2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency = 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity = 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency = 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static unsigned int get_update_sysctl_factor(void)
146 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling) {
150 case SCHED_TUNABLESCALING_NONE:
153 case SCHED_TUNABLESCALING_LINEAR:
156 case SCHED_TUNABLESCALING_LOG:
158 factor = 1 + ilog2(cpus);
165 static void update_sysctl(void)
167 unsigned int factor = get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity);
172 SET_SYSCTL(sched_latency);
173 SET_SYSCTL(sched_wakeup_granularity);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight *lw)
189 if (likely(lw->inv_weight))
192 w = scale_load_down(lw->weight);
194 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
196 else if (unlikely(!w))
197 lw->inv_weight = WMULT_CONST;
199 lw->inv_weight = WMULT_CONST / w;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
216 u64 fact = scale_load_down(weight);
217 int shift = WMULT_SHIFT;
219 __update_inv_weight(lw);
221 if (unlikely(fact >> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact = (u64)(u32)fact * lw->inv_weight;
236 return mul_u64_u32_shr(delta_exec, fact, shift);
240 const struct sched_class fair_sched_class;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct *task_of(struct sched_entity *se)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se));
262 return container_of(se, struct task_struct, se);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
310 if (cfs_rq->on_list) {
311 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq *
322 is_same_group(struct sched_entity *se, struct sched_entity *pse)
324 if (se->cfs_rq == pse->cfs_rq)
330 static inline struct sched_entity *parent_entity(struct sched_entity *se)
336 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
338 int se_depth, pse_depth;
341 * preemption test can be made between sibling entities who are in the
342 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343 * both tasks until we find their ancestors who are siblings of common
347 /* First walk up until both entities are at same depth */
348 se_depth = (*se)->depth;
349 pse_depth = (*pse)->depth;
351 while (se_depth > pse_depth) {
353 *se = parent_entity(*se);
356 while (pse_depth > se_depth) {
358 *pse = parent_entity(*pse);
361 while (!is_same_group(*se, *pse)) {
362 *se = parent_entity(*se);
363 *pse = parent_entity(*pse);
367 #else /* !CONFIG_FAIR_GROUP_SCHED */
369 static inline struct task_struct *task_of(struct sched_entity *se)
371 return container_of(se, struct task_struct, se);
374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
376 return container_of(cfs_rq, struct rq, cfs);
379 #define entity_is_task(se) 1
381 #define for_each_sched_entity(se) \
382 for (; se; se = NULL)
384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
386 return &task_rq(p)->cfs;
389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
391 struct task_struct *p = task_of(se);
392 struct rq *rq = task_rq(p);
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
414 static inline struct sched_entity *parent_entity(struct sched_entity *se)
420 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
424 #endif /* CONFIG_FAIR_GROUP_SCHED */
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
429 /**************************************************************
430 * Scheduling class tree data structure manipulation methods:
433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
435 s64 delta = (s64)(vruntime - max_vruntime);
437 max_vruntime = vruntime;
442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
444 s64 delta = (s64)(vruntime - min_vruntime);
446 min_vruntime = vruntime;
451 static inline int entity_before(struct sched_entity *a,
452 struct sched_entity *b)
454 return (s64)(a->vruntime - b->vruntime) < 0;
457 static void update_min_vruntime(struct cfs_rq *cfs_rq)
459 u64 vruntime = cfs_rq->min_vruntime;
462 vruntime = cfs_rq->curr->vruntime;
464 if (cfs_rq->rb_leftmost) {
465 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
470 vruntime = se->vruntime;
472 vruntime = min_vruntime(vruntime, se->vruntime);
475 /* ensure we never gain time by being placed backwards. */
476 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
479 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484 * Enqueue an entity into the rb-tree:
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
488 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489 struct rb_node *parent = NULL;
490 struct sched_entity *entry;
494 * Find the right place in the rbtree:
498 entry = rb_entry(parent, struct sched_entity, run_node);
500 * We dont care about collisions. Nodes with
501 * the same key stay together.
503 if (entity_before(se, entry)) {
504 link = &parent->rb_left;
506 link = &parent->rb_right;
512 * Maintain a cache of leftmost tree entries (it is frequently
516 cfs_rq->rb_leftmost = &se->run_node;
518 rb_link_node(&se->run_node, parent, link);
519 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
524 if (cfs_rq->rb_leftmost == &se->run_node) {
525 struct rb_node *next_node;
527 next_node = rb_next(&se->run_node);
528 cfs_rq->rb_leftmost = next_node;
531 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
536 struct rb_node *left = cfs_rq->rb_leftmost;
541 return rb_entry(left, struct sched_entity, run_node);
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
546 struct rb_node *next = rb_next(&se->run_node);
551 return rb_entry(next, struct sched_entity, run_node);
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
557 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
562 return rb_entry(last, struct sched_entity, run_node);
565 /**************************************************************
566 * Scheduling class statistics methods:
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570 void __user *buffer, size_t *lenp,
573 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574 unsigned int factor = get_update_sysctl_factor();
579 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580 sysctl_sched_min_granularity);
582 #define WRT_SYSCTL(name) \
583 (normalized_sysctl_##name = sysctl_##name / (factor))
584 WRT_SYSCTL(sched_min_granularity);
585 WRT_SYSCTL(sched_latency);
586 WRT_SYSCTL(sched_wakeup_granularity);
596 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
598 if (unlikely(se->load.weight != NICE_0_LOAD))
599 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
605 * The idea is to set a period in which each task runs once.
607 * When there are too many tasks (sched_nr_latency) we have to stretch
608 * this period because otherwise the slices get too small.
610 * p = (nr <= nl) ? l : l*nr/nl
612 static u64 __sched_period(unsigned long nr_running)
614 if (unlikely(nr_running > sched_nr_latency))
615 return nr_running * sysctl_sched_min_granularity;
617 return sysctl_sched_latency;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
630 for_each_sched_entity(se) {
631 struct load_weight *load;
632 struct load_weight lw;
634 cfs_rq = cfs_rq_of(se);
635 load = &cfs_rq->load;
637 if (unlikely(!se->on_rq)) {
640 update_load_add(&lw, se->load.weight);
643 slice = __calc_delta(slice, se->load.weight, load);
649 * We calculate the vruntime slice of a to-be-inserted task.
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
655 return calc_delta_fair(sched_slice(cfs_rq, se), se);
659 static int select_idle_sibling(struct task_struct *p, int cpu);
660 static unsigned long task_h_load(struct task_struct *p);
663 * We choose a half-life close to 1 scheduling period.
664 * Note: The tables below are dependent on this value.
666 #define LOAD_AVG_PERIOD 32
667 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
668 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
670 /* Give new sched_entity start runnable values to heavy its load in infant time */
671 void init_entity_runnable_average(struct sched_entity *se)
673 struct sched_avg *sa = &se->avg;
675 sa->last_update_time = 0;
677 * sched_avg's period_contrib should be strictly less then 1024, so
678 * we give it 1023 to make sure it is almost a period (1024us), and
679 * will definitely be update (after enqueue).
681 sa->period_contrib = 1023;
682 sa->load_avg = scale_load_down(se->load.weight);
683 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
684 sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
685 sa->util_sum = LOAD_AVG_MAX;
686 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
689 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
690 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
692 void init_entity_runnable_average(struct sched_entity *se)
698 * Update the current task's runtime statistics.
700 static void update_curr(struct cfs_rq *cfs_rq)
702 struct sched_entity *curr = cfs_rq->curr;
703 u64 now = rq_clock_task(rq_of(cfs_rq));
709 delta_exec = now - curr->exec_start;
710 if (unlikely((s64)delta_exec <= 0))
713 curr->exec_start = now;
715 schedstat_set(curr->statistics.exec_max,
716 max(delta_exec, curr->statistics.exec_max));
718 curr->sum_exec_runtime += delta_exec;
719 schedstat_add(cfs_rq, exec_clock, delta_exec);
721 curr->vruntime += calc_delta_fair(delta_exec, curr);
722 update_min_vruntime(cfs_rq);
724 if (entity_is_task(curr)) {
725 struct task_struct *curtask = task_of(curr);
727 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
728 cpuacct_charge(curtask, delta_exec);
729 account_group_exec_runtime(curtask, delta_exec);
732 account_cfs_rq_runtime(cfs_rq, delta_exec);
735 static void update_curr_fair(struct rq *rq)
737 update_curr(cfs_rq_of(&rq->curr->se));
741 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
743 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
747 * Task is being enqueued - update stats:
749 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
752 * Are we enqueueing a waiting task? (for current tasks
753 * a dequeue/enqueue event is a NOP)
755 if (se != cfs_rq->curr)
756 update_stats_wait_start(cfs_rq, se);
760 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
762 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
763 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
764 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
765 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
766 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
767 #ifdef CONFIG_SCHEDSTATS
768 if (entity_is_task(se)) {
769 trace_sched_stat_wait(task_of(se),
770 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
773 schedstat_set(se->statistics.wait_start, 0);
777 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
780 * Mark the end of the wait period if dequeueing a
783 if (se != cfs_rq->curr)
784 update_stats_wait_end(cfs_rq, se);
788 * We are picking a new current task - update its stats:
791 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
794 * We are starting a new run period:
796 se->exec_start = rq_clock_task(rq_of(cfs_rq));
799 /**************************************************
800 * Scheduling class queueing methods:
803 #ifdef CONFIG_NUMA_BALANCING
805 * Approximate time to scan a full NUMA task in ms. The task scan period is
806 * calculated based on the tasks virtual memory size and
807 * numa_balancing_scan_size.
809 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
810 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
812 /* Portion of address space to scan in MB */
813 unsigned int sysctl_numa_balancing_scan_size = 256;
815 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
816 unsigned int sysctl_numa_balancing_scan_delay = 1000;
818 static unsigned int task_nr_scan_windows(struct task_struct *p)
820 unsigned long rss = 0;
821 unsigned long nr_scan_pages;
824 * Calculations based on RSS as non-present and empty pages are skipped
825 * by the PTE scanner and NUMA hinting faults should be trapped based
828 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
829 rss = get_mm_rss(p->mm);
833 rss = round_up(rss, nr_scan_pages);
834 return rss / nr_scan_pages;
837 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
838 #define MAX_SCAN_WINDOW 2560
840 static unsigned int task_scan_min(struct task_struct *p)
842 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
843 unsigned int scan, floor;
844 unsigned int windows = 1;
846 if (scan_size < MAX_SCAN_WINDOW)
847 windows = MAX_SCAN_WINDOW / scan_size;
848 floor = 1000 / windows;
850 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
851 return max_t(unsigned int, floor, scan);
854 static unsigned int task_scan_max(struct task_struct *p)
856 unsigned int smin = task_scan_min(p);
859 /* Watch for min being lower than max due to floor calculations */
860 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
861 return max(smin, smax);
864 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
866 rq->nr_numa_running += (p->numa_preferred_nid != -1);
867 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
870 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
872 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
873 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
879 spinlock_t lock; /* nr_tasks, tasks */
884 nodemask_t active_nodes;
885 unsigned long total_faults;
887 * Faults_cpu is used to decide whether memory should move
888 * towards the CPU. As a consequence, these stats are weighted
889 * more by CPU use than by memory faults.
891 unsigned long *faults_cpu;
892 unsigned long faults[0];
895 /* Shared or private faults. */
896 #define NR_NUMA_HINT_FAULT_TYPES 2
898 /* Memory and CPU locality */
899 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
901 /* Averaged statistics, and temporary buffers. */
902 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
904 pid_t task_numa_group_id(struct task_struct *p)
906 return p->numa_group ? p->numa_group->gid : 0;
910 * The averaged statistics, shared & private, memory & cpu,
911 * occupy the first half of the array. The second half of the
912 * array is for current counters, which are averaged into the
913 * first set by task_numa_placement.
915 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
917 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
920 static inline unsigned long task_faults(struct task_struct *p, int nid)
925 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
926 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
929 static inline unsigned long group_faults(struct task_struct *p, int nid)
934 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
935 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
938 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
940 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
941 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
944 /* Handle placement on systems where not all nodes are directly connected. */
945 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
946 int maxdist, bool task)
948 unsigned long score = 0;
952 * All nodes are directly connected, and the same distance
953 * from each other. No need for fancy placement algorithms.
955 if (sched_numa_topology_type == NUMA_DIRECT)
959 * This code is called for each node, introducing N^2 complexity,
960 * which should be ok given the number of nodes rarely exceeds 8.
962 for_each_online_node(node) {
963 unsigned long faults;
964 int dist = node_distance(nid, node);
967 * The furthest away nodes in the system are not interesting
968 * for placement; nid was already counted.
970 if (dist == sched_max_numa_distance || node == nid)
974 * On systems with a backplane NUMA topology, compare groups
975 * of nodes, and move tasks towards the group with the most
976 * memory accesses. When comparing two nodes at distance
977 * "hoplimit", only nodes closer by than "hoplimit" are part
978 * of each group. Skip other nodes.
980 if (sched_numa_topology_type == NUMA_BACKPLANE &&
984 /* Add up the faults from nearby nodes. */
986 faults = task_faults(p, node);
988 faults = group_faults(p, node);
991 * On systems with a glueless mesh NUMA topology, there are
992 * no fixed "groups of nodes". Instead, nodes that are not
993 * directly connected bounce traffic through intermediate
994 * nodes; a numa_group can occupy any set of nodes.
995 * The further away a node is, the less the faults count.
996 * This seems to result in good task placement.
998 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
999 faults *= (sched_max_numa_distance - dist);
1000 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1010 * These return the fraction of accesses done by a particular task, or
1011 * task group, on a particular numa node. The group weight is given a
1012 * larger multiplier, in order to group tasks together that are almost
1013 * evenly spread out between numa nodes.
1015 static inline unsigned long task_weight(struct task_struct *p, int nid,
1018 unsigned long faults, total_faults;
1020 if (!p->numa_faults)
1023 total_faults = p->total_numa_faults;
1028 faults = task_faults(p, nid);
1029 faults += score_nearby_nodes(p, nid, dist, true);
1031 return 1000 * faults / total_faults;
1034 static inline unsigned long group_weight(struct task_struct *p, int nid,
1037 unsigned long faults, total_faults;
1042 total_faults = p->numa_group->total_faults;
1047 faults = group_faults(p, nid);
1048 faults += score_nearby_nodes(p, nid, dist, false);
1050 return 1000 * faults / total_faults;
1053 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1054 int src_nid, int dst_cpu)
1056 struct numa_group *ng = p->numa_group;
1057 int dst_nid = cpu_to_node(dst_cpu);
1058 int last_cpupid, this_cpupid;
1060 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1063 * Multi-stage node selection is used in conjunction with a periodic
1064 * migration fault to build a temporal task<->page relation. By using
1065 * a two-stage filter we remove short/unlikely relations.
1067 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1068 * a task's usage of a particular page (n_p) per total usage of this
1069 * page (n_t) (in a given time-span) to a probability.
1071 * Our periodic faults will sample this probability and getting the
1072 * same result twice in a row, given these samples are fully
1073 * independent, is then given by P(n)^2, provided our sample period
1074 * is sufficiently short compared to the usage pattern.
1076 * This quadric squishes small probabilities, making it less likely we
1077 * act on an unlikely task<->page relation.
1079 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1080 if (!cpupid_pid_unset(last_cpupid) &&
1081 cpupid_to_nid(last_cpupid) != dst_nid)
1084 /* Always allow migrate on private faults */
1085 if (cpupid_match_pid(p, last_cpupid))
1088 /* A shared fault, but p->numa_group has not been set up yet. */
1093 * Do not migrate if the destination is not a node that
1094 * is actively used by this numa group.
1096 if (!node_isset(dst_nid, ng->active_nodes))
1100 * Source is a node that is not actively used by this
1101 * numa group, while the destination is. Migrate.
1103 if (!node_isset(src_nid, ng->active_nodes))
1107 * Both source and destination are nodes in active
1108 * use by this numa group. Maximize memory bandwidth
1109 * by migrating from more heavily used groups, to less
1110 * heavily used ones, spreading the load around.
1111 * Use a 1/4 hysteresis to avoid spurious page movement.
1113 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1116 static unsigned long weighted_cpuload(const int cpu);
1117 static unsigned long source_load(int cpu, int type);
1118 static unsigned long target_load(int cpu, int type);
1119 static unsigned long capacity_of(int cpu);
1120 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1122 /* Cached statistics for all CPUs within a node */
1124 unsigned long nr_running;
1127 /* Total compute capacity of CPUs on a node */
1128 unsigned long compute_capacity;
1130 /* Approximate capacity in terms of runnable tasks on a node */
1131 unsigned long task_capacity;
1132 int has_free_capacity;
1136 * XXX borrowed from update_sg_lb_stats
1138 static void update_numa_stats(struct numa_stats *ns, int nid)
1140 int smt, cpu, cpus = 0;
1141 unsigned long capacity;
1143 memset(ns, 0, sizeof(*ns));
1144 for_each_cpu(cpu, cpumask_of_node(nid)) {
1145 struct rq *rq = cpu_rq(cpu);
1147 ns->nr_running += rq->nr_running;
1148 ns->load += weighted_cpuload(cpu);
1149 ns->compute_capacity += capacity_of(cpu);
1155 * If we raced with hotplug and there are no CPUs left in our mask
1156 * the @ns structure is NULL'ed and task_numa_compare() will
1157 * not find this node attractive.
1159 * We'll either bail at !has_free_capacity, or we'll detect a huge
1160 * imbalance and bail there.
1165 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1166 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1167 capacity = cpus / smt; /* cores */
1169 ns->task_capacity = min_t(unsigned, capacity,
1170 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1171 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1174 struct task_numa_env {
1175 struct task_struct *p;
1177 int src_cpu, src_nid;
1178 int dst_cpu, dst_nid;
1180 struct numa_stats src_stats, dst_stats;
1185 struct task_struct *best_task;
1190 static void task_numa_assign(struct task_numa_env *env,
1191 struct task_struct *p, long imp)
1194 put_task_struct(env->best_task);
1199 env->best_imp = imp;
1200 env->best_cpu = env->dst_cpu;
1203 static bool load_too_imbalanced(long src_load, long dst_load,
1204 struct task_numa_env *env)
1207 long orig_src_load, orig_dst_load;
1208 long src_capacity, dst_capacity;
1211 * The load is corrected for the CPU capacity available on each node.
1214 * ------------ vs ---------
1215 * src_capacity dst_capacity
1217 src_capacity = env->src_stats.compute_capacity;
1218 dst_capacity = env->dst_stats.compute_capacity;
1220 /* We care about the slope of the imbalance, not the direction. */
1221 if (dst_load < src_load)
1222 swap(dst_load, src_load);
1224 /* Is the difference below the threshold? */
1225 imb = dst_load * src_capacity * 100 -
1226 src_load * dst_capacity * env->imbalance_pct;
1231 * The imbalance is above the allowed threshold.
1232 * Compare it with the old imbalance.
1234 orig_src_load = env->src_stats.load;
1235 orig_dst_load = env->dst_stats.load;
1237 if (orig_dst_load < orig_src_load)
1238 swap(orig_dst_load, orig_src_load);
1240 old_imb = orig_dst_load * src_capacity * 100 -
1241 orig_src_load * dst_capacity * env->imbalance_pct;
1243 /* Would this change make things worse? */
1244 return (imb > old_imb);
1248 * This checks if the overall compute and NUMA accesses of the system would
1249 * be improved if the source tasks was migrated to the target dst_cpu taking
1250 * into account that it might be best if task running on the dst_cpu should
1251 * be exchanged with the source task
1253 static void task_numa_compare(struct task_numa_env *env,
1254 long taskimp, long groupimp)
1256 struct rq *src_rq = cpu_rq(env->src_cpu);
1257 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1258 struct task_struct *cur;
1259 long src_load, dst_load;
1261 long imp = env->p->numa_group ? groupimp : taskimp;
1263 int dist = env->dist;
1267 raw_spin_lock_irq(&dst_rq->lock);
1270 * No need to move the exiting task, and this ensures that ->curr
1271 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1272 * is safe under RCU read lock.
1273 * Note that rcu_read_lock() itself can't protect from the final
1274 * put_task_struct() after the last schedule().
1276 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1278 raw_spin_unlock_irq(&dst_rq->lock);
1281 * Because we have preemption enabled we can get migrated around and
1282 * end try selecting ourselves (current == env->p) as a swap candidate.
1288 * "imp" is the fault differential for the source task between the
1289 * source and destination node. Calculate the total differential for
1290 * the source task and potential destination task. The more negative
1291 * the value is, the more rmeote accesses that would be expected to
1292 * be incurred if the tasks were swapped.
1295 /* Skip this swap candidate if cannot move to the source cpu */
1296 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1300 * If dst and source tasks are in the same NUMA group, or not
1301 * in any group then look only at task weights.
1303 if (cur->numa_group == env->p->numa_group) {
1304 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1305 task_weight(cur, env->dst_nid, dist);
1307 * Add some hysteresis to prevent swapping the
1308 * tasks within a group over tiny differences.
1310 if (cur->numa_group)
1314 * Compare the group weights. If a task is all by
1315 * itself (not part of a group), use the task weight
1318 if (cur->numa_group)
1319 imp += group_weight(cur, env->src_nid, dist) -
1320 group_weight(cur, env->dst_nid, dist);
1322 imp += task_weight(cur, env->src_nid, dist) -
1323 task_weight(cur, env->dst_nid, dist);
1327 if (imp <= env->best_imp && moveimp <= env->best_imp)
1331 /* Is there capacity at our destination? */
1332 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1333 !env->dst_stats.has_free_capacity)
1339 /* Balance doesn't matter much if we're running a task per cpu */
1340 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1341 dst_rq->nr_running == 1)
1345 * In the overloaded case, try and keep the load balanced.
1348 load = task_h_load(env->p);
1349 dst_load = env->dst_stats.load + load;
1350 src_load = env->src_stats.load - load;
1352 if (moveimp > imp && moveimp > env->best_imp) {
1354 * If the improvement from just moving env->p direction is
1355 * better than swapping tasks around, check if a move is
1356 * possible. Store a slightly smaller score than moveimp,
1357 * so an actually idle CPU will win.
1359 if (!load_too_imbalanced(src_load, dst_load, env)) {
1366 if (imp <= env->best_imp)
1370 load = task_h_load(cur);
1375 if (load_too_imbalanced(src_load, dst_load, env))
1379 * One idle CPU per node is evaluated for a task numa move.
1380 * Call select_idle_sibling to maybe find a better one.
1383 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1386 task_numa_assign(env, cur, imp);
1391 static void task_numa_find_cpu(struct task_numa_env *env,
1392 long taskimp, long groupimp)
1396 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1397 /* Skip this CPU if the source task cannot migrate */
1398 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1402 task_numa_compare(env, taskimp, groupimp);
1406 /* Only move tasks to a NUMA node less busy than the current node. */
1407 static bool numa_has_capacity(struct task_numa_env *env)
1409 struct numa_stats *src = &env->src_stats;
1410 struct numa_stats *dst = &env->dst_stats;
1412 if (src->has_free_capacity && !dst->has_free_capacity)
1416 * Only consider a task move if the source has a higher load
1417 * than the destination, corrected for CPU capacity on each node.
1419 * src->load dst->load
1420 * --------------------- vs ---------------------
1421 * src->compute_capacity dst->compute_capacity
1423 if (src->load * dst->compute_capacity * env->imbalance_pct >
1425 dst->load * src->compute_capacity * 100)
1431 static int task_numa_migrate(struct task_struct *p)
1433 struct task_numa_env env = {
1436 .src_cpu = task_cpu(p),
1437 .src_nid = task_node(p),
1439 .imbalance_pct = 112,
1445 struct sched_domain *sd;
1446 unsigned long taskweight, groupweight;
1448 long taskimp, groupimp;
1451 * Pick the lowest SD_NUMA domain, as that would have the smallest
1452 * imbalance and would be the first to start moving tasks about.
1454 * And we want to avoid any moving of tasks about, as that would create
1455 * random movement of tasks -- counter the numa conditions we're trying
1459 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1461 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1465 * Cpusets can break the scheduler domain tree into smaller
1466 * balance domains, some of which do not cross NUMA boundaries.
1467 * Tasks that are "trapped" in such domains cannot be migrated
1468 * elsewhere, so there is no point in (re)trying.
1470 if (unlikely(!sd)) {
1471 p->numa_preferred_nid = task_node(p);
1475 env.dst_nid = p->numa_preferred_nid;
1476 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1477 taskweight = task_weight(p, env.src_nid, dist);
1478 groupweight = group_weight(p, env.src_nid, dist);
1479 update_numa_stats(&env.src_stats, env.src_nid);
1480 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1481 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1482 update_numa_stats(&env.dst_stats, env.dst_nid);
1484 /* Try to find a spot on the preferred nid. */
1485 if (numa_has_capacity(&env))
1486 task_numa_find_cpu(&env, taskimp, groupimp);
1489 * Look at other nodes in these cases:
1490 * - there is no space available on the preferred_nid
1491 * - the task is part of a numa_group that is interleaved across
1492 * multiple NUMA nodes; in order to better consolidate the group,
1493 * we need to check other locations.
1495 if (env.best_cpu == -1 || (p->numa_group &&
1496 nodes_weight(p->numa_group->active_nodes) > 1)) {
1497 for_each_online_node(nid) {
1498 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1501 dist = node_distance(env.src_nid, env.dst_nid);
1502 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1504 taskweight = task_weight(p, env.src_nid, dist);
1505 groupweight = group_weight(p, env.src_nid, dist);
1508 /* Only consider nodes where both task and groups benefit */
1509 taskimp = task_weight(p, nid, dist) - taskweight;
1510 groupimp = group_weight(p, nid, dist) - groupweight;
1511 if (taskimp < 0 && groupimp < 0)
1516 update_numa_stats(&env.dst_stats, env.dst_nid);
1517 if (numa_has_capacity(&env))
1518 task_numa_find_cpu(&env, taskimp, groupimp);
1523 * If the task is part of a workload that spans multiple NUMA nodes,
1524 * and is migrating into one of the workload's active nodes, remember
1525 * this node as the task's preferred numa node, so the workload can
1527 * A task that migrated to a second choice node will be better off
1528 * trying for a better one later. Do not set the preferred node here.
1530 if (p->numa_group) {
1531 if (env.best_cpu == -1)
1536 if (node_isset(nid, p->numa_group->active_nodes))
1537 sched_setnuma(p, env.dst_nid);
1540 /* No better CPU than the current one was found. */
1541 if (env.best_cpu == -1)
1545 * Reset the scan period if the task is being rescheduled on an
1546 * alternative node to recheck if the tasks is now properly placed.
1548 p->numa_scan_period = task_scan_min(p);
1550 if (env.best_task == NULL) {
1551 ret = migrate_task_to(p, env.best_cpu);
1553 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1557 ret = migrate_swap(p, env.best_task);
1559 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1560 put_task_struct(env.best_task);
1564 /* Attempt to migrate a task to a CPU on the preferred node. */
1565 static void numa_migrate_preferred(struct task_struct *p)
1567 unsigned long interval = HZ;
1569 /* This task has no NUMA fault statistics yet */
1570 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1573 /* Periodically retry migrating the task to the preferred node */
1574 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1575 p->numa_migrate_retry = jiffies + interval;
1577 /* Success if task is already running on preferred CPU */
1578 if (task_node(p) == p->numa_preferred_nid)
1581 /* Otherwise, try migrate to a CPU on the preferred node */
1582 task_numa_migrate(p);
1586 * Find the nodes on which the workload is actively running. We do this by
1587 * tracking the nodes from which NUMA hinting faults are triggered. This can
1588 * be different from the set of nodes where the workload's memory is currently
1591 * The bitmask is used to make smarter decisions on when to do NUMA page
1592 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1593 * are added when they cause over 6/16 of the maximum number of faults, but
1594 * only removed when they drop below 3/16.
1596 static void update_numa_active_node_mask(struct numa_group *numa_group)
1598 unsigned long faults, max_faults = 0;
1601 for_each_online_node(nid) {
1602 faults = group_faults_cpu(numa_group, nid);
1603 if (faults > max_faults)
1604 max_faults = faults;
1607 for_each_online_node(nid) {
1608 faults = group_faults_cpu(numa_group, nid);
1609 if (!node_isset(nid, numa_group->active_nodes)) {
1610 if (faults > max_faults * 6 / 16)
1611 node_set(nid, numa_group->active_nodes);
1612 } else if (faults < max_faults * 3 / 16)
1613 node_clear(nid, numa_group->active_nodes);
1618 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1619 * increments. The more local the fault statistics are, the higher the scan
1620 * period will be for the next scan window. If local/(local+remote) ratio is
1621 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1622 * the scan period will decrease. Aim for 70% local accesses.
1624 #define NUMA_PERIOD_SLOTS 10
1625 #define NUMA_PERIOD_THRESHOLD 7
1628 * Increase the scan period (slow down scanning) if the majority of
1629 * our memory is already on our local node, or if the majority of
1630 * the page accesses are shared with other processes.
1631 * Otherwise, decrease the scan period.
1633 static void update_task_scan_period(struct task_struct *p,
1634 unsigned long shared, unsigned long private)
1636 unsigned int period_slot;
1640 unsigned long remote = p->numa_faults_locality[0];
1641 unsigned long local = p->numa_faults_locality[1];
1644 * If there were no record hinting faults then either the task is
1645 * completely idle or all activity is areas that are not of interest
1646 * to automatic numa balancing. Related to that, if there were failed
1647 * migration then it implies we are migrating too quickly or the local
1648 * node is overloaded. In either case, scan slower
1650 if (local + shared == 0 || p->numa_faults_locality[2]) {
1651 p->numa_scan_period = min(p->numa_scan_period_max,
1652 p->numa_scan_period << 1);
1654 p->mm->numa_next_scan = jiffies +
1655 msecs_to_jiffies(p->numa_scan_period);
1661 * Prepare to scale scan period relative to the current period.
1662 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1663 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1664 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1666 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1667 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1668 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1669 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1672 diff = slot * period_slot;
1674 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1677 * Scale scan rate increases based on sharing. There is an
1678 * inverse relationship between the degree of sharing and
1679 * the adjustment made to the scanning period. Broadly
1680 * speaking the intent is that there is little point
1681 * scanning faster if shared accesses dominate as it may
1682 * simply bounce migrations uselessly
1684 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1685 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1688 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1689 task_scan_min(p), task_scan_max(p));
1690 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1694 * Get the fraction of time the task has been running since the last
1695 * NUMA placement cycle. The scheduler keeps similar statistics, but
1696 * decays those on a 32ms period, which is orders of magnitude off
1697 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1698 * stats only if the task is so new there are no NUMA statistics yet.
1700 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1702 u64 runtime, delta, now;
1703 /* Use the start of this time slice to avoid calculations. */
1704 now = p->se.exec_start;
1705 runtime = p->se.sum_exec_runtime;
1707 if (p->last_task_numa_placement) {
1708 delta = runtime - p->last_sum_exec_runtime;
1709 *period = now - p->last_task_numa_placement;
1711 delta = p->se.avg.load_sum / p->se.load.weight;
1712 *period = LOAD_AVG_MAX;
1715 p->last_sum_exec_runtime = runtime;
1716 p->last_task_numa_placement = now;
1722 * Determine the preferred nid for a task in a numa_group. This needs to
1723 * be done in a way that produces consistent results with group_weight,
1724 * otherwise workloads might not converge.
1726 static int preferred_group_nid(struct task_struct *p, int nid)
1731 /* Direct connections between all NUMA nodes. */
1732 if (sched_numa_topology_type == NUMA_DIRECT)
1736 * On a system with glueless mesh NUMA topology, group_weight
1737 * scores nodes according to the number of NUMA hinting faults on
1738 * both the node itself, and on nearby nodes.
1740 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1741 unsigned long score, max_score = 0;
1742 int node, max_node = nid;
1744 dist = sched_max_numa_distance;
1746 for_each_online_node(node) {
1747 score = group_weight(p, node, dist);
1748 if (score > max_score) {
1757 * Finding the preferred nid in a system with NUMA backplane
1758 * interconnect topology is more involved. The goal is to locate
1759 * tasks from numa_groups near each other in the system, and
1760 * untangle workloads from different sides of the system. This requires
1761 * searching down the hierarchy of node groups, recursively searching
1762 * inside the highest scoring group of nodes. The nodemask tricks
1763 * keep the complexity of the search down.
1765 nodes = node_online_map;
1766 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1767 unsigned long max_faults = 0;
1768 nodemask_t max_group = NODE_MASK_NONE;
1771 /* Are there nodes at this distance from each other? */
1772 if (!find_numa_distance(dist))
1775 for_each_node_mask(a, nodes) {
1776 unsigned long faults = 0;
1777 nodemask_t this_group;
1778 nodes_clear(this_group);
1780 /* Sum group's NUMA faults; includes a==b case. */
1781 for_each_node_mask(b, nodes) {
1782 if (node_distance(a, b) < dist) {
1783 faults += group_faults(p, b);
1784 node_set(b, this_group);
1785 node_clear(b, nodes);
1789 /* Remember the top group. */
1790 if (faults > max_faults) {
1791 max_faults = faults;
1792 max_group = this_group;
1794 * subtle: at the smallest distance there is
1795 * just one node left in each "group", the
1796 * winner is the preferred nid.
1801 /* Next round, evaluate the nodes within max_group. */
1809 static void task_numa_placement(struct task_struct *p)
1811 int seq, nid, max_nid = -1, max_group_nid = -1;
1812 unsigned long max_faults = 0, max_group_faults = 0;
1813 unsigned long fault_types[2] = { 0, 0 };
1814 unsigned long total_faults;
1815 u64 runtime, period;
1816 spinlock_t *group_lock = NULL;
1819 * The p->mm->numa_scan_seq field gets updated without
1820 * exclusive access. Use READ_ONCE() here to ensure
1821 * that the field is read in a single access:
1823 seq = READ_ONCE(p->mm->numa_scan_seq);
1824 if (p->numa_scan_seq == seq)
1826 p->numa_scan_seq = seq;
1827 p->numa_scan_period_max = task_scan_max(p);
1829 total_faults = p->numa_faults_locality[0] +
1830 p->numa_faults_locality[1];
1831 runtime = numa_get_avg_runtime(p, &period);
1833 /* If the task is part of a group prevent parallel updates to group stats */
1834 if (p->numa_group) {
1835 group_lock = &p->numa_group->lock;
1836 spin_lock_irq(group_lock);
1839 /* Find the node with the highest number of faults */
1840 for_each_online_node(nid) {
1841 /* Keep track of the offsets in numa_faults array */
1842 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1843 unsigned long faults = 0, group_faults = 0;
1846 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1847 long diff, f_diff, f_weight;
1849 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1850 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1851 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1852 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1854 /* Decay existing window, copy faults since last scan */
1855 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1856 fault_types[priv] += p->numa_faults[membuf_idx];
1857 p->numa_faults[membuf_idx] = 0;
1860 * Normalize the faults_from, so all tasks in a group
1861 * count according to CPU use, instead of by the raw
1862 * number of faults. Tasks with little runtime have
1863 * little over-all impact on throughput, and thus their
1864 * faults are less important.
1866 f_weight = div64_u64(runtime << 16, period + 1);
1867 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1869 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1870 p->numa_faults[cpubuf_idx] = 0;
1872 p->numa_faults[mem_idx] += diff;
1873 p->numa_faults[cpu_idx] += f_diff;
1874 faults += p->numa_faults[mem_idx];
1875 p->total_numa_faults += diff;
1876 if (p->numa_group) {
1878 * safe because we can only change our own group
1880 * mem_idx represents the offset for a given
1881 * nid and priv in a specific region because it
1882 * is at the beginning of the numa_faults array.
1884 p->numa_group->faults[mem_idx] += diff;
1885 p->numa_group->faults_cpu[mem_idx] += f_diff;
1886 p->numa_group->total_faults += diff;
1887 group_faults += p->numa_group->faults[mem_idx];
1891 if (faults > max_faults) {
1892 max_faults = faults;
1896 if (group_faults > max_group_faults) {
1897 max_group_faults = group_faults;
1898 max_group_nid = nid;
1902 update_task_scan_period(p, fault_types[0], fault_types[1]);
1904 if (p->numa_group) {
1905 update_numa_active_node_mask(p->numa_group);
1906 spin_unlock_irq(group_lock);
1907 max_nid = preferred_group_nid(p, max_group_nid);
1911 /* Set the new preferred node */
1912 if (max_nid != p->numa_preferred_nid)
1913 sched_setnuma(p, max_nid);
1915 if (task_node(p) != p->numa_preferred_nid)
1916 numa_migrate_preferred(p);
1920 static inline int get_numa_group(struct numa_group *grp)
1922 return atomic_inc_not_zero(&grp->refcount);
1925 static inline void put_numa_group(struct numa_group *grp)
1927 if (atomic_dec_and_test(&grp->refcount))
1928 kfree_rcu(grp, rcu);
1931 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1934 struct numa_group *grp, *my_grp;
1935 struct task_struct *tsk;
1937 int cpu = cpupid_to_cpu(cpupid);
1940 if (unlikely(!p->numa_group)) {
1941 unsigned int size = sizeof(struct numa_group) +
1942 4*nr_node_ids*sizeof(unsigned long);
1944 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1948 atomic_set(&grp->refcount, 1);
1949 spin_lock_init(&grp->lock);
1951 /* Second half of the array tracks nids where faults happen */
1952 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1955 node_set(task_node(current), grp->active_nodes);
1957 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1958 grp->faults[i] = p->numa_faults[i];
1960 grp->total_faults = p->total_numa_faults;
1963 rcu_assign_pointer(p->numa_group, grp);
1967 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1969 if (!cpupid_match_pid(tsk, cpupid))
1972 grp = rcu_dereference(tsk->numa_group);
1976 my_grp = p->numa_group;
1981 * Only join the other group if its bigger; if we're the bigger group,
1982 * the other task will join us.
1984 if (my_grp->nr_tasks > grp->nr_tasks)
1988 * Tie-break on the grp address.
1990 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1993 /* Always join threads in the same process. */
1994 if (tsk->mm == current->mm)
1997 /* Simple filter to avoid false positives due to PID collisions */
1998 if (flags & TNF_SHARED)
2001 /* Update priv based on whether false sharing was detected */
2004 if (join && !get_numa_group(grp))
2012 BUG_ON(irqs_disabled());
2013 double_lock_irq(&my_grp->lock, &grp->lock);
2015 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2016 my_grp->faults[i] -= p->numa_faults[i];
2017 grp->faults[i] += p->numa_faults[i];
2019 my_grp->total_faults -= p->total_numa_faults;
2020 grp->total_faults += p->total_numa_faults;
2025 spin_unlock(&my_grp->lock);
2026 spin_unlock_irq(&grp->lock);
2028 rcu_assign_pointer(p->numa_group, grp);
2030 put_numa_group(my_grp);
2038 void task_numa_free(struct task_struct *p)
2040 struct numa_group *grp = p->numa_group;
2041 void *numa_faults = p->numa_faults;
2042 unsigned long flags;
2046 spin_lock_irqsave(&grp->lock, flags);
2047 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2048 grp->faults[i] -= p->numa_faults[i];
2049 grp->total_faults -= p->total_numa_faults;
2052 spin_unlock_irqrestore(&grp->lock, flags);
2053 RCU_INIT_POINTER(p->numa_group, NULL);
2054 put_numa_group(grp);
2057 p->numa_faults = NULL;
2062 * Got a PROT_NONE fault for a page on @node.
2064 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2066 struct task_struct *p = current;
2067 bool migrated = flags & TNF_MIGRATED;
2068 int cpu_node = task_node(current);
2069 int local = !!(flags & TNF_FAULT_LOCAL);
2072 if (!static_branch_likely(&sched_numa_balancing))
2075 /* for example, ksmd faulting in a user's mm */
2079 /* Allocate buffer to track faults on a per-node basis */
2080 if (unlikely(!p->numa_faults)) {
2081 int size = sizeof(*p->numa_faults) *
2082 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2084 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2085 if (!p->numa_faults)
2088 p->total_numa_faults = 0;
2089 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2093 * First accesses are treated as private, otherwise consider accesses
2094 * to be private if the accessing pid has not changed
2096 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2099 priv = cpupid_match_pid(p, last_cpupid);
2100 if (!priv && !(flags & TNF_NO_GROUP))
2101 task_numa_group(p, last_cpupid, flags, &priv);
2105 * If a workload spans multiple NUMA nodes, a shared fault that
2106 * occurs wholly within the set of nodes that the workload is
2107 * actively using should be counted as local. This allows the
2108 * scan rate to slow down when a workload has settled down.
2110 if (!priv && !local && p->numa_group &&
2111 node_isset(cpu_node, p->numa_group->active_nodes) &&
2112 node_isset(mem_node, p->numa_group->active_nodes))
2115 task_numa_placement(p);
2118 * Retry task to preferred node migration periodically, in case it
2119 * case it previously failed, or the scheduler moved us.
2121 if (time_after(jiffies, p->numa_migrate_retry))
2122 numa_migrate_preferred(p);
2125 p->numa_pages_migrated += pages;
2126 if (flags & TNF_MIGRATE_FAIL)
2127 p->numa_faults_locality[2] += pages;
2129 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2130 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2131 p->numa_faults_locality[local] += pages;
2134 static void reset_ptenuma_scan(struct task_struct *p)
2137 * We only did a read acquisition of the mmap sem, so
2138 * p->mm->numa_scan_seq is written to without exclusive access
2139 * and the update is not guaranteed to be atomic. That's not
2140 * much of an issue though, since this is just used for
2141 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2142 * expensive, to avoid any form of compiler optimizations:
2144 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2145 p->mm->numa_scan_offset = 0;
2149 * The expensive part of numa migration is done from task_work context.
2150 * Triggered from task_tick_numa().
2152 void task_numa_work(struct callback_head *work)
2154 unsigned long migrate, next_scan, now = jiffies;
2155 struct task_struct *p = current;
2156 struct mm_struct *mm = p->mm;
2157 struct vm_area_struct *vma;
2158 unsigned long start, end;
2159 unsigned long nr_pte_updates = 0;
2162 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2164 work->next = work; /* protect against double add */
2166 * Who cares about NUMA placement when they're dying.
2168 * NOTE: make sure not to dereference p->mm before this check,
2169 * exit_task_work() happens _after_ exit_mm() so we could be called
2170 * without p->mm even though we still had it when we enqueued this
2173 if (p->flags & PF_EXITING)
2176 if (!mm->numa_next_scan) {
2177 mm->numa_next_scan = now +
2178 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2182 * Enforce maximal scan/migration frequency..
2184 migrate = mm->numa_next_scan;
2185 if (time_before(now, migrate))
2188 if (p->numa_scan_period == 0) {
2189 p->numa_scan_period_max = task_scan_max(p);
2190 p->numa_scan_period = task_scan_min(p);
2193 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2194 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2198 * Delay this task enough that another task of this mm will likely win
2199 * the next time around.
2201 p->node_stamp += 2 * TICK_NSEC;
2203 start = mm->numa_scan_offset;
2204 pages = sysctl_numa_balancing_scan_size;
2205 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2209 down_read(&mm->mmap_sem);
2210 vma = find_vma(mm, start);
2212 reset_ptenuma_scan(p);
2216 for (; vma; vma = vma->vm_next) {
2217 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2218 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2223 * Shared library pages mapped by multiple processes are not
2224 * migrated as it is expected they are cache replicated. Avoid
2225 * hinting faults in read-only file-backed mappings or the vdso
2226 * as migrating the pages will be of marginal benefit.
2229 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2233 * Skip inaccessible VMAs to avoid any confusion between
2234 * PROT_NONE and NUMA hinting ptes
2236 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2240 start = max(start, vma->vm_start);
2241 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2242 end = min(end, vma->vm_end);
2243 nr_pte_updates += change_prot_numa(vma, start, end);
2246 * Scan sysctl_numa_balancing_scan_size but ensure that
2247 * at least one PTE is updated so that unused virtual
2248 * address space is quickly skipped.
2251 pages -= (end - start) >> PAGE_SHIFT;
2258 } while (end != vma->vm_end);
2263 * It is possible to reach the end of the VMA list but the last few
2264 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2265 * would find the !migratable VMA on the next scan but not reset the
2266 * scanner to the start so check it now.
2269 mm->numa_scan_offset = start;
2271 reset_ptenuma_scan(p);
2272 up_read(&mm->mmap_sem);
2276 * Drive the periodic memory faults..
2278 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2280 struct callback_head *work = &curr->numa_work;
2284 * We don't care about NUMA placement if we don't have memory.
2286 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2290 * Using runtime rather than walltime has the dual advantage that
2291 * we (mostly) drive the selection from busy threads and that the
2292 * task needs to have done some actual work before we bother with
2295 now = curr->se.sum_exec_runtime;
2296 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2298 if (now - curr->node_stamp > period) {
2299 if (!curr->node_stamp)
2300 curr->numa_scan_period = task_scan_min(curr);
2301 curr->node_stamp += period;
2303 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2304 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2305 task_work_add(curr, work, true);
2310 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2314 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2318 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2321 #endif /* CONFIG_NUMA_BALANCING */
2324 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2326 update_load_add(&cfs_rq->load, se->load.weight);
2327 if (!parent_entity(se))
2328 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2330 if (entity_is_task(se)) {
2331 struct rq *rq = rq_of(cfs_rq);
2333 account_numa_enqueue(rq, task_of(se));
2334 list_add(&se->group_node, &rq->cfs_tasks);
2337 cfs_rq->nr_running++;
2341 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2343 update_load_sub(&cfs_rq->load, se->load.weight);
2344 if (!parent_entity(se))
2345 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2346 if (entity_is_task(se)) {
2347 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2348 list_del_init(&se->group_node);
2350 cfs_rq->nr_running--;
2353 #ifdef CONFIG_FAIR_GROUP_SCHED
2355 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2360 * Use this CPU's real-time load instead of the last load contribution
2361 * as the updating of the contribution is delayed, and we will use the
2362 * the real-time load to calc the share. See update_tg_load_avg().
2364 tg_weight = atomic_long_read(&tg->load_avg);
2365 tg_weight -= cfs_rq->tg_load_avg_contrib;
2366 tg_weight += cfs_rq_load_avg(cfs_rq);
2371 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2373 long tg_weight, load, shares;
2375 tg_weight = calc_tg_weight(tg, cfs_rq);
2376 load = cfs_rq_load_avg(cfs_rq);
2378 shares = (tg->shares * load);
2380 shares /= tg_weight;
2382 if (shares < MIN_SHARES)
2383 shares = MIN_SHARES;
2384 if (shares > tg->shares)
2385 shares = tg->shares;
2389 # else /* CONFIG_SMP */
2390 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2394 # endif /* CONFIG_SMP */
2395 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2396 unsigned long weight)
2399 /* commit outstanding execution time */
2400 if (cfs_rq->curr == se)
2401 update_curr(cfs_rq);
2402 account_entity_dequeue(cfs_rq, se);
2405 update_load_set(&se->load, weight);
2408 account_entity_enqueue(cfs_rq, se);
2411 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2413 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2415 struct task_group *tg;
2416 struct sched_entity *se;
2420 se = tg->se[cpu_of(rq_of(cfs_rq))];
2421 if (!se || throttled_hierarchy(cfs_rq))
2424 if (likely(se->load.weight == tg->shares))
2427 shares = calc_cfs_shares(cfs_rq, tg);
2429 reweight_entity(cfs_rq_of(se), se, shares);
2431 #else /* CONFIG_FAIR_GROUP_SCHED */
2432 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2435 #endif /* CONFIG_FAIR_GROUP_SCHED */
2438 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2439 static const u32 runnable_avg_yN_inv[] = {
2440 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2441 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2442 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2443 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2444 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2445 0x85aac367, 0x82cd8698,
2449 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2450 * over-estimates when re-combining.
2452 static const u32 runnable_avg_yN_sum[] = {
2453 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2454 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2455 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2460 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2462 static __always_inline u64 decay_load(u64 val, u64 n)
2464 unsigned int local_n;
2468 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2471 /* after bounds checking we can collapse to 32-bit */
2475 * As y^PERIOD = 1/2, we can combine
2476 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2477 * With a look-up table which covers y^n (n<PERIOD)
2479 * To achieve constant time decay_load.
2481 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2482 val >>= local_n / LOAD_AVG_PERIOD;
2483 local_n %= LOAD_AVG_PERIOD;
2486 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2491 * For updates fully spanning n periods, the contribution to runnable
2492 * average will be: \Sum 1024*y^n
2494 * We can compute this reasonably efficiently by combining:
2495 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2497 static u32 __compute_runnable_contrib(u64 n)
2501 if (likely(n <= LOAD_AVG_PERIOD))
2502 return runnable_avg_yN_sum[n];
2503 else if (unlikely(n >= LOAD_AVG_MAX_N))
2504 return LOAD_AVG_MAX;
2506 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2508 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2509 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2511 n -= LOAD_AVG_PERIOD;
2512 } while (n > LOAD_AVG_PERIOD);
2514 contrib = decay_load(contrib, n);
2515 return contrib + runnable_avg_yN_sum[n];
2518 #define scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2521 * We can represent the historical contribution to runnable average as the
2522 * coefficients of a geometric series. To do this we sub-divide our runnable
2523 * history into segments of approximately 1ms (1024us); label the segment that
2524 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2526 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2528 * (now) (~1ms ago) (~2ms ago)
2530 * Let u_i denote the fraction of p_i that the entity was runnable.
2532 * We then designate the fractions u_i as our co-efficients, yielding the
2533 * following representation of historical load:
2534 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2536 * We choose y based on the with of a reasonably scheduling period, fixing:
2539 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2540 * approximately half as much as the contribution to load within the last ms
2543 * When a period "rolls over" and we have new u_0`, multiplying the previous
2544 * sum again by y is sufficient to update:
2545 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2546 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2548 static __always_inline int
2549 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2550 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2552 u64 delta, scaled_delta, periods;
2554 int delta_w, scaled_delta_w, decayed = 0;
2555 unsigned long scale_freq = arch_scale_freq_capacity(NULL, cpu);
2556 unsigned long scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2558 delta = now - sa->last_update_time;
2560 * This should only happen when time goes backwards, which it
2561 * unfortunately does during sched clock init when we swap over to TSC.
2563 if ((s64)delta < 0) {
2564 sa->last_update_time = now;
2569 * Use 1024ns as the unit of measurement since it's a reasonable
2570 * approximation of 1us and fast to compute.
2575 sa->last_update_time = now;
2577 /* delta_w is the amount already accumulated against our next period */
2578 delta_w = sa->period_contrib;
2579 if (delta + delta_w >= 1024) {
2582 /* how much left for next period will start over, we don't know yet */
2583 sa->period_contrib = 0;
2586 * Now that we know we're crossing a period boundary, figure
2587 * out how much from delta we need to complete the current
2588 * period and accrue it.
2590 delta_w = 1024 - delta_w;
2591 scaled_delta_w = scale(delta_w, scale_freq);
2593 sa->load_sum += weight * scaled_delta_w;
2595 cfs_rq->runnable_load_sum +=
2596 weight * scaled_delta_w;
2600 sa->util_sum += scale(scaled_delta_w, scale_cpu);
2604 /* Figure out how many additional periods this update spans */
2605 periods = delta / 1024;
2608 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2610 cfs_rq->runnable_load_sum =
2611 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2613 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2615 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2616 contrib = __compute_runnable_contrib(periods);
2617 contrib = scale(contrib, scale_freq);
2619 sa->load_sum += weight * contrib;
2621 cfs_rq->runnable_load_sum += weight * contrib;
2624 sa->util_sum += scale(contrib, scale_cpu);
2627 /* Remainder of delta accrued against u_0` */
2628 scaled_delta = scale(delta, scale_freq);
2630 sa->load_sum += weight * scaled_delta;
2632 cfs_rq->runnable_load_sum += weight * scaled_delta;
2635 sa->util_sum += scale(scaled_delta, scale_cpu);
2637 sa->period_contrib += delta;
2640 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2642 cfs_rq->runnable_load_avg =
2643 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2645 sa->util_avg = (sa->util_sum << SCHED_LOAD_SHIFT) / LOAD_AVG_MAX;
2651 #ifdef CONFIG_FAIR_GROUP_SCHED
2653 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2654 * and effective_load (which is not done because it is too costly).
2656 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2658 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2660 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2661 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2662 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2666 #else /* CONFIG_FAIR_GROUP_SCHED */
2667 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2668 #endif /* CONFIG_FAIR_GROUP_SCHED */
2670 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2672 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2673 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2675 struct sched_avg *sa = &cfs_rq->avg;
2678 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2679 long r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2680 sa->load_avg = max_t(long, sa->load_avg - r, 0);
2681 sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
2684 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2685 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2686 sa->util_avg = max_t(long, sa->util_avg - r, 0);
2687 sa->util_sum = max_t(s32, sa->util_sum -
2688 ((r * LOAD_AVG_MAX) >> SCHED_LOAD_SHIFT), 0);
2691 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2692 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2694 #ifndef CONFIG_64BIT
2696 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2702 /* Update task and its cfs_rq load average */
2703 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2705 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2706 u64 now = cfs_rq_clock_task(cfs_rq);
2707 int cpu = cpu_of(rq_of(cfs_rq));
2710 * Track task load average for carrying it to new CPU after migrated, and
2711 * track group sched_entity load average for task_h_load calc in migration
2713 __update_load_avg(now, cpu, &se->avg,
2714 se->on_rq * scale_load_down(se->load.weight),
2715 cfs_rq->curr == se, NULL);
2717 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2718 update_tg_load_avg(cfs_rq, 0);
2721 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2723 if (!sched_feat(ATTACH_AGE_LOAD))
2727 * If we got migrated (either between CPUs or between cgroups) we'll
2728 * have aged the average right before clearing @last_update_time.
2730 if (se->avg.last_update_time) {
2731 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2732 &se->avg, 0, 0, NULL);
2735 * XXX: we could have just aged the entire load away if we've been
2736 * absent from the fair class for too long.
2741 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2742 cfs_rq->avg.load_avg += se->avg.load_avg;
2743 cfs_rq->avg.load_sum += se->avg.load_sum;
2744 cfs_rq->avg.util_avg += se->avg.util_avg;
2745 cfs_rq->avg.util_sum += se->avg.util_sum;
2748 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2750 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2751 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2752 cfs_rq->curr == se, NULL);
2754 cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
2755 cfs_rq->avg.load_sum = max_t(s64, cfs_rq->avg.load_sum - se->avg.load_sum, 0);
2756 cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
2757 cfs_rq->avg.util_sum = max_t(s32, cfs_rq->avg.util_sum - se->avg.util_sum, 0);
2760 /* Add the load generated by se into cfs_rq's load average */
2762 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2764 struct sched_avg *sa = &se->avg;
2765 u64 now = cfs_rq_clock_task(cfs_rq);
2766 int migrated, decayed;
2768 migrated = !sa->last_update_time;
2770 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2771 se->on_rq * scale_load_down(se->load.weight),
2772 cfs_rq->curr == se, NULL);
2775 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2777 cfs_rq->runnable_load_avg += sa->load_avg;
2778 cfs_rq->runnable_load_sum += sa->load_sum;
2781 attach_entity_load_avg(cfs_rq, se);
2783 if (decayed || migrated)
2784 update_tg_load_avg(cfs_rq, 0);
2787 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2789 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2791 update_load_avg(se, 1);
2793 cfs_rq->runnable_load_avg =
2794 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2795 cfs_rq->runnable_load_sum =
2796 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2800 * Task first catches up with cfs_rq, and then subtract
2801 * itself from the cfs_rq (task must be off the queue now).
2803 void remove_entity_load_avg(struct sched_entity *se)
2805 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2806 u64 last_update_time;
2808 #ifndef CONFIG_64BIT
2809 u64 last_update_time_copy;
2812 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2814 last_update_time = cfs_rq->avg.last_update_time;
2815 } while (last_update_time != last_update_time_copy);
2817 last_update_time = cfs_rq->avg.last_update_time;
2820 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2821 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2822 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2826 * Update the rq's load with the elapsed running time before entering
2827 * idle. if the last scheduled task is not a CFS task, idle_enter will
2828 * be the only way to update the runnable statistic.
2830 void idle_enter_fair(struct rq *this_rq)
2835 * Update the rq's load with the elapsed idle time before a task is
2836 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2837 * be the only way to update the runnable statistic.
2839 void idle_exit_fair(struct rq *this_rq)
2843 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2845 return cfs_rq->runnable_load_avg;
2848 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2850 return cfs_rq->avg.load_avg;
2853 static int idle_balance(struct rq *this_rq);
2855 #else /* CONFIG_SMP */
2857 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2859 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2861 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2862 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2865 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2867 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2869 static inline int idle_balance(struct rq *rq)
2874 #endif /* CONFIG_SMP */
2876 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2878 #ifdef CONFIG_SCHEDSTATS
2879 struct task_struct *tsk = NULL;
2881 if (entity_is_task(se))
2884 if (se->statistics.sleep_start) {
2885 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2890 if (unlikely(delta > se->statistics.sleep_max))
2891 se->statistics.sleep_max = delta;
2893 se->statistics.sleep_start = 0;
2894 se->statistics.sum_sleep_runtime += delta;
2897 account_scheduler_latency(tsk, delta >> 10, 1);
2898 trace_sched_stat_sleep(tsk, delta);
2901 if (se->statistics.block_start) {
2902 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2907 if (unlikely(delta > se->statistics.block_max))
2908 se->statistics.block_max = delta;
2910 se->statistics.block_start = 0;
2911 se->statistics.sum_sleep_runtime += delta;
2914 if (tsk->in_iowait) {
2915 se->statistics.iowait_sum += delta;
2916 se->statistics.iowait_count++;
2917 trace_sched_stat_iowait(tsk, delta);
2920 trace_sched_stat_blocked(tsk, delta);
2923 * Blocking time is in units of nanosecs, so shift by
2924 * 20 to get a milliseconds-range estimation of the
2925 * amount of time that the task spent sleeping:
2927 if (unlikely(prof_on == SLEEP_PROFILING)) {
2928 profile_hits(SLEEP_PROFILING,
2929 (void *)get_wchan(tsk),
2932 account_scheduler_latency(tsk, delta >> 10, 0);
2938 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2940 #ifdef CONFIG_SCHED_DEBUG
2941 s64 d = se->vruntime - cfs_rq->min_vruntime;
2946 if (d > 3*sysctl_sched_latency)
2947 schedstat_inc(cfs_rq, nr_spread_over);
2952 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2954 u64 vruntime = cfs_rq->min_vruntime;
2957 * The 'current' period is already promised to the current tasks,
2958 * however the extra weight of the new task will slow them down a
2959 * little, place the new task so that it fits in the slot that
2960 * stays open at the end.
2962 if (initial && sched_feat(START_DEBIT))
2963 vruntime += sched_vslice(cfs_rq, se);
2965 /* sleeps up to a single latency don't count. */
2967 unsigned long thresh = sysctl_sched_latency;
2970 * Halve their sleep time's effect, to allow
2971 * for a gentler effect of sleepers:
2973 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2979 /* ensure we never gain time by being placed backwards. */
2980 se->vruntime = max_vruntime(se->vruntime, vruntime);
2983 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2986 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2989 * Update the normalized vruntime before updating min_vruntime
2990 * through calling update_curr().
2992 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2993 se->vruntime += cfs_rq->min_vruntime;
2996 * Update run-time statistics of the 'current'.
2998 update_curr(cfs_rq);
2999 enqueue_entity_load_avg(cfs_rq, se);
3000 account_entity_enqueue(cfs_rq, se);
3001 update_cfs_shares(cfs_rq);
3003 if (flags & ENQUEUE_WAKEUP) {
3004 place_entity(cfs_rq, se, 0);
3005 enqueue_sleeper(cfs_rq, se);
3008 update_stats_enqueue(cfs_rq, se);
3009 check_spread(cfs_rq, se);
3010 if (se != cfs_rq->curr)
3011 __enqueue_entity(cfs_rq, se);
3014 if (cfs_rq->nr_running == 1) {
3015 list_add_leaf_cfs_rq(cfs_rq);
3016 check_enqueue_throttle(cfs_rq);
3020 static void __clear_buddies_last(struct sched_entity *se)
3022 for_each_sched_entity(se) {
3023 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3024 if (cfs_rq->last != se)
3027 cfs_rq->last = NULL;
3031 static void __clear_buddies_next(struct sched_entity *se)
3033 for_each_sched_entity(se) {
3034 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3035 if (cfs_rq->next != se)
3038 cfs_rq->next = NULL;
3042 static void __clear_buddies_skip(struct sched_entity *se)
3044 for_each_sched_entity(se) {
3045 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3046 if (cfs_rq->skip != se)
3049 cfs_rq->skip = NULL;
3053 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3055 if (cfs_rq->last == se)
3056 __clear_buddies_last(se);
3058 if (cfs_rq->next == se)
3059 __clear_buddies_next(se);
3061 if (cfs_rq->skip == se)
3062 __clear_buddies_skip(se);
3065 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3068 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3071 * Update run-time statistics of the 'current'.
3073 update_curr(cfs_rq);
3074 dequeue_entity_load_avg(cfs_rq, se);
3076 update_stats_dequeue(cfs_rq, se);
3077 if (flags & DEQUEUE_SLEEP) {
3078 #ifdef CONFIG_SCHEDSTATS
3079 if (entity_is_task(se)) {
3080 struct task_struct *tsk = task_of(se);
3082 if (tsk->state & TASK_INTERRUPTIBLE)
3083 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3084 if (tsk->state & TASK_UNINTERRUPTIBLE)
3085 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3090 clear_buddies(cfs_rq, se);
3092 if (se != cfs_rq->curr)
3093 __dequeue_entity(cfs_rq, se);
3095 account_entity_dequeue(cfs_rq, se);
3098 * Normalize the entity after updating the min_vruntime because the
3099 * update can refer to the ->curr item and we need to reflect this
3100 * movement in our normalized position.
3102 if (!(flags & DEQUEUE_SLEEP))
3103 se->vruntime -= cfs_rq->min_vruntime;
3105 /* return excess runtime on last dequeue */
3106 return_cfs_rq_runtime(cfs_rq);
3108 update_min_vruntime(cfs_rq);
3109 update_cfs_shares(cfs_rq);
3113 * Preempt the current task with a newly woken task if needed:
3116 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3118 unsigned long ideal_runtime, delta_exec;
3119 struct sched_entity *se;
3122 ideal_runtime = sched_slice(cfs_rq, curr);
3123 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3124 if (delta_exec > ideal_runtime) {
3125 resched_curr(rq_of(cfs_rq));
3127 * The current task ran long enough, ensure it doesn't get
3128 * re-elected due to buddy favours.
3130 clear_buddies(cfs_rq, curr);
3135 * Ensure that a task that missed wakeup preemption by a
3136 * narrow margin doesn't have to wait for a full slice.
3137 * This also mitigates buddy induced latencies under load.
3139 if (delta_exec < sysctl_sched_min_granularity)
3142 se = __pick_first_entity(cfs_rq);
3143 delta = curr->vruntime - se->vruntime;
3148 if (delta > ideal_runtime)
3149 resched_curr(rq_of(cfs_rq));
3153 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3155 /* 'current' is not kept within the tree. */
3158 * Any task has to be enqueued before it get to execute on
3159 * a CPU. So account for the time it spent waiting on the
3162 update_stats_wait_end(cfs_rq, se);
3163 __dequeue_entity(cfs_rq, se);
3164 update_load_avg(se, 1);
3167 update_stats_curr_start(cfs_rq, se);
3169 #ifdef CONFIG_SCHEDSTATS
3171 * Track our maximum slice length, if the CPU's load is at
3172 * least twice that of our own weight (i.e. dont track it
3173 * when there are only lesser-weight tasks around):
3175 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3176 se->statistics.slice_max = max(se->statistics.slice_max,
3177 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3180 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3184 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3187 * Pick the next process, keeping these things in mind, in this order:
3188 * 1) keep things fair between processes/task groups
3189 * 2) pick the "next" process, since someone really wants that to run
3190 * 3) pick the "last" process, for cache locality
3191 * 4) do not run the "skip" process, if something else is available
3193 static struct sched_entity *
3194 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3196 struct sched_entity *left = __pick_first_entity(cfs_rq);
3197 struct sched_entity *se;
3200 * If curr is set we have to see if its left of the leftmost entity
3201 * still in the tree, provided there was anything in the tree at all.
3203 if (!left || (curr && entity_before(curr, left)))
3206 se = left; /* ideally we run the leftmost entity */
3209 * Avoid running the skip buddy, if running something else can
3210 * be done without getting too unfair.
3212 if (cfs_rq->skip == se) {
3213 struct sched_entity *second;
3216 second = __pick_first_entity(cfs_rq);
3218 second = __pick_next_entity(se);
3219 if (!second || (curr && entity_before(curr, second)))
3223 if (second && wakeup_preempt_entity(second, left) < 1)
3228 * Prefer last buddy, try to return the CPU to a preempted task.
3230 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3234 * Someone really wants this to run. If it's not unfair, run it.
3236 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3239 clear_buddies(cfs_rq, se);
3244 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3246 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3249 * If still on the runqueue then deactivate_task()
3250 * was not called and update_curr() has to be done:
3253 update_curr(cfs_rq);
3255 /* throttle cfs_rqs exceeding runtime */
3256 check_cfs_rq_runtime(cfs_rq);
3258 check_spread(cfs_rq, prev);
3260 update_stats_wait_start(cfs_rq, prev);
3261 /* Put 'current' back into the tree. */
3262 __enqueue_entity(cfs_rq, prev);
3263 /* in !on_rq case, update occurred at dequeue */
3264 update_load_avg(prev, 0);
3266 cfs_rq->curr = NULL;
3270 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3273 * Update run-time statistics of the 'current'.
3275 update_curr(cfs_rq);
3278 * Ensure that runnable average is periodically updated.
3280 update_load_avg(curr, 1);
3281 update_cfs_shares(cfs_rq);
3283 #ifdef CONFIG_SCHED_HRTICK
3285 * queued ticks are scheduled to match the slice, so don't bother
3286 * validating it and just reschedule.
3289 resched_curr(rq_of(cfs_rq));
3293 * don't let the period tick interfere with the hrtick preemption
3295 if (!sched_feat(DOUBLE_TICK) &&
3296 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3300 if (cfs_rq->nr_running > 1)
3301 check_preempt_tick(cfs_rq, curr);
3305 /**************************************************
3306 * CFS bandwidth control machinery
3309 #ifdef CONFIG_CFS_BANDWIDTH
3311 #ifdef HAVE_JUMP_LABEL
3312 static struct static_key __cfs_bandwidth_used;
3314 static inline bool cfs_bandwidth_used(void)
3316 return static_key_false(&__cfs_bandwidth_used);
3319 void cfs_bandwidth_usage_inc(void)
3321 static_key_slow_inc(&__cfs_bandwidth_used);
3324 void cfs_bandwidth_usage_dec(void)
3326 static_key_slow_dec(&__cfs_bandwidth_used);
3328 #else /* HAVE_JUMP_LABEL */
3329 static bool cfs_bandwidth_used(void)
3334 void cfs_bandwidth_usage_inc(void) {}
3335 void cfs_bandwidth_usage_dec(void) {}
3336 #endif /* HAVE_JUMP_LABEL */
3339 * default period for cfs group bandwidth.
3340 * default: 0.1s, units: nanoseconds
3342 static inline u64 default_cfs_period(void)
3344 return 100000000ULL;
3347 static inline u64 sched_cfs_bandwidth_slice(void)
3349 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3353 * Replenish runtime according to assigned quota and update expiration time.
3354 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3355 * additional synchronization around rq->lock.
3357 * requires cfs_b->lock
3359 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3363 if (cfs_b->quota == RUNTIME_INF)
3366 now = sched_clock_cpu(smp_processor_id());
3367 cfs_b->runtime = cfs_b->quota;
3368 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3371 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3373 return &tg->cfs_bandwidth;
3376 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3377 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3379 if (unlikely(cfs_rq->throttle_count))
3380 return cfs_rq->throttled_clock_task;
3382 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3385 /* returns 0 on failure to allocate runtime */
3386 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3388 struct task_group *tg = cfs_rq->tg;
3389 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3390 u64 amount = 0, min_amount, expires;
3392 /* note: this is a positive sum as runtime_remaining <= 0 */
3393 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3395 raw_spin_lock(&cfs_b->lock);
3396 if (cfs_b->quota == RUNTIME_INF)
3397 amount = min_amount;
3399 start_cfs_bandwidth(cfs_b);
3401 if (cfs_b->runtime > 0) {
3402 amount = min(cfs_b->runtime, min_amount);
3403 cfs_b->runtime -= amount;
3407 expires = cfs_b->runtime_expires;
3408 raw_spin_unlock(&cfs_b->lock);
3410 cfs_rq->runtime_remaining += amount;
3412 * we may have advanced our local expiration to account for allowed
3413 * spread between our sched_clock and the one on which runtime was
3416 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3417 cfs_rq->runtime_expires = expires;
3419 return cfs_rq->runtime_remaining > 0;
3423 * Note: This depends on the synchronization provided by sched_clock and the
3424 * fact that rq->clock snapshots this value.
3426 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3428 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3430 /* if the deadline is ahead of our clock, nothing to do */
3431 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3434 if (cfs_rq->runtime_remaining < 0)
3438 * If the local deadline has passed we have to consider the
3439 * possibility that our sched_clock is 'fast' and the global deadline
3440 * has not truly expired.
3442 * Fortunately we can check determine whether this the case by checking
3443 * whether the global deadline has advanced. It is valid to compare
3444 * cfs_b->runtime_expires without any locks since we only care about
3445 * exact equality, so a partial write will still work.
3448 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3449 /* extend local deadline, drift is bounded above by 2 ticks */
3450 cfs_rq->runtime_expires += TICK_NSEC;
3452 /* global deadline is ahead, expiration has passed */
3453 cfs_rq->runtime_remaining = 0;
3457 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3459 /* dock delta_exec before expiring quota (as it could span periods) */
3460 cfs_rq->runtime_remaining -= delta_exec;
3461 expire_cfs_rq_runtime(cfs_rq);
3463 if (likely(cfs_rq->runtime_remaining > 0))
3467 * if we're unable to extend our runtime we resched so that the active
3468 * hierarchy can be throttled
3470 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3471 resched_curr(rq_of(cfs_rq));
3474 static __always_inline
3475 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3477 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3480 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3483 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3485 return cfs_bandwidth_used() && cfs_rq->throttled;
3488 /* check whether cfs_rq, or any parent, is throttled */
3489 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3491 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3495 * Ensure that neither of the group entities corresponding to src_cpu or
3496 * dest_cpu are members of a throttled hierarchy when performing group
3497 * load-balance operations.
3499 static inline int throttled_lb_pair(struct task_group *tg,
3500 int src_cpu, int dest_cpu)
3502 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3504 src_cfs_rq = tg->cfs_rq[src_cpu];
3505 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3507 return throttled_hierarchy(src_cfs_rq) ||
3508 throttled_hierarchy(dest_cfs_rq);
3511 /* updated child weight may affect parent so we have to do this bottom up */
3512 static int tg_unthrottle_up(struct task_group *tg, void *data)
3514 struct rq *rq = data;
3515 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3517 cfs_rq->throttle_count--;
3519 if (!cfs_rq->throttle_count) {
3520 /* adjust cfs_rq_clock_task() */
3521 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3522 cfs_rq->throttled_clock_task;
3529 static int tg_throttle_down(struct task_group *tg, void *data)
3531 struct rq *rq = data;
3532 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3534 /* group is entering throttled state, stop time */
3535 if (!cfs_rq->throttle_count)
3536 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3537 cfs_rq->throttle_count++;
3542 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3544 struct rq *rq = rq_of(cfs_rq);
3545 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3546 struct sched_entity *se;
3547 long task_delta, dequeue = 1;
3550 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3552 /* freeze hierarchy runnable averages while throttled */
3554 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3557 task_delta = cfs_rq->h_nr_running;
3558 for_each_sched_entity(se) {
3559 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3560 /* throttled entity or throttle-on-deactivate */
3565 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3566 qcfs_rq->h_nr_running -= task_delta;
3568 if (qcfs_rq->load.weight)
3573 sub_nr_running(rq, task_delta);
3575 cfs_rq->throttled = 1;
3576 cfs_rq->throttled_clock = rq_clock(rq);
3577 raw_spin_lock(&cfs_b->lock);
3578 empty = list_empty(&cfs_b->throttled_cfs_rq);
3581 * Add to the _head_ of the list, so that an already-started
3582 * distribute_cfs_runtime will not see us
3584 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3587 * If we're the first throttled task, make sure the bandwidth
3591 start_cfs_bandwidth(cfs_b);
3593 raw_spin_unlock(&cfs_b->lock);
3596 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3598 struct rq *rq = rq_of(cfs_rq);
3599 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3600 struct sched_entity *se;
3604 se = cfs_rq->tg->se[cpu_of(rq)];
3606 cfs_rq->throttled = 0;
3608 update_rq_clock(rq);
3610 raw_spin_lock(&cfs_b->lock);
3611 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3612 list_del_rcu(&cfs_rq->throttled_list);
3613 raw_spin_unlock(&cfs_b->lock);
3615 /* update hierarchical throttle state */
3616 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3618 if (!cfs_rq->load.weight)
3621 task_delta = cfs_rq->h_nr_running;
3622 for_each_sched_entity(se) {
3626 cfs_rq = cfs_rq_of(se);
3628 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3629 cfs_rq->h_nr_running += task_delta;
3631 if (cfs_rq_throttled(cfs_rq))
3636 add_nr_running(rq, task_delta);
3638 /* determine whether we need to wake up potentially idle cpu */
3639 if (rq->curr == rq->idle && rq->cfs.nr_running)
3643 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3644 u64 remaining, u64 expires)
3646 struct cfs_rq *cfs_rq;
3648 u64 starting_runtime = remaining;
3651 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3653 struct rq *rq = rq_of(cfs_rq);
3655 raw_spin_lock(&rq->lock);
3656 if (!cfs_rq_throttled(cfs_rq))
3659 runtime = -cfs_rq->runtime_remaining + 1;
3660 if (runtime > remaining)
3661 runtime = remaining;
3662 remaining -= runtime;
3664 cfs_rq->runtime_remaining += runtime;
3665 cfs_rq->runtime_expires = expires;
3667 /* we check whether we're throttled above */
3668 if (cfs_rq->runtime_remaining > 0)
3669 unthrottle_cfs_rq(cfs_rq);
3672 raw_spin_unlock(&rq->lock);
3679 return starting_runtime - remaining;
3683 * Responsible for refilling a task_group's bandwidth and unthrottling its
3684 * cfs_rqs as appropriate. If there has been no activity within the last
3685 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3686 * used to track this state.
3688 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3690 u64 runtime, runtime_expires;
3693 /* no need to continue the timer with no bandwidth constraint */
3694 if (cfs_b->quota == RUNTIME_INF)
3695 goto out_deactivate;
3697 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3698 cfs_b->nr_periods += overrun;
3701 * idle depends on !throttled (for the case of a large deficit), and if
3702 * we're going inactive then everything else can be deferred
3704 if (cfs_b->idle && !throttled)
3705 goto out_deactivate;
3707 __refill_cfs_bandwidth_runtime(cfs_b);
3710 /* mark as potentially idle for the upcoming period */
3715 /* account preceding periods in which throttling occurred */
3716 cfs_b->nr_throttled += overrun;
3718 runtime_expires = cfs_b->runtime_expires;
3721 * This check is repeated as we are holding onto the new bandwidth while
3722 * we unthrottle. This can potentially race with an unthrottled group
3723 * trying to acquire new bandwidth from the global pool. This can result
3724 * in us over-using our runtime if it is all used during this loop, but
3725 * only by limited amounts in that extreme case.
3727 while (throttled && cfs_b->runtime > 0) {
3728 runtime = cfs_b->runtime;
3729 raw_spin_unlock(&cfs_b->lock);
3730 /* we can't nest cfs_b->lock while distributing bandwidth */
3731 runtime = distribute_cfs_runtime(cfs_b, runtime,
3733 raw_spin_lock(&cfs_b->lock);
3735 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3737 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3741 * While we are ensured activity in the period following an
3742 * unthrottle, this also covers the case in which the new bandwidth is
3743 * insufficient to cover the existing bandwidth deficit. (Forcing the
3744 * timer to remain active while there are any throttled entities.)
3754 /* a cfs_rq won't donate quota below this amount */
3755 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3756 /* minimum remaining period time to redistribute slack quota */
3757 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3758 /* how long we wait to gather additional slack before distributing */
3759 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3762 * Are we near the end of the current quota period?
3764 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3765 * hrtimer base being cleared by hrtimer_start. In the case of
3766 * migrate_hrtimers, base is never cleared, so we are fine.
3768 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3770 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3773 /* if the call-back is running a quota refresh is already occurring */
3774 if (hrtimer_callback_running(refresh_timer))
3777 /* is a quota refresh about to occur? */
3778 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3779 if (remaining < min_expire)
3785 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3787 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3789 /* if there's a quota refresh soon don't bother with slack */
3790 if (runtime_refresh_within(cfs_b, min_left))
3793 hrtimer_start(&cfs_b->slack_timer,
3794 ns_to_ktime(cfs_bandwidth_slack_period),
3798 /* we know any runtime found here is valid as update_curr() precedes return */
3799 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3801 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3802 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3804 if (slack_runtime <= 0)
3807 raw_spin_lock(&cfs_b->lock);
3808 if (cfs_b->quota != RUNTIME_INF &&
3809 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3810 cfs_b->runtime += slack_runtime;
3812 /* we are under rq->lock, defer unthrottling using a timer */
3813 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3814 !list_empty(&cfs_b->throttled_cfs_rq))
3815 start_cfs_slack_bandwidth(cfs_b);
3817 raw_spin_unlock(&cfs_b->lock);
3819 /* even if it's not valid for return we don't want to try again */
3820 cfs_rq->runtime_remaining -= slack_runtime;
3823 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3825 if (!cfs_bandwidth_used())
3828 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3831 __return_cfs_rq_runtime(cfs_rq);
3835 * This is done with a timer (instead of inline with bandwidth return) since
3836 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3838 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3840 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3843 /* confirm we're still not at a refresh boundary */
3844 raw_spin_lock(&cfs_b->lock);
3845 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3846 raw_spin_unlock(&cfs_b->lock);
3850 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3851 runtime = cfs_b->runtime;
3853 expires = cfs_b->runtime_expires;
3854 raw_spin_unlock(&cfs_b->lock);
3859 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3861 raw_spin_lock(&cfs_b->lock);
3862 if (expires == cfs_b->runtime_expires)
3863 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3864 raw_spin_unlock(&cfs_b->lock);
3868 * When a group wakes up we want to make sure that its quota is not already
3869 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3870 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3872 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3874 if (!cfs_bandwidth_used())
3877 /* an active group must be handled by the update_curr()->put() path */
3878 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3881 /* ensure the group is not already throttled */
3882 if (cfs_rq_throttled(cfs_rq))
3885 /* update runtime allocation */
3886 account_cfs_rq_runtime(cfs_rq, 0);
3887 if (cfs_rq->runtime_remaining <= 0)
3888 throttle_cfs_rq(cfs_rq);
3891 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3892 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3894 if (!cfs_bandwidth_used())
3897 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3901 * it's possible for a throttled entity to be forced into a running
3902 * state (e.g. set_curr_task), in this case we're finished.
3904 if (cfs_rq_throttled(cfs_rq))
3907 throttle_cfs_rq(cfs_rq);
3911 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3913 struct cfs_bandwidth *cfs_b =
3914 container_of(timer, struct cfs_bandwidth, slack_timer);
3916 do_sched_cfs_slack_timer(cfs_b);
3918 return HRTIMER_NORESTART;
3921 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3923 struct cfs_bandwidth *cfs_b =
3924 container_of(timer, struct cfs_bandwidth, period_timer);
3928 raw_spin_lock(&cfs_b->lock);
3930 overrun = hrtimer_forward_now(timer, cfs_b->period);
3934 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3937 cfs_b->period_active = 0;
3938 raw_spin_unlock(&cfs_b->lock);
3940 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3943 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3945 raw_spin_lock_init(&cfs_b->lock);
3947 cfs_b->quota = RUNTIME_INF;
3948 cfs_b->period = ns_to_ktime(default_cfs_period());
3950 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3951 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
3952 cfs_b->period_timer.function = sched_cfs_period_timer;
3953 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3954 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3957 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3959 cfs_rq->runtime_enabled = 0;
3960 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3963 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3965 lockdep_assert_held(&cfs_b->lock);
3967 if (!cfs_b->period_active) {
3968 cfs_b->period_active = 1;
3969 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
3970 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
3974 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3976 /* init_cfs_bandwidth() was not called */
3977 if (!cfs_b->throttled_cfs_rq.next)
3980 hrtimer_cancel(&cfs_b->period_timer);
3981 hrtimer_cancel(&cfs_b->slack_timer);
3984 static void __maybe_unused update_runtime_enabled(struct rq *rq)
3986 struct cfs_rq *cfs_rq;
3988 for_each_leaf_cfs_rq(rq, cfs_rq) {
3989 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
3991 raw_spin_lock(&cfs_b->lock);
3992 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
3993 raw_spin_unlock(&cfs_b->lock);
3997 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3999 struct cfs_rq *cfs_rq;
4001 for_each_leaf_cfs_rq(rq, cfs_rq) {
4002 if (!cfs_rq->runtime_enabled)
4006 * clock_task is not advancing so we just need to make sure
4007 * there's some valid quota amount
4009 cfs_rq->runtime_remaining = 1;
4011 * Offline rq is schedulable till cpu is completely disabled
4012 * in take_cpu_down(), so we prevent new cfs throttling here.
4014 cfs_rq->runtime_enabled = 0;
4016 if (cfs_rq_throttled(cfs_rq))
4017 unthrottle_cfs_rq(cfs_rq);
4021 #else /* CONFIG_CFS_BANDWIDTH */
4022 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4024 return rq_clock_task(rq_of(cfs_rq));
4027 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4028 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4029 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4030 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4032 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4037 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4042 static inline int throttled_lb_pair(struct task_group *tg,
4043 int src_cpu, int dest_cpu)
4048 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4050 #ifdef CONFIG_FAIR_GROUP_SCHED
4051 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4054 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4058 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4059 static inline void update_runtime_enabled(struct rq *rq) {}
4060 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4062 #endif /* CONFIG_CFS_BANDWIDTH */
4064 /**************************************************
4065 * CFS operations on tasks:
4068 #ifdef CONFIG_SCHED_HRTICK
4069 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4071 struct sched_entity *se = &p->se;
4072 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4074 WARN_ON(task_rq(p) != rq);
4076 if (cfs_rq->nr_running > 1) {
4077 u64 slice = sched_slice(cfs_rq, se);
4078 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4079 s64 delta = slice - ran;
4086 hrtick_start(rq, delta);
4091 * called from enqueue/dequeue and updates the hrtick when the
4092 * current task is from our class and nr_running is low enough
4095 static void hrtick_update(struct rq *rq)
4097 struct task_struct *curr = rq->curr;
4099 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4102 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4103 hrtick_start_fair(rq, curr);
4105 #else /* !CONFIG_SCHED_HRTICK */
4107 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4111 static inline void hrtick_update(struct rq *rq)
4117 * The enqueue_task method is called before nr_running is
4118 * increased. Here we update the fair scheduling stats and
4119 * then put the task into the rbtree:
4122 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4124 struct cfs_rq *cfs_rq;
4125 struct sched_entity *se = &p->se;
4127 for_each_sched_entity(se) {
4130 cfs_rq = cfs_rq_of(se);
4131 enqueue_entity(cfs_rq, se, flags);
4134 * end evaluation on encountering a throttled cfs_rq
4136 * note: in the case of encountering a throttled cfs_rq we will
4137 * post the final h_nr_running increment below.
4139 if (cfs_rq_throttled(cfs_rq))
4141 cfs_rq->h_nr_running++;
4143 flags = ENQUEUE_WAKEUP;
4146 for_each_sched_entity(se) {
4147 cfs_rq = cfs_rq_of(se);
4148 cfs_rq->h_nr_running++;
4150 if (cfs_rq_throttled(cfs_rq))
4153 update_load_avg(se, 1);
4154 update_cfs_shares(cfs_rq);
4158 add_nr_running(rq, 1);
4163 static void set_next_buddy(struct sched_entity *se);
4166 * The dequeue_task method is called before nr_running is
4167 * decreased. We remove the task from the rbtree and
4168 * update the fair scheduling stats:
4170 static void dequeue_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_sleep = flags & DEQUEUE_SLEEP;
4176 for_each_sched_entity(se) {
4177 cfs_rq = cfs_rq_of(se);
4178 dequeue_entity(cfs_rq, se, flags);
4181 * end evaluation on encountering a throttled cfs_rq
4183 * note: in the case of encountering a throttled cfs_rq we will
4184 * post the final h_nr_running decrement below.
4186 if (cfs_rq_throttled(cfs_rq))
4188 cfs_rq->h_nr_running--;
4190 /* Don't dequeue parent if it has other entities besides us */
4191 if (cfs_rq->load.weight) {
4193 * Bias pick_next to pick a task from this cfs_rq, as
4194 * p is sleeping when it is within its sched_slice.
4196 if (task_sleep && parent_entity(se))
4197 set_next_buddy(parent_entity(se));
4199 /* avoid re-evaluating load for this entity */
4200 se = parent_entity(se);
4203 flags |= DEQUEUE_SLEEP;
4206 for_each_sched_entity(se) {
4207 cfs_rq = cfs_rq_of(se);
4208 cfs_rq->h_nr_running--;
4210 if (cfs_rq_throttled(cfs_rq))
4213 update_load_avg(se, 1);
4214 update_cfs_shares(cfs_rq);
4218 sub_nr_running(rq, 1);
4226 * per rq 'load' arrray crap; XXX kill this.
4230 * The exact cpuload at various idx values, calculated at every tick would be
4231 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4233 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4234 * on nth tick when cpu may be busy, then we have:
4235 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4236 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4238 * decay_load_missed() below does efficient calculation of
4239 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4240 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4242 * The calculation is approximated on a 128 point scale.
4243 * degrade_zero_ticks is the number of ticks after which load at any
4244 * particular idx is approximated to be zero.
4245 * degrade_factor is a precomputed table, a row for each load idx.
4246 * Each column corresponds to degradation factor for a power of two ticks,
4247 * based on 128 point scale.
4249 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4250 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4252 * With this power of 2 load factors, we can degrade the load n times
4253 * by looking at 1 bits in n and doing as many mult/shift instead of
4254 * n mult/shifts needed by the exact degradation.
4256 #define DEGRADE_SHIFT 7
4257 static const unsigned char
4258 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4259 static const unsigned char
4260 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4261 {0, 0, 0, 0, 0, 0, 0, 0},
4262 {64, 32, 8, 0, 0, 0, 0, 0},
4263 {96, 72, 40, 12, 1, 0, 0},
4264 {112, 98, 75, 43, 15, 1, 0},
4265 {120, 112, 98, 76, 45, 16, 2} };
4268 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4269 * would be when CPU is idle and so we just decay the old load without
4270 * adding any new load.
4272 static unsigned long
4273 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4277 if (!missed_updates)
4280 if (missed_updates >= degrade_zero_ticks[idx])
4284 return load >> missed_updates;
4286 while (missed_updates) {
4287 if (missed_updates % 2)
4288 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4290 missed_updates >>= 1;
4297 * Update rq->cpu_load[] statistics. This function is usually called every
4298 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4299 * every tick. We fix it up based on jiffies.
4301 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4302 unsigned long pending_updates)
4306 this_rq->nr_load_updates++;
4308 /* Update our load: */
4309 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4310 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4311 unsigned long old_load, new_load;
4313 /* scale is effectively 1 << i now, and >> i divides by scale */
4315 old_load = this_rq->cpu_load[i];
4316 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4317 new_load = this_load;
4319 * Round up the averaging division if load is increasing. This
4320 * prevents us from getting stuck on 9 if the load is 10, for
4323 if (new_load > old_load)
4324 new_load += scale - 1;
4326 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4329 sched_avg_update(this_rq);
4332 /* Used instead of source_load when we know the type == 0 */
4333 static unsigned long weighted_cpuload(const int cpu)
4335 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4338 #ifdef CONFIG_NO_HZ_COMMON
4340 * There is no sane way to deal with nohz on smp when using jiffies because the
4341 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4342 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4344 * Therefore we cannot use the delta approach from the regular tick since that
4345 * would seriously skew the load calculation. However we'll make do for those
4346 * updates happening while idle (nohz_idle_balance) or coming out of idle
4347 * (tick_nohz_idle_exit).
4349 * This means we might still be one tick off for nohz periods.
4353 * Called from nohz_idle_balance() to update the load ratings before doing the
4356 static void update_idle_cpu_load(struct rq *this_rq)
4358 unsigned long curr_jiffies = READ_ONCE(jiffies);
4359 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4360 unsigned long pending_updates;
4363 * bail if there's load or we're actually up-to-date.
4365 if (load || curr_jiffies == this_rq->last_load_update_tick)
4368 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4369 this_rq->last_load_update_tick = curr_jiffies;
4371 __update_cpu_load(this_rq, load, pending_updates);
4375 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4377 void update_cpu_load_nohz(void)
4379 struct rq *this_rq = this_rq();
4380 unsigned long curr_jiffies = READ_ONCE(jiffies);
4381 unsigned long pending_updates;
4383 if (curr_jiffies == this_rq->last_load_update_tick)
4386 raw_spin_lock(&this_rq->lock);
4387 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4388 if (pending_updates) {
4389 this_rq->last_load_update_tick = curr_jiffies;
4391 * We were idle, this means load 0, the current load might be
4392 * !0 due to remote wakeups and the sort.
4394 __update_cpu_load(this_rq, 0, pending_updates);
4396 raw_spin_unlock(&this_rq->lock);
4398 #endif /* CONFIG_NO_HZ */
4401 * Called from scheduler_tick()
4403 void update_cpu_load_active(struct rq *this_rq)
4405 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4407 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4409 this_rq->last_load_update_tick = jiffies;
4410 __update_cpu_load(this_rq, load, 1);
4414 * Return a low guess at the load of a migration-source cpu weighted
4415 * according to the scheduling class and "nice" value.
4417 * We want to under-estimate the load of migration sources, to
4418 * balance conservatively.
4420 static unsigned long source_load(int cpu, int type)
4422 struct rq *rq = cpu_rq(cpu);
4423 unsigned long total = weighted_cpuload(cpu);
4425 if (type == 0 || !sched_feat(LB_BIAS))
4428 return min(rq->cpu_load[type-1], total);
4432 * Return a high guess at the load of a migration-target cpu weighted
4433 * according to the scheduling class and "nice" value.
4435 static unsigned long target_load(int cpu, int type)
4437 struct rq *rq = cpu_rq(cpu);
4438 unsigned long total = weighted_cpuload(cpu);
4440 if (type == 0 || !sched_feat(LB_BIAS))
4443 return max(rq->cpu_load[type-1], total);
4446 static unsigned long capacity_of(int cpu)
4448 return cpu_rq(cpu)->cpu_capacity;
4451 static unsigned long capacity_orig_of(int cpu)
4453 return cpu_rq(cpu)->cpu_capacity_orig;
4456 static unsigned long cpu_avg_load_per_task(int cpu)
4458 struct rq *rq = cpu_rq(cpu);
4459 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4460 unsigned long load_avg = weighted_cpuload(cpu);
4463 return load_avg / nr_running;
4468 static void record_wakee(struct task_struct *p)
4471 * Rough decay (wiping) for cost saving, don't worry
4472 * about the boundary, really active task won't care
4475 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4476 current->wakee_flips >>= 1;
4477 current->wakee_flip_decay_ts = jiffies;
4480 if (current->last_wakee != p) {
4481 current->last_wakee = p;
4482 current->wakee_flips++;
4486 static void task_waking_fair(struct task_struct *p)
4488 struct sched_entity *se = &p->se;
4489 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4492 #ifndef CONFIG_64BIT
4493 u64 min_vruntime_copy;
4496 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4498 min_vruntime = cfs_rq->min_vruntime;
4499 } while (min_vruntime != min_vruntime_copy);
4501 min_vruntime = cfs_rq->min_vruntime;
4504 se->vruntime -= min_vruntime;
4508 #ifdef CONFIG_FAIR_GROUP_SCHED
4510 * effective_load() calculates the load change as seen from the root_task_group
4512 * Adding load to a group doesn't make a group heavier, but can cause movement
4513 * of group shares between cpus. Assuming the shares were perfectly aligned one
4514 * can calculate the shift in shares.
4516 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4517 * on this @cpu and results in a total addition (subtraction) of @wg to the
4518 * total group weight.
4520 * Given a runqueue weight distribution (rw_i) we can compute a shares
4521 * distribution (s_i) using:
4523 * s_i = rw_i / \Sum rw_j (1)
4525 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4526 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4527 * shares distribution (s_i):
4529 * rw_i = { 2, 4, 1, 0 }
4530 * s_i = { 2/7, 4/7, 1/7, 0 }
4532 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4533 * task used to run on and the CPU the waker is running on), we need to
4534 * compute the effect of waking a task on either CPU and, in case of a sync
4535 * wakeup, compute the effect of the current task going to sleep.
4537 * So for a change of @wl to the local @cpu with an overall group weight change
4538 * of @wl we can compute the new shares distribution (s'_i) using:
4540 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4542 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4543 * differences in waking a task to CPU 0. The additional task changes the
4544 * weight and shares distributions like:
4546 * rw'_i = { 3, 4, 1, 0 }
4547 * s'_i = { 3/8, 4/8, 1/8, 0 }
4549 * We can then compute the difference in effective weight by using:
4551 * dw_i = S * (s'_i - s_i) (3)
4553 * Where 'S' is the group weight as seen by its parent.
4555 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4556 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4557 * 4/7) times the weight of the group.
4559 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4561 struct sched_entity *se = tg->se[cpu];
4563 if (!tg->parent) /* the trivial, non-cgroup case */
4566 for_each_sched_entity(se) {
4572 * W = @wg + \Sum rw_j
4574 W = wg + calc_tg_weight(tg, se->my_q);
4579 w = cfs_rq_load_avg(se->my_q) + wl;
4582 * wl = S * s'_i; see (2)
4585 wl = (w * (long)tg->shares) / W;
4590 * Per the above, wl is the new se->load.weight value; since
4591 * those are clipped to [MIN_SHARES, ...) do so now. See
4592 * calc_cfs_shares().
4594 if (wl < MIN_SHARES)
4598 * wl = dw_i = S * (s'_i - s_i); see (3)
4600 wl -= se->avg.load_avg;
4603 * Recursively apply this logic to all parent groups to compute
4604 * the final effective load change on the root group. Since
4605 * only the @tg group gets extra weight, all parent groups can
4606 * only redistribute existing shares. @wl is the shift in shares
4607 * resulting from this level per the above.
4616 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4624 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4625 * A waker of many should wake a different task than the one last awakened
4626 * at a frequency roughly N times higher than one of its wakees. In order
4627 * to determine whether we should let the load spread vs consolodating to
4628 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4629 * partner, and a factor of lls_size higher frequency in the other. With
4630 * both conditions met, we can be relatively sure that the relationship is
4631 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4632 * being client/server, worker/dispatcher, interrupt source or whatever is
4633 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4635 static int wake_wide(struct task_struct *p)
4637 unsigned int master = current->wakee_flips;
4638 unsigned int slave = p->wakee_flips;
4639 int factor = this_cpu_read(sd_llc_size);
4642 swap(master, slave);
4643 if (slave < factor || master < slave * factor)
4648 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4650 s64 this_load, load;
4651 s64 this_eff_load, prev_eff_load;
4652 int idx, this_cpu, prev_cpu;
4653 struct task_group *tg;
4654 unsigned long weight;
4658 this_cpu = smp_processor_id();
4659 prev_cpu = task_cpu(p);
4660 load = source_load(prev_cpu, idx);
4661 this_load = target_load(this_cpu, idx);
4664 * If sync wakeup then subtract the (maximum possible)
4665 * effect of the currently running task from the load
4666 * of the current CPU:
4669 tg = task_group(current);
4670 weight = current->se.avg.load_avg;
4672 this_load += effective_load(tg, this_cpu, -weight, -weight);
4673 load += effective_load(tg, prev_cpu, 0, -weight);
4677 weight = p->se.avg.load_avg;
4680 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4681 * due to the sync cause above having dropped this_load to 0, we'll
4682 * always have an imbalance, but there's really nothing you can do
4683 * about that, so that's good too.
4685 * Otherwise check if either cpus are near enough in load to allow this
4686 * task to be woken on this_cpu.
4688 this_eff_load = 100;
4689 this_eff_load *= capacity_of(prev_cpu);
4691 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4692 prev_eff_load *= capacity_of(this_cpu);
4694 if (this_load > 0) {
4695 this_eff_load *= this_load +
4696 effective_load(tg, this_cpu, weight, weight);
4698 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4701 balanced = this_eff_load <= prev_eff_load;
4703 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4708 schedstat_inc(sd, ttwu_move_affine);
4709 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4715 * find_idlest_group finds and returns the least busy CPU group within the
4718 static struct sched_group *
4719 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4720 int this_cpu, int sd_flag)
4722 struct sched_group *idlest = NULL, *group = sd->groups;
4723 unsigned long min_load = ULONG_MAX, this_load = 0;
4724 int load_idx = sd->forkexec_idx;
4725 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4727 if (sd_flag & SD_BALANCE_WAKE)
4728 load_idx = sd->wake_idx;
4731 unsigned long load, avg_load;
4735 /* Skip over this group if it has no CPUs allowed */
4736 if (!cpumask_intersects(sched_group_cpus(group),
4737 tsk_cpus_allowed(p)))
4740 local_group = cpumask_test_cpu(this_cpu,
4741 sched_group_cpus(group));
4743 /* Tally up the load of all CPUs in the group */
4746 for_each_cpu(i, sched_group_cpus(group)) {
4747 /* Bias balancing toward cpus of our domain */
4749 load = source_load(i, load_idx);
4751 load = target_load(i, load_idx);
4756 /* Adjust by relative CPU capacity of the group */
4757 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4760 this_load = avg_load;
4761 } else if (avg_load < min_load) {
4762 min_load = avg_load;
4765 } while (group = group->next, group != sd->groups);
4767 if (!idlest || 100*this_load < imbalance*min_load)
4773 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4776 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4778 unsigned long load, min_load = ULONG_MAX;
4779 unsigned int min_exit_latency = UINT_MAX;
4780 u64 latest_idle_timestamp = 0;
4781 int least_loaded_cpu = this_cpu;
4782 int shallowest_idle_cpu = -1;
4785 /* Traverse only the allowed CPUs */
4786 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4788 struct rq *rq = cpu_rq(i);
4789 struct cpuidle_state *idle = idle_get_state(rq);
4790 if (idle && idle->exit_latency < min_exit_latency) {
4792 * We give priority to a CPU whose idle state
4793 * has the smallest exit latency irrespective
4794 * of any idle timestamp.
4796 min_exit_latency = idle->exit_latency;
4797 latest_idle_timestamp = rq->idle_stamp;
4798 shallowest_idle_cpu = i;
4799 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4800 rq->idle_stamp > latest_idle_timestamp) {
4802 * If equal or no active idle state, then
4803 * the most recently idled CPU might have
4806 latest_idle_timestamp = rq->idle_stamp;
4807 shallowest_idle_cpu = i;
4809 } else if (shallowest_idle_cpu == -1) {
4810 load = weighted_cpuload(i);
4811 if (load < min_load || (load == min_load && i == this_cpu)) {
4813 least_loaded_cpu = i;
4818 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4822 * Try and locate an idle CPU in the sched_domain.
4824 static int select_idle_sibling(struct task_struct *p, int target)
4826 struct sched_domain *sd;
4827 struct sched_group *sg;
4828 int i = task_cpu(p);
4830 if (idle_cpu(target))
4834 * If the prevous cpu is cache affine and idle, don't be stupid.
4836 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4840 * Otherwise, iterate the domains and find an elegible idle cpu.
4842 sd = rcu_dereference(per_cpu(sd_llc, target));
4843 for_each_lower_domain(sd) {
4846 if (!cpumask_intersects(sched_group_cpus(sg),
4847 tsk_cpus_allowed(p)))
4850 for_each_cpu(i, sched_group_cpus(sg)) {
4851 if (i == target || !idle_cpu(i))
4855 target = cpumask_first_and(sched_group_cpus(sg),
4856 tsk_cpus_allowed(p));
4860 } while (sg != sd->groups);
4867 * cpu_util returns the amount of capacity of a CPU that is used by CFS
4868 * tasks. The unit of the return value must be the one of capacity so we can
4869 * compare the utilization with the capacity of the CPU that is available for
4870 * CFS task (ie cpu_capacity).
4872 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
4873 * recent utilization of currently non-runnable tasks on a CPU. It represents
4874 * the amount of utilization of a CPU in the range [0..capacity_orig] where
4875 * capacity_orig is the cpu_capacity available at the highest frequency
4876 * (arch_scale_freq_capacity()).
4877 * The utilization of a CPU converges towards a sum equal to or less than the
4878 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
4879 * the running time on this CPU scaled by capacity_curr.
4881 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
4882 * higher than capacity_orig because of unfortunate rounding in
4883 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
4884 * the average stabilizes with the new running time. We need to check that the
4885 * utilization stays within the range of [0..capacity_orig] and cap it if
4886 * necessary. Without utilization capping, a group could be seen as overloaded
4887 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
4888 * available capacity. We allow utilization to overshoot capacity_curr (but not
4889 * capacity_orig) as it useful for predicting the capacity required after task
4890 * migrations (scheduler-driven DVFS).
4892 static int cpu_util(int cpu)
4894 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
4895 unsigned long capacity = capacity_orig_of(cpu);
4897 return (util >= capacity) ? capacity : util;
4901 * select_task_rq_fair: Select target runqueue for the waking task in domains
4902 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4903 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4905 * Balances load by selecting the idlest cpu in the idlest group, or under
4906 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4908 * Returns the target cpu number.
4910 * preempt must be disabled.
4913 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4915 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4916 int cpu = smp_processor_id();
4917 int new_cpu = prev_cpu;
4918 int want_affine = 0;
4919 int sync = wake_flags & WF_SYNC;
4921 if (sd_flag & SD_BALANCE_WAKE)
4922 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4925 for_each_domain(cpu, tmp) {
4926 if (!(tmp->flags & SD_LOAD_BALANCE))
4930 * If both cpu and prev_cpu are part of this domain,
4931 * cpu is a valid SD_WAKE_AFFINE target.
4933 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4934 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4939 if (tmp->flags & sd_flag)
4941 else if (!want_affine)
4946 sd = NULL; /* Prefer wake_affine over balance flags */
4947 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4952 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
4953 new_cpu = select_idle_sibling(p, new_cpu);
4956 struct sched_group *group;
4959 if (!(sd->flags & sd_flag)) {
4964 group = find_idlest_group(sd, p, cpu, sd_flag);
4970 new_cpu = find_idlest_cpu(group, p, cpu);
4971 if (new_cpu == -1 || new_cpu == cpu) {
4972 /* Now try balancing at a lower domain level of cpu */
4977 /* Now try balancing at a lower domain level of new_cpu */
4979 weight = sd->span_weight;
4981 for_each_domain(cpu, tmp) {
4982 if (weight <= tmp->span_weight)
4984 if (tmp->flags & sd_flag)
4987 /* while loop will break here if sd == NULL */
4995 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4996 * cfs_rq_of(p) references at time of call are still valid and identify the
4997 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4998 * other assumptions, including the state of rq->lock, should be made.
5000 static void migrate_task_rq_fair(struct task_struct *p, int next_cpu)
5003 * We are supposed to update the task to "current" time, then its up to date
5004 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5005 * what current time is, so simply throw away the out-of-date time. This
5006 * will result in the wakee task is less decayed, but giving the wakee more
5007 * load sounds not bad.
5009 remove_entity_load_avg(&p->se);
5011 /* Tell new CPU we are migrated */
5012 p->se.avg.last_update_time = 0;
5014 /* We have migrated, no longer consider this task hot */
5015 p->se.exec_start = 0;
5018 static void task_dead_fair(struct task_struct *p)
5020 remove_entity_load_avg(&p->se);
5022 #endif /* CONFIG_SMP */
5024 static unsigned long
5025 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5027 unsigned long gran = sysctl_sched_wakeup_granularity;
5030 * Since its curr running now, convert the gran from real-time
5031 * to virtual-time in his units.
5033 * By using 'se' instead of 'curr' we penalize light tasks, so
5034 * they get preempted easier. That is, if 'se' < 'curr' then
5035 * the resulting gran will be larger, therefore penalizing the
5036 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5037 * be smaller, again penalizing the lighter task.
5039 * This is especially important for buddies when the leftmost
5040 * task is higher priority than the buddy.
5042 return calc_delta_fair(gran, se);
5046 * Should 'se' preempt 'curr'.
5060 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5062 s64 gran, vdiff = curr->vruntime - se->vruntime;
5067 gran = wakeup_gran(curr, se);
5074 static void set_last_buddy(struct sched_entity *se)
5076 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5079 for_each_sched_entity(se)
5080 cfs_rq_of(se)->last = se;
5083 static void set_next_buddy(struct sched_entity *se)
5085 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5088 for_each_sched_entity(se)
5089 cfs_rq_of(se)->next = se;
5092 static void set_skip_buddy(struct sched_entity *se)
5094 for_each_sched_entity(se)
5095 cfs_rq_of(se)->skip = se;
5099 * Preempt the current task with a newly woken task if needed:
5101 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5103 struct task_struct *curr = rq->curr;
5104 struct sched_entity *se = &curr->se, *pse = &p->se;
5105 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5106 int scale = cfs_rq->nr_running >= sched_nr_latency;
5107 int next_buddy_marked = 0;
5109 if (unlikely(se == pse))
5113 * This is possible from callers such as attach_tasks(), in which we
5114 * unconditionally check_prempt_curr() after an enqueue (which may have
5115 * lead to a throttle). This both saves work and prevents false
5116 * next-buddy nomination below.
5118 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5121 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5122 set_next_buddy(pse);
5123 next_buddy_marked = 1;
5127 * We can come here with TIF_NEED_RESCHED already set from new task
5130 * Note: this also catches the edge-case of curr being in a throttled
5131 * group (e.g. via set_curr_task), since update_curr() (in the
5132 * enqueue of curr) will have resulted in resched being set. This
5133 * prevents us from potentially nominating it as a false LAST_BUDDY
5136 if (test_tsk_need_resched(curr))
5139 /* Idle tasks are by definition preempted by non-idle tasks. */
5140 if (unlikely(curr->policy == SCHED_IDLE) &&
5141 likely(p->policy != SCHED_IDLE))
5145 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5146 * is driven by the tick):
5148 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5151 find_matching_se(&se, &pse);
5152 update_curr(cfs_rq_of(se));
5154 if (wakeup_preempt_entity(se, pse) == 1) {
5156 * Bias pick_next to pick the sched entity that is
5157 * triggering this preemption.
5159 if (!next_buddy_marked)
5160 set_next_buddy(pse);
5169 * Only set the backward buddy when the current task is still
5170 * on the rq. This can happen when a wakeup gets interleaved
5171 * with schedule on the ->pre_schedule() or idle_balance()
5172 * point, either of which can * drop the rq lock.
5174 * Also, during early boot the idle thread is in the fair class,
5175 * for obvious reasons its a bad idea to schedule back to it.
5177 if (unlikely(!se->on_rq || curr == rq->idle))
5180 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5184 static struct task_struct *
5185 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5187 struct cfs_rq *cfs_rq = &rq->cfs;
5188 struct sched_entity *se;
5189 struct task_struct *p;
5193 #ifdef CONFIG_FAIR_GROUP_SCHED
5194 if (!cfs_rq->nr_running)
5197 if (prev->sched_class != &fair_sched_class)
5201 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5202 * likely that a next task is from the same cgroup as the current.
5204 * Therefore attempt to avoid putting and setting the entire cgroup
5205 * hierarchy, only change the part that actually changes.
5209 struct sched_entity *curr = cfs_rq->curr;
5212 * Since we got here without doing put_prev_entity() we also
5213 * have to consider cfs_rq->curr. If it is still a runnable
5214 * entity, update_curr() will update its vruntime, otherwise
5215 * forget we've ever seen it.
5219 update_curr(cfs_rq);
5224 * This call to check_cfs_rq_runtime() will do the
5225 * throttle and dequeue its entity in the parent(s).
5226 * Therefore the 'simple' nr_running test will indeed
5229 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5233 se = pick_next_entity(cfs_rq, curr);
5234 cfs_rq = group_cfs_rq(se);
5240 * Since we haven't yet done put_prev_entity and if the selected task
5241 * is a different task than we started out with, try and touch the
5242 * least amount of cfs_rqs.
5245 struct sched_entity *pse = &prev->se;
5247 while (!(cfs_rq = is_same_group(se, pse))) {
5248 int se_depth = se->depth;
5249 int pse_depth = pse->depth;
5251 if (se_depth <= pse_depth) {
5252 put_prev_entity(cfs_rq_of(pse), pse);
5253 pse = parent_entity(pse);
5255 if (se_depth >= pse_depth) {
5256 set_next_entity(cfs_rq_of(se), se);
5257 se = parent_entity(se);
5261 put_prev_entity(cfs_rq, pse);
5262 set_next_entity(cfs_rq, se);
5265 if (hrtick_enabled(rq))
5266 hrtick_start_fair(rq, p);
5273 if (!cfs_rq->nr_running)
5276 put_prev_task(rq, prev);
5279 se = pick_next_entity(cfs_rq, NULL);
5280 set_next_entity(cfs_rq, se);
5281 cfs_rq = group_cfs_rq(se);
5286 if (hrtick_enabled(rq))
5287 hrtick_start_fair(rq, p);
5293 * This is OK, because current is on_cpu, which avoids it being picked
5294 * for load-balance and preemption/IRQs are still disabled avoiding
5295 * further scheduler activity on it and we're being very careful to
5296 * re-start the picking loop.
5298 lockdep_unpin_lock(&rq->lock);
5299 new_tasks = idle_balance(rq);
5300 lockdep_pin_lock(&rq->lock);
5302 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5303 * possible for any higher priority task to appear. In that case we
5304 * must re-start the pick_next_entity() loop.
5316 * Account for a descheduled task:
5318 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5320 struct sched_entity *se = &prev->se;
5321 struct cfs_rq *cfs_rq;
5323 for_each_sched_entity(se) {
5324 cfs_rq = cfs_rq_of(se);
5325 put_prev_entity(cfs_rq, se);
5330 * sched_yield() is very simple
5332 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5334 static void yield_task_fair(struct rq *rq)
5336 struct task_struct *curr = rq->curr;
5337 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5338 struct sched_entity *se = &curr->se;
5341 * Are we the only task in the tree?
5343 if (unlikely(rq->nr_running == 1))
5346 clear_buddies(cfs_rq, se);
5348 if (curr->policy != SCHED_BATCH) {
5349 update_rq_clock(rq);
5351 * Update run-time statistics of the 'current'.
5353 update_curr(cfs_rq);
5355 * Tell update_rq_clock() that we've just updated,
5356 * so we don't do microscopic update in schedule()
5357 * and double the fastpath cost.
5359 rq_clock_skip_update(rq, true);
5365 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5367 struct sched_entity *se = &p->se;
5369 /* throttled hierarchies are not runnable */
5370 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5373 /* Tell the scheduler that we'd really like pse to run next. */
5376 yield_task_fair(rq);
5382 /**************************************************
5383 * Fair scheduling class load-balancing methods.
5387 * The purpose of load-balancing is to achieve the same basic fairness the
5388 * per-cpu scheduler provides, namely provide a proportional amount of compute
5389 * time to each task. This is expressed in the following equation:
5391 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5393 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5394 * W_i,0 is defined as:
5396 * W_i,0 = \Sum_j w_i,j (2)
5398 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5399 * is derived from the nice value as per prio_to_weight[].
5401 * The weight average is an exponential decay average of the instantaneous
5404 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5406 * C_i is the compute capacity of cpu i, typically it is the
5407 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5408 * can also include other factors [XXX].
5410 * To achieve this balance we define a measure of imbalance which follows
5411 * directly from (1):
5413 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5415 * We them move tasks around to minimize the imbalance. In the continuous
5416 * function space it is obvious this converges, in the discrete case we get
5417 * a few fun cases generally called infeasible weight scenarios.
5420 * - infeasible weights;
5421 * - local vs global optima in the discrete case. ]
5426 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5427 * for all i,j solution, we create a tree of cpus that follows the hardware
5428 * topology where each level pairs two lower groups (or better). This results
5429 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5430 * tree to only the first of the previous level and we decrease the frequency
5431 * of load-balance at each level inv. proportional to the number of cpus in
5437 * \Sum { --- * --- * 2^i } = O(n) (5)
5439 * `- size of each group
5440 * | | `- number of cpus doing load-balance
5442 * `- sum over all levels
5444 * Coupled with a limit on how many tasks we can migrate every balance pass,
5445 * this makes (5) the runtime complexity of the balancer.
5447 * An important property here is that each CPU is still (indirectly) connected
5448 * to every other cpu in at most O(log n) steps:
5450 * The adjacency matrix of the resulting graph is given by:
5453 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5456 * And you'll find that:
5458 * A^(log_2 n)_i,j != 0 for all i,j (7)
5460 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5461 * The task movement gives a factor of O(m), giving a convergence complexity
5464 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5469 * In order to avoid CPUs going idle while there's still work to do, new idle
5470 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5471 * tree itself instead of relying on other CPUs to bring it work.
5473 * This adds some complexity to both (5) and (8) but it reduces the total idle
5481 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5484 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5489 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5491 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5493 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5496 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5497 * rewrite all of this once again.]
5500 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5502 enum fbq_type { regular, remote, all };
5504 #define LBF_ALL_PINNED 0x01
5505 #define LBF_NEED_BREAK 0x02
5506 #define LBF_DST_PINNED 0x04
5507 #define LBF_SOME_PINNED 0x08
5510 struct sched_domain *sd;
5518 struct cpumask *dst_grpmask;
5520 enum cpu_idle_type idle;
5522 /* The set of CPUs under consideration for load-balancing */
5523 struct cpumask *cpus;
5528 unsigned int loop_break;
5529 unsigned int loop_max;
5531 enum fbq_type fbq_type;
5532 struct list_head tasks;
5536 * Is this task likely cache-hot:
5538 static int task_hot(struct task_struct *p, struct lb_env *env)
5542 lockdep_assert_held(&env->src_rq->lock);
5544 if (p->sched_class != &fair_sched_class)
5547 if (unlikely(p->policy == SCHED_IDLE))
5551 * Buddy candidates are cache hot:
5553 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5554 (&p->se == cfs_rq_of(&p->se)->next ||
5555 &p->se == cfs_rq_of(&p->se)->last))
5558 if (sysctl_sched_migration_cost == -1)
5560 if (sysctl_sched_migration_cost == 0)
5563 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5565 return delta < (s64)sysctl_sched_migration_cost;
5568 #ifdef CONFIG_NUMA_BALANCING
5570 * Returns 1, if task migration degrades locality
5571 * Returns 0, if task migration improves locality i.e migration preferred.
5572 * Returns -1, if task migration is not affected by locality.
5574 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5576 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5577 unsigned long src_faults, dst_faults;
5578 int src_nid, dst_nid;
5580 if (!static_branch_likely(&sched_numa_balancing))
5583 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5586 src_nid = cpu_to_node(env->src_cpu);
5587 dst_nid = cpu_to_node(env->dst_cpu);
5589 if (src_nid == dst_nid)
5592 /* Migrating away from the preferred node is always bad. */
5593 if (src_nid == p->numa_preferred_nid) {
5594 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5600 /* Encourage migration to the preferred node. */
5601 if (dst_nid == p->numa_preferred_nid)
5605 src_faults = group_faults(p, src_nid);
5606 dst_faults = group_faults(p, dst_nid);
5608 src_faults = task_faults(p, src_nid);
5609 dst_faults = task_faults(p, dst_nid);
5612 return dst_faults < src_faults;
5616 static inline int migrate_degrades_locality(struct task_struct *p,
5624 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5627 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5631 lockdep_assert_held(&env->src_rq->lock);
5634 * We do not migrate tasks that are:
5635 * 1) throttled_lb_pair, or
5636 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5637 * 3) running (obviously), or
5638 * 4) are cache-hot on their current CPU.
5640 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5643 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5646 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5648 env->flags |= LBF_SOME_PINNED;
5651 * Remember if this task can be migrated to any other cpu in
5652 * our sched_group. We may want to revisit it if we couldn't
5653 * meet load balance goals by pulling other tasks on src_cpu.
5655 * Also avoid computing new_dst_cpu if we have already computed
5656 * one in current iteration.
5658 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5661 /* Prevent to re-select dst_cpu via env's cpus */
5662 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5663 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5664 env->flags |= LBF_DST_PINNED;
5665 env->new_dst_cpu = cpu;
5673 /* Record that we found atleast one task that could run on dst_cpu */
5674 env->flags &= ~LBF_ALL_PINNED;
5676 if (task_running(env->src_rq, p)) {
5677 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5682 * Aggressive migration if:
5683 * 1) destination numa is preferred
5684 * 2) task is cache cold, or
5685 * 3) too many balance attempts have failed.
5687 tsk_cache_hot = migrate_degrades_locality(p, env);
5688 if (tsk_cache_hot == -1)
5689 tsk_cache_hot = task_hot(p, env);
5691 if (tsk_cache_hot <= 0 ||
5692 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5693 if (tsk_cache_hot == 1) {
5694 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5695 schedstat_inc(p, se.statistics.nr_forced_migrations);
5700 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5705 * detach_task() -- detach the task for the migration specified in env
5707 static void detach_task(struct task_struct *p, struct lb_env *env)
5709 lockdep_assert_held(&env->src_rq->lock);
5711 deactivate_task(env->src_rq, p, 0);
5712 p->on_rq = TASK_ON_RQ_MIGRATING;
5713 set_task_cpu(p, env->dst_cpu);
5717 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5718 * part of active balancing operations within "domain".
5720 * Returns a task if successful and NULL otherwise.
5722 static struct task_struct *detach_one_task(struct lb_env *env)
5724 struct task_struct *p, *n;
5726 lockdep_assert_held(&env->src_rq->lock);
5728 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5729 if (!can_migrate_task(p, env))
5732 detach_task(p, env);
5735 * Right now, this is only the second place where
5736 * lb_gained[env->idle] is updated (other is detach_tasks)
5737 * so we can safely collect stats here rather than
5738 * inside detach_tasks().
5740 schedstat_inc(env->sd, lb_gained[env->idle]);
5746 static const unsigned int sched_nr_migrate_break = 32;
5749 * detach_tasks() -- tries to detach up to imbalance weighted load from
5750 * busiest_rq, as part of a balancing operation within domain "sd".
5752 * Returns number of detached tasks if successful and 0 otherwise.
5754 static int detach_tasks(struct lb_env *env)
5756 struct list_head *tasks = &env->src_rq->cfs_tasks;
5757 struct task_struct *p;
5761 lockdep_assert_held(&env->src_rq->lock);
5763 if (env->imbalance <= 0)
5766 while (!list_empty(tasks)) {
5768 * We don't want to steal all, otherwise we may be treated likewise,
5769 * which could at worst lead to a livelock crash.
5771 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5774 p = list_first_entry(tasks, struct task_struct, se.group_node);
5777 /* We've more or less seen every task there is, call it quits */
5778 if (env->loop > env->loop_max)
5781 /* take a breather every nr_migrate tasks */
5782 if (env->loop > env->loop_break) {
5783 env->loop_break += sched_nr_migrate_break;
5784 env->flags |= LBF_NEED_BREAK;
5788 if (!can_migrate_task(p, env))
5791 load = task_h_load(p);
5793 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5796 if ((load / 2) > env->imbalance)
5799 detach_task(p, env);
5800 list_add(&p->se.group_node, &env->tasks);
5803 env->imbalance -= load;
5805 #ifdef CONFIG_PREEMPT
5807 * NEWIDLE balancing is a source of latency, so preemptible
5808 * kernels will stop after the first task is detached to minimize
5809 * the critical section.
5811 if (env->idle == CPU_NEWLY_IDLE)
5816 * We only want to steal up to the prescribed amount of
5819 if (env->imbalance <= 0)
5824 list_move_tail(&p->se.group_node, tasks);
5828 * Right now, this is one of only two places we collect this stat
5829 * so we can safely collect detach_one_task() stats here rather
5830 * than inside detach_one_task().
5832 schedstat_add(env->sd, lb_gained[env->idle], detached);
5838 * attach_task() -- attach the task detached by detach_task() to its new rq.
5840 static void attach_task(struct rq *rq, struct task_struct *p)
5842 lockdep_assert_held(&rq->lock);
5844 BUG_ON(task_rq(p) != rq);
5845 p->on_rq = TASK_ON_RQ_QUEUED;
5846 activate_task(rq, p, 0);
5847 check_preempt_curr(rq, p, 0);
5851 * attach_one_task() -- attaches the task returned from detach_one_task() to
5854 static void attach_one_task(struct rq *rq, struct task_struct *p)
5856 raw_spin_lock(&rq->lock);
5858 raw_spin_unlock(&rq->lock);
5862 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5865 static void attach_tasks(struct lb_env *env)
5867 struct list_head *tasks = &env->tasks;
5868 struct task_struct *p;
5870 raw_spin_lock(&env->dst_rq->lock);
5872 while (!list_empty(tasks)) {
5873 p = list_first_entry(tasks, struct task_struct, se.group_node);
5874 list_del_init(&p->se.group_node);
5876 attach_task(env->dst_rq, p);
5879 raw_spin_unlock(&env->dst_rq->lock);
5882 #ifdef CONFIG_FAIR_GROUP_SCHED
5883 static void update_blocked_averages(int cpu)
5885 struct rq *rq = cpu_rq(cpu);
5886 struct cfs_rq *cfs_rq;
5887 unsigned long flags;
5889 raw_spin_lock_irqsave(&rq->lock, flags);
5890 update_rq_clock(rq);
5893 * Iterates the task_group tree in a bottom up fashion, see
5894 * list_add_leaf_cfs_rq() for details.
5896 for_each_leaf_cfs_rq(rq, cfs_rq) {
5897 /* throttled entities do not contribute to load */
5898 if (throttled_hierarchy(cfs_rq))
5901 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
5902 update_tg_load_avg(cfs_rq, 0);
5904 raw_spin_unlock_irqrestore(&rq->lock, flags);
5908 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5909 * This needs to be done in a top-down fashion because the load of a child
5910 * group is a fraction of its parents load.
5912 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5914 struct rq *rq = rq_of(cfs_rq);
5915 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5916 unsigned long now = jiffies;
5919 if (cfs_rq->last_h_load_update == now)
5922 cfs_rq->h_load_next = NULL;
5923 for_each_sched_entity(se) {
5924 cfs_rq = cfs_rq_of(se);
5925 cfs_rq->h_load_next = se;
5926 if (cfs_rq->last_h_load_update == now)
5931 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
5932 cfs_rq->last_h_load_update = now;
5935 while ((se = cfs_rq->h_load_next) != NULL) {
5936 load = cfs_rq->h_load;
5937 load = div64_ul(load * se->avg.load_avg,
5938 cfs_rq_load_avg(cfs_rq) + 1);
5939 cfs_rq = group_cfs_rq(se);
5940 cfs_rq->h_load = load;
5941 cfs_rq->last_h_load_update = now;
5945 static unsigned long task_h_load(struct task_struct *p)
5947 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5949 update_cfs_rq_h_load(cfs_rq);
5950 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
5951 cfs_rq_load_avg(cfs_rq) + 1);
5954 static inline void update_blocked_averages(int cpu)
5956 struct rq *rq = cpu_rq(cpu);
5957 struct cfs_rq *cfs_rq = &rq->cfs;
5958 unsigned long flags;
5960 raw_spin_lock_irqsave(&rq->lock, flags);
5961 update_rq_clock(rq);
5962 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
5963 raw_spin_unlock_irqrestore(&rq->lock, flags);
5966 static unsigned long task_h_load(struct task_struct *p)
5968 return p->se.avg.load_avg;
5972 /********** Helpers for find_busiest_group ************************/
5981 * sg_lb_stats - stats of a sched_group required for load_balancing
5983 struct sg_lb_stats {
5984 unsigned long avg_load; /*Avg load across the CPUs of the group */
5985 unsigned long group_load; /* Total load over the CPUs of the group */
5986 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5987 unsigned long load_per_task;
5988 unsigned long group_capacity;
5989 unsigned long group_util; /* Total utilization of the group */
5990 unsigned int sum_nr_running; /* Nr tasks running in the group */
5991 unsigned int idle_cpus;
5992 unsigned int group_weight;
5993 enum group_type group_type;
5994 int group_no_capacity;
5995 #ifdef CONFIG_NUMA_BALANCING
5996 unsigned int nr_numa_running;
5997 unsigned int nr_preferred_running;
6002 * sd_lb_stats - Structure to store the statistics of a sched_domain
6003 * during load balancing.
6005 struct sd_lb_stats {
6006 struct sched_group *busiest; /* Busiest group in this sd */
6007 struct sched_group *local; /* Local group in this sd */
6008 unsigned long total_load; /* Total load of all groups in sd */
6009 unsigned long total_capacity; /* Total capacity of all groups in sd */
6010 unsigned long avg_load; /* Average load across all groups in sd */
6012 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6013 struct sg_lb_stats local_stat; /* Statistics of the local group */
6016 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6019 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6020 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6021 * We must however clear busiest_stat::avg_load because
6022 * update_sd_pick_busiest() reads this before assignment.
6024 *sds = (struct sd_lb_stats){
6028 .total_capacity = 0UL,
6031 .sum_nr_running = 0,
6032 .group_type = group_other,
6038 * get_sd_load_idx - Obtain the load index for a given sched domain.
6039 * @sd: The sched_domain whose load_idx is to be obtained.
6040 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6042 * Return: The load index.
6044 static inline int get_sd_load_idx(struct sched_domain *sd,
6045 enum cpu_idle_type idle)
6051 load_idx = sd->busy_idx;
6054 case CPU_NEWLY_IDLE:
6055 load_idx = sd->newidle_idx;
6058 load_idx = sd->idle_idx;
6065 static unsigned long scale_rt_capacity(int cpu)
6067 struct rq *rq = cpu_rq(cpu);
6068 u64 total, used, age_stamp, avg;
6072 * Since we're reading these variables without serialization make sure
6073 * we read them once before doing sanity checks on them.
6075 age_stamp = READ_ONCE(rq->age_stamp);
6076 avg = READ_ONCE(rq->rt_avg);
6077 delta = __rq_clock_broken(rq) - age_stamp;
6079 if (unlikely(delta < 0))
6082 total = sched_avg_period() + delta;
6084 used = div_u64(avg, total);
6086 if (likely(used < SCHED_CAPACITY_SCALE))
6087 return SCHED_CAPACITY_SCALE - used;
6092 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6094 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6095 struct sched_group *sdg = sd->groups;
6097 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6099 capacity *= scale_rt_capacity(cpu);
6100 capacity >>= SCHED_CAPACITY_SHIFT;
6105 cpu_rq(cpu)->cpu_capacity = capacity;
6106 sdg->sgc->capacity = capacity;
6109 void update_group_capacity(struct sched_domain *sd, int cpu)
6111 struct sched_domain *child = sd->child;
6112 struct sched_group *group, *sdg = sd->groups;
6113 unsigned long capacity;
6114 unsigned long interval;
6116 interval = msecs_to_jiffies(sd->balance_interval);
6117 interval = clamp(interval, 1UL, max_load_balance_interval);
6118 sdg->sgc->next_update = jiffies + interval;
6121 update_cpu_capacity(sd, cpu);
6127 if (child->flags & SD_OVERLAP) {
6129 * SD_OVERLAP domains cannot assume that child groups
6130 * span the current group.
6133 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6134 struct sched_group_capacity *sgc;
6135 struct rq *rq = cpu_rq(cpu);
6138 * build_sched_domains() -> init_sched_groups_capacity()
6139 * gets here before we've attached the domains to the
6142 * Use capacity_of(), which is set irrespective of domains
6143 * in update_cpu_capacity().
6145 * This avoids capacity from being 0 and
6146 * causing divide-by-zero issues on boot.
6148 if (unlikely(!rq->sd)) {
6149 capacity += capacity_of(cpu);
6153 sgc = rq->sd->groups->sgc;
6154 capacity += sgc->capacity;
6158 * !SD_OVERLAP domains can assume that child groups
6159 * span the current group.
6162 group = child->groups;
6164 capacity += group->sgc->capacity;
6165 group = group->next;
6166 } while (group != child->groups);
6169 sdg->sgc->capacity = capacity;
6173 * Check whether the capacity of the rq has been noticeably reduced by side
6174 * activity. The imbalance_pct is used for the threshold.
6175 * Return true is the capacity is reduced
6178 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6180 return ((rq->cpu_capacity * sd->imbalance_pct) <
6181 (rq->cpu_capacity_orig * 100));
6185 * Group imbalance indicates (and tries to solve) the problem where balancing
6186 * groups is inadequate due to tsk_cpus_allowed() constraints.
6188 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6189 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6192 * { 0 1 2 3 } { 4 5 6 7 }
6195 * If we were to balance group-wise we'd place two tasks in the first group and
6196 * two tasks in the second group. Clearly this is undesired as it will overload
6197 * cpu 3 and leave one of the cpus in the second group unused.
6199 * The current solution to this issue is detecting the skew in the first group
6200 * by noticing the lower domain failed to reach balance and had difficulty
6201 * moving tasks due to affinity constraints.
6203 * When this is so detected; this group becomes a candidate for busiest; see
6204 * update_sd_pick_busiest(). And calculate_imbalance() and
6205 * find_busiest_group() avoid some of the usual balance conditions to allow it
6206 * to create an effective group imbalance.
6208 * This is a somewhat tricky proposition since the next run might not find the
6209 * group imbalance and decide the groups need to be balanced again. A most
6210 * subtle and fragile situation.
6213 static inline int sg_imbalanced(struct sched_group *group)
6215 return group->sgc->imbalance;
6219 * group_has_capacity returns true if the group has spare capacity that could
6220 * be used by some tasks.
6221 * We consider that a group has spare capacity if the * number of task is
6222 * smaller than the number of CPUs or if the utilization is lower than the
6223 * available capacity for CFS tasks.
6224 * For the latter, we use a threshold to stabilize the state, to take into
6225 * account the variance of the tasks' load and to return true if the available
6226 * capacity in meaningful for the load balancer.
6227 * As an example, an available capacity of 1% can appear but it doesn't make
6228 * any benefit for the load balance.
6231 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6233 if (sgs->sum_nr_running < sgs->group_weight)
6236 if ((sgs->group_capacity * 100) >
6237 (sgs->group_util * env->sd->imbalance_pct))
6244 * group_is_overloaded returns true if the group has more tasks than it can
6246 * group_is_overloaded is not equals to !group_has_capacity because a group
6247 * with the exact right number of tasks, has no more spare capacity but is not
6248 * overloaded so both group_has_capacity and group_is_overloaded return
6252 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6254 if (sgs->sum_nr_running <= sgs->group_weight)
6257 if ((sgs->group_capacity * 100) <
6258 (sgs->group_util * env->sd->imbalance_pct))
6264 static enum group_type group_classify(struct lb_env *env,
6265 struct sched_group *group,
6266 struct sg_lb_stats *sgs)
6268 if (sgs->group_no_capacity)
6269 return group_overloaded;
6271 if (sg_imbalanced(group))
6272 return group_imbalanced;
6278 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6279 * @env: The load balancing environment.
6280 * @group: sched_group whose statistics are to be updated.
6281 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6282 * @local_group: Does group contain this_cpu.
6283 * @sgs: variable to hold the statistics for this group.
6284 * @overload: Indicate more than one runnable task for any CPU.
6286 static inline void update_sg_lb_stats(struct lb_env *env,
6287 struct sched_group *group, int load_idx,
6288 int local_group, struct sg_lb_stats *sgs,
6294 memset(sgs, 0, sizeof(*sgs));
6296 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6297 struct rq *rq = cpu_rq(i);
6299 /* Bias balancing toward cpus of our domain */
6301 load = target_load(i, load_idx);
6303 load = source_load(i, load_idx);
6305 sgs->group_load += load;
6306 sgs->group_util += cpu_util(i);
6307 sgs->sum_nr_running += rq->cfs.h_nr_running;
6309 if (rq->nr_running > 1)
6312 #ifdef CONFIG_NUMA_BALANCING
6313 sgs->nr_numa_running += rq->nr_numa_running;
6314 sgs->nr_preferred_running += rq->nr_preferred_running;
6316 sgs->sum_weighted_load += weighted_cpuload(i);
6321 /* Adjust by relative CPU capacity of the group */
6322 sgs->group_capacity = group->sgc->capacity;
6323 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6325 if (sgs->sum_nr_running)
6326 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6328 sgs->group_weight = group->group_weight;
6330 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6331 sgs->group_type = group_classify(env, group, sgs);
6335 * update_sd_pick_busiest - return 1 on busiest group
6336 * @env: The load balancing environment.
6337 * @sds: sched_domain statistics
6338 * @sg: sched_group candidate to be checked for being the busiest
6339 * @sgs: sched_group statistics
6341 * Determine if @sg is a busier group than the previously selected
6344 * Return: %true if @sg is a busier group than the previously selected
6345 * busiest group. %false otherwise.
6347 static bool update_sd_pick_busiest(struct lb_env *env,
6348 struct sd_lb_stats *sds,
6349 struct sched_group *sg,
6350 struct sg_lb_stats *sgs)
6352 struct sg_lb_stats *busiest = &sds->busiest_stat;
6354 if (sgs->group_type > busiest->group_type)
6357 if (sgs->group_type < busiest->group_type)
6360 if (sgs->avg_load <= busiest->avg_load)
6363 /* This is the busiest node in its class. */
6364 if (!(env->sd->flags & SD_ASYM_PACKING))
6368 * ASYM_PACKING needs to move all the work to the lowest
6369 * numbered CPUs in the group, therefore mark all groups
6370 * higher than ourself as busy.
6372 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6376 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6383 #ifdef CONFIG_NUMA_BALANCING
6384 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6386 if (sgs->sum_nr_running > sgs->nr_numa_running)
6388 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6393 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6395 if (rq->nr_running > rq->nr_numa_running)
6397 if (rq->nr_running > rq->nr_preferred_running)
6402 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6407 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6411 #endif /* CONFIG_NUMA_BALANCING */
6414 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6415 * @env: The load balancing environment.
6416 * @sds: variable to hold the statistics for this sched_domain.
6418 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6420 struct sched_domain *child = env->sd->child;
6421 struct sched_group *sg = env->sd->groups;
6422 struct sg_lb_stats tmp_sgs;
6423 int load_idx, prefer_sibling = 0;
6424 bool overload = false;
6426 if (child && child->flags & SD_PREFER_SIBLING)
6429 load_idx = get_sd_load_idx(env->sd, env->idle);
6432 struct sg_lb_stats *sgs = &tmp_sgs;
6435 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6438 sgs = &sds->local_stat;
6440 if (env->idle != CPU_NEWLY_IDLE ||
6441 time_after_eq(jiffies, sg->sgc->next_update))
6442 update_group_capacity(env->sd, env->dst_cpu);
6445 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6452 * In case the child domain prefers tasks go to siblings
6453 * first, lower the sg capacity so that we'll try
6454 * and move all the excess tasks away. We lower the capacity
6455 * of a group only if the local group has the capacity to fit
6456 * these excess tasks. The extra check prevents the case where
6457 * you always pull from the heaviest group when it is already
6458 * under-utilized (possible with a large weight task outweighs
6459 * the tasks on the system).
6461 if (prefer_sibling && sds->local &&
6462 group_has_capacity(env, &sds->local_stat) &&
6463 (sgs->sum_nr_running > 1)) {
6464 sgs->group_no_capacity = 1;
6465 sgs->group_type = group_overloaded;
6468 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6470 sds->busiest_stat = *sgs;
6474 /* Now, start updating sd_lb_stats */
6475 sds->total_load += sgs->group_load;
6476 sds->total_capacity += sgs->group_capacity;
6479 } while (sg != env->sd->groups);
6481 if (env->sd->flags & SD_NUMA)
6482 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6484 if (!env->sd->parent) {
6485 /* update overload indicator if we are at root domain */
6486 if (env->dst_rq->rd->overload != overload)
6487 env->dst_rq->rd->overload = overload;
6493 * check_asym_packing - Check to see if the group is packed into the
6496 * This is primarily intended to used at the sibling level. Some
6497 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6498 * case of POWER7, it can move to lower SMT modes only when higher
6499 * threads are idle. When in lower SMT modes, the threads will
6500 * perform better since they share less core resources. Hence when we
6501 * have idle threads, we want them to be the higher ones.
6503 * This packing function is run on idle threads. It checks to see if
6504 * the busiest CPU in this domain (core in the P7 case) has a higher
6505 * CPU number than the packing function is being run on. Here we are
6506 * assuming lower CPU number will be equivalent to lower a SMT thread
6509 * Return: 1 when packing is required and a task should be moved to
6510 * this CPU. The amount of the imbalance is returned in *imbalance.
6512 * @env: The load balancing environment.
6513 * @sds: Statistics of the sched_domain which is to be packed
6515 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6519 if (!(env->sd->flags & SD_ASYM_PACKING))
6525 busiest_cpu = group_first_cpu(sds->busiest);
6526 if (env->dst_cpu > busiest_cpu)
6529 env->imbalance = DIV_ROUND_CLOSEST(
6530 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6531 SCHED_CAPACITY_SCALE);
6537 * fix_small_imbalance - Calculate the minor imbalance that exists
6538 * amongst the groups of a sched_domain, during
6540 * @env: The load balancing environment.
6541 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6544 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6546 unsigned long tmp, capa_now = 0, capa_move = 0;
6547 unsigned int imbn = 2;
6548 unsigned long scaled_busy_load_per_task;
6549 struct sg_lb_stats *local, *busiest;
6551 local = &sds->local_stat;
6552 busiest = &sds->busiest_stat;
6554 if (!local->sum_nr_running)
6555 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6556 else if (busiest->load_per_task > local->load_per_task)
6559 scaled_busy_load_per_task =
6560 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6561 busiest->group_capacity;
6563 if (busiest->avg_load + scaled_busy_load_per_task >=
6564 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6565 env->imbalance = busiest->load_per_task;
6570 * OK, we don't have enough imbalance to justify moving tasks,
6571 * however we may be able to increase total CPU capacity used by
6575 capa_now += busiest->group_capacity *
6576 min(busiest->load_per_task, busiest->avg_load);
6577 capa_now += local->group_capacity *
6578 min(local->load_per_task, local->avg_load);
6579 capa_now /= SCHED_CAPACITY_SCALE;
6581 /* Amount of load we'd subtract */
6582 if (busiest->avg_load > scaled_busy_load_per_task) {
6583 capa_move += busiest->group_capacity *
6584 min(busiest->load_per_task,
6585 busiest->avg_load - scaled_busy_load_per_task);
6588 /* Amount of load we'd add */
6589 if (busiest->avg_load * busiest->group_capacity <
6590 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6591 tmp = (busiest->avg_load * busiest->group_capacity) /
6592 local->group_capacity;
6594 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6595 local->group_capacity;
6597 capa_move += local->group_capacity *
6598 min(local->load_per_task, local->avg_load + tmp);
6599 capa_move /= SCHED_CAPACITY_SCALE;
6601 /* Move if we gain throughput */
6602 if (capa_move > capa_now)
6603 env->imbalance = busiest->load_per_task;
6607 * calculate_imbalance - Calculate the amount of imbalance present within the
6608 * groups of a given sched_domain during load balance.
6609 * @env: load balance environment
6610 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6612 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6614 unsigned long max_pull, load_above_capacity = ~0UL;
6615 struct sg_lb_stats *local, *busiest;
6617 local = &sds->local_stat;
6618 busiest = &sds->busiest_stat;
6620 if (busiest->group_type == group_imbalanced) {
6622 * In the group_imb case we cannot rely on group-wide averages
6623 * to ensure cpu-load equilibrium, look at wider averages. XXX
6625 busiest->load_per_task =
6626 min(busiest->load_per_task, sds->avg_load);
6630 * In the presence of smp nice balancing, certain scenarios can have
6631 * max load less than avg load(as we skip the groups at or below
6632 * its cpu_capacity, while calculating max_load..)
6634 if (busiest->avg_load <= sds->avg_load ||
6635 local->avg_load >= sds->avg_load) {
6637 return fix_small_imbalance(env, sds);
6641 * If there aren't any idle cpus, avoid creating some.
6643 if (busiest->group_type == group_overloaded &&
6644 local->group_type == group_overloaded) {
6645 load_above_capacity = busiest->sum_nr_running *
6647 if (load_above_capacity > busiest->group_capacity)
6648 load_above_capacity -= busiest->group_capacity;
6650 load_above_capacity = ~0UL;
6654 * We're trying to get all the cpus to the average_load, so we don't
6655 * want to push ourselves above the average load, nor do we wish to
6656 * reduce the max loaded cpu below the average load. At the same time,
6657 * we also don't want to reduce the group load below the group capacity
6658 * (so that we can implement power-savings policies etc). Thus we look
6659 * for the minimum possible imbalance.
6661 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6663 /* How much load to actually move to equalise the imbalance */
6664 env->imbalance = min(
6665 max_pull * busiest->group_capacity,
6666 (sds->avg_load - local->avg_load) * local->group_capacity
6667 ) / SCHED_CAPACITY_SCALE;
6670 * if *imbalance is less than the average load per runnable task
6671 * there is no guarantee that any tasks will be moved so we'll have
6672 * a think about bumping its value to force at least one task to be
6675 if (env->imbalance < busiest->load_per_task)
6676 return fix_small_imbalance(env, sds);
6679 /******* find_busiest_group() helpers end here *********************/
6682 * find_busiest_group - Returns the busiest group within the sched_domain
6683 * if there is an imbalance. If there isn't an imbalance, and
6684 * the user has opted for power-savings, it returns a group whose
6685 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6686 * such a group exists.
6688 * Also calculates the amount of weighted load which should be moved
6689 * to restore balance.
6691 * @env: The load balancing environment.
6693 * Return: - The busiest group if imbalance exists.
6694 * - If no imbalance and user has opted for power-savings balance,
6695 * return the least loaded group whose CPUs can be
6696 * put to idle by rebalancing its tasks onto our group.
6698 static struct sched_group *find_busiest_group(struct lb_env *env)
6700 struct sg_lb_stats *local, *busiest;
6701 struct sd_lb_stats sds;
6703 init_sd_lb_stats(&sds);
6706 * Compute the various statistics relavent for load balancing at
6709 update_sd_lb_stats(env, &sds);
6710 local = &sds.local_stat;
6711 busiest = &sds.busiest_stat;
6713 /* ASYM feature bypasses nice load balance check */
6714 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6715 check_asym_packing(env, &sds))
6718 /* There is no busy sibling group to pull tasks from */
6719 if (!sds.busiest || busiest->sum_nr_running == 0)
6722 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6723 / sds.total_capacity;
6726 * If the busiest group is imbalanced the below checks don't
6727 * work because they assume all things are equal, which typically
6728 * isn't true due to cpus_allowed constraints and the like.
6730 if (busiest->group_type == group_imbalanced)
6733 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6734 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6735 busiest->group_no_capacity)
6739 * If the local group is busier than the selected busiest group
6740 * don't try and pull any tasks.
6742 if (local->avg_load >= busiest->avg_load)
6746 * Don't pull any tasks if this group is already above the domain
6749 if (local->avg_load >= sds.avg_load)
6752 if (env->idle == CPU_IDLE) {
6754 * This cpu is idle. If the busiest group is not overloaded
6755 * and there is no imbalance between this and busiest group
6756 * wrt idle cpus, it is balanced. The imbalance becomes
6757 * significant if the diff is greater than 1 otherwise we
6758 * might end up to just move the imbalance on another group
6760 if ((busiest->group_type != group_overloaded) &&
6761 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6765 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6766 * imbalance_pct to be conservative.
6768 if (100 * busiest->avg_load <=
6769 env->sd->imbalance_pct * local->avg_load)
6774 /* Looks like there is an imbalance. Compute it */
6775 calculate_imbalance(env, &sds);
6784 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6786 static struct rq *find_busiest_queue(struct lb_env *env,
6787 struct sched_group *group)
6789 struct rq *busiest = NULL, *rq;
6790 unsigned long busiest_load = 0, busiest_capacity = 1;
6793 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6794 unsigned long capacity, wl;
6798 rt = fbq_classify_rq(rq);
6801 * We classify groups/runqueues into three groups:
6802 * - regular: there are !numa tasks
6803 * - remote: there are numa tasks that run on the 'wrong' node
6804 * - all: there is no distinction
6806 * In order to avoid migrating ideally placed numa tasks,
6807 * ignore those when there's better options.
6809 * If we ignore the actual busiest queue to migrate another
6810 * task, the next balance pass can still reduce the busiest
6811 * queue by moving tasks around inside the node.
6813 * If we cannot move enough load due to this classification
6814 * the next pass will adjust the group classification and
6815 * allow migration of more tasks.
6817 * Both cases only affect the total convergence complexity.
6819 if (rt > env->fbq_type)
6822 capacity = capacity_of(i);
6824 wl = weighted_cpuload(i);
6827 * When comparing with imbalance, use weighted_cpuload()
6828 * which is not scaled with the cpu capacity.
6831 if (rq->nr_running == 1 && wl > env->imbalance &&
6832 !check_cpu_capacity(rq, env->sd))
6836 * For the load comparisons with the other cpu's, consider
6837 * the weighted_cpuload() scaled with the cpu capacity, so
6838 * that the load can be moved away from the cpu that is
6839 * potentially running at a lower capacity.
6841 * Thus we're looking for max(wl_i / capacity_i), crosswise
6842 * multiplication to rid ourselves of the division works out
6843 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6844 * our previous maximum.
6846 if (wl * busiest_capacity > busiest_load * capacity) {
6848 busiest_capacity = capacity;
6857 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6858 * so long as it is large enough.
6860 #define MAX_PINNED_INTERVAL 512
6862 /* Working cpumask for load_balance and load_balance_newidle. */
6863 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6865 static int need_active_balance(struct lb_env *env)
6867 struct sched_domain *sd = env->sd;
6869 if (env->idle == CPU_NEWLY_IDLE) {
6872 * ASYM_PACKING needs to force migrate tasks from busy but
6873 * higher numbered CPUs in order to pack all tasks in the
6874 * lowest numbered CPUs.
6876 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6881 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
6882 * It's worth migrating the task if the src_cpu's capacity is reduced
6883 * because of other sched_class or IRQs if more capacity stays
6884 * available on dst_cpu.
6886 if ((env->idle != CPU_NOT_IDLE) &&
6887 (env->src_rq->cfs.h_nr_running == 1)) {
6888 if ((check_cpu_capacity(env->src_rq, sd)) &&
6889 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
6893 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6896 static int active_load_balance_cpu_stop(void *data);
6898 static int should_we_balance(struct lb_env *env)
6900 struct sched_group *sg = env->sd->groups;
6901 struct cpumask *sg_cpus, *sg_mask;
6902 int cpu, balance_cpu = -1;
6905 * In the newly idle case, we will allow all the cpu's
6906 * to do the newly idle load balance.
6908 if (env->idle == CPU_NEWLY_IDLE)
6911 sg_cpus = sched_group_cpus(sg);
6912 sg_mask = sched_group_mask(sg);
6913 /* Try to find first idle cpu */
6914 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6915 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6922 if (balance_cpu == -1)
6923 balance_cpu = group_balance_cpu(sg);
6926 * First idle cpu or the first cpu(busiest) in this sched group
6927 * is eligible for doing load balancing at this and above domains.
6929 return balance_cpu == env->dst_cpu;
6933 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6934 * tasks if there is an imbalance.
6936 static int load_balance(int this_cpu, struct rq *this_rq,
6937 struct sched_domain *sd, enum cpu_idle_type idle,
6938 int *continue_balancing)
6940 int ld_moved, cur_ld_moved, active_balance = 0;
6941 struct sched_domain *sd_parent = sd->parent;
6942 struct sched_group *group;
6944 unsigned long flags;
6945 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
6947 struct lb_env env = {
6949 .dst_cpu = this_cpu,
6951 .dst_grpmask = sched_group_cpus(sd->groups),
6953 .loop_break = sched_nr_migrate_break,
6956 .tasks = LIST_HEAD_INIT(env.tasks),
6960 * For NEWLY_IDLE load_balancing, we don't need to consider
6961 * other cpus in our group
6963 if (idle == CPU_NEWLY_IDLE)
6964 env.dst_grpmask = NULL;
6966 cpumask_copy(cpus, cpu_active_mask);
6968 schedstat_inc(sd, lb_count[idle]);
6971 if (!should_we_balance(&env)) {
6972 *continue_balancing = 0;
6976 group = find_busiest_group(&env);
6978 schedstat_inc(sd, lb_nobusyg[idle]);
6982 busiest = find_busiest_queue(&env, group);
6984 schedstat_inc(sd, lb_nobusyq[idle]);
6988 BUG_ON(busiest == env.dst_rq);
6990 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6992 env.src_cpu = busiest->cpu;
6993 env.src_rq = busiest;
6996 if (busiest->nr_running > 1) {
6998 * Attempt to move tasks. If find_busiest_group has found
6999 * an imbalance but busiest->nr_running <= 1, the group is
7000 * still unbalanced. ld_moved simply stays zero, so it is
7001 * correctly treated as an imbalance.
7003 env.flags |= LBF_ALL_PINNED;
7004 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7007 raw_spin_lock_irqsave(&busiest->lock, flags);
7010 * cur_ld_moved - load moved in current iteration
7011 * ld_moved - cumulative load moved across iterations
7013 cur_ld_moved = detach_tasks(&env);
7016 * We've detached some tasks from busiest_rq. Every
7017 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7018 * unlock busiest->lock, and we are able to be sure
7019 * that nobody can manipulate the tasks in parallel.
7020 * See task_rq_lock() family for the details.
7023 raw_spin_unlock(&busiest->lock);
7027 ld_moved += cur_ld_moved;
7030 local_irq_restore(flags);
7032 if (env.flags & LBF_NEED_BREAK) {
7033 env.flags &= ~LBF_NEED_BREAK;
7038 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7039 * us and move them to an alternate dst_cpu in our sched_group
7040 * where they can run. The upper limit on how many times we
7041 * iterate on same src_cpu is dependent on number of cpus in our
7044 * This changes load balance semantics a bit on who can move
7045 * load to a given_cpu. In addition to the given_cpu itself
7046 * (or a ilb_cpu acting on its behalf where given_cpu is
7047 * nohz-idle), we now have balance_cpu in a position to move
7048 * load to given_cpu. In rare situations, this may cause
7049 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7050 * _independently_ and at _same_ time to move some load to
7051 * given_cpu) causing exceess load to be moved to given_cpu.
7052 * This however should not happen so much in practice and
7053 * moreover subsequent load balance cycles should correct the
7054 * excess load moved.
7056 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7058 /* Prevent to re-select dst_cpu via env's cpus */
7059 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7061 env.dst_rq = cpu_rq(env.new_dst_cpu);
7062 env.dst_cpu = env.new_dst_cpu;
7063 env.flags &= ~LBF_DST_PINNED;
7065 env.loop_break = sched_nr_migrate_break;
7068 * Go back to "more_balance" rather than "redo" since we
7069 * need to continue with same src_cpu.
7075 * We failed to reach balance because of affinity.
7078 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7080 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7081 *group_imbalance = 1;
7084 /* All tasks on this runqueue were pinned by CPU affinity */
7085 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7086 cpumask_clear_cpu(cpu_of(busiest), cpus);
7087 if (!cpumask_empty(cpus)) {
7089 env.loop_break = sched_nr_migrate_break;
7092 goto out_all_pinned;
7097 schedstat_inc(sd, lb_failed[idle]);
7099 * Increment the failure counter only on periodic balance.
7100 * We do not want newidle balance, which can be very
7101 * frequent, pollute the failure counter causing
7102 * excessive cache_hot migrations and active balances.
7104 if (idle != CPU_NEWLY_IDLE)
7105 sd->nr_balance_failed++;
7107 if (need_active_balance(&env)) {
7108 raw_spin_lock_irqsave(&busiest->lock, flags);
7110 /* don't kick the active_load_balance_cpu_stop,
7111 * if the curr task on busiest cpu can't be
7114 if (!cpumask_test_cpu(this_cpu,
7115 tsk_cpus_allowed(busiest->curr))) {
7116 raw_spin_unlock_irqrestore(&busiest->lock,
7118 env.flags |= LBF_ALL_PINNED;
7119 goto out_one_pinned;
7123 * ->active_balance synchronizes accesses to
7124 * ->active_balance_work. Once set, it's cleared
7125 * only after active load balance is finished.
7127 if (!busiest->active_balance) {
7128 busiest->active_balance = 1;
7129 busiest->push_cpu = this_cpu;
7132 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7134 if (active_balance) {
7135 stop_one_cpu_nowait(cpu_of(busiest),
7136 active_load_balance_cpu_stop, busiest,
7137 &busiest->active_balance_work);
7141 * We've kicked active balancing, reset the failure
7144 sd->nr_balance_failed = sd->cache_nice_tries+1;
7147 sd->nr_balance_failed = 0;
7149 if (likely(!active_balance)) {
7150 /* We were unbalanced, so reset the balancing interval */
7151 sd->balance_interval = sd->min_interval;
7154 * If we've begun active balancing, start to back off. This
7155 * case may not be covered by the all_pinned logic if there
7156 * is only 1 task on the busy runqueue (because we don't call
7159 if (sd->balance_interval < sd->max_interval)
7160 sd->balance_interval *= 2;
7167 * We reach balance although we may have faced some affinity
7168 * constraints. Clear the imbalance flag if it was set.
7171 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7173 if (*group_imbalance)
7174 *group_imbalance = 0;
7179 * We reach balance because all tasks are pinned at this level so
7180 * we can't migrate them. Let the imbalance flag set so parent level
7181 * can try to migrate them.
7183 schedstat_inc(sd, lb_balanced[idle]);
7185 sd->nr_balance_failed = 0;
7188 /* tune up the balancing interval */
7189 if (((env.flags & LBF_ALL_PINNED) &&
7190 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7191 (sd->balance_interval < sd->max_interval))
7192 sd->balance_interval *= 2;
7199 static inline unsigned long
7200 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7202 unsigned long interval = sd->balance_interval;
7205 interval *= sd->busy_factor;
7207 /* scale ms to jiffies */
7208 interval = msecs_to_jiffies(interval);
7209 interval = clamp(interval, 1UL, max_load_balance_interval);
7215 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7217 unsigned long interval, next;
7219 interval = get_sd_balance_interval(sd, cpu_busy);
7220 next = sd->last_balance + interval;
7222 if (time_after(*next_balance, next))
7223 *next_balance = next;
7227 * idle_balance is called by schedule() if this_cpu is about to become
7228 * idle. Attempts to pull tasks from other CPUs.
7230 static int idle_balance(struct rq *this_rq)
7232 unsigned long next_balance = jiffies + HZ;
7233 int this_cpu = this_rq->cpu;
7234 struct sched_domain *sd;
7235 int pulled_task = 0;
7238 idle_enter_fair(this_rq);
7241 * We must set idle_stamp _before_ calling idle_balance(), such that we
7242 * measure the duration of idle_balance() as idle time.
7244 this_rq->idle_stamp = rq_clock(this_rq);
7246 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7247 !this_rq->rd->overload) {
7249 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7251 update_next_balance(sd, 0, &next_balance);
7257 raw_spin_unlock(&this_rq->lock);
7259 update_blocked_averages(this_cpu);
7261 for_each_domain(this_cpu, sd) {
7262 int continue_balancing = 1;
7263 u64 t0, domain_cost;
7265 if (!(sd->flags & SD_LOAD_BALANCE))
7268 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7269 update_next_balance(sd, 0, &next_balance);
7273 if (sd->flags & SD_BALANCE_NEWIDLE) {
7274 t0 = sched_clock_cpu(this_cpu);
7276 pulled_task = load_balance(this_cpu, this_rq,
7278 &continue_balancing);
7280 domain_cost = sched_clock_cpu(this_cpu) - t0;
7281 if (domain_cost > sd->max_newidle_lb_cost)
7282 sd->max_newidle_lb_cost = domain_cost;
7284 curr_cost += domain_cost;
7287 update_next_balance(sd, 0, &next_balance);
7290 * Stop searching for tasks to pull if there are
7291 * now runnable tasks on this rq.
7293 if (pulled_task || this_rq->nr_running > 0)
7298 raw_spin_lock(&this_rq->lock);
7300 if (curr_cost > this_rq->max_idle_balance_cost)
7301 this_rq->max_idle_balance_cost = curr_cost;
7304 * While browsing the domains, we released the rq lock, a task could
7305 * have been enqueued in the meantime. Since we're not going idle,
7306 * pretend we pulled a task.
7308 if (this_rq->cfs.h_nr_running && !pulled_task)
7312 /* Move the next balance forward */
7313 if (time_after(this_rq->next_balance, next_balance))
7314 this_rq->next_balance = next_balance;
7316 /* Is there a task of a high priority class? */
7317 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7321 idle_exit_fair(this_rq);
7322 this_rq->idle_stamp = 0;
7329 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7330 * running tasks off the busiest CPU onto idle CPUs. It requires at
7331 * least 1 task to be running on each physical CPU where possible, and
7332 * avoids physical / logical imbalances.
7334 static int active_load_balance_cpu_stop(void *data)
7336 struct rq *busiest_rq = data;
7337 int busiest_cpu = cpu_of(busiest_rq);
7338 int target_cpu = busiest_rq->push_cpu;
7339 struct rq *target_rq = cpu_rq(target_cpu);
7340 struct sched_domain *sd;
7341 struct task_struct *p = NULL;
7343 raw_spin_lock_irq(&busiest_rq->lock);
7345 /* make sure the requested cpu hasn't gone down in the meantime */
7346 if (unlikely(busiest_cpu != smp_processor_id() ||
7347 !busiest_rq->active_balance))
7350 /* Is there any task to move? */
7351 if (busiest_rq->nr_running <= 1)
7355 * This condition is "impossible", if it occurs
7356 * we need to fix it. Originally reported by
7357 * Bjorn Helgaas on a 128-cpu setup.
7359 BUG_ON(busiest_rq == target_rq);
7361 /* Search for an sd spanning us and the target CPU. */
7363 for_each_domain(target_cpu, sd) {
7364 if ((sd->flags & SD_LOAD_BALANCE) &&
7365 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7370 struct lb_env env = {
7372 .dst_cpu = target_cpu,
7373 .dst_rq = target_rq,
7374 .src_cpu = busiest_rq->cpu,
7375 .src_rq = busiest_rq,
7379 schedstat_inc(sd, alb_count);
7381 p = detach_one_task(&env);
7383 schedstat_inc(sd, alb_pushed);
7385 schedstat_inc(sd, alb_failed);
7389 busiest_rq->active_balance = 0;
7390 raw_spin_unlock(&busiest_rq->lock);
7393 attach_one_task(target_rq, p);
7400 static inline int on_null_domain(struct rq *rq)
7402 return unlikely(!rcu_dereference_sched(rq->sd));
7405 #ifdef CONFIG_NO_HZ_COMMON
7407 * idle load balancing details
7408 * - When one of the busy CPUs notice that there may be an idle rebalancing
7409 * needed, they will kick the idle load balancer, which then does idle
7410 * load balancing for all the idle CPUs.
7413 cpumask_var_t idle_cpus_mask;
7415 unsigned long next_balance; /* in jiffy units */
7416 } nohz ____cacheline_aligned;
7418 static inline int find_new_ilb(void)
7420 int ilb = cpumask_first(nohz.idle_cpus_mask);
7422 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7429 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7430 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7431 * CPU (if there is one).
7433 static void nohz_balancer_kick(void)
7437 nohz.next_balance++;
7439 ilb_cpu = find_new_ilb();
7441 if (ilb_cpu >= nr_cpu_ids)
7444 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7447 * Use smp_send_reschedule() instead of resched_cpu().
7448 * This way we generate a sched IPI on the target cpu which
7449 * is idle. And the softirq performing nohz idle load balance
7450 * will be run before returning from the IPI.
7452 smp_send_reschedule(ilb_cpu);
7456 static inline void nohz_balance_exit_idle(int cpu)
7458 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7460 * Completely isolated CPUs don't ever set, so we must test.
7462 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7463 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7464 atomic_dec(&nohz.nr_cpus);
7466 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7470 static inline void set_cpu_sd_state_busy(void)
7472 struct sched_domain *sd;
7473 int cpu = smp_processor_id();
7476 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7478 if (!sd || !sd->nohz_idle)
7482 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7487 void set_cpu_sd_state_idle(void)
7489 struct sched_domain *sd;
7490 int cpu = smp_processor_id();
7493 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7495 if (!sd || sd->nohz_idle)
7499 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7505 * This routine will record that the cpu is going idle with tick stopped.
7506 * This info will be used in performing idle load balancing in the future.
7508 void nohz_balance_enter_idle(int cpu)
7511 * If this cpu is going down, then nothing needs to be done.
7513 if (!cpu_active(cpu))
7516 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7520 * If we're a completely isolated CPU, we don't play.
7522 if (on_null_domain(cpu_rq(cpu)))
7525 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7526 atomic_inc(&nohz.nr_cpus);
7527 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7530 static int sched_ilb_notifier(struct notifier_block *nfb,
7531 unsigned long action, void *hcpu)
7533 switch (action & ~CPU_TASKS_FROZEN) {
7535 nohz_balance_exit_idle(smp_processor_id());
7543 static DEFINE_SPINLOCK(balancing);
7546 * Scale the max load_balance interval with the number of CPUs in the system.
7547 * This trades load-balance latency on larger machines for less cross talk.
7549 void update_max_interval(void)
7551 max_load_balance_interval = HZ*num_online_cpus()/10;
7555 * It checks each scheduling domain to see if it is due to be balanced,
7556 * and initiates a balancing operation if so.
7558 * Balancing parameters are set up in init_sched_domains.
7560 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7562 int continue_balancing = 1;
7564 unsigned long interval;
7565 struct sched_domain *sd;
7566 /* Earliest time when we have to do rebalance again */
7567 unsigned long next_balance = jiffies + 60*HZ;
7568 int update_next_balance = 0;
7569 int need_serialize, need_decay = 0;
7572 update_blocked_averages(cpu);
7575 for_each_domain(cpu, sd) {
7577 * Decay the newidle max times here because this is a regular
7578 * visit to all the domains. Decay ~1% per second.
7580 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7581 sd->max_newidle_lb_cost =
7582 (sd->max_newidle_lb_cost * 253) / 256;
7583 sd->next_decay_max_lb_cost = jiffies + HZ;
7586 max_cost += sd->max_newidle_lb_cost;
7588 if (!(sd->flags & SD_LOAD_BALANCE))
7592 * Stop the load balance at this level. There is another
7593 * CPU in our sched group which is doing load balancing more
7596 if (!continue_balancing) {
7602 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7604 need_serialize = sd->flags & SD_SERIALIZE;
7605 if (need_serialize) {
7606 if (!spin_trylock(&balancing))
7610 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7611 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7613 * The LBF_DST_PINNED logic could have changed
7614 * env->dst_cpu, so we can't know our idle
7615 * state even if we migrated tasks. Update it.
7617 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7619 sd->last_balance = jiffies;
7620 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7623 spin_unlock(&balancing);
7625 if (time_after(next_balance, sd->last_balance + interval)) {
7626 next_balance = sd->last_balance + interval;
7627 update_next_balance = 1;
7632 * Ensure the rq-wide value also decays but keep it at a
7633 * reasonable floor to avoid funnies with rq->avg_idle.
7635 rq->max_idle_balance_cost =
7636 max((u64)sysctl_sched_migration_cost, max_cost);
7641 * next_balance will be updated only when there is a need.
7642 * When the cpu is attached to null domain for ex, it will not be
7645 if (likely(update_next_balance)) {
7646 rq->next_balance = next_balance;
7648 #ifdef CONFIG_NO_HZ_COMMON
7650 * If this CPU has been elected to perform the nohz idle
7651 * balance. Other idle CPUs have already rebalanced with
7652 * nohz_idle_balance() and nohz.next_balance has been
7653 * updated accordingly. This CPU is now running the idle load
7654 * balance for itself and we need to update the
7655 * nohz.next_balance accordingly.
7657 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
7658 nohz.next_balance = rq->next_balance;
7663 #ifdef CONFIG_NO_HZ_COMMON
7665 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7666 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7668 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7670 int this_cpu = this_rq->cpu;
7673 /* Earliest time when we have to do rebalance again */
7674 unsigned long next_balance = jiffies + 60*HZ;
7675 int update_next_balance = 0;
7677 if (idle != CPU_IDLE ||
7678 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7681 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7682 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7686 * If this cpu gets work to do, stop the load balancing
7687 * work being done for other cpus. Next load
7688 * balancing owner will pick it up.
7693 rq = cpu_rq(balance_cpu);
7696 * If time for next balance is due,
7699 if (time_after_eq(jiffies, rq->next_balance)) {
7700 raw_spin_lock_irq(&rq->lock);
7701 update_rq_clock(rq);
7702 update_idle_cpu_load(rq);
7703 raw_spin_unlock_irq(&rq->lock);
7704 rebalance_domains(rq, CPU_IDLE);
7707 if (time_after(next_balance, rq->next_balance)) {
7708 next_balance = rq->next_balance;
7709 update_next_balance = 1;
7714 * next_balance will be updated only when there is a need.
7715 * When the CPU is attached to null domain for ex, it will not be
7718 if (likely(update_next_balance))
7719 nohz.next_balance = next_balance;
7721 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7725 * Current heuristic for kicking the idle load balancer in the presence
7726 * of an idle cpu in the system.
7727 * - This rq has more than one task.
7728 * - This rq has at least one CFS task and the capacity of the CPU is
7729 * significantly reduced because of RT tasks or IRQs.
7730 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7731 * multiple busy cpu.
7732 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7733 * domain span are idle.
7735 static inline bool nohz_kick_needed(struct rq *rq)
7737 unsigned long now = jiffies;
7738 struct sched_domain *sd;
7739 struct sched_group_capacity *sgc;
7740 int nr_busy, cpu = rq->cpu;
7743 if (unlikely(rq->idle_balance))
7747 * We may be recently in ticked or tickless idle mode. At the first
7748 * busy tick after returning from idle, we will update the busy stats.
7750 set_cpu_sd_state_busy();
7751 nohz_balance_exit_idle(cpu);
7754 * None are in tickless mode and hence no need for NOHZ idle load
7757 if (likely(!atomic_read(&nohz.nr_cpus)))
7760 if (time_before(now, nohz.next_balance))
7763 if (rq->nr_running >= 2)
7767 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7769 sgc = sd->groups->sgc;
7770 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7779 sd = rcu_dereference(rq->sd);
7781 if ((rq->cfs.h_nr_running >= 1) &&
7782 check_cpu_capacity(rq, sd)) {
7788 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7789 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7790 sched_domain_span(sd)) < cpu)) {
7800 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7804 * run_rebalance_domains is triggered when needed from the scheduler tick.
7805 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7807 static void run_rebalance_domains(struct softirq_action *h)
7809 struct rq *this_rq = this_rq();
7810 enum cpu_idle_type idle = this_rq->idle_balance ?
7811 CPU_IDLE : CPU_NOT_IDLE;
7814 * If this cpu has a pending nohz_balance_kick, then do the
7815 * balancing on behalf of the other idle cpus whose ticks are
7816 * stopped. Do nohz_idle_balance *before* rebalance_domains to
7817 * give the idle cpus a chance to load balance. Else we may
7818 * load balance only within the local sched_domain hierarchy
7819 * and abort nohz_idle_balance altogether if we pull some load.
7821 nohz_idle_balance(this_rq, idle);
7822 rebalance_domains(this_rq, idle);
7826 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7828 void trigger_load_balance(struct rq *rq)
7830 /* Don't need to rebalance while attached to NULL domain */
7831 if (unlikely(on_null_domain(rq)))
7834 if (time_after_eq(jiffies, rq->next_balance))
7835 raise_softirq(SCHED_SOFTIRQ);
7836 #ifdef CONFIG_NO_HZ_COMMON
7837 if (nohz_kick_needed(rq))
7838 nohz_balancer_kick();
7842 static void rq_online_fair(struct rq *rq)
7846 update_runtime_enabled(rq);
7849 static void rq_offline_fair(struct rq *rq)
7853 /* Ensure any throttled groups are reachable by pick_next_task */
7854 unthrottle_offline_cfs_rqs(rq);
7857 #endif /* CONFIG_SMP */
7860 * scheduler tick hitting a task of our scheduling class:
7862 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7864 struct cfs_rq *cfs_rq;
7865 struct sched_entity *se = &curr->se;
7867 for_each_sched_entity(se) {
7868 cfs_rq = cfs_rq_of(se);
7869 entity_tick(cfs_rq, se, queued);
7872 if (!static_branch_unlikely(&sched_numa_balancing))
7873 task_tick_numa(rq, curr);
7877 * called on fork with the child task as argument from the parent's context
7878 * - child not yet on the tasklist
7879 * - preemption disabled
7881 static void task_fork_fair(struct task_struct *p)
7883 struct cfs_rq *cfs_rq;
7884 struct sched_entity *se = &p->se, *curr;
7885 int this_cpu = smp_processor_id();
7886 struct rq *rq = this_rq();
7887 unsigned long flags;
7889 raw_spin_lock_irqsave(&rq->lock, flags);
7891 update_rq_clock(rq);
7893 cfs_rq = task_cfs_rq(current);
7894 curr = cfs_rq->curr;
7897 * Not only the cpu but also the task_group of the parent might have
7898 * been changed after parent->se.parent,cfs_rq were copied to
7899 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7900 * of child point to valid ones.
7903 __set_task_cpu(p, this_cpu);
7906 update_curr(cfs_rq);
7909 se->vruntime = curr->vruntime;
7910 place_entity(cfs_rq, se, 1);
7912 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7914 * Upon rescheduling, sched_class::put_prev_task() will place
7915 * 'current' within the tree based on its new key value.
7917 swap(curr->vruntime, se->vruntime);
7921 se->vruntime -= cfs_rq->min_vruntime;
7923 raw_spin_unlock_irqrestore(&rq->lock, flags);
7927 * Priority of the task has changed. Check to see if we preempt
7931 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7933 if (!task_on_rq_queued(p))
7937 * Reschedule if we are currently running on this runqueue and
7938 * our priority decreased, or if we are not currently running on
7939 * this runqueue and our priority is higher than the current's
7941 if (rq->curr == p) {
7942 if (p->prio > oldprio)
7945 check_preempt_curr(rq, p, 0);
7948 static inline bool vruntime_normalized(struct task_struct *p)
7950 struct sched_entity *se = &p->se;
7953 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
7954 * the dequeue_entity(.flags=0) will already have normalized the
7961 * When !on_rq, vruntime of the task has usually NOT been normalized.
7962 * But there are some cases where it has already been normalized:
7964 * - A forked child which is waiting for being woken up by
7965 * wake_up_new_task().
7966 * - A task which has been woken up by try_to_wake_up() and
7967 * waiting for actually being woken up by sched_ttwu_pending().
7969 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
7975 static void detach_task_cfs_rq(struct task_struct *p)
7977 struct sched_entity *se = &p->se;
7978 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7980 if (!vruntime_normalized(p)) {
7982 * Fix up our vruntime so that the current sleep doesn't
7983 * cause 'unlimited' sleep bonus.
7985 place_entity(cfs_rq, se, 0);
7986 se->vruntime -= cfs_rq->min_vruntime;
7989 /* Catch up with the cfs_rq and remove our load when we leave */
7990 detach_entity_load_avg(cfs_rq, se);
7993 static void attach_task_cfs_rq(struct task_struct *p)
7995 struct sched_entity *se = &p->se;
7996 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7998 #ifdef CONFIG_FAIR_GROUP_SCHED
8000 * Since the real-depth could have been changed (only FAIR
8001 * class maintain depth value), reset depth properly.
8003 se->depth = se->parent ? se->parent->depth + 1 : 0;
8006 /* Synchronize task with its cfs_rq */
8007 attach_entity_load_avg(cfs_rq, se);
8009 if (!vruntime_normalized(p))
8010 se->vruntime += cfs_rq->min_vruntime;
8013 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8015 detach_task_cfs_rq(p);
8018 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8020 attach_task_cfs_rq(p);
8022 if (task_on_rq_queued(p)) {
8024 * We were most likely switched from sched_rt, so
8025 * kick off the schedule if running, otherwise just see
8026 * if we can still preempt the current task.
8031 check_preempt_curr(rq, p, 0);
8035 /* Account for a task changing its policy or group.
8037 * This routine is mostly called to set cfs_rq->curr field when a task
8038 * migrates between groups/classes.
8040 static void set_curr_task_fair(struct rq *rq)
8042 struct sched_entity *se = &rq->curr->se;
8044 for_each_sched_entity(se) {
8045 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8047 set_next_entity(cfs_rq, se);
8048 /* ensure bandwidth has been allocated on our new cfs_rq */
8049 account_cfs_rq_runtime(cfs_rq, 0);
8053 void init_cfs_rq(struct cfs_rq *cfs_rq)
8055 cfs_rq->tasks_timeline = RB_ROOT;
8056 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8057 #ifndef CONFIG_64BIT
8058 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8061 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8062 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8066 #ifdef CONFIG_FAIR_GROUP_SCHED
8067 static void task_move_group_fair(struct task_struct *p)
8069 detach_task_cfs_rq(p);
8070 set_task_rq(p, task_cpu(p));
8073 /* Tell se's cfs_rq has been changed -- migrated */
8074 p->se.avg.last_update_time = 0;
8076 attach_task_cfs_rq(p);
8079 void free_fair_sched_group(struct task_group *tg)
8083 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8085 for_each_possible_cpu(i) {
8087 kfree(tg->cfs_rq[i]);
8090 remove_entity_load_avg(tg->se[i]);
8099 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8101 struct cfs_rq *cfs_rq;
8102 struct sched_entity *se;
8105 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8108 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8112 tg->shares = NICE_0_LOAD;
8114 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8116 for_each_possible_cpu(i) {
8117 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8118 GFP_KERNEL, cpu_to_node(i));
8122 se = kzalloc_node(sizeof(struct sched_entity),
8123 GFP_KERNEL, cpu_to_node(i));
8127 init_cfs_rq(cfs_rq);
8128 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8129 init_entity_runnable_average(se);
8140 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8142 struct rq *rq = cpu_rq(cpu);
8143 unsigned long flags;
8146 * Only empty task groups can be destroyed; so we can speculatively
8147 * check on_list without danger of it being re-added.
8149 if (!tg->cfs_rq[cpu]->on_list)
8152 raw_spin_lock_irqsave(&rq->lock, flags);
8153 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8154 raw_spin_unlock_irqrestore(&rq->lock, flags);
8157 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8158 struct sched_entity *se, int cpu,
8159 struct sched_entity *parent)
8161 struct rq *rq = cpu_rq(cpu);
8165 init_cfs_rq_runtime(cfs_rq);
8167 tg->cfs_rq[cpu] = cfs_rq;
8170 /* se could be NULL for root_task_group */
8175 se->cfs_rq = &rq->cfs;
8178 se->cfs_rq = parent->my_q;
8179 se->depth = parent->depth + 1;
8183 /* guarantee group entities always have weight */
8184 update_load_set(&se->load, NICE_0_LOAD);
8185 se->parent = parent;
8188 static DEFINE_MUTEX(shares_mutex);
8190 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8193 unsigned long flags;
8196 * We can't change the weight of the root cgroup.
8201 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8203 mutex_lock(&shares_mutex);
8204 if (tg->shares == shares)
8207 tg->shares = shares;
8208 for_each_possible_cpu(i) {
8209 struct rq *rq = cpu_rq(i);
8210 struct sched_entity *se;
8213 /* Propagate contribution to hierarchy */
8214 raw_spin_lock_irqsave(&rq->lock, flags);
8216 /* Possible calls to update_curr() need rq clock */
8217 update_rq_clock(rq);
8218 for_each_sched_entity(se)
8219 update_cfs_shares(group_cfs_rq(se));
8220 raw_spin_unlock_irqrestore(&rq->lock, flags);
8224 mutex_unlock(&shares_mutex);
8227 #else /* CONFIG_FAIR_GROUP_SCHED */
8229 void free_fair_sched_group(struct task_group *tg) { }
8231 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8236 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8238 #endif /* CONFIG_FAIR_GROUP_SCHED */
8241 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8243 struct sched_entity *se = &task->se;
8244 unsigned int rr_interval = 0;
8247 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8250 if (rq->cfs.load.weight)
8251 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8257 * All the scheduling class methods:
8259 const struct sched_class fair_sched_class = {
8260 .next = &idle_sched_class,
8261 .enqueue_task = enqueue_task_fair,
8262 .dequeue_task = dequeue_task_fair,
8263 .yield_task = yield_task_fair,
8264 .yield_to_task = yield_to_task_fair,
8266 .check_preempt_curr = check_preempt_wakeup,
8268 .pick_next_task = pick_next_task_fair,
8269 .put_prev_task = put_prev_task_fair,
8272 .select_task_rq = select_task_rq_fair,
8273 .migrate_task_rq = migrate_task_rq_fair,
8275 .rq_online = rq_online_fair,
8276 .rq_offline = rq_offline_fair,
8278 .task_waking = task_waking_fair,
8279 .task_dead = task_dead_fair,
8280 .set_cpus_allowed = set_cpus_allowed_common,
8283 .set_curr_task = set_curr_task_fair,
8284 .task_tick = task_tick_fair,
8285 .task_fork = task_fork_fair,
8287 .prio_changed = prio_changed_fair,
8288 .switched_from = switched_from_fair,
8289 .switched_to = switched_to_fair,
8291 .get_rr_interval = get_rr_interval_fair,
8293 .update_curr = update_curr_fair,
8295 #ifdef CONFIG_FAIR_GROUP_SCHED
8296 .task_move_group = task_move_group_fair,
8300 #ifdef CONFIG_SCHED_DEBUG
8301 void print_cfs_stats(struct seq_file *m, int cpu)
8303 struct cfs_rq *cfs_rq;
8306 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8307 print_cfs_rq(m, cpu, cfs_rq);
8311 #ifdef CONFIG_NUMA_BALANCING
8312 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8315 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8317 for_each_online_node(node) {
8318 if (p->numa_faults) {
8319 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8320 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8322 if (p->numa_group) {
8323 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8324 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8326 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8329 #endif /* CONFIG_NUMA_BALANCING */
8330 #endif /* CONFIG_SCHED_DEBUG */
8332 __init void init_sched_fair_class(void)
8335 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8337 #ifdef CONFIG_NO_HZ_COMMON
8338 nohz.next_balance = jiffies;
8339 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8340 cpu_notifier(sched_ilb_notifier, 0);