 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3568 cfs_rq->throttle_count++;
3573 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3575 struct rq *rq = rq_of(cfs_rq);
3576 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3577 struct sched_entity *se;
3578 long task_delta, dequeue = 1;
3581 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3583 /* freeze hierarchy runnable averages while throttled */
3585 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3588 task_delta = cfs_rq->h_nr_running;
3589 for_each_sched_entity(se) {
3590 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3591 /* throttled entity or throttle-on-deactivate */
3596 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3597 qcfs_rq->h_nr_running -= task_delta;
3599 if (qcfs_rq->load.weight)
3604 sub_nr_running(rq, task_delta);
3606 cfs_rq->throttled = 1;
3607 cfs_rq->throttled_clock = rq_clock(rq);
3608 raw_spin_lock(&cfs_b->lock);
3609 empty = list_empty(&cfs_b->throttled_cfs_rq);
3612 * Add to the _head_ of the list, so that an already-started
3613 * distribute_cfs_runtime will not see us
3615 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3618 * If we're the first throttled task, make sure the bandwidth
3622 start_cfs_bandwidth(cfs_b);
3624 raw_spin_unlock(&cfs_b->lock);
3627 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3629 struct rq *rq = rq_of(cfs_rq);
3630 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3631 struct sched_entity *se;
3635 se = cfs_rq->tg->se[cpu_of(rq)];
3637 cfs_rq->throttled = 0;
3639 update_rq_clock(rq);
3641 raw_spin_lock(&cfs_b->lock);
3642 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3643 list_del_rcu(&cfs_rq->throttled_list);
3644 raw_spin_unlock(&cfs_b->lock);
3646 /* update hierarchical throttle state */
3647 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3649 if (!cfs_rq->load.weight)
3652 task_delta = cfs_rq->h_nr_running;
3653 for_each_sched_entity(se) {
3657 cfs_rq = cfs_rq_of(se);
3659 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3660 cfs_rq->h_nr_running += task_delta;
3662 if (cfs_rq_throttled(cfs_rq))
3667 add_nr_running(rq, task_delta);
3669 /* determine whether we need to wake up potentially idle cpu */
3670 if (rq->curr == rq->idle && rq->cfs.nr_running)
3674 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3675 u64 remaining, u64 expires)
3677 struct cfs_rq *cfs_rq;
3679 u64 starting_runtime = remaining;
3682 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3684 struct rq *rq = rq_of(cfs_rq);
3686 raw_spin_lock(&rq->lock);
3687 if (!cfs_rq_throttled(cfs_rq))
3690 runtime = -cfs_rq->runtime_remaining + 1;
3691 if (runtime > remaining)
3692 runtime = remaining;
3693 remaining -= runtime;
3695 cfs_rq->runtime_remaining += runtime;
3696 cfs_rq->runtime_expires = expires;
3698 /* we check whether we're throttled above */
3699 if (cfs_rq->runtime_remaining > 0)
3700 unthrottle_cfs_rq(cfs_rq);
3703 raw_spin_unlock(&rq->lock);
3710 return starting_runtime - remaining;
3714 * Responsible for refilling a task_group's bandwidth and unthrottling its
3715 * cfs_rqs as appropriate. If there has been no activity within the last
3716 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3717 * used to track this state.
3719 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3721 u64 runtime, runtime_expires;
3724 /* no need to continue the timer with no bandwidth constraint */
3725 if (cfs_b->quota == RUNTIME_INF)
3726 goto out_deactivate;
3728 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3729 cfs_b->nr_periods += overrun;
3732 * idle depends on !throttled (for the case of a large deficit), and if
3733 * we're going inactive then everything else can be deferred
3735 if (cfs_b->idle && !throttled)
3736 goto out_deactivate;
3738 __refill_cfs_bandwidth_runtime(cfs_b);
3741 /* mark as potentially idle for the upcoming period */
3746 /* account preceding periods in which throttling occurred */
3747 cfs_b->nr_throttled += overrun;
3749 runtime_expires = cfs_b->runtime_expires;
3752 * This check is repeated as we are holding onto the new bandwidth while
3753 * we unthrottle. This can potentially race with an unthrottled group
3754 * trying to acquire new bandwidth from the global pool. This can result
3755 * in us over-using our runtime if it is all used during this loop, but
3756 * only by limited amounts in that extreme case.
3758 while (throttled && cfs_b->runtime > 0) {
3759 runtime = cfs_b->runtime;
3760 raw_spin_unlock(&cfs_b->lock);
3761 /* we can't nest cfs_b->lock while distributing bandwidth */
3762 runtime = distribute_cfs_runtime(cfs_b, runtime,
3764 raw_spin_lock(&cfs_b->lock);
3766 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3768 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3772 * While we are ensured activity in the period following an
3773 * unthrottle, this also covers the case in which the new bandwidth is
3774 * insufficient to cover the existing bandwidth deficit. (Forcing the
3775 * timer to remain active while there are any throttled entities.)
3785 /* a cfs_rq won't donate quota below this amount */
3786 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3787 /* minimum remaining period time to redistribute slack quota */
3788 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3789 /* how long we wait to gather additional slack before distributing */
3790 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3793 * Are we near the end of the current quota period?
3795 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3796 * hrtimer base being cleared by hrtimer_start. In the case of
3797 * migrate_hrtimers, base is never cleared, so we are fine.
3799 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3801 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3804 /* if the call-back is running a quota refresh is already occurring */
3805 if (hrtimer_callback_running(refresh_timer))
3808 /* is a quota refresh about to occur? */
3809 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3810 if (remaining < min_expire)
3816 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3818 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3820 /* if there's a quota refresh soon don't bother with slack */
3821 if (runtime_refresh_within(cfs_b, min_left))
3824 hrtimer_start(&cfs_b->slack_timer,
3825 ns_to_ktime(cfs_bandwidth_slack_period),
3829 /* we know any runtime found here is valid as update_curr() precedes return */
3830 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3832 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3833 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3835 if (slack_runtime <= 0)
3838 raw_spin_lock(&cfs_b->lock);
3839 if (cfs_b->quota != RUNTIME_INF &&
3840 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3841 cfs_b->runtime += slack_runtime;
3843 /* we are under rq->lock, defer unthrottling using a timer */
3844 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3845 !list_empty(&cfs_b->throttled_cfs_rq))
3846 start_cfs_slack_bandwidth(cfs_b);
3848 raw_spin_unlock(&cfs_b->lock);
3850 /* even if it's not valid for return we don't want to try again */
3851 cfs_rq->runtime_remaining -= slack_runtime;
3854 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3856 if (!cfs_bandwidth_used())
3859 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3862 __return_cfs_rq_runtime(cfs_rq);
3866 * This is done with a timer (instead of inline with bandwidth return) since
3867 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3869 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3871 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3874 /* confirm we're still not at a refresh boundary */
3875 raw_spin_lock(&cfs_b->lock);
3876 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3877 raw_spin_unlock(&cfs_b->lock);
3881 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3882 runtime = cfs_b->runtime;
3884 expires = cfs_b->runtime_expires;
3885 raw_spin_unlock(&cfs_b->lock);
3890 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3892 raw_spin_lock(&cfs_b->lock);
3893 if (expires == cfs_b->runtime_expires)
3894 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3895 raw_spin_unlock(&cfs_b->lock);
3899 * When a group wakes up we want to make sure that its quota is not already
3900 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3901 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3903 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3905 if (!cfs_bandwidth_used())
3908 /* an active group must be handled by the update_curr()->put() path */
3909 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3912 /* ensure the group is not already throttled */
3913 if (cfs_rq_throttled(cfs_rq))
3916 /* update runtime allocation */
3917 account_cfs_rq_runtime(cfs_rq, 0);
3918 if (cfs_rq->runtime_remaining <= 0)
3919 throttle_cfs_rq(cfs_rq);
3922 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3923 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3925 if (!cfs_bandwidth_used())
3928 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3932 * it's possible for a throttled entity to be forced into a running
3933 * state (e.g. set_curr_task), in this case we're finished.
3935 if (cfs_rq_throttled(cfs_rq))
3938 throttle_cfs_rq(cfs_rq);
3942 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3944 struct cfs_bandwidth *cfs_b =
3945 container_of(timer, struct cfs_bandwidth, slack_timer);
3947 do_sched_cfs_slack_timer(cfs_b);
3949 return HRTIMER_NORESTART;
3952 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3954 struct cfs_bandwidth *cfs_b =
3955 container_of(timer, struct cfs_bandwidth, period_timer);
3959 raw_spin_lock(&cfs_b->lock);
3961 overrun = hrtimer_forward_now(timer, cfs_b->period);
3965 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3968 cfs_b->period_active = 0;
3969 raw_spin_unlock(&cfs_b->lock);
3971 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3974 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3976 raw_spin_lock_init(&cfs_b->lock);
3978 cfs_b->quota = RUNTIME_INF;
3979 cfs_b->period = ns_to_ktime(default_cfs_period());
3981 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3982 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
3983 cfs_b->period_timer.function = sched_cfs_period_timer;
3984 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3985 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3988 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3990 cfs_rq->runtime_enabled = 0;
3991 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3994 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3996 lockdep_assert_held(&cfs_b->lock);
3998 if (!cfs_b->period_active) {
3999 cfs_b->period_active = 1;
4000 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4001 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4005 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4007 /* init_cfs_bandwidth() was not called */
4008 if (!cfs_b->throttled_cfs_rq.next)
4011 hrtimer_cancel(&cfs_b->period_timer);
4012 hrtimer_cancel(&cfs_b->slack_timer);
4015 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4017 struct cfs_rq *cfs_rq;
4019 for_each_leaf_cfs_rq(rq, cfs_rq) {
4020 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4022 raw_spin_lock(&cfs_b->lock);
4023 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4024 raw_spin_unlock(&cfs_b->lock);
4028 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4030 struct cfs_rq *cfs_rq;
4032 for_each_leaf_cfs_rq(rq, cfs_rq) {
4033 if (!cfs_rq->runtime_enabled)
4037 * clock_task is not advancing so we just need to make sure
4038 * there's some valid quota amount
4040 cfs_rq->runtime_remaining = 1;
4042 * Offline rq is schedulable till cpu is completely disabled
4043 * in take_cpu_down(), so we prevent new cfs throttling here.
4045 cfs_rq->runtime_enabled = 0;
4047 if (cfs_rq_throttled(cfs_rq))
4048 unthrottle_cfs_rq(cfs_rq);
4052 #else /* CONFIG_CFS_BANDWIDTH */
4053 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4055 return rq_clock_task(rq_of(cfs_rq));
4058 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4059 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4060 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4061 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4063 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4068 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4073 static inline int throttled_lb_pair(struct task_group *tg,
4074 int src_cpu, int dest_cpu)
4079 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4081 #ifdef CONFIG_FAIR_GROUP_SCHED
4082 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4085 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4089 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4090 static inline void update_runtime_enabled(struct rq *rq) {}
4091 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4093 #endif /* CONFIG_CFS_BANDWIDTH */
4095 /**************************************************
4096 * CFS operations on tasks:
4099 #ifdef CONFIG_SCHED_HRTICK
4100 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4102 struct sched_entity *se = &p->se;
4103 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4105 WARN_ON(task_rq(p) != rq);
4107 if (cfs_rq->nr_running > 1) {
4108 u64 slice = sched_slice(cfs_rq, se);
4109 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4110 s64 delta = slice - ran;
4117 hrtick_start(rq, delta);
4122 * called from enqueue/dequeue and updates the hrtick when the
4123 * current task is from our class and nr_running is low enough
4126 static void hrtick_update(struct rq *rq)
4128 struct task_struct *curr = rq->curr;
4130 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4133 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4134 hrtick_start_fair(rq, curr);
4136 #else /* !CONFIG_SCHED_HRTICK */
4138 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4142 static inline void hrtick_update(struct rq *rq)
4147 static unsigned long capacity_orig_of(int cpu);
4148 static int cpu_util(int cpu);
4150 static void update_capacity_of(int cpu)
4152 unsigned long req_cap;
4157 /* Convert scale-invariant capacity to cpu. */
4158 req_cap = cpu_util(cpu) * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4159 set_cfs_cpu_capacity(cpu, true, req_cap);
4162 static bool cpu_overutilized(int cpu);
4165 * The enqueue_task method is called before nr_running is
4166 * increased. Here we update the fair scheduling stats and
4167 * then put the task into the rbtree:
4170 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4172 struct cfs_rq *cfs_rq;
4173 struct sched_entity *se = &p->se;
4174 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4175 int task_wakeup = flags & ENQUEUE_WAKEUP;
4177 for_each_sched_entity(se) {
4180 cfs_rq = cfs_rq_of(se);
4181 enqueue_entity(cfs_rq, se, flags);
4184 * end evaluation on encountering a throttled cfs_rq
4186 * note: in the case of encountering a throttled cfs_rq we will
4187 * post the final h_nr_running increment below.
4189 if (cfs_rq_throttled(cfs_rq))
4191 cfs_rq->h_nr_running++;
4193 flags = ENQUEUE_WAKEUP;
4196 for_each_sched_entity(se) {
4197 cfs_rq = cfs_rq_of(se);
4198 cfs_rq->h_nr_running++;
4200 if (cfs_rq_throttled(cfs_rq))
4203 update_load_avg(se, 1);
4204 update_cfs_shares(cfs_rq);
4208 add_nr_running(rq, 1);
4209 if (!task_new && !rq->rd->overutilized &&
4210 cpu_overutilized(rq->cpu))
4211 rq->rd->overutilized = true;
4214 * We want to potentially trigger a freq switch
4215 * request only for tasks that are waking up; this is
4216 * because we get here also during load balancing, but
4217 * in these cases it seems wise to trigger as single
4218 * request after load balancing is done.
4220 if (task_new || task_wakeup)
4221 update_capacity_of(cpu_of(rq));
4226 static void set_next_buddy(struct sched_entity *se);
4229 * The dequeue_task method is called before nr_running is
4230 * decreased. We remove the task from the rbtree and
4231 * update the fair scheduling stats:
4233 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4235 struct cfs_rq *cfs_rq;
4236 struct sched_entity *se = &p->se;
4237 int task_sleep = flags & DEQUEUE_SLEEP;
4239 for_each_sched_entity(se) {
4240 cfs_rq = cfs_rq_of(se);
4241 dequeue_entity(cfs_rq, se, flags);
4244 * end evaluation on encountering a throttled cfs_rq
4246 * note: in the case of encountering a throttled cfs_rq we will
4247 * post the final h_nr_running decrement below.
4249 if (cfs_rq_throttled(cfs_rq))
4251 cfs_rq->h_nr_running--;
4253 /* Don't dequeue parent if it has other entities besides us */
4254 if (cfs_rq->load.weight) {
4256 * Bias pick_next to pick a task from this cfs_rq, as
4257 * p is sleeping when it is within its sched_slice.
4259 if (task_sleep && parent_entity(se))
4260 set_next_buddy(parent_entity(se));
4262 /* avoid re-evaluating load for this entity */
4263 se = parent_entity(se);
4266 flags |= DEQUEUE_SLEEP;
4269 for_each_sched_entity(se) {
4270 cfs_rq = cfs_rq_of(se);
4271 cfs_rq->h_nr_running--;
4273 if (cfs_rq_throttled(cfs_rq))
4276 update_load_avg(se, 1);
4277 update_cfs_shares(cfs_rq);
4281 sub_nr_running(rq, 1);
4284 * We want to potentially trigger a freq switch
4285 * request only for tasks that are going to sleep;
4286 * this is because we get here also during load
4287 * balancing, but in these cases it seems wise to
4288 * trigger as single request after load balancing is
4292 if (rq->cfs.nr_running)
4293 update_capacity_of(cpu_of(rq));
4294 else if (sched_freq())
4295 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4304 * per rq 'load' arrray crap; XXX kill this.
