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 (!numabalancing_enabled)
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];
2519 * We can represent the historical contribution to runnable average as the
2520 * coefficients of a geometric series. To do this we sub-divide our runnable
2521 * history into segments of approximately 1ms (1024us); label the segment that
2522 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2524 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2526 * (now) (~1ms ago) (~2ms ago)
2528 * Let u_i denote the fraction of p_i that the entity was runnable.
2530 * We then designate the fractions u_i as our co-efficients, yielding the
2531 * following representation of historical load:
2532 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2534 * We choose y based on the with of a reasonably scheduling period, fixing:
2537 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2538 * approximately half as much as the contribution to load within the last ms
2541 * When a period "rolls over" and we have new u_0`, multiplying the previous
2542 * sum again by y is sufficient to update:
2543 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2544 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2546 static __always_inline int
2547 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2548 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2552 int delta_w, decayed = 0;
2553 unsigned long scale_freq = arch_scale_freq_capacity(NULL, cpu);
2555 delta = now - sa->last_update_time;
2557 * This should only happen when time goes backwards, which it
2558 * unfortunately does during sched clock init when we swap over to TSC.
2560 if ((s64)delta < 0) {
2561 sa->last_update_time = now;
2566 * Use 1024ns as the unit of measurement since it's a reasonable
2567 * approximation of 1us and fast to compute.
2572 sa->last_update_time = now;
2574 /* delta_w is the amount already accumulated against our next period */
2575 delta_w = sa->period_contrib;
2576 if (delta + delta_w >= 1024) {
2579 /* how much left for next period will start over, we don't know yet */
2580 sa->period_contrib = 0;
2583 * Now that we know we're crossing a period boundary, figure
2584 * out how much from delta we need to complete the current
2585 * period and accrue it.
2587 delta_w = 1024 - delta_w;
2589 sa->load_sum += weight * delta_w;
2591 cfs_rq->runnable_load_sum += weight * delta_w;
2594 sa->util_sum += delta_w * scale_freq >> SCHED_CAPACITY_SHIFT;
2598 /* Figure out how many additional periods this update spans */
2599 periods = delta / 1024;
2602 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2604 cfs_rq->runnable_load_sum =
2605 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2607 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2609 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2610 contrib = __compute_runnable_contrib(periods);
2612 sa->load_sum += weight * contrib;
2614 cfs_rq->runnable_load_sum += weight * contrib;
2617 sa->util_sum += contrib * scale_freq >> SCHED_CAPACITY_SHIFT;
2620 /* Remainder of delta accrued against u_0` */
2622 sa->load_sum += weight * delta;
2624 cfs_rq->runnable_load_sum += weight * delta;
2627 sa->util_sum += delta * scale_freq >> SCHED_CAPACITY_SHIFT;
2629 sa->period_contrib += delta;
2632 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2634 cfs_rq->runnable_load_avg =
2635 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2637 sa->util_avg = (sa->util_sum << SCHED_LOAD_SHIFT) / LOAD_AVG_MAX;
2643 #ifdef CONFIG_FAIR_GROUP_SCHED
2645 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2646 * and effective_load (which is not done because it is too costly).
2648 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2650 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2652 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2653 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2654 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2658 #else /* CONFIG_FAIR_GROUP_SCHED */
2659 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2660 #endif /* CONFIG_FAIR_GROUP_SCHED */
2662 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2664 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2665 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2667 struct sched_avg *sa = &cfs_rq->avg;
2670 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2671 long r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2672 sa->load_avg = max_t(long, sa->load_avg - r, 0);
2673 sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
2676 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2677 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2678 sa->util_avg = max_t(long, sa->util_avg - r, 0);
2679 sa->util_sum = max_t(s32, sa->util_sum -
2680 ((r * LOAD_AVG_MAX) >> SCHED_LOAD_SHIFT), 0);
2683 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2684 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2686 #ifndef CONFIG_64BIT
2688 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2694 /* Update task and its cfs_rq load average */
2695 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2697 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2698 u64 now = cfs_rq_clock_task(cfs_rq);
2699 int cpu = cpu_of(rq_of(cfs_rq));
2702 * Track task load average for carrying it to new CPU after migrated, and
2703 * track group sched_entity load average for task_h_load calc in migration
2705 __update_load_avg(now, cpu, &se->avg,
2706 se->on_rq * scale_load_down(se->load.weight),
2707 cfs_rq->curr == se, NULL);
2709 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2710 update_tg_load_avg(cfs_rq, 0);
2713 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2715 if (!sched_feat(ATTACH_AGE_LOAD))
2719 * If we got migrated (either between CPUs or between cgroups) we'll
2720 * have aged the average right before clearing @last_update_time.
2722 if (se->avg.last_update_time) {
2723 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2724 &se->avg, 0, 0, NULL);
2727 * XXX: we could have just aged the entire load away if we've been
2728 * absent from the fair class for too long.
2733 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2734 cfs_rq->avg.load_avg += se->avg.load_avg;
2735 cfs_rq->avg.load_sum += se->avg.load_sum;
2736 cfs_rq->avg.util_avg += se->avg.util_avg;
2737 cfs_rq->avg.util_sum += se->avg.util_sum;
2740 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2742 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2743 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2744 cfs_rq->curr == se, NULL);
2746 cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
2747 cfs_rq->avg.load_sum = max_t(s64, cfs_rq->avg.load_sum - se->avg.load_sum, 0);
2748 cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
2749 cfs_rq->avg.util_sum = max_t(s32, cfs_rq->avg.util_sum - se->avg.util_sum, 0);
2752 /* Add the load generated by se into cfs_rq's load average */
2754 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2756 struct sched_avg *sa = &se->avg;
2757 u64 now = cfs_rq_clock_task(cfs_rq);
2758 int migrated, decayed;
2760 migrated = !sa->last_update_time;
2762 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2763 se->on_rq * scale_load_down(se->load.weight),
2764 cfs_rq->curr == se, NULL);
2767 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2769 cfs_rq->runnable_load_avg += sa->load_avg;
2770 cfs_rq->runnable_load_sum += sa->load_sum;
2773 attach_entity_load_avg(cfs_rq, se);
2775 if (decayed || migrated)
2776 update_tg_load_avg(cfs_rq, 0);
2779 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2781 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2783 update_load_avg(se, 1);
2785 cfs_rq->runnable_load_avg =
2786 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2787 cfs_rq->runnable_load_sum =
2788 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2792 * Task first catches up with cfs_rq, and then subtract
2793 * itself from the cfs_rq (task must be off the queue now).
2795 void remove_entity_load_avg(struct sched_entity *se)
2797 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2798 u64 last_update_time;
2800 #ifndef CONFIG_64BIT
2801 u64 last_update_time_copy;
2804 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2806 last_update_time = cfs_rq->avg.last_update_time;
2807 } while (last_update_time != last_update_time_copy);
2809 last_update_time = cfs_rq->avg.last_update_time;
2812 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2813 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2814 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2818 * Update the rq's load with the elapsed running time before entering
2819 * idle. if the last scheduled task is not a CFS task, idle_enter will
2820 * be the only way to update the runnable statistic.
2822 void idle_enter_fair(struct rq *this_rq)
2827 * Update the rq's load with the elapsed idle time before a task is
2828 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2829 * be the only way to update the runnable statistic.
2831 void idle_exit_fair(struct rq *this_rq)
2835 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2837 return cfs_rq->runnable_load_avg;
2840 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2842 return cfs_rq->avg.load_avg;
2845 static int idle_balance(struct rq *this_rq);
2847 #else /* CONFIG_SMP */
2849 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2851 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2853 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2854 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2857 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2859 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2861 static inline int idle_balance(struct rq *rq)
2866 #endif /* CONFIG_SMP */
2868 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2870 #ifdef CONFIG_SCHEDSTATS
2871 struct task_struct *tsk = NULL;
2873 if (entity_is_task(se))
2876 if (se->statistics.sleep_start) {
2877 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2882 if (unlikely(delta > se->statistics.sleep_max))
2883 se->statistics.sleep_max = delta;
2885 se->statistics.sleep_start = 0;
2886 se->statistics.sum_sleep_runtime += delta;
2889 account_scheduler_latency(tsk, delta >> 10, 1);
2890 trace_sched_stat_sleep(tsk, delta);
2893 if (se->statistics.block_start) {
2894 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2899 if (unlikely(delta > se->statistics.block_max))
2900 se->statistics.block_max = delta;
2902 se->statistics.block_start = 0;
2903 se->statistics.sum_sleep_runtime += delta;
2906 if (tsk->in_iowait) {
2907 se->statistics.iowait_sum += delta;
2908 se->statistics.iowait_count++;
2909 trace_sched_stat_iowait(tsk, delta);
2912 trace_sched_stat_blocked(tsk, delta);
2915 * Blocking time is in units of nanosecs, so shift by
2916 * 20 to get a milliseconds-range estimation of the
2917 * amount of time that the task spent sleeping:
2919 if (unlikely(prof_on == SLEEP_PROFILING)) {
2920 profile_hits(SLEEP_PROFILING,
2921 (void *)get_wchan(tsk),
2924 account_scheduler_latency(tsk, delta >> 10, 0);
2930 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2932 #ifdef CONFIG_SCHED_DEBUG
2933 s64 d = se->vruntime - cfs_rq->min_vruntime;
2938 if (d > 3*sysctl_sched_latency)
2939 schedstat_inc(cfs_rq, nr_spread_over);
2944 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2946 u64 vruntime = cfs_rq->min_vruntime;
2949 * The 'current' period is already promised to the current tasks,
2950 * however the extra weight of the new task will slow them down a
2951 * little, place the new task so that it fits in the slot that
2952 * stays open at the end.
2954 if (initial && sched_feat(START_DEBIT))
2955 vruntime += sched_vslice(cfs_rq, se);
2957 /* sleeps up to a single latency don't count. */
2959 unsigned long thresh = sysctl_sched_latency;
2962 * Halve their sleep time's effect, to allow
2963 * for a gentler effect of sleepers:
2965 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2971 /* ensure we never gain time by being placed backwards. */
2972 se->vruntime = max_vruntime(se->vruntime, vruntime);
2975 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2978 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2981 * Update the normalized vruntime before updating min_vruntime
2982 * through calling update_curr().
2984 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2985 se->vruntime += cfs_rq->min_vruntime;
2988 * Update run-time statistics of the 'current'.
2990 update_curr(cfs_rq);
2991 enqueue_entity_load_avg(cfs_rq, se);
2992 account_entity_enqueue(cfs_rq, se);
2993 update_cfs_shares(cfs_rq);
2995 if (flags & ENQUEUE_WAKEUP) {
2996 place_entity(cfs_rq, se, 0);
2997 enqueue_sleeper(cfs_rq, se);
3000 update_stats_enqueue(cfs_rq, se);
3001 check_spread(cfs_rq, se);
3002 if (se != cfs_rq->curr)
3003 __enqueue_entity(cfs_rq, se);
3006 if (cfs_rq->nr_running == 1) {
3007 list_add_leaf_cfs_rq(cfs_rq);
3008 check_enqueue_throttle(cfs_rq);
3012 static void __clear_buddies_last(struct sched_entity *se)
3014 for_each_sched_entity(se) {
3015 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3016 if (cfs_rq->last != se)
3019 cfs_rq->last = NULL;
3023 static void __clear_buddies_next(struct sched_entity *se)
3025 for_each_sched_entity(se) {
3026 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3027 if (cfs_rq->next != se)
3030 cfs_rq->next = NULL;
3034 static void __clear_buddies_skip(struct sched_entity *se)
3036 for_each_sched_entity(se) {
3037 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3038 if (cfs_rq->skip != se)
3041 cfs_rq->skip = NULL;
3045 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3047 if (cfs_rq->last == se)
3048 __clear_buddies_last(se);
3050 if (cfs_rq->next == se)
3051 __clear_buddies_next(se);
3053 if (cfs_rq->skip == se)
3054 __clear_buddies_skip(se);
3057 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3060 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3063 * Update run-time statistics of the 'current'.
3065 update_curr(cfs_rq);
3066 dequeue_entity_load_avg(cfs_rq, se);
3068 update_stats_dequeue(cfs_rq, se);
3069 if (flags & DEQUEUE_SLEEP) {
3070 #ifdef CONFIG_SCHEDSTATS
3071 if (entity_is_task(se)) {
3072 struct task_struct *tsk = task_of(se);
3074 if (tsk->state & TASK_INTERRUPTIBLE)
3075 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3076 if (tsk->state & TASK_UNINTERRUPTIBLE)
3077 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3082 clear_buddies(cfs_rq, se);
3084 if (se != cfs_rq->curr)
3085 __dequeue_entity(cfs_rq, se);
3087 account_entity_dequeue(cfs_rq, se);
3090 * Normalize the entity after updating the min_vruntime because the
3091 * update can refer to the ->curr item and we need to reflect this
3092 * movement in our normalized position.
3094 if (!(flags & DEQUEUE_SLEEP))
3095 se->vruntime -= cfs_rq->min_vruntime;
3097 /* return excess runtime on last dequeue */
3098 return_cfs_rq_runtime(cfs_rq);
3100 update_min_vruntime(cfs_rq);
3101 update_cfs_shares(cfs_rq);
3105 * Preempt the current task with a newly woken task if needed:
3108 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3110 unsigned long ideal_runtime, delta_exec;
3111 struct sched_entity *se;
3114 ideal_runtime = sched_slice(cfs_rq, curr);
3115 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3116 if (delta_exec > ideal_runtime) {
3117 resched_curr(rq_of(cfs_rq));
3119 * The current task ran long enough, ensure it doesn't get
3120 * re-elected due to buddy favours.
3122 clear_buddies(cfs_rq, curr);
3127 * Ensure that a task that missed wakeup preemption by a
3128 * narrow margin doesn't have to wait for a full slice.
3129 * This also mitigates buddy induced latencies under load.
3131 if (delta_exec < sysctl_sched_min_granularity)
3134 se = __pick_first_entity(cfs_rq);
3135 delta = curr->vruntime - se->vruntime;
3140 if (delta > ideal_runtime)
3141 resched_curr(rq_of(cfs_rq));
3145 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3147 /* 'current' is not kept within the tree. */
3150 * Any task has to be enqueued before it get to execute on
3151 * a CPU. So account for the time it spent waiting on the
3154 update_stats_wait_end(cfs_rq, se);
3155 __dequeue_entity(cfs_rq, se);
3156 update_load_avg(se, 1);
3159 update_stats_curr_start(cfs_rq, se);
3161 #ifdef CONFIG_SCHEDSTATS
3163 * Track our maximum slice length, if the CPU's load is at
3164 * least twice that of our own weight (i.e. dont track it
3165 * when there are only lesser-weight tasks around):
3167 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3168 se->statistics.slice_max = max(se->statistics.slice_max,
3169 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3172 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3176 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3179 * Pick the next process, keeping these things in mind, in this order:
3180 * 1) keep things fair between processes/task groups
3181 * 2) pick the "next" process, since someone really wants that to run
3182 * 3) pick the "last" process, for cache locality
3183 * 4) do not run the "skip" process, if something else is available
3185 static struct sched_entity *
3186 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3188 struct sched_entity *left = __pick_first_entity(cfs_rq);
3189 struct sched_entity *se;
3192 * If curr is set we have to see if its left of the leftmost entity
3193 * still in the tree, provided there was anything in the tree at all.
3195 if (!left || (curr && entity_before(curr, left)))
3198 se = left; /* ideally we run the leftmost entity */
3201 * Avoid running the skip buddy, if running something else can
3202 * be done without getting too unfair.
3204 if (cfs_rq->skip == se) {
3205 struct sched_entity *second;
3208 second = __pick_first_entity(cfs_rq);
3210 second = __pick_next_entity(se);
3211 if (!second || (curr && entity_before(curr, second)))
3215 if (second && wakeup_preempt_entity(second, left) < 1)
3220 * Prefer last buddy, try to return the CPU to a preempted task.
3222 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3226 * Someone really wants this to run. If it's not unfair, run it.
3228 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3231 clear_buddies(cfs_rq, se);
3236 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3238 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3241 * If still on the runqueue then deactivate_task()
3242 * was not called and update_curr() has to be done:
3245 update_curr(cfs_rq);
3247 /* throttle cfs_rqs exceeding runtime */
3248 check_cfs_rq_runtime(cfs_rq);
3250 check_spread(cfs_rq, prev);
3252 update_stats_wait_start(cfs_rq, prev);
3253 /* Put 'current' back into the tree. */
3254 __enqueue_entity(cfs_rq, prev);
3255 /* in !on_rq case, update occurred at dequeue */
3256 update_load_avg(prev, 0);
3258 cfs_rq->curr = NULL;
3262 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3265 * Update run-time statistics of the 'current'.
