2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
40 * Targeted preemption latency for CPU-bound tasks:
41 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
43 * NOTE: this latency value is not the same as the concept of
44 * 'timeslice length' - timeslices in CFS are of variable length
45 * and have no persistent notion like in traditional, time-slice
46 * based scheduling concepts.
48 * (to see the precise effective timeslice length of your workload,
49 * run vmstat and monitor the context-switches (cs) field)
51 unsigned int sysctl_sched_latency = 6000000ULL;
52 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
55 * The initial- and re-scaling of tunables is configurable
56 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
59 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
60 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
61 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
63 enum sched_tunable_scaling sysctl_sched_tunable_scaling
64 = SCHED_TUNABLESCALING_LOG;
67 * Minimal preemption granularity for CPU-bound tasks:
68 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
70 unsigned int sysctl_sched_min_granularity = 750000ULL;
71 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
74 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
76 static unsigned int sched_nr_latency = 8;
79 * After fork, child runs first. If set to 0 (default) then
80 * parent will (try to) run first.
82 unsigned int sysctl_sched_child_runs_first __read_mostly;
85 * SCHED_OTHER wake-up granularity.
86 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
88 * This option delays the preemption effects of decoupled workloads
89 * and reduces their over-scheduling. Synchronous workloads will still
90 * have immediate wakeup/sleep latencies.
92 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
93 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
95 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
98 * The exponential sliding window over which load is averaged for shares
102 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
104 #ifdef CONFIG_CFS_BANDWIDTH
106 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
107 * each time a cfs_rq requests quota.
109 * Note: in the case that the slice exceeds the runtime remaining (either due
110 * to consumption or the quota being specified to be smaller than the slice)
111 * we will always only issue the remaining available time.
113 * default: 5 msec, units: microseconds
115 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
118 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
124 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
130 static inline void update_load_set(struct load_weight *lw, unsigned long w)
137 * Increase the granularity value when there are more CPUs,
138 * because with more CPUs the 'effective latency' as visible
139 * to users decreases. But the relationship is not linear,
140 * so pick a second-best guess by going with the log2 of the
143 * This idea comes from the SD scheduler of Con Kolivas:
145 static unsigned int get_update_sysctl_factor(void)
147 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
150 switch (sysctl_sched_tunable_scaling) {
151 case SCHED_TUNABLESCALING_NONE:
154 case SCHED_TUNABLESCALING_LINEAR:
157 case SCHED_TUNABLESCALING_LOG:
159 factor = 1 + ilog2(cpus);
166 static void update_sysctl(void)
168 unsigned int factor = get_update_sysctl_factor();
170 #define SET_SYSCTL(name) \
171 (sysctl_##name = (factor) * normalized_sysctl_##name)
172 SET_SYSCTL(sched_min_granularity);
173 SET_SYSCTL(sched_latency);
174 SET_SYSCTL(sched_wakeup_granularity);
178 void sched_init_granularity(void)
183 #define WMULT_CONST (~0U)
184 #define WMULT_SHIFT 32
186 static void __update_inv_weight(struct load_weight *lw)
190 if (likely(lw->inv_weight))
193 w = scale_load_down(lw->weight);
195 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
197 else if (unlikely(!w))
198 lw->inv_weight = WMULT_CONST;
200 lw->inv_weight = WMULT_CONST / w;
204 * delta_exec * weight / lw.weight
206 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
208 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
209 * we're guaranteed shift stays positive because inv_weight is guaranteed to
210 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
212 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
213 * weight/lw.weight <= 1, and therefore our shift will also be positive.
215 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
217 u64 fact = scale_load_down(weight);
218 int shift = WMULT_SHIFT;
220 __update_inv_weight(lw);
222 if (unlikely(fact >> 32)) {
229 /* hint to use a 32x32->64 mul */
230 fact = (u64)(u32)fact * lw->inv_weight;
237 return mul_u64_u32_shr(delta_exec, fact, shift);
241 const struct sched_class fair_sched_class;
243 /**************************************************************
244 * CFS operations on generic schedulable entities:
247 #ifdef CONFIG_FAIR_GROUP_SCHED
249 /* cpu runqueue to which this cfs_rq is attached */
250 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
255 /* An entity is a task if it doesn't "own" a runqueue */
256 #define entity_is_task(se) (!se->my_q)
258 static inline struct task_struct *task_of(struct sched_entity *se)
260 #ifdef CONFIG_SCHED_DEBUG
261 WARN_ON_ONCE(!entity_is_task(se));
263 return container_of(se, struct task_struct, se);
266 /* Walk up scheduling entities hierarchy */
267 #define for_each_sched_entity(se) \
268 for (; se; se = se->parent)
270 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
275 /* runqueue on which this entity is (to be) queued */
276 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
281 /* runqueue "owned" by this group */
282 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
287 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
289 if (!cfs_rq->on_list) {
291 * Ensure we either appear before our parent (if already
292 * enqueued) or force our parent to appear after us when it is
293 * enqueued. The fact that we always enqueue bottom-up
294 * reduces this to two cases.
296 if (cfs_rq->tg->parent &&
297 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
298 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
299 &rq_of(cfs_rq)->leaf_cfs_rq_list);
301 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
302 &rq_of(cfs_rq)->leaf_cfs_rq_list);
309 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
311 if (cfs_rq->on_list) {
312 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
317 /* Iterate thr' all leaf cfs_rq's on a runqueue */
318 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
319 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
321 /* Do the two (enqueued) entities belong to the same group ? */
322 static inline struct cfs_rq *
323 is_same_group(struct sched_entity *se, struct sched_entity *pse)
325 if (se->cfs_rq == pse->cfs_rq)
331 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
339 int se_depth, pse_depth;
342 * preemption test can be made between sibling entities who are in the
343 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
344 * both tasks until we find their ancestors who are siblings of common
348 /* First walk up until both entities are at same depth */
349 se_depth = (*se)->depth;
350 pse_depth = (*pse)->depth;
352 while (se_depth > pse_depth) {
354 *se = parent_entity(*se);
357 while (pse_depth > se_depth) {
359 *pse = parent_entity(*pse);
362 while (!is_same_group(*se, *pse)) {
363 *se = parent_entity(*se);
364 *pse = parent_entity(*pse);
368 #else /* !CONFIG_FAIR_GROUP_SCHED */
370 static inline struct task_struct *task_of(struct sched_entity *se)
372 return container_of(se, struct task_struct, se);
375 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
377 return container_of(cfs_rq, struct rq, cfs);
380 #define entity_is_task(se) 1
382 #define for_each_sched_entity(se) \
383 for (; se; se = NULL)
385 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
387 return &task_rq(p)->cfs;
390 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
392 struct task_struct *p = task_of(se);
393 struct rq *rq = task_rq(p);
398 /* runqueue "owned" by this group */
399 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
404 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
408 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
412 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
413 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
415 static inline struct sched_entity *parent_entity(struct sched_entity *se)
421 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
425 #endif /* CONFIG_FAIR_GROUP_SCHED */
427 static __always_inline
428 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
430 /**************************************************************
431 * Scheduling class tree data structure manipulation methods:
434 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
436 s64 delta = (s64)(vruntime - max_vruntime);
438 max_vruntime = vruntime;
443 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
445 s64 delta = (s64)(vruntime - min_vruntime);
447 min_vruntime = vruntime;
452 static inline int entity_before(struct sched_entity *a,
453 struct sched_entity *b)
455 return (s64)(a->vruntime - b->vruntime) < 0;
458 static void update_min_vruntime(struct cfs_rq *cfs_rq)
460 u64 vruntime = cfs_rq->min_vruntime;
463 vruntime = cfs_rq->curr->vruntime;
465 if (cfs_rq->rb_leftmost) {
466 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
471 vruntime = se->vruntime;
473 vruntime = min_vruntime(vruntime, se->vruntime);
476 /* ensure we never gain time by being placed backwards. */
477 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
480 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
485 * Enqueue an entity into the rb-tree:
487 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
489 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
490 struct rb_node *parent = NULL;
491 struct sched_entity *entry;
495 * Find the right place in the rbtree:
499 entry = rb_entry(parent, struct sched_entity, run_node);
501 * We dont care about collisions. Nodes with
502 * the same key stay together.
504 if (entity_before(se, entry)) {
505 link = &parent->rb_left;
507 link = &parent->rb_right;
513 * Maintain a cache of leftmost tree entries (it is frequently
517 cfs_rq->rb_leftmost = &se->run_node;
519 rb_link_node(&se->run_node, parent, link);
520 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
523 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
525 if (cfs_rq->rb_leftmost == &se->run_node) {
526 struct rb_node *next_node;
528 next_node = rb_next(&se->run_node);
529 cfs_rq->rb_leftmost = next_node;
532 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
535 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
537 struct rb_node *left = cfs_rq->rb_leftmost;
542 return rb_entry(left, struct sched_entity, run_node);
545 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
547 struct rb_node *next = rb_next(&se->run_node);
552 return rb_entry(next, struct sched_entity, run_node);
555 #ifdef CONFIG_SCHED_DEBUG
556 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
558 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
563 return rb_entry(last, struct sched_entity, run_node);
566 /**************************************************************
567 * Scheduling class statistics methods:
570 int sched_proc_update_handler(struct ctl_table *table, int write,
571 void __user *buffer, size_t *lenp,
574 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
575 unsigned int factor = get_update_sysctl_factor();
580 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
581 sysctl_sched_min_granularity);
583 #define WRT_SYSCTL(name) \
584 (normalized_sysctl_##name = sysctl_##name / (factor))
585 WRT_SYSCTL(sched_min_granularity);
586 WRT_SYSCTL(sched_latency);
587 WRT_SYSCTL(sched_wakeup_granularity);
597 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
599 if (unlikely(se->load.weight != NICE_0_LOAD))
600 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
606 * The idea is to set a period in which each task runs once.
608 * When there are too many tasks (sched_nr_latency) we have to stretch
609 * this period because otherwise the slices get too small.
611 * p = (nr <= nl) ? l : l*nr/nl
613 static u64 __sched_period(unsigned long nr_running)
615 if (unlikely(nr_running > sched_nr_latency))
616 return nr_running * sysctl_sched_min_granularity;
618 return sysctl_sched_latency;
622 * We calculate the wall-time slice from the period by taking a part
623 * proportional to the weight.
627 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
629 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
631 for_each_sched_entity(se) {
632 struct load_weight *load;
633 struct load_weight lw;
635 cfs_rq = cfs_rq_of(se);
636 load = &cfs_rq->load;
638 if (unlikely(!se->on_rq)) {
641 update_load_add(&lw, se->load.weight);
644 slice = __calc_delta(slice, se->load.weight, load);
650 * We calculate the vruntime slice of a to-be-inserted task.
654 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
656 return calc_delta_fair(sched_slice(cfs_rq, se), se);
660 static int select_idle_sibling(struct task_struct *p, int cpu);
661 static unsigned long task_h_load(struct task_struct *p);
664 * We choose a half-life close to 1 scheduling period.
665 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
666 * dependent on this value.
668 #define LOAD_AVG_PERIOD 32
669 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
670 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
672 /* Give new sched_entity start runnable values to heavy its load in infant time */
673 void init_entity_runnable_average(struct sched_entity *se)
675 struct sched_avg *sa = &se->avg;
677 sa->last_update_time = 0;
679 * sched_avg's period_contrib should be strictly less then 1024, so
680 * we give it 1023 to make sure it is almost a period (1024us), and
681 * will definitely be update (after enqueue).
683 sa->period_contrib = 1023;
684 sa->load_avg = scale_load_down(se->load.weight);
685 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
686 sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
687 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
688 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
691 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
692 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
694 void init_entity_runnable_average(struct sched_entity *se)
700 * Update the current task's runtime statistics.
702 static void update_curr(struct cfs_rq *cfs_rq)
704 struct sched_entity *curr = cfs_rq->curr;
705 u64 now = rq_clock_task(rq_of(cfs_rq));
711 delta_exec = now - curr->exec_start;
712 if (unlikely((s64)delta_exec <= 0))
715 curr->exec_start = now;
717 schedstat_set(curr->statistics.exec_max,
718 max(delta_exec, curr->statistics.exec_max));
720 curr->sum_exec_runtime += delta_exec;
721 schedstat_add(cfs_rq, exec_clock, delta_exec);
723 curr->vruntime += calc_delta_fair(delta_exec, curr);
724 update_min_vruntime(cfs_rq);
726 if (entity_is_task(curr)) {
727 struct task_struct *curtask = task_of(curr);
729 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
730 cpuacct_charge(curtask, delta_exec);
731 account_group_exec_runtime(curtask, delta_exec);
734 account_cfs_rq_runtime(cfs_rq, delta_exec);
737 static void update_curr_fair(struct rq *rq)
739 update_curr(cfs_rq_of(&rq->curr->se));
743 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
745 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
749 * Task is being enqueued - update stats:
751 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
754 * Are we enqueueing a waiting task? (for current tasks
755 * a dequeue/enqueue event is a NOP)
757 if (se != cfs_rq->curr)
758 update_stats_wait_start(cfs_rq, se);
762 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
764 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
765 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
766 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
767 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
768 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
769 #ifdef CONFIG_SCHEDSTATS
770 if (entity_is_task(se)) {
771 trace_sched_stat_wait(task_of(se),
772 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
775 schedstat_set(se->statistics.wait_start, 0);
779 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
782 * Mark the end of the wait period if dequeueing a
785 if (se != cfs_rq->curr)
786 update_stats_wait_end(cfs_rq, se);
790 * We are picking a new current task - update its stats:
793 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
796 * We are starting a new run period:
798 se->exec_start = rq_clock_task(rq_of(cfs_rq));
801 /**************************************************
802 * Scheduling class queueing methods:
805 #ifdef CONFIG_NUMA_BALANCING
807 * Approximate time to scan a full NUMA task in ms. The task scan period is
808 * calculated based on the tasks virtual memory size and
809 * numa_balancing_scan_size.
811 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
812 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
814 /* Portion of address space to scan in MB */
815 unsigned int sysctl_numa_balancing_scan_size = 256;
817 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
818 unsigned int sysctl_numa_balancing_scan_delay = 1000;
820 static unsigned int task_nr_scan_windows(struct task_struct *p)
822 unsigned long rss = 0;
823 unsigned long nr_scan_pages;
826 * Calculations based on RSS as non-present and empty pages are skipped
827 * by the PTE scanner and NUMA hinting faults should be trapped based
830 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
831 rss = get_mm_rss(p->mm);
835 rss = round_up(rss, nr_scan_pages);
836 return rss / nr_scan_pages;
839 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
840 #define MAX_SCAN_WINDOW 2560
842 static unsigned int task_scan_min(struct task_struct *p)
844 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
845 unsigned int scan, floor;
846 unsigned int windows = 1;
848 if (scan_size < MAX_SCAN_WINDOW)
849 windows = MAX_SCAN_WINDOW / scan_size;
850 floor = 1000 / windows;
852 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
853 return max_t(unsigned int, floor, scan);
856 static unsigned int task_scan_max(struct task_struct *p)
858 unsigned int smin = task_scan_min(p);
861 /* Watch for min being lower than max due to floor calculations */
862 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
863 return max(smin, smax);
866 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
868 rq->nr_numa_running += (p->numa_preferred_nid != -1);
869 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
872 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
874 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
875 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
881 spinlock_t lock; /* nr_tasks, tasks */
886 nodemask_t active_nodes;
887 unsigned long total_faults;
889 * Faults_cpu is used to decide whether memory should move
890 * towards the CPU. As a consequence, these stats are weighted
891 * more by CPU use than by memory faults.
893 unsigned long *faults_cpu;
894 unsigned long faults[0];
897 /* Shared or private faults. */
898 #define NR_NUMA_HINT_FAULT_TYPES 2
900 /* Memory and CPU locality */
901 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
903 /* Averaged statistics, and temporary buffers. */
904 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
906 pid_t task_numa_group_id(struct task_struct *p)
908 return p->numa_group ? p->numa_group->gid : 0;
912 * The averaged statistics, shared & private, memory & cpu,
913 * occupy the first half of the array. The second half of the
914 * array is for current counters, which are averaged into the
915 * first set by task_numa_placement.
917 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
919 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
922 static inline unsigned long task_faults(struct task_struct *p, int nid)
927 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
928 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
931 static inline unsigned long group_faults(struct task_struct *p, int nid)
936 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
937 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
940 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
942 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
943 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
946 /* Handle placement on systems where not all nodes are directly connected. */
947 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
948 int maxdist, bool task)
950 unsigned long score = 0;
954 * All nodes are directly connected, and the same distance
955 * from each other. No need for fancy placement algorithms.
957 if (sched_numa_topology_type == NUMA_DIRECT)
961 * This code is called for each node, introducing N^2 complexity,
962 * which should be ok given the number of nodes rarely exceeds 8.
964 for_each_online_node(node) {
965 unsigned long faults;
966 int dist = node_distance(nid, node);
969 * The furthest away nodes in the system are not interesting
970 * for placement; nid was already counted.
972 if (dist == sched_max_numa_distance || node == nid)
976 * On systems with a backplane NUMA topology, compare groups
977 * of nodes, and move tasks towards the group with the most
978 * memory accesses. When comparing two nodes at distance
979 * "hoplimit", only nodes closer by than "hoplimit" are part
980 * of each group. Skip other nodes.
982 if (sched_numa_topology_type == NUMA_BACKPLANE &&
986 /* Add up the faults from nearby nodes. */
988 faults = task_faults(p, node);
990 faults = group_faults(p, node);
993 * On systems with a glueless mesh NUMA topology, there are
994 * no fixed "groups of nodes". Instead, nodes that are not
995 * directly connected bounce traffic through intermediate
996 * nodes; a numa_group can occupy any set of nodes.
997 * The further away a node is, the less the faults count.
998 * This seems to result in good task placement.
1000 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1001 faults *= (sched_max_numa_distance - dist);
1002 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1012 * These return the fraction of accesses done by a particular task, or
1013 * task group, on a particular numa node. The group weight is given a
1014 * larger multiplier, in order to group tasks together that are almost
1015 * evenly spread out between numa nodes.
1017 static inline unsigned long task_weight(struct task_struct *p, int nid,
1020 unsigned long faults, total_faults;
1022 if (!p->numa_faults)
1025 total_faults = p->total_numa_faults;
1030 faults = task_faults(p, nid);
1031 faults += score_nearby_nodes(p, nid, dist, true);
1033 return 1000 * faults / total_faults;
1036 static inline unsigned long group_weight(struct task_struct *p, int nid,
1039 unsigned long faults, total_faults;
1044 total_faults = p->numa_group->total_faults;
1049 faults = group_faults(p, nid);
1050 faults += score_nearby_nodes(p, nid, dist, false);
1052 return 1000 * faults / total_faults;
1055 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1056 int src_nid, int dst_cpu)
1058 struct numa_group *ng = p->numa_group;
1059 int dst_nid = cpu_to_node(dst_cpu);
1060 int last_cpupid, this_cpupid;
1062 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1065 * Multi-stage node selection is used in conjunction with a periodic
1066 * migration fault to build a temporal task<->page relation. By using
1067 * a two-stage filter we remove short/unlikely relations.
1069 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1070 * a task's usage of a particular page (n_p) per total usage of this
1071 * page (n_t) (in a given time-span) to a probability.
1073 * Our periodic faults will sample this probability and getting the
1074 * same result twice in a row, given these samples are fully
1075 * independent, is then given by P(n)^2, provided our sample period
1076 * is sufficiently short compared to the usage pattern.
1078 * This quadric squishes small probabilities, making it less likely we
1079 * act on an unlikely task<->page relation.
1081 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1082 if (!cpupid_pid_unset(last_cpupid) &&
1083 cpupid_to_nid(last_cpupid) != dst_nid)
1086 /* Always allow migrate on private faults */
1087 if (cpupid_match_pid(p, last_cpupid))
1090 /* A shared fault, but p->numa_group has not been set up yet. */
1095 * Do not migrate if the destination is not a node that
1096 * is actively used by this numa group.
1098 if (!node_isset(dst_nid, ng->active_nodes))
1102 * Source is a node that is not actively used by this
1103 * numa group, while the destination is. Migrate.
1105 if (!node_isset(src_nid, ng->active_nodes))
1109 * Both source and destination are nodes in active
1110 * use by this numa group. Maximize memory bandwidth
1111 * by migrating from more heavily used groups, to less
1112 * heavily used ones, spreading the load around.
1113 * Use a 1/4 hysteresis to avoid spurious page movement.
1115 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1118 static unsigned long weighted_cpuload(const int cpu);
1119 static unsigned long source_load(int cpu, int type);
1120 static unsigned long target_load(int cpu, int type);
1121 static unsigned long capacity_of(int cpu);
1122 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1124 /* Cached statistics for all CPUs within a node */
1126 unsigned long nr_running;
1129 /* Total compute capacity of CPUs on a node */
1130 unsigned long compute_capacity;
1132 /* Approximate capacity in terms of runnable tasks on a node */
1133 unsigned long task_capacity;
1134 int has_free_capacity;
1138 * XXX borrowed from update_sg_lb_stats
1140 static void update_numa_stats(struct numa_stats *ns, int nid)
1142 int smt, cpu, cpus = 0;
1143 unsigned long capacity;
1145 memset(ns, 0, sizeof(*ns));
1146 for_each_cpu(cpu, cpumask_of_node(nid)) {
1147 struct rq *rq = cpu_rq(cpu);
1149 ns->nr_running += rq->nr_running;
1150 ns->load += weighted_cpuload(cpu);
1151 ns->compute_capacity += capacity_of(cpu);
1157 * If we raced with hotplug and there are no CPUs left in our mask
1158 * the @ns structure is NULL'ed and task_numa_compare() will
1159 * not find this node attractive.
1161 * We'll either bail at !has_free_capacity, or we'll detect a huge
1162 * imbalance and bail there.
1167 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1168 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1169 capacity = cpus / smt; /* cores */
1171 ns->task_capacity = min_t(unsigned, capacity,
1172 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1173 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1176 struct task_numa_env {
1177 struct task_struct *p;
1179 int src_cpu, src_nid;
1180 int dst_cpu, dst_nid;
1182 struct numa_stats src_stats, dst_stats;
1187 struct task_struct *best_task;
1192 static void task_numa_assign(struct task_numa_env *env,
1193 struct task_struct *p, long imp)
1196 put_task_struct(env->best_task);
1201 env->best_imp = imp;
1202 env->best_cpu = env->dst_cpu;
1205 static bool load_too_imbalanced(long src_load, long dst_load,
1206 struct task_numa_env *env)
1209 long orig_src_load, orig_dst_load;
1210 long src_capacity, dst_capacity;
1213 * The load is corrected for the CPU capacity available on each node.
1216 * ------------ vs ---------
1217 * src_capacity dst_capacity
1219 src_capacity = env->src_stats.compute_capacity;
1220 dst_capacity = env->dst_stats.compute_capacity;
1222 /* We care about the slope of the imbalance, not the direction. */
1223 if (dst_load < src_load)
1224 swap(dst_load, src_load);
1226 /* Is the difference below the threshold? */
1227 imb = dst_load * src_capacity * 100 -
1228 src_load * dst_capacity * env->imbalance_pct;
1233 * The imbalance is above the allowed threshold.
1234 * Compare it with the old imbalance.
1236 orig_src_load = env->src_stats.load;
1237 orig_dst_load = env->dst_stats.load;
1239 if (orig_dst_load < orig_src_load)
1240 swap(orig_dst_load, orig_src_load);
1242 old_imb = orig_dst_load * src_capacity * 100 -
1243 orig_src_load * dst_capacity * env->imbalance_pct;
1245 /* Would this change make things worse? */
1246 return (imb > old_imb);
1250 * This checks if the overall compute and NUMA accesses of the system would
1251 * be improved if the source tasks was migrated to the target dst_cpu taking
1252 * into account that it might be best if task running on the dst_cpu should
1253 * be exchanged with the source task
1255 static void task_numa_compare(struct task_numa_env *env,
1256 long taskimp, long groupimp)
1258 struct rq *src_rq = cpu_rq(env->src_cpu);
1259 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1260 struct task_struct *cur;
1261 long src_load, dst_load;
1263 long imp = env->p->numa_group ? groupimp : taskimp;
1265 int dist = env->dist;
1269 raw_spin_lock_irq(&dst_rq->lock);
1272 * No need to move the exiting task, and this ensures that ->curr
1273 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1274 * is safe under RCU read lock.
1275 * Note that rcu_read_lock() itself can't protect from the final
1276 * put_task_struct() after the last schedule().
1278 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1280 raw_spin_unlock_irq(&dst_rq->lock);
1283 * Because we have preemption enabled we can get migrated around and
1284 * end try selecting ourselves (current == env->p) as a swap candidate.
1290 * "imp" is the fault differential for the source task between the
1291 * source and destination node. Calculate the total differential for
1292 * the source task and potential destination task. The more negative
1293 * the value is, the more rmeote accesses that would be expected to
1294 * be incurred if the tasks were swapped.
1297 /* Skip this swap candidate if cannot move to the source cpu */
1298 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1302 * If dst and source tasks are in the same NUMA group, or not
1303 * in any group then look only at task weights.