4308 * The exact cpuload at various idx values, calculated at every tick would be
4309 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4311 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4312 * on nth tick when cpu may be busy, then we have:
4313 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4314 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4316 * decay_load_missed() below does efficient calculation of
4317 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4318 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4320 * The calculation is approximated on a 128 point scale.
4321 * degrade_zero_ticks is the number of ticks after which load at any
4322 * particular idx is approximated to be zero.
4323 * degrade_factor is a precomputed table, a row for each load idx.
4324 * Each column corresponds to degradation factor for a power of two ticks,
4325 * based on 128 point scale.
4327 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4328 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4330 * With this power of 2 load factors, we can degrade the load n times
4331 * by looking at 1 bits in n and doing as many mult/shift instead of
4332 * n mult/shifts needed by the exact degradation.
4334 #define DEGRADE_SHIFT 7
4335 static const unsigned char
4336 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4337 static const unsigned char
4338 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4339 {0, 0, 0, 0, 0, 0, 0, 0},
4340 {64, 32, 8, 0, 0, 0, 0, 0},
4341 {96, 72, 40, 12, 1, 0, 0},
4342 {112, 98, 75, 43, 15, 1, 0},
4343 {120, 112, 98, 76, 45, 16, 2} };
4346 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4347 * would be when CPU is idle and so we just decay the old load without
4348 * adding any new load.
4350 static unsigned long
4351 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4355 if (!missed_updates)
4358 if (missed_updates >= degrade_zero_ticks[idx])
4362 return load >> missed_updates;
4364 while (missed_updates) {
4365 if (missed_updates % 2)
4366 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4368 missed_updates >>= 1;
4375 * Update rq->cpu_load[] statistics. This function is usually called every
4376 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4377 * every tick. We fix it up based on jiffies.
4379 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4380 unsigned long pending_updates)
4384 this_rq->nr_load_updates++;
4386 /* Update our load: */
4387 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4388 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4389 unsigned long old_load, new_load;
4391 /* scale is effectively 1 << i now, and >> i divides by scale */
4393 old_load = this_rq->cpu_load[i];
4394 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4395 new_load = this_load;
4397 * Round up the averaging division if load is increasing. This
4398 * prevents us from getting stuck on 9 if the load is 10, for
4401 if (new_load > old_load)
4402 new_load += scale - 1;
4404 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4407 sched_avg_update(this_rq);
4410 /* Used instead of source_load when we know the type == 0 */
4411 static unsigned long weighted_cpuload(const int cpu)
4413 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4416 #ifdef CONFIG_NO_HZ_COMMON
4418 * There is no sane way to deal with nohz on smp when using jiffies because the
4419 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4420 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4422 * Therefore we cannot use the delta approach from the regular tick since that
4423 * would seriously skew the load calculation. However we'll make do for those
4424 * updates happening while idle (nohz_idle_balance) or coming out of idle
4425 * (tick_nohz_idle_exit).
4427 * This means we might still be one tick off for nohz periods.
4431 * Called from nohz_idle_balance() to update the load ratings before doing the
4434 static void update_idle_cpu_load(struct rq *this_rq)
4436 unsigned long curr_jiffies = READ_ONCE(jiffies);
4437 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4438 unsigned long pending_updates;
4441 * bail if there's load or we're actually up-to-date.
4443 if (load || curr_jiffies == this_rq->last_load_update_tick)
4446 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4447 this_rq->last_load_update_tick = curr_jiffies;
4449 __update_cpu_load(this_rq, load, pending_updates);
4453 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4455 void update_cpu_load_nohz(void)
4457 struct rq *this_rq = this_rq();
4458 unsigned long curr_jiffies = READ_ONCE(jiffies);
4459 unsigned long pending_updates;
4461 if (curr_jiffies == this_rq->last_load_update_tick)
4464 raw_spin_lock(&this_rq->lock);
4465 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4466 if (pending_updates) {
4467 this_rq->last_load_update_tick = curr_jiffies;
4469 * We were idle, this means load 0, the current load might be
4470 * !0 due to remote wakeups and the sort.
4472 __update_cpu_load(this_rq, 0, pending_updates);
4474 raw_spin_unlock(&this_rq->lock);
4476 #endif /* CONFIG_NO_HZ */
4479 * Called from scheduler_tick()
4481 void update_cpu_load_active(struct rq *this_rq)
4483 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4485 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4487 this_rq->last_load_update_tick = jiffies;
4488 __update_cpu_load(this_rq, load, 1);
4492 * Return a low guess at the load of a migration-source cpu weighted
4493 * according to the scheduling class and "nice" value.
4495 * We want to under-estimate the load of migration sources, to
4496 * balance conservatively.
4498 static unsigned long source_load(int cpu, int type)
4500 struct rq *rq = cpu_rq(cpu);
4501 unsigned long total = weighted_cpuload(cpu);
4503 if (type == 0 || !sched_feat(LB_BIAS))
4506 return min(rq->cpu_load[type-1], total);
4510 * Return a high guess at the load of a migration-target cpu weighted
4511 * according to the scheduling class and "nice" value.
4513 static unsigned long target_load(int cpu, int type)
4515 struct rq *rq = cpu_rq(cpu);
4516 unsigned long total = weighted_cpuload(cpu);
4518 if (type == 0 || !sched_feat(LB_BIAS))
4521 return max(rq->cpu_load[type-1], total);
4524 static unsigned long capacity_of(int cpu)
4526 return cpu_rq(cpu)->cpu_capacity;
4529 static unsigned long capacity_orig_of(int cpu)
4531 return cpu_rq(cpu)->cpu_capacity_orig;
4534 static unsigned long cpu_avg_load_per_task(int cpu)
4536 struct rq *rq = cpu_rq(cpu);
4537 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4538 unsigned long load_avg = weighted_cpuload(cpu);
4541 return load_avg / nr_running;
4546 static void record_wakee(struct task_struct *p)
4549 * Rough decay (wiping) for cost saving, don't worry
4550 * about the boundary, really active task won't care
4553 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4554 current->wakee_flips >>= 1;
4555 current->wakee_flip_decay_ts = jiffies;
4558 if (current->last_wakee != p) {
4559 current->last_wakee = p;
4560 current->wakee_flips++;
4564 static void task_waking_fair(struct task_struct *p)
4566 struct sched_entity *se = &p->se;
4567 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4570 #ifndef CONFIG_64BIT
4571 u64 min_vruntime_copy;
4574 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4576 min_vruntime = cfs_rq->min_vruntime;
4577 } while (min_vruntime != min_vruntime_copy);
4579 min_vruntime = cfs_rq->min_vruntime;
4582 se->vruntime -= min_vruntime;
4586 #ifdef CONFIG_FAIR_GROUP_SCHED
4588 * effective_load() calculates the load change as seen from the root_task_group
4590 * Adding load to a group doesn't make a group heavier, but can cause movement
4591 * of group shares between cpus. Assuming the shares were perfectly aligned one
4592 * can calculate the shift in shares.
4594 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4595 * on this @cpu and results in a total addition (subtraction) of @wg to the
4596 * total group weight.
4598 * Given a runqueue weight distribution (rw_i) we can compute a shares
4599 * distribution (s_i) using:
4601 * s_i = rw_i / \Sum rw_j (1)
4603 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4604 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4605 * shares distribution (s_i):
4607 * rw_i = { 2, 4, 1, 0 }
4608 * s_i = { 2/7, 4/7, 1/7, 0 }
4610 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4611 * task used to run on and the CPU the waker is running on), we need to
4612 * compute the effect of waking a task on either CPU and, in case of a sync
4613 * wakeup, compute the effect of the current task going to sleep.
4615 * So for a change of @wl to the local @cpu with an overall group weight change
4616 * of @wl we can compute the new shares distribution (s'_i) using:
4618 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4620 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4621 * differences in waking a task to CPU 0. The additional task changes the
4622 * weight and shares distributions like:
4624 * rw'_i = { 3, 4, 1, 0 }
4625 * s'_i = { 3/8, 4/8, 1/8, 0 }
4627 * We can then compute the difference in effective weight by using:
4629 * dw_i = S * (s'_i - s_i) (3)
4631 * Where 'S' is the group weight as seen by its parent.
4633 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4634 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4635 * 4/7) times the weight of the group.
4637 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4639 struct sched_entity *se = tg->se[cpu];
4641 if (!tg->parent) /* the trivial, non-cgroup case */
4644 for_each_sched_entity(se) {
4650 * W = @wg + \Sum rw_j
4652 W = wg + calc_tg_weight(tg, se->my_q);
4657 w = cfs_rq_load_avg(se->my_q) + wl;
4660 * wl = S * s'_i; see (2)
4663 wl = (w * (long)tg->shares) / W;
4668 * Per the above, wl is the new se->load.weight value; since
4669 * those are clipped to [MIN_SHARES, ...) do so now. See
4670 * calc_cfs_shares().
4672 if (wl < MIN_SHARES)
4676 * wl = dw_i = S * (s'_i - s_i); see (3)
4678 wl -= se->avg.load_avg;
4681 * Recursively apply this logic to all parent groups to compute
4682 * the final effective load change on the root group. Since
4683 * only the @tg group gets extra weight, all parent groups can
4684 * only redistribute existing shares. @wl is the shift in shares
4685 * resulting from this level per the above.
4694 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4702 * Returns the current capacity of cpu after applying both
4703 * cpu and freq scaling.
4705 static unsigned long capacity_curr_of(int cpu)
4707 return cpu_rq(cpu)->cpu_capacity_orig *
4708 arch_scale_freq_capacity(NULL, cpu)
4709 >> SCHED_CAPACITY_SHIFT;
4713 * cpu_util returns the amount of capacity of a CPU that is used by CFS
4714 * tasks. The unit of the return value must be the one of capacity so we can
4715 * compare the utilization with the capacity of the CPU that is available for
4716 * CFS task (ie cpu_capacity).
4718 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
4719 * recent utilization of currently non-runnable tasks on a CPU. It represents
4720 * the amount of utilization of a CPU in the range [0..capacity_orig] where
4721 * capacity_orig is the cpu_capacity available at the highest frequency
4722 * (arch_scale_freq_capacity()).
4723 * The utilization of a CPU converges towards a sum equal to or less than the
4724 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
4725 * the running time on this CPU scaled by capacity_curr.
4727 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
4728 * higher than capacity_orig because of unfortunate rounding in
4729 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
4730 * the average stabilizes with the new running time. We need to check that the
4731 * utilization stays within the range of [0..capacity_orig] and cap it if
4732 * necessary. Without utilization capping, a group could be seen as overloaded
4733 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
4734 * available capacity. We allow utilization to overshoot capacity_curr (but not
4735 * capacity_orig) as it useful for predicting the capacity required after task
4736 * migrations (scheduler-driven DVFS).
4738 static unsigned long __cpu_util(int cpu, int delta)
4740 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
4741 unsigned long capacity = capacity_orig_of(cpu);
4747 return (delta >= capacity) ? capacity : delta;
4750 static unsigned long cpu_util(int cpu)
4752 return __cpu_util(cpu, 0);
4755 static inline bool energy_aware(void)
4757 return sched_feat(ENERGY_AWARE);
4761 struct sched_group *sg_top;
4762 struct sched_group *sg_cap;
4771 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4772 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4773 * energy calculations. Using the scale-invariant util returned by
4774 * cpu_util() and approximating scale-invariant util by:
4776 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4778 * the normalized util can be found using the specific capacity.
4780 * capacity = capacity_orig * curr_freq/max_freq
4782 * norm_util = running_time/time ~ util/capacity
4784 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4786 int util = __cpu_util(cpu, delta);
4788 if (util >= capacity)
4789 return SCHED_CAPACITY_SCALE;
4791 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4794 static int calc_util_delta(struct energy_env *eenv, int cpu)
4796 if (cpu == eenv->src_cpu)
4797 return -eenv->util_delta;
4798 if (cpu == eenv->dst_cpu)
4799 return eenv->util_delta;
4804 unsigned long group_max_util(struct energy_env *eenv)
4807 unsigned long max_util = 0;
4809 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4810 delta = calc_util_delta(eenv, i);
4811 max_util = max(max_util, __cpu_util(i, delta));
4818 * group_norm_util() returns the approximated group util relative to it's
4819 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4820 * energy calculations. Since task executions may or may not overlap in time in
4821 * the group the true normalized util is between max(cpu_norm_util(i)) and
4822 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4823 * latter is used as the estimate as it leads to a more pessimistic energy
4824 * estimate (more busy).
4827 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4830 unsigned long util_sum = 0;
4831 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4833 for_each_cpu(i, sched_group_cpus(sg)) {
4834 delta = calc_util_delta(eenv, i);
4835 util_sum += __cpu_norm_util(i, capacity, delta);
4838 if (util_sum > SCHED_CAPACITY_SCALE)
4839 return SCHED_CAPACITY_SCALE;
4843 static int find_new_capacity(struct energy_env *eenv,
4844 const struct sched_group_energy const *sge)
4847 unsigned long util = group_max_util(eenv);
4849 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4850 if (sge->cap_states[idx].cap >= util)
4854 eenv->cap_idx = idx;
4859 static int group_idle_state(struct sched_group *sg)
4861 int i, state = INT_MAX;
4863 /* Find the shallowest idle state in the sched group. */
4864 for_each_cpu(i, sched_group_cpus(sg))
4865 state = min(state, idle_get_state_idx(cpu_rq(i)));
4867 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4874 * sched_group_energy(): Computes the absolute energy consumption of cpus
4875 * belonging to the sched_group including shared resources shared only by
4876 * members of the group. Iterates over all cpus in the hierarchy below the
4877 * sched_group starting from the bottom working it's way up before going to
4878 * the next cpu until all cpus are covered at all levels. The current
4879 * implementation is likely to gather the same util statistics multiple times.
4880 * This can probably be done in a faster but more complex way.
4881 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4883 static int sched_group_energy(struct energy_env *eenv)
4885 struct sched_domain *sd;
4886 int cpu, total_energy = 0;
4887 struct cpumask visit_cpus;
4888 struct sched_group *sg;
4890 WARN_ON(!eenv->sg_top->sge);
4892 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4894 while (!cpumask_empty(&visit_cpus)) {
4895 struct sched_group *sg_shared_cap = NULL;
4897 cpu = cpumask_first(&visit_cpus);
4900 * Is the group utilization affected by cpus outside this
4903 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4907 * We most probably raced with hotplug; returning a
4908 * wrong energy estimation is better than entering an
4914 sg_shared_cap = sd->parent->groups;
4916 for_each_domain(cpu, sd) {
4919 /* Has this sched_domain already been visited? */
4920 if (sd->child && group_first_cpu(sg) != cpu)
4924 unsigned long group_util;
4925 int sg_busy_energy, sg_idle_energy;
4926 int cap_idx, idle_idx;
4928 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
4929 eenv->sg_cap = sg_shared_cap;
4933 cap_idx = find_new_capacity(eenv, sg->sge);
4934 idle_idx = group_idle_state(sg);
4935 group_util = group_norm_util(eenv, sg);
4936 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
4937 >> SCHED_CAPACITY_SHIFT;
4938 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
4939 * sg->sge->idle_states[idle_idx].power)
4940 >> SCHED_CAPACITY_SHIFT;
4942 total_energy += sg_busy_energy + sg_idle_energy;
4945 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
4947 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
4950 } while (sg = sg->next, sg != sd->groups);
4956 eenv->energy = total_energy;
4960 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
4962 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
4966 * energy_diff(): Estimate the energy impact of changing the utilization
4967 * distribution. eenv specifies the change: utilisation amount, source, and
4968 * destination cpu. Source or destination cpu may be -1 in which case the
4969 * utilization is removed from or added to the system (e.g. task wake-up). If
4970 * both are specified, the utilization is migrated.