3267 update_curr(cfs_rq);
3270 * Ensure that runnable average is periodically updated.
3272 update_load_avg(curr, 1);
3273 update_cfs_shares(cfs_rq);
3275 #ifdef CONFIG_SCHED_HRTICK
3277 * queued ticks are scheduled to match the slice, so don't bother
3278 * validating it and just reschedule.
3281 resched_curr(rq_of(cfs_rq));
3285 * don't let the period tick interfere with the hrtick preemption
3287 if (!sched_feat(DOUBLE_TICK) &&
3288 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3292 if (cfs_rq->nr_running > 1)
3293 check_preempt_tick(cfs_rq, curr);
3297 /**************************************************
3298 * CFS bandwidth control machinery
3301 #ifdef CONFIG_CFS_BANDWIDTH
3303 #ifdef HAVE_JUMP_LABEL
3304 static struct static_key __cfs_bandwidth_used;
3306 static inline bool cfs_bandwidth_used(void)
3308 return static_key_false(&__cfs_bandwidth_used);
3311 void cfs_bandwidth_usage_inc(void)
3313 static_key_slow_inc(&__cfs_bandwidth_used);
3316 void cfs_bandwidth_usage_dec(void)
3318 static_key_slow_dec(&__cfs_bandwidth_used);
3320 #else /* HAVE_JUMP_LABEL */
3321 static bool cfs_bandwidth_used(void)
3326 void cfs_bandwidth_usage_inc(void) {}
3327 void cfs_bandwidth_usage_dec(void) {}
3328 #endif /* HAVE_JUMP_LABEL */
3331 * default period for cfs group bandwidth.
3332 * default: 0.1s, units: nanoseconds
3334 static inline u64 default_cfs_period(void)
3336 return 100000000ULL;
3339 static inline u64 sched_cfs_bandwidth_slice(void)
3341 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3345 * Replenish runtime according to assigned quota and update expiration time.
3346 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3347 * additional synchronization around rq->lock.
3349 * requires cfs_b->lock
3351 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3355 if (cfs_b->quota == RUNTIME_INF)
3358 now = sched_clock_cpu(smp_processor_id());
3359 cfs_b->runtime = cfs_b->quota;
3360 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3363 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3365 return &tg->cfs_bandwidth;
3368 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3369 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3371 if (unlikely(cfs_rq->throttle_count))
3372 return cfs_rq->throttled_clock_task;
3374 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3377 /* returns 0 on failure to allocate runtime */
3378 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3380 struct task_group *tg = cfs_rq->tg;
3381 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3382 u64 amount = 0, min_amount, expires;
3384 /* note: this is a positive sum as runtime_remaining <= 0 */
3385 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3387 raw_spin_lock(&cfs_b->lock);
3388 if (cfs_b->quota == RUNTIME_INF)
3389 amount = min_amount;
3391 start_cfs_bandwidth(cfs_b);
3393 if (cfs_b->runtime > 0) {
3394 amount = min(cfs_b->runtime, min_amount);
3395 cfs_b->runtime -= amount;
3399 expires = cfs_b->runtime_expires;
3400 raw_spin_unlock(&cfs_b->lock);
3402 cfs_rq->runtime_remaining += amount;
3404 * we may have advanced our local expiration to account for allowed
3405 * spread between our sched_clock and the one on which runtime was
3408 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3409 cfs_rq->runtime_expires = expires;
3411 return cfs_rq->runtime_remaining > 0;
3415 * Note: This depends on the synchronization provided by sched_clock and the
3416 * fact that rq->clock snapshots this value.
3418 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3420 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3422 /* if the deadline is ahead of our clock, nothing to do */
3423 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3426 if (cfs_rq->runtime_remaining < 0)
3430 * If the local deadline has passed we have to consider the
3431 * possibility that our sched_clock is 'fast' and the global deadline
3432 * has not truly expired.
3434 * Fortunately we can check determine whether this the case by checking
3435 * whether the global deadline has advanced. It is valid to compare
3436 * cfs_b->runtime_expires without any locks since we only care about
3437 * exact equality, so a partial write will still work.
3440 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3441 /* extend local deadline, drift is bounded above by 2 ticks */
3442 cfs_rq->runtime_expires += TICK_NSEC;
3444 /* global deadline is ahead, expiration has passed */
3445 cfs_rq->runtime_remaining = 0;
3449 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3451 /* dock delta_exec before expiring quota (as it could span periods) */
3452 cfs_rq->runtime_remaining -= delta_exec;
3453 expire_cfs_rq_runtime(cfs_rq);
3455 if (likely(cfs_rq->runtime_remaining > 0))
3459 * if we're unable to extend our runtime we resched so that the active
3460 * hierarchy can be throttled
3462 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3463 resched_curr(rq_of(cfs_rq));
3466 static __always_inline
3467 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3469 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3472 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3475 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3477 return cfs_bandwidth_used() && cfs_rq->throttled;
3480 /* check whether cfs_rq, or any parent, is throttled */
3481 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3483 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3487 * Ensure that neither of the group entities corresponding to src_cpu or
3488 * dest_cpu are members of a throttled hierarchy when performing group
3489 * load-balance operations.
3491 static inline int throttled_lb_pair(struct task_group *tg,
3492 int src_cpu, int dest_cpu)
3494 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3496 src_cfs_rq = tg->cfs_rq[src_cpu];
3497 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3499 return throttled_hierarchy(src_cfs_rq) ||
3500 throttled_hierarchy(dest_cfs_rq);
3503 /* updated child weight may affect parent so we have to do this bottom up */
3504 static int tg_unthrottle_up(struct task_group *tg, void *data)
3506 struct rq *rq = data;
3507 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3509 cfs_rq->throttle_count--;
3511 if (!cfs_rq->throttle_count) {
3512 /* adjust cfs_rq_clock_task() */
3513 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3514 cfs_rq->throttled_clock_task;
3521 static int tg_throttle_down(struct task_group *tg, void *data)
3523 struct rq *rq = data;
3524 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3526 /* group is entering throttled state, stop time */
3527 if (!cfs_rq->throttle_count)
3528 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3529 cfs_rq->throttle_count++;
3534 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3536 struct rq *rq = rq_of(cfs_rq);
3537 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3538 struct sched_entity *se;
3539 long task_delta, dequeue = 1;
3542 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3544 /* freeze hierarchy runnable averages while throttled */
3546 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3549 task_delta = cfs_rq->h_nr_running;
3550 for_each_sched_entity(se) {
3551 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3552 /* throttled entity or throttle-on-deactivate */
3557 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3558 qcfs_rq->h_nr_running -= task_delta;
3560 if (qcfs_rq->load.weight)
3565 sub_nr_running(rq, task_delta);
3567 cfs_rq->throttled = 1;
3568 cfs_rq->throttled_clock = rq_clock(rq);
3569 raw_spin_lock(&cfs_b->lock);
3570 empty = list_empty(&cfs_b->throttled_cfs_rq);
3573 * Add to the _head_ of the list, so that an already-started
3574 * distribute_cfs_runtime will not see us
3576 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3579 * If we're the first throttled task, make sure the bandwidth
3583 start_cfs_bandwidth(cfs_b);
3585 raw_spin_unlock(&cfs_b->lock);
3588 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3590 struct rq *rq = rq_of(cfs_rq);
3591 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3592 struct sched_entity *se;
3596 se = cfs_rq->tg->se[cpu_of(rq)];
3598 cfs_rq->throttled = 0;
3600 update_rq_clock(rq);
3602 raw_spin_lock(&cfs_b->lock);
3603 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3604 list_del_rcu(&cfs_rq->throttled_list);
3605 raw_spin_unlock(&cfs_b->lock);
3607 /* update hierarchical throttle state */
3608 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3610 if (!cfs_rq->load.weight)
3613 task_delta = cfs_rq->h_nr_running;
3614 for_each_sched_entity(se) {
3618 cfs_rq = cfs_rq_of(se);
3620 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3621 cfs_rq->h_nr_running += task_delta;
3623 if (cfs_rq_throttled(cfs_rq))
3628 add_nr_running(rq, task_delta);
3630 /* determine whether we need to wake up potentially idle cpu */
3631 if (rq->curr == rq->idle && rq->cfs.nr_running)
3635 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3636 u64 remaining, u64 expires)
3638 struct cfs_rq *cfs_rq;
3640 u64 starting_runtime = remaining;
3643 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3645 struct rq *rq = rq_of(cfs_rq);
3647 raw_spin_lock(&rq->lock);
3648 if (!cfs_rq_throttled(cfs_rq))
3651 runtime = -cfs_rq->runtime_remaining + 1;
3652 if (runtime > remaining)
3653 runtime = remaining;
3654 remaining -= runtime;
3656 cfs_rq->runtime_remaining += runtime;
3657 cfs_rq->runtime_expires = expires;
3659 /* we check whether we're throttled above */
3660 if (cfs_rq->runtime_remaining > 0)
3661 unthrottle_cfs_rq(cfs_rq);
3664 raw_spin_unlock(&rq->lock);
3671 return starting_runtime - remaining;
3675 * Responsible for refilling a task_group's bandwidth and unthrottling its
3676 * cfs_rqs as appropriate. If there has been no activity within the last
3677 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3678 * used to track this state.
3680 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3682 u64 runtime, runtime_expires;
3685 /* no need to continue the timer with no bandwidth constraint */
3686 if (cfs_b->quota == RUNTIME_INF)
3687 goto out_deactivate;
3689 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3690 cfs_b->nr_periods += overrun;
3693 * idle depends on !throttled (for the case of a large deficit), and if
3694 * we're going inactive then everything else can be deferred
3696 if (cfs_b->idle && !throttled)
3697 goto out_deactivate;
3699 __refill_cfs_bandwidth_runtime(cfs_b);
3702 /* mark as potentially idle for the upcoming period */
3707 /* account preceding periods in which throttling occurred */
3708 cfs_b->nr_throttled += overrun;
3710 runtime_expires = cfs_b->runtime_expires;
3713 * This check is repeated as we are holding onto the new bandwidth while
3714 * we unthrottle. This can potentially race with an unthrottled group
3715 * trying to acquire new bandwidth from the global pool. This can result
3716 * in us over-using our runtime if it is all used during this loop, but
3717 * only by limited amounts in that extreme case.
3719 while (throttled && cfs_b->runtime > 0) {
3720 runtime = cfs_b->runtime;
3721 raw_spin_unlock(&cfs_b->lock);
3722 /* we can't nest cfs_b->lock while distributing bandwidth */
3723 runtime = distribute_cfs_runtime(cfs_b, runtime,
3725 raw_spin_lock(&cfs_b->lock);
3727 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3729 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3733 * While we are ensured activity in the period following an
3734 * unthrottle, this also covers the case in which the new bandwidth is
3735 * insufficient to cover the existing bandwidth deficit. (Forcing the
3736 * timer to remain active while there are any throttled entities.)
3746 /* a cfs_rq won't donate quota below this amount */
3747 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3748 /* minimum remaining period time to redistribute slack quota */
3749 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3750 /* how long we wait to gather additional slack before distributing */
3751 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3754 * Are we near the end of the current quota period?
3756 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3757 * hrtimer base being cleared by hrtimer_start. In the case of
3758 * migrate_hrtimers, base is never cleared, so we are fine.
3760 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3762 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3765 /* if the call-back is running a quota refresh is already occurring */
3766 if (hrtimer_callback_running(refresh_timer))
3769 /* is a quota refresh about to occur? */
3770 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3771 if (remaining < min_expire)
3777 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3779 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3781 /* if there's a quota refresh soon don't bother with slack */
3782 if (runtime_refresh_within(cfs_b, min_left))
3785 hrtimer_start(&cfs_b->slack_timer,
3786 ns_to_ktime(cfs_bandwidth_slack_period),
3790 /* we know any runtime found here is valid as update_curr() precedes return */
3791 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3793 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3794 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3796 if (slack_runtime <= 0)
3799 raw_spin_lock(&cfs_b->lock);
3800 if (cfs_b->quota != RUNTIME_INF &&
3801 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3802 cfs_b->runtime += slack_runtime;
3804 /* we are under rq->lock, defer unthrottling using a timer */
3805 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3806 !list_empty(&cfs_b->throttled_cfs_rq))
3807 start_cfs_slack_bandwidth(cfs_b);
3809 raw_spin_unlock(&cfs_b->lock);
3811 /* even if it's not valid for return we don't want to try again */
3812 cfs_rq->runtime_remaining -= slack_runtime;
3815 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3817 if (!cfs_bandwidth_used())
3820 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3823 __return_cfs_rq_runtime(cfs_rq);
3827 * This is done with a timer (instead of inline with bandwidth return) since
3828 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3830 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3832 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3835 /* confirm we're still not at a refresh boundary */
3836 raw_spin_lock(&cfs_b->lock);
3837 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3838 raw_spin_unlock(&cfs_b->lock);
3842 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3843 runtime = cfs_b->runtime;
3845 expires = cfs_b->runtime_expires;
3846 raw_spin_unlock(&cfs_b->lock);
3851 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3853 raw_spin_lock(&cfs_b->lock);
3854 if (expires == cfs_b->runtime_expires)
3855 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3856 raw_spin_unlock(&cfs_b->lock);
3860 * When a group wakes up we want to make sure that its quota is not already
3861 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3862 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3864 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3866 if (!cfs_bandwidth_used())
3869 /* an active group must be handled by the update_curr()->put() path */
3870 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3873 /* ensure the group is not already throttled */
3874 if (cfs_rq_throttled(cfs_rq))
3877 /* update runtime allocation */
3878 account_cfs_rq_runtime(cfs_rq, 0);
3879 if (cfs_rq->runtime_remaining <= 0)
3880 throttle_cfs_rq(cfs_rq);
3883 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3884 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3886 if (!cfs_bandwidth_used())
3889 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3893 * it's possible for a throttled entity to be forced into a running
3894 * state (e.g. set_curr_task), in this case we're finished.
3896 if (cfs_rq_throttled(cfs_rq))
3899 throttle_cfs_rq(cfs_rq);
3903 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3905 struct cfs_bandwidth *cfs_b =
3906 container_of(timer, struct cfs_bandwidth, slack_timer);
3908 do_sched_cfs_slack_timer(cfs_b);
3910 return HRTIMER_NORESTART;
3913 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3915 struct cfs_bandwidth *cfs_b =
3916 container_of(timer, struct cfs_bandwidth, period_timer);
3920 raw_spin_lock(&cfs_b->lock);
3922 overrun = hrtimer_forward_now(timer, cfs_b->period);
3926 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3929 cfs_b->period_active = 0;
3930 raw_spin_unlock(&cfs_b->lock);
3932 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3935 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3937 raw_spin_lock_init(&cfs_b->lock);
3939 cfs_b->quota = RUNTIME_INF;
3940 cfs_b->period = ns_to_ktime(default_cfs_period());
3942 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3943 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
3944 cfs_b->period_timer.function = sched_cfs_period_timer;
3945 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3946 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3949 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3951 cfs_rq->runtime_enabled = 0;
3952 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3955 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3957 lockdep_assert_held(&cfs_b->lock);
3959 if (!cfs_b->period_active) {
3960 cfs_b->period_active = 1;
3961 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
3962 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
3966 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3968 /* init_cfs_bandwidth() was not called */
3969 if (!cfs_b->throttled_cfs_rq.next)
3972 hrtimer_cancel(&cfs_b->period_timer);
3973 hrtimer_cancel(&cfs_b->slack_timer);
3976 static void __maybe_unused update_runtime_enabled(struct rq *rq)
3978 struct cfs_rq *cfs_rq;
3980 for_each_leaf_cfs_rq(rq, cfs_rq) {
3981 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
3983 raw_spin_lock(&cfs_b->lock);
3984 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
3985 raw_spin_unlock(&cfs_b->lock);
3989 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3991 struct cfs_rq *cfs_rq;
3993 for_each_leaf_cfs_rq(rq, cfs_rq) {
3994 if (!cfs_rq->runtime_enabled)
3998 * clock_task is not advancing so we just need to make sure
3999 * there's some valid quota amount
4001 cfs_rq->runtime_remaining = 1;
4003 * Offline rq is schedulable till cpu is completely disabled
4004 * in take_cpu_down(), so we prevent new cfs throttling here.