1305 if (cur->numa_group == env->p->numa_group) {
1306 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1307 task_weight(cur, env->dst_nid, dist);
1309 * Add some hysteresis to prevent swapping the
1310 * tasks within a group over tiny differences.
1312 if (cur->numa_group)
1316 * Compare the group weights. If a task is all by
1317 * itself (not part of a group), use the task weight
1320 if (cur->numa_group)
1321 imp += group_weight(cur, env->src_nid, dist) -
1322 group_weight(cur, env->dst_nid, dist);
1324 imp += task_weight(cur, env->src_nid, dist) -
1325 task_weight(cur, env->dst_nid, dist);
1329 if (imp <= env->best_imp && moveimp <= env->best_imp)
1333 /* Is there capacity at our destination? */
1334 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1335 !env->dst_stats.has_free_capacity)
1341 /* Balance doesn't matter much if we're running a task per cpu */
1342 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1343 dst_rq->nr_running == 1)
1347 * In the overloaded case, try and keep the load balanced.
1350 load = task_h_load(env->p);
1351 dst_load = env->dst_stats.load + load;
1352 src_load = env->src_stats.load - load;
1354 if (moveimp > imp && moveimp > env->best_imp) {
1356 * If the improvement from just moving env->p direction is
1357 * better than swapping tasks around, check if a move is
1358 * possible. Store a slightly smaller score than moveimp,
1359 * so an actually idle CPU will win.
1361 if (!load_too_imbalanced(src_load, dst_load, env)) {
1368 if (imp <= env->best_imp)
1372 load = task_h_load(cur);
1377 if (load_too_imbalanced(src_load, dst_load, env))
1381 * One idle CPU per node is evaluated for a task numa move.
1382 * Call select_idle_sibling to maybe find a better one.
1385 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1388 task_numa_assign(env, cur, imp);
1393 static void task_numa_find_cpu(struct task_numa_env *env,
1394 long taskimp, long groupimp)
1398 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1399 /* Skip this CPU if the source task cannot migrate */
1400 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1404 task_numa_compare(env, taskimp, groupimp);
1408 /* Only move tasks to a NUMA node less busy than the current node. */
1409 static bool numa_has_capacity(struct task_numa_env *env)
1411 struct numa_stats *src = &env->src_stats;
1412 struct numa_stats *dst = &env->dst_stats;
1414 if (src->has_free_capacity && !dst->has_free_capacity)
1418 * Only consider a task move if the source has a higher load
1419 * than the destination, corrected for CPU capacity on each node.
1421 * src->load dst->load
1422 * --------------------- vs ---------------------
1423 * src->compute_capacity dst->compute_capacity
1425 if (src->load * dst->compute_capacity * env->imbalance_pct >
1427 dst->load * src->compute_capacity * 100)
1433 static int task_numa_migrate(struct task_struct *p)
1435 struct task_numa_env env = {
1438 .src_cpu = task_cpu(p),
1439 .src_nid = task_node(p),
1441 .imbalance_pct = 112,
1447 struct sched_domain *sd;
1448 unsigned long taskweight, groupweight;
1450 long taskimp, groupimp;
1453 * Pick the lowest SD_NUMA domain, as that would have the smallest
1454 * imbalance and would be the first to start moving tasks about.
1456 * And we want to avoid any moving of tasks about, as that would create
1457 * random movement of tasks -- counter the numa conditions we're trying
1461 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1463 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1467 * Cpusets can break the scheduler domain tree into smaller
1468 * balance domains, some of which do not cross NUMA boundaries.
1469 * Tasks that are "trapped" in such domains cannot be migrated
1470 * elsewhere, so there is no point in (re)trying.
1472 if (unlikely(!sd)) {
1473 p->numa_preferred_nid = task_node(p);
1477 env.dst_nid = p->numa_preferred_nid;
1478 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1479 taskweight = task_weight(p, env.src_nid, dist);
1480 groupweight = group_weight(p, env.src_nid, dist);
1481 update_numa_stats(&env.src_stats, env.src_nid);
1482 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1483 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1484 update_numa_stats(&env.dst_stats, env.dst_nid);
1486 /* Try to find a spot on the preferred nid. */
1487 if (numa_has_capacity(&env))
1488 task_numa_find_cpu(&env, taskimp, groupimp);
1491 * Look at other nodes in these cases:
1492 * - there is no space available on the preferred_nid
1493 * - the task is part of a numa_group that is interleaved across
1494 * multiple NUMA nodes; in order to better consolidate the group,
1495 * we need to check other locations.
1497 if (env.best_cpu == -1 || (p->numa_group &&
1498 nodes_weight(p->numa_group->active_nodes) > 1)) {
1499 for_each_online_node(nid) {
1500 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1503 dist = node_distance(env.src_nid, env.dst_nid);
1504 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1506 taskweight = task_weight(p, env.src_nid, dist);
1507 groupweight = group_weight(p, env.src_nid, dist);
1510 /* Only consider nodes where both task and groups benefit */
1511 taskimp = task_weight(p, nid, dist) - taskweight;
1512 groupimp = group_weight(p, nid, dist) - groupweight;
1513 if (taskimp < 0 && groupimp < 0)
1518 update_numa_stats(&env.dst_stats, env.dst_nid);
1519 if (numa_has_capacity(&env))
1520 task_numa_find_cpu(&env, taskimp, groupimp);
1525 * If the task is part of a workload that spans multiple NUMA nodes,
1526 * and is migrating into one of the workload's active nodes, remember
1527 * this node as the task's preferred numa node, so the workload can
1529 * A task that migrated to a second choice node will be better off
1530 * trying for a better one later. Do not set the preferred node here.
1532 if (p->numa_group) {
1533 if (env.best_cpu == -1)
1538 if (node_isset(nid, p->numa_group->active_nodes))
1539 sched_setnuma(p, env.dst_nid);
1542 /* No better CPU than the current one was found. */
1543 if (env.best_cpu == -1)
1547 * Reset the scan period if the task is being rescheduled on an
1548 * alternative node to recheck if the tasks is now properly placed.
1550 p->numa_scan_period = task_scan_min(p);
1552 if (env.best_task == NULL) {
1553 ret = migrate_task_to(p, env.best_cpu);
1555 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1559 ret = migrate_swap(p, env.best_task);
1561 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1562 put_task_struct(env.best_task);
1566 /* Attempt to migrate a task to a CPU on the preferred node. */
1567 static void numa_migrate_preferred(struct task_struct *p)
1569 unsigned long interval = HZ;
1571 /* This task has no NUMA fault statistics yet */
1572 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1575 /* Periodically retry migrating the task to the preferred node */
1576 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1577 p->numa_migrate_retry = jiffies + interval;
1579 /* Success if task is already running on preferred CPU */
1580 if (task_node(p) == p->numa_preferred_nid)
1583 /* Otherwise, try migrate to a CPU on the preferred node */
1584 task_numa_migrate(p);
1588 * Find the nodes on which the workload is actively running. We do this by
1589 * tracking the nodes from which NUMA hinting faults are triggered. This can
1590 * be different from the set of nodes where the workload's memory is currently
1593 * The bitmask is used to make smarter decisions on when to do NUMA page
1594 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1595 * are added when they cause over 6/16 of the maximum number of faults, but
1596 * only removed when they drop below 3/16.
1598 static void update_numa_active_node_mask(struct numa_group *numa_group)
1600 unsigned long faults, max_faults = 0;
1603 for_each_online_node(nid) {
1604 faults = group_faults_cpu(numa_group, nid);
1605 if (faults > max_faults)
1606 max_faults = faults;
1609 for_each_online_node(nid) {
1610 faults = group_faults_cpu(numa_group, nid);
1611 if (!node_isset(nid, numa_group->active_nodes)) {
1612 if (faults > max_faults * 6 / 16)
1613 node_set(nid, numa_group->active_nodes);
1614 } else if (faults < max_faults * 3 / 16)
1615 node_clear(nid, numa_group->active_nodes);
1620 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1621 * increments. The more local the fault statistics are, the higher the scan
1622 * period will be for the next scan window. If local/(local+remote) ratio is
1623 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1624 * the scan period will decrease. Aim for 70% local accesses.
1626 #define NUMA_PERIOD_SLOTS 10
1627 #define NUMA_PERIOD_THRESHOLD 7
1630 * Increase the scan period (slow down scanning) if the majority of
1631 * our memory is already on our local node, or if the majority of
1632 * the page accesses are shared with other processes.
1633 * Otherwise, decrease the scan period.
1635 static void update_task_scan_period(struct task_struct *p,
1636 unsigned long shared, unsigned long private)
1638 unsigned int period_slot;
1642 unsigned long remote = p->numa_faults_locality[0];
1643 unsigned long local = p->numa_faults_locality[1];
1646 * If there were no record hinting faults then either the task is
1647 * completely idle or all activity is areas that are not of interest
1648 * to automatic numa balancing. Related to that, if there were failed
1649 * migration then it implies we are migrating too quickly or the local
1650 * node is overloaded. In either case, scan slower
1652 if (local + shared == 0 || p->numa_faults_locality[2]) {
1653 p->numa_scan_period = min(p->numa_scan_period_max,
1654 p->numa_scan_period << 1);
1656 p->mm->numa_next_scan = jiffies +
1657 msecs_to_jiffies(p->numa_scan_period);
1663 * Prepare to scale scan period relative to the current period.
1664 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1665 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1666 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1668 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1669 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1670 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1671 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1674 diff = slot * period_slot;
1676 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1679 * Scale scan rate increases based on sharing. There is an
1680 * inverse relationship between the degree of sharing and
1681 * the adjustment made to the scanning period. Broadly
1682 * speaking the intent is that there is little point
1683 * scanning faster if shared accesses dominate as it may
1684 * simply bounce migrations uselessly
1686 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1687 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1690 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1691 task_scan_min(p), task_scan_max(p));
1692 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1696 * Get the fraction of time the task has been running since the last
1697 * NUMA placement cycle. The scheduler keeps similar statistics, but
1698 * decays those on a 32ms period, which is orders of magnitude off
1699 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1700 * stats only if the task is so new there are no NUMA statistics yet.
1702 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1704 u64 runtime, delta, now;
1705 /* Use the start of this time slice to avoid calculations. */
1706 now = p->se.exec_start;
1707 runtime = p->se.sum_exec_runtime;
1709 if (p->last_task_numa_placement) {
1710 delta = runtime - p->last_sum_exec_runtime;
1711 *period = now - p->last_task_numa_placement;
1713 delta = p->se.avg.load_sum / p->se.load.weight;
1714 *period = LOAD_AVG_MAX;
1717 p->last_sum_exec_runtime = runtime;
1718 p->last_task_numa_placement = now;
1724 * Determine the preferred nid for a task in a numa_group. This needs to
1725 * be done in a way that produces consistent results with group_weight,
1726 * otherwise workloads might not converge.
1728 static int preferred_group_nid(struct task_struct *p, int nid)
1733 /* Direct connections between all NUMA nodes. */
1734 if (sched_numa_topology_type == NUMA_DIRECT)
1738 * On a system with glueless mesh NUMA topology, group_weight
1739 * scores nodes according to the number of NUMA hinting faults on
1740 * both the node itself, and on nearby nodes.
1742 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1743 unsigned long score, max_score = 0;
1744 int node, max_node = nid;
1746 dist = sched_max_numa_distance;
1748 for_each_online_node(node) {
1749 score = group_weight(p, node, dist);
1750 if (score > max_score) {
1759 * Finding the preferred nid in a system with NUMA backplane
1760 * interconnect topology is more involved. The goal is to locate
1761 * tasks from numa_groups near each other in the system, and
1762 * untangle workloads from different sides of the system. This requires
1763 * searching down the hierarchy of node groups, recursively searching
1764 * inside the highest scoring group of nodes. The nodemask tricks
1765 * keep the complexity of the search down.
1767 nodes = node_online_map;
1768 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1769 unsigned long max_faults = 0;
1770 nodemask_t max_group = NODE_MASK_NONE;
1773 /* Are there nodes at this distance from each other? */
1774 if (!find_numa_distance(dist))
1777 for_each_node_mask(a, nodes) {
1778 unsigned long faults = 0;
1779 nodemask_t this_group;
1780 nodes_clear(this_group);
1782 /* Sum group's NUMA faults; includes a==b case. */
1783 for_each_node_mask(b, nodes) {
1784 if (node_distance(a, b) < dist) {
1785 faults += group_faults(p, b);
1786 node_set(b, this_group);
1787 node_clear(b, nodes);
1791 /* Remember the top group. */
1792 if (faults > max_faults) {
1793 max_faults = faults;
1794 max_group = this_group;
1796 * subtle: at the smallest distance there is
1797 * just one node left in each "group", the
1798 * winner is the preferred nid.
1803 /* Next round, evaluate the nodes within max_group. */
1811 static void task_numa_placement(struct task_struct *p)
1813 int seq, nid, max_nid = -1, max_group_nid = -1;
1814 unsigned long max_faults = 0, max_group_faults = 0;
1815 unsigned long fault_types[2] = { 0, 0 };
1816 unsigned long total_faults;
1817 u64 runtime, period;
1818 spinlock_t *group_lock = NULL;
1821 * The p->mm->numa_scan_seq field gets updated without
1822 * exclusive access. Use READ_ONCE() here to ensure
1823 * that the field is read in a single access:
1825 seq = READ_ONCE(p->mm->numa_scan_seq);
1826 if (p->numa_scan_seq == seq)
1828 p->numa_scan_seq = seq;
1829 p->numa_scan_period_max = task_scan_max(p);
1831 total_faults = p->numa_faults_locality[0] +
1832 p->numa_faults_locality[1];
1833 runtime = numa_get_avg_runtime(p, &period);
1835 /* If the task is part of a group prevent parallel updates to group stats */
1836 if (p->numa_group) {
1837 group_lock = &p->numa_group->lock;
1838 spin_lock_irq(group_lock);
1841 /* Find the node with the highest number of faults */
1842 for_each_online_node(nid) {
1843 /* Keep track of the offsets in numa_faults array */
1844 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1845 unsigned long faults = 0, group_faults = 0;
1848 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1849 long diff, f_diff, f_weight;
1851 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1852 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1853 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1854 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1856 /* Decay existing window, copy faults since last scan */
1857 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1858 fault_types[priv] += p->numa_faults[membuf_idx];
1859 p->numa_faults[membuf_idx] = 0;
1862 * Normalize the faults_from, so all tasks in a group
1863 * count according to CPU use, instead of by the raw
1864 * number of faults. Tasks with little runtime have
1865 * little over-all impact on throughput, and thus their
1866 * faults are less important.
1868 f_weight = div64_u64(runtime << 16, period + 1);
1869 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1871 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1872 p->numa_faults[cpubuf_idx] = 0;
1874 p->numa_faults[mem_idx] += diff;
1875 p->numa_faults[cpu_idx] += f_diff;
1876 faults += p->numa_faults[mem_idx];
1877 p->total_numa_faults += diff;
1878 if (p->numa_group) {
1880 * safe because we can only change our own group
1882 * mem_idx represents the offset for a given
1883 * nid and priv in a specific region because it
1884 * is at the beginning of the numa_faults array.
1886 p->numa_group->faults[mem_idx] += diff;
1887 p->numa_group->faults_cpu[mem_idx] += f_diff;
1888 p->numa_group->total_faults += diff;
1889 group_faults += p->numa_group->faults[mem_idx];
1893 if (faults > max_faults) {
1894 max_faults = faults;
1898 if (group_faults > max_group_faults) {
1899 max_group_faults = group_faults;
1900 max_group_nid = nid;
1904 update_task_scan_period(p, fault_types[0], fault_types[1]);
1906 if (p->numa_group) {
1907 update_numa_active_node_mask(p->numa_group);
1908 spin_unlock_irq(group_lock);
1909 max_nid = preferred_group_nid(p, max_group_nid);
1913 /* Set the new preferred node */
1914 if (max_nid != p->numa_preferred_nid)
1915 sched_setnuma(p, max_nid);
1917 if (task_node(p) != p->numa_preferred_nid)
1918 numa_migrate_preferred(p);
1922 static inline int get_numa_group(struct numa_group *grp)
1924 return atomic_inc_not_zero(&grp->refcount);
1927 static inline void put_numa_group(struct numa_group *grp)
1929 if (atomic_dec_and_test(&grp->refcount))
1930 kfree_rcu(grp, rcu);
1933 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1936 struct numa_group *grp, *my_grp;
1937 struct task_struct *tsk;
1939 int cpu = cpupid_to_cpu(cpupid);
1942 if (unlikely(!p->numa_group)) {
1943 unsigned int size = sizeof(struct numa_group) +
1944 4*nr_node_ids*sizeof(unsigned long);
1946 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1950 atomic_set(&grp->refcount, 1);
1951 spin_lock_init(&grp->lock);
1953 /* Second half of the array tracks nids where faults happen */
1954 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1957 node_set(task_node(current), grp->active_nodes);
1959 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1960 grp->faults[i] = p->numa_faults[i];
1962 grp->total_faults = p->total_numa_faults;
1965 rcu_assign_pointer(p->numa_group, grp);
1969 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1971 if (!cpupid_match_pid(tsk, cpupid))
1974 grp = rcu_dereference(tsk->numa_group);
1978 my_grp = p->numa_group;
1983 * Only join the other group if its bigger; if we're the bigger group,
1984 * the other task will join us.
1986 if (my_grp->nr_tasks > grp->nr_tasks)
1990 * Tie-break on the grp address.
1992 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1995 /* Always join threads in the same process. */
1996 if (tsk->mm == current->mm)
1999 /* Simple filter to avoid false positives due to PID collisions */
2000 if (flags & TNF_SHARED)
2003 /* Update priv based on whether false sharing was detected */
2006 if (join && !get_numa_group(grp))
2014 BUG_ON(irqs_disabled());
2015 double_lock_irq(&my_grp->lock, &grp->lock);
2017 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2018 my_grp->faults[i] -= p->numa_faults[i];
2019 grp->faults[i] += p->numa_faults[i];
2021 my_grp->total_faults -= p->total_numa_faults;
2022 grp->total_faults += p->total_numa_faults;
2027 spin_unlock(&my_grp->lock);
2028 spin_unlock_irq(&grp->lock);
2030 rcu_assign_pointer(p->numa_group, grp);
2032 put_numa_group(my_grp);
2040 void task_numa_free(struct task_struct *p)
2042 struct numa_group *grp = p->numa_group;
2043 void *numa_faults = p->numa_faults;
2044 unsigned long flags;
2048 spin_lock_irqsave(&grp->lock, flags);
2049 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2050 grp->faults[i] -= p->numa_faults[i];
2051 grp->total_faults -= p->total_numa_faults;
2054 spin_unlock_irqrestore(&grp->lock, flags);
2055 RCU_INIT_POINTER(p->numa_group, NULL);
2056 put_numa_group(grp);
2059 p->numa_faults = NULL;
2064 * Got a PROT_NONE fault for a page on @node.
2066 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2068 struct task_struct *p = current;
2069 bool migrated = flags & TNF_MIGRATED;
2070 int cpu_node = task_node(current);
2071 int local = !!(flags & TNF_FAULT_LOCAL);
2074 if (!static_branch_likely(&sched_numa_balancing))
2077 /* for example, ksmd faulting in a user's mm */
2081 /* Allocate buffer to track faults on a per-node basis */
2082 if (unlikely(!p->numa_faults)) {
2083 int size = sizeof(*p->numa_faults) *
2084 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2086 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2087 if (!p->numa_faults)
2090 p->total_numa_faults = 0;
2091 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2095 * First accesses are treated as private, otherwise consider accesses
2096 * to be private if the accessing pid has not changed
2098 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2101 priv = cpupid_match_pid(p, last_cpupid);
2102 if (!priv && !(flags & TNF_NO_GROUP))
2103 task_numa_group(p, last_cpupid, flags, &priv);
2107 * If a workload spans multiple NUMA nodes, a shared fault that
2108 * occurs wholly within the set of nodes that the workload is
2109 * actively using should be counted as local. This allows the
2110 * scan rate to slow down when a workload has settled down.
2112 if (!priv && !local && p->numa_group &&
2113 node_isset(cpu_node, p->numa_group->active_nodes) &&
2114 node_isset(mem_node, p->numa_group->active_nodes))
2117 task_numa_placement(p);
2120 * Retry task to preferred node migration periodically, in case it
2121 * case it previously failed, or the scheduler moved us.
2123 if (time_after(jiffies, p->numa_migrate_retry))
2124 numa_migrate_preferred(p);
2127 p->numa_pages_migrated += pages;
2128 if (flags & TNF_MIGRATE_FAIL)
2129 p->numa_faults_locality[2] += pages;
2131 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2132 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2133 p->numa_faults_locality[local] += pages;
2136 static void reset_ptenuma_scan(struct task_struct *p)
2139 * We only did a read acquisition of the mmap sem, so
2140 * p->mm->numa_scan_seq is written to without exclusive access
2141 * and the update is not guaranteed to be atomic. That's not
2142 * much of an issue though, since this is just used for
2143 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2144 * expensive, to avoid any form of compiler optimizations:
2146 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2147 p->mm->numa_scan_offset = 0;
2151 * The expensive part of numa migration is done from task_work context.
2152 * Triggered from task_tick_numa().
2154 void task_numa_work(struct callback_head *work)
2156 unsigned long migrate, next_scan, now = jiffies;
2157 struct task_struct *p = current;
2158 struct mm_struct *mm = p->mm;
2159 struct vm_area_struct *vma;
2160 unsigned long start, end;
2161 unsigned long nr_pte_updates = 0;
2162 long pages, virtpages;
2164 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2166 work->next = work; /* protect against double add */
2168 * Who cares about NUMA placement when they're dying.
2170 * NOTE: make sure not to dereference p->mm before this check,
2171 * exit_task_work() happens _after_ exit_mm() so we could be called
2172 * without p->mm even though we still had it when we enqueued this
2175 if (p->flags & PF_EXITING)
2178 if (!mm->numa_next_scan) {
2179 mm->numa_next_scan = now +
2180 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2184 * Enforce maximal scan/migration frequency..
2186 migrate = mm->numa_next_scan;
2187 if (time_before(now, migrate))
2190 if (p->numa_scan_period == 0) {
2191 p->numa_scan_period_max = task_scan_max(p);
2192 p->numa_scan_period = task_scan_min(p);
2195 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2196 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2200 * Delay this task enough that another task of this mm will likely win
2201 * the next time around.
2203 p->node_stamp += 2 * TICK_NSEC;
2205 start = mm->numa_scan_offset;
2206 pages = sysctl_numa_balancing_scan_size;
2207 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2208 virtpages = pages * 8; /* Scan up to this much virtual space */
2213 down_read(&mm->mmap_sem);
2214 vma = find_vma(mm, start);
2216 reset_ptenuma_scan(p);
2220 for (; vma; vma = vma->vm_next) {
2221 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2222 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2227 * Shared library pages mapped by multiple processes are not
2228 * migrated as it is expected they are cache replicated. Avoid
2229 * hinting faults in read-only file-backed mappings or the vdso
2230 * as migrating the pages will be of marginal benefit.
2233 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2237 * Skip inaccessible VMAs to avoid any confusion between
2238 * PROT_NONE and NUMA hinting ptes
2240 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2244 start = max(start, vma->vm_start);
2245 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2246 end = min(end, vma->vm_end);
2247 nr_pte_updates = change_prot_numa(vma, start, end);
2250 * Try to scan sysctl_numa_balancing_size worth of
2251 * hpages that have at least one present PTE that
2252 * is not already pte-numa. If the VMA contains
2253 * areas that are unused or already full of prot_numa
2254 * PTEs, scan up to virtpages, to skip through those
2258 pages -= (end - start) >> PAGE_SHIFT;
2259 virtpages -= (end - start) >> PAGE_SHIFT;
2262 if (pages <= 0 || virtpages <= 0)
2266 } while (end != vma->vm_end);
2271 * It is possible to reach the end of the VMA list but the last few
2272 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2273 * would find the !migratable VMA on the next scan but not reset the
2274 * scanner to the start so check it now.
2277 mm->numa_scan_offset = start;
2279 reset_ptenuma_scan(p);
2280 up_read(&mm->mmap_sem);
2284 * Drive the periodic memory faults..
2286 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2288 struct callback_head *work = &curr->numa_work;
2292 * We don't care about NUMA placement if we don't have memory.
2294 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2298 * Using runtime rather than walltime has the dual advantage that
2299 * we (mostly) drive the selection from busy threads and that the
2300 * task needs to have done some actual work before we bother with
2303 now = curr->se.sum_exec_runtime;
2304 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2306 if (now > curr->node_stamp + period) {
2307 if (!curr->node_stamp)
2308 curr->numa_scan_period = task_scan_min(curr);
2309 curr->node_stamp += period;
2311 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2312 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2313 task_work_add(curr, work, true);
2318 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2322 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2326 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2329 #endif /* CONFIG_NUMA_BALANCING */
2332 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2334 update_load_add(&cfs_rq->load, se->load.weight);
2335 if (!parent_entity(se))
2336 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2338 if (entity_is_task(se)) {
2339 struct rq *rq = rq_of(cfs_rq);
2341 account_numa_enqueue(rq, task_of(se));
2342 list_add(&se->group_node, &rq->cfs_tasks);
2345 cfs_rq->nr_running++;
2349 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2351 update_load_sub(&cfs_rq->load, se->load.weight);
2352 if (!parent_entity(se))
2353 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2354 if (entity_is_task(se)) {
2355 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2356 list_del_init(&se->group_node);
2358 cfs_rq->nr_running--;
2361 #ifdef CONFIG_FAIR_GROUP_SCHED
2363 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2368 * Use this CPU's real-time load instead of the last load contribution
2369 * as the updating of the contribution is delayed, and we will use the
2370 * the real-time load to calc the share. See update_tg_load_avg().
2372 tg_weight = atomic_long_read(&tg->load_avg);
2373 tg_weight -= cfs_rq->tg_load_avg_contrib;
2374 tg_weight += cfs_rq->load.weight;
2379 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2381 long tg_weight, load, shares;
2383 tg_weight = calc_tg_weight(tg, cfs_rq);
2384 load = cfs_rq->load.weight;
2386 shares = (tg->shares * load);
2388 shares /= tg_weight;
2390 if (shares < MIN_SHARES)
2391 shares = MIN_SHARES;
2392 if (shares > tg->shares)
2393 shares = tg->shares;
2397 # else /* CONFIG_SMP */
2398 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2402 # endif /* CONFIG_SMP */
2403 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2404 unsigned long weight)
2407 /* commit outstanding execution time */
2408 if (cfs_rq->curr == se)
2409 update_curr(cfs_rq);
2410 account_entity_dequeue(cfs_rq, se);
2413 update_load_set(&se->load, weight);
2416 account_entity_enqueue(cfs_rq, se);
2419 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2421 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2423 struct task_group *tg;
2424 struct sched_entity *se;
2428 se = tg->se[cpu_of(rq_of(cfs_rq))];
2429 if (!se || throttled_hierarchy(cfs_rq))
2432 if (likely(se->load.weight == tg->shares))
2435 shares = calc_cfs_shares(cfs_rq, tg);
2437 reweight_entity(cfs_rq_of(se), se, shares);
2439 #else /* CONFIG_FAIR_GROUP_SCHED */
2440 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2443 #endif /* CONFIG_FAIR_GROUP_SCHED */
2446 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2447 static const u32 runnable_avg_yN_inv[] = {
2448 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2449 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2450 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2451 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2452 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2453 0x85aac367, 0x82cd8698,
2457 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2458 * over-estimates when re-combining.
2460 static const u32 runnable_avg_yN_sum[] = {
2461 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2462 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2463 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2468 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2470 static __always_inline u64 decay_load(u64 val, u64 n)
2472 unsigned int local_n;
2476 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2479 /* after bounds checking we can collapse to 32-bit */
2483 * As y^PERIOD = 1/2, we can combine
2484 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2485 * With a look-up table which covers y^n (n<PERIOD)
2487 * To achieve constant time decay_load.
2489 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2490 val >>= local_n / LOAD_AVG_PERIOD;
2491 local_n %= LOAD_AVG_PERIOD;
2494 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2499 * For updates fully spanning n periods, the contribution to runnable
2500 * average will be: \Sum 1024*y^n
2502 * We can compute this reasonably efficiently by combining:
2503 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2505 static u32 __compute_runnable_contrib(u64 n)
2509 if (likely(n <= LOAD_AVG_PERIOD))
2510 return runnable_avg_yN_sum[n];
2511 else if (unlikely(n >= LOAD_AVG_MAX_N))
2512 return LOAD_AVG_MAX;
2514 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2516 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2517 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2519 n -= LOAD_AVG_PERIOD;
2520 } while (n > LOAD_AVG_PERIOD);
2522 contrib = decay_load(contrib, n);
2523 return contrib + runnable_avg_yN_sum[n];
2526 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2527 #error "load tracking assumes 2^10 as unit"
2530 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2533 * We can represent the historical contribution to runnable average as the
2534 * coefficients of a geometric series. To do this we sub-divide our runnable
2535 * history into segments of approximately 1ms (1024us); label the segment that
2536 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2538 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2540 * (now) (~1ms ago) (~2ms ago)
2542 * Let u_i denote the fraction of p_i that the entity was runnable.
2544 * We then designate the fractions u_i as our co-efficients, yielding the
2545 * following representation of historical load:
2546 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2548 * We choose y based on the with of a reasonably scheduling period, fixing:
2551 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2552 * approximately half as much as the contribution to load within the last ms
2555 * When a period "rolls over" and we have new u_0`, multiplying the previous
2556 * sum again by y is sufficient to update:
2557 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2558 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2560 static __always_inline int
2561 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2562 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2564 u64 delta, scaled_delta, periods;
2566 unsigned int delta_w, scaled_delta_w, decayed = 0;
2567 unsigned long scale_freq, scale_cpu;
2569 delta = now - sa->last_update_time;
2571 * This should only happen when time goes backwards, which it
2572 * unfortunately does during sched clock init when we swap over to TSC.
2574 if ((s64)delta < 0) {
2575 sa->last_update_time = now;
2580 * Use 1024ns as the unit of measurement since it's a reasonable
2581 * approximation of 1us and fast to compute.
2586 sa->last_update_time = now;
2588 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2589 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2590 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2592 /* delta_w is the amount already accumulated against our next period */
2593 delta_w = sa->period_contrib;
2594 if (delta + delta_w >= 1024) {
2597 /* how much left for next period will start over, we don't know yet */
2598 sa->period_contrib = 0;
2601 * Now that we know we're crossing a period boundary, figure
2602 * out how much from delta we need to complete the current
2603 * period and accrue it.
2605 delta_w = 1024 - delta_w;
2606 scaled_delta_w = cap_scale(delta_w, scale_freq);
2608 sa->load_sum += weight * scaled_delta_w;
2610 cfs_rq->runnable_load_sum +=
2611 weight * scaled_delta_w;
2615 sa->util_sum += scaled_delta_w * scale_cpu;
2619 /* Figure out how many additional periods this update spans */
2620 periods = delta / 1024;
2623 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2625 cfs_rq->runnable_load_sum =
2626 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2628 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2630 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2631 contrib = __compute_runnable_contrib(periods);
2632 contrib = cap_scale(contrib, scale_freq);
2634 sa->load_sum += weight * contrib;
2636 cfs_rq->runnable_load_sum += weight * contrib;
2639 sa->util_sum += contrib * scale_cpu;
2642 /* Remainder of delta accrued against u_0` */
2643 scaled_delta = cap_scale(delta, scale_freq);
2645 sa->load_sum += weight * scaled_delta;
2647 cfs_rq->runnable_load_sum += weight * scaled_delta;
2650 sa->util_sum += scaled_delta * scale_cpu;
2652 sa->period_contrib += delta;
2655 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2657 cfs_rq->runnable_load_avg =
2658 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2660 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2666 #ifdef CONFIG_FAIR_GROUP_SCHED
2668 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2669 * and effective_load (which is not done because it is too costly).
2671 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2673 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2675 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2676 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2677 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2681 #else /* CONFIG_FAIR_GROUP_SCHED */
2682 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2683 #endif /* CONFIG_FAIR_GROUP_SCHED */
2685 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2687 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2688 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2690 struct sched_avg *sa = &cfs_rq->avg;
2691 int decayed, removed = 0;
2693 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2694 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2695 sa->load_avg = max_t(long, sa->load_avg - r, 0);
2696 sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
2700 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2701 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2702 sa->util_avg = max_t(long, sa->util_avg - r, 0);
2703 sa->util_sum = max_t(s32, sa->util_sum - r * LOAD_AVG_MAX, 0);
2706 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2707 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2709 #ifndef CONFIG_64BIT
2711 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2714 return decayed || removed;
2717 /* Update task and its cfs_rq load average */
2718 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2720 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2721 u64 now = cfs_rq_clock_task(cfs_rq);
2722 int cpu = cpu_of(rq_of(cfs_rq));
2725 * Track task load average for carrying it to new CPU after migrated, and
2726 * track group sched_entity load average for task_h_load calc in migration
2728 __update_load_avg(now, cpu, &se->avg,
2729 se->on_rq * scale_load_down(se->load.weight),
2730 cfs_rq->curr == se, NULL);
2732 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2733 update_tg_load_avg(cfs_rq, 0);
2735 if (entity_is_task(se))
2736 trace_sched_load_avg_task(task_of(se), &se->avg);
2737 trace_sched_load_avg_cpu(cpu, cfs_rq);
2740 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2742 if (!sched_feat(ATTACH_AGE_LOAD))
2746 * If we got migrated (either between CPUs or between cgroups) we'll
2747 * have aged the average right before clearing @last_update_time.
2749 if (se->avg.last_update_time) {
2750 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2751 &se->avg, 0, 0, NULL);
2754 * XXX: we could have just aged the entire load away if we've been
2755 * absent from the fair class for too long.
2760 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2761 cfs_rq->avg.load_avg += se->avg.load_avg;
2762 cfs_rq->avg.load_sum += se->avg.load_sum;
2763 cfs_rq->avg.util_avg += se->avg.util_avg;
2764 cfs_rq->avg.util_sum += se->avg.util_sum;
2767 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2769 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2770 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2771 cfs_rq->curr == se, NULL);
2773 cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
2774 cfs_rq->avg.load_sum = max_t(s64, cfs_rq->avg.load_sum - se->avg.load_sum, 0);
2775 cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
2776 cfs_rq->avg.util_sum = max_t(s32, cfs_rq->avg.util_sum - se->avg.util_sum, 0);
2779 /* Add the load generated by se into cfs_rq's load average */
2781 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2783 struct sched_avg *sa = &se->avg;
2784 u64 now = cfs_rq_clock_task(cfs_rq);
2785 int migrated, decayed;
2787 migrated = !sa->last_update_time;
2789 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2790 se->on_rq * scale_load_down(se->load.weight),
2791 cfs_rq->curr == se, NULL);
2794 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2796 cfs_rq->runnable_load_avg += sa->load_avg;
2797 cfs_rq->runnable_load_sum += sa->load_sum;
2800 attach_entity_load_avg(cfs_rq, se);
2802 if (decayed || migrated)
2803 update_tg_load_avg(cfs_rq, 0);
2806 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2808 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2810 update_load_avg(se, 1);
2812 cfs_rq->runnable_load_avg =
2813 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2814 cfs_rq->runnable_load_sum =
2815 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2818 #ifndef CONFIG_64BIT
2819 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2821 u64 last_update_time_copy;
2822 u64 last_update_time;
2825 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2827 last_update_time = cfs_rq->avg.last_update_time;
2828 } while (last_update_time != last_update_time_copy);
2830 return last_update_time;
2833 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2835 return cfs_rq->avg.last_update_time;
2840 * Task first catches up with cfs_rq, and then subtract
2841 * itself from the cfs_rq (task must be off the queue now).
2843 void remove_entity_load_avg(struct sched_entity *se)
2845 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2846 u64 last_update_time;
2849 * Newly created task or never used group entity should not be removed
2850 * from its (source) cfs_rq
2852 if (se->avg.last_update_time == 0)
2855 last_update_time = cfs_rq_last_update_time(cfs_rq);
2857 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2858 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2859 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2863 * Update the rq's load with the elapsed running time before entering
2864 * idle. if the last scheduled task is not a CFS task, idle_enter will
2865 * be the only way to update the runnable statistic.
2867 void idle_enter_fair(struct rq *this_rq)
2872 * Update the rq's load with the elapsed idle time before a task is
2873 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2874 * be the only way to update the runnable statistic.
2876 void idle_exit_fair(struct rq *this_rq)
2880 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2882 return cfs_rq->runnable_load_avg;
2885 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2887 return cfs_rq->avg.load_avg;
2890 static int idle_balance(struct rq *this_rq);
2892 #else /* CONFIG_SMP */
2894 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2896 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2898 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2899 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2902 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2904 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2906 static inline int idle_balance(struct rq *rq)
2911 #endif /* CONFIG_SMP */
2913 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2915 #ifdef CONFIG_SCHEDSTATS
2916 struct task_struct *tsk = NULL;
2918 if (entity_is_task(se))
2921 if (se->statistics.sleep_start) {
2922 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2927 if (unlikely(delta > se->statistics.sleep_max))
2928 se->statistics.sleep_max = delta;
2930 se->statistics.sleep_start = 0;
2931 se->statistics.sum_sleep_runtime += delta;
2934 account_scheduler_latency(tsk, delta >> 10, 1);
2935 trace_sched_stat_sleep(tsk, delta);
2938 if (se->statistics.block_start) {
2939 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2944 if (unlikely(delta > se->statistics.block_max))
2945 se->statistics.block_max = delta;
2947 se->statistics.block_start = 0;
2948 se->statistics.sum_sleep_runtime += delta;
2951 if (tsk->in_iowait) {
2952 se->statistics.iowait_sum += delta;
2953 se->statistics.iowait_count++;
2954 trace_sched_stat_iowait(tsk, delta);
2957 trace_sched_stat_blocked(tsk, delta);
2960 * Blocking time is in units of nanosecs, so shift by
2961 * 20 to get a milliseconds-range estimation of the
2962 * amount of time that the task spent sleeping:
2964 if (unlikely(prof_on == SLEEP_PROFILING)) {
2965 profile_hits(SLEEP_PROFILING,
2966 (void *)get_wchan(tsk),
2969 account_scheduler_latency(tsk, delta >> 10, 0);
2975 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2977 #ifdef CONFIG_SCHED_DEBUG
2978 s64 d = se->vruntime - cfs_rq->min_vruntime;
2983 if (d > 3*sysctl_sched_latency)
2984 schedstat_inc(cfs_rq, nr_spread_over);
2989 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2991 u64 vruntime = cfs_rq->min_vruntime;
2994 * The 'current' period is already promised to the current tasks,
2995 * however the extra weight of the new task will slow them down a
2996 * little, place the new task so that it fits in the slot that
2997 * stays open at the end.
2999 if (initial && sched_feat(START_DEBIT))
3000 vruntime += sched_vslice(cfs_rq, se);
3002 /* sleeps up to a single latency don't count. */
3004 unsigned long thresh = sysctl_sched_latency;
3007 * Halve their sleep time's effect, to allow
3008 * for a gentler effect of sleepers:
3010 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3016 /* ensure we never gain time by being placed backwards. */
3017 se->vruntime = max_vruntime(se->vruntime, vruntime);
3020 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3023 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3026 * Update the normalized vruntime before updating min_vruntime
3027 * through calling update_curr().
3029 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3030 se->vruntime += cfs_rq->min_vruntime;
3033 * Update run-time statistics of the 'current'.
3035 update_curr(cfs_rq);
3036 enqueue_entity_load_avg(cfs_rq, se);
3037 account_entity_enqueue(cfs_rq, se);
3038 update_cfs_shares(cfs_rq);
3040 if (flags & ENQUEUE_WAKEUP) {
3041 place_entity(cfs_rq, se, 0);
3042 enqueue_sleeper(cfs_rq, se);
3045 update_stats_enqueue(cfs_rq, se);
3046 check_spread(cfs_rq, se);
3047 if (se != cfs_rq->curr)
3048 __enqueue_entity(cfs_rq, se);
3051 if (cfs_rq->nr_running == 1) {
3052 list_add_leaf_cfs_rq(cfs_rq);
3053 check_enqueue_throttle(cfs_rq);
3057 static void __clear_buddies_last(struct sched_entity *se)
3059 for_each_sched_entity(se) {
3060 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3061 if (cfs_rq->last != se)
3064 cfs_rq->last = NULL;
3068 static void __clear_buddies_next(struct sched_entity *se)
3070 for_each_sched_entity(se) {
3071 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3072 if (cfs_rq->next != se)
3075 cfs_rq->next = NULL;
3079 static void __clear_buddies_skip(struct sched_entity *se)
3081 for_each_sched_entity(se) {
3082 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3083 if (cfs_rq->skip != se)
3086 cfs_rq->skip = NULL;
3090 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3092 if (cfs_rq->last == se)
3093 __clear_buddies_last(se);
3095 if (cfs_rq->next == se)
3096 __clear_buddies_next(se);
3098 if (cfs_rq->skip == se)
3099 __clear_buddies_skip(se);
3102 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3105 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3108 * Update run-time statistics of the 'current'.
3110 update_curr(cfs_rq);
3111 dequeue_entity_load_avg(cfs_rq, se);
3113 update_stats_dequeue(cfs_rq, se);
3114 if (flags & DEQUEUE_SLEEP) {
3115 #ifdef CONFIG_SCHEDSTATS
3116 if (entity_is_task(se)) {
3117 struct task_struct *tsk = task_of(se);
3119 if (tsk->state & TASK_INTERRUPTIBLE)
3120 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3121 if (tsk->state & TASK_UNINTERRUPTIBLE)
3122 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3127 clear_buddies(cfs_rq, se);
3129 if (se != cfs_rq->curr)
3130 __dequeue_entity(cfs_rq, se);
3132 account_entity_dequeue(cfs_rq, se);
3135 * Normalize the entity after updating the min_vruntime because the
3136 * update can refer to the ->curr item and we need to reflect this
3137 * movement in our normalized position.
3139 if (!(flags & DEQUEUE_SLEEP))
3140 se->vruntime -= cfs_rq->min_vruntime;
3142 /* return excess runtime on last dequeue */
3143 return_cfs_rq_runtime(cfs_rq);
3145 update_min_vruntime(cfs_rq);
3146 update_cfs_shares(cfs_rq);
3150 * Preempt the current task with a newly woken task if needed:
3153 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3155 unsigned long ideal_runtime, delta_exec;
3156 struct sched_entity *se;
3159 ideal_runtime = sched_slice(cfs_rq, curr);
3160 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3161 if (delta_exec > ideal_runtime) {
3162 resched_curr(rq_of(cfs_rq));
3164 * The current task ran long enough, ensure it doesn't get
3165 * re-elected due to buddy favours.
3167 clear_buddies(cfs_rq, curr);
3172 * Ensure that a task that missed wakeup preemption by a
3173 * narrow margin doesn't have to wait for a full slice.
3174 * This also mitigates buddy induced latencies under load.
3176 if (delta_exec < sysctl_sched_min_granularity)
3179 se = __pick_first_entity(cfs_rq);
3180 delta = curr->vruntime - se->vruntime;
3185 if (delta > ideal_runtime)
3186 resched_curr(rq_of(cfs_rq));
3190 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3192 /* 'current' is not kept within the tree. */
3195 * Any task has to be enqueued before it get to execute on
3196 * a CPU. So account for the time it spent waiting on the
3199 update_stats_wait_end(cfs_rq, se);
3200 __dequeue_entity(cfs_rq, se);
3201 update_load_avg(se, 1);
3204 update_stats_curr_start(cfs_rq, se);
3206 #ifdef CONFIG_SCHEDSTATS
3208 * Track our maximum slice length, if the CPU's load is at
3209 * least twice that of our own weight (i.e. dont track it
3210 * when there are only lesser-weight tasks around):
3212 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3213 se->statistics.slice_max = max(se->statistics.slice_max,
3214 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3217 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3221 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3224 * Pick the next process, keeping these things in mind, in this order:
3225 * 1) keep things fair between processes/task groups
3226 * 2) pick the "next" process, since someone really wants that to run
3227 * 3) pick the "last" process, for cache locality
3228 * 4) do not run the "skip" process, if something else is available
3230 static struct sched_entity *
3231 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3233 struct sched_entity *left = __pick_first_entity(cfs_rq);
3234 struct sched_entity *se;
3237 * If curr is set we have to see if its left of the leftmost entity
3238 * still in the tree, provided there was anything in the tree at all.
3240 if (!left || (curr && entity_before(curr, left)))
3243 se = left; /* ideally we run the leftmost entity */
3246 * Avoid running the skip buddy, if running something else can
3247 * be done without getting too unfair.
3249 if (cfs_rq->skip == se) {
3250 struct sched_entity *second;
3253 second = __pick_first_entity(cfs_rq);
3255 second = __pick_next_entity(se);
3256 if (!second || (curr && entity_before(curr, second)))
3260 if (second && wakeup_preempt_entity(second, left) < 1)
3265 * Prefer last buddy, try to return the CPU to a preempted task.
3267 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3271 * Someone really wants this to run. If it's not unfair, run it.
3273 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3276 clear_buddies(cfs_rq, se);
3281 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3283 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3286 * If still on the runqueue then deactivate_task()
3287 * was not called and update_curr() has to be done:
3290 update_curr(cfs_rq);
3292 /* throttle cfs_rqs exceeding runtime */
3293 check_cfs_rq_runtime(cfs_rq);
3295 check_spread(cfs_rq, prev);
3297 update_stats_wait_start(cfs_rq, prev);
3298 /* Put 'current' back into the tree. */
3299 __enqueue_entity(cfs_rq, prev);
3300 /* in !on_rq case, update occurred at dequeue */
3301 update_load_avg(prev, 0);
3303 cfs_rq->curr = NULL;
3307 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3310 * Update run-time statistics of the 'current'.
3312 update_curr(cfs_rq);
3315 * Ensure that runnable average is periodically updated.
3317 update_load_avg(curr, 1);
3318 update_cfs_shares(cfs_rq);
3320 #ifdef CONFIG_SCHED_HRTICK
3322 * queued ticks are scheduled to match the slice, so don't bother
3323 * validating it and just reschedule.
3326 resched_curr(rq_of(cfs_rq));
3330 * don't let the period tick interfere with the hrtick preemption
3332 if (!sched_feat(DOUBLE_TICK) &&
3333 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3337 if (cfs_rq->nr_running > 1)
3338 check_preempt_tick(cfs_rq, curr);
3342 /**************************************************
3343 * CFS bandwidth control machinery
3346 #ifdef CONFIG_CFS_BANDWIDTH
3348 #ifdef HAVE_JUMP_LABEL
3349 static struct static_key __cfs_bandwidth_used;
3351 static inline bool cfs_bandwidth_used(void)
3353 return static_key_false(&__cfs_bandwidth_used);
3356 void cfs_bandwidth_usage_inc(void)
3358 static_key_slow_inc(&__cfs_bandwidth_used);
3361 void cfs_bandwidth_usage_dec(void)
3363 static_key_slow_dec(&__cfs_bandwidth_used);
3365 #else /* HAVE_JUMP_LABEL */
3366 static bool cfs_bandwidth_used(void)
3371 void cfs_bandwidth_usage_inc(void) {}
3372 void cfs_bandwidth_usage_dec(void) {}
3373 #endif /* HAVE_JUMP_LABEL */
3376 * default period for cfs group bandwidth.
3377 * default: 0.1s, units: nanoseconds
3379 static inline u64 default_cfs_period(void)
3381 return 100000000ULL;
3384 static inline u64 sched_cfs_bandwidth_slice(void)
3386 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3390 * Replenish runtime according to assigned quota and update expiration time.
3391 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3392 * additional synchronization around rq->lock.
3394 * requires cfs_b->lock
3396 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3400 if (cfs_b->quota == RUNTIME_INF)
3403 now = sched_clock_cpu(smp_processor_id());
3404 cfs_b->runtime = cfs_b->quota;
3405 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3408 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3410 return &tg->cfs_bandwidth;
3413 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3414 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3416 if (unlikely(cfs_rq->throttle_count))
3417 return cfs_rq->throttled_clock_task;
3419 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3422 /* returns 0 on failure to allocate runtime */
3423 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3425 struct task_group *tg = cfs_rq->tg;
3426 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3427 u64 amount = 0, min_amount, expires;
3429 /* note: this is a positive sum as runtime_remaining <= 0 */
3430 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3432 raw_spin_lock(&cfs_b->lock);
3433 if (cfs_b->quota == RUNTIME_INF)
3434 amount = min_amount;
3436 start_cfs_bandwidth(cfs_b);
3438 if (cfs_b->runtime > 0) {
3439 amount = min(cfs_b->runtime, min_amount);
3440 cfs_b->runtime -= amount;
3444 expires = cfs_b->runtime_expires;
3445 raw_spin_unlock(&cfs_b->lock);
3447 cfs_rq->runtime_remaining += amount;
3449 * we may have advanced our local expiration to account for allowed
3450 * spread between our sched_clock and the one on which runtime was
3453 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3454 cfs_rq->runtime_expires = expires;
3456 return cfs_rq->runtime_remaining > 0;
3460 * Note: This depends on the synchronization provided by sched_clock and the
3461 * fact that rq->clock snapshots this value.
3463 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3465 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3467 /* if the deadline is ahead of our clock, nothing to do */
3468 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3471 if (cfs_rq->runtime_remaining < 0)
3475 * If the local deadline has passed we have to consider the
3476 * possibility that our sched_clock is 'fast' and the global deadline
3477 * has not truly expired.
3479 * Fortunately we can check determine whether this the case by checking
3480 * whether the global deadline has advanced. It is valid to compare
3481 * cfs_b->runtime_expires without any locks since we only care about
3482 * exact equality, so a partial write will still work.
3485 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3486 /* extend local deadline, drift is bounded above by 2 ticks */
3487 cfs_rq->runtime_expires += TICK_NSEC;
3489 /* global deadline is ahead, expiration has passed */
3490 cfs_rq->runtime_remaining = 0;
3494 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3496 /* dock delta_exec before expiring quota (as it could span periods) */
3497 cfs_rq->runtime_remaining -= delta_exec;
3498 expire_cfs_rq_runtime(cfs_rq);
3500 if (likely(cfs_rq->runtime_remaining > 0))
3504 * if we're unable to extend our runtime we resched so that the active
3505 * hierarchy can be throttled
3507 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3508 resched_curr(rq_of(cfs_rq));
3511 static __always_inline
3512 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3514 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3517 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3520 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3522 return cfs_bandwidth_used() && cfs_rq->throttled;
3525 /* check whether cfs_rq, or any parent, is throttled */
3526 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3528 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3532 * Ensure that neither of the group entities corresponding to src_cpu or
3533 * dest_cpu are members of a throttled hierarchy when performing group
3534 * load-balance operations.
3536 static inline int throttled_lb_pair(struct task_group *tg,
3537 int src_cpu, int dest_cpu)
3539 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3541 src_cfs_rq = tg->cfs_rq[src_cpu];
3542 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3544 return throttled_hierarchy(src_cfs_rq) ||
3545 throttled_hierarchy(dest_cfs_rq);
3548 /* updated child weight may affect parent so we have to do this bottom up */
3549 static int tg_unthrottle_up(struct task_group *tg, void *data)
3551 struct rq *rq = data;
3552 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3554 cfs_rq->throttle_count--;
3556 if (!cfs_rq->throttle_count) {
3557 /* adjust cfs_rq_clock_task() */
3558 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3559 cfs_rq->throttled_clock_task;
3566 static int tg_throttle_down(struct task_group *tg, void *data)
3568 struct rq *rq = data;
3569 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3571 /* group is entering throttled state, stop time */
3572 if (!cfs_rq->throttle_count)
3573 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3574 cfs_rq->throttle_count++;
3579 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3581 struct rq *rq = rq_of(cfs_rq);
3582 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3583 struct sched_entity *se;
3584 long task_delta, dequeue = 1;
3587 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3589 /* freeze hierarchy runnable averages while throttled */
3591 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3594 task_delta = cfs_rq->h_nr_running;
3595 for_each_sched_entity(se) {
3596 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3597 /* throttled entity or throttle-on-deactivate */
3602 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3603 qcfs_rq->h_nr_running -= task_delta;
3605 if (qcfs_rq->load.weight)
3610 sub_nr_running(rq, task_delta);
3612 cfs_rq->throttled = 1;
3613 cfs_rq->throttled_clock = rq_clock(rq);
3614 raw_spin_lock(&cfs_b->lock);
3615 empty = list_empty(&cfs_b->throttled_cfs_rq);
3618 * Add to the _head_ of the list, so that an already-started
3619 * distribute_cfs_runtime will not see us
3621 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3624 * If we're the first throttled task, make sure the bandwidth
3628 start_cfs_bandwidth(cfs_b);
3630 raw_spin_unlock(&cfs_b->lock);
3633 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3635 struct rq *rq = rq_of(cfs_rq);
3636 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3637 struct sched_entity *se;
3641 se = cfs_rq->tg->se[cpu_of(rq)];
3643 cfs_rq->throttled = 0;
3645 update_rq_clock(rq);
3647 raw_spin_lock(&cfs_b->lock);
3648 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3649 list_del_rcu(&cfs_rq->throttled_list);
3650 raw_spin_unlock(&cfs_b->lock);
3652 /* update hierarchical throttle state */
3653 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3655 if (!cfs_rq->load.weight)
3658 task_delta = cfs_rq->h_nr_running;
3659 for_each_sched_entity(se) {
3663 cfs_rq = cfs_rq_of(se);
3665 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3666 cfs_rq->h_nr_running += task_delta;
3668 if (cfs_rq_throttled(cfs_rq))
3673 add_nr_running(rq, task_delta);
3675 /* determine whether we need to wake up potentially idle cpu */
3676 if (rq->curr == rq->idle && rq->cfs.nr_running)
3680 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3681 u64 remaining, u64 expires)
3683 struct cfs_rq *cfs_rq;
3685 u64 starting_runtime = remaining;
3688 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3690 struct rq *rq = rq_of(cfs_rq);
3692 raw_spin_lock(&rq->lock);
3693 if (!cfs_rq_throttled(cfs_rq))
3696 runtime = -cfs_rq->runtime_remaining + 1;
3697 if (runtime > remaining)
3698 runtime = remaining;
3699 remaining -= runtime;
3701 cfs_rq->runtime_remaining += runtime;
3702 cfs_rq->runtime_expires = expires;
3704 /* we check whether we're throttled above */
3705 if (cfs_rq->runtime_remaining > 0)
3706 unthrottle_cfs_rq(cfs_rq);
3709 raw_spin_unlock(&rq->lock);
3716 return starting_runtime - remaining;
3720 * Responsible for refilling a task_group's bandwidth and unthrottling its
3721 * cfs_rqs as appropriate. If there has been no activity within the last
3722 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3723 * used to track this state.
3725 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3727 u64 runtime, runtime_expires;
3730 /* no need to continue the timer with no bandwidth constraint */
3731 if (cfs_b->quota == RUNTIME_INF)
3732 goto out_deactivate;
3734 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3735 cfs_b->nr_periods += overrun;
3738 * idle depends on !throttled (for the case of a large deficit), and if
3739 * we're going inactive then everything else can be deferred
3741 if (cfs_b->idle && !throttled)
3742 goto out_deactivate;
3744 __refill_cfs_bandwidth_runtime(cfs_b);
3747 /* mark as potentially idle for the upcoming period */
3752 /* account preceding periods in which throttling occurred */
3753 cfs_b->nr_throttled += overrun;
3755 runtime_expires = cfs_b->runtime_expires;
3758 * This check is repeated as we are holding onto the new bandwidth while
3759 * we unthrottle. This can potentially race with an unthrottled group
3760 * trying to acquire new bandwidth from the global pool. This can result
3761 * in us over-using our runtime if it is all used during this loop, but
3762 * only by limited amounts in that extreme case.
3764 while (throttled && cfs_b->runtime > 0) {
3765 runtime = cfs_b->runtime;
3766 raw_spin_unlock(&cfs_b->lock);
3767 /* we can't nest cfs_b->lock while distributing bandwidth */
3768 runtime = distribute_cfs_runtime(cfs_b, runtime,
3770 raw_spin_lock(&cfs_b->lock);
3772 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3774 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3778 * While we are ensured activity in the period following an
3779 * unthrottle, this also covers the case in which the new bandwidth is
3780 * insufficient to cover the existing bandwidth deficit. (Forcing the
3781 * timer to remain active while there are any throttled entities.)
3791 /* a cfs_rq won't donate quota below this amount */
3792 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3793 /* minimum remaining period time to redistribute slack quota */
3794 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3795 /* how long we wait to gather additional slack before distributing */
3796 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3799 * Are we near the end of the current quota period?
3801 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3802 * hrtimer base being cleared by hrtimer_start. In the case of
3803 * migrate_hrtimers, base is never cleared, so we are fine.
3805 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3807 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3810 /* if the call-back is running a quota refresh is already occurring */
3811 if (hrtimer_callback_running(refresh_timer))
3814 /* is a quota refresh about to occur? */
3815 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3816 if (remaining < min_expire)
3822 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3824 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3826 /* if there's a quota refresh soon don't bother with slack */
3827 if (runtime_refresh_within(cfs_b, min_left))
3830 hrtimer_start(&cfs_b->slack_timer,
3831 ns_to_ktime(cfs_bandwidth_slack_period),
3835 /* we know any runtime found here is valid as update_curr() precedes return */
3836 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3838 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3839 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3841 if (slack_runtime <= 0)
3844 raw_spin_lock(&cfs_b->lock);
3845 if (cfs_b->quota != RUNTIME_INF &&
3846 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3847 cfs_b->runtime += slack_runtime;
3849 /* we are under rq->lock, defer unthrottling using a timer */
3850 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3851 !list_empty(&cfs_b->throttled_cfs_rq))
3852 start_cfs_slack_bandwidth(cfs_b);
3854 raw_spin_unlock(&cfs_b->lock);
3856 /* even if it's not valid for return we don't want to try again */
3857 cfs_rq->runtime_remaining -= slack_runtime;
3860 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3862 if (!cfs_bandwidth_used())
3865 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3868 __return_cfs_rq_runtime(cfs_rq);
3872 * This is done with a timer (instead of inline with bandwidth return) since
3873 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3875 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3877 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3880 /* confirm we're still not at a refresh boundary */
3881 raw_spin_lock(&cfs_b->lock);
3882 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3883 raw_spin_unlock(&cfs_b->lock);
3887 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3888 runtime = cfs_b->runtime;
3890 expires = cfs_b->runtime_expires;
3891 raw_spin_unlock(&cfs_b->lock);
3896 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3898 raw_spin_lock(&cfs_b->lock);
3899 if (expires == cfs_b->runtime_expires)
3900 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3901 raw_spin_unlock(&cfs_b->lock);
3905 * When a group wakes up we want to make sure that its quota is not already
3906 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3907 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3909 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3911 if (!cfs_bandwidth_used())
3914 /* an active group must be handled by the update_curr()->put() path */
3915 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3918 /* ensure the group is not already throttled */
3919 if (cfs_rq_throttled(cfs_rq))
3922 /* update runtime allocation */
3923 account_cfs_rq_runtime(cfs_rq, 0);
3924 if (cfs_rq->runtime_remaining <= 0)
3925 throttle_cfs_rq(cfs_rq);
3928 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3929 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3931 if (!cfs_bandwidth_used())
3934 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3938 * it's possible for a throttled entity to be forced into a running
3939 * state (e.g. set_curr_task), in this case we're finished.
3941 if (cfs_rq_throttled(cfs_rq))
3944 throttle_cfs_rq(cfs_rq);
3948 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3950 struct cfs_bandwidth *cfs_b =
3951 container_of(timer, struct cfs_bandwidth, slack_timer);
3953 do_sched_cfs_slack_timer(cfs_b);
3955 return HRTIMER_NORESTART;
3958 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3960 struct cfs_bandwidth *cfs_b =
3961 container_of(timer, struct cfs_bandwidth, period_timer);
3965 raw_spin_lock(&cfs_b->lock);
3967 overrun = hrtimer_forward_now(timer, cfs_b->period);
3971 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3974 cfs_b->period_active = 0;
3975 raw_spin_unlock(&cfs_b->lock);
3977 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3980 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3982 raw_spin_lock_init(&cfs_b->lock);
3984 cfs_b->quota = RUNTIME_INF;
3985 cfs_b->period = ns_to_ktime(default_cfs_period());
3987 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3988 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
3989 cfs_b->period_timer.function = sched_cfs_period_timer;
3990 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3991 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3994 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3996 cfs_rq->runtime_enabled = 0;
3997 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4000 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4002 lockdep_assert_held(&cfs_b->lock);
4004 if (!cfs_b->period_active) {
4005 cfs_b->period_active = 1;
4006 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4007 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4011 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4013 /* init_cfs_bandwidth() was not called */
4014 if (!cfs_b->throttled_cfs_rq.next)
4017 hrtimer_cancel(&cfs_b->period_timer);
4018 hrtimer_cancel(&cfs_b->slack_timer);
4021 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4023 struct cfs_rq *cfs_rq;
4025 for_each_leaf_cfs_rq(rq, cfs_rq) {
4026 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4028 raw_spin_lock(&cfs_b->lock);
4029 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4030 raw_spin_unlock(&cfs_b->lock);
4034 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4036 struct cfs_rq *cfs_rq;
4038 for_each_leaf_cfs_rq(rq, cfs_rq) {
4039 if (!cfs_rq->runtime_enabled)
4043 * clock_task is not advancing so we just need to make sure
4044 * there's some valid quota amount
4046 cfs_rq->runtime_remaining = 1;
4048 * Offline rq is schedulable till cpu is completely disabled
4049 * in take_cpu_down(), so we prevent new cfs throttling here.
4051 cfs_rq->runtime_enabled = 0;
4053 if (cfs_rq_throttled(cfs_rq))
4054 unthrottle_cfs_rq(cfs_rq);
4058 #else /* CONFIG_CFS_BANDWIDTH */
4059 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4061 return rq_clock_task(rq_of(cfs_rq));
4064 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4065 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4066 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4067 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4069 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4074 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4079 static inline int throttled_lb_pair(struct task_group *tg,
4080 int src_cpu, int dest_cpu)
4085 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4087 #ifdef CONFIG_FAIR_GROUP_SCHED
4088 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4091 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4095 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4096 static inline void update_runtime_enabled(struct rq *rq) {}
4097 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4099 #endif /* CONFIG_CFS_BANDWIDTH */
4101 /**************************************************
4102 * CFS operations on tasks:
4105 #ifdef CONFIG_SCHED_HRTICK
4106 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4108 struct sched_entity *se = &p->se;
4109 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4111 WARN_ON(task_rq(p) != rq);
4113 if (cfs_rq->nr_running > 1) {
4114 u64 slice = sched_slice(cfs_rq, se);
4115 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4116 s64 delta = slice - ran;
4123 hrtick_start(rq, delta);
4128 * called from enqueue/dequeue and updates the hrtick when the
4129 * current task is from our class and nr_running is low enough
4132 static void hrtick_update(struct rq *rq)
4134 struct task_struct *curr = rq->curr;
4136 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4139 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4140 hrtick_start_fair(rq, curr);
4142 #else /* !CONFIG_SCHED_HRTICK */
4144 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4148 static inline void hrtick_update(struct rq *rq)
4153 static inline unsigned long boosted_cpu_util(int cpu);
4155 static void update_capacity_of(int cpu)
4157 unsigned long req_cap;
4162 /* Convert scale-invariant capacity to cpu. */
4163 req_cap = boosted_cpu_util(cpu);
4164 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4165 set_cfs_cpu_capacity(cpu, true, req_cap);
4168 static bool cpu_overutilized(int cpu);
4171 * The enqueue_task method is called before nr_running is
4172 * increased. Here we update the fair scheduling stats and
4173 * then put the task into the rbtree:
4176 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4178 struct cfs_rq *cfs_rq;
4179 struct sched_entity *se = &p->se;
4180 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4181 int task_wakeup = flags & ENQUEUE_WAKEUP;
4183 for_each_sched_entity(se) {
4186 cfs_rq = cfs_rq_of(se);
4187 enqueue_entity(cfs_rq, se, flags);
4190 * end evaluation on encountering a throttled cfs_rq
4192 * note: in the case of encountering a throttled cfs_rq we will
4193 * post the final h_nr_running increment below.
4195 if (cfs_rq_throttled(cfs_rq))
4197 cfs_rq->h_nr_running++;
4199 flags = ENQUEUE_WAKEUP;
4202 for_each_sched_entity(se) {
4203 cfs_rq = cfs_rq_of(se);
4204 cfs_rq->h_nr_running++;
4206 if (cfs_rq_throttled(cfs_rq))
4209 update_load_avg(se, 1);
4210 update_cfs_shares(cfs_rq);
4214 add_nr_running(rq, 1);
4215 if (!task_new && !rq->rd->overutilized &&
4216 cpu_overutilized(rq->cpu))
4217 rq->rd->overutilized = true;
4219 schedtune_enqueue_task(p, cpu_of(rq));
4222 * We want to potentially trigger a freq switch
4223 * request only for tasks that are waking up; this is
4224 * because we get here also during load balancing, but
4225 * in these cases it seems wise to trigger as single
4226 * request after load balancing is done.
4228 if (task_new || task_wakeup)
4229 update_capacity_of(cpu_of(rq));
4234 static void set_next_buddy(struct sched_entity *se);
4237 * The dequeue_task method is called before nr_running is
4238 * decreased. We remove the task from the rbtree and
4239 * update the fair scheduling stats:
4241 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4243 struct cfs_rq *cfs_rq;
4244 struct sched_entity *se = &p->se;
4245 int task_sleep = flags & DEQUEUE_SLEEP;
4247 for_each_sched_entity(se) {
4248 cfs_rq = cfs_rq_of(se);
4249 dequeue_entity(cfs_rq, se, flags);
4252 * end evaluation on encountering a throttled cfs_rq
4254 * note: in the case of encountering a throttled cfs_rq we will
4255 * post the final h_nr_running decrement below.
4257 if (cfs_rq_throttled(cfs_rq))
4259 cfs_rq->h_nr_running--;
4261 /* Don't dequeue parent if it has other entities besides us */
4262 if (cfs_rq->load.weight) {
4264 * Bias pick_next to pick a task from this cfs_rq, as
4265 * p is sleeping when it is within its sched_slice.
4267 if (task_sleep && parent_entity(se))
4268 set_next_buddy(parent_entity(se));
4270 /* avoid re-evaluating load for this entity */
4271 se = parent_entity(se);
4274 flags |= DEQUEUE_SLEEP;
4277 for_each_sched_entity(se) {
4278 cfs_rq = cfs_rq_of(se);
4279 cfs_rq->h_nr_running--;
4281 if (cfs_rq_throttled(cfs_rq))
4284 update_load_avg(se, 1);
4285 update_cfs_shares(cfs_rq);
4289 sub_nr_running(rq, 1);
4290 schedtune_dequeue_task(p, cpu_of(rq));
4293 * We want to potentially trigger a freq switch
4294 * request only for tasks that are going to sleep;
4295 * this is because we get here also during load
4296 * balancing, but in these cases it seems wise to
4297 * trigger as single request after load balancing is
4301 if (rq->cfs.nr_running)
4302 update_capacity_of(cpu_of(rq));
4303 else if (sched_freq())
4304 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4313 * per rq 'load' arrray crap; XXX kill this.
4317 * The exact cpuload at various idx values, calculated at every tick would be
4318 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4320 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4321 * on nth tick when cpu may be busy, then we have:
4322 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4323 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4325 * decay_load_missed() below does efficient calculation of
4326 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4327 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4329 * The calculation is approximated on a 128 point scale.
4330 * degrade_zero_ticks is the number of ticks after which load at any
4331 * particular idx is approximated to be zero.
4332 * degrade_factor is a precomputed table, a row for each load idx.
4333 * Each column corresponds to degradation factor for a power of two ticks,
4334 * based on 128 point scale.
4336 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4337 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4339 * With this power of 2 load factors, we can degrade the load n times
4340 * by looking at 1 bits in n and doing as many mult/shift instead of
4341 * n mult/shifts needed by the exact degradation.
4343 #define DEGRADE_SHIFT 7
4344 static const unsigned char
4345 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4346 static const unsigned char
4347 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4348 {0, 0, 0, 0, 0, 0, 0, 0},
4349 {64, 32, 8, 0, 0, 0, 0, 0},
4350 {96, 72, 40, 12, 1, 0, 0},
4351 {112, 98, 75, 43, 15, 1, 0},
4352 {120, 112, 98, 76, 45, 16, 2} };
4355 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4356 * would be when CPU is idle and so we just decay the old load without
4357 * adding any new load.
4359 static unsigned long
4360 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4364 if (!missed_updates)
4367 if (missed_updates >= degrade_zero_ticks[idx])
4371 return load >> missed_updates;
4373 while (missed_updates) {
4374 if (missed_updates % 2)
4375 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4377 missed_updates >>= 1;
4384 * Update rq->cpu_load[] statistics. This function is usually called every
4385 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4386 * every tick. We fix it up based on jiffies.
4388 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4389 unsigned long pending_updates)
4393 this_rq->nr_load_updates++;
4395 /* Update our load: */
4396 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4397 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4398 unsigned long old_load, new_load;
4400 /* scale is effectively 1 << i now, and >> i divides by scale */
4402 old_load = this_rq->cpu_load[i];
4403 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4404 new_load = this_load;
4406 * Round up the averaging division if load is increasing. This
4407 * prevents us from getting stuck on 9 if the load is 10, for
4410 if (new_load > old_load)
4411 new_load += scale - 1;
4413 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4416 sched_avg_update(this_rq);
4419 /* Used instead of source_load when we know the type == 0 */
4420 static unsigned long weighted_cpuload(const int cpu)
4422 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4425 #ifdef CONFIG_NO_HZ_COMMON
4427 * There is no sane way to deal with nohz on smp when using jiffies because the
4428 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4429 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4431 * Therefore we cannot use the delta approach from the regular tick since that
4432 * would seriously skew the load calculation. However we'll make do for those
4433 * updates happening while idle (nohz_idle_balance) or coming out of idle
4434 * (tick_nohz_idle_exit).
4436 * This means we might still be one tick off for nohz periods.
4440 * Called from nohz_idle_balance() to update the load ratings before doing the
4443 static void update_idle_cpu_load(struct rq *this_rq)
4445 unsigned long curr_jiffies = READ_ONCE(jiffies);
4446 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4447 unsigned long pending_updates;
4450 * bail if there's load or we're actually up-to-date.
4452 if (load || curr_jiffies == this_rq->last_load_update_tick)
4455 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4456 this_rq->last_load_update_tick = curr_jiffies;
4458 __update_cpu_load(this_rq, load, pending_updates);
4462 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4464 void update_cpu_load_nohz(void)
4466 struct rq *this_rq = this_rq();
4467 unsigned long curr_jiffies = READ_ONCE(jiffies);
4468 unsigned long pending_updates;
4470 if (curr_jiffies == this_rq->last_load_update_tick)
4473 raw_spin_lock(&this_rq->lock);
4474 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4475 if (pending_updates) {
4476 this_rq->last_load_update_tick = curr_jiffies;
4478 * We were idle, this means load 0, the current load might be
4479 * !0 due to remote wakeups and the sort.
4481 __update_cpu_load(this_rq, 0, pending_updates);
4483 raw_spin_unlock(&this_rq->lock);
4485 #endif /* CONFIG_NO_HZ */
4488 * Called from scheduler_tick()
4490 void update_cpu_load_active(struct rq *this_rq)
4492 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4494 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4496 this_rq->last_load_update_tick = jiffies;
4497 __update_cpu_load(this_rq, load, 1);
4501 * Return a low guess at the load of a migration-source cpu weighted
4502 * according to the scheduling class and "nice" value.
4504 * We want to under-estimate the load of migration sources, to
4505 * balance conservatively.
4507 static unsigned long source_load(int cpu, int type)
4509 struct rq *rq = cpu_rq(cpu);
4510 unsigned long total = weighted_cpuload(cpu);
4512 if (type == 0 || !sched_feat(LB_BIAS))
4515 return min(rq->cpu_load[type-1], total);
4519 * Return a high guess at the load of a migration-target cpu weighted
4520 * according to the scheduling class and "nice" value.
4522 static unsigned long target_load(int cpu, int type)
4524 struct rq *rq = cpu_rq(cpu);
4525 unsigned long total = weighted_cpuload(cpu);
4527 if (type == 0 || !sched_feat(LB_BIAS))
4530 return max(rq->cpu_load[type-1], total);
4534 static unsigned long cpu_avg_load_per_task(int cpu)
4536 struct rq *rq = cpu_rq(cpu);
4537 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4538 unsigned long load_avg = weighted_cpuload(cpu);
4541 return load_avg / nr_running;
4546 static void record_wakee(struct task_struct *p)
4549 * Rough decay (wiping) for cost saving, don't worry
4550 * about the boundary, really active task won't care
4553 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4554 current->wakee_flips >>= 1;
4555 current->wakee_flip_decay_ts = jiffies;
4558 if (current->last_wakee != p) {
4559 current->last_wakee = p;
4560 current->wakee_flips++;
4564 static void task_waking_fair(struct task_struct *p)
4566 struct sched_entity *se = &p->se;
4567 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4570 #ifndef CONFIG_64BIT
4571 u64 min_vruntime_copy;
4574 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4576 min_vruntime = cfs_rq->min_vruntime;
4577 } while (min_vruntime != min_vruntime_copy);
4579 min_vruntime = cfs_rq->min_vruntime;
4582 se->vruntime -= min_vruntime;
4586 #ifdef CONFIG_FAIR_GROUP_SCHED
4588 * effective_load() calculates the load change as seen from the root_task_group
4590 * Adding load to a group doesn't make a group heavier, but can cause movement
4591 * of group shares between cpus. Assuming the shares were perfectly aligned one
4592 * can calculate the shift in shares.
4594 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4595 * on this @cpu and results in a total addition (subtraction) of @wg to the
4596 * total group weight.
4598 * Given a runqueue weight distribution (rw_i) we can compute a shares
4599 * distribution (s_i) using:
4601 * s_i = rw_i / \Sum rw_j (1)
4603 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4604 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4605 * shares distribution (s_i):
4607 * rw_i = { 2, 4, 1, 0 }
4608 * s_i = { 2/7, 4/7, 1/7, 0 }
4610 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4611 * task used to run on and the CPU the waker is running on), we need to
4612 * compute the effect of waking a task on either CPU and, in case of a sync
4613 * wakeup, compute the effect of the current task going to sleep.
4615 * So for a change of @wl to the local @cpu with an overall group weight change
4616 * of @wl we can compute the new shares distribution (s'_i) using:
4618 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4620 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4621 * differences in waking a task to CPU 0. The additional task changes the
4622 * weight and shares distributions like:
4624 * rw'_i = { 3, 4, 1, 0 }
4625 * s'_i = { 3/8, 4/8, 1/8, 0 }
4627 * We can then compute the difference in effective weight by using:
4629 * dw_i = S * (s'_i - s_i) (3)
4631 * Where 'S' is the group weight as seen by its parent.
4633 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4634 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4635 * 4/7) times the weight of the group.
4637 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4639 struct sched_entity *se = tg->se[cpu];
4641 if (!tg->parent) /* the trivial, non-cgroup case */
4644 for_each_sched_entity(se) {
4650 * W = @wg + \Sum rw_j
4652 W = wg + calc_tg_weight(tg, se->my_q);
4657 w = cfs_rq_load_avg(se->my_q) + wl;
4660 * wl = S * s'_i; see (2)
4663 wl = (w * (long)tg->shares) / W;
4668 * Per the above, wl is the new se->load.weight value; since
4669 * those are clipped to [MIN_SHARES, ...) do so now. See
4670 * calc_cfs_shares().
4672 if (wl < MIN_SHARES)
4676 * wl = dw_i = S * (s'_i - s_i); see (3)
4678 wl -= se->avg.load_avg;
4681 * Recursively apply this logic to all parent groups to compute
4682 * the final effective load change on the root group. Since
4683 * only the @tg group gets extra weight, all parent groups can
4684 * only redistribute existing shares. @wl is the shift in shares
4685 * resulting from this level per the above.
4694 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4701 static inline bool energy_aware(void)
4703 return sched_feat(ENERGY_AWARE);
4707 struct sched_group *sg_top;
4708 struct sched_group *sg_cap;
4715 struct task_struct *task;
4730 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4731 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4732 * energy calculations. Using the scale-invariant util returned by
4733 * cpu_util() and approximating scale-invariant util by:
4735 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4737 * the normalized util can be found using the specific capacity.
4739 * capacity = capacity_orig * curr_freq/max_freq
4741 * norm_util = running_time/time ~ util/capacity
4743 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4745 int util = __cpu_util(cpu, delta);
4747 if (util >= capacity)
4748 return SCHED_CAPACITY_SCALE;
4750 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4753 static int calc_util_delta(struct energy_env *eenv, int cpu)
4755 if (cpu == eenv->src_cpu)
4756 return -eenv->util_delta;
4757 if (cpu == eenv->dst_cpu)
4758 return eenv->util_delta;
4763 unsigned long group_max_util(struct energy_env *eenv)
4766 unsigned long max_util = 0;
4768 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4769 delta = calc_util_delta(eenv, i);
4770 max_util = max(max_util, __cpu_util(i, delta));
4777 * group_norm_util() returns the approximated group util relative to it's
4778 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4779 * energy calculations. Since task executions may or may not overlap in time in
4780 * the group the true normalized util is between max(cpu_norm_util(i)) and
4781 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4782 * latter is used as the estimate as it leads to a more pessimistic energy
4783 * estimate (more busy).
4786 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4789 unsigned long util_sum = 0;
4790 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4792 for_each_cpu(i, sched_group_cpus(sg)) {
4793 delta = calc_util_delta(eenv, i);
4794 util_sum += __cpu_norm_util(i, capacity, delta);
4797 if (util_sum > SCHED_CAPACITY_SCALE)
4798 return SCHED_CAPACITY_SCALE;
4802 static int find_new_capacity(struct energy_env *eenv,
4803 const struct sched_group_energy const *sge)
4806 unsigned long util = group_max_util(eenv);
4808 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4809 if (sge->cap_states[idx].cap >= util)
4813 eenv->cap_idx = idx;
4818 static int group_idle_state(struct sched_group *sg)
4820 int i, state = INT_MAX;
4822 /* Find the shallowest idle state in the sched group. */
4823 for_each_cpu(i, sched_group_cpus(sg))
4824 state = min(state, idle_get_state_idx(cpu_rq(i)));
4826 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4833 * sched_group_energy(): Computes the absolute energy consumption of cpus
4834 * belonging to the sched_group including shared resources shared only by
4835 * members of the group. Iterates over all cpus in the hierarchy below the
4836 * sched_group starting from the bottom working it's way up before going to
4837 * the next cpu until all cpus are covered at all levels. The current
4838 * implementation is likely to gather the same util statistics multiple times.
4839 * This can probably be done in a faster but more complex way.
4840 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4842 static int sched_group_energy(struct energy_env *eenv)
4844 struct sched_domain *sd;
4845 int cpu, total_energy = 0;
4846 struct cpumask visit_cpus;
4847 struct sched_group *sg;
4849 WARN_ON(!eenv->sg_top->sge);
4851 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4853 while (!cpumask_empty(&visit_cpus)) {
4854 struct sched_group *sg_shared_cap = NULL;
4856 cpu = cpumask_first(&visit_cpus);
4859 * Is the group utilization affected by cpus outside this
4862 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4866 * We most probably raced with hotplug; returning a
4867 * wrong energy estimation is better than entering an
4873 sg_shared_cap = sd->parent->groups;
4875 for_each_domain(cpu, sd) {
4878 /* Has this sched_domain already been visited? */
4879 if (sd->child && group_first_cpu(sg) != cpu)
4883 unsigned long group_util;
4884 int sg_busy_energy, sg_idle_energy;
4885 int cap_idx, idle_idx;
4887 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
4888 eenv->sg_cap = sg_shared_cap;
4892 cap_idx = find_new_capacity(eenv, sg->sge);
4894 if (sg->group_weight == 1) {
4895 /* Remove capacity of src CPU (before task move) */
4896 if (eenv->util_delta == 0 &&
4897 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
4898 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
4899 eenv->cap.delta -= eenv->cap.before;
4901 /* Add capacity of dst CPU (after task move) */
4902 if (eenv->util_delta != 0 &&
4903 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
4904 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
4905 eenv->cap.delta += eenv->cap.after;
4909 idle_idx = group_idle_state(sg);
4910 group_util = group_norm_util(eenv, sg);
4911 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
4912 >> SCHED_CAPACITY_SHIFT;
4913 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
4914 * sg->sge->idle_states[idle_idx].power)
4915 >> SCHED_CAPACITY_SHIFT;
4917 total_energy += sg_busy_energy + sg_idle_energy;
4920 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
4922 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
4925 } while (sg = sg->next, sg != sd->groups);
4931 eenv->energy = total_energy;
4935 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
4937 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
4940 #ifdef CONFIG_SCHED_TUNE
4941 static int energy_diff_evaluate(struct energy_env *eenv)
4946 /* Return energy diff when boost margin is 0 */
4947 #ifdef CONFIG_CGROUP_SCHEDTUNE
4948 boost = schedtune_task_boost(eenv->task);
4950 boost = get_sysctl_sched_cfs_boost();
4953 return eenv->nrg.diff;
4955 /* Compute normalized energy diff */
4956 nrg_delta = schedtune_normalize_energy(eenv->nrg.diff);
4957 eenv->nrg.delta = nrg_delta;
4959 eenv->payoff = schedtune_accept_deltas(
4965 * When SchedTune is enabled, the energy_diff() function will return
4966 * the computed energy payoff value. Since the energy_diff() return
4967 * value is expected to be negative by its callers, this evaluation
4968 * function return a negative value each time the evaluation return a
4969 * positive payoff, which is the condition for the acceptance of
4970 * a scheduling decision
4972 return -eenv->payoff;
4974 #else /* CONFIG_SCHED_TUNE */
4975 #define energy_diff_evaluate(eenv) eenv->nrg.diff
4979 * energy_diff(): Estimate the energy impact of changing the utilization
4980 * distribution. eenv specifies the change: utilisation amount, source, and
4981 * destination cpu. Source or destination cpu may be -1 in which case the
4982 * utilization is removed from or added to the system (e.g. task wake-up). If
4983 * both are specified, the utilization is migrated.
4985 static int energy_diff(struct energy_env *eenv)
4987 struct sched_domain *sd;
4988 struct sched_group *sg;
4989 int sd_cpu = -1, energy_before = 0, energy_after = 0;
4991 struct energy_env eenv_before = {
4993 .src_cpu = eenv->src_cpu,
4994 .dst_cpu = eenv->dst_cpu,
4995 .nrg = { 0, 0, 0, 0},
4999 if (eenv->src_cpu == eenv->dst_cpu)
5002 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5003 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5006 return 0; /* Error */
5011 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5012 eenv_before.sg_top = eenv->sg_top = sg;
5014 if (sched_group_energy(&eenv_before))
5015 return 0; /* Invalid result abort */
5016 energy_before += eenv_before.energy;
5018 /* Keep track of SRC cpu (before) capacity */
5019 eenv->cap.before = eenv_before.cap.before;
5020 eenv->cap.delta = eenv_before.cap.delta;
5022 if (sched_group_energy(eenv))
5023 return 0; /* Invalid result abort */
5024 energy_after += eenv->energy;
5026 } while (sg = sg->next, sg != sd->groups);
5028 eenv->nrg.before = energy_before;
5029 eenv->nrg.after = energy_after;
5030 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5033 return energy_diff_evaluate(eenv);
5037 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5038 * A waker of many should wake a different task than the one last awakened
5039 * at a frequency roughly N times higher than one of its wakees. In order
5040 * to determine whether we should let the load spread vs consolodating to
5041 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5042 * partner, and a factor of lls_size higher frequency in the other. With
5043 * both conditions met, we can be relatively sure that the relationship is
5044 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5045 * being client/server, worker/dispatcher, interrupt source or whatever is
5046 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5048 static int wake_wide(struct task_struct *p)
5050 unsigned int master = current->wakee_flips;
5051 unsigned int slave = p->wakee_flips;
5052 int factor = this_cpu_read(sd_llc_size);
5055 swap(master, slave);
5056 if (slave < factor || master < slave * factor)
5061 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5063 s64 this_load, load;
5064 s64 this_eff_load, prev_eff_load;
5065 int idx, this_cpu, prev_cpu;
5066 struct task_group *tg;
5067 unsigned long weight;
5071 this_cpu = smp_processor_id();
5072 prev_cpu = task_cpu(p);
5073 load = source_load(prev_cpu, idx);
5074 this_load = target_load(this_cpu, idx);
5077 * If sync wakeup then subtract the (maximum possible)
5078 * effect of the currently running task from the load
5079 * of the current CPU:
5082 tg = task_group(current);
5083 weight = current->se.avg.load_avg;
5085 this_load += effective_load(tg, this_cpu, -weight, -weight);
5086 load += effective_load(tg, prev_cpu, 0, -weight);
5090 weight = p->se.avg.load_avg;
5093 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5094 * due to the sync cause above having dropped this_load to 0, we'll
5095 * always have an imbalance, but there's really nothing you can do
5096 * about that, so that's good too.
5098 * Otherwise check if either cpus are near enough in load to allow this
5099 * task to be woken on this_cpu.
5101 this_eff_load = 100;
5102 this_eff_load *= capacity_of(prev_cpu);
5104 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5105 prev_eff_load *= capacity_of(this_cpu);
5107 if (this_load > 0) {
5108 this_eff_load *= this_load +
5109 effective_load(tg, this_cpu, weight, weight);
5111 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5114 balanced = this_eff_load <= prev_eff_load;
5116 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5121 schedstat_inc(sd, ttwu_move_affine);
5122 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5127 static inline unsigned long task_util(struct task_struct *p)
5129 return p->se.avg.util_avg;
5132 unsigned int capacity_margin = 1280; /* ~20% margin */
5134 static inline unsigned long boosted_task_util(struct task_struct *task);
5136 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5138 unsigned long capacity = capacity_of(cpu);
5140 util += boosted_task_util(p);
5142 return (capacity * 1024) > (util * capacity_margin);
5145 static inline bool task_fits_max(struct task_struct *p, int cpu)
5147 unsigned long capacity = capacity_of(cpu);
5148 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5150 if (capacity == max_capacity)
5153 if (capacity * capacity_margin > max_capacity * 1024)
5156 return __task_fits(p, cpu, 0);
5159 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5161 return __task_fits(p, cpu, cpu_util(cpu));
5164 static bool cpu_overutilized(int cpu)
5166 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5169 #ifdef CONFIG_SCHED_TUNE
5171 static unsigned long
5172 schedtune_margin(unsigned long signal, unsigned long boost)
5174 unsigned long long margin = 0;
5177 * Signal proportional compensation (SPC)
5179 * The Boost (B) value is used to compute a Margin (M) which is
5180 * proportional to the complement of the original Signal (S):
5181 * M = B * (SCHED_LOAD_SCALE - S)
5182 * The obtained M could be used by the caller to "boost" S.
5184 margin = SCHED_LOAD_SCALE - signal;
5188 * Fast integer division by constant:
5189 * Constant : (C) = 100
5190 * Precision : 0.1% (P) = 0.1
5191 * Reference : C * 100 / P (R) = 100000
5194 * Shift bits : ceil(log(R,2)) (S) = 17
5195 * Mult const : round(2^S/C) (M) = 1311
5205 static inline unsigned int
5206 schedtune_cpu_margin(unsigned long util, int cpu)
5210 #ifdef CONFIG_CGROUP_SCHEDTUNE
5211 boost = schedtune_cpu_boost(cpu);
5213 boost = get_sysctl_sched_cfs_boost();
5218 return schedtune_margin(util, boost);
5221 static inline unsigned long
5222 schedtune_task_margin(struct task_struct *task)
5226 unsigned long margin;
5228 #ifdef CONFIG_CGROUP_SCHEDTUNE
5229 boost = schedtune_task_boost(task);
5231 boost = get_sysctl_sched_cfs_boost();
5236 util = task_util(task);
5237 margin = schedtune_margin(util, boost);
5242 #else /* CONFIG_SCHED_TUNE */
5244 static inline unsigned int
5245 schedtune_cpu_margin(unsigned long util, int cpu)
5250 static inline unsigned int
5251 schedtune_task_margin(struct task_struct *task)
5256 #endif /* CONFIG_SCHED_TUNE */
5258 static inline unsigned long
5259 boosted_cpu_util(int cpu)
5261 unsigned long util = cpu_util(cpu);
5262 unsigned long margin = schedtune_cpu_margin(util, cpu);
5264 return util + margin;
5267 static inline unsigned long
5268 boosted_task_util(struct task_struct *task)
5270 unsigned long util = task_util(task);
5271 unsigned long margin = schedtune_task_margin(task);
5273 return util + margin;
5277 * find_idlest_group finds and returns the least busy CPU group within the
5280 static struct sched_group *
5281 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5282 int this_cpu, int sd_flag)
5284 struct sched_group *idlest = NULL, *group = sd->groups;
5285 struct sched_group *fit_group = NULL, *spare_group = NULL;
5286 unsigned long min_load = ULONG_MAX, this_load = 0;
5287 unsigned long fit_capacity = ULONG_MAX;
5288 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5289 int load_idx = sd->forkexec_idx;
5290 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5292 if (sd_flag & SD_BALANCE_WAKE)
5293 load_idx = sd->wake_idx;
5296 unsigned long load, avg_load, spare_capacity;
5300 /* Skip over this group if it has no CPUs allowed */
5301 if (!cpumask_intersects(sched_group_cpus(group),
5302 tsk_cpus_allowed(p)))
5305 local_group = cpumask_test_cpu(this_cpu,
5306 sched_group_cpus(group));
5308 /* Tally up the load of all CPUs in the group */
5311 for_each_cpu(i, sched_group_cpus(group)) {
5312 /* Bias balancing toward cpus of our domain */
5314 load = source_load(i, load_idx);
5316 load = target_load(i, load_idx);
5321 * Look for most energy-efficient group that can fit
5322 * that can fit the task.
5324 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5325 fit_capacity = capacity_of(i);
5330 * Look for group which has most spare capacity on a
5333 spare_capacity = capacity_of(i) - cpu_util(i);
5334 if (spare_capacity > max_spare_capacity) {
5335 max_spare_capacity = spare_capacity;
5336 spare_group = group;
5340 /* Adjust by relative CPU capacity of the group */
5341 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5344 this_load = avg_load;
5345 } else if (avg_load < min_load) {
5346 min_load = avg_load;
5349 } while (group = group->next, group != sd->groups);
5357 if (!idlest || 100*this_load < imbalance*min_load)
5363 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5366 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5368 unsigned long load, min_load = ULONG_MAX;
5369 unsigned int min_exit_latency = UINT_MAX;
5370 u64 latest_idle_timestamp = 0;
5371 int least_loaded_cpu = this_cpu;
5372 int shallowest_idle_cpu = -1;
5375 /* Traverse only the allowed CPUs */
5376 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5377 if (task_fits_spare(p, i)) {
5378 struct rq *rq = cpu_rq(i);
5379 struct cpuidle_state *idle = idle_get_state(rq);
5380 if (idle && idle->exit_latency < min_exit_latency) {
5382 * We give priority to a CPU whose idle state
5383 * has the smallest exit latency irrespective
5384 * of any idle timestamp.
5386 min_exit_latency = idle->exit_latency;
5387 latest_idle_timestamp = rq->idle_stamp;
5388 shallowest_idle_cpu = i;
5389 } else if (idle_cpu(i) &&
5390 (!idle || idle->exit_latency == min_exit_latency) &&
5391 rq->idle_stamp > latest_idle_timestamp) {
5393 * If equal or no active idle state, then
5394 * the most recently idled CPU might have
5397 latest_idle_timestamp = rq->idle_stamp;
5398 shallowest_idle_cpu = i;
5399 } else if (shallowest_idle_cpu == -1) {
5401 * If we haven't found an idle CPU yet
5402 * pick a non-idle one that can fit the task as
5405 shallowest_idle_cpu = i;
5407 } else if (shallowest_idle_cpu == -1) {
5408 load = weighted_cpuload(i);
5409 if (load < min_load || (load == min_load && i == this_cpu)) {
5411 least_loaded_cpu = i;
5416 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5420 * Try and locate an idle CPU in the sched_domain.
5422 static int select_idle_sibling(struct task_struct *p, int target)
5424 struct sched_domain *sd;
5425 struct sched_group *sg;
5426 int i = task_cpu(p);
5428 if (idle_cpu(target))
5432 * If the prevous cpu is cache affine and idle, don't be stupid.
5434 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5438 * Otherwise, iterate the domains and find an elegible idle cpu.
5440 sd = rcu_dereference(per_cpu(sd_llc, target));
5441 for_each_lower_domain(sd) {
5444 if (!cpumask_intersects(sched_group_cpus(sg),
5445 tsk_cpus_allowed(p)))
5448 for_each_cpu(i, sched_group_cpus(sg)) {
5449 if (i == target || !idle_cpu(i))
5453 target = cpumask_first_and(sched_group_cpus(sg),
5454 tsk_cpus_allowed(p));
5458 } while (sg != sd->groups);
5464 static int energy_aware_wake_cpu(struct task_struct *p, int target)
5466 struct sched_domain *sd;
5467 struct sched_group *sg, *sg_target;
5468 int target_max_cap = INT_MAX;
5469 int target_cpu = task_cpu(p);
5472 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5481 * Find group with sufficient capacity. We only get here if no cpu is
5482 * overutilized. We may end up overutilizing a cpu by adding the task,
5483 * but that should not be any worse than select_idle_sibling().
5484 * load_balance() should sort it out later as we get above the tipping
5488 /* Assuming all cpus are the same in group */
5489 int max_cap_cpu = group_first_cpu(sg);
5492 * Assume smaller max capacity means more energy-efficient.
5493 * Ideally we should query the energy model for the right
5494 * answer but it easily ends up in an exhaustive search.
5496 if (capacity_of(max_cap_cpu) < target_max_cap &&
5497 task_fits_max(p, max_cap_cpu)) {
5499 target_max_cap = capacity_of(max_cap_cpu);
5501 } while (sg = sg->next, sg != sd->groups);
5503 /* Find cpu with sufficient capacity */
5504 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5506 * p's blocked utilization is still accounted for on prev_cpu
5507 * so prev_cpu will receive a negative bias due to the double
5508 * accounting. However, the blocked utilization may be zero.
5510 int new_util = cpu_util(i) + boosted_task_util(p);
5512 if (new_util > capacity_orig_of(i))
5515 if (new_util < capacity_curr_of(i)) {
5517 if (cpu_rq(i)->nr_running)
5521 /* cpu has capacity at higher OPP, keep it as fallback */
5522 if (target_cpu == task_cpu(p))
5526 if (target_cpu != task_cpu(p)) {
5527 struct energy_env eenv = {
5528 .util_delta = task_util(p),
5529 .src_cpu = task_cpu(p),
5530 .dst_cpu = target_cpu,
5534 /* Not enough spare capacity on previous cpu */
5535 if (cpu_overutilized(task_cpu(p)))
5538 if (energy_diff(&eenv) >= 0)
5546 * select_task_rq_fair: Select target runqueue for the waking task in domains
5547 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5548 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5550 * Balances load by selecting the idlest cpu in the idlest group, or under
5551 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5553 * Returns the target cpu number.
5555 * preempt must be disabled.
5558 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5560 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5561 int cpu = smp_processor_id();
5562 int new_cpu = prev_cpu;
5563 int want_affine = 0;
5564 int sync = wake_flags & WF_SYNC;
5566 if (sd_flag & SD_BALANCE_WAKE)
5567 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5568 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5572 for_each_domain(cpu, tmp) {
5573 if (!(tmp->flags & SD_LOAD_BALANCE))
5577 * If both cpu and prev_cpu are part of this domain,
5578 * cpu is a valid SD_WAKE_AFFINE target.
5580 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5581 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5586 if (tmp->flags & sd_flag)
5588 else if (!want_affine)
5593 sd = NULL; /* Prefer wake_affine over balance flags */
5594 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5599 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5600 new_cpu = energy_aware_wake_cpu(p, prev_cpu);
5601 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5602 new_cpu = select_idle_sibling(p, new_cpu);
5605 struct sched_group *group;
5608 if (!(sd->flags & sd_flag)) {
5613 group = find_idlest_group(sd, p, cpu, sd_flag);
5619 new_cpu = find_idlest_cpu(group, p, cpu);
5620 if (new_cpu == -1 || new_cpu == cpu) {
5621 /* Now try balancing at a lower domain level of cpu */
5626 /* Now try balancing at a lower domain level of new_cpu */
5628 weight = sd->span_weight;
5630 for_each_domain(cpu, tmp) {
5631 if (weight <= tmp->span_weight)
5633 if (tmp->flags & sd_flag)
5636 /* while loop will break here if sd == NULL */
5644 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5645 * cfs_rq_of(p) references at time of call are still valid and identify the
5646 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5647 * other assumptions, including the state of rq->lock, should be made.
5649 static void migrate_task_rq_fair(struct task_struct *p)
5652 * We are supposed to update the task to "current" time, then its up to date
5653 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5654 * what current time is, so simply throw away the out-of-date time. This
5655 * will result in the wakee task is less decayed, but giving the wakee more
5656 * load sounds not bad.
5658 remove_entity_load_avg(&p->se);
5660 /* Tell new CPU we are migrated */
5661 p->se.avg.last_update_time = 0;
5663 /* We have migrated, no longer consider this task hot */
5664 p->se.exec_start = 0;
5667 static void task_dead_fair(struct task_struct *p)
5669 remove_entity_load_avg(&p->se);
5671 #endif /* CONFIG_SMP */
5673 static unsigned long
5674 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5676 unsigned long gran = sysctl_sched_wakeup_granularity;
5679 * Since its curr running now, convert the gran from real-time
5680 * to virtual-time in his units.
5682 * By using 'se' instead of 'curr' we penalize light tasks, so
5683 * they get preempted easier. That is, if 'se' < 'curr' then
5684 * the resulting gran will be larger, therefore penalizing the
5685 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5686 * be smaller, again penalizing the lighter task.
5688 * This is especially important for buddies when the leftmost
5689 * task is higher priority than the buddy.
5691 return calc_delta_fair(gran, se);
5695 * Should 'se' preempt 'curr'.
5709 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5711 s64 gran, vdiff = curr->vruntime - se->vruntime;
5716 gran = wakeup_gran(curr, se);
5723 static void set_last_buddy(struct sched_entity *se)
5725 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5728 for_each_sched_entity(se)
5729 cfs_rq_of(se)->last = se;
5732 static void set_next_buddy(struct sched_entity *se)
5734 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5737 for_each_sched_entity(se)
5738 cfs_rq_of(se)->next = se;
5741 static void set_skip_buddy(struct sched_entity *se)
5743 for_each_sched_entity(se)
5744 cfs_rq_of(se)->skip = se;
5748 * Preempt the current task with a newly woken task if needed:
5750 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5752 struct task_struct *curr = rq->curr;
5753 struct sched_entity *se = &curr->se, *pse = &p->se;
5754 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5755 int scale = cfs_rq->nr_running >= sched_nr_latency;
5756 int next_buddy_marked = 0;
5758 if (unlikely(se == pse))
5762 * This is possible from callers such as attach_tasks(), in which we
5763 * unconditionally check_prempt_curr() after an enqueue (which may have
5764 * lead to a throttle). This both saves work and prevents false
5765 * next-buddy nomination below.
5767 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5770 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5771 set_next_buddy(pse);
5772 next_buddy_marked = 1;
5776 * We can come here with TIF_NEED_RESCHED already set from new task
5779 * Note: this also catches the edge-case of curr being in a throttled
5780 * group (e.g. via set_curr_task), since update_curr() (in the
5781 * enqueue of curr) will have resulted in resched being set. This
5782 * prevents us from potentially nominating it as a false LAST_BUDDY
5785 if (test_tsk_need_resched(curr))
5788 /* Idle tasks are by definition preempted by non-idle tasks. */
5789 if (unlikely(curr->policy == SCHED_IDLE) &&
5790 likely(p->policy != SCHED_IDLE))
5794 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5795 * is driven by the tick):
5797 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5800 find_matching_se(&se, &pse);
5801 update_curr(cfs_rq_of(se));
5803 if (wakeup_preempt_entity(se, pse) == 1) {
5805 * Bias pick_next to pick the sched entity that is
5806 * triggering this preemption.
5808 if (!next_buddy_marked)
5809 set_next_buddy(pse);
5818 * Only set the backward buddy when the current task is still
5819 * on the rq. This can happen when a wakeup gets interleaved
5820 * with schedule on the ->pre_schedule() or idle_balance()
5821 * point, either of which can * drop the rq lock.
5823 * Also, during early boot the idle thread is in the fair class,
5824 * for obvious reasons its a bad idea to schedule back to it.
5826 if (unlikely(!se->on_rq || curr == rq->idle))
5829 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5833 static struct task_struct *
5834 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5836 struct cfs_rq *cfs_rq = &rq->cfs;
5837 struct sched_entity *se;
5838 struct task_struct *p;
5842 #ifdef CONFIG_FAIR_GROUP_SCHED
5843 if (!cfs_rq->nr_running)
5846 if (prev->sched_class != &fair_sched_class)
5850 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5851 * likely that a next task is from the same cgroup as the current.
5853 * Therefore attempt to avoid putting and setting the entire cgroup
5854 * hierarchy, only change the part that actually changes.
5858 struct sched_entity *curr = cfs_rq->curr;
5861 * Since we got here without doing put_prev_entity() we also
5862 * have to consider cfs_rq->curr. If it is still a runnable
5863 * entity, update_curr() will update its vruntime, otherwise
5864 * forget we've ever seen it.
5868 update_curr(cfs_rq);
5873 * This call to check_cfs_rq_runtime() will do the
5874 * throttle and dequeue its entity in the parent(s).
5875 * Therefore the 'simple' nr_running test will indeed
5878 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5882 se = pick_next_entity(cfs_rq, curr);
5883 cfs_rq = group_cfs_rq(se);
5889 * Since we haven't yet done put_prev_entity and if the selected task
5890 * is a different task than we started out with, try and touch the
5891 * least amount of cfs_rqs.
5894 struct sched_entity *pse = &prev->se;
5896 while (!(cfs_rq = is_same_group(se, pse))) {
5897 int se_depth = se->depth;
5898 int pse_depth = pse->depth;
5900 if (se_depth <= pse_depth) {
5901 put_prev_entity(cfs_rq_of(pse), pse);
5902 pse = parent_entity(pse);
5904 if (se_depth >= pse_depth) {
5905 set_next_entity(cfs_rq_of(se), se);
5906 se = parent_entity(se);
5910 put_prev_entity(cfs_rq, pse);
5911 set_next_entity(cfs_rq, se);
5914 if (hrtick_enabled(rq))
5915 hrtick_start_fair(rq, p);
5917 rq->misfit_task = !task_fits_max(p, rq->cpu);
5924 if (!cfs_rq->nr_running)
5927 put_prev_task(rq, prev);
5930 se = pick_next_entity(cfs_rq, NULL);
5931 set_next_entity(cfs_rq, se);
5932 cfs_rq = group_cfs_rq(se);
5937 if (hrtick_enabled(rq))
5938 hrtick_start_fair(rq, p);
5940 rq->misfit_task = !task_fits_max(p, rq->cpu);
5945 rq->misfit_task = 0;
5947 * This is OK, because current is on_cpu, which avoids it being picked
5948 * for load-balance and preemption/IRQs are still disabled avoiding
5949 * further scheduler activity on it and we're being very careful to
5950 * re-start the picking loop.
5952 lockdep_unpin_lock(&rq->lock);
5953 new_tasks = idle_balance(rq);
5954 lockdep_pin_lock(&rq->lock);
5956 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5957 * possible for any higher priority task to appear. In that case we
5958 * must re-start the pick_next_entity() loop.
5970 * Account for a descheduled task:
5972 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5974 struct sched_entity *se = &prev->se;
5975 struct cfs_rq *cfs_rq;
5977 for_each_sched_entity(se) {
5978 cfs_rq = cfs_rq_of(se);
5979 put_prev_entity(cfs_rq, se);
5984 * sched_yield() is very simple
5986 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5988 static void yield_task_fair(struct rq *rq)
5990 struct task_struct *curr = rq->curr;
5991 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5992 struct sched_entity *se = &curr->se;
5995 * Are we the only task in the tree?
5997 if (unlikely(rq->nr_running == 1))
6000 clear_buddies(cfs_rq, se);
6002 if (curr->policy != SCHED_BATCH) {
6003 update_rq_clock(rq);
6005 * Update run-time statistics of the 'current'.
6007 update_curr(cfs_rq);
6009 * Tell update_rq_clock() that we've just updated,
6010 * so we don't do microscopic update in schedule()
6011 * and double the fastpath cost.
6013 rq_clock_skip_update(rq, true);
6019 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6021 struct sched_entity *se = &p->se;
6023 /* throttled hierarchies are not runnable */
6024 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6027 /* Tell the scheduler that we'd really like pse to run next. */
6030 yield_task_fair(rq);
6036 /**************************************************
6037 * Fair scheduling class load-balancing methods.
6041 * The purpose of load-balancing is to achieve the same basic fairness the
6042 * per-cpu scheduler provides, namely provide a proportional amount of compute
6043 * time to each task. This is expressed in the following equation:
6045 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6047 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6048 * W_i,0 is defined as:
6050 * W_i,0 = \Sum_j w_i,j (2)
6052 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6053 * is derived from the nice value as per prio_to_weight[].
6055 * The weight average is an exponential decay average of the instantaneous
6058 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6060 * C_i is the compute capacity of cpu i, typically it is the
6061 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6062 * can also include other factors [XXX].
6064 * To achieve this balance we define a measure of imbalance which follows
6065 * directly from (1):
6067 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6069 * We them move tasks around to minimize the imbalance. In the continuous
6070 * function space it is obvious this converges, in the discrete case we get
6071 * a few fun cases generally called infeasible weight scenarios.
6074 * - infeasible weights;
6075 * - local vs global optima in the discrete case. ]
6080 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6081 * for all i,j solution, we create a tree of cpus that follows the hardware
6082 * topology where each level pairs two lower groups (or better). This results
6083 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6084 * tree to only the first of the previous level and we decrease the frequency
6085 * of load-balance at each level inv. proportional to the number of cpus in
6091 * \Sum { --- * --- * 2^i } = O(n) (5)
6093 * `- size of each group
6094 * | | `- number of cpus doing load-balance
6096 * `- sum over all levels
6098 * Coupled with a limit on how many tasks we can migrate every balance pass,
6099 * this makes (5) the runtime complexity of the balancer.
6101 * An important property here is that each CPU is still (indirectly) connected
6102 * to every other cpu in at most O(log n) steps:
6104 * The adjacency matrix of the resulting graph is given by:
6107 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6110 * And you'll find that:
6112 * A^(log_2 n)_i,j != 0 for all i,j (7)
6114 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6115 * The task movement gives a factor of O(m), giving a convergence complexity
6118 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6123 * In order to avoid CPUs going idle while there's still work to do, new idle
6124 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6125 * tree itself instead of relying on other CPUs to bring it work.
6127 * This adds some complexity to both (5) and (8) but it reduces the total idle
6135 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6138 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6143 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6145 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6147 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6150 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6151 * rewrite all of this once again.]
6154 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6156 enum fbq_type { regular, remote, all };
6165 #define LBF_ALL_PINNED 0x01
6166 #define LBF_NEED_BREAK 0x02
6167 #define LBF_DST_PINNED 0x04
6168 #define LBF_SOME_PINNED 0x08
6171 struct sched_domain *sd;
6179 struct cpumask *dst_grpmask;
6181 enum cpu_idle_type idle;
6183 unsigned int src_grp_nr_running;
6184 /* The set of CPUs under consideration for load-balancing */
6185 struct cpumask *cpus;
6190 unsigned int loop_break;
6191 unsigned int loop_max;
6193 enum fbq_type fbq_type;
6194 enum group_type busiest_group_type;
6195 struct list_head tasks;
6199 * Is this task likely cache-hot:
6201 static int task_hot(struct task_struct *p, struct lb_env *env)
6205 lockdep_assert_held(&env->src_rq->lock);
6207 if (p->sched_class != &fair_sched_class)
6210 if (unlikely(p->policy == SCHED_IDLE))
6214 * Buddy candidates are cache hot:
6216 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6217 (&p->se == cfs_rq_of(&p->se)->next ||
6218 &p->se == cfs_rq_of(&p->se)->last))
6221 if (sysctl_sched_migration_cost == -1)
6223 if (sysctl_sched_migration_cost == 0)
6226 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6228 return delta < (s64)sysctl_sched_migration_cost;
6231 #ifdef CONFIG_NUMA_BALANCING
6233 * Returns 1, if task migration degrades locality
6234 * Returns 0, if task migration improves locality i.e migration preferred.
6235 * Returns -1, if task migration is not affected by locality.
6237 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6239 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6240 unsigned long src_faults, dst_faults;
6241 int src_nid, dst_nid;
6243 if (!static_branch_likely(&sched_numa_balancing))
6246 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6249 src_nid = cpu_to_node(env->src_cpu);
6250 dst_nid = cpu_to_node(env->dst_cpu);
6252 if (src_nid == dst_nid)
6255 /* Migrating away from the preferred node is always bad. */
6256 if (src_nid == p->numa_preferred_nid) {
6257 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6263 /* Encourage migration to the preferred node. */
6264 if (dst_nid == p->numa_preferred_nid)
6268 src_faults = group_faults(p, src_nid);
6269 dst_faults = group_faults(p, dst_nid);
6271 src_faults = task_faults(p, src_nid);
6272 dst_faults = task_faults(p, dst_nid);
6275 return dst_faults < src_faults;
6279 static inline int migrate_degrades_locality(struct task_struct *p,
6287 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6290 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6294 lockdep_assert_held(&env->src_rq->lock);
6297 * We do not migrate tasks that are:
6298 * 1) throttled_lb_pair, or
6299 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6300 * 3) running (obviously), or
6301 * 4) are cache-hot on their current CPU.
6303 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6306 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6309 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6311 env->flags |= LBF_SOME_PINNED;
6314 * Remember if this task can be migrated to any other cpu in
6315 * our sched_group. We may want to revisit it if we couldn't
6316 * meet load balance goals by pulling other tasks on src_cpu.
6318 * Also avoid computing new_dst_cpu if we have already computed
6319 * one in current iteration.
6321 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6324 /* Prevent to re-select dst_cpu via env's cpus */
6325 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6326 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6327 env->flags |= LBF_DST_PINNED;
6328 env->new_dst_cpu = cpu;
6336 /* Record that we found atleast one task that could run on dst_cpu */
6337 env->flags &= ~LBF_ALL_PINNED;
6339 if (task_running(env->src_rq, p)) {
6340 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6345 * Aggressive migration if:
6346 * 1) destination numa is preferred
6347 * 2) task is cache cold, or
6348 * 3) too many balance attempts have failed.
6350 tsk_cache_hot = migrate_degrades_locality(p, env);
6351 if (tsk_cache_hot == -1)
6352 tsk_cache_hot = task_hot(p, env);
6354 if (tsk_cache_hot <= 0 ||
6355 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6356 if (tsk_cache_hot == 1) {
6357 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6358 schedstat_inc(p, se.statistics.nr_forced_migrations);
6363 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6368 * detach_task() -- detach the task for the migration specified in env
6370 static void detach_task(struct task_struct *p, struct lb_env *env)
6372 lockdep_assert_held(&env->src_rq->lock);
6374 deactivate_task(env->src_rq, p, 0);
6375 p->on_rq = TASK_ON_RQ_MIGRATING;
6376 set_task_cpu(p, env->dst_cpu);
6380 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6381 * part of active balancing operations within "domain".
6383 * Returns a task if successful and NULL otherwise.
6385 static struct task_struct *detach_one_task(struct lb_env *env)
6387 struct task_struct *p, *n;
6389 lockdep_assert_held(&env->src_rq->lock);
6391 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6392 if (!can_migrate_task(p, env))
6395 detach_task(p, env);
6398 * Right now, this is only the second place where
6399 * lb_gained[env->idle] is updated (other is detach_tasks)
6400 * so we can safely collect stats here rather than
6401 * inside detach_tasks().
6403 schedstat_inc(env->sd, lb_gained[env->idle]);
6409 static const unsigned int sched_nr_migrate_break = 32;
6412 * detach_tasks() -- tries to detach up to imbalance weighted load from
6413 * busiest_rq, as part of a balancing operation within domain "sd".
6415 * Returns number of detached tasks if successful and 0 otherwise.
6417 static int detach_tasks(struct lb_env *env)
6419 struct list_head *tasks = &env->src_rq->cfs_tasks;
6420 struct task_struct *p;
6424 lockdep_assert_held(&env->src_rq->lock);
6426 if (env->imbalance <= 0)
6429 while (!list_empty(tasks)) {
6431 * We don't want to steal all, otherwise we may be treated likewise,
6432 * which could at worst lead to a livelock crash.
6434 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6437 p = list_first_entry(tasks, struct task_struct, se.group_node);
6440 /* We've more or less seen every task there is, call it quits */
6441 if (env->loop > env->loop_max)
6444 /* take a breather every nr_migrate tasks */
6445 if (env->loop > env->loop_break) {
6446 env->loop_break += sched_nr_migrate_break;
6447 env->flags |= LBF_NEED_BREAK;
6451 if (!can_migrate_task(p, env))
6454 load = task_h_load(p);
6456 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6459 if ((load / 2) > env->imbalance)
6462 detach_task(p, env);
6463 list_add(&p->se.group_node, &env->tasks);
6466 env->imbalance -= load;
6468 #ifdef CONFIG_PREEMPT
6470 * NEWIDLE balancing is a source of latency, so preemptible
6471 * kernels will stop after the first task is detached to minimize
6472 * the critical section.
6474 if (env->idle == CPU_NEWLY_IDLE)
6479 * We only want to steal up to the prescribed amount of
6482 if (env->imbalance <= 0)
6487 list_move_tail(&p->se.group_node, tasks);
6491 * Right now, this is one of only two places we collect this stat
6492 * so we can safely collect detach_one_task() stats here rather
6493 * than inside detach_one_task().
6495 schedstat_add(env->sd, lb_gained[env->idle], detached);
6501 * attach_task() -- attach the task detached by detach_task() to its new rq.
6503 static void attach_task(struct rq *rq, struct task_struct *p)
6505 lockdep_assert_held(&rq->lock);
6507 BUG_ON(task_rq(p) != rq);
6508 p->on_rq = TASK_ON_RQ_QUEUED;
6509 activate_task(rq, p, 0);
6510 check_preempt_curr(rq, p, 0);
6514 * attach_one_task() -- attaches the task returned from detach_one_task() to
6517 static void attach_one_task(struct rq *rq, struct task_struct *p)
6519 raw_spin_lock(&rq->lock);
6522 * We want to potentially raise target_cpu's OPP.
6524 update_capacity_of(cpu_of(rq));
6525 raw_spin_unlock(&rq->lock);
6529 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6532 static void attach_tasks(struct lb_env *env)
6534 struct list_head *tasks = &env->tasks;
6535 struct task_struct *p;
6537 raw_spin_lock(&env->dst_rq->lock);
6539 while (!list_empty(tasks)) {
6540 p = list_first_entry(tasks, struct task_struct, se.group_node);
6541 list_del_init(&p->se.group_node);
6543 attach_task(env->dst_rq, p);
6547 * We want to potentially raise env.dst_cpu's OPP.
6549 update_capacity_of(env->dst_cpu);
6551 raw_spin_unlock(&env->dst_rq->lock);
6554 #ifdef CONFIG_FAIR_GROUP_SCHED
6555 static void update_blocked_averages(int cpu)
6557 struct rq *rq = cpu_rq(cpu);
6558 struct cfs_rq *cfs_rq;
6559 unsigned long flags;
6561 raw_spin_lock_irqsave(&rq->lock, flags);
6562 update_rq_clock(rq);
6565 * Iterates the task_group tree in a bottom up fashion, see
6566 * list_add_leaf_cfs_rq() for details.
6568 for_each_leaf_cfs_rq(rq, cfs_rq) {
6569 /* throttled entities do not contribute to load */
6570 if (throttled_hierarchy(cfs_rq))
6573 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6574 update_tg_load_avg(cfs_rq, 0);
6576 raw_spin_unlock_irqrestore(&rq->lock, flags);
6580 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6581 * This needs to be done in a top-down fashion because the load of a child
6582 * group is a fraction of its parents load.
6584 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6586 struct rq *rq = rq_of(cfs_rq);
6587 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6588 unsigned long now = jiffies;
6591 if (cfs_rq->last_h_load_update == now)
6594 cfs_rq->h_load_next = NULL;
6595 for_each_sched_entity(se) {
6596 cfs_rq = cfs_rq_of(se);
6597 cfs_rq->h_load_next = se;
6598 if (cfs_rq->last_h_load_update == now)
6603 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6604 cfs_rq->last_h_load_update = now;
6607 while ((se = cfs_rq->h_load_next) != NULL) {
6608 load = cfs_rq->h_load;
6609 load = div64_ul(load * se->avg.load_avg,
6610 cfs_rq_load_avg(cfs_rq) + 1);
6611 cfs_rq = group_cfs_rq(se);
6612 cfs_rq->h_load = load;
6613 cfs_rq->last_h_load_update = now;
6617 static unsigned long task_h_load(struct task_struct *p)
6619 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6621 update_cfs_rq_h_load(cfs_rq);
6622 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6623 cfs_rq_load_avg(cfs_rq) + 1);
6626 static inline void update_blocked_averages(int cpu)
6628 struct rq *rq = cpu_rq(cpu);
6629 struct cfs_rq *cfs_rq = &rq->cfs;
6630 unsigned long flags;
6632 raw_spin_lock_irqsave(&rq->lock, flags);
6633 update_rq_clock(rq);
6634 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6635 raw_spin_unlock_irqrestore(&rq->lock, flags);
6638 static unsigned long task_h_load(struct task_struct *p)
6640 return p->se.avg.load_avg;
6644 /********** Helpers for find_busiest_group ************************/
6647 * sg_lb_stats - stats of a sched_group required for load_balancing
6649 struct sg_lb_stats {
6650 unsigned long avg_load; /*Avg load across the CPUs of the group */
6651 unsigned long group_load; /* Total load over the CPUs of the group */
6652 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6653 unsigned long load_per_task;
6654 unsigned long group_capacity;
6655 unsigned long group_util; /* Total utilization of the group */
6656 unsigned int sum_nr_running; /* Nr tasks running in the group */
6657 unsigned int idle_cpus;
6658 unsigned int group_weight;
6659 enum group_type group_type;
6660 int group_no_capacity;
6661 int group_misfit_task; /* A cpu has a task too big for its capacity */
6662 #ifdef CONFIG_NUMA_BALANCING
6663 unsigned int nr_numa_running;
6664 unsigned int nr_preferred_running;
6669 * sd_lb_stats - Structure to store the statistics of a sched_domain
6670 * during load balancing.
6672 struct sd_lb_stats {
6673 struct sched_group *busiest; /* Busiest group in this sd */
6674 struct sched_group *local; /* Local group in this sd */
6675 unsigned long total_load; /* Total load of all groups in sd */
6676 unsigned long total_capacity; /* Total capacity of all groups in sd */
6677 unsigned long avg_load; /* Average load across all groups in sd */
6679 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6680 struct sg_lb_stats local_stat; /* Statistics of the local group */
6683 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6686 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6687 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6688 * We must however clear busiest_stat::avg_load because
6689 * update_sd_pick_busiest() reads this before assignment.
6691 *sds = (struct sd_lb_stats){
6695 .total_capacity = 0UL,
6698 .sum_nr_running = 0,
6699 .group_type = group_other,
6705 * get_sd_load_idx - Obtain the load index for a given sched domain.
6706 * @sd: The sched_domain whose load_idx is to be obtained.
6707 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6709 * Return: The load index.
6711 static inline int get_sd_load_idx(struct sched_domain *sd,
6712 enum cpu_idle_type idle)
6718 load_idx = sd->busy_idx;
6721 case CPU_NEWLY_IDLE:
6722 load_idx = sd->newidle_idx;
6725 load_idx = sd->idle_idx;
6732 static unsigned long scale_rt_capacity(int cpu)
6734 struct rq *rq = cpu_rq(cpu);
6735 u64 total, used, age_stamp, avg;
6739 * Since we're reading these variables without serialization make sure
6740 * we read them once before doing sanity checks on them.
6742 age_stamp = READ_ONCE(rq->age_stamp);
6743 avg = READ_ONCE(rq->rt_avg);
6744 delta = __rq_clock_broken(rq) - age_stamp;
6746 if (unlikely(delta < 0))
6749 total = sched_avg_period() + delta;
6751 used = div_u64(avg, total);
6754 * deadline bandwidth is defined at system level so we must
6755 * weight this bandwidth with the max capacity of the system.
6756 * As a reminder, avg_bw is 20bits width and
6757 * scale_cpu_capacity is 10 bits width
6759 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
6761 if (likely(used < SCHED_CAPACITY_SCALE))
6762 return SCHED_CAPACITY_SCALE - used;
6767 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
6769 raw_spin_lock_init(&mcc->lock);
6774 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6776 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6777 struct sched_group *sdg = sd->groups;
6778 struct max_cpu_capacity *mcc;
6779 unsigned long max_capacity;
6781 unsigned long flags;
6783 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6785 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
6787 raw_spin_lock_irqsave(&mcc->lock, flags);
6788 max_capacity = mcc->val;
6789 max_cap_cpu = mcc->cpu;
6791 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
6792 (max_capacity < capacity)) {
6793 mcc->val = capacity;
6795 #ifdef CONFIG_SCHED_DEBUG
6796 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6797 pr_info("CPU%d: update max cpu_capacity %lu\n", cpu, capacity);
6801 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6803 skip_unlock: __attribute__ ((unused));
6804 capacity *= scale_rt_capacity(cpu);
6805 capacity >>= SCHED_CAPACITY_SHIFT;
6810 cpu_rq(cpu)->cpu_capacity = capacity;
6811 sdg->sgc->capacity = capacity;
6812 sdg->sgc->max_capacity = capacity;
6815 void update_group_capacity(struct sched_domain *sd, int cpu)
6817 struct sched_domain *child = sd->child;
6818 struct sched_group *group, *sdg = sd->groups;
6819 unsigned long capacity, max_capacity;
6820 unsigned long interval;
6822 interval = msecs_to_jiffies(sd->balance_interval);
6823 interval = clamp(interval, 1UL, max_load_balance_interval);
6824 sdg->sgc->next_update = jiffies + interval;
6827 update_cpu_capacity(sd, cpu);
6834 if (child->flags & SD_OVERLAP) {
6836 * SD_OVERLAP domains cannot assume that child groups
6837 * span the current group.
6840 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6841 struct sched_group_capacity *sgc;
6842 struct rq *rq = cpu_rq(cpu);
6845 * build_sched_domains() -> init_sched_groups_capacity()
6846 * gets here before we've attached the domains to the
6849 * Use capacity_of(), which is set irrespective of domains
6850 * in update_cpu_capacity().
6852 * This avoids capacity from being 0 and
6853 * causing divide-by-zero issues on boot.
6855 if (unlikely(!rq->sd)) {
6856 capacity += capacity_of(cpu);
6858 sgc = rq->sd->groups->sgc;
6859 capacity += sgc->capacity;
6862 max_capacity = max(capacity, max_capacity);
6866 * !SD_OVERLAP domains can assume that child groups
6867 * span the current group.
6870 group = child->groups;
6872 struct sched_group_capacity *sgc = group->sgc;
6874 capacity += sgc->capacity;
6875 max_capacity = max(sgc->max_capacity, max_capacity);
6876 group = group->next;
6877 } while (group != child->groups);
6880 sdg->sgc->capacity = capacity;
6881 sdg->sgc->max_capacity = max_capacity;
6885 * Check whether the capacity of the rq has been noticeably reduced by side
6886 * activity. The imbalance_pct is used for the threshold.
6887 * Return true is the capacity is reduced
6890 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6892 return ((rq->cpu_capacity * sd->imbalance_pct) <
6893 (rq->cpu_capacity_orig * 100));
6897 * Group imbalance indicates (and tries to solve) the problem where balancing
6898 * groups is inadequate due to tsk_cpus_allowed() constraints.
6900 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6901 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6904 * { 0 1 2 3 } { 4 5 6 7 }
6907 * If we were to balance group-wise we'd place two tasks in the first group and
6908 * two tasks in the second group. Clearly this is undesired as it will overload
6909 * cpu 3 and leave one of the cpus in the second group unused.
6911 * The current solution to this issue is detecting the skew in the first group
6912 * by noticing the lower domain failed to reach balance and had difficulty
6913 * moving tasks due to affinity constraints.
6915 * When this is so detected; this group becomes a candidate for busiest; see
6916 * update_sd_pick_busiest(). And calculate_imbalance() and
6917 * find_busiest_group() avoid some of the usual balance conditions to allow it
6918 * to create an effective group imbalance.
6920 * This is a somewhat tricky proposition since the next run might not find the
6921 * group imbalance and decide the groups need to be balanced again. A most
6922 * subtle and fragile situation.
6925 static inline int sg_imbalanced(struct sched_group *group)
6927 return group->sgc->imbalance;
6931 * group_has_capacity returns true if the group has spare capacity that could
6932 * be used by some tasks.
6933 * We consider that a group has spare capacity if the * number of task is
6934 * smaller than the number of CPUs or if the utilization is lower than the
6935 * available capacity for CFS tasks.
6936 * For the latter, we use a threshold to stabilize the state, to take into
6937 * account the variance of the tasks' load and to return true if the available
6938 * capacity in meaningful for the load balancer.
6939 * As an example, an available capacity of 1% can appear but it doesn't make
6940 * any benefit for the load balance.
6943 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6945 if (sgs->sum_nr_running < sgs->group_weight)
6948 if ((sgs->group_capacity * 100) >
6949 (sgs->group_util * env->sd->imbalance_pct))
6956 * group_is_overloaded returns true if the group has more tasks than it can
6958 * group_is_overloaded is not equals to !group_has_capacity because a group
6959 * with the exact right number of tasks, has no more spare capacity but is not
6960 * overloaded so both group_has_capacity and group_is_overloaded return
6964 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6966 if (sgs->sum_nr_running <= sgs->group_weight)
6969 if ((sgs->group_capacity * 100) <
6970 (sgs->group_util * env->sd->imbalance_pct))
6978 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
6979 * per-cpu capacity than sched_group ref.
6982 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
6984 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
6985 ref->sgc->max_capacity;
6989 group_type group_classify(struct sched_group *group,
6990 struct sg_lb_stats *sgs)
6992 if (sgs->group_no_capacity)
6993 return group_overloaded;
6995 if (sg_imbalanced(group))
6996 return group_imbalanced;
6998 if (sgs->group_misfit_task)
6999 return group_misfit_task;
7005 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7006 * @env: The load balancing environment.
7007 * @group: sched_group whose statistics are to be updated.
7008 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7009 * @local_group: Does group contain this_cpu.
7010 * @sgs: variable to hold the statistics for this group.
7011 * @overload: Indicate more than one runnable task for any CPU.
7012 * @overutilized: Indicate overutilization for any CPU.
7014 static inline void update_sg_lb_stats(struct lb_env *env,
7015 struct sched_group *group, int load_idx,
7016 int local_group, struct sg_lb_stats *sgs,
7017 bool *overload, bool *overutilized)
7022 memset(sgs, 0, sizeof(*sgs));
7024 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7025 struct rq *rq = cpu_rq(i);
7027 /* Bias balancing toward cpus of our domain */
7029 load = target_load(i, load_idx);
7031 load = source_load(i, load_idx);
7033 sgs->group_load += load;
7034 sgs->group_util += cpu_util(i);
7035 sgs->sum_nr_running += rq->cfs.h_nr_running;
7037 if (rq->nr_running > 1)
7040 #ifdef CONFIG_NUMA_BALANCING
7041 sgs->nr_numa_running += rq->nr_numa_running;
7042 sgs->nr_preferred_running += rq->nr_preferred_running;
7044 sgs->sum_weighted_load += weighted_cpuload(i);
7048 if (cpu_overutilized(i)) {
7049 *overutilized = true;
7050 if (!sgs->group_misfit_task && rq->misfit_task)
7051 sgs->group_misfit_task = capacity_of(i);
7055 /* Adjust by relative CPU capacity of the group */
7056 sgs->group_capacity = group->sgc->capacity;
7057 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7059 if (sgs->sum_nr_running)
7060 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7062 sgs->group_weight = group->group_weight;
7064 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7065 sgs->group_type = group_classify(group, sgs);
7069 * update_sd_pick_busiest - return 1 on busiest group
7070 * @env: The load balancing environment.
7071 * @sds: sched_domain statistics
7072 * @sg: sched_group candidate to be checked for being the busiest
7073 * @sgs: sched_group statistics
7075 * Determine if @sg is a busier group than the previously selected
7078 * Return: %true if @sg is a busier group than the previously selected
7079 * busiest group. %false otherwise.
7081 static bool update_sd_pick_busiest(struct lb_env *env,
7082 struct sd_lb_stats *sds,
7083 struct sched_group *sg,
7084 struct sg_lb_stats *sgs)
7086 struct sg_lb_stats *busiest = &sds->busiest_stat;
7088 if (sgs->group_type > busiest->group_type)
7091 if (sgs->group_type < busiest->group_type)
7095 * Candidate sg doesn't face any serious load-balance problems
7096 * so don't pick it if the local sg is already filled up.
7098 if (sgs->group_type == group_other &&
7099 !group_has_capacity(env, &sds->local_stat))
7102 if (sgs->avg_load <= busiest->avg_load)
7106 * Candiate sg has no more than one task per cpu and has higher
7107 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7109 if (sgs->sum_nr_running <= sgs->group_weight &&
7110 group_smaller_cpu_capacity(sds->local, sg))
7113 /* This is the busiest node in its class. */
7114 if (!(env->sd->flags & SD_ASYM_PACKING))
7118 * ASYM_PACKING needs to move all the work to the lowest
7119 * numbered CPUs in the group, therefore mark all groups
7120 * higher than ourself as busy.
7122 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7126 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7133 #ifdef CONFIG_NUMA_BALANCING
7134 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7136 if (sgs->sum_nr_running > sgs->nr_numa_running)
7138 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7143 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7145 if (rq->nr_running > rq->nr_numa_running)
7147 if (rq->nr_running > rq->nr_preferred_running)
7152 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7157 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7161 #endif /* CONFIG_NUMA_BALANCING */
7164 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7165 * @env: The load balancing environment.
7166 * @sds: variable to hold the statistics for this sched_domain.
7168 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7170 struct sched_domain *child = env->sd->child;
7171 struct sched_group *sg = env->sd->groups;
7172 struct sg_lb_stats tmp_sgs;
7173 int load_idx, prefer_sibling = 0;
7174 bool overload = false, overutilized = false;
7176 if (child && child->flags & SD_PREFER_SIBLING)
7179 load_idx = get_sd_load_idx(env->sd, env->idle);
7182 struct sg_lb_stats *sgs = &tmp_sgs;
7185 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7188 sgs = &sds->local_stat;
7190 if (env->idle != CPU_NEWLY_IDLE ||
7191 time_after_eq(jiffies, sg->sgc->next_update))
7192 update_group_capacity(env->sd, env->dst_cpu);
7195 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7196 &overload, &overutilized);
7202 * In case the child domain prefers tasks go to siblings
7203 * first, lower the sg capacity so that we'll try
7204 * and move all the excess tasks away. We lower the capacity
7205 * of a group only if the local group has the capacity to fit
7206 * these excess tasks. The extra check prevents the case where
7207 * you always pull from the heaviest group when it is already
7208 * under-utilized (possible with a large weight task outweighs
7209 * the tasks on the system).
7211 if (prefer_sibling && sds->local &&
7212 group_has_capacity(env, &sds->local_stat) &&
7213 (sgs->sum_nr_running > 1)) {
7214 sgs->group_no_capacity = 1;
7215 sgs->group_type = group_classify(sg, sgs);
7219 * Ignore task groups with misfit tasks if local group has no
7220 * capacity or if per-cpu capacity isn't higher.
7222 if (sgs->group_type == group_misfit_task &&
7223 (!group_has_capacity(env, &sds->local_stat) ||
7224 !group_smaller_cpu_capacity(sg, sds->local)))
7225 sgs->group_type = group_other;
7227 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7229 sds->busiest_stat = *sgs;
7233 /* Now, start updating sd_lb_stats */
7234 sds->total_load += sgs->group_load;
7235 sds->total_capacity += sgs->group_capacity;
7238 } while (sg != env->sd->groups);
7240 if (env->sd->flags & SD_NUMA)
7241 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7243 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7245 if (!env->sd->parent) {
7246 /* update overload indicator if we are at root domain */
7247 if (env->dst_rq->rd->overload != overload)
7248 env->dst_rq->rd->overload = overload;
7250 /* Update over-utilization (tipping point, U >= 0) indicator */
7251 if (env->dst_rq->rd->overutilized != overutilized)
7252 env->dst_rq->rd->overutilized = overutilized;
7254 if (!env->dst_rq->rd->overutilized && overutilized)
7255 env->dst_rq->rd->overutilized = true;
7260 * check_asym_packing - Check to see if the group is packed into the
7263 * This is primarily intended to used at the sibling level. Some
7264 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7265 * case of POWER7, it can move to lower SMT modes only when higher
7266 * threads are idle. When in lower SMT modes, the threads will
7267 * perform better since they share less core resources. Hence when we
7268 * have idle threads, we want them to be the higher ones.
7270 * This packing function is run on idle threads. It checks to see if
7271 * the busiest CPU in this domain (core in the P7 case) has a higher
7272 * CPU number than the packing function is being run on. Here we are
7273 * assuming lower CPU number will be equivalent to lower a SMT thread
7276 * Return: 1 when packing is required and a task should be moved to
7277 * this CPU. The amount of the imbalance is returned in *imbalance.
7279 * @env: The load balancing environment.
7280 * @sds: Statistics of the sched_domain which is to be packed
7282 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7286 if (!(env->sd->flags & SD_ASYM_PACKING))
7292 busiest_cpu = group_first_cpu(sds->busiest);
7293 if (env->dst_cpu > busiest_cpu)
7296 env->imbalance = DIV_ROUND_CLOSEST(
7297 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7298 SCHED_CAPACITY_SCALE);
7304 * fix_small_imbalance - Calculate the minor imbalance that exists
7305 * amongst the groups of a sched_domain, during
7307 * @env: The load balancing environment.
7308 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7311 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7313 unsigned long tmp, capa_now = 0, capa_move = 0;
7314 unsigned int imbn = 2;
7315 unsigned long scaled_busy_load_per_task;
7316 struct sg_lb_stats *local, *busiest;
7318 local = &sds->local_stat;
7319 busiest = &sds->busiest_stat;
7321 if (!local->sum_nr_running)
7322 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7323 else if (busiest->load_per_task > local->load_per_task)
7326 scaled_busy_load_per_task =
7327 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7328 busiest->group_capacity;
7330 if (busiest->avg_load + scaled_busy_load_per_task >=
7331 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7332 env->imbalance = busiest->load_per_task;
7337 * OK, we don't have enough imbalance to justify moving tasks,
7338 * however we may be able to increase total CPU capacity used by
7342 capa_now += busiest->group_capacity *
7343 min(busiest->load_per_task, busiest->avg_load);
7344 capa_now += local->group_capacity *
7345 min(local->load_per_task, local->avg_load);
7346 capa_now /= SCHED_CAPACITY_SCALE;
7348 /* Amount of load we'd subtract */
7349 if (busiest->avg_load > scaled_busy_load_per_task) {
7350 capa_move += busiest->group_capacity *
7351 min(busiest->load_per_task,
7352 busiest->avg_load - scaled_busy_load_per_task);
7355 /* Amount of load we'd add */
7356 if (busiest->avg_load * busiest->group_capacity <
7357 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7358 tmp = (busiest->avg_load * busiest->group_capacity) /
7359 local->group_capacity;
7361 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7362 local->group_capacity;
7364 capa_move += local->group_capacity *
7365 min(local->load_per_task, local->avg_load + tmp);
7366 capa_move /= SCHED_CAPACITY_SCALE;
7368 /* Move if we gain throughput */
7369 if (capa_move > capa_now)
7370 env->imbalance = busiest->load_per_task;
7374 * calculate_imbalance - Calculate the amount of imbalance present within the
7375 * groups of a given sched_domain during load balance.
7376 * @env: load balance environment
7377 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7379 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7381 unsigned long max_pull, load_above_capacity = ~0UL;
7382 struct sg_lb_stats *local, *busiest;
7384 local = &sds->local_stat;
7385 busiest = &sds->busiest_stat;
7387 if (busiest->group_type == group_imbalanced) {
7389 * In the group_imb case we cannot rely on group-wide averages
7390 * to ensure cpu-load equilibrium, look at wider averages. XXX
7392 busiest->load_per_task =
7393 min(busiest->load_per_task, sds->avg_load);
7397 * In the presence of smp nice balancing, certain scenarios can have
7398 * max load less than avg load(as we skip the groups at or below
7399 * its cpu_capacity, while calculating max_load..)
7401 if (busiest->avg_load <= sds->avg_load ||
7402 local->avg_load >= sds->avg_load) {
7403 /* Misfitting tasks should be migrated in any case */
7404 if (busiest->group_type == group_misfit_task) {
7405 env->imbalance = busiest->group_misfit_task;
7410 * Busiest group is overloaded, local is not, use the spare
7411 * cycles to maximize throughput
7413 if (busiest->group_type == group_overloaded &&
7414 local->group_type <= group_misfit_task) {
7415 env->imbalance = busiest->load_per_task;
7420 return fix_small_imbalance(env, sds);
7424 * If there aren't any idle cpus, avoid creating some.
7426 if (busiest->group_type == group_overloaded &&
7427 local->group_type == group_overloaded) {
7428 load_above_capacity = busiest->sum_nr_running *
7430 if (load_above_capacity > busiest->group_capacity)
7431 load_above_capacity -= busiest->group_capacity;
7433 load_above_capacity = ~0UL;
7437 * We're trying to get all the cpus to the average_load, so we don't
7438 * want to push ourselves above the average load, nor do we wish to
7439 * reduce the max loaded cpu below the average load. At the same time,
7440 * we also don't want to reduce the group load below the group capacity
7441 * (so that we can implement power-savings policies etc). Thus we look
7442 * for the minimum possible imbalance.
7444 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7446 /* How much load to actually move to equalise the imbalance */
7447 env->imbalance = min(
7448 max_pull * busiest->group_capacity,
7449 (sds->avg_load - local->avg_load) * local->group_capacity
7450 ) / SCHED_CAPACITY_SCALE;
7452 /* Boost imbalance to allow misfit task to be balanced. */
7453 if (busiest->group_type == group_misfit_task)
7454 env->imbalance = max_t(long, env->imbalance,
7455 busiest->group_misfit_task);
7458 * if *imbalance is less than the average load per runnable task
7459 * there is no guarantee that any tasks will be moved so we'll have
7460 * a think about bumping its value to force at least one task to be
7463 if (env->imbalance < busiest->load_per_task)
7464 return fix_small_imbalance(env, sds);
7467 /******* find_busiest_group() helpers end here *********************/
7470 * find_busiest_group - Returns the busiest group within the sched_domain
7471 * if there is an imbalance. If there isn't an imbalance, and
7472 * the user has opted for power-savings, it returns a group whose
7473 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7474 * such a group exists.
7476 * Also calculates the amount of weighted load which should be moved
7477 * to restore balance.
7479 * @env: The load balancing environment.
7481 * Return: - The busiest group if imbalance exists.
7482 * - If no imbalance and user has opted for power-savings balance,
7483 * return the least loaded group whose CPUs can be
7484 * put to idle by rebalancing its tasks onto our group.
7486 static struct sched_group *find_busiest_group(struct lb_env *env)
7488 struct sg_lb_stats *local, *busiest;
7489 struct sd_lb_stats sds;
7491 init_sd_lb_stats(&sds);
7494 * Compute the various statistics relavent for load balancing at
7497 update_sd_lb_stats(env, &sds);
7499 if (energy_aware() && !env->dst_rq->rd->overutilized)
7502 local = &sds.local_stat;
7503 busiest = &sds.busiest_stat;
7505 /* ASYM feature bypasses nice load balance check */
7506 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7507 check_asym_packing(env, &sds))
7510 /* There is no busy sibling group to pull tasks from */
7511 if (!sds.busiest || busiest->sum_nr_running == 0)
7514 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7515 / sds.total_capacity;
7518 * If the busiest group is imbalanced the below checks don't
7519 * work because they assume all things are equal, which typically
7520 * isn't true due to cpus_allowed constraints and the like.
7522 if (busiest->group_type == group_imbalanced)
7525 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7526 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7527 busiest->group_no_capacity)
7530 /* Misfitting tasks should be dealt with regardless of the avg load */
7531 if (busiest->group_type == group_misfit_task) {
7536 * If the local group is busier than the selected busiest group
7537 * don't try and pull any tasks.
7539 if (local->avg_load >= busiest->avg_load)
7543 * Don't pull any tasks if this group is already above the domain
7546 if (local->avg_load >= sds.avg_load)
7549 if (env->idle == CPU_IDLE) {
7551 * This cpu is idle. If the busiest group is not overloaded
7552 * and there is no imbalance between this and busiest group
7553 * wrt idle cpus, it is balanced. The imbalance becomes
7554 * significant if the diff is greater than 1 otherwise we
7555 * might end up to just move the imbalance on another group
7557 if ((busiest->group_type != group_overloaded) &&
7558 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7559 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7563 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7564 * imbalance_pct to be conservative.
7566 if (100 * busiest->avg_load <=
7567 env->sd->imbalance_pct * local->avg_load)
7572 env->busiest_group_type = busiest->group_type;
7573 /* Looks like there is an imbalance. Compute it */
7574 calculate_imbalance(env, &sds);
7583 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7585 static struct rq *find_busiest_queue(struct lb_env *env,
7586 struct sched_group *group)
7588 struct rq *busiest = NULL, *rq;
7589 unsigned long busiest_load = 0, busiest_capacity = 1;
7592 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7593 unsigned long capacity, wl;
7597 rt = fbq_classify_rq(rq);
7600 * We classify groups/runqueues into three groups:
7601 * - regular: there are !numa tasks
7602 * - remote: there are numa tasks that run on the 'wrong' node
7603 * - all: there is no distinction
7605 * In order to avoid migrating ideally placed numa tasks,
7606 * ignore those when there's better options.
7608 * If we ignore the actual busiest queue to migrate another
7609 * task, the next balance pass can still reduce the busiest
7610 * queue by moving tasks around inside the node.
7612 * If we cannot move enough load due to this classification
7613 * the next pass will adjust the group classification and
7614 * allow migration of more tasks.
7616 * Both cases only affect the total convergence complexity.
7618 if (rt > env->fbq_type)
7621 capacity = capacity_of(i);
7623 wl = weighted_cpuload(i);
7626 * When comparing with imbalance, use weighted_cpuload()
7627 * which is not scaled with the cpu capacity.
7630 if (rq->nr_running == 1 && wl > env->imbalance &&
7631 !check_cpu_capacity(rq, env->sd) &&
7632 env->busiest_group_type != group_misfit_task)
7636 * For the load comparisons with the other cpu's, consider
7637 * the weighted_cpuload() scaled with the cpu capacity, so
7638 * that the load can be moved away from the cpu that is
7639 * potentially running at a lower capacity.
7641 * Thus we're looking for max(wl_i / capacity_i), crosswise
7642 * multiplication to rid ourselves of the division works out
7643 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7644 * our previous maximum.
7646 if (wl * busiest_capacity > busiest_load * capacity) {
7648 busiest_capacity = capacity;
7657 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7658 * so long as it is large enough.
7660 #define MAX_PINNED_INTERVAL 512
7662 /* Working cpumask for load_balance and load_balance_newidle. */
7663 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7665 static int need_active_balance(struct lb_env *env)
7667 struct sched_domain *sd = env->sd;
7669 if (env->idle == CPU_NEWLY_IDLE) {
7672 * ASYM_PACKING needs to force migrate tasks from busy but
7673 * higher numbered CPUs in order to pack all tasks in the
7674 * lowest numbered CPUs.
7676 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7681 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7682 * It's worth migrating the task if the src_cpu's capacity is reduced
7683 * because of other sched_class or IRQs if more capacity stays
7684 * available on dst_cpu.
7686 if ((env->idle != CPU_NOT_IDLE) &&
7687 (env->src_rq->cfs.h_nr_running == 1)) {
7688 if ((check_cpu_capacity(env->src_rq, sd)) &&
7689 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7693 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
7694 env->src_rq->cfs.h_nr_running == 1 &&
7695 cpu_overutilized(env->src_cpu) &&
7696 !cpu_overutilized(env->dst_cpu)) {
7700 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7703 static int active_load_balance_cpu_stop(void *data);
7705 static int should_we_balance(struct lb_env *env)
7707 struct sched_group *sg = env->sd->groups;
7708 struct cpumask *sg_cpus, *sg_mask;
7709 int cpu, balance_cpu = -1;
7712 * In the newly idle case, we will allow all the cpu's
7713 * to do the newly idle load balance.
7715 if (env->idle == CPU_NEWLY_IDLE)
7718 sg_cpus = sched_group_cpus(sg);
7719 sg_mask = sched_group_mask(sg);
7720 /* Try to find first idle cpu */
7721 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7722 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7729 if (balance_cpu == -1)
7730 balance_cpu = group_balance_cpu(sg);
7733 * First idle cpu or the first cpu(busiest) in this sched group
7734 * is eligible for doing load balancing at this and above domains.
7736 return balance_cpu == env->dst_cpu;
7740 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7741 * tasks if there is an imbalance.
7743 static int load_balance(int this_cpu, struct rq *this_rq,
7744 struct sched_domain *sd, enum cpu_idle_type idle,
7745 int *continue_balancing)
7747 int ld_moved, cur_ld_moved, active_balance = 0;
7748 struct sched_domain *sd_parent = sd->parent;
7749 struct sched_group *group;
7751 unsigned long flags;
7752 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7754 struct lb_env env = {
7756 .dst_cpu = this_cpu,
7758 .dst_grpmask = sched_group_cpus(sd->groups),
7760 .loop_break = sched_nr_migrate_break,
7763 .tasks = LIST_HEAD_INIT(env.tasks),
7767 * For NEWLY_IDLE load_balancing, we don't need to consider
7768 * other cpus in our group
7770 if (idle == CPU_NEWLY_IDLE)
7771 env.dst_grpmask = NULL;
7773 cpumask_copy(cpus, cpu_active_mask);
7775 schedstat_inc(sd, lb_count[idle]);
7778 if (!should_we_balance(&env)) {
7779 *continue_balancing = 0;
7783 group = find_busiest_group(&env);
7785 schedstat_inc(sd, lb_nobusyg[idle]);
7789 busiest = find_busiest_queue(&env, group);
7791 schedstat_inc(sd, lb_nobusyq[idle]);
7795 BUG_ON(busiest == env.dst_rq);
7797 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7799 env.src_cpu = busiest->cpu;
7800 env.src_rq = busiest;
7803 if (busiest->nr_running > 1) {
7805 * Attempt to move tasks. If find_busiest_group has found
7806 * an imbalance but busiest->nr_running <= 1, the group is
7807 * still unbalanced. ld_moved simply stays zero, so it is
7808 * correctly treated as an imbalance.
7810 env.flags |= LBF_ALL_PINNED;
7811 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7814 raw_spin_lock_irqsave(&busiest->lock, flags);
7817 * cur_ld_moved - load moved in current iteration
7818 * ld_moved - cumulative load moved across iterations
7820 cur_ld_moved = detach_tasks(&env);
7822 * We want to potentially lower env.src_cpu's OPP.
7825 update_capacity_of(env.src_cpu);
7828 * We've detached some tasks from busiest_rq. Every
7829 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7830 * unlock busiest->lock, and we are able to be sure
7831 * that nobody can manipulate the tasks in parallel.
7832 * See task_rq_lock() family for the details.
7835 raw_spin_unlock(&busiest->lock);
7839 ld_moved += cur_ld_moved;
7842 local_irq_restore(flags);
7844 if (env.flags & LBF_NEED_BREAK) {
7845 env.flags &= ~LBF_NEED_BREAK;
7850 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7851 * us and move them to an alternate dst_cpu in our sched_group
7852 * where they can run. The upper limit on how many times we
7853 * iterate on same src_cpu is dependent on number of cpus in our
7856 * This changes load balance semantics a bit on who can move
7857 * load to a given_cpu. In addition to the given_cpu itself
7858 * (or a ilb_cpu acting on its behalf where given_cpu is
7859 * nohz-idle), we now have balance_cpu in a position to move
7860 * load to given_cpu. In rare situations, this may cause
7861 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7862 * _independently_ and at _same_ time to move some load to
7863 * given_cpu) causing exceess load to be moved to given_cpu.
7864 * This however should not happen so much in practice and
7865 * moreover subsequent load balance cycles should correct the
7866 * excess load moved.
7868 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7870 /* Prevent to re-select dst_cpu via env's cpus */
7871 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7873 env.dst_rq = cpu_rq(env.new_dst_cpu);
7874 env.dst_cpu = env.new_dst_cpu;
7875 env.flags &= ~LBF_DST_PINNED;
7877 env.loop_break = sched_nr_migrate_break;
7880 * Go back to "more_balance" rather than "redo" since we
7881 * need to continue with same src_cpu.
7887 * We failed to reach balance because of affinity.
7890 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7892 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7893 *group_imbalance = 1;
7896 /* All tasks on this runqueue were pinned by CPU affinity */
7897 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7898 cpumask_clear_cpu(cpu_of(busiest), cpus);
7899 if (!cpumask_empty(cpus)) {
7901 env.loop_break = sched_nr_migrate_break;
7904 goto out_all_pinned;
7909 schedstat_inc(sd, lb_failed[idle]);
7911 * Increment the failure counter only on periodic balance.
7912 * We do not want newidle balance, which can be very
7913 * frequent, pollute the failure counter causing
7914 * excessive cache_hot migrations and active balances.
7916 if (idle != CPU_NEWLY_IDLE)
7917 if (env.src_grp_nr_running > 1)
7918 sd->nr_balance_failed++;
7920 if (need_active_balance(&env)) {
7921 raw_spin_lock_irqsave(&busiest->lock, flags);
7923 /* don't kick the active_load_balance_cpu_stop,
7924 * if the curr task on busiest cpu can't be
7927 if (!cpumask_test_cpu(this_cpu,
7928 tsk_cpus_allowed(busiest->curr))) {
7929 raw_spin_unlock_irqrestore(&busiest->lock,
7931 env.flags |= LBF_ALL_PINNED;
7932 goto out_one_pinned;
7936 * ->active_balance synchronizes accesses to
7937 * ->active_balance_work. Once set, it's cleared
7938 * only after active load balance is finished.
7940 if (!busiest->active_balance) {
7941 busiest->active_balance = 1;
7942 busiest->push_cpu = this_cpu;
7945 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7947 if (active_balance) {
7948 stop_one_cpu_nowait(cpu_of(busiest),
7949 active_load_balance_cpu_stop, busiest,
7950 &busiest->active_balance_work);
7954 * We've kicked active balancing, reset the failure
7957 sd->nr_balance_failed = sd->cache_nice_tries+1;
7960 sd->nr_balance_failed = 0;
7962 if (likely(!active_balance)) {
7963 /* We were unbalanced, so reset the balancing interval */
7964 sd->balance_interval = sd->min_interval;
7967 * If we've begun active balancing, start to back off. This
7968 * case may not be covered by the all_pinned logic if there
7969 * is only 1 task on the busy runqueue (because we don't call
7972 if (sd->balance_interval < sd->max_interval)
7973 sd->balance_interval *= 2;
7980 * We reach balance although we may have faced some affinity
7981 * constraints. Clear the imbalance flag if it was set.
7984 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7986 if (*group_imbalance)
7987 *group_imbalance = 0;
7992 * We reach balance because all tasks are pinned at this level so
7993 * we can't migrate them. Let the imbalance flag set so parent level
7994 * can try to migrate them.
7996 schedstat_inc(sd, lb_balanced[idle]);
7998 sd->nr_balance_failed = 0;
8001 /* tune up the balancing interval */
8002 if (((env.flags & LBF_ALL_PINNED) &&
8003 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8004 (sd->balance_interval < sd->max_interval))
8005 sd->balance_interval *= 2;
8012 static inline unsigned long
8013 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8015 unsigned long interval = sd->balance_interval;
8018 interval *= sd->busy_factor;
8020 /* scale ms to jiffies */
8021 interval = msecs_to_jiffies(interval);
8022 interval = clamp(interval, 1UL, max_load_balance_interval);
8028 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8030 unsigned long interval, next;
8032 interval = get_sd_balance_interval(sd, cpu_busy);
8033 next = sd->last_balance + interval;
8035 if (time_after(*next_balance, next))
8036 *next_balance = next;
8040 * idle_balance is called by schedule() if this_cpu is about to become
8041 * idle. Attempts to pull tasks from other CPUs.
8043 static int idle_balance(struct rq *this_rq)
8045 unsigned long next_balance = jiffies + HZ;
8046 int this_cpu = this_rq->cpu;
8047 struct sched_domain *sd;
8048 int pulled_task = 0;
8051 idle_enter_fair(this_rq);
8054 * We must set idle_stamp _before_ calling idle_balance(), such that we
8055 * measure the duration of idle_balance() as idle time.
8057 this_rq->idle_stamp = rq_clock(this_rq);
8059 if (!energy_aware() &&
8060 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8061 !this_rq->rd->overload)) {
8063 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8065 update_next_balance(sd, 0, &next_balance);
8071 raw_spin_unlock(&this_rq->lock);
8073 update_blocked_averages(this_cpu);
8075 for_each_domain(this_cpu, sd) {
8076 int continue_balancing = 1;
8077 u64 t0, domain_cost;
8079 if (!(sd->flags & SD_LOAD_BALANCE))
8082 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8083 update_next_balance(sd, 0, &next_balance);
8087 if (sd->flags & SD_BALANCE_NEWIDLE) {
8088 t0 = sched_clock_cpu(this_cpu);
8090 pulled_task = load_balance(this_cpu, this_rq,
8092 &continue_balancing);
8094 domain_cost = sched_clock_cpu(this_cpu) - t0;
8095 if (domain_cost > sd->max_newidle_lb_cost)
8096 sd->max_newidle_lb_cost = domain_cost;
8098 curr_cost += domain_cost;
8101 update_next_balance(sd, 0, &next_balance);
8104 * Stop searching for tasks to pull if there are
8105 * now runnable tasks on this rq.
8107 if (pulled_task || this_rq->nr_running > 0)
8112 raw_spin_lock(&this_rq->lock);
8114 if (curr_cost > this_rq->max_idle_balance_cost)
8115 this_rq->max_idle_balance_cost = curr_cost;
8118 * While browsing the domains, we released the rq lock, a task could
8119 * have been enqueued in the meantime. Since we're not going idle,
8120 * pretend we pulled a task.
8122 if (this_rq->cfs.h_nr_running && !pulled_task)
8126 /* Move the next balance forward */
8127 if (time_after(this_rq->next_balance, next_balance))
8128 this_rq->next_balance = next_balance;
8130 /* Is there a task of a high priority class? */
8131 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8135 idle_exit_fair(this_rq);
8136 this_rq->idle_stamp = 0;
8143 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8144 * running tasks off the busiest CPU onto idle CPUs. It requires at
8145 * least 1 task to be running on each physical CPU where possible, and
8146 * avoids physical / logical imbalances.
8148 static int active_load_balance_cpu_stop(void *data)
8150 struct rq *busiest_rq = data;
8151 int busiest_cpu = cpu_of(busiest_rq);
8152 int target_cpu = busiest_rq->push_cpu;
8153 struct rq *target_rq = cpu_rq(target_cpu);
8154 struct sched_domain *sd;
8155 struct task_struct *p = NULL;
8157 raw_spin_lock_irq(&busiest_rq->lock);
8159 /* make sure the requested cpu hasn't gone down in the meantime */
8160 if (unlikely(busiest_cpu != smp_processor_id() ||
8161 !busiest_rq->active_balance))
8164 /* Is there any task to move? */
8165 if (busiest_rq->nr_running <= 1)
8169 * This condition is "impossible", if it occurs
8170 * we need to fix it. Originally reported by
8171 * Bjorn Helgaas on a 128-cpu setup.
8173 BUG_ON(busiest_rq == target_rq);
8175 /* Search for an sd spanning us and the target CPU. */
8177 for_each_domain(target_cpu, sd) {
8178 if ((sd->flags & SD_LOAD_BALANCE) &&
8179 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8184 struct lb_env env = {
8186 .dst_cpu = target_cpu,
8187 .dst_rq = target_rq,
8188 .src_cpu = busiest_rq->cpu,
8189 .src_rq = busiest_rq,
8193 schedstat_inc(sd, alb_count);
8195 p = detach_one_task(&env);
8197 schedstat_inc(sd, alb_pushed);
8199 * We want to potentially lower env.src_cpu's OPP.
8201 update_capacity_of(env.src_cpu);
8204 schedstat_inc(sd, alb_failed);
8208 busiest_rq->active_balance = 0;
8209 raw_spin_unlock(&busiest_rq->lock);
8212 attach_one_task(target_rq, p);
8219 static inline int on_null_domain(struct rq *rq)
8221 return unlikely(!rcu_dereference_sched(rq->sd));
8224 #ifdef CONFIG_NO_HZ_COMMON
8226 * idle load balancing details
8227 * - When one of the busy CPUs notice that there may be an idle rebalancing
8228 * needed, they will kick the idle load balancer, which then does idle
8229 * load balancing for all the idle CPUs.
8232 cpumask_var_t idle_cpus_mask;
8234 unsigned long next_balance; /* in jiffy units */
8235 } nohz ____cacheline_aligned;
8237 static inline int find_new_ilb(void)
8239 int ilb = cpumask_first(nohz.idle_cpus_mask);
8241 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8248 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8249 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8250 * CPU (if there is one).
8252 static void nohz_balancer_kick(void)
8256 nohz.next_balance++;
8258 ilb_cpu = find_new_ilb();
8260 if (ilb_cpu >= nr_cpu_ids)
8263 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8266 * Use smp_send_reschedule() instead of resched_cpu().
8267 * This way we generate a sched IPI on the target cpu which
8268 * is idle. And the softirq performing nohz idle load balance
8269 * will be run before returning from the IPI.
8271 smp_send_reschedule(ilb_cpu);
8275 static inline void nohz_balance_exit_idle(int cpu)
8277 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8279 * Completely isolated CPUs don't ever set, so we must test.
8281 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8282 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8283 atomic_dec(&nohz.nr_cpus);
8285 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8289 static inline void set_cpu_sd_state_busy(void)
8291 struct sched_domain *sd;
8292 int cpu = smp_processor_id();
8295 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8297 if (!sd || !sd->nohz_idle)
8301 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8306 void set_cpu_sd_state_idle(void)
8308 struct sched_domain *sd;
8309 int cpu = smp_processor_id();
8312 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8314 if (!sd || sd->nohz_idle)
8318 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8324 * This routine will record that the cpu is going idle with tick stopped.
8325 * This info will be used in performing idle load balancing in the future.
8327 void nohz_balance_enter_idle(int cpu)
8330 * If this cpu is going down, then nothing needs to be done.
8332 if (!cpu_active(cpu))
8335 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8339 * If we're a completely isolated CPU, we don't play.
8341 if (on_null_domain(cpu_rq(cpu)))
8344 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8345 atomic_inc(&nohz.nr_cpus);
8346 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8349 static int sched_ilb_notifier(struct notifier_block *nfb,
8350 unsigned long action, void *hcpu)
8352 switch (action & ~CPU_TASKS_FROZEN) {
8354 nohz_balance_exit_idle(smp_processor_id());
8362 static DEFINE_SPINLOCK(balancing);
8365 * Scale the max load_balance interval with the number of CPUs in the system.
8366 * This trades load-balance latency on larger machines for less cross talk.
8368 void update_max_interval(void)
8370 max_load_balance_interval = HZ*num_online_cpus()/10;
8374 * It checks each scheduling domain to see if it is due to be balanced,
8375 * and initiates a balancing operation if so.
8377 * Balancing parameters are set up in init_sched_domains.
8379 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8381 int continue_balancing = 1;
8383 unsigned long interval;
8384 struct sched_domain *sd;
8385 /* Earliest time when we have to do rebalance again */
8386 unsigned long next_balance = jiffies + 60*HZ;
8387 int update_next_balance = 0;
8388 int need_serialize, need_decay = 0;
8391 update_blocked_averages(cpu);
8394 for_each_domain(cpu, sd) {
8396 * Decay the newidle max times here because this is a regular
8397 * visit to all the domains. Decay ~1% per second.
8399 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8400 sd->max_newidle_lb_cost =
8401 (sd->max_newidle_lb_cost * 253) / 256;
8402 sd->next_decay_max_lb_cost = jiffies + HZ;
8405 max_cost += sd->max_newidle_lb_cost;
8407 if (!(sd->flags & SD_LOAD_BALANCE))
8411 * Stop the load balance at this level. There is another
8412 * CPU in our sched group which is doing load balancing more
8415 if (!continue_balancing) {
8421 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8423 need_serialize = sd->flags & SD_SERIALIZE;
8424 if (need_serialize) {
8425 if (!spin_trylock(&balancing))
8429 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8430 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8432 * The LBF_DST_PINNED logic could have changed
8433 * env->dst_cpu, so we can't know our idle
8434 * state even if we migrated tasks. Update it.
8436 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8438 sd->last_balance = jiffies;
8439 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8442 spin_unlock(&balancing);
8444 if (time_after(next_balance, sd->last_balance + interval)) {
8445 next_balance = sd->last_balance + interval;
8446 update_next_balance = 1;
8451 * Ensure the rq-wide value also decays but keep it at a
8452 * reasonable floor to avoid funnies with rq->avg_idle.
8454 rq->max_idle_balance_cost =
8455 max((u64)sysctl_sched_migration_cost, max_cost);
8460 * next_balance will be updated only when there is a need.
8461 * When the cpu is attached to null domain for ex, it will not be
8464 if (likely(update_next_balance)) {
8465 rq->next_balance = next_balance;
8467 #ifdef CONFIG_NO_HZ_COMMON
8469 * If this CPU has been elected to perform the nohz idle
8470 * balance. Other idle CPUs have already rebalanced with
8471 * nohz_idle_balance() and nohz.next_balance has been
8472 * updated accordingly. This CPU is now running the idle load
8473 * balance for itself and we need to update the
8474 * nohz.next_balance accordingly.
8476 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8477 nohz.next_balance = rq->next_balance;
8482 #ifdef CONFIG_NO_HZ_COMMON
8484 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8485 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8487 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8489 int this_cpu = this_rq->cpu;
8492 /* Earliest time when we have to do rebalance again */
8493 unsigned long next_balance = jiffies + 60*HZ;
8494 int update_next_balance = 0;
8496 if (idle != CPU_IDLE ||
8497 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8500 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8501 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8505 * If this cpu gets work to do, stop the load balancing
8506 * work being done for other cpus. Next load
8507 * balancing owner will pick it up.
8512 rq = cpu_rq(balance_cpu);
8515 * If time for next balance is due,
8518 if (time_after_eq(jiffies, rq->next_balance)) {
8519 raw_spin_lock_irq(&rq->lock);
8520 update_rq_clock(rq);
8521 update_idle_cpu_load(rq);
8522 raw_spin_unlock_irq(&rq->lock);
8523 rebalance_domains(rq, CPU_IDLE);
8526 if (time_after(next_balance, rq->next_balance)) {
8527 next_balance = rq->next_balance;
8528 update_next_balance = 1;
8533 * next_balance will be updated only when there is a need.
8534 * When the CPU is attached to null domain for ex, it will not be
8537 if (likely(update_next_balance))
8538 nohz.next_balance = next_balance;
8540 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8544 * Current heuristic for kicking the idle load balancer in the presence
8545 * of an idle cpu in the system.
8546 * - This rq has more than one task.
8547 * - This rq has at least one CFS task and the capacity of the CPU is
8548 * significantly reduced because of RT tasks or IRQs.
8549 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8550 * multiple busy cpu.
8551 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8552 * domain span are idle.
8554 static inline bool nohz_kick_needed(struct rq *rq)
8556 unsigned long now = jiffies;
8557 struct sched_domain *sd;
8558 struct sched_group_capacity *sgc;
8559 int nr_busy, cpu = rq->cpu;
8562 if (unlikely(rq->idle_balance))
8566 * We may be recently in ticked or tickless idle mode. At the first
8567 * busy tick after returning from idle, we will update the busy stats.
8569 set_cpu_sd_state_busy();
8570 nohz_balance_exit_idle(cpu);
8573 * None are in tickless mode and hence no need for NOHZ idle load
8576 if (likely(!atomic_read(&nohz.nr_cpus)))
8579 if (time_before(now, nohz.next_balance))
8582 if (rq->nr_running >= 2 &&
8583 (!energy_aware() || cpu_overutilized(cpu)))
8587 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8588 if (sd && !energy_aware()) {
8589 sgc = sd->groups->sgc;
8590 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8599 sd = rcu_dereference(rq->sd);
8601 if ((rq->cfs.h_nr_running >= 1) &&
8602 check_cpu_capacity(rq, sd)) {
8608 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8609 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8610 sched_domain_span(sd)) < cpu)) {
8620 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8624 * run_rebalance_domains is triggered when needed from the scheduler tick.
8625 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8627 static void run_rebalance_domains(struct softirq_action *h)
8629 struct rq *this_rq = this_rq();
8630 enum cpu_idle_type idle = this_rq->idle_balance ?
8631 CPU_IDLE : CPU_NOT_IDLE;
8634 * If this cpu has a pending nohz_balance_kick, then do the
8635 * balancing on behalf of the other idle cpus whose ticks are
8636 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8637 * give the idle cpus a chance to load balance. Else we may
8638 * load balance only within the local sched_domain hierarchy
8639 * and abort nohz_idle_balance altogether if we pull some load.
8641 nohz_idle_balance(this_rq, idle);
8642 rebalance_domains(this_rq, idle);
8646 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8648 void trigger_load_balance(struct rq *rq)
8650 /* Don't need to rebalance while attached to NULL domain */
8651 if (unlikely(on_null_domain(rq)))
8654 if (time_after_eq(jiffies, rq->next_balance))
8655 raise_softirq(SCHED_SOFTIRQ);
8656 #ifdef CONFIG_NO_HZ_COMMON
8657 if (nohz_kick_needed(rq))
8658 nohz_balancer_kick();
8662 static void rq_online_fair(struct rq *rq)
8666 update_runtime_enabled(rq);
8669 static void rq_offline_fair(struct rq *rq)
8673 /* Ensure any throttled groups are reachable by pick_next_task */
8674 unthrottle_offline_cfs_rqs(rq);
8677 #endif /* CONFIG_SMP */
8680 * scheduler tick hitting a task of our scheduling class:
8682 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8684 struct cfs_rq *cfs_rq;
8685 struct sched_entity *se = &curr->se;
8687 for_each_sched_entity(se) {
8688 cfs_rq = cfs_rq_of(se);
8689 entity_tick(cfs_rq, se, queued);
8692 if (static_branch_unlikely(&sched_numa_balancing))
8693 task_tick_numa(rq, curr);
8695 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr)))
8696 rq->rd->overutilized = true;
8698 rq->misfit_task = !task_fits_max(curr, rq->cpu);
8702 * called on fork with the child task as argument from the parent's context
8703 * - child not yet on the tasklist
8704 * - preemption disabled
8706 static void task_fork_fair(struct task_struct *p)
8708 struct cfs_rq *cfs_rq;
8709 struct sched_entity *se = &p->se, *curr;
8710 int this_cpu = smp_processor_id();
8711 struct rq *rq = this_rq();
8712 unsigned long flags;
8714 raw_spin_lock_irqsave(&rq->lock, flags);
8716 update_rq_clock(rq);
8718 cfs_rq = task_cfs_rq(current);
8719 curr = cfs_rq->curr;
8722 * Not only the cpu but also the task_group of the parent might have
8723 * been changed after parent->se.parent,cfs_rq were copied to
8724 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8725 * of child point to valid ones.
8728 __set_task_cpu(p, this_cpu);
8731 update_curr(cfs_rq);
8734 se->vruntime = curr->vruntime;
8735 place_entity(cfs_rq, se, 1);
8737 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8739 * Upon rescheduling, sched_class::put_prev_task() will place
8740 * 'current' within the tree based on its new key value.
8742 swap(curr->vruntime, se->vruntime);
8746 se->vruntime -= cfs_rq->min_vruntime;
8748 raw_spin_unlock_irqrestore(&rq->lock, flags);
8752 * Priority of the task has changed. Check to see if we preempt
8756 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8758 if (!task_on_rq_queued(p))
8762 * Reschedule if we are currently running on this runqueue and
8763 * our priority decreased, or if we are not currently running on
8764 * this runqueue and our priority is higher than the current's
8766 if (rq->curr == p) {
8767 if (p->prio > oldprio)
8770 check_preempt_curr(rq, p, 0);
8773 static inline bool vruntime_normalized(struct task_struct *p)
8775 struct sched_entity *se = &p->se;
8778 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8779 * the dequeue_entity(.flags=0) will already have normalized the
8786 * When !on_rq, vruntime of the task has usually NOT been normalized.
8787 * But there are some cases where it has already been normalized:
8789 * - A forked child which is waiting for being woken up by
8790 * wake_up_new_task().
8791 * - A task which has been woken up by try_to_wake_up() and
8792 * waiting for actually being woken up by sched_ttwu_pending().
8794 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8800 static void detach_task_cfs_rq(struct task_struct *p)
8802 struct sched_entity *se = &p->se;
8803 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8805 if (!vruntime_normalized(p)) {
8807 * Fix up our vruntime so that the current sleep doesn't
8808 * cause 'unlimited' sleep bonus.
8810 place_entity(cfs_rq, se, 0);
8811 se->vruntime -= cfs_rq->min_vruntime;
8814 /* Catch up with the cfs_rq and remove our load when we leave */
8815 detach_entity_load_avg(cfs_rq, se);
8818 static void attach_task_cfs_rq(struct task_struct *p)
8820 struct sched_entity *se = &p->se;
8821 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8823 #ifdef CONFIG_FAIR_GROUP_SCHED
8825 * Since the real-depth could have been changed (only FAIR
8826 * class maintain depth value), reset depth properly.
8828 se->depth = se->parent ? se->parent->depth + 1 : 0;
8831 /* Synchronize task with its cfs_rq */
8832 attach_entity_load_avg(cfs_rq, se);
8834 if (!vruntime_normalized(p))
8835 se->vruntime += cfs_rq->min_vruntime;
8838 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8840 detach_task_cfs_rq(p);
8843 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8845 attach_task_cfs_rq(p);
8847 if (task_on_rq_queued(p)) {
8849 * We were most likely switched from sched_rt, so
8850 * kick off the schedule if running, otherwise just see
8851 * if we can still preempt the current task.
8856 check_preempt_curr(rq, p, 0);
8860 /* Account for a task changing its policy or group.
8862 * This routine is mostly called to set cfs_rq->curr field when a task
8863 * migrates between groups/classes.
8865 static void set_curr_task_fair(struct rq *rq)
8867 struct sched_entity *se = &rq->curr->se;
8869 for_each_sched_entity(se) {
8870 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8872 set_next_entity(cfs_rq, se);
8873 /* ensure bandwidth has been allocated on our new cfs_rq */
8874 account_cfs_rq_runtime(cfs_rq, 0);
8878 void init_cfs_rq(struct cfs_rq *cfs_rq)
8880 cfs_rq->tasks_timeline = RB_ROOT;
8881 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8882 #ifndef CONFIG_64BIT
8883 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8886 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8887 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8891 #ifdef CONFIG_FAIR_GROUP_SCHED
8892 static void task_move_group_fair(struct task_struct *p)
8894 detach_task_cfs_rq(p);
8895 set_task_rq(p, task_cpu(p));
8898 /* Tell se's cfs_rq has been changed -- migrated */
8899 p->se.avg.last_update_time = 0;
8901 attach_task_cfs_rq(p);
8904 void free_fair_sched_group(struct task_group *tg)
8908 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8910 for_each_possible_cpu(i) {
8912 kfree(tg->cfs_rq[i]);
8915 remove_entity_load_avg(tg->se[i]);
8924 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8926 struct cfs_rq *cfs_rq;
8927 struct sched_entity *se;
8930 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8933 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8937 tg->shares = NICE_0_LOAD;
8939 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8941 for_each_possible_cpu(i) {
8942 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8943 GFP_KERNEL, cpu_to_node(i));
8947 se = kzalloc_node(sizeof(struct sched_entity),
8948 GFP_KERNEL, cpu_to_node(i));
8952 init_cfs_rq(cfs_rq);
8953 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8954 init_entity_runnable_average(se);
8965 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8967 struct rq *rq = cpu_rq(cpu);
8968 unsigned long flags;
8971 * Only empty task groups can be destroyed; so we can speculatively
8972 * check on_list without danger of it being re-added.
8974 if (!tg->cfs_rq[cpu]->on_list)
8977 raw_spin_lock_irqsave(&rq->lock, flags);
8978 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8979 raw_spin_unlock_irqrestore(&rq->lock, flags);
8982 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8983 struct sched_entity *se, int cpu,
8984 struct sched_entity *parent)
8986 struct rq *rq = cpu_rq(cpu);
8990 init_cfs_rq_runtime(cfs_rq);
8992 tg->cfs_rq[cpu] = cfs_rq;
8995 /* se could be NULL for root_task_group */
9000 se->cfs_rq = &rq->cfs;
9003 se->cfs_rq = parent->my_q;
9004 se->depth = parent->depth + 1;
9008 /* guarantee group entities always have weight */
9009 update_load_set(&se->load, NICE_0_LOAD);
9010 se->parent = parent;
9013 static DEFINE_MUTEX(shares_mutex);
9015 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9018 unsigned long flags;
9021 * We can't change the weight of the root cgroup.
9026 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9028 mutex_lock(&shares_mutex);
9029 if (tg->shares == shares)
9032 tg->shares = shares;
9033 for_each_possible_cpu(i) {
9034 struct rq *rq = cpu_