4972 static int energy_diff(struct energy_env *eenv)
4974 struct sched_domain *sd;
4975 struct sched_group *sg;
4976 int sd_cpu = -1, energy_before = 0, energy_after = 0;
4978 struct energy_env eenv_before = {
4980 .src_cpu = eenv->src_cpu,
4981 .dst_cpu = eenv->dst_cpu,
4984 if (eenv->src_cpu == eenv->dst_cpu)
4987 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
4988 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
4991 return 0; /* Error */
4996 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
4997 eenv_before.sg_top = eenv->sg_top = sg;
4999 if (sched_group_energy(&eenv_before))
5000 return 0; /* Invalid result abort */
5001 energy_before += eenv_before.energy;
5003 if (sched_group_energy(eenv))
5004 return 0; /* Invalid result abort */
5005 energy_after += eenv->energy;
5007 } while (sg = sg->next, sg != sd->groups);
5009 return energy_after-energy_before;
5013 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5014 * A waker of many should wake a different task than the one last awakened
5015 * at a frequency roughly N times higher than one of its wakees. In order
5016 * to determine whether we should let the load spread vs consolodating to
5017 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5018 * partner, and a factor of lls_size higher frequency in the other. With
5019 * both conditions met, we can be relatively sure that the relationship is
5020 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5021 * being client/server, worker/dispatcher, interrupt source or whatever is
5022 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5024 static int wake_wide(struct task_struct *p)
5026 unsigned int master = current->wakee_flips;
5027 unsigned int slave = p->wakee_flips;
5028 int factor = this_cpu_read(sd_llc_size);
5031 swap(master, slave);
5032 if (slave < factor || master < slave * factor)
5037 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5039 s64 this_load, load;
5040 s64 this_eff_load, prev_eff_load;
5041 int idx, this_cpu, prev_cpu;
5042 struct task_group *tg;
5043 unsigned long weight;
5047 this_cpu = smp_processor_id();
5048 prev_cpu = task_cpu(p);
5049 load = source_load(prev_cpu, idx);
5050 this_load = target_load(this_cpu, idx);
5053 * If sync wakeup then subtract the (maximum possible)
5054 * effect of the currently running task from the load
5055 * of the current CPU:
5058 tg = task_group(current);
5059 weight = current->se.avg.load_avg;
5061 this_load += effective_load(tg, this_cpu, -weight, -weight);
5062 load += effective_load(tg, prev_cpu, 0, -weight);
5066 weight = p->se.avg.load_avg;
5069 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5070 * due to the sync cause above having dropped this_load to 0, we'll
5071 * always have an imbalance, but there's really nothing you can do
5072 * about that, so that's good too.
5074 * Otherwise check if either cpus are near enough in load to allow this
5075 * task to be woken on this_cpu.
5077 this_eff_load = 100;
5078 this_eff_load *= capacity_of(prev_cpu);
5080 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5081 prev_eff_load *= capacity_of(this_cpu);
5083 if (this_load > 0) {
5084 this_eff_load *= this_load +
5085 effective_load(tg, this_cpu, weight, weight);
5087 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5090 balanced = this_eff_load <= prev_eff_load;
5092 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5097 schedstat_inc(sd, ttwu_move_affine);
5098 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5103 static inline unsigned long task_util(struct task_struct *p)
5105 return p->se.avg.util_avg;
5108 unsigned int capacity_margin = 1280; /* ~20% margin */
5110 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5112 unsigned long capacity = capacity_of(cpu);
5114 util += task_util(p);
5116 return (capacity * 1024) > (util * capacity_margin);
5119 static inline bool task_fits_max(struct task_struct *p, int cpu)
5121 unsigned long capacity = capacity_of(cpu);
5122 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5124 if (capacity == max_capacity)
5127 if (capacity * capacity_margin > max_capacity * 1024)
5130 return __task_fits(p, cpu, 0);
5133 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5135 return __task_fits(p, cpu, cpu_util(cpu));
5138 static bool cpu_overutilized(int cpu)
5140 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5144 * find_idlest_group finds and returns the least busy CPU group within the
5147 static struct sched_group *
5148 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5149 int this_cpu, int sd_flag)
5151 struct sched_group *idlest = NULL, *group = sd->groups;
5152 struct sched_group *fit_group = NULL, *spare_group = NULL;
5153 unsigned long min_load = ULONG_MAX, this_load = 0;
5154 unsigned long fit_capacity = ULONG_MAX;
5155 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5156 int load_idx = sd->forkexec_idx;
5157 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5159 if (sd_flag & SD_BALANCE_WAKE)
5160 load_idx = sd->wake_idx;
5163 unsigned long load, avg_load, spare_capacity;
5167 /* Skip over this group if it has no CPUs allowed */
5168 if (!cpumask_intersects(sched_group_cpus(group),
5169 tsk_cpus_allowed(p)))
5172 local_group = cpumask_test_cpu(this_cpu,
5173 sched_group_cpus(group));
5175 /* Tally up the load of all CPUs in the group */
5178 for_each_cpu(i, sched_group_cpus(group)) {
5179 /* Bias balancing toward cpus of our domain */
5181 load = source_load(i, load_idx);
5183 load = target_load(i, load_idx);
5188 * Look for most energy-efficient group that can fit
5189 * that can fit the task.
5191 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5192 fit_capacity = capacity_of(i);
5197 * Look for group which has most spare capacity on a
5200 spare_capacity = capacity_of(i) - cpu_util(i);
5201 if (spare_capacity > max_spare_capacity) {
5202 max_spare_capacity = spare_capacity;
5203 spare_group = group;
5207 /* Adjust by relative CPU capacity of the group */
5208 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5211 this_load = avg_load;
5212 } else if (avg_load < min_load) {
5213 min_load = avg_load;
5216 } while (group = group->next, group != sd->groups);
5224 if (!idlest || 100*this_load < imbalance*min_load)
5230 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5233 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5235 unsigned long load, min_load = ULONG_MAX;
5236 unsigned int min_exit_latency = UINT_MAX;
5237 u64 latest_idle_timestamp = 0;
5238 int least_loaded_cpu = this_cpu;
5239 int shallowest_idle_cpu = -1;
5242 /* Traverse only the allowed CPUs */
5243 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5244 if (task_fits_spare(p, i)) {
5245 struct rq *rq = cpu_rq(i);
5246 struct cpuidle_state *idle = idle_get_state(rq);
5247 if (idle && idle->exit_latency < min_exit_latency) {
5249 * We give priority to a CPU whose idle state
5250 * has the smallest exit latency irrespective
5251 * of any idle timestamp.
5253 min_exit_latency = idle->exit_latency;
5254 latest_idle_timestamp = rq->idle_stamp;
5255 shallowest_idle_cpu = i;
5256 } else if (idle_cpu(i) &&
5257 (!idle || idle->exit_latency == min_exit_latency) &&
5258 rq->idle_stamp > latest_idle_timestamp) {
5260 * If equal or no active idle state, then
5261 * the most recently idled CPU might have
5264 latest_idle_timestamp = rq->idle_stamp;
5265 shallowest_idle_cpu = i;
5266 } else if (shallowest_idle_cpu == -1) {
5268 * If we haven't found an idle CPU yet
5269 * pick a non-idle one that can fit the task as
5272 shallowest_idle_cpu = i;
5274 } else if (shallowest_idle_cpu == -1) {
5275 load = weighted_cpuload(i);
5276 if (load < min_load || (load == min_load && i == this_cpu)) {
5278 least_loaded_cpu = i;
5283 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5287 * Try and locate an idle CPU in the sched_domain.
5289 static int select_idle_sibling(struct task_struct *p, int target)
5291 struct sched_domain *sd;
5292 struct sched_group *sg;
5293 int i = task_cpu(p);
5295 if (idle_cpu(target))
5299 * If the prevous cpu is cache affine and idle, don't be stupid.
5301 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5305 * Otherwise, iterate the domains and find an elegible idle cpu.
5307 sd = rcu_dereference(per_cpu(sd_llc, target));
5308 for_each_lower_domain(sd) {
5311 if (!cpumask_intersects(sched_group_cpus(sg),
5312 tsk_cpus_allowed(p)))
5315 for_each_cpu(i, sched_group_cpus(sg)) {
5316 if (i == target || !idle_cpu(i))
5320 target = cpumask_first_and(sched_group_cpus(sg),
5321 tsk_cpus_allowed(p));
5325 } while (sg != sd->groups);
5331 static int energy_aware_wake_cpu(struct task_struct *p, int target)
5333 struct sched_domain *sd;
5334 struct sched_group *sg, *sg_target;
5335 int target_max_cap = INT_MAX;
5336 int target_cpu = task_cpu(p);
5339 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5348 * Find group with sufficient capacity. We only get here if no cpu is
5349 * overutilized. We may end up overutilizing a cpu by adding the task,
5350 * but that should not be any worse than select_idle_sibling().
5351 * load_balance() should sort it out later as we get above the tipping
5355 /* Assuming all cpus are the same in group */
5356 int max_cap_cpu = group_first_cpu(sg);
5359 * Assume smaller max capacity means more energy-efficient.
5360 * Ideally we should query the energy model for the right
5361 * answer but it easily ends up in an exhaustive search.
5363 if (capacity_of(max_cap_cpu) < target_max_cap &&
5364 task_fits_max(p, max_cap_cpu)) {
5366 target_max_cap = capacity_of(max_cap_cpu);
5368 } while (sg = sg->next, sg != sd->groups);
5370 /* Find cpu with sufficient capacity */
5371 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5373 * p's blocked utilization is still accounted for on prev_cpu
5374 * so prev_cpu will receive a negative bias due to the double
5375 * accounting. However, the blocked utilization may be zero.
5377 int new_util = cpu_util(i) + task_util(p);
5379 if (new_util > capacity_orig_of(i))
5382 if (new_util < capacity_curr_of(i)) {
5384 if (cpu_rq(i)->nr_running)
5388 /* cpu has capacity at higher OPP, keep it as fallback */
5389 if (target_cpu == task_cpu(p))
5393 if (target_cpu != task_cpu(p)) {
5394 struct energy_env eenv = {
5395 .util_delta = task_util(p),
5396 .src_cpu = task_cpu(p),
5397 .dst_cpu = target_cpu,
5400 /* Not enough spare capacity on previous cpu */
5401 if (cpu_overutilized(task_cpu(p)))
5404 if (energy_diff(&eenv) >= 0)
5412 * select_task_rq_fair: Select target runqueue for the waking task in domains
5413 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5414 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5416 * Balances load by selecting the idlest cpu in the idlest group, or under
5417 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5419 * Returns the target cpu number.
5421 * preempt must be disabled.
5424 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5426 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5427 int cpu = smp_processor_id();
5428 int new_cpu = prev_cpu;
5429 int want_affine = 0;
5430 int sync = wake_flags & WF_SYNC;
5432 if (sd_flag & SD_BALANCE_WAKE)
5433 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5434 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5438 for_each_domain(cpu, tmp) {
5439 if (!(tmp->flags & SD_LOAD_BALANCE))
5443 * If both cpu and prev_cpu are part of this domain,
5444 * cpu is a valid SD_WAKE_AFFINE target.
5446 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5447 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5452 if (tmp->flags & sd_flag)
5454 else if (!want_affine)
5459 sd = NULL; /* Prefer wake_affine over balance flags */
5460 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5465 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5466 new_cpu = energy_aware_wake_cpu(p, prev_cpu);
5467 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5468 new_cpu = select_idle_sibling(p, new_cpu);
5471 struct sched_group *group;
5474 if (!(sd->flags & sd_flag)) {
5479 group = find_idlest_group(sd, p, cpu, sd_flag);
5485 new_cpu = find_idlest_cpu(group, p, cpu);
5486 if (new_cpu == -1 || new_cpu == cpu) {
5487 /* Now try balancing at a lower domain level of cpu */
5492 /* Now try balancing at a lower domain level of new_cpu */
5494 weight = sd->span_weight;
5496 for_each_domain(cpu, tmp) {
5497 if (weight <= tmp->span_weight)
5499 if (tmp->flags & sd_flag)
5502 /* while loop will break here if sd == NULL */
5510 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5511 * cfs_rq_of(p) references at time of call are still valid and identify the
5512 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5513 * other assumptions, including the state of rq->lock, should be made.
5515 static void migrate_task_rq_fair(struct task_struct *p)
5518 * We are supposed to update the task to "current" time, then its up to date
5519 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5520 * what current time is, so simply throw away the out-of-date time. This
5521 * will result in the wakee task is less decayed, but giving the wakee more
5522 * load sounds not bad.
5524 remove_entity_load_avg(&p->se);
5526 /* Tell new CPU we are migrated */
5527 p->se.avg.last_update_time = 0;
5529 /* We have migrated, no longer consider this task hot */
5530 p->se.exec_start = 0;
5533 static void task_dead_fair(struct task_struct *p)
5535 remove_entity_load_avg(&p->se);
5537 #endif /* CONFIG_SMP */
5539 static unsigned long
5540 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5542 unsigned long gran = sysctl_sched_wakeup_granularity;
5545 * Since its curr running now, convert the gran from real-time
5546 * to virtual-time in his units.
5548 * By using 'se' instead of 'curr' we penalize light tasks, so
5549 * they get preempted easier. That is, if 'se' < 'curr' then
5550 * the resulting gran will be larger, therefore penalizing the
5551 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5552 * be smaller, again penalizing the lighter task.
5554 * This is especially important for buddies when the leftmost
5555 * task is higher priority than the buddy.
5557 return calc_delta_fair(gran, se);
5561 * Should 'se' preempt 'curr'.
5575 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5577 s64 gran, vdiff = curr->vruntime - se->vruntime;
5582 gran = wakeup_gran(curr, se);
5589 static void set_last_buddy(struct sched_entity *se)
5591 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5594 for_each_sched_entity(se)
5595 cfs_rq_of(se)->last = se;
5598 static void set_next_buddy(struct sched_entity *se)
5600 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5603 for_each_sched_entity(se)
5604 cfs_rq_of(se)->next = se;
5607 static void set_skip_buddy(struct sched_entity *se)
5609 for_each_sched_entity(se)
5610 cfs_rq_of(se)->skip = se;
5614 * Preempt the current task with a newly woken task if needed:
5616 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5618 struct task_struct *curr = rq->curr;
5619 struct sched_entity *se = &curr->se, *pse = &p->se;
5620 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5621 int scale = cfs_rq->nr_running >= sched_nr_latency;
5622 int next_buddy_marked = 0;
5624 if (unlikely(se == pse))
5628 * This is possible from callers such as attach_tasks(), in which we
5629 * unconditionally check_prempt_curr() after an enqueue (which may have
5630 * lead to a throttle). This both saves work and prevents false
5631 * next-buddy nomination below.
5633 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5636 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5637 set_next_buddy(pse);
5638 next_buddy_marked = 1;
5642 * We can come here with TIF_NEED_RESCHED already set from new task
5645 * Note: this also catches the edge-case of curr being in a throttled
5646 * group (e.g. via set_curr_task), since update_curr() (in the
5647 * enqueue of curr) will have resulted in resched being set. This
5648 * prevents us from potentially nominating it as a false LAST_BUDDY
5651 if (test_tsk_need_resched(curr))
5654 /* Idle tasks are by definition preempted by non-idle tasks. */
5655 if (unlikely(curr->policy == SCHED_IDLE) &&
5656 likely(p->policy != SCHED_IDLE))
5660 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5661 * is driven by the tick):
5663 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5666 find_matching_se(&se, &pse);
5667 update_curr(cfs_rq_of(se));
5669 if (wakeup_preempt_entity(se, pse) == 1) {
5671 * Bias pick_next to pick the sched entity that is
5672 * triggering this preemption.
5674 if (!next_buddy_marked)
5675 set_next_buddy(pse);
5684 * Only set the backward buddy when the current task is still
5685 * on the rq. This can happen when a wakeup gets interleaved
5686 * with schedule on the ->pre_schedule() or idle_balance()
5687 * point, either of which can * drop the rq lock.
5689 * Also, during early boot the idle thread is in the fair class,
5690 * for obvious reasons its a bad idea to schedule back to it.
5692 if (unlikely(!se->on_rq || curr == rq->idle))
5695 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5699 static struct task_struct *
5700 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5702 struct cfs_rq *cfs_rq = &rq->cfs;
5703 struct sched_entity *se;
5704 struct task_struct *p;
5708 #ifdef CONFIG_FAIR_GROUP_SCHED
5709 if (!cfs_rq->nr_running)
5712 if (prev->sched_class != &fair_sched_class)
5716 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5717 * likely that a next task is from the same cgroup as the current.
5719 * Therefore attempt to avoid putting and setting the entire cgroup
5720 * hierarchy, only change the part that actually changes.
5724 struct sched_entity *curr = cfs_rq->curr;
5727 * Since we got here without doing put_prev_entity() we also
5728 * have to consider cfs_rq->curr. If it is still a runnable
5729 * entity, update_curr() will update its vruntime, otherwise
5730 * forget we've ever seen it.
5734 update_curr(cfs_rq);
5739 * This call to check_cfs_rq_runtime() will do the
5740 * throttle and dequeue its entity in the parent(s).
5741 * Therefore the 'simple' nr_running test will indeed
5744 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5748 se = pick_next_entity(cfs_rq, curr);
5749 cfs_rq = group_cfs_rq(se);
5755 * Since we haven't yet done put_prev_entity and if the selected task
5756 * is a different task than we started out with, try and touch the
5757 * least amount of cfs_rqs.
5760 struct sched_entity *pse = &prev->se;
5762 while (!(cfs_rq = is_same_group(se, pse))) {
5763 int se_depth = se->depth;
5764 int pse_depth = pse->depth;
5766 if (se_depth <= pse_depth) {
5767 put_prev_entity(cfs_rq_of(pse), pse);
5768 pse = parent_entity(pse);
5770 if (se_depth >= pse_depth) {
5771 set_next_entity(cfs_rq_of(se), se);
5772 se = parent_entity(se);
5776 put_prev_entity(cfs_rq, pse);
5777 set_next_entity(cfs_rq, se);
5780 if (hrtick_enabled(rq))
5781 hrtick_start_fair(rq, p);
5783 rq->misfit_task = !task_fits_max(p, rq->cpu);
5790 if (!cfs_rq->nr_running)
5793 put_prev_task(rq, prev);
5796 se = pick_next_entity(cfs_rq, NULL);
5797 set_next_entity(cfs_rq, se);
5798 cfs_rq = group_cfs_rq(se);
5803 if (hrtick_enabled(rq))
5804 hrtick_start_fair(rq, p);
5806 rq->misfit_task = !task_fits_max(p, rq->cpu);
5811 rq->misfit_task = 0;
5813 * This is OK, because current is on_cpu, which avoids it being picked
5814 * for load-balance and preemption/IRQs are still disabled avoiding
5815 * further scheduler activity on it and we're being very careful to
5816 * re-start the picking loop.
5818 lockdep_unpin_lock(&rq->lock);
5819 new_tasks = idle_balance(rq);
5820 lockdep_pin_lock(&rq->lock);
5822 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5823 * possible for any higher priority task to appear. In that case we
5824 * must re-start the pick_next_entity() loop.
5836 * Account for a descheduled task:
5838 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5840 struct sched_entity *se = &prev->se;
5841 struct cfs_rq *cfs_rq;
5843 for_each_sched_entity(se) {
5844 cfs_rq = cfs_rq_of(se);
5845 put_prev_entity(cfs_rq, se);
5850 * sched_yield() is very simple
5852 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5854 static void yield_task_fair(struct rq *rq)
5856 struct task_struct *curr = rq->curr;
5857 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5858 struct sched_entity *se = &curr->se;
5861 * Are we the only task in the tree?
5863 if (unlikely(rq->nr_running == 1))
5866 clear_buddies(cfs_rq, se);
5868 if (curr->policy != SCHED_BATCH) {
5869 update_rq_clock(rq);
5871 * Update run-time statistics of the 'current'.
5873 update_curr(cfs_rq);
5875 * Tell update_rq_clock() that we've just updated,
5876 * so we don't do microscopic update in schedule()
5877 * and double the fastpath cost.
5879 rq_clock_skip_update(rq, true);
5885 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5887 struct sched_entity *se = &p->se;
5889 /* throttled hierarchies are not runnable */
5890 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5893 /* Tell the scheduler that we'd really like pse to run next. */
5896 yield_task_fair(rq);
5902 /**************************************************
5903 * Fair scheduling class load-balancing methods.
5907 * The purpose of load-balancing is to achieve the same basic fairness the
5908 * per-cpu scheduler provides, namely provide a proportional amount of compute
5909 * time to each task. This is expressed in the following equation:
5911 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5913 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5914 * W_i,0 is defined as:
5916 * W_i,0 = \Sum_j w_i,j (2)
5918 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5919 * is derived from the nice value as per prio_to_weight[].
5921 * The weight average is an exponential decay average of the instantaneous
5924 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5926 * C_i is the compute capacity of cpu i, typically it is the
5927 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5928 * can also include other factors [XXX].
5930 * To achieve this balance we define a measure of imbalance which follows
5931 * directly from (1):
5933 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5935 * We them move tasks around to minimize the imbalance. In the continuous
5936 * function space it is obvious this converges, in the discrete case we get
5937 * a few fun cases generally called infeasible weight scenarios.
5940 * - infeasible weights;
5941 * - local vs global optima in the discrete case. ]
5946 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5947 * for all i,j solution, we create a tree of cpus that follows the hardware
5948 * topology where each level pairs two lower groups (or better). This results
5949 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5950 * tree to only the first of the previous level and we decrease the frequency
5951 * of load-balance at each level inv. proportional to the number of cpus in
5957 * \Sum { --- * --- * 2^i } = O(n) (5)
5959 * `- size of each group
5960 * | | `- number of cpus doing load-balance
5962 * `- sum over all levels
5964 * Coupled with a limit on how many tasks we can migrate every balance pass,
5965 * this makes (5) the runtime complexity of the balancer.
5967 * An important property here is that each CPU is still (indirectly) connected
5968 * to every other cpu in at most O(log n) steps:
5970 * The adjacency matrix of the resulting graph is given by:
5973 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5976 * And you'll find that:
5978 * A^(log_2 n)_i,j != 0 for all i,j (7)
5980 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5981 * The task movement gives a factor of O(m), giving a convergence complexity
5984 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5989 * In order to avoid CPUs going idle while there's still work to do, new idle
5990 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5991 * tree itself instead of relying on other CPUs to bring it work.
5993 * This adds some complexity to both (5) and (8) but it reduces the total idle
6001 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6004 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6009 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6011 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6013 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6016 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6017 * rewrite all of this once again.]
6020 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6022 enum fbq_type { regular, remote, all };
6031 #define LBF_ALL_PINNED 0x01
6032 #define LBF_NEED_BREAK 0x02
6033 #define LBF_DST_PINNED 0x04
6034 #define LBF_SOME_PINNED 0x08
6037 struct sched_domain *sd;
6045 struct cpumask *dst_grpmask;
6047 enum cpu_idle_type idle;
6049 unsigned int src_grp_nr_running;
6050 /* The set of CPUs under consideration for load-balancing */
6051 struct cpumask *cpus;
6056 unsigned int loop_break;
6057 unsigned int loop_max;
6059 enum fbq_type fbq_type;
6060 enum group_type busiest_group_type;
6061 struct list_head tasks;
6065 * Is this task likely cache-hot:
6067 static int task_hot(struct task_struct *p, struct lb_env *env)
6071 lockdep_assert_held(&env->src_rq->lock);
6073 if (p->sched_class != &fair_sched_class)
6076 if (unlikely(p->policy == SCHED_IDLE))
6080 * Buddy candidates are cache hot:
6082 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6083 (&p->se == cfs_rq_of(&p->se)->next ||
6084 &p->se == cfs_rq_of(&p->se)->last))
6087 if (sysctl_sched_migration_cost == -1)
6089 if (sysctl_sched_migration_cost == 0)
6092 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6094 return delta < (s64)sysctl_sched_migration_cost;
6097 #ifdef CONFIG_NUMA_BALANCING
6099 * Returns 1, if task migration degrades locality
6100 * Returns 0, if task migration improves locality i.e migration preferred.
6101 * Returns -1, if task migration is not affected by locality.
6103 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6105 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6106 unsigned long src_faults, dst_faults;
6107 int src_nid, dst_nid;
6109 if (!static_branch_likely(&sched_numa_balancing))
6112 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6115 src_nid = cpu_to_node(env->src_cpu);
6116 dst_nid = cpu_to_node(env->dst_cpu);
6118 if (src_nid == dst_nid)
6121 /* Migrating away from the preferred node is always bad. */
6122 if (src_nid == p->numa_preferred_nid) {
6123 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6129 /* Encourage migration to the preferred node. */
6130 if (dst_nid == p->numa_preferred_nid)
6134 src_faults = group_faults(p, src_nid);
6135 dst_faults = group_faults(p, dst_nid);
6137 src_faults = task_faults(p, src_nid);
6138 dst_faults = task_faults(p, dst_nid);
6141 return dst_faults < src_faults;
6145 static inline int migrate_degrades_locality(struct task_struct *p,
6153 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6156 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6160 lockdep_assert_held(&env->src_rq->lock);
6163 * We do not migrate tasks that are:
6164 * 1) throttled_lb_pair, or
6165 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6166 * 3) running (obviously), or
6167 * 4) are cache-hot on their current CPU.
6169 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6172 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6175 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6177 env->flags |= LBF_SOME_PINNED;
6180 * Remember if this task can be migrated to any other cpu in
6181 * our sched_group. We may want to revisit it if we couldn't
6182 * meet load balance goals by pulling other tasks on src_cpu.
6184 * Also avoid computing new_dst_cpu if we have already computed
6185 * one in current iteration.
6187 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6190 /* Prevent to re-select dst_cpu via env's cpus */
6191 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6192 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6193 env->flags |= LBF_DST_PINNED;
6194 env->new_dst_cpu = cpu;
6202 /* Record that we found atleast one task that could run on dst_cpu */
6203 env->flags &= ~LBF_ALL_PINNED;
6205 if (task_running(env->src_rq, p)) {
6206 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6211 * Aggressive migration if:
6212 * 1) destination numa is preferred
6213 * 2) task is cache cold, or
6214 * 3) too many balance attempts have failed.
6216 tsk_cache_hot = migrate_degrades_locality(p, env);
6217 if (tsk_cache_hot == -1)
6218 tsk_cache_hot = task_hot(p, env);
6220 if (tsk_cache_hot <= 0 ||
6221 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6222 if (tsk_cache_hot == 1) {
6223 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6224 schedstat_inc(p, se.statistics.nr_forced_migrations);
6229 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6234 * detach_task() -- detach the task for the migration specified in env
6236 static void detach_task(struct task_struct *p, struct lb_env *env)
6238 lockdep_assert_held(&env->src_rq->lock);
6240 deactivate_task(env->src_rq, p, 0);
6241 p->on_rq = TASK_ON_RQ_MIGRATING;
6242 set_task_cpu(p, env->dst_cpu);
6246 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6247 * part of active balancing operations within "domain".
6249 * Returns a task if successful and NULL otherwise.
6251 static struct task_struct *detach_one_task(struct lb_env *env)
6253 struct task_struct *p, *n;
6255 lockdep_assert_held(&env->src_rq->lock);
6257 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6258 if (!can_migrate_task(p, env))
6261 detach_task(p, env);
6264 * Right now, this is only the second place where
6265 * lb_gained[env->idle] is updated (other is detach_tasks)
6266 * so we can safely collect stats here rather than
6267 * inside detach_tasks().
6269 schedstat_inc(env->sd, lb_gained[env->idle]);
6275 static const unsigned int sched_nr_migrate_break = 32;
6278 * detach_tasks() -- tries to detach up to imbalance weighted load from
6279 * busiest_rq, as part of a balancing operation within domain "sd".
6281 * Returns number of detached tasks if successful and 0 otherwise.
6283 static int detach_tasks(struct lb_env *env)
6285 struct list_head *tasks = &env->src_rq->cfs_tasks;
6286 struct task_struct *p;
6290 lockdep_assert_held(&env->src_rq->lock);
6292 if (env->imbalance <= 0)
6295 while (!list_empty(tasks)) {
6297 * We don't want to steal all, otherwise we may be treated likewise,
6298 * which could at worst lead to a livelock crash.
6300 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6303 p = list_first_entry(tasks, struct task_struct, se.group_node);
6306 /* We've more or less seen every task there is, call it quits */
6307 if (env->loop > env->loop_max)
6310 /* take a breather every nr_migrate tasks */
6311 if (env->loop > env->loop_break) {
6312 env->loop_break += sched_nr_migrate_break;
6313 env->flags |= LBF_NEED_BREAK;
6317 if (!can_migrate_task(p, env))
6320 load = task_h_load(p);
6322 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6325 if ((load / 2) > env->imbalance)
6328 detach_task(p, env);
6329 list_add(&p->se.group_node, &env->tasks);
6332 env->imbalance -= load;
6334 #ifdef CONFIG_PREEMPT
6336 * NEWIDLE balancing is a source of latency, so preemptible
6337 * kernels will stop after the first task is detached to minimize
6338 * the critical section.
6340 if (env->idle == CPU_NEWLY_IDLE)
6345 * We only want to steal up to the prescribed amount of
6348 if (env->imbalance <= 0)
6353 list_move_tail(&p->se.group_node, tasks);
6357 * Right now, this is one of only two places we collect this stat
6358 * so we can safely collect detach_one_task() stats here rather
6359 * than inside detach_one_task().
6361 schedstat_add(env->sd, lb_gained[env->idle], detached);
6367 * attach_task() -- attach the task detached by detach_task() to its new rq.
6369 static void attach_task(struct rq *rq, struct task_struct *p)
6371 lockdep_assert_held(&rq->lock);
6373 BUG_ON(task_rq(p) != rq);
6374 p->on_rq = TASK_ON_RQ_QUEUED;
6375 activate_task(rq, p, 0);
6376 check_preempt_curr(rq, p, 0);
6380 * attach_one_task() -- attaches the task returned from detach_one_task() to
6383 static void attach_one_task(struct rq *rq, struct task_struct *p)
6385 raw_spin_lock(&rq->lock);
6388 * We want to potentially raise target_cpu's OPP.
6390 update_capacity_of(cpu_of(rq));
6391 raw_spin_unlock(&rq->lock);
6395 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6398 static void attach_tasks(struct lb_env *env)
6400 struct list_head *tasks = &env->tasks;
6401 struct task_struct *p;
6403 raw_spin_lock(&env->dst_rq->lock);
6405 while (!list_empty(tasks)) {
6406 p = list_first_entry(tasks, struct task_struct, se.group_node);
6407 list_del_init(&p->se.group_node);
6409 attach_task(env->dst_rq, p);
6413 * We want to potentially raise env.dst_cpu's OPP.
6415 update_capacity_of(env->dst_cpu);
6417 raw_spin_unlock(&env->dst_rq->lock);
6420 #ifdef CONFIG_FAIR_GROUP_SCHED
6421 static void update_blocked_averages(int cpu)
6423 struct rq *rq = cpu_rq(cpu);
6424 struct cfs_rq *cfs_rq;
6425 unsigned long flags;
6427 raw_spin_lock_irqsave(&rq->lock, flags);
6428 update_rq_clock(rq);
6431 * Iterates the task_group tree in a bottom up fashion, see
6432 * list_add_leaf_cfs_rq() for details.
6434 for_each_leaf_cfs_rq(rq, cfs_rq) {
6435 /* throttled entities do not contribute to load */
6436 if (throttled_hierarchy(cfs_rq))
6439 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6440 update_tg_load_avg(cfs_rq, 0);
6442 raw_spin_unlock_irqrestore(&rq->lock, flags);
6446 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6447 * This needs to be done in a top-down fashion because the load of a child
6448 * group is a fraction of its parents load.
6450 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6452 struct rq *rq = rq_of(cfs_rq);
6453 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6454 unsigned long now = jiffies;
6457 if (cfs_rq->last_h_load_update == now)
6460 cfs_rq->h_load_next = NULL;
6461 for_each_sched_entity(se) {
6462 cfs_rq = cfs_rq_of(se);
6463 cfs_rq->h_load_next = se;
6464 if (cfs_rq->last_h_load_update == now)
6469 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6470 cfs_rq->last_h_load_update = now;
6473 while ((se = cfs_rq->h_load_next) != NULL) {
6474 load = cfs_rq->h_load;
6475 load = div64_ul(load * se->avg.load_avg,
6476 cfs_rq_load_avg(cfs_rq) + 1);
6477 cfs_rq = group_cfs_rq(se);
6478 cfs_rq->h_load = load;
6479 cfs_rq->last_h_load_update = now;
6483 static unsigned long task_h_load(struct task_struct *p)
6485 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6487 update_cfs_rq_h_load(cfs_rq);
6488 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6489 cfs_rq_load_avg(cfs_rq) + 1);
6492 static inline void update_blocked_averages(int cpu)
6494 struct rq *rq = cpu_rq(cpu);
6495 struct cfs_rq *cfs_rq = &rq->cfs;
6496 unsigned long flags;
6498 raw_spin_lock_irqsave(&rq->lock, flags);
6499 update_rq_clock(rq);
6500 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6501 raw_spin_unlock_irqrestore(&rq->lock, flags);
6504 static unsigned long task_h_load(struct task_struct *p)
6506 return p->se.avg.load_avg;
6510 /********** Helpers for find_busiest_group ************************/
6513 * sg_lb_stats - stats of a sched_group required for load_balancing
6515 struct sg_lb_stats {
6516 unsigned long avg_load; /*Avg load across the CPUs of the group */
6517 unsigned long group_load; /* Total load over the CPUs of the group */
6518 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6519 unsigned long load_per_task;
6520 unsigned long group_capacity;
6521 unsigned long group_util; /* Total utilization of the group */
6522 unsigned int sum_nr_running; /* Nr tasks running in the group */
6523 unsigned int idle_cpus;
6524 unsigned int group_weight;
6525 enum group_type group_type;
6526 int group_no_capacity;
6527 int group_misfit_task; /* A cpu has a task too big for its capacity */
6528 #ifdef CONFIG_NUMA_BALANCING
6529 unsigned int nr_numa_running;
6530 unsigned int nr_preferred_running;
6535 * sd_lb_stats - Structure to store the statistics of a sched_domain
6536 * during load balancing.
6538 struct sd_lb_stats {
6539 struct sched_group *busiest; /* Busiest group in this sd */
6540 struct sched_group *local; /* Local group in this sd */
6541 unsigned long total_load; /* Total load of all groups in sd */
6542 unsigned long total_capacity; /* Total capacity of all groups in sd */
6543 unsigned long avg_load; /* Average load across all groups in sd */
6545 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6546 struct sg_lb_stats local_stat; /* Statistics of the local group */
6549 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6552 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6553 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6554 * We must however clear busiest_stat::avg_load because
6555 * update_sd_pick_busiest() reads this before assignment.
6557 *sds = (struct sd_lb_stats){
6561 .total_capacity = 0UL,
6564 .sum_nr_running = 0,
6565 .group_type = group_other,
6571 * get_sd_load_idx - Obtain the load index for a given sched domain.
6572 * @sd: The sched_domain whose load_idx is to be obtained.
6573 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6575 * Return: The load index.
6577 static inline int get_sd_load_idx(struct sched_domain *sd,
6578 enum cpu_idle_type idle)
6584 load_idx = sd->busy_idx;
6587 case CPU_NEWLY_IDLE:
6588 load_idx = sd->newidle_idx;
6591 load_idx = sd->idle_idx;
6598 static unsigned long scale_rt_capacity(int cpu)
6600 struct rq *rq = cpu_rq(cpu);
6601 u64 total, used, age_stamp, avg;
6605 * Since we're reading these variables without serialization make sure
6606 * we read them once before doing sanity checks on them.
6608 age_stamp = READ_ONCE(rq->age_stamp);
6609 avg = READ_ONCE(rq->rt_avg);
6610 delta = __rq_clock_broken(rq) - age_stamp;
6612 if (unlikely(delta < 0))
6615 total = sched_avg_period() + delta;
6617 used = div_u64(avg, total);
6619 if (likely(used < SCHED_CAPACITY_SCALE))
6620 return SCHED_CAPACITY_SCALE - used;
6625 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
6627 raw_spin_lock_init(&mcc->lock);
6632 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6634 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6635 struct sched_group *sdg = sd->groups;
6636 struct max_cpu_capacity *mcc;
6637 unsigned long max_capacity;
6639 unsigned long flags;
6641 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6643 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
6645 raw_spin_lock_irqsave(&mcc->lock, flags);
6646 max_capacity = mcc->val;
6647 max_cap_cpu = mcc->cpu;
6649 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
6650 (max_capacity < capacity)) {
6651 mcc->val = capacity;
6653 #ifdef CONFIG_SCHED_DEBUG
6654 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6655 pr_info("CPU%d: update max cpu_capacity %lu\n", cpu, capacity);
6659 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6661 skip_unlock: __attribute__ ((unused));
6662 capacity *= scale_rt_capacity(cpu);
6663 capacity >>= SCHED_CAPACITY_SHIFT;
6668 cpu_rq(cpu)->cpu_capacity = capacity;
6669 sdg->sgc->capacity = capacity;
6670 sdg->sgc->max_capacity = capacity;
6673 void update_group_capacity(struct sched_domain *sd, int cpu)
6675 struct sched_domain *child = sd->child;
6676 struct sched_group *group, *sdg = sd->groups;
6677 unsigned long capacity, max_capacity;
6678 unsigned long interval;
6680 interval = msecs_to_jiffies(sd->balance_interval);
6681 interval = clamp(interval, 1UL, max_load_balance_interval);
6682 sdg->sgc->next_update = jiffies + interval;
6685 update_cpu_capacity(sd, cpu);
6692 if (child->flags & SD_OVERLAP) {
6694 * SD_OVERLAP domains cannot assume that child groups
6695 * span the current group.
6698 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6699 struct sched_group_capacity *sgc;
6700 struct rq *rq = cpu_rq(cpu);
6703 * build_sched_domains() -> init_sched_groups_capacity()
6704 * gets here before we've attached the domains to the
6707 * Use capacity_of(), which is set irrespective of domains
6708 * in update_cpu_capacity().
6710 * This avoids capacity from being 0 and
6711 * causing divide-by-zero issues on boot.
6713 if (unlikely(!rq->sd)) {
6714 capacity += capacity_of(cpu);
6716 sgc = rq->sd->groups->sgc;
6717 capacity += sgc->capacity;
6720 max_capacity = max(capacity, max_capacity);
6724 * !SD_OVERLAP domains can assume that child groups
6725 * span the current group.
6728 group = child->groups;
6730 struct sched_group_capacity *sgc = group->sgc;
6732 capacity += sgc->capacity;
6733 max_capacity = max(sgc->max_capacity, max_capacity);
6734 group = group->next;
6735 } while (group != child->groups);
6738 sdg->sgc->capacity = capacity;
6739 sdg->sgc->max_capacity = max_capacity;
6743 * Check whether the capacity of the rq has been noticeably reduced by side
6744 * activity. The imbalance_pct is used for the threshold.
6745 * Return true is the capacity is reduced
6748 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6750 return ((rq->cpu_capacity * sd->imbalance_pct) <
6751 (rq->cpu_capacity_orig * 100));
6755 * Group imbalance indicates (and tries to solve) the problem where balancing
6756 * groups is inadequate due to tsk_cpus_allowed() constraints.
6758 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6759 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6762 * { 0 1 2 3 } { 4 5 6 7 }
6765 * If we were to balance group-wise we'd place two tasks in the first group and
6766 * two tasks in the second group. Clearly this is undesired as it will overload
6767 * cpu 3 and leave one of the cpus in the second group unused.
6769 * The current solution to this issue is detecting the skew in the first group
6770 * by noticing the lower domain failed to reach balance and had difficulty
6771 * moving tasks due to affinity constraints.
6773 * When this is so detected; this group becomes a candidate for busiest; see
6774 * update_sd_pick_busiest(). And calculate_imbalance() and
6775 * find_busiest_group() avoid some of the usual balance conditions to allow it
6776 * to create an effective group imbalance.
6778 * This is a somewhat tricky proposition since the next run might not find the
6779 * group imbalance and decide the groups need to be balanced again. A most
6780 * subtle and fragile situation.
6783 static inline int sg_imbalanced(struct sched_group *group)
6785 return group->sgc->imbalance;
6789 * group_has_capacity returns true if the group has spare capacity that could
6790 * be used by some tasks.
6791 * We consider that a group has spare capacity if the * number of task is
6792 * smaller than the number of CPUs or if the utilization is lower than the
6793 * available capacity for CFS tasks.
6794 * For the latter, we use a threshold to stabilize the state, to take into
6795 * account the variance of the tasks' load and to return true if the available
6796 * capacity in meaningful for the load balancer.
6797 * As an example, an available capacity of 1% can appear but it doesn't make
6798 * any benefit for the load balance.
6801 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6803 if (sgs->sum_nr_running < sgs->group_weight)
6806 if ((sgs->group_capacity * 100) >
6807 (sgs->group_util * env->sd->imbalance_pct))
6814 * group_is_overloaded returns true if the group has more tasks than it can
6816 * group_is_overloaded is not equals to !group_has_capacity because a group
6817 * with the exact right number of tasks, has no more spare capacity but is not
6818 * overloaded so both group_has_capacity and group_is_overloaded return
6822 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6824 if (sgs->sum_nr_running <= sgs->group_weight)
6827 if ((sgs->group_capacity * 100) <
6828 (sgs->group_util * env->sd->imbalance_pct))
6836 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
6837 * per-cpu capacity than sched_group ref.
6840 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
6842 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
6843 ref->sgc->max_capacity;
6847 group_type group_classify(struct sched_group *group,
6848 struct sg_lb_stats *sgs)
6850 if (sgs->group_no_capacity)
6851 return group_overloaded;
6853 if (sg_imbalanced(group))
6854 return group_imbalanced;
6856 if (sgs->group_misfit_task)
6857 return group_misfit_task;
6863 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6864 * @env: The load balancing environment.
6865 * @group: sched_group whose statistics are to be updated.
6866 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6867 * @local_group: Does group contain this_cpu.
6868 * @sgs: variable to hold the statistics for this group.
6869 * @overload: Indicate more than one runnable task for any CPU.
6870 * @overutilized: Indicate overutilization for any CPU.
6872 static inline void update_sg_lb_stats(struct lb_env *env,
6873 struct sched_group *group, int load_idx,
6874 int local_group, struct sg_lb_stats *sgs,
6875 bool *overload, bool *overutilized)
6880 memset(sgs, 0, sizeof(*sgs));
6882 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6883 struct rq *rq = cpu_rq(i);
6885 /* Bias balancing toward cpus of our domain */
6887 load = target_load(i, load_idx);
6889 load = source_load(i, load_idx);
6891 sgs->group_load += load;
6892 sgs->group_util += cpu_util(i);
6893 sgs->sum_nr_running += rq->cfs.h_nr_running;
6895 if (rq->nr_running > 1)
6898 #ifdef CONFIG_NUMA_BALANCING
6899 sgs->nr_numa_running += rq->nr_numa_running;
6900 sgs->nr_preferred_running += rq->nr_preferred_running;
6902 sgs->sum_weighted_load += weighted_cpuload(i);
6906 if (cpu_overutilized(i)) {
6907 *overutilized = true;
6908 if (!sgs->group_misfit_task && rq->misfit_task)
6909 sgs->group_misfit_task = capacity_of(i);
6913 /* Adjust by relative CPU capacity of the group */
6914 sgs->group_capacity = group->sgc->capacity;
6915 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6917 if (sgs->sum_nr_running)
6918 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6920 sgs->group_weight = group->group_weight;
6922 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6923 sgs->group_type = group_classify(group, sgs);
6927 * update_sd_pick_busiest - return 1 on busiest group
6928 * @env: The load balancing environment.
6929 * @sds: sched_domain statistics
6930 * @sg: sched_group candidate to be checked for being the busiest
6931 * @sgs: sched_group statistics
6933 * Determine if @sg is a busier group than the previously selected
6936 * Return: %true if @sg is a busier group than the previously selected
6937 * busiest group. %false otherwise.
6939 static bool update_sd_pick_busiest(struct lb_env *env,
6940 struct sd_lb_stats *sds,
6941 struct sched_group *sg,
6942 struct sg_lb_stats *sgs)
6944 struct sg_lb_stats *busiest = &sds->busiest_stat;
6946 if (sgs->group_type > busiest->group_type)
6949 if (sgs->group_type < busiest->group_type)
6953 * Candidate sg doesn't face any serious load-balance problems
6954 * so don't pick it if the local sg is already filled up.
6956 if (sgs->group_type == group_other &&
6957 !group_has_capacity(env, &sds->local_stat))
6960 if (sgs->avg_load <= busiest->avg_load)
6964 * Candiate sg has no more than one task per cpu and has higher
6965 * per-cpu capacity. No reason to pull tasks to less capable cpus.
6967 if (sgs->sum_nr_running <= sgs->group_weight &&
6968 group_smaller_cpu_capacity(sds->local, sg))
6971 /* This is the busiest node in its class. */
6972 if (!(env->sd->flags & SD_ASYM_PACKING))
6976 * ASYM_PACKING needs to move all the work to the lowest
6977 * numbered CPUs in the group, therefore mark all groups
6978 * higher than ourself as busy.
6980 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6984 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6991 #ifdef CONFIG_NUMA_BALANCING
6992 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6994 if (sgs->sum_nr_running > sgs->nr_numa_running)
6996 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7001 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7003 if (rq->nr_running > rq->nr_numa_running)
7005 if (rq->nr_running > rq->nr_preferred_running)
7010 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7015 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7019 #endif /* CONFIG_NUMA_BALANCING */
7022 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7023 * @env: The load balancing environment.
7024 * @sds: variable to hold the statistics for this sched_domain.
7026 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7028 struct sched_domain *child = env->sd->child;
7029 struct sched_group *sg = env->sd->groups;
7030 struct sg_lb_stats tmp_sgs;
7031 int load_idx, prefer_sibling = 0;
7032 bool overload = false, overutilized = false;
7034 if (child && child->flags & SD_PREFER_SIBLING)
7037 load_idx = get_sd_load_idx(env->sd, env->idle);
7040 struct sg_lb_stats *sgs = &tmp_sgs;
7043 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7046 sgs = &sds->local_stat;
7048 if (env->idle != CPU_NEWLY_IDLE ||
7049 time_after_eq(jiffies, sg->sgc->next_update))
7050 update_group_capacity(env->sd, env->dst_cpu);
7053 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7054 &overload, &overutilized);
7060 * In case the child domain prefers tasks go to siblings
7061 * first, lower the sg capacity so that we'll try
7062 * and move all the excess tasks away. We lower the capacity
7063 * of a group only if the local group has the capacity to fit
7064 * these excess tasks. The extra check prevents the case where
7065 * you always pull from the heaviest group when it is already
7066 * under-utilized (possible with a large weight task outweighs
7067 * the tasks on the system).
7069 if (prefer_sibling && sds->local &&
7070 group_has_capacity(env, &sds->local_stat) &&
7071 (sgs->sum_nr_running > 1)) {
7072 sgs->group_no_capacity = 1;
7073 sgs->group_type = group_classify(sg, sgs);
7077 * Ignore task groups with misfit tasks if local group has no
7078 * capacity or if per-cpu capacity isn't higher.
7080 if (sgs->group_type == group_misfit_task &&
7081 (!group_has_capacity(env, &sds->local_stat) ||
7082 !group_smaller_cpu_capacity(sg, sds->local)))
7083 sgs->group_type = group_other;
7085 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7087 sds->busiest_stat = *sgs;
7091 /* Now, start updating sd_lb_stats */
7092 sds->total_load += sgs->group_load;
7093 sds->total_capacity += sgs->group_capacity;
7096 } while (sg != env->sd->groups);
7098 if (env->sd->flags & SD_NUMA)
7099 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7101 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7103 if (!env->sd->parent) {
7104 /* update overload indicator if we are at root domain */
7105 if (env->dst_rq->rd->overload != overload)
7106 env->dst_rq->rd->overload = overload;
7108 /* Update over-utilization (tipping point, U >= 0) indicator */
7109 if (env->dst_rq->rd->overutilized != overutilized)
7110 env->dst_rq->rd->overutilized = overutilized;
7112 if (!env->dst_rq->rd->overutilized && overutilized)
7113 env->dst_rq->rd->overutilized = true;
7118 * check_asym_packing - Check to see if the group is packed into the
7121 * This is primarily intended to used at the sibling level. Some
7122 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7123 * case of POWER7, it can move to lower SMT modes only when higher
7124 * threads are idle. When in lower SMT modes, the threads will
7125 * perform better since they share less core resources. Hence when we
7126 * have idle threads, we want them to be the higher ones.
7128 * This packing function is run on idle threads. It checks to see if
7129 * the busiest CPU in this domain (core in the P7 case) has a higher
7130 * CPU number than the packing function is being run on. Here we are
7131 * assuming lower CPU number will be equivalent to lower a SMT thread
7134 * Return: 1 when packing is required and a task should be moved to
7135 * this CPU. The amount of the imbalance is returned in *imbalance.
7137 * @env: The load balancing environment.
7138 * @sds: Statistics of the sched_domain which is to be packed
7140 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7144 if (!(env->sd->flags & SD_ASYM_PACKING))
7150 busiest_cpu = group_first_cpu(sds->busiest);
7151 if (env->dst_cpu > busiest_cpu)
7154 env->imbalance = DIV_ROUND_CLOSEST(
7155 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7156 SCHED_CAPACITY_SCALE);
7162 * fix_small_imbalance - Calculate the minor imbalance that exists
7163 * amongst the groups of a sched_domain, during
7165 * @env: The load balancing environment.
7166 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7169 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7171 unsigned long tmp, capa_now = 0, capa_move = 0;
7172 unsigned int imbn = 2;
7173 unsigned long scaled_busy_load_per_task;
7174 struct sg_lb_stats *local, *busiest;
7176 local = &sds->local_stat;
7177 busiest = &sds->busiest_stat;
7179 if (!local->sum_nr_running)
7180 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7181 else if (busiest->load_per_task > local->load_per_task)
7184 scaled_busy_load_per_task =
7185 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7186 busiest->group_capacity;
7188 if (busiest->avg_load + scaled_busy_load_per_task >=
7189 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7190 env->imbalance = busiest->load_per_task;
7195 * OK, we don't have enough imbalance to justify moving tasks,
7196 * however we may be able to increase total CPU capacity used by
7200 capa_now += busiest->group_capacity *
7201 min(busiest->load_per_task, busiest->avg_load);
7202 capa_now += local->group_capacity *
7203 min(local->load_per_task, local->avg_load);
7204 capa_now /= SCHED_CAPACITY_SCALE;
7206 /* Amount of load we'd subtract */
7207 if (busiest->avg_load > scaled_busy_load_per_task) {
7208 capa_move += busiest->group_capacity *
7209 min(busiest->load_per_task,
7210 busiest->avg_load - scaled_busy_load_per_task);
7213 /* Amount of load we'd add */
7214 if (busiest->avg_load * busiest->group_capacity <
7215 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7216 tmp = (busiest->avg_load * busiest->group_capacity) /
7217 local->group_capacity;
7219 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7220 local->group_capacity;
7222 capa_move += local->group_capacity *
7223 min(local->load_per_task, local->avg_load + tmp);
7224 capa_move /= SCHED_CAPACITY_SCALE;
7226 /* Move if we gain throughput */
7227 if (capa_move > capa_now)
7228 env->imbalance = busiest->load_per_task;
7232 * calculate_imbalance - Calculate the amount of imbalance present within the
7233 * groups of a given sched_domain during load balance.
7234 * @env: load balance environment
7235 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7237 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7239 unsigned long max_pull, load_above_capacity = ~0UL;
7240 struct sg_lb_stats *local, *busiest;
7242 local = &sds->local_stat;
7243 busiest = &sds->busiest_stat;
7245 if (busiest->group_type == group_imbalanced) {
7247 * In the group_imb case we cannot rely on group-wide averages
7248 * to ensure cpu-load equilibrium, look at wider averages. XXX
7250 busiest->load_per_task =
7251 min(busiest->load_per_task, sds->avg_load);
7255 * In the presence of smp nice balancing, certain scenarios can have
7256 * max load less than avg load(as we skip the groups at or below
7257 * its cpu_capacity, while calculating max_load..)
7259 if (busiest->avg_load <= sds->avg_load ||
7260 local->avg_load >= sds->avg_load) {
7261 /* Misfitting tasks should be migrated in any case */
7262 if (busiest->group_type == group_misfit_task) {
7263 env->imbalance = busiest->group_misfit_task;
7268 * Busiest group is overloaded, local is not, use the spare
7269 * cycles to maximize throughput
7271 if (busiest->group_type == group_overloaded &&
7272 local->group_type <= group_misfit_task) {
7273 env->imbalance = busiest->load_per_task;
7278 return fix_small_imbalance(env, sds);
7282 * If there aren't any idle cpus, avoid creating some.
7284 if (busiest->group_type == group_overloaded &&
7285 local->group_type == group_overloaded) {
7286 load_above_capacity = busiest->sum_nr_running *
7288 if (load_above_capacity > busiest->group_capacity)
7289 load_above_capacity -= busiest->group_capacity;
7291 load_above_capacity = ~0UL;
7295 * We're trying to get all the cpus to the average_load, so we don't
7296 * want to push ourselves above the average load, nor do we wish to
7297 * reduce the max loaded cpu below the average load. At the same time,
7298 * we also don't want to reduce the group load below the group capacity
7299 * (so that we can implement power-savings policies etc). Thus we look
7300 * for the minimum possible imbalance.
7302 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7304 /* How much load to actually move to equalise the imbalance */
7305 env->imbalance = min(
7306 max_pull * busiest->group_capacity,
7307 (sds->avg_load - local->avg_load) * local->group_capacity
7308 ) / SCHED_CAPACITY_SCALE;
7310 /* Boost imbalance to allow misfit task to be balanced. */
7311 if (busiest->group_type == group_misfit_task)
7312 env->imbalance = max_t(long, env->imbalance,
7313 busiest->group_misfit_task);
7316 * if *imbalance is less than the average load per runnable task
7317 * there is no guarantee that any tasks will be moved so we'll have
7318 * a think about bumping its value to force at least one task to be
7321 if (env->imbalance < busiest->load_per_task)
7322 return fix_small_imbalance(env, sds);
7325 /******* find_busiest_group() helpers end here *********************/
7328 * find_busiest_group - Returns the busiest group within the sched_domain
7329 * if there is an imbalance. If there isn't an imbalance, and
7330 * the user has opted for power-savings, it returns a group whose
7331 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7332 * such a group exists.
7334 * Also calculates the amount of weighted load which should be moved
7335 * to restore balance.
7337 * @env: The load balancing environment.
7339 * Return: - The busiest group if imbalance exists.
7340 * - If no imbalance and user has opted for power-savings balance,
7341 * return the least loaded group whose CPUs can be
7342 * put to idle by rebalancing its tasks onto our group.
7344 static struct sched_group *find_busiest_group(struct lb_env *env)
7346 struct sg_lb_stats *local, *busiest;
7347 struct sd_lb_stats sds;
7349 init_sd_lb_stats(&sds);
7352 * Compute the various statistics relavent for load balancing at
7355 update_sd_lb_stats(env, &sds);
7357 if (energy_aware() && !env->dst_rq->rd->overutilized)
7360 local = &sds.local_stat;
7361 busiest = &sds.busiest_stat;
7363 /* ASYM feature bypasses nice load balance check */
7364 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7365 check_asym_packing(env, &sds))
7368 /* There is no busy sibling group to pull tasks from */
7369 if (!sds.busiest || busiest->sum_nr_running == 0)
7372 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7373 / sds.total_capacity;
7376 * If the busiest group is imbalanced the below checks don't
7377 * work because they assume all things are equal, which typically
7378 * isn't true due to cpus_allowed constraints and the like.
7380 if (busiest->group_type == group_imbalanced)
7383 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7384 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7385 busiest->group_no_capacity)
7388 /* Misfitting tasks should be dealt with regardless of the avg load */
7389 if (busiest->group_type == group_misfit_task) {
7394 * If the local group is busier than the selected busiest group
7395 * don't try and pull any tasks.
7397 if (local->avg_load >= busiest->avg_load)
7401 * Don't pull any tasks if this group is already above the domain
7404 if (local->avg_load >= sds.avg_load)
7407 if (env->idle == CPU_IDLE) {
7409 * This cpu is idle. If the busiest group is not overloaded
7410 * and there is no imbalance between this and busiest group
7411 * wrt idle cpus, it is balanced. The imbalance becomes
7412 * significant if the diff is greater than 1 otherwise we
7413 * might end up to just move the imbalance on another group
7415 if ((busiest->group_type != group_overloaded) &&
7416 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7417 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7421 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7422 * imbalance_pct to be conservative.
7424 if (100 * busiest->avg_load <=
7425 env->sd->imbalance_pct * local->avg_load)
7430 env->busiest_group_type = busiest->group_type;
7431 /* Looks like there is an imbalance. Compute it */
7432 calculate_imbalance(env, &sds);
7441 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7443 static struct rq *find_busiest_queue(struct lb_env *env,
7444 struct sched_group *group)
7446 struct rq *busiest = NULL, *rq;
7447 unsigned long busiest_load = 0, busiest_capacity = 1;
7450 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7451 unsigned long capacity, wl;
7455 rt = fbq_classify_rq(rq);
7458 * We classify groups/runqueues into three groups:
7459 * - regular: there are !numa tasks
7460 * - remote: there are numa tasks that run on the 'wrong' node
7461 * - all: there is no distinction
7463 * In order to avoid migrating ideally placed numa tasks,
7464 * ignore those when there's better options.
7466 * If we ignore the actual busiest queue to migrate another
7467 * task, the next balance pass can still reduce the busiest
7468 * queue by moving tasks around inside the node.
7470 * If we cannot move enough load due to this classification
7471 * the next pass will adjust the group classification and
7472 * allow migration of more tasks.
7474 * Both cases only affect the total convergence complexity.
7476 if (rt > env->fbq_type)
7479 capacity = capacity_of(i);
7481 wl = weighted_cpuload(i);
7484 * When comparing with imbalance, use weighted_cpuload()
7485 * which is not scaled with the cpu capacity.
7488 if (rq->nr_running == 1 && wl > env->imbalance &&
7489 !check_cpu_capacity(rq, env->sd) &&
7490 env->busiest_group_type != group_misfit_task)
7494 * For the load comparisons with the other cpu's, consider
7495 * the weighted_cpuload() scaled with the cpu capacity, so
7496 * that the load can be moved away from the cpu that is
7497 * potentially running at a lower capacity.
7499 * Thus we're looking for max(wl_i / capacity_i), crosswise
7500 * multiplication to rid ourselves of the division works out
7501 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7502 * our previous maximum.
7504 if (wl * busiest_capacity > busiest_load * capacity) {
7506 busiest_capacity = capacity;
7515 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7516 * so long as it is large enough.
7518 #define MAX_PINNED_INTERVAL 512
7520 /* Working cpumask for load_balance and load_balance_newidle. */
7521 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7523 static int need_active_balance(struct lb_env *env)
7525 struct sched_domain *sd = env->sd;
7527 if (env->idle == CPU_NEWLY_IDLE) {
7530 * ASYM_PACKING needs to force migrate tasks from busy but
7531 * higher numbered CPUs in order to pack all tasks in the
7532 * lowest numbered CPUs.
7534 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7539 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7540 * It's worth migrating the task if the src_cpu's capacity is reduced
7541 * because of other sched_class or IRQs if more capacity stays
7542 * available on dst_cpu.
7544 if ((env->idle != CPU_NOT_IDLE) &&
7545 (env->src_rq->cfs.h_nr_running == 1)) {
7546 if ((check_cpu_capacity(env->src_rq, sd)) &&
7547 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7551 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
7552 env->src_rq->cfs.h_nr_running == 1 &&
7553 cpu_overutilized(env->src_cpu) &&
7554 !cpu_overutilized(env->dst_cpu)) {
7558 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7561 static int active_load_balance_cpu_stop(void *data);
7563 static int should_we_balance(struct lb_env *env)
7565 struct sched_group *sg = env->sd->groups;
7566 struct cpumask *sg_cpus, *sg_mask;
7567 int cpu, balance_cpu = -1;
7570 * In the newly idle case, we will allow all the cpu's
7571 * to do the newly idle load balance.
7573 if (env->idle == CPU_NEWLY_IDLE)
7576 sg_cpus = sched_group_cpus(sg);
7577 sg_mask = sched_group_mask(sg);
7578 /* Try to find first idle cpu */
7579 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7580 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7587 if (balance_cpu == -1)
7588 balance_cpu = group_balance_cpu(sg);
7591 * First idle cpu or the first cpu(busiest) in this sched group
7592 * is eligible for doing load balancing at this and above domains.
7594 return balance_cpu == env->dst_cpu;
7598 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7599 * tasks if there is an imbalance.
7601 static int load_balance(int this_cpu, struct rq *this_rq,
7602 struct sched_domain *sd, enum cpu_idle_type idle,
7603 int *continue_balancing)
7605 int ld_moved, cur_ld_moved, active_balance = 0;
7606 struct sched_domain *sd_parent = sd->parent;
7607 struct sched_group *group;
7609 unsigned long flags;
7610 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7612 struct lb_env env = {
7614 .dst_cpu = this_cpu,
7616 .dst_grpmask = sched_group_cpus(sd->groups),
7618 .loop_break = sched_nr_migrate_break,
7621 .tasks = LIST_HEAD_INIT(env.tasks),
7625 * For NEWLY_IDLE load_balancing, we don't need to consider
7626 * other cpus in our group
7628 if (idle == CPU_NEWLY_IDLE)
7629 env.dst_grpmask = NULL;
7631 cpumask_copy(cpus, cpu_active_mask);
7633 schedstat_inc(sd, lb_count[idle]);
7636 if (!should_we_balance(&env)) {
7637 *continue_balancing = 0;
7641 group = find_busiest_group(&env);
7643 schedstat_inc(sd, lb_nobusyg[idle]);
7647 busiest = find_busiest_queue(&env, group);
7649 schedstat_inc(sd, lb_nobusyq[idle]);
7653 BUG_ON(busiest == env.dst_rq);
7655 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7657 env.src_cpu = busiest->cpu;
7658 env.src_rq = busiest;
7661 if (busiest->nr_running > 1) {
7663 * Attempt to move tasks. If find_busiest_group has found
7664 * an imbalance but busiest->nr_running <= 1, the group is
7665 * still unbalanced. ld_moved simply stays zero, so it is
7666 * correctly treated as an imbalance.
7668 env.flags |= LBF_ALL_PINNED;
7669 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7672 raw_spin_lock_irqsave(&busiest->lock, flags);
7675 * cur_ld_moved - load moved in current iteration
7676 * ld_moved - cumulative load moved across iterations
7678 cur_ld_moved = detach_tasks(&env);
7680 * We want to potentially lower env.src_cpu's OPP.
7683 update_capacity_of(env.src_cpu);
7686 * We've detached some tasks from busiest_rq. Every
7687 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7688 * unlock busiest->lock, and we are able to be sure
7689 * that nobody can manipulate the tasks in parallel.
7690 * See task_rq_lock() family for the details.
7693 raw_spin_unlock(&busiest->lock);
7697 ld_moved += cur_ld_moved;
7700 local_irq_restore(flags);
7702 if (env.flags & LBF_NEED_BREAK) {
7703 env.flags &= ~LBF_NEED_BREAK;
7708 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7709 * us and move them to an alternate dst_cpu in our sched_group
7710 * where they can run. The upper limit on how many times we
7711 * iterate on same src_cpu is dependent on number of cpus in our
7714 * This changes load balance semantics a bit on who can move
7715 * load to a given_cpu. In addition to the given_cpu itself
7716 * (or a ilb_cpu acting on its behalf where given_cpu is
7717 * nohz-idle), we now have balance_cpu in a position to move
7718 * load to given_cpu. In rare situations, this may cause
7719 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7720 * _independently_ and at _same_ time to move some load to
7721 * given_cpu) causing exceess load to be moved to given_cpu.
7722 * This however should not happen so much in practice and
7723 * moreover subsequent load balance cycles should correct the
7724 * excess load moved.
7726 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7728 /* Prevent to re-select dst_cpu via env's cpus */
7729 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7731 env.dst_rq = cpu_rq(env.new_dst_cpu);
7732 env.dst_cpu = env.new_dst_cpu;
7733 env.flags &= ~LBF_DST_PINNED;
7735 env.loop_break = sched_nr_migrate_break;
7738 * Go back to "more_balance" rather than "redo" since we
7739 * need to continue with same src_cpu.
7745 * We failed to reach balance because of affinity.
7748 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7750 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7751 *group_imbalance = 1;
7754 /* All tasks on this runqueue were pinned by CPU affinity */
7755 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7756 cpumask_clear_cpu(cpu_of(busiest), cpus);
7757 if (!cpumask_empty(cpus)) {
7759 env.loop_break = sched_nr_migrate_break;
7762 goto out_all_pinned;
7767 schedstat_inc(sd, lb_failed[idle]);
7769 * Increment the failure counter only on periodic balance.
7770 * We do not want newidle balance, which can be very
7771 * frequent, pollute the failure counter causing
7772 * excessive cache_hot migrations and active balances.
7774 if (idle != CPU_NEWLY_IDLE)
7775 if (env.src_grp_nr_running > 1)
7776 sd->nr_balance_failed++;
7778 if (need_active_balance(&env)) {
7779 raw_spin_lock_irqsave(&busiest->lock, flags);
7781 /* don't kick the active_load_balance_cpu_stop,
7782 * if the curr task on busiest cpu can't be
7785 if (!cpumask_test_cpu(this_cpu,
7786 tsk_cpus_allowed(busiest->curr))) {
7787 raw_spin_unlock_irqrestore(&busiest->lock,
7789 env.flags |= LBF_ALL_PINNED;
7790 goto out_one_pinned;
7794 * ->active_balance synchronizes accesses to
7795 * ->active_balance_work. Once set, it's cleared
7796 * only after active load balance is finished.
7798 if (!busiest->active_balance) {
7799 busiest->active_balance = 1;
7800 busiest->push_cpu = this_cpu;
7803 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7805 if (active_balance) {
7806 stop_one_cpu_nowait(cpu_of(busiest),
7807 active_load_balance_cpu_stop, busiest,
7808 &busiest->active_balance_work);
7812 * We've kicked active balancing, reset the failure
7815 sd->nr_balance_failed = sd->cache_nice_tries+1;
7818 sd->nr_balance_failed = 0;
7820 if (likely(!active_balance)) {
7821 /* We were unbalanced, so reset the balancing interval */
7822 sd->balance_interval = sd->min_interval;
7825 * If we've begun active balancing, start to back off. This
7826 * case may not be covered by the all_pinned logic if there
7827 * is only 1 task on the busy runqueue (because we don't call
7830 if (sd->balance_interval < sd->max_interval)
7831 sd->balance_interval *= 2;
7838 * We reach balance although we may have faced some affinity
7839 * constraints. Clear the imbalance flag if it was set.
7842 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7844 if (*group_imbalance)
7845 *group_imbalance = 0;
7850 * We reach balance because all tasks are pinned at this level so
7851 * we can't migrate them. Let the imbalance flag set so parent level
7852 * can try to migrate them.
7854 schedstat_inc(sd, lb_balanced[idle]);
7856 sd->nr_balance_failed = 0;
7859 /* tune up the balancing interval */
7860 if (((env.flags & LBF_ALL_PINNED) &&
7861 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7862 (sd->balance_interval < sd->max_interval))
7863 sd->balance_interval *= 2;
7870 static inline unsigned long
7871 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7873 unsigned long interval = sd->balance_interval;
7876 interval *= sd->busy_factor;
7878 /* scale ms to jiffies */
7879 interval = msecs_to_jiffies(interval);
7880 interval = clamp(interval, 1UL, max_load_balance_interval);
7886 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7888 unsigned long interval, next;
7890 interval = get_sd_balance_interval(sd, cpu_busy);
7891 next = sd->last_balance + interval;
7893 if (time_after(*next_balance, next))
7894 *next_balance = next;
7898 * idle_balance is called by schedule() if this_cpu is about to become
7899 * idle. Attempts to pull tasks from other CPUs.
7901 static int idle_balance(struct rq *this_rq)
7903 unsigned long next_balance = jiffies + HZ;
7904 int this_cpu = this_rq->cpu;
7905 struct sched_domain *sd;
7906 int pulled_task = 0;
7909 idle_enter_fair(this_rq);
7912 * We must set idle_stamp _before_ calling idle_balance(), such that we
7913 * measure the duration of idle_balance() as idle time.
7915 this_rq->idle_stamp = rq_clock(this_rq);
7917 if (!energy_aware() &&
7918 (this_rq->avg_idle < sysctl_sched_migration_cost ||
7919 !this_rq->rd->overload)) {
7921 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7923 update_next_balance(sd, 0, &next_balance);
7929 raw_spin_unlock(&this_rq->lock);
7931 update_blocked_averages(this_cpu);
7933 for_each_domain(this_cpu, sd) {
7934 int continue_balancing = 1;
7935 u64 t0, domain_cost;
7937 if (!(sd->flags & SD_LOAD_BALANCE))
7940 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7941 update_next_balance(sd, 0, &next_balance);
7945 if (sd->flags & SD_BALANCE_NEWIDLE) {
7946 t0 = sched_clock_cpu(this_cpu);
7948 pulled_task = load_balance(this_cpu, this_rq,
7950 &continue_balancing);
7952 domain_cost = sched_clock_cpu(this_cpu) - t0;
7953 if (domain_cost > sd->max_newidle_lb_cost)
7954 sd->max_newidle_lb_cost = domain_cost;
7956 curr_cost += domain_cost;
7959 update_next_balance(sd, 0, &next_balance);
7962 * Stop searching for tasks to pull if there are
7963 * now runnable tasks on this rq.
7965 if (pulled_task || this_rq->nr_running > 0)
7970 raw_spin_lock(&this_rq->lock);
7972 if (curr_cost > this_rq->max_idle_balance_cost)
7973 this_rq->max_idle_balance_cost = curr_cost;
7976 * While browsing the domains, we released the rq lock, a task could
7977 * have been enqueued in the meantime. Since we're not going idle,
7978 * pretend we pulled a task.
7980 if (this_rq->cfs.h_nr_running && !pulled_task)
7984 /* Move the next balance forward */
7985 if (time_after(this_rq->next_balance, next_balance))
7986 this_rq->next_balance = next_balance;
7988 /* Is there a task of a high priority class? */
7989 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7993 idle_exit_fair(this_rq);
7994 this_rq->idle_stamp = 0;
8001 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8002 * running tasks off the busiest CPU onto idle CPUs. It requires at
8003 * least 1 task to be running on each physical CPU where possible, and
8004 * avoids physical / logical imbalances.
8006 static int active_load_balance_cpu_stop(void *data)
8008 struct rq *busiest_rq = data;
8009 int busiest_cpu = cpu_of(busiest_rq);
8010 int target_cpu = busiest_rq->push_cpu;
8011 struct rq *target_rq = cpu_rq(target_cpu);
8012 struct sched_domain *sd;
8013 struct task_struct *p = NULL;
8015 raw_spin_lock_irq(&busiest_rq->lock);
8017 /* make sure the requested cpu hasn't gone down in the meantime */
8018 if (unlikely(busiest_cpu != smp_processor_id() ||
8019 !busiest_rq->active_balance))
8022 /* Is there any task to move? */
8023 if (busiest_rq->nr_running <= 1)
8027 * This condition is "impossible", if it occurs
8028 * we need to fix it. Originally reported by
8029 * Bjorn Helgaas on a 128-cpu setup.
8031 BUG_ON(busiest_rq == target_rq);
8033 /* Search for an sd spanning us and the target CPU. */
8035 for_each_domain(target_cpu, sd) {
8036 if ((sd->flags & SD_LOAD_BALANCE) &&
8037 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8042 struct lb_env env = {
8044 .dst_cpu = target_cpu,
8045 .dst_rq = target_rq,
8046 .src_cpu = busiest_rq->cpu,
8047 .src_rq = busiest_rq,
8051 schedstat_inc(sd, alb_count);
8053 p = detach_one_task(&env);
8055 schedstat_inc(sd, alb_pushed);
8057 * We want to potentially lower env.src_cpu's OPP.
8059 update_capacity_of(env.src_cpu);
8062 schedstat_inc(sd, alb_failed);
8066 busiest_rq->active_balance = 0;
8067 raw_spin_unlock(&busiest_rq->lock);
8070 attach_one_task(target_rq, p);
8077 static inline int on_null_domain(struct rq *rq)
8079 return unlikely(!rcu_dereference_sched(rq->sd));
8082 #ifdef CONFIG_NO_HZ_COMMON
8084 * idle load balancing details
8085 * - When one of the busy CPUs notice that there may be an idle rebalancing
8086 * needed, they will kick the idle load balancer, which then does idle
8087 * load balancing for all the idle CPUs.
8090 cpumask_var_t idle_cpus_mask;
8092 unsigned long next_balance; /* in jiffy units */
8093 } nohz ____cacheline_aligned;
8095 static inline int find_new_ilb(void)
8097 int ilb = cpumask_first(nohz.idle_cpus_mask);
8099 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8106 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8107 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8108 * CPU (if there is one).
8110 static void nohz_balancer_kick(void)
8114 nohz.next_balance++;
8116 ilb_cpu = find_new_ilb();
8118 if (ilb_cpu >= nr_cpu_ids)
8121 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8124 * Use smp_send_reschedule() instead of resched_cpu().
8125 * This way we generate a sched IPI on the target cpu which
8126 * is idle. And the softirq performing nohz idle load balance
8127 * will be run before returning from the IPI.
8129 smp_send_reschedule(ilb_cpu);
8133 static inline void nohz_balance_exit_idle(int cpu)
8135 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8137 * Completely isolated CPUs don't ever set, so we must test.
8139 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8140 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8141 atomic_dec(&nohz.nr_cpus);
8143 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8147 static inline void set_cpu_sd_state_busy(void)
8149 struct sched_domain *sd;
8150 int cpu = smp_processor_id();
8153 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8155 if (!sd || !sd->nohz_idle)
8159 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8164 void set_cpu_sd_state_idle(void)
8166 struct sched_domain *sd;
8167 int cpu = smp_processor_id();
8170 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8172 if (!sd || sd->nohz_idle)
8176 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8182 * This routine will record that the cpu is going idle with tick stopped.
8183 * This info will be used in performing idle load balancing in the future.
8185 void nohz_balance_enter_idle(int cpu)
8188 * If this cpu is going down, then nothing needs to be done.
8190 if (!cpu_active(cpu))
8193 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8197 * If we're a completely isolated CPU, we don't play.
8199 if (on_null_domain(cpu_rq(cpu)))
8202 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8203 atomic_inc(&nohz.nr_cpus);
8204 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8207 static int sched_ilb_notifier(struct notifier_block *nfb,
8208 unsigned long action, void *hcpu)
8210 switch (action & ~CPU_TASKS_FROZEN) {
8212 nohz_balance_exit_idle(smp_processor_id());
8220 static DEFINE_SPINLOCK(balancing);
8223 * Scale the max load_balance interval with the number of CPUs in the system.
8224 * This trades load-balance latency on larger machines for less cross talk.
8226 void update_max_interval(void)
8228 max_load_balance_interval = HZ*num_online_cpus()/10;
8232 * It checks each scheduling domain to see if it is due to be balanced,
8233 * and initiates a balancing operation if so.
8235 * Balancing parameters are set up in init_sched_domains.
8237 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8239 int continue_balancing = 1;
8241 unsigned long interval;
8242 struct sched_domain *sd;
8243 /* Earliest time when we have to do rebalance again */
8244 unsigned long next_balance = jiffies + 60*HZ;
8245 int update_next_balance = 0;
8246 int need_serialize, need_decay = 0;
8249 update_blocked_averages(cpu);
8252 for_each_domain(cpu, sd) {
8254 * Decay the newidle max times here because this is a regular
8255 * visit to all the domains. Decay ~1% per second.
8257 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8258 sd->max_newidle_lb_cost =
8259 (sd->max_newidle_lb_cost * 253) / 256;
8260 sd->next_decay_max_lb_cost = jiffies + HZ;
8263 max_cost += sd->max_newidle_lb_cost;
8265 if (!(sd->flags & SD_LOAD_BALANCE))
8269 * Stop the load balance at this level. There is another
8270 * CPU in our sched group which is doing load balancing more
8273 if (!continue_balancing) {
8279 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8281 need_serialize = sd->flags & SD_SERIALIZE;
8282 if (need_serialize) {
8283 if (!spin_trylock(&balancing))
8287 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8288 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8290 * The LBF_DST_PINNED logic could have changed
8291 * env->dst_cpu, so we can't know our idle
8292 * state even if we migrated tasks. Update it.
8294 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8296 sd->last_balance = jiffies;
8297 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8300 spin_unlock(&balancing);
8302 if (time_after(next_balance, sd->last_balance + interval)) {
8303 next_balance = sd->last_balance + interval;
8304 update_next_balance = 1;
8309 * Ensure the rq-wide value also decays but keep it at a
8310 * reasonable floor to avoid funnies with rq->avg_idle.
8312 rq->max_idle_balance_cost =
8313 max((u64)sysctl_sched_migration_cost, max_cost);
8318 * next_balance will be updated only when there is a need.
8319 * When the cpu is attached to null domain for ex, it will not be
8322 if (likely(update_next_balance)) {
8323 rq->next_balance = next_balance;
8325 #ifdef CONFIG_NO_HZ_COMMON
8327 * If this CPU has been elected to perform the nohz idle
8328 * balance. Other idle CPUs have already rebalanced with
8329 * nohz_idle_balance() and nohz.next_balance has been
8330 * updated accordingly. This CPU is now running the idle load
8331 * balance for itself and we need to update the
8332 * nohz.next_balance accordingly.
8334 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8335 nohz.next_balance = rq->next_balance;
8340 #ifdef CONFIG_NO_HZ_COMMON
8342 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8343 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8345 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8347 int this_cpu = this_rq->cpu;
8350 /* Earliest time when we have to do rebalance again */
8351 unsigned long next_balance = jiffies + 60*HZ;
8352 int update_next_balance = 0;
8354 if (idle != CPU_IDLE ||
8355 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8358 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8359 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8363 * If this cpu gets work to do, stop the load balancing
8364 * work being done for other cpus. Next load
8365 * balancing owner will pick it up.
8370 rq = cpu_rq(balance_cpu);
8373 * If time for next balance is due,
8376 if (time_after_eq(jiffies, rq->next_balance)) {
8377 raw_spin_lock_irq(&rq->lock);
8378 update_rq_clock(rq);
8379 update_idle_cpu_load(rq);
8380 raw_spin_unlock_irq(&rq->lock);
8381 rebalance_domains(rq, CPU_IDLE);
8384 if (time_after(next_balance, rq->next_balance)) {
8385 next_balance = rq->next_balance;
8386 update_next_balance = 1;
8391 * next_balance will be updated only when there is a need.
8392 * When the CPU is attached to null domain for ex, it will not be
8395 if (likely(update_next_balance))
8396 nohz.next_balance = next_balance;
8398 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8402 * Current heuristic for kicking the idle load balancer in the presence
8403 * of an idle cpu in the system.
8404 * - This rq has more than one task.
8405 * - This rq has at least one CFS task and the capacity of the CPU is
8406 * significantly reduced because of RT tasks or IRQs.
8407 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8408 * multiple busy cpu.
8409 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8410 * domain span are idle.
8412 static inline bool nohz_kick_needed(struct rq *rq)
8414 unsigned long now = jiffies;
8415 struct sched_domain *sd;
8416 struct sched_group_capacity *sgc;
8417 int nr_busy, cpu = rq->cpu;
8420 if (unlikely(rq->idle_balance))
8424 * We may be recently in ticked or tickless idle mode. At the first
8425 * busy tick after returning from idle, we will update the busy stats.
8427 set_cpu_sd_state_busy();
8428 nohz_balance_exit_idle(cpu);
8431 * None are in tickless mode and hence no need for NOHZ idle load
8434 if (likely(!atomic_read(&nohz.nr_cpus)))
8437 if (time_before(now, nohz.next_balance))
8440 if (rq->nr_running >= 2 &&
8441 (!energy_aware() || cpu_overutilized(cpu)))
8445 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8446 if (sd && !energy_aware()) {
8447 sgc = sd->groups->sgc;
8448 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8457 sd = rcu_dereference(rq->sd);
8459 if ((rq->cfs.h_nr_running >= 1) &&
8460 check_cpu_capacity(rq, sd)) {
8466 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8467 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8468 sched_domain_span(sd)) < cpu)) {
8478 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8482 * run_rebalance_domains is triggered when needed from the scheduler tick.
8483 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8485 static void run_rebalance_domains(struct softirq_action *h)
8487 struct rq *this_rq = this_rq();
8488 enum cpu_idle_type idle = this_rq->idle_balance ?
8489 CPU_IDLE : CPU_NOT_IDLE;
8492 * If this cpu has a pending nohz_balance_kick, then do the
8493 * balancing on behalf of the other idle cpus whose ticks are
8494 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8495 * give the idle cpus a chance to load balance. Else we may
8496 * load balance only within the local sched_domain hierarchy
8497 * and abort nohz_idle_balance altogether if we pull some load.
8499 nohz_idle_balance(this_rq, idle);
8500 rebalance_domains(this_rq, idle);
8504 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8506 void trigger_load_balance(struct rq *rq)
8508 /* Don't need to rebalance while attached to NULL domain */
8509 if (unlikely(on_null_domain(rq)))
8512 if (time_after_eq(jiffies, rq->next_balance))
8513 raise_softirq(SCHED_SOFTIRQ);
8514 #ifdef CONFIG_NO_HZ_COMMON
8515 if (nohz_kick_needed(rq))
8516 nohz_balancer_kick();
8520 static void rq_online_fair(struct rq *rq)
8524 update_runtime_enabled(rq);
8527 static void rq_offline_fair(struct rq *rq)
8531 /* Ensure any throttled groups are reachable by pick_next_task */
8532 unthrottle_offline_cfs_rqs(rq);
8535 #endif /* CONFIG_SMP */
8538 * scheduler tick hitting a task of our scheduling class:
8540 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8542 struct cfs_rq *cfs_rq;
8543 struct sched_entity *se = &curr->se;
8545 for_each_sched_entity(se) {
8546 cfs_rq = cfs_rq_of(se);
8547 entity_tick(cfs_rq, se, queued);
8550 if (static_branch_unlikely(&sched_numa_balancing))
8551 task_tick_numa(rq, curr);
8553 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr)))
8554 rq->rd->overutilized = true;
8556 rq->misfit_task = !task_fits_max(curr, rq->cpu);
8560 * called on fork with the child task as argument from the parent's context
8561 * - child not yet on the tasklist
8562 * - preemption disabled
8564 static void task_fork_fair(struct task_struct *p)
8566 struct cfs_rq *cfs_rq;
8567 struct sched_entity *se = &p->se, *curr;
8568 int this_cpu = smp_processor_id();
8569 struct rq *rq = this_rq();
8570 unsigned long flags;
8572 raw_spin_lock_irqsave(&rq->lock, flags);
8574 update_rq_clock(rq);
8576 cfs_rq = task_cfs_rq(current);
8577 curr = cfs_rq->curr;
8580 * Not only the cpu but also the task_group of the parent might have
8581 * been changed after parent->se.parent,cfs_rq were copied to
8582 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8583 * of child point to valid ones.
8586 __set_task_cpu(p, this_cpu);
8589 update_curr(cfs_rq);
8592 se->vruntime = curr->vruntime;
8593 place_entity(cfs_rq, se, 1);
8595 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8597 * Upon rescheduling, sched_class::put_prev_task() will place
8598 * 'current' within the tree based on its new key value.
8600 swap(curr->vruntime, se->vruntime);
8604 se->vruntime -= cfs_rq->min_vruntime;
8606 raw_spin_unlock_irqrestore(&rq->lock, flags);
8610 * Priority of the task has changed. Check to see if we preempt
8614 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8616 if (!task_on_rq_queued(p))
8620 * Reschedule if we are currently running on this runqueue and
8621 * our priority decreased, or if we are not currently running on
8622 * this runqueue and our priority is higher than the current's
8624 if (rq->curr == p) {
8625 if (p->prio > oldprio)
8628 check_preempt_curr(rq, p, 0);
8631 static inline bool vruntime_normalized(struct task_struct *p)
8633 struct sched_entity *se = &p->se;
8636 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8637 * the dequeue_entity(.flags=0) will already have normalized the
8644 * When !on_rq, vruntime of the task has usually NOT been normalized.
8645 * But there are some cases where it has already been normalized:
8647 * - A forked child which is waiting for being woken up by
8648 * wake_up_new_task().
8649 * - A task which has been woken up by try_to_wake_up() and
8650 * waiting for actually being woken up by sched_ttwu_pending().
8652 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8658 static void detach_task_cfs_rq(struct task_struct *p)
8660 struct sched_entity *se = &p->se;
8661 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8663 if (!vruntime_normalized(p)) {
8665 * Fix up our vruntime so that the current sleep doesn't
8666 * cause 'unlimited' sleep bonus.
8668 place_entity(cfs_rq, se, 0);
8669 se->vruntime -= cfs_rq->min_vruntime;
8672 /* Catch up with the cfs_rq and remove our load when we leave */
8673 detach_entity_load_avg(cfs_rq, se);
8676 static void attach_task_cfs_rq(struct task_struct *p)
8678 struct sched_entity *se = &p->se;
8679 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8681 #ifdef CONFIG_FAIR_GROUP_SCHED
8683 * Since the real-depth could have been changed (only FAIR
8684 * class maintain depth value), reset depth properly.
8686 se->depth = se->parent ? se->parent->depth + 1 : 0;
8689 /* Synchronize task with its cfs_rq */
8690 attach_entity_load_avg(cfs_rq, se);
8692 if (!vruntime_normalized(p))
8693 se->vruntime += cfs_rq->min_vruntime;
8696 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8698 detach_task_cfs_rq(p);
8701 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8703 attach_task_cfs_rq(p);
8705 if (task_on_rq_queued(p)) {
8707 * We were most likely switched from sched_rt, so
8708 * kick off the schedule if running, otherwise just see
8709 * if we can still preempt the current task.
8714 check_preempt_curr(rq, p, 0);
8718 /* Account for a task changing its policy or group.
8720 * This routine is mostly called to set cfs_rq->curr field when a task
8721 * migrates between groups/classes.
8723 static void set_curr_task_fair(struct rq *rq)
8725 struct sched_entity *se = &rq->curr->se;
8727 for_each_sched_entity(se) {
8728 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8730 set_next_entity(cfs_rq, se);
8731 /* ensure bandwidth has been allocated on our new cfs_rq */
8732 account_cfs_rq_runtime(cfs_rq, 0);
8736 void init_cfs_rq(struct cfs_rq *cfs_rq)
8738 cfs_rq->tasks_timeline = RB_ROOT;
8739 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8740 #ifndef CONFIG_64BIT
8741 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8744 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8745 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8749 #ifdef CONFIG_FAIR_GROUP_SCHED
8750 static void task_move_group_fair(struct task_struct *p)
8752 detach_task_cfs_rq(p);
8753 set_task_rq(p, task_cpu(p));
8756 /* Tell se's cfs_rq has been changed -- migrated */
8757 p->se.avg.last_update_time = 0;
8759 attach_task_cfs_rq(p);
8762 void free_fair_sched_group(struct task_group *tg)
8766 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8768 for_each_possible_cpu(i) {
8770 kfree(tg->cfs_rq[i]);
8773 remove_entity_load_avg(tg->se[i]);
8782 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8784 struct cfs_rq *cfs_rq;
8785 struct sched_entity *se;
8788 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8791 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8795 tg->shares = NICE_0_LOAD;
8797 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8799 for_each_possible_cpu(i) {
8800 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8801 GFP_KERNEL, cpu_to_node(i));
8805 se = kzalloc_node(sizeof(struct sched_entity),
8806 GFP_KERNEL, cpu_to_node(i));
8810 init_cfs_rq(cfs_rq);
8811 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8812 init_entity_runnable_average(se);
8823 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8825 struct rq *rq = cpu_rq(cpu);
8826 unsigned long flags;
8829 * Only empty task groups can be destroyed; so we can speculatively
8830 * check on_list without danger of it being re-added.
8832 if (!tg->cfs_rq[cpu]->on_list)
8835 raw_spin_lock_irqsave(&rq->lock, flags);
8836 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8837 raw_spin_unlock_irqrestore(&rq->lock, flags);
8840 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8841 struct sched_entity *se, int cpu,
8842 struct sched_entity *parent)
8844 struct rq *rq = cpu_rq(cpu);
8848 init_cfs_rq_runtime(cfs_rq);
8850 tg->cfs_rq[cpu] = cfs_rq;
8853 /* se could be NULL for root_task_group */
8858 se->cfs_rq = &rq->cfs;
8861 se->cfs_rq = parent->my_q;
8862 se->depth = parent->depth + 1;
8866 /* guarantee group entities always have weight */
8867 update_load_set(&se->load, NICE_0_LOAD);
8868 se->parent = parent;
8871 static DEFINE_MUTEX(shares_mutex);
8873 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8876 unsigned long flags;
8879 * We can't change the weight of the root cgroup.
8884 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8886 mutex_lock(&shares_mutex);
8887 if (tg->shares == shares)
8890 tg->shares = shares;
8891 for_each_possible_cpu(i) {
8892 struct rq *rq = cpu_rq(i);
8893 struct sched_entity *se;
8896 /* Propagate contribution to hierarchy */
8897 raw_spin_lock_irqsave(&rq->lock, flags);
8899 /* Possible calls to update_curr() need rq clock */
8900 update_rq_clock(rq);
8901 for_each_sched_entity(se)
8902 update_cfs_shares(group_cfs_rq(se));
8903 raw_spin_unlock_irqrestore(&rq->lock, flags);
8907 mutex_unlock(&shares_mutex);
8910 #else /* CONFIG_FAIR_GROUP_SCHED */
8912 void free_fair_sched_group(struct task_group *tg) { }
8914 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8919 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8921 #endif /* CONFIG_FAIR_GROUP_SCHED */
8924 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8926 struct sched_entity *se = &task->se;
8927 unsigned int rr_interval = 0;
8930 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8933 if (rq->cfs.load.weight)
8934 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8940 * All the scheduling class methods:
8942 const struct sched_class fair_sched_class = {
8943 .next = &idle_sched_class,
8944 .enqueue_task = enqueue_task_fair,
8945 .dequeue_task = dequeue_task_fair,
8946 .yield_task = yield_task_fair,
8947 .yield_to_task = yield_to_task_fair,
8949 .check_preempt_curr = check_preempt_wakeup,
8951 .pick_next_task = pick_next_task_fair,
8952 .put_prev_task = put_prev_task_fair,
8955 .select_task_rq = select_task_rq_fair,
8956 .migrate_task_rq = migrate_task_rq_fair,
8958 .rq_online = rq_online_fair,
8959 .rq_offline = rq_offline_fair,
8961 .task_waking = task_waking_fair,
8962 .task_dead = task_dead_fair,
8963 .set_cpus_allowed = set_cpus_allowed_common,
8966 .set_curr_task = set_curr_task_fair,
8967 .task_tick = task_tick_fair,
8968 .task_fork = task_fork_fair,
8970 .prio_changed = prio_changed_fair,
8971 .switched_from = switched_from_fair,
8972 .switched_to = switched_to_fair,
8974 .get_rr_interval = get_rr_interval_fair,
8976 .update_curr = update_curr_fair,
8978 #ifdef CONFIG_FAIR_GROUP_SCHED
8979 .task_move_group = task_move_group_fair,
8983 #ifdef CONFIG_SCHED_DEBUG
8984 void print_cfs_stats(struct seq_file *m, int cpu)
8986 struct cfs_rq *cfs_rq;
8989 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8990 print_cfs_rq(m, cpu, cfs_rq);
8994 #ifdef CONFIG_NUMA_BALANCING
8995 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8998 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9000 for_each_online_node(node) {
9001 if (p->numa_faults) {
9002 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9003 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9005 if (p->numa_group) {
9006 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9007 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9009 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9012 #endif /* CONFIG_NUMA_BALANCING */
9013 #endif /* CONFIG_SCHED_DEBUG */
9015 __init void init_sched_fair_class(void)
9018 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9020 #ifdef CONFIG_NO_HZ_COMMON
9021 nohz.next_balance = jiffies;
9022 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9023 cpu_notifier(sched_ilb_notifier, 0);