4006 cfs_rq->runtime_enabled = 0;
4008 if (cfs_rq_throttled(cfs_rq))
4009 unthrottle_cfs_rq(cfs_rq);
4013 #else /* CONFIG_CFS_BANDWIDTH */
4014 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4016 return rq_clock_task(rq_of(cfs_rq));
4019 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4020 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4021 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4022 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4024 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4029 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4034 static inline int throttled_lb_pair(struct task_group *tg,
4035 int src_cpu, int dest_cpu)
4040 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4042 #ifdef CONFIG_FAIR_GROUP_SCHED
4043 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4046 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4050 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4051 static inline void update_runtime_enabled(struct rq *rq) {}
4052 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4054 #endif /* CONFIG_CFS_BANDWIDTH */
4056 /**************************************************
4057 * CFS operations on tasks:
4060 #ifdef CONFIG_SCHED_HRTICK
4061 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4063 struct sched_entity *se = &p->se;
4064 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4066 WARN_ON(task_rq(p) != rq);
4068 if (cfs_rq->nr_running > 1) {
4069 u64 slice = sched_slice(cfs_rq, se);
4070 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4071 s64 delta = slice - ran;
4078 hrtick_start(rq, delta);
4083 * called from enqueue/dequeue and updates the hrtick when the
4084 * current task is from our class and nr_running is low enough
4087 static void hrtick_update(struct rq *rq)
4089 struct task_struct *curr = rq->curr;
4091 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4094 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4095 hrtick_start_fair(rq, curr);
4097 #else /* !CONFIG_SCHED_HRTICK */
4099 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4103 static inline void hrtick_update(struct rq *rq)
4109 * The enqueue_task method is called before nr_running is
4110 * increased. Here we update the fair scheduling stats and
4111 * then put the task into the rbtree:
4114 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4116 struct cfs_rq *cfs_rq;
4117 struct sched_entity *se = &p->se;
4119 for_each_sched_entity(se) {
4122 cfs_rq = cfs_rq_of(se);
4123 enqueue_entity(cfs_rq, se, flags);
4126 * end evaluation on encountering a throttled cfs_rq
4128 * note: in the case of encountering a throttled cfs_rq we will
4129 * post the final h_nr_running increment below.
4131 if (cfs_rq_throttled(cfs_rq))
4133 cfs_rq->h_nr_running++;
4135 flags = ENQUEUE_WAKEUP;
4138 for_each_sched_entity(se) {
4139 cfs_rq = cfs_rq_of(se);
4140 cfs_rq->h_nr_running++;
4142 if (cfs_rq_throttled(cfs_rq))
4145 update_load_avg(se, 1);
4146 update_cfs_shares(cfs_rq);
4150 add_nr_running(rq, 1);
4155 static void set_next_buddy(struct sched_entity *se);
4158 * The dequeue_task method is called before nr_running is
4159 * decreased. We remove the task from the rbtree and
4160 * update the fair scheduling stats:
4162 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4164 struct cfs_rq *cfs_rq;
4165 struct sched_entity *se = &p->se;
4166 int task_sleep = flags & DEQUEUE_SLEEP;
4168 for_each_sched_entity(se) {
4169 cfs_rq = cfs_rq_of(se);
4170 dequeue_entity(cfs_rq, se, flags);
4173 * end evaluation on encountering a throttled cfs_rq
4175 * note: in the case of encountering a throttled cfs_rq we will
4176 * post the final h_nr_running decrement below.
4178 if (cfs_rq_throttled(cfs_rq))
4180 cfs_rq->h_nr_running--;
4182 /* Don't dequeue parent if it has other entities besides us */
4183 if (cfs_rq->load.weight) {
4185 * Bias pick_next to pick a task from this cfs_rq, as
4186 * p is sleeping when it is within its sched_slice.
4188 if (task_sleep && parent_entity(se))
4189 set_next_buddy(parent_entity(se));
4191 /* avoid re-evaluating load for this entity */
4192 se = parent_entity(se);
4195 flags |= DEQUEUE_SLEEP;
4198 for_each_sched_entity(se) {
4199 cfs_rq = cfs_rq_of(se);
4200 cfs_rq->h_nr_running--;
4202 if (cfs_rq_throttled(cfs_rq))
4205 update_load_avg(se, 1);
4206 update_cfs_shares(cfs_rq);
4210 sub_nr_running(rq, 1);
4218 * per rq 'load' arrray crap; XXX kill this.
4222 * The exact cpuload at various idx values, calculated at every tick would be
4223 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4225 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4226 * on nth tick when cpu may be busy, then we have:
4227 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4228 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4230 * decay_load_missed() below does efficient calculation of
4231 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4232 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4234 * The calculation is approximated on a 128 point scale.
4235 * degrade_zero_ticks is the number of ticks after which load at any
4236 * particular idx is approximated to be zero.
4237 * degrade_factor is a precomputed table, a row for each load idx.
4238 * Each column corresponds to degradation factor for a power of two ticks,
4239 * based on 128 point scale.
4241 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4242 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4244 * With this power of 2 load factors, we can degrade the load n times
4245 * by looking at 1 bits in n and doing as many mult/shift instead of
4246 * n mult/shifts needed by the exact degradation.
4248 #define DEGRADE_SHIFT 7
4249 static const unsigned char
4250 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4251 static const unsigned char
4252 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4253 {0, 0, 0, 0, 0, 0, 0, 0},
4254 {64, 32, 8, 0, 0, 0, 0, 0},
4255 {96, 72, 40, 12, 1, 0, 0},
4256 {112, 98, 75, 43, 15, 1, 0},
4257 {120, 112, 98, 76, 45, 16, 2} };
4260 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4261 * would be when CPU is idle and so we just decay the old load without
4262 * adding any new load.
4264 static unsigned long
4265 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4269 if (!missed_updates)
4272 if (missed_updates >= degrade_zero_ticks[idx])
4276 return load >> missed_updates;
4278 while (missed_updates) {
4279 if (missed_updates % 2)
4280 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4282 missed_updates >>= 1;
4289 * Update rq->cpu_load[] statistics. This function is usually called every
4290 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4291 * every tick. We fix it up based on jiffies.
4293 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4294 unsigned long pending_updates)
4298 this_rq->nr_load_updates++;
4300 /* Update our load: */
4301 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4302 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4303 unsigned long old_load, new_load;
4305 /* scale is effectively 1 << i now, and >> i divides by scale */
4307 old_load = this_rq->cpu_load[i];
4308 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4309 new_load = this_load;
4311 * Round up the averaging division if load is increasing. This
4312 * prevents us from getting stuck on 9 if the load is 10, for
4315 if (new_load > old_load)
4316 new_load += scale - 1;
4318 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4321 sched_avg_update(this_rq);
4324 /* Used instead of source_load when we know the type == 0 */
4325 static unsigned long weighted_cpuload(const int cpu)
4327 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4330 #ifdef CONFIG_NO_HZ_COMMON
4332 * There is no sane way to deal with nohz on smp when using jiffies because the
4333 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4334 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4336 * Therefore we cannot use the delta approach from the regular tick since that
4337 * would seriously skew the load calculation. However we'll make do for those
4338 * updates happening while idle (nohz_idle_balance) or coming out of idle
4339 * (tick_nohz_idle_exit).
4341 * This means we might still be one tick off for nohz periods.
4345 * Called from nohz_idle_balance() to update the load ratings before doing the
4348 static void update_idle_cpu_load(struct rq *this_rq)
4350 unsigned long curr_jiffies = READ_ONCE(jiffies);
4351 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4352 unsigned long pending_updates;
4355 * bail if there's load or we're actually up-to-date.
4357 if (load || curr_jiffies == this_rq->last_load_update_tick)
4360 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4361 this_rq->last_load_update_tick = curr_jiffies;
4363 __update_cpu_load(this_rq, load, pending_updates);
4367 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4369 void update_cpu_load_nohz(void)
4371 struct rq *this_rq = this_rq();
4372 unsigned long curr_jiffies = READ_ONCE(jiffies);
4373 unsigned long pending_updates;
4375 if (curr_jiffies == this_rq->last_load_update_tick)
4378 raw_spin_lock(&this_rq->lock);
4379 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4380 if (pending_updates) {
4381 this_rq->last_load_update_tick = curr_jiffies;
4383 * We were idle, this means load 0, the current load might be
4384 * !0 due to remote wakeups and the sort.
4386 __update_cpu_load(this_rq, 0, pending_updates);
4388 raw_spin_unlock(&this_rq->lock);
4390 #endif /* CONFIG_NO_HZ */
4393 * Called from scheduler_tick()
4395 void update_cpu_load_active(struct rq *this_rq)
4397 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4399 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4401 this_rq->last_load_update_tick = jiffies;
4402 __update_cpu_load(this_rq, load, 1);
4406 * Return a low guess at the load of a migration-source cpu weighted
4407 * according to the scheduling class and "nice" value.
4409 * We want to under-estimate the load of migration sources, to
4410 * balance conservatively.
4412 static unsigned long source_load(int cpu, int type)
4414 struct rq *rq = cpu_rq(cpu);
4415 unsigned long total = weighted_cpuload(cpu);
4417 if (type == 0 || !sched_feat(LB_BIAS))
4420 return min(rq->cpu_load[type-1], total);
4424 * Return a high guess at the load of a migration-target cpu weighted
4425 * according to the scheduling class and "nice" value.
4427 static unsigned long target_load(int cpu, int type)
4429 struct rq *rq = cpu_rq(cpu);
4430 unsigned long total = weighted_cpuload(cpu);
4432 if (type == 0 || !sched_feat(LB_BIAS))
4435 return max(rq->cpu_load[type-1], total);
4438 static unsigned long capacity_of(int cpu)
4440 return cpu_rq(cpu)->cpu_capacity;
4443 static unsigned long capacity_orig_of(int cpu)
4445 return cpu_rq(cpu)->cpu_capacity_orig;
4448 static unsigned long cpu_avg_load_per_task(int cpu)
4450 struct rq *rq = cpu_rq(cpu);
4451 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4452 unsigned long load_avg = weighted_cpuload(cpu);
4455 return load_avg / nr_running;
4460 static void record_wakee(struct task_struct *p)
4463 * Rough decay (wiping) for cost saving, don't worry
4464 * about the boundary, really active task won't care
4467 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4468 current->wakee_flips >>= 1;
4469 current->wakee_flip_decay_ts = jiffies;
4472 if (current->last_wakee != p) {
4473 current->last_wakee = p;
4474 current->wakee_flips++;
4478 static void task_waking_fair(struct task_struct *p)
4480 struct sched_entity *se = &p->se;
4481 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4484 #ifndef CONFIG_64BIT
4485 u64 min_vruntime_copy;
4488 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4490 min_vruntime = cfs_rq->min_vruntime;
4491 } while (min_vruntime != min_vruntime_copy);
4493 min_vruntime = cfs_rq->min_vruntime;
4496 se->vruntime -= min_vruntime;
4500 #ifdef CONFIG_FAIR_GROUP_SCHED
4502 * effective_load() calculates the load change as seen from the root_task_group
4504 * Adding load to a group doesn't make a group heavier, but can cause movement
4505 * of group shares between cpus. Assuming the shares were perfectly aligned one
4506 * can calculate the shift in shares.
4508 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4509 * on this @cpu and results in a total addition (subtraction) of @wg to the
4510 * total group weight.
4512 * Given a runqueue weight distribution (rw_i) we can compute a shares
4513 * distribution (s_i) using:
4515 * s_i = rw_i / \Sum rw_j (1)
4517 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4518 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4519 * shares distribution (s_i):
4521 * rw_i = { 2, 4, 1, 0 }
4522 * s_i = { 2/7, 4/7, 1/7, 0 }
4524 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4525 * task used to run on and the CPU the waker is running on), we need to
4526 * compute the effect of waking a task on either CPU and, in case of a sync
4527 * wakeup, compute the effect of the current task going to sleep.
4529 * So for a change of @wl to the local @cpu with an overall group weight change
4530 * of @wl we can compute the new shares distribution (s'_i) using:
4532 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4534 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4535 * differences in waking a task to CPU 0. The additional task changes the
4536 * weight and shares distributions like:
4538 * rw'_i = { 3, 4, 1, 0 }
4539 * s'_i = { 3/8, 4/8, 1/8, 0 }
4541 * We can then compute the difference in effective weight by using:
4543 * dw_i = S * (s'_i - s_i) (3)
4545 * Where 'S' is the group weight as seen by its parent.
4547 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4548 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4549 * 4/7) times the weight of the group.
4551 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4553 struct sched_entity *se = tg->se[cpu];
4555 if (!tg->parent) /* the trivial, non-cgroup case */
4558 for_each_sched_entity(se) {
4564 * W = @wg + \Sum rw_j
4566 W = wg + calc_tg_weight(tg, se->my_q);
4571 w = cfs_rq_load_avg(se->my_q) + wl;
4574 * wl = S * s'_i; see (2)
4577 wl = (w * (long)tg->shares) / W;
4582 * Per the above, wl is the new se->load.weight value; since
4583 * those are clipped to [MIN_SHARES, ...) do so now. See
4584 * calc_cfs_shares().
4586 if (wl < MIN_SHARES)
4590 * wl = dw_i = S * (s'_i - s_i); see (3)
4592 wl -= se->avg.load_avg;
4595 * Recursively apply this logic to all parent groups to compute
4596 * the final effective load change on the root group. Since
4597 * only the @tg group gets extra weight, all parent groups can
4598 * only redistribute existing shares. @wl is the shift in shares
4599 * resulting from this level per the above.
4608 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4616 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4617 * A waker of many should wake a different task than the one last awakened
4618 * at a frequency roughly N times higher than one of its wakees. In order
4619 * to determine whether we should let the load spread vs consolodating to
4620 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4621 * partner, and a factor of lls_size higher frequency in the other. With
4622 * both conditions met, we can be relatively sure that the relationship is
4623 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4624 * being client/server, worker/dispatcher, interrupt source or whatever is
4625 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4627 static int wake_wide(struct task_struct *p)
4629 unsigned int master = current->wakee_flips;
4630 unsigned int slave = p->wakee_flips;
4631 int factor = this_cpu_read(sd_llc_size);
4634 swap(master, slave);
4635 if (slave < factor || master < slave * factor)
4640 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4642 s64 this_load, load;
4643 s64 this_eff_load, prev_eff_load;
4644 int idx, this_cpu, prev_cpu;
4645 struct task_group *tg;
4646 unsigned long weight;
4650 this_cpu = smp_processor_id();
4651 prev_cpu = task_cpu(p);
4652 load = source_load(prev_cpu, idx);
4653 this_load = target_load(this_cpu, idx);
4656 * If sync wakeup then subtract the (maximum possible)
4657 * effect of the currently running task from the load
4658 * of the current CPU:
4661 tg = task_group(current);
4662 weight = current->se.avg.load_avg;
4664 this_load += effective_load(tg, this_cpu, -weight, -weight);
4665 load += effective_load(tg, prev_cpu, 0, -weight);
4669 weight = p->se.avg.load_avg;
4672 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4673 * due to the sync cause above having dropped this_load to 0, we'll
4674 * always have an imbalance, but there's really nothing you can do
4675 * about that, so that's good too.
4677 * Otherwise check if either cpus are near enough in load to allow this
4678 * task to be woken on this_cpu.
4680 this_eff_load = 100;
4681 this_eff_load *= capacity_of(prev_cpu);
4683 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4684 prev_eff_load *= capacity_of(this_cpu);
4686 if (this_load > 0) {
4687 this_eff_load *= this_load +
4688 effective_load(tg, this_cpu, weight, weight);
4690 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4693 balanced = this_eff_load <= prev_eff_load;
4695 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4700 schedstat_inc(sd, ttwu_move_affine);
4701 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4707 * find_idlest_group finds and returns the least busy CPU group within the
4710 static struct sched_group *
4711 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4712 int this_cpu, int sd_flag)
4714 struct sched_group *idlest = NULL, *group = sd->groups;
4715 unsigned long min_load = ULONG_MAX, this_load = 0;
4716 int load_idx = sd->forkexec_idx;
4717 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4719 if (sd_flag & SD_BALANCE_WAKE)
4720 load_idx = sd->wake_idx;
4723 unsigned long load, avg_load;
4727 /* Skip over this group if it has no CPUs allowed */
4728 if (!cpumask_intersects(sched_group_cpus(group),
4729 tsk_cpus_allowed(p)))
4732 local_group = cpumask_test_cpu(this_cpu,
4733 sched_group_cpus(group));
4735 /* Tally up the load of all CPUs in the group */
4738 for_each_cpu(i, sched_group_cpus(group)) {
4739 /* Bias balancing toward cpus of our domain */
4741 load = source_load(i, load_idx);
4743 load = target_load(i, load_idx);
4748 /* Adjust by relative CPU capacity of the group */
4749 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4752 this_load = avg_load;
4753 } else if (avg_load < min_load) {
4754 min_load = avg_load;
4757 } while (group = group->next, group != sd->groups);
4759 if (!idlest || 100*this_load < imbalance*min_load)
4765 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4768 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4770 unsigned long load, min_load = ULONG_MAX;
4771 unsigned int min_exit_latency = UINT_MAX;
4772 u64 latest_idle_timestamp = 0;
4773 int least_loaded_cpu = this_cpu;
4774 int shallowest_idle_cpu = -1;
4777 /* Traverse only the allowed CPUs */
4778 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4780 struct rq *rq = cpu_rq(i);
4781 struct cpuidle_state *idle = idle_get_state(rq);
4782 if (idle && idle->exit_latency < min_exit_latency) {
4784 * We give priority to a CPU whose idle state
4785 * has the smallest exit latency irrespective
4786 * of any idle timestamp.
4788 min_exit_latency = idle->exit_latency;
4789 latest_idle_timestamp = rq->idle_stamp;
4790 shallowest_idle_cpu = i;
4791 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4792 rq->idle_stamp > latest_idle_timestamp) {
4794 * If equal or no active idle state, then
4795 * the most recently idled CPU might have
4798 latest_idle_timestamp = rq->idle_stamp;
4799 shallowest_idle_cpu = i;
4801 } else if (shallowest_idle_cpu == -1) {
4802 load = weighted_cpuload(i);
4803 if (load < min_load || (load == min_load && i == this_cpu)) {
4805 least_loaded_cpu = i;
4810 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4814 * Try and locate an idle CPU in the sched_domain.
4816 static int select_idle_sibling(struct task_struct *p, int target)
4818 struct sched_domain *sd;
4819 struct sched_group *sg;
4820 int i = task_cpu(p);
4822 if (idle_cpu(target))
4826 * If the prevous cpu is cache affine and idle, don't be stupid.
4828 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4832 * Otherwise, iterate the domains and find an elegible idle cpu.
4834 sd = rcu_dereference(per_cpu(sd_llc, target));
4835 for_each_lower_domain(sd) {
4838 if (!cpumask_intersects(sched_group_cpus(sg),
4839 tsk_cpus_allowed(p)))
4842 for_each_cpu(i, sched_group_cpus(sg)) {
4843 if (i == target || !idle_cpu(i))
4847 target = cpumask_first_and(sched_group_cpus(sg),
4848 tsk_cpus_allowed(p));
4852 } while (sg != sd->groups);
4858 * get_cpu_usage returns the amount of capacity of a CPU that is used by CFS
4859 * tasks. The unit of the return value must be the one of capacity so we can
4860 * compare the usage with the capacity of the CPU that is available for CFS
4861 * task (ie cpu_capacity).
4862 * cfs.avg.util_avg is the sum of running time of runnable tasks on a
4863 * CPU. It represents the amount of utilization of a CPU in the range
4864 * [0..SCHED_LOAD_SCALE]. The usage of a CPU can't be higher than the full
4865 * capacity of the CPU because it's about the running time on this CPU.
4866 * Nevertheless, cfs.avg.util_avg can be higher than SCHED_LOAD_SCALE
4867 * because of unfortunate rounding in util_avg or just
4868 * after migrating tasks until the average stabilizes with the new running
4869 * time. So we need to check that the usage stays into the range
4870 * [0..cpu_capacity_orig] and cap if necessary.
4871 * Without capping the usage, a group could be seen as overloaded (CPU0 usage
4872 * at 121% + CPU1 usage at 80%) whereas CPU1 has 20% of available capacity
4874 static int get_cpu_usage(int cpu)
4876 unsigned long usage = cpu_rq(cpu)->cfs.avg.util_avg;
4877 unsigned long capacity = capacity_orig_of(cpu);
4879 if (usage >= SCHED_LOAD_SCALE)
4882 return (usage * capacity) >> SCHED_LOAD_SHIFT;
4886 * select_task_rq_fair: Select target runqueue for the waking task in domains
4887 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4888 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4890 * Balances load by selecting the idlest cpu in the idlest group, or under
4891 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4893 * Returns the target cpu number.
4895 * preempt must be disabled.
4898 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4900 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4901 int cpu = smp_processor_id();
4902 int new_cpu = prev_cpu;
4903 int want_affine = 0;
4904 int sync = wake_flags & WF_SYNC;
4906 if (sd_flag & SD_BALANCE_WAKE)
4907 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4910 for_each_domain(cpu, tmp) {
4911 if (!(tmp->flags & SD_LOAD_BALANCE))
4915 * If both cpu and prev_cpu are part of this domain,
4916 * cpu is a valid SD_WAKE_AFFINE target.
4918 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4919 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4924 if (tmp->flags & sd_flag)
4926 else if (!want_affine)
4931 sd = NULL; /* Prefer wake_affine over balance flags */
4932 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4937 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
4938 new_cpu = select_idle_sibling(p, new_cpu);
4941 struct sched_group *group;
4944 if (!(sd->flags & sd_flag)) {
4949 group = find_idlest_group(sd, p, cpu, sd_flag);
4955 new_cpu = find_idlest_cpu(group, p, cpu);
4956 if (new_cpu == -1 || new_cpu == cpu) {
4957 /* Now try balancing at a lower domain level of cpu */
4962 /* Now try balancing at a lower domain level of new_cpu */
4964 weight = sd->span_weight;
4966 for_each_domain(cpu, tmp) {
4967 if (weight <= tmp->span_weight)
4969 if (tmp->flags & sd_flag)
4972 /* while loop will break here if sd == NULL */
4980 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4981 * cfs_rq_of(p) references at time of call are still valid and identify the
4982 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4983 * other assumptions, including the state of rq->lock, should be made.
4985 static void migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4988 * We are supposed to update the task to "current" time, then its up to date
4989 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
4990 * what current time is, so simply throw away the out-of-date time. This
4991 * will result in the wakee task is less decayed, but giving the wakee more
4992 * load sounds not bad.
4994 remove_entity_load_avg(&p->se);
4996 /* Tell new CPU we are migrated */
4997 p->se.avg.last_update_time = 0;
4999 /* We have migrated, no longer consider this task hot */
5000 p->se.exec_start = 0;
5003 static void task_dead_fair(struct task_struct *p)
5005 remove_entity_load_avg(&p->se);
5007 #endif /* CONFIG_SMP */
5009 static unsigned long
5010 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5012 unsigned long gran = sysctl_sched_wakeup_granularity;
5015 * Since its curr running now, convert the gran from real-time
5016 * to virtual-time in his units.
5018 * By using 'se' instead of 'curr' we penalize light tasks, so
5019 * they get preempted easier. That is, if 'se' < 'curr' then
5020 * the resulting gran will be larger, therefore penalizing the
5021 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5022 * be smaller, again penalizing the lighter task.
5024 * This is especially important for buddies when the leftmost
5025 * task is higher priority than the buddy.
5027 return calc_delta_fair(gran, se);
5031 * Should 'se' preempt 'curr'.
5045 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5047 s64 gran, vdiff = curr->vruntime - se->vruntime;
5052 gran = wakeup_gran(curr, se);
5059 static void set_last_buddy(struct sched_entity *se)
5061 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5064 for_each_sched_entity(se)
5065 cfs_rq_of(se)->last = se;
5068 static void set_next_buddy(struct sched_entity *se)
5070 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5073 for_each_sched_entity(se)
5074 cfs_rq_of(se)->next = se;
5077 static void set_skip_buddy(struct sched_entity *se)
5079 for_each_sched_entity(se)
5080 cfs_rq_of(se)->skip = se;
5084 * Preempt the current task with a newly woken task if needed:
5086 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5088 struct task_struct *curr = rq->curr;
5089 struct sched_entity *se = &curr->se, *pse = &p->se;
5090 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5091 int scale = cfs_rq->nr_running >= sched_nr_latency;
5092 int next_buddy_marked = 0;
5094 if (unlikely(se == pse))
5098 * This is possible from callers such as attach_tasks(), in which we
5099 * unconditionally check_prempt_curr() after an enqueue (which may have
5100 * lead to a throttle). This both saves work and prevents false
5101 * next-buddy nomination below.
5103 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5106 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5107 set_next_buddy(pse);
5108 next_buddy_marked = 1;
5112 * We can come here with TIF_NEED_RESCHED already set from new task
5115 * Note: this also catches the edge-case of curr being in a throttled
5116 * group (e.g. via set_curr_task), since update_curr() (in the
5117 * enqueue of curr) will have resulted in resched being set. This
5118 * prevents us from potentially nominating it as a false LAST_BUDDY
5121 if (test_tsk_need_resched(curr))
5124 /* Idle tasks are by definition preempted by non-idle tasks. */
5125 if (unlikely(curr->policy == SCHED_IDLE) &&
5126 likely(p->policy != SCHED_IDLE))
5130 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5131 * is driven by the tick):
5133 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5136 find_matching_se(&se, &pse);
5137 update_curr(cfs_rq_of(se));
5139 if (wakeup_preempt_entity(se, pse) == 1) {
5141 * Bias pick_next to pick the sched entity that is
5142 * triggering this preemption.
5144 if (!next_buddy_marked)
5145 set_next_buddy(pse);
5154 * Only set the backward buddy when the current task is still
5155 * on the rq. This can happen when a wakeup gets interleaved
5156 * with schedule on the ->pre_schedule() or idle_balance()
5157 * point, either of which can * drop the rq lock.
5159 * Also, during early boot the idle thread is in the fair class,
5160 * for obvious reasons its a bad idea to schedule back to it.
5162 if (unlikely(!se->on_rq || curr == rq->idle))
5165 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5169 static struct task_struct *
5170 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5172 struct cfs_rq *cfs_rq = &rq->cfs;
5173 struct sched_entity *se;
5174 struct task_struct *p;
5178 #ifdef CONFIG_FAIR_GROUP_SCHED
5179 if (!cfs_rq->nr_running)
5182 if (prev->sched_class != &fair_sched_class)
5186 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5187 * likely that a next task is from the same cgroup as the current.
5189 * Therefore attempt to avoid putting and setting the entire cgroup
5190 * hierarchy, only change the part that actually changes.
5194 struct sched_entity *curr = cfs_rq->curr;
5197 * Since we got here without doing put_prev_entity() we also
5198 * have to consider cfs_rq->curr. If it is still a runnable
5199 * entity, update_curr() will update its vruntime, otherwise
5200 * forget we've ever seen it.
5204 update_curr(cfs_rq);
5209 * This call to check_cfs_rq_runtime() will do the
5210 * throttle and dequeue its entity in the parent(s).
5211 * Therefore the 'simple' nr_running test will indeed
5214 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5218 se = pick_next_entity(cfs_rq, curr);
5219 cfs_rq = group_cfs_rq(se);
5225 * Since we haven't yet done put_prev_entity and if the selected task
5226 * is a different task than we started out with, try and touch the
5227 * least amount of cfs_rqs.
5230 struct sched_entity *pse = &prev->se;
5232 while (!(cfs_rq = is_same_group(se, pse))) {
5233 int se_depth = se->depth;
5234 int pse_depth = pse->depth;
5236 if (se_depth <= pse_depth) {
5237 put_prev_entity(cfs_rq_of(pse), pse);
5238 pse = parent_entity(pse);
5240 if (se_depth >= pse_depth) {
5241 set_next_entity(cfs_rq_of(se), se);
5242 se = parent_entity(se);
5246 put_prev_entity(cfs_rq, pse);
5247 set_next_entity(cfs_rq, se);
5250 if (hrtick_enabled(rq))
5251 hrtick_start_fair(rq, p);
5258 if (!cfs_rq->nr_running)
5261 put_prev_task(rq, prev);
5264 se = pick_next_entity(cfs_rq, NULL);
5265 set_next_entity(cfs_rq, se);
5266 cfs_rq = group_cfs_rq(se);
5271 if (hrtick_enabled(rq))
5272 hrtick_start_fair(rq, p);
5278 * This is OK, because current is on_cpu, which avoids it being picked
5279 * for load-balance and preemption/IRQs are still disabled avoiding
5280 * further scheduler activity on it and we're being very careful to
5281 * re-start the picking loop.
5283 lockdep_unpin_lock(&rq->lock);
5284 new_tasks = idle_balance(rq);
5285 lockdep_pin_lock(&rq->lock);
5287 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5288 * possible for any higher priority task to appear. In that case we
5289 * must re-start the pick_next_entity() loop.
5301 * Account for a descheduled task:
5303 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5305 struct sched_entity *se = &prev->se;
5306 struct cfs_rq *cfs_rq;
5308 for_each_sched_entity(se) {
5309 cfs_rq = cfs_rq_of(se);
5310 put_prev_entity(cfs_rq, se);
5315 * sched_yield() is very simple
5317 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5319 static void yield_task_fair(struct rq *rq)
5321 struct task_struct *curr = rq->curr;
5322 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5323 struct sched_entity *se = &curr->se;
5326 * Are we the only task in the tree?
5328 if (unlikely(rq->nr_running == 1))
5331 clear_buddies(cfs_rq, se);
5333 if (curr->policy != SCHED_BATCH) {
5334 update_rq_clock(rq);
5336 * Update run-time statistics of the 'current'.
5338 update_curr(cfs_rq);
5340 * Tell update_rq_clock() that we've just updated,
5341 * so we don't do microscopic update in schedule()
5342 * and double the fastpath cost.
5344 rq_clock_skip_update(rq, true);
5350 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5352 struct sched_entity *se = &p->se;
5354 /* throttled hierarchies are not runnable */
5355 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5358 /* Tell the scheduler that we'd really like pse to run next. */
5361 yield_task_fair(rq);
5367 /**************************************************
5368 * Fair scheduling class load-balancing methods.
5372 * The purpose of load-balancing is to achieve the same basic fairness the
5373 * per-cpu scheduler provides, namely provide a proportional amount of compute
5374 * time to each task. This is expressed in the following equation:
5376 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5378 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5379 * W_i,0 is defined as:
5381 * W_i,0 = \Sum_j w_i,j (2)
5383 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5384 * is derived from the nice value as per prio_to_weight[].
5386 * The weight average is an exponential decay average of the instantaneous
5389 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5391 * C_i is the compute capacity of cpu i, typically it is the
5392 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5393 * can also include other factors [XXX].
5395 * To achieve this balance we define a measure of imbalance which follows
5396 * directly from (1):
5398 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5400 * We them move tasks around to minimize the imbalance. In the continuous
5401 * function space it is obvious this converges, in the discrete case we get
5402 * a few fun cases generally called infeasible weight scenarios.
5405 * - infeasible weights;
5406 * - local vs global optima in the discrete case. ]
5411 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5412 * for all i,j solution, we create a tree of cpus that follows the hardware
5413 * topology where each level pairs two lower groups (or better). This results
5414 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5415 * tree to only the first of the previous level and we decrease the frequency
5416 * of load-balance at each level inv. proportional to the number of cpus in
5422 * \Sum { --- * --- * 2^i } = O(n) (5)
5424 * `- size of each group
5425 * | | `- number of cpus doing load-balance
5427 * `- sum over all levels
5429 * Coupled with a limit on how many tasks we can migrate every balance pass,
5430 * this makes (5) the runtime complexity of the balancer.
5432 * An important property here is that each CPU is still (indirectly) connected
5433 * to every other cpu in at most O(log n) steps:
5435 * The adjacency matrix of the resulting graph is given by:
5438 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5441 * And you'll find that:
5443 * A^(log_2 n)_i,j != 0 for all i,j (7)
5445 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5446 * The task movement gives a factor of O(m), giving a convergence complexity
5449 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5454 * In order to avoid CPUs going idle while there's still work to do, new idle
5455 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5456 * tree itself instead of relying on other CPUs to bring it work.
5458 * This adds some complexity to both (5) and (8) but it reduces the total idle
5466 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5469 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5474 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5476 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5478 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5481 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5482 * rewrite all of this once again.]
5485 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5487 enum fbq_type { regular, remote, all };
5489 #define LBF_ALL_PINNED 0x01
5490 #define LBF_NEED_BREAK 0x02
5491 #define LBF_DST_PINNED 0x04
5492 #define LBF_SOME_PINNED 0x08
5495 struct sched_domain *sd;
5503 struct cpumask *dst_grpmask;
5505 enum cpu_idle_type idle;
5507 /* The set of CPUs under consideration for load-balancing */
5508 struct cpumask *cpus;
5513 unsigned int loop_break;
5514 unsigned int loop_max;
5516 enum fbq_type fbq_type;
5517 struct list_head tasks;
5521 * Is this task likely cache-hot:
5523 static int task_hot(struct task_struct *p, struct lb_env *env)
5527 lockdep_assert_held(&env->src_rq->lock);
5529 if (p->sched_class != &fair_sched_class)
5532 if (unlikely(p->policy == SCHED_IDLE))
5536 * Buddy candidates are cache hot:
5538 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5539 (&p->se == cfs_rq_of(&p->se)->next ||
5540 &p->se == cfs_rq_of(&p->se)->last))
5543 if (sysctl_sched_migration_cost == -1)
5545 if (sysctl_sched_migration_cost == 0)
5548 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5550 return delta < (s64)sysctl_sched_migration_cost;
5553 #ifdef CONFIG_NUMA_BALANCING
5555 * Returns 1, if task migration degrades locality
5556 * Returns 0, if task migration improves locality i.e migration preferred.
5557 * Returns -1, if task migration is not affected by locality.
5559 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5561 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5562 unsigned long src_faults, dst_faults;
5563 int src_nid, dst_nid;
5565 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5568 if (!sched_feat(NUMA))
5571 src_nid = cpu_to_node(env->src_cpu);
5572 dst_nid = cpu_to_node(env->dst_cpu);
5574 if (src_nid == dst_nid)
5577 /* Migrating away from the preferred node is always bad. */
5578 if (src_nid == p->numa_preferred_nid) {
5579 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5585 /* Encourage migration to the preferred node. */
5586 if (dst_nid == p->numa_preferred_nid)
5590 src_faults = group_faults(p, src_nid);
5591 dst_faults = group_faults(p, dst_nid);
5593 src_faults = task_faults(p, src_nid);
5594 dst_faults = task_faults(p, dst_nid);
5597 return dst_faults < src_faults;
5601 static inline int migrate_degrades_locality(struct task_struct *p,
5609 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5612 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5616 lockdep_assert_held(&env->src_rq->lock);
5619 * We do not migrate tasks that are:
5620 * 1) throttled_lb_pair, or
5621 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5622 * 3) running (obviously), or
5623 * 4) are cache-hot on their current CPU.
5625 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5628 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5631 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5633 env->flags |= LBF_SOME_PINNED;
5636 * Remember if this task can be migrated to any other cpu in
5637 * our sched_group. We may want to revisit it if we couldn't
5638 * meet load balance goals by pulling other tasks on src_cpu.
5640 * Also avoid computing new_dst_cpu if we have already computed
5641 * one in current iteration.
5643 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5646 /* Prevent to re-select dst_cpu via env's cpus */
5647 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5648 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5649 env->flags |= LBF_DST_PINNED;
5650 env->new_dst_cpu = cpu;
5658 /* Record that we found atleast one task that could run on dst_cpu */
5659 env->flags &= ~LBF_ALL_PINNED;
5661 if (task_running(env->src_rq, p)) {
5662 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5667 * Aggressive migration if:
5668 * 1) destination numa is preferred
5669 * 2) task is cache cold, or
5670 * 3) too many balance attempts have failed.
5672 tsk_cache_hot = migrate_degrades_locality(p, env);
5673 if (tsk_cache_hot == -1)
5674 tsk_cache_hot = task_hot(p, env);
5676 if (tsk_cache_hot <= 0 ||
5677 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5678 if (tsk_cache_hot == 1) {
5679 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5680 schedstat_inc(p, se.statistics.nr_forced_migrations);
5685 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5690 * detach_task() -- detach the task for the migration specified in env
5692 static void detach_task(struct task_struct *p, struct lb_env *env)
5694 lockdep_assert_held(&env->src_rq->lock);
5696 deactivate_task(env->src_rq, p, 0);
5697 p->on_rq = TASK_ON_RQ_MIGRATING;
5698 set_task_cpu(p, env->dst_cpu);
5702 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5703 * part of active balancing operations within "domain".
5705 * Returns a task if successful and NULL otherwise.
5707 static struct task_struct *detach_one_task(struct lb_env *env)
5709 struct task_struct *p, *n;
5711 lockdep_assert_held(&env->src_rq->lock);
5713 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5714 if (!can_migrate_task(p, env))
5717 detach_task(p, env);
5720 * Right now, this is only the second place where
5721 * lb_gained[env->idle] is updated (other is detach_tasks)
5722 * so we can safely collect stats here rather than
5723 * inside detach_tasks().
5725 schedstat_inc(env->sd, lb_gained[env->idle]);
5731 static const unsigned int sched_nr_migrate_break = 32;
5734 * detach_tasks() -- tries to detach up to imbalance weighted load from
5735 * busiest_rq, as part of a balancing operation within domain "sd".
5737 * Returns number of detached tasks if successful and 0 otherwise.
5739 static int detach_tasks(struct lb_env *env)
5741 struct list_head *tasks = &env->src_rq->cfs_tasks;
5742 struct task_struct *p;
5746 lockdep_assert_held(&env->src_rq->lock);
5748 if (env->imbalance <= 0)
5751 while (!list_empty(tasks)) {
5753 * We don't want to steal all, otherwise we may be treated likewise,
5754 * which could at worst lead to a livelock crash.
5756 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5759 p = list_first_entry(tasks, struct task_struct, se.group_node);
5762 /* We've more or less seen every task there is, call it quits */
5763 if (env->loop > env->loop_max)
5766 /* take a breather every nr_migrate tasks */
5767 if (env->loop > env->loop_break) {
5768 env->loop_break += sched_nr_migrate_break;
5769 env->flags |= LBF_NEED_BREAK;
5773 if (!can_migrate_task(p, env))
5776 load = task_h_load(p);
5778 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5781 if ((load / 2) > env->imbalance)
5784 detach_task(p, env);
5785 list_add(&p->se.group_node, &env->tasks);
5788 env->imbalance -= load;
5790 #ifdef CONFIG_PREEMPT
5792 * NEWIDLE balancing is a source of latency, so preemptible
5793 * kernels will stop after the first task is detached to minimize
5794 * the critical section.
5796 if (env->idle == CPU_NEWLY_IDLE)
5801 * We only want to steal up to the prescribed amount of
5804 if (env->imbalance <= 0)
5809 list_move_tail(&p->se.group_node, tasks);
5813 * Right now, this is one of only two places we collect this stat
5814 * so we can safely collect detach_one_task() stats here rather
5815 * than inside detach_one_task().
5817 schedstat_add(env->sd, lb_gained[env->idle], detached);
5823 * attach_task() -- attach the task detached by detach_task() to its new rq.
5825 static void attach_task(struct rq *rq, struct task_struct *p)
5827 lockdep_assert_held(&rq->lock);
5829 BUG_ON(task_rq(p) != rq);
5830 p->on_rq = TASK_ON_RQ_QUEUED;
5831 activate_task(rq, p, 0);
5832 check_preempt_curr(rq, p, 0);
5836 * attach_one_task() -- attaches the task returned from detach_one_task() to
5839 static void attach_one_task(struct rq *rq, struct task_struct *p)
5841 raw_spin_lock(&rq->lock);
5843 raw_spin_unlock(&rq->lock);
5847 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5850 static void attach_tasks(struct lb_env *env)
5852 struct list_head *tasks = &env->tasks;
5853 struct task_struct *p;
5855 raw_spin_lock(&env->dst_rq->lock);
5857 while (!list_empty(tasks)) {
5858 p = list_first_entry(tasks, struct task_struct, se.group_node);
5859 list_del_init(&p->se.group_node);
5861 attach_task(env->dst_rq, p);
5864 raw_spin_unlock(&env->dst_rq->lock);
5867 #ifdef CONFIG_FAIR_GROUP_SCHED
5868 static void update_blocked_averages(int cpu)
5870 struct rq *rq = cpu_rq(cpu);
5871 struct cfs_rq *cfs_rq;
5872 unsigned long flags;
5874 raw_spin_lock_irqsave(&rq->lock, flags);
5875 update_rq_clock(rq);
5878 * Iterates the task_group tree in a bottom up fashion, see
5879 * list_add_leaf_cfs_rq() for details.
5881 for_each_leaf_cfs_rq(rq, cfs_rq) {
5882 /* throttled entities do not contribute to load */
5883 if (throttled_hierarchy(cfs_rq))
5886 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
5887 update_tg_load_avg(cfs_rq, 0);
5889 raw_spin_unlock_irqrestore(&rq->lock, flags);
5893 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5894 * This needs to be done in a top-down fashion because the load of a child
5895 * group is a fraction of its parents load.
5897 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5899 struct rq *rq = rq_of(cfs_rq);
5900 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5901 unsigned long now = jiffies;
5904 if (cfs_rq->last_h_load_update == now)
5907 cfs_rq->h_load_next = NULL;
5908 for_each_sched_entity(se) {
5909 cfs_rq = cfs_rq_of(se);
5910 cfs_rq->h_load_next = se;
5911 if (cfs_rq->last_h_load_update == now)
5916 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
5917 cfs_rq->last_h_load_update = now;
5920 while ((se = cfs_rq->h_load_next) != NULL) {
5921 load = cfs_rq->h_load;
5922 load = div64_ul(load * se->avg.load_avg,
5923 cfs_rq_load_avg(cfs_rq) + 1);
5924 cfs_rq = group_cfs_rq(se);
5925 cfs_rq->h_load = load;
5926 cfs_rq->last_h_load_update = now;
5930 static unsigned long task_h_load(struct task_struct *p)
5932 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5934 update_cfs_rq_h_load(cfs_rq);
5935 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
5936 cfs_rq_load_avg(cfs_rq) + 1);
5939 static inline void update_blocked_averages(int cpu)
5941 struct rq *rq = cpu_rq(cpu);
5942 struct cfs_rq *cfs_rq = &rq->cfs;
5943 unsigned long flags;
5945 raw_spin_lock_irqsave(&rq->lock, flags);
5946 update_rq_clock(rq);
5947 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
5948 raw_spin_unlock_irqrestore(&rq->lock, flags);
5951 static unsigned long task_h_load(struct task_struct *p)
5953 return p->se.avg.load_avg;
5957 /********** Helpers for find_busiest_group ************************/
5966 * sg_lb_stats - stats of a sched_group required for load_balancing
5968 struct sg_lb_stats {
5969 unsigned long avg_load; /*Avg load across the CPUs of the group */
5970 unsigned long group_load; /* Total load over the CPUs of the group */
5971 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5972 unsigned long load_per_task;
5973 unsigned long group_capacity;
5974 unsigned long group_usage; /* Total usage of the group */
5975 unsigned int sum_nr_running; /* Nr tasks running in the group */
5976 unsigned int idle_cpus;
5977 unsigned int group_weight;
5978 enum group_type group_type;
5979 int group_no_capacity;
5980 #ifdef CONFIG_NUMA_BALANCING
5981 unsigned int nr_numa_running;
5982 unsigned int nr_preferred_running;
5987 * sd_lb_stats - Structure to store the statistics of a sched_domain
5988 * during load balancing.
5990 struct sd_lb_stats {
5991 struct sched_group *busiest; /* Busiest group in this sd */
5992 struct sched_group *local; /* Local group in this sd */
5993 unsigned long total_load; /* Total load of all groups in sd */
5994 unsigned long total_capacity; /* Total capacity of all groups in sd */
5995 unsigned long avg_load; /* Average load across all groups in sd */
5997 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5998 struct sg_lb_stats local_stat; /* Statistics of the local group */
6001 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6004 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6005 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6006 * We must however clear busiest_stat::avg_load because
6007 * update_sd_pick_busiest() reads this before assignment.
6009 *sds = (struct sd_lb_stats){
6013 .total_capacity = 0UL,
6016 .sum_nr_running = 0,
6017 .group_type = group_other,
6023 * get_sd_load_idx - Obtain the load index for a given sched domain.
6024 * @sd: The sched_domain whose load_idx is to be obtained.
6025 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6027 * Return: The load index.
6029 static inline int get_sd_load_idx(struct sched_domain *sd,
6030 enum cpu_idle_type idle)
6036 load_idx = sd->busy_idx;
6039 case CPU_NEWLY_IDLE:
6040 load_idx = sd->newidle_idx;
6043 load_idx = sd->idle_idx;
6050 static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu)
6052 if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1))
6053 return sd->smt_gain / sd->span_weight;
6055 return SCHED_CAPACITY_SCALE;
6058 unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
6060 return default_scale_cpu_capacity(sd, cpu);
6063 static unsigned long scale_rt_capacity(int cpu)
6065 struct rq *rq = cpu_rq(cpu);
6066 u64 total, used, age_stamp, avg;
6070 * Since we're reading these variables without serialization make sure
6071 * we read them once before doing sanity checks on them.
6073 age_stamp = READ_ONCE(rq->age_stamp);
6074 avg = READ_ONCE(rq->rt_avg);
6075 delta = __rq_clock_broken(rq) - age_stamp;
6077 if (unlikely(delta < 0))
6080 total = sched_avg_period() + delta;
6082 used = div_u64(avg, total);
6084 if (likely(used < SCHED_CAPACITY_SCALE))
6085 return SCHED_CAPACITY_SCALE - used;
6090 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6092 unsigned long capacity = SCHED_CAPACITY_SCALE;
6093 struct sched_group *sdg = sd->groups;
6095 if (sched_feat(ARCH_CAPACITY))
6096 capacity *= arch_scale_cpu_capacity(sd, cpu);
6098 capacity *= default_scale_cpu_capacity(sd, cpu);
6100 capacity >>= SCHED_CAPACITY_SHIFT;
6102 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6104 capacity *= scale_rt_capacity(cpu);
6105 capacity >>= SCHED_CAPACITY_SHIFT;
6110 cpu_rq(cpu)->cpu_capacity = capacity;
6111 sdg->sgc->capacity = capacity;
6114 void update_group_capacity(struct sched_domain *sd, int cpu)
6116 struct sched_domain *child = sd->child;
6117 struct sched_group *group, *sdg = sd->groups;
6118 unsigned long capacity;
6119 unsigned long interval;
6121 interval = msecs_to_jiffies(sd->balance_interval);
6122 interval = clamp(interval, 1UL, max_load_balance_interval);
6123 sdg->sgc->next_update = jiffies + interval;
6126 update_cpu_capacity(sd, cpu);
6132 if (child->flags & SD_OVERLAP) {
6134 * SD_OVERLAP domains cannot assume that child groups
6135 * span the current group.
6138 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6139 struct sched_group_capacity *sgc;
6140 struct rq *rq = cpu_rq(cpu);
6143 * build_sched_domains() -> init_sched_groups_capacity()
6144 * gets here before we've attached the domains to the
6147 * Use capacity_of(), which is set irrespective of domains
6148 * in update_cpu_capacity().
6150 * This avoids capacity from being 0 and
6151 * causing divide-by-zero issues on boot.
6153 if (unlikely(!rq->sd)) {
6154 capacity += capacity_of(cpu);
6158 sgc = rq->sd->groups->sgc;
6159 capacity += sgc->capacity;
6163 * !SD_OVERLAP domains can assume that child groups
6164 * span the current group.
6167 group = child->groups;
6169 capacity += group->sgc->capacity;
6170 group = group->next;
6171 } while (group != child->groups);
6174 sdg->sgc->capacity = capacity;
6178 * Check whether the capacity of the rq has been noticeably reduced by side
6179 * activity. The imbalance_pct is used for the threshold.
6180 * Return true is the capacity is reduced
6183 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6185 return ((rq->cpu_capacity * sd->imbalance_pct) <
6186 (rq->cpu_capacity_orig * 100));
6190 * Group imbalance indicates (and tries to solve) the problem where balancing
6191 * groups is inadequate due to tsk_cpus_allowed() constraints.
6193 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6194 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6197 * { 0 1 2 3 } { 4 5 6 7 }
6200 * If we were to balance group-wise we'd place two tasks in the first group and
6201 * two tasks in the second group. Clearly this is undesired as it will overload
6202 * cpu 3 and leave one of the cpus in the second group unused.
6204 * The current solution to this issue is detecting the skew in the first group
6205 * by noticing the lower domain failed to reach balance and had difficulty
6206 * moving tasks due to affinity constraints.
6208 * When this is so detected; this group becomes a candidate for busiest; see
6209 * update_sd_pick_busiest(). And calculate_imbalance() and
6210 * find_busiest_group() avoid some of the usual balance conditions to allow it
6211 * to create an effective group imbalance.
6213 * This is a somewhat tricky proposition since the next run might not find the
6214 * group imbalance and decide the groups need to be balanced again. A most
6215 * subtle and fragile situation.
6218 static inline int sg_imbalanced(struct sched_group *group)
6220 return group->sgc->imbalance;
6224 * group_has_capacity returns true if the group has spare capacity that could
6225 * be used by some tasks.
6226 * We consider that a group has spare capacity if the * number of task is
6227 * smaller than the number of CPUs or if the usage is lower than the available
6228 * capacity for CFS tasks.
6229 * For the latter, we use a threshold to stabilize the state, to take into
6230 * account the variance of the tasks' load and to return true if the available
6231 * capacity in meaningful for the load balancer.
6232 * As an example, an available capacity of 1% can appear but it doesn't make
6233 * any benefit for the load balance.
6236 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6238 if (sgs->sum_nr_running < sgs->group_weight)
6241 if ((sgs->group_capacity * 100) >
6242 (sgs->group_usage * env->sd->imbalance_pct))
6249 * group_is_overloaded returns true if the group has more tasks than it can
6251 * group_is_overloaded is not equals to !group_has_capacity because a group
6252 * with the exact right number of tasks, has no more spare capacity but is not
6253 * overloaded so both group_has_capacity and group_is_overloaded return
6257 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6259 if (sgs->sum_nr_running <= sgs->group_weight)
6262 if ((sgs->group_capacity * 100) <
6263 (sgs->group_usage * env->sd->imbalance_pct))
6269 static enum group_type group_classify(struct lb_env *env,
6270 struct sched_group *group,
6271 struct sg_lb_stats *sgs)
6273 if (sgs->group_no_capacity)
6274 return group_overloaded;
6276 if (sg_imbalanced(group))
6277 return group_imbalanced;
6283 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6284 * @env: The load balancing environment.
6285 * @group: sched_group whose statistics are to be updated.
6286 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6287 * @local_group: Does group contain this_cpu.
6288 * @sgs: variable to hold the statistics for this group.
6289 * @overload: Indicate more than one runnable task for any CPU.
6291 static inline void update_sg_lb_stats(struct lb_env *env,
6292 struct sched_group *group, int load_idx,
6293 int local_group, struct sg_lb_stats *sgs,
6299 memset(sgs, 0, sizeof(*sgs));
6301 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6302 struct rq *rq = cpu_rq(i);
6304 /* Bias balancing toward cpus of our domain */
6306 load = target_load(i, load_idx);
6308 load = source_load(i, load_idx);
6310 sgs->group_load += load;
6311 sgs->group_usage += get_cpu_usage(i);
6312 sgs->sum_nr_running += rq->cfs.h_nr_running;
6314 if (rq->nr_running > 1)
6317 #ifdef CONFIG_NUMA_BALANCING
6318 sgs->nr_numa_running += rq->nr_numa_running;
6319 sgs->nr_preferred_running += rq->nr_preferred_running;
6321 sgs->sum_weighted_load += weighted_cpuload(i);
6326 /* Adjust by relative CPU capacity of the group */
6327 sgs->group_capacity = group->sgc->capacity;
6328 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6330 if (sgs->sum_nr_running)
6331 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6333 sgs->group_weight = group->group_weight;
6335 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6336 sgs->group_type = group_classify(env, group, sgs);
6340 * update_sd_pick_busiest - return 1 on busiest group
6341 * @env: The load balancing environment.
6342 * @sds: sched_domain statistics
6343 * @sg: sched_group candidate to be checked for being the busiest
6344 * @sgs: sched_group statistics
6346 * Determine if @sg is a busier group than the previously selected
6349 * Return: %true if @sg is a busier group than the previously selected
6350 * busiest group. %false otherwise.
6352 static bool update_sd_pick_busiest(struct lb_env *env,
6353 struct sd_lb_stats *sds,
6354 struct sched_group *sg,
6355 struct sg_lb_stats *sgs)
6357 struct sg_lb_stats *busiest = &sds->busiest_stat;
6359 if (sgs->group_type > busiest->group_type)
6362 if (sgs->group_type < busiest->group_type)
6365 if (sgs->avg_load <= busiest->avg_load)
6368 /* This is the busiest node in its class. */
6369 if (!(env->sd->flags & SD_ASYM_PACKING))
6373 * ASYM_PACKING needs to move all the work to the lowest
6374 * numbered CPUs in the group, therefore mark all groups
6375 * higher than ourself as busy.
6377 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6381 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6388 #ifdef CONFIG_NUMA_BALANCING
6389 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6391 if (sgs->sum_nr_running > sgs->nr_numa_running)
6393 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6398 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6400 if (rq->nr_running > rq->nr_numa_running)
6402 if (rq->nr_running > rq->nr_preferred_running)
6407 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6412 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6416 #endif /* CONFIG_NUMA_BALANCING */
6419 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6420 * @env: The load balancing environment.
6421 * @sds: variable to hold the statistics for this sched_domain.
6423 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6425 struct sched_domain *child = env->sd->child;
6426 struct sched_group *sg = env->sd->groups;
6427 struct sg_lb_stats tmp_sgs;
6428 int load_idx, prefer_sibling = 0;
6429 bool overload = false;
6431 if (child && child->flags & SD_PREFER_SIBLING)
6434 load_idx = get_sd_load_idx(env->sd, env->idle);
6437 struct sg_lb_stats *sgs = &tmp_sgs;
6440 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6443 sgs = &sds->local_stat;
6445 if (env->idle != CPU_NEWLY_IDLE ||
6446 time_after_eq(jiffies, sg->sgc->next_update))
6447 update_group_capacity(env->sd, env->dst_cpu);
6450 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6457 * In case the child domain prefers tasks go to siblings
6458 * first, lower the sg capacity so that we'll try
6459 * and move all the excess tasks away. We lower the capacity
6460 * of a group only if the local group has the capacity to fit
6461 * these excess tasks. The extra check prevents the case where
6462 * you always pull from the heaviest group when it is already
6463 * under-utilized (possible with a large weight task outweighs
6464 * the tasks on the system).
6466 if (prefer_sibling && sds->local &&
6467 group_has_capacity(env, &sds->local_stat) &&
6468 (sgs->sum_nr_running > 1)) {
6469 sgs->group_no_capacity = 1;
6470 sgs->group_type = group_overloaded;
6473 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6475 sds->busiest_stat = *sgs;
6479 /* Now, start updating sd_lb_stats */
6480 sds->total_load += sgs->group_load;
6481 sds->total_capacity += sgs->group_capacity;
6484 } while (sg != env->sd->groups);
6486 if (env->sd->flags & SD_NUMA)
6487 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6489 if (!env->sd->parent) {
6490 /* update overload indicator if we are at root domain */
6491 if (env->dst_rq->rd->overload != overload)
6492 env->dst_rq->rd->overload = overload;
6498 * check_asym_packing - Check to see if the group is packed into the
6501 * This is primarily intended to used at the sibling level. Some
6502 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6503 * case of POWER7, it can move to lower SMT modes only when higher
6504 * threads are idle. When in lower SMT modes, the threads will
6505 * perform better since they share less core resources. Hence when we
6506 * have idle threads, we want them to be the higher ones.
6508 * This packing function is run on idle threads. It checks to see if
6509 * the busiest CPU in this domain (core in the P7 case) has a higher
6510 * CPU number than the packing function is being run on. Here we are
6511 * assuming lower CPU number will be equivalent to lower a SMT thread
6514 * Return: 1 when packing is required and a task should be moved to
6515 * this CPU. The amount of the imbalance is returned in *imbalance.
6517 * @env: The load balancing environment.
6518 * @sds: Statistics of the sched_domain which is to be packed
6520 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6524 if (!(env->sd->flags & SD_ASYM_PACKING))
6530 busiest_cpu = group_first_cpu(sds->busiest);
6531 if (env->dst_cpu > busiest_cpu)
6534 env->imbalance = DIV_ROUND_CLOSEST(
6535 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6536 SCHED_CAPACITY_SCALE);
6542 * fix_small_imbalance - Calculate the minor imbalance that exists
6543 * amongst the groups of a sched_domain, during
6545 * @env: The load balancing environment.
6546 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6549 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6551 unsigned long tmp, capa_now = 0, capa_move = 0;
6552 unsigned int imbn = 2;
6553 unsigned long scaled_busy_load_per_task;
6554 struct sg_lb_stats *local, *busiest;
6556 local = &sds->local_stat;
6557 busiest = &sds->busiest_stat;
6559 if (!local->sum_nr_running)
6560 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6561 else if (busiest->load_per_task > local->load_per_task)
6564 scaled_busy_load_per_task =
6565 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6566 busiest->group_capacity;
6568 if (busiest->avg_load + scaled_busy_load_per_task >=
6569 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6570 env->imbalance = busiest->load_per_task;
6575 * OK, we don't have enough imbalance to justify moving tasks,
6576 * however we may be able to increase total CPU capacity used by
6580 capa_now += busiest->group_capacity *
6581 min(busiest->load_per_task, busiest->avg_load);
6582 capa_now += local->group_capacity *
6583 min(local->load_per_task, local->avg_load);
6584 capa_now /= SCHED_CAPACITY_SCALE;
6586 /* Amount of load we'd subtract */
6587 if (busiest->avg_load > scaled_busy_load_per_task) {
6588 capa_move += busiest->group_capacity *
6589 min(busiest->load_per_task,
6590 busiest->avg_load - scaled_busy_load_per_task);
6593 /* Amount of load we'd add */
6594 if (busiest->avg_load * busiest->group_capacity <
6595 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6596 tmp = (busiest->avg_load * busiest->group_capacity) /
6597 local->group_capacity;
6599 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6600 local->group_capacity;
6602 capa_move += local->group_capacity *
6603 min(local->load_per_task, local->avg_load + tmp);
6604 capa_move /= SCHED_CAPACITY_SCALE;
6606 /* Move if we gain throughput */
6607 if (capa_move > capa_now)
6608 env->imbalance = busiest->load_per_task;
6612 * calculate_imbalance - Calculate the amount of imbalance present within the
6613 * groups of a given sched_domain during load balance.
6614 * @env: load balance environment
6615 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6617 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6619 unsigned long max_pull, load_above_capacity = ~0UL;
6620 struct sg_lb_stats *local, *busiest;
6622 local = &sds->local_stat;
6623 busiest = &sds->busiest_stat;
6625 if (busiest->group_type == group_imbalanced) {
6627 * In the group_imb case we cannot rely on group-wide averages
6628 * to ensure cpu-load equilibrium, look at wider averages. XXX
6630 busiest->load_per_task =
6631 min(busiest->load_per_task, sds->avg_load);
6635 * In the presence of smp nice balancing, certain scenarios can have
6636 * max load less than avg load(as we skip the groups at or below
6637 * its cpu_capacity, while calculating max_load..)
6639 if (busiest->avg_load <= sds->avg_load ||
6640 local->avg_load >= sds->avg_load) {
6642 return fix_small_imbalance(env, sds);
6646 * If there aren't any idle cpus, avoid creating some.
6648 if (busiest->group_type == group_overloaded &&
6649 local->group_type == group_overloaded) {
6650 load_above_capacity = busiest->sum_nr_running *
6652 if (load_above_capacity > busiest->group_capacity)
6653 load_above_capacity -= busiest->group_capacity;
6655 load_above_capacity = ~0UL;
6659 * We're trying to get all the cpus to the average_load, so we don't
6660 * want to push ourselves above the average load, nor do we wish to
6661 * reduce the max loaded cpu below the average load. At the same time,
6662 * we also don't want to reduce the group load below the group capacity
6663 * (so that we can implement power-savings policies etc). Thus we look
6664 * for the minimum possible imbalance.
6666 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6668 /* How much load to actually move to equalise the imbalance */
6669 env->imbalance = min(
6670 max_pull * busiest->group_capacity,
6671 (sds->avg_load - local->avg_load) * local->group_capacity
6672 ) / SCHED_CAPACITY_SCALE;
6675 * if *imbalance is less than the average load per runnable task
6676 * there is no guarantee that any tasks will be moved so we'll have
6677 * a think about bumping its value to force at least one task to be
6680 if (env->imbalance < busiest->load_per_task)
6681 return fix_small_imbalance(env, sds);
6684 /******* find_busiest_group() helpers end here *********************/
6687 * find_busiest_group - Returns the busiest group within the sched_domain
6688 * if there is an imbalance. If there isn't an imbalance, and
6689 * the user has opted for power-savings, it returns a group whose
6690 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6691 * such a group exists.
6693 * Also calculates the amount of weighted load which should be moved
6694 * to restore balance.
6696 * @env: The load balancing environment.
6698 * Return: - The busiest group if imbalance exists.
6699 * - If no imbalance and user has opted for power-savings balance,
6700 * return the least loaded group whose CPUs can be
6701 * put to idle by rebalancing its tasks onto our group.
6703 static struct sched_group *find_busiest_group(struct lb_env *env)
6705 struct sg_lb_stats *local, *busiest;
6706 struct sd_lb_stats sds;
6708 init_sd_lb_stats(&sds);
6711 * Compute the various statistics relavent for load balancing at
6714 update_sd_lb_stats(env, &sds);
6715 local = &sds.local_stat;
6716 busiest = &sds.busiest_stat;
6718 /* ASYM feature bypasses nice load balance check */
6719 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6720 check_asym_packing(env, &sds))
6723 /* There is no busy sibling group to pull tasks from */
6724 if (!sds.busiest || busiest->sum_nr_running == 0)
6727 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6728 / sds.total_capacity;
6731 * If the busiest group is imbalanced the below checks don't
6732 * work because they assume all things are equal, which typically
6733 * isn't true due to cpus_allowed constraints and the like.
6735 if (busiest->group_type == group_imbalanced)
6738 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6739 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6740 busiest->group_no_capacity)
6744 * If the local group is busier than the selected busiest group
6745 * don't try and pull any tasks.
6747 if (local->avg_load >= busiest->avg_load)
6751 * Don't pull any tasks if this group is already above the domain
6754 if (local->avg_load >= sds.avg_load)
6757 if (env->idle == CPU_IDLE) {
6759 * This cpu is idle. If the busiest group is not overloaded
6760 * and there is no imbalance between this and busiest group
6761 * wrt idle cpus, it is balanced. The imbalance becomes
6762 * significant if the diff is greater than 1 otherwise we
6763 * might end up to just move the imbalance on another group
6765 if ((busiest->group_type != group_overloaded) &&
6766 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6770 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6771 * imbalance_pct to be conservative.
6773 if (100 * busiest->avg_load <=
6774 env->sd->imbalance_pct * local->avg_load)
6779 /* Looks like there is an imbalance. Compute it */
6780 calculate_imbalance(env, &sds);
6789 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6791 static struct rq *find_busiest_queue(struct lb_env *env,
6792 struct sched_group *group)
6794 struct rq *busiest = NULL, *rq;
6795 unsigned long busiest_load = 0, busiest_capacity = 1;
6798 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6799 unsigned long capacity, wl;
6803 rt = fbq_classify_rq(rq);
6806 * We classify groups/runqueues into three groups:
6807 * - regular: there are !numa tasks
6808 * - remote: there are numa tasks that run on the 'wrong' node
6809 * - all: there is no distinction
6811 * In order to avoid migrating ideally placed numa tasks,
6812 * ignore those when there's better options.
6814 * If we ignore the actual busiest queue to migrate another
6815 * task, the next balance pass can still reduce the busiest
6816 * queue by moving tasks around inside the node.
6818 * If we cannot move enough load due to this classification
6819 * the next pass will adjust the group classification and
6820 * allow migration of more tasks.
6822 * Both cases only affect the total convergence complexity.
6824 if (rt > env->fbq_type)
6827 capacity = capacity_of(i);
6829 wl = weighted_cpuload(i);
6832 * When comparing with imbalance, use weighted_cpuload()
6833 * which is not scaled with the cpu capacity.
6836 if (rq->nr_running == 1 && wl > env->imbalance &&
6837 !check_cpu_capacity(rq, env->sd))
6841 * For the load comparisons with the other cpu's, consider
6842 * the weighted_cpuload() scaled with the cpu capacity, so
6843 * that the load can be moved away from the cpu that is
6844 * potentially running at a lower capacity.
6846 * Thus we're looking for max(wl_i / capacity_i), crosswise
6847 * multiplication to rid ourselves of the division works out
6848 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6849 * our previous maximum.
6851 if (wl * busiest_capacity > busiest_load * capacity) {
6853 busiest_capacity = capacity;
6862 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6863 * so long as it is large enough.
6865 #define MAX_PINNED_INTERVAL 512
6867 /* Working cpumask for load_balance and load_balance_newidle. */
6868 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6870 static int need_active_balance(struct lb_env *env)
6872 struct sched_domain *sd = env->sd;
6874 if (env->idle == CPU_NEWLY_IDLE) {
6877 * ASYM_PACKING needs to force migrate tasks from busy but
6878 * higher numbered CPUs in order to pack all tasks in the
6879 * lowest numbered CPUs.
6881 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6886 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
6887 * It's worth migrating the task if the src_cpu's capacity is reduced
6888 * because of other sched_class or IRQs if more capacity stays
6889 * available on dst_cpu.
6891 if ((env->idle != CPU_NOT_IDLE) &&
6892 (env->src_rq->cfs.h_nr_running == 1)) {
6893 if ((check_cpu_capacity(env->src_rq, sd)) &&
6894 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
6898 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6901 static int active_load_balance_cpu_stop(void *data);
6903 static int should_we_balance(struct lb_env *env)
6905 struct sched_group *sg = env->sd->groups;
6906 struct cpumask *sg_cpus, *sg_mask;
6907 int cpu, balance_cpu = -1;
6910 * In the newly idle case, we will allow all the cpu's
6911 * to do the newly idle load balance.
6913 if (env->idle == CPU_NEWLY_IDLE)
6916 sg_cpus = sched_group_cpus(sg);
6917 sg_mask = sched_group_mask(sg);
6918 /* Try to find first idle cpu */
6919 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6920 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6927 if (balance_cpu == -1)
6928 balance_cpu = group_balance_cpu(sg);
6931 * First idle cpu or the first cpu(busiest) in this sched group
6932 * is eligible for doing load balancing at this and above domains.
6934 return balance_cpu == env->dst_cpu;
6938 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6939 * tasks if there is an imbalance.
6941 static int load_balance(int this_cpu, struct rq *this_rq,
6942 struct sched_domain *sd, enum cpu_idle_type idle,
6943 int *continue_balancing)
6945 int ld_moved, cur_ld_moved, active_balance = 0;
6946 struct sched_domain *sd_parent = sd->parent;
6947 struct sched_group *group;
6949 unsigned long flags;
6950 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
6952 struct lb_env env = {
6954 .dst_cpu = this_cpu,
6956 .dst_grpmask = sched_group_cpus(sd->groups),
6958 .loop_break = sched_nr_migrate_break,
6961 .tasks = LIST_HEAD_INIT(env.tasks),
6965 * For NEWLY_IDLE load_balancing, we don't need to consider
6966 * other cpus in our group
6968 if (idle == CPU_NEWLY_IDLE)
6969 env.dst_grpmask = NULL;
6971 cpumask_copy(cpus, cpu_active_mask);
6973 schedstat_inc(sd, lb_count[idle]);
6976 if (!should_we_balance(&env)) {
6977 *continue_balancing = 0;
6981 group = find_busiest_group(&env);
6983 schedstat_inc(sd, lb_nobusyg[idle]);
6987 busiest = find_busiest_queue(&env, group);
6989 schedstat_inc(sd, lb_nobusyq[idle]);
6993 BUG_ON(busiest == env.dst_rq);
6995 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6997 env.src_cpu = busiest->cpu;
6998 env.src_rq = busiest;
7001 if (busiest->nr_running > 1) {
7003 * Attempt to move tasks. If find_busiest_group has found
7004 * an imbalance but busiest->nr_running <= 1, the group is
7005 * still unbalanced. ld_moved simply stays zero, so it is
7006 * correctly treated as an imbalance.
7008 env.flags |= LBF_ALL_PINNED;
7009 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7012 raw_spin_lock_irqsave(&busiest->lock, flags);
7015 * cur_ld_moved - load moved in current iteration
7016 * ld_moved - cumulative load moved across iterations
7018 cur_ld_moved = detach_tasks(&env);
7021 * We've detached some tasks from busiest_rq. Every
7022 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7023 * unlock busiest->lock, and we are able to be sure
7024 * that nobody can manipulate the tasks in parallel.
7025 * See task_rq_lock() family for the details.
7028 raw_spin_unlock(&busiest->lock);
7032 ld_moved += cur_ld_moved;
7035 local_irq_restore(flags);
7037 if (env.flags & LBF_NEED_BREAK) {
7038 env.flags &= ~LBF_NEED_BREAK;
7043 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7044 * us and move them to an alternate dst_cpu in our sched_group
7045 * where they can run. The upper limit on how many times we
7046 * iterate on same src_cpu is dependent on number of cpus in our
7049 * This changes load balance semantics a bit on who can move
7050 * load to a given_cpu. In addition to the given_cpu itself
7051 * (or a ilb_cpu acting on its behalf where given_cpu is
7052 * nohz-idle), we now have balance_cpu in a position to move
7053 * load to given_cpu. In rare situations, this may cause
7054 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7055 * _independently_ and at _same_ time to move some load to
7056 * given_cpu) causing exceess load to be moved to given_cpu.
7057 * This however should not happen so much in practice and
7058 * moreover subsequent load balance cycles should correct the
7059 * excess load moved.
7061 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7063 /* Prevent to re-select dst_cpu via env's cpus */
7064 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7066 env.dst_rq = cpu_rq(env.new_dst_cpu);
7067 env.dst_cpu = env.new_dst_cpu;
7068 env.flags &= ~LBF_DST_PINNED;
7070 env.loop_break = sched_nr_migrate_break;
7073 * Go back to "more_balance" rather than "redo" since we
7074 * need to continue with same src_cpu.
7080 * We failed to reach balance because of affinity.
7083 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7085 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7086 *group_imbalance = 1;
7089 /* All tasks on this runqueue were pinned by CPU affinity */
7090 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7091 cpumask_clear_cpu(cpu_of(busiest), cpus);
7092 if (!cpumask_empty(cpus)) {
7094 env.loop_break = sched_nr_migrate_break;
7097 goto out_all_pinned;
7102 schedstat_inc(sd, lb_failed[idle]);
7104 * Increment the failure counter only on periodic balance.
7105 * We do not want newidle balance, which can be very
7106 * frequent, pollute the failure counter causing
7107 * excessive cache_hot migrations and active balances.
7109 if (idle != CPU_NEWLY_IDLE)
7110 sd->nr_balance_failed++;
7112 if (need_active_balance(&env)) {
7113 raw_spin_lock_irqsave(&busiest->lock, flags);
7115 /* don't kick the active_load_balance_cpu_stop,
7116 * if the curr task on busiest cpu can't be
7119 if (!cpumask_test_cpu(this_cpu,
7120 tsk_cpus_allowed(busiest->curr))) {
7121 raw_spin_unlock_irqrestore(&busiest->lock,
7123 env.flags |= LBF_ALL_PINNED;
7124 goto out_one_pinned;
7128 * ->active_balance synchronizes accesses to
7129 * ->active_balance_work. Once set, it's cleared
7130 * only after active load balance is finished.
7132 if (!busiest->active_balance) {
7133 busiest->active_balance = 1;
7134 busiest->push_cpu = this_cpu;
7137 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7139 if (active_balance) {
7140 stop_one_cpu_nowait(cpu_of(busiest),
7141 active_load_balance_cpu_stop, busiest,
7142 &busiest->active_balance_work);
7146 * We've kicked active balancing, reset the failure
7149 sd->nr_balance_failed = sd->cache_nice_tries+1;
7152 sd->nr_balance_failed = 0;
7154 if (likely(!active_balance)) {
7155 /* We were unbalanced, so reset the balancing interval */
7156 sd->balance_interval = sd->min_interval;
7159 * If we've begun active balancing, start to back off. This
7160 * case may not be covered by the all_pinned logic if there
7161 * is only 1 task on the busy runqueue (because we don't call
7164 if (sd->balance_interval < sd->max_interval)
7165 sd->balance_interval *= 2;
7172 * We reach balance although we may have faced some affinity
7173 * constraints. Clear the imbalance flag if it was set.
7176 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7178 if (*group_imbalance)
7179 *group_imbalance = 0;
7184 * We reach balance because all tasks are pinned at this level so
7185 * we can't migrate them. Let the imbalance flag set so parent level
7186 * can try to migrate them.
7188 schedstat_inc(sd, lb_balanced[idle]);
7190 sd->nr_balance_failed = 0;
7193 /* tune up the balancing interval */
7194 if (((env.flags & LBF_ALL_PINNED) &&
7195 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7196 (sd->balance_interval < sd->max_interval))
7197 sd->balance_interval *= 2;
7204 static inline unsigned long
7205 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7207 unsigned long interval = sd->balance_interval;
7210 interval *= sd->busy_factor;
7212 /* scale ms to jiffies */
7213 interval = msecs_to_jiffies(interval);
7214 interval = clamp(interval, 1UL, max_load_balance_interval);
7220 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7222 unsigned long interval, next;
7224 interval = get_sd_balance_interval(sd, cpu_busy);
7225 next = sd->last_balance + interval;
7227 if (time_after(*next_balance, next))
7228 *next_balance = next;
7232 * idle_balance is called by schedule() if this_cpu is about to become
7233 * idle. Attempts to pull tasks from other CPUs.
7235 static int idle_balance(struct rq *this_rq)
7237 unsigned long next_balance = jiffies + HZ;
7238 int this_cpu = this_rq->cpu;
7239 struct sched_domain *sd;
7240 int pulled_task = 0;
7243 idle_enter_fair(this_rq);
7246 * We must set idle_stamp _before_ calling idle_balance(), such that we
7247 * measure the duration of idle_balance() as idle time.
7249 this_rq->idle_stamp = rq_clock(this_rq);
7251 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7252 !this_rq->rd->overload) {
7254 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7256 update_next_balance(sd, 0, &next_balance);
7262 raw_spin_unlock(&this_rq->lock);
7264 update_blocked_averages(this_cpu);
7266 for_each_domain(this_cpu, sd) {
7267 int continue_balancing = 1;
7268 u64 t0, domain_cost;
7270 if (!(sd->flags & SD_LOAD_BALANCE))
7273 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7274 update_next_balance(sd, 0, &next_balance);
7278 if (sd->flags & SD_BALANCE_NEWIDLE) {
7279 t0 = sched_clock_cpu(this_cpu);
7281 pulled_task = load_balance(this_cpu, this_rq,
7283 &continue_balancing);
7285 domain_cost = sched_clock_cpu(this_cpu) - t0;
7286 if (domain_cost > sd->max_newidle_lb_cost)
7287 sd->max_newidle_lb_cost = domain_cost;
7289 curr_cost += domain_cost;
7292 update_next_balance(sd, 0, &next_balance);
7295 * Stop searching for tasks to pull if there are
7296 * now runnable tasks on this rq.
7298 if (pulled_task || this_rq->nr_running > 0)
7303 raw_spin_lock(&this_rq->lock);
7305 if (curr_cost > this_rq->max_idle_balance_cost)
7306 this_rq->max_idle_balance_cost = curr_cost;
7309 * While browsing the domains, we released the rq lock, a task could
7310 * have been enqueued in the meantime. Since we're not going idle,
7311 * pretend we pulled a task.
7313 if (this_rq->cfs.h_nr_running && !pulled_task)
7317 /* Move the next balance forward */
7318 if (time_after(this_rq->next_balance, next_balance))
7319 this_rq->next_balance = next_balance;
7321 /* Is there a task of a high priority class? */
7322 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7326 idle_exit_fair(this_rq);
7327 this_rq->idle_stamp = 0;
7334 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7335 * running tasks off the busiest CPU onto idle CPUs. It requires at
7336 * least 1 task to be running on each physical CPU where possible, and
7337 * avoids physical / logical imbalances.
7339 static int active_load_balance_cpu_stop(void *data)
7341 struct rq *busiest_rq = data;
7342 int busiest_cpu = cpu_of(busiest_rq);
7343 int target_cpu = busiest_rq->push_cpu;
7344 struct rq *target_rq = cpu_rq(target_cpu);
7345 struct sched_domain *sd;
7346 struct task_struct *p = NULL;
7348 raw_spin_lock_irq(&busiest_rq->lock);
7350 /* make sure the requested cpu hasn't gone down in the meantime */
7351 if (unlikely(busiest_cpu != smp_processor_id() ||
7352 !busiest_rq->active_balance))
7355 /* Is there any task to move? */
7356 if (busiest_rq->nr_running <= 1)
7360 * This condition is "impossible", if it occurs
7361 * we need to fix it. Originally reported by
7362 * Bjorn Helgaas on a 128-cpu setup.
7364 BUG_ON(busiest_rq == target_rq);
7366 /* Search for an sd spanning us and the target CPU. */
7368 for_each_domain(target_cpu, sd) {
7369 if ((sd->flags & SD_LOAD_BALANCE) &&
7370 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7375 struct lb_env env = {
7377 .dst_cpu = target_cpu,
7378 .dst_rq = target_rq,
7379 .src_cpu = busiest_rq->cpu,
7380 .src_rq = busiest_rq,
7384 schedstat_inc(sd, alb_count);
7386 p = detach_one_task(&env);
7388 schedstat_inc(sd, alb_pushed);
7390 schedstat_inc(sd, alb_failed);
7394 busiest_rq->active_balance = 0;
7395 raw_spin_unlock(&busiest_rq->lock);
7398 attach_one_task(target_rq, p);
7405 static inline int on_null_domain(struct rq *rq)
7407 return unlikely(!rcu_dereference_sched(rq->sd));
7410 #ifdef CONFIG_NO_HZ_COMMON
7412 * idle load balancing details
7413 * - When one of the busy CPUs notice that there may be an idle rebalancing
7414 * needed, they will kick the idle load balancer, which then does idle
7415 * load balancing for all the idle CPUs.
7418 cpumask_var_t idle_cpus_mask;
7420 unsigned long next_balance; /* in jiffy units */
7421 } nohz ____cacheline_aligned;
7423 static inline int find_new_ilb(void)
7425 int ilb = cpumask_first(nohz.idle_cpus_mask);
7427 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7434 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7435 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7436 * CPU (if there is one).
7438 static void nohz_balancer_kick(void)
7442 nohz.next_balance++;
7444 ilb_cpu = find_new_ilb();
7446 if (ilb_cpu >= nr_cpu_ids)
7449 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7452 * Use smp_send_reschedule() instead of resched_cpu().
7453 * This way we generate a sched IPI on the target cpu which
7454 * is idle. And the softirq performing nohz idle load balance
7455 * will be run before returning from the IPI.
7457 smp_send_reschedule(ilb_cpu);
7461 static inline void nohz_balance_exit_idle(int cpu)
7463 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7465 * Completely isolated CPUs don't ever set, so we must test.
7467 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7468 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7469 atomic_dec(&nohz.nr_cpus);
7471 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7475 static inline void set_cpu_sd_state_busy(void)
7477 struct sched_domain *sd;
7478 int cpu = smp_processor_id();
7481 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7483 if (!sd || !sd->nohz_idle)
7487 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7492 void set_cpu_sd_state_idle(void)
7494 struct sched_domain *sd;
7495 int cpu = smp_processor_id();
7498 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7500 if (!sd || sd->nohz_idle)
7504 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7510 * This routine will record that the cpu is going idle with tick stopped.
7511 * This info will be used in performing idle load balancing in the future.
7513 void nohz_balance_enter_idle(int cpu)
7516 * If this cpu is going down, then nothing needs to be done.
7518 if (!cpu_active(cpu))
7521 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7525 * If we're a completely isolated CPU, we don't play.
7527 if (on_null_domain(cpu_rq(cpu)))
7530 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7531 atomic_inc(&nohz.nr_cpus);
7532 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7535 static int sched_ilb_notifier(struct notifier_block *nfb,
7536 unsigned long action, void *hcpu)
7538 switch (action & ~CPU_TASKS_FROZEN) {
7540 nohz_balance_exit_idle(smp_processor_id());
7548 static DEFINE_SPINLOCK(balancing);
7551 * Scale the max load_balance interval with the number of CPUs in the system.
7552 * This trades load-balance latency on larger machines for less cross talk.
7554 void update_max_interval(void)
7556 max_load_balance_interval = HZ*num_online_cpus()/10;
7560 * It checks each scheduling domain to see if it is due to be balanced,
7561 * and initiates a balancing operation if so.
7563 * Balancing parameters are set up in init_sched_domains.
7565 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7567 int continue_balancing = 1;
7569 unsigned long interval;
7570 struct sched_domain *sd;
7571 /* Earliest time when we have to do rebalance again */
7572 unsigned long next_balance = jiffies + 60*HZ;
7573 int update_next_balance = 0;
7574 int need_serialize, need_decay = 0;
7577 update_blocked_averages(cpu);
7580 for_each_domain(cpu, sd) {
7582 * Decay the newidle max times here because this is a regular
7583 * visit to all the domains. Decay ~1% per second.
7585 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7586 sd->max_newidle_lb_cost =
7587 (sd->max_newidle_lb_cost * 253) / 256;
7588 sd->next_decay_max_lb_cost = jiffies + HZ;
7591 max_cost += sd->max_newidle_lb_cost;
7593 if (!(sd->flags & SD_LOAD_BALANCE))
7597 * Stop the load balance at this level. There is another
7598 * CPU in our sched group which is doing load balancing more
7601 if (!continue_balancing) {
7607 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7609 need_serialize = sd->flags & SD_SERIALIZE;
7610 if (need_serialize) {
7611 if (!spin_trylock(&balancing))
7615 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7616 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7618 * The LBF_DST_PINNED logic could have changed
7619 * env->dst_cpu, so we can't know our idle
7620 * state even if we migrated tasks. Update it.
7622 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7624 sd->last_balance = jiffies;
7625 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7628 spin_unlock(&balancing);
7630 if (time_after(next_balance, sd->last_balance + interval)) {
7631 next_balance = sd->last_balance + interval;
7632 update_next_balance = 1;
7637 * Ensure the rq-wide value also decays but keep it at a
7638 * reasonable floor to avoid funnies with rq->avg_idle.
7640 rq->max_idle_balance_cost =
7641 max((u64)sysctl_sched_migration_cost, max_cost);
7646 * next_balance will be updated only when there is a need.
7647 * When the cpu is attached to null domain for ex, it will not be
7650 if (likely(update_next_balance))
7651 rq->next_balance = next_balance;
7654 #ifdef CONFIG_NO_HZ_COMMON
7656 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7657 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7659 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7661 int this_cpu = this_rq->cpu;
7665 if (idle != CPU_IDLE ||
7666 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7669 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7670 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7674 * If this cpu gets work to do, stop the load balancing
7675 * work being done for other cpus. Next load
7676 * balancing owner will pick it up.
7681 rq = cpu_rq(balance_cpu);
7684 * If time for next balance is due,
7687 if (time_after_eq(jiffies, rq->next_balance)) {
7688 raw_spin_lock_irq(&rq->lock);
7689 update_rq_clock(rq);
7690 update_idle_cpu_load(rq);
7691 raw_spin_unlock_irq(&rq->lock);
7692 rebalance_domains(rq, CPU_IDLE);
7695 if (time_after(this_rq->next_balance, rq->next_balance))
7696 this_rq->next_balance = rq->next_balance;
7698 nohz.next_balance = this_rq->next_balance;
7700 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7704 * Current heuristic for kicking the idle load balancer in the presence
7705 * of an idle cpu in the system.
7706 * - This rq has more than one task.
7707 * - This rq has at least one CFS task and the capacity of the CPU is
7708 * significantly reduced because of RT tasks or IRQs.
7709 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7710 * multiple busy cpu.
7711 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7712 * domain span are idle.
7714 static inline bool nohz_kick_needed(struct rq *rq)
7716 unsigned long now = jiffies;
7717 struct sched_domain *sd;
7718 struct sched_group_capacity *sgc;
7719 int nr_busy, cpu = rq->cpu;
7722 if (unlikely(rq->idle_balance))
7726 * We may be recently in ticked or tickless idle mode. At the first
7727 * busy tick after returning from idle, we will update the busy stats.
7729 set_cpu_sd_state_busy();
7730 nohz_balance_exit_idle(cpu);
7733 * None are in tickless mode and hence no need for NOHZ idle load
7736 if (likely(!atomic_read(&nohz.nr_cpus)))
7739 if (time_before(now, nohz.next_balance))
7742 if (rq->nr_running >= 2)
7746 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7748 sgc = sd->groups->sgc;
7749 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7758 sd = rcu_dereference(rq->sd);
7760 if ((rq->cfs.h_nr_running >= 1) &&
7761 check_cpu_capacity(rq, sd)) {
7767 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7768 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7769 sched_domain_span(sd)) < cpu)) {
7779 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7783 * run_rebalance_domains is triggered when needed from the scheduler tick.
7784 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7786 static void run_rebalance_domains(struct softirq_action *h)
7788 struct rq *this_rq = this_rq();
7789 enum cpu_idle_type idle = this_rq->idle_balance ?
7790 CPU_IDLE : CPU_NOT_IDLE;
7793 * If this cpu has a pending nohz_balance_kick, then do the
7794 * balancing on behalf of the other idle cpus whose ticks are
7795 * stopped. Do nohz_idle_balance *before* rebalance_domains to
7796 * give the idle cpus a chance to load balance. Else we may
7797 * load balance only within the local sched_domain hierarchy
7798 * and abort nohz_idle_balance altogether if we pull some load.
7800 nohz_idle_balance(this_rq, idle);
7801 rebalance_domains(this_rq, idle);
7805 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7807 void trigger_load_balance(struct rq *rq)
7809 /* Don't need to rebalance while attached to NULL domain */
7810 if (unlikely(on_null_domain(rq)))
7813 if (time_after_eq(jiffies, rq->next_balance))
7814 raise_softirq(SCHED_SOFTIRQ);
7815 #ifdef CONFIG_NO_HZ_COMMON
7816 if (nohz_kick_needed(rq))
7817 nohz_balancer_kick();
7821 static void rq_online_fair(struct rq *rq)
7825 update_runtime_enabled(rq);
7828 static void rq_offline_fair(struct rq *rq)
7832 /* Ensure any throttled groups are reachable by pick_next_task */
7833 unthrottle_offline_cfs_rqs(rq);
7836 #endif /* CONFIG_SMP */
7839 * scheduler tick hitting a task of our scheduling class:
7841 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7843 struct cfs_rq *cfs_rq;
7844 struct sched_entity *se = &curr->se;
7846 for_each_sched_entity(se) {
7847 cfs_rq = cfs_rq_of(se);
7848 entity_tick(cfs_rq, se, queued);
7851 if (numabalancing_enabled)
7852 task_tick_numa(rq, curr);
7856 * called on fork with the child task as argument from the parent's context
7857 * - child not yet on the tasklist
7858 * - preemption disabled
7860 static void task_fork_fair(struct task_struct *p)
7862 struct cfs_rq *cfs_rq;
7863 struct sched_entity *se = &p->se, *curr;
7864 int this_cpu = smp_processor_id();
7865 struct rq *rq = this_rq();
7866 unsigned long flags;
7868 raw_spin_lock_irqsave(&rq->lock, flags);
7870 update_rq_clock(rq);
7872 cfs_rq = task_cfs_rq(current);
7873 curr = cfs_rq->curr;
7876 * Not only the cpu but also the task_group of the parent might have
7877 * been changed after parent->se.parent,cfs_rq were copied to
7878 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7879 * of child point to valid ones.
7882 __set_task_cpu(p, this_cpu);
7885 update_curr(cfs_rq);
7888 se->vruntime = curr->vruntime;
7889 place_entity(cfs_rq, se, 1);
7891 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7893 * Upon rescheduling, sched_class::put_prev_task() will place
7894 * 'current' within the tree based on its new key value.
7896 swap(curr->vruntime, se->vruntime);
7900 se->vruntime -= cfs_rq->min_vruntime;
7902 raw_spin_unlock_irqrestore(&rq->lock, flags);
7906 * Priority of the task has changed. Check to see if we preempt
7910 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7912 if (!task_on_rq_queued(p))
7916 * Reschedule if we are currently running on this runqueue and
7917 * our priority decreased, or if we are not currently running on
7918 * this runqueue and our priority is higher than the current's
7920 if (rq->curr == p) {
7921 if (p->prio > oldprio)
7924 check_preempt_curr(rq, p, 0);
7927 static inline bool vruntime_normalized(struct task_struct *p)
7929 struct sched_entity *se = &p->se;
7932 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
7933 * the dequeue_entity(.flags=0) will already have normalized the
7940 * When !on_rq, vruntime of the task has usually NOT been normalized.
7941 * But there are some cases where it has already been normalized:
7943 * - A forked child which is waiting for being woken up by
7944 * wake_up_new_task().
7945 * - A task which has been woken up by try_to_wake_up() and
7946 * waiting for actually being woken up by sched_ttwu_pending().
7948 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
7954 static void detach_task_cfs_rq(struct task_struct *p)
7956 struct sched_entity *se = &p->se;
7957 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7959 if (!vruntime_normalized(p)) {
7961 * Fix up our vruntime so that the current sleep doesn't
7962 * cause 'unlimited' sleep bonus.
7964 place_entity(cfs_rq, se, 0);
7965 se->vruntime -= cfs_rq->min_vruntime;
7968 /* Catch up with the cfs_rq and remove our load when we leave */
7969 detach_entity_load_avg(cfs_rq, se);
7972 static void attach_task_cfs_rq(struct task_struct *p)
7974 struct sched_entity *se = &p->se;
7975 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7977 #ifdef CONFIG_FAIR_GROUP_SCHED
7979 * Since the real-depth could have been changed (only FAIR
7980 * class maintain depth value), reset depth properly.
7982 se->depth = se->parent ? se->parent->depth + 1 : 0;
7985 /* Synchronize task with its cfs_rq */
7986 attach_entity_load_avg(cfs_rq, se);
7988 if (!vruntime_normalized(p))
7989 se->vruntime += cfs_rq->min_vruntime;
7992 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7994 detach_task_cfs_rq(p);
7997 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7999 attach_task_cfs_rq(p);
8001 if (task_on_rq_queued(p)) {
8003 * We were most likely switched from sched_rt, so
8004 * kick off the schedule if running, otherwise just see
8005 * if we can still preempt the current task.
8010 check_preempt_curr(rq, p, 0);
8014 /* Account for a task changing its policy or group.
8016 * This routine is mostly called to set cfs_rq->curr field when a task
8017 * migrates between groups/classes.
8019 static void set_curr_task_fair(struct rq *rq)
8021 struct sched_entity *se = &rq->curr->se;
8023 for_each_sched_entity(se) {
8024 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8026 set_next_entity(cfs_rq, se);
8027 /* ensure bandwidth has been allocated on our new cfs_rq */
8028 account_cfs_rq_runtime(cfs_rq, 0);
8032 void init_cfs_rq(struct cfs_rq *cfs_rq)
8034 cfs_rq->tasks_timeline = RB_ROOT;
8035 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8036 #ifndef CONFIG_64BIT
8037 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8040 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8041 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8045 #ifdef CONFIG_FAIR_GROUP_SCHED
8046 static void task_move_group_fair(struct task_struct *p, int queued)
8048 detach_task_cfs_rq(p);
8049 set_task_rq(p, task_cpu(p));
8052 /* Tell se's cfs_rq has been changed -- migrated */
8053 p->se.avg.last_update_time = 0;
8055 attach_task_cfs_rq(p);
8058 void free_fair_sched_group(struct task_group *tg)
8062 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8064 for_each_possible_cpu(i) {
8066 kfree(tg->cfs_rq[i]);
8069 remove_entity_load_avg(tg->se[i]);
8078 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8080 struct cfs_rq *cfs_rq;
8081 struct sched_entity *se;
8084 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8087 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8091 tg->shares = NICE_0_LOAD;
8093 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8095 for_each_possible_cpu(i) {
8096 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8097 GFP_KERNEL, cpu_to_node(i));
8101 se = kzalloc_node(sizeof(struct sched_entity),
8102 GFP_KERNEL, cpu_to_node(i));
8106 init_cfs_rq(cfs_rq);
8107 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8108 init_entity_runnable_average(se);
8119 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8121 struct rq *rq = cpu_rq(cpu);
8122 unsigned long flags;
8125 * Only empty task groups can be destroyed; so we can speculatively
8126 * check on_list without danger of it being re-added.
8128 if (!tg->cfs_rq[cpu]->on_list)
8131 raw_spin_lock_irqsave(&rq->lock, flags);
8132 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8133 raw_spin_unlock_irqrestore(&rq->lock, flags);
8136 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8137 struct sched_entity *se, int cpu,
8138 struct sched_entity *parent)
8140 struct rq *rq = cpu_rq(cpu);
8144 init_cfs_rq_runtime(cfs_rq);
8146 tg->cfs_rq[cpu] = cfs_rq;
8149 /* se could be NULL for root_task_group */
8154 se->cfs_rq = &rq->cfs;
8157 se->cfs_rq = parent->my_q;
8158 se->depth = parent->depth + 1;
8162 /* guarantee group entities always have weight */
8163 update_load_set(&se->load, NICE_0_LOAD);
8164 se->parent = parent;
8167 static DEFINE_MUTEX(shares_mutex);
8169 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8172 unsigned long flags;
8175 * We can't change the weight of the root cgroup.
8180 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8182 mutex_lock(&shares_mutex);
8183 if (tg->shares == shares)
8186 tg->shares = shares;
8187 for_each_possible_cpu(i) {
8188 struct rq *rq = cpu_rq(i);
8189 struct sched_entity *se;
8192 /* Propagate contribution to hierarchy */
8193 raw_spin_lock_irqsave(&rq->lock, flags);
8195 /* Possible calls to update_curr() need rq clock */
8196 update_rq_clock(rq);
8197 for_each_sched_entity(se)
8198 update_cfs_shares(group_cfs_rq(se));
8199 raw_spin_unlock_irqrestore(&rq->lock, flags);
8203 mutex_unlock(&shares_mutex);
8206 #else /* CONFIG_FAIR_GROUP_SCHED */
8208 void free_fair_sched_group(struct task_group *tg) { }
8210 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8215 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8217 #endif /* CONFIG_FAIR_GROUP_SCHED */
8220 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8222 struct sched_entity *se = &task->se;
8223 unsigned int rr_interval = 0;
8226 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8229 if (rq->cfs.load.weight)
8230 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8236 * All the scheduling class methods:
8238 const struct sched_class fair_sched_class = {
8239 .next = &idle_sched_class,
8240 .enqueue_task = enqueue_task_fair,
8241 .dequeue_task = dequeue_task_fair,
8242 .yield_task = yield_task_fair,
8243 .yield_to_task = yield_to_task_fair,
8245 .check_preempt_curr = check_preempt_wakeup,
8247 .pick_next_task = pick_next_task_fair,
8248 .put_prev_task = put_prev_task_fair,
8251 .select_task_rq = select_task_rq_fair,
8252 .migrate_task_rq = migrate_task_rq_fair,
8254 .rq_online = rq_online_fair,
8255 .rq_offline = rq_offline_fair,
8257 .task_waking = task_waking_fair,
8258 .task_dead = task_dead_fair,
8259 .set_cpus_allowed = set_cpus_allowed_common,
8262 .set_curr_task = set_curr_task_fair,
8263 .task_tick = task_tick_fair,
8264 .task_fork = task_fork_fair,
8266 .prio_changed = prio_changed_fair,
8267 .switched_from = switched_from_fair,
8268 .switched_to = switched_to_fair,
8270 .get_rr_interval = get_rr_interval_fair,
8272 .update_curr = update_curr_fair,
8274 #ifdef CONFIG_FAIR_GROUP_SCHED
8275 .task_move_group = task_move_group_fair,
8279 #ifdef CONFIG_SCHED_DEBUG
8280 void print_cfs_stats(struct seq_file *m, int cpu)
8282 struct cfs_rq *cfs_rq;
8285 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8286 print_cfs_rq(m, cpu, cfs_rq);
8290 #ifdef CONFIG_NUMA_BALANCING
8291 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8294 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8296 for_each_online_node(node) {
8297 if (p->numa_faults) {
8298 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8299 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8301 if (p->numa_group) {
8302 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8303 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8305 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8308 #endif /* CONFIG_NUMA_BALANCING */
8309 #endif /* CONFIG_SCHED_DEBUG */
8311 __init void init_sched_fair_class(void)
8314 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8316 #ifdef CONFIG_NO_HZ_COMMON
8317 nohz.next_balance = jiffies;
8318 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8319 cpu_notifier(sched_ilb_notifier, 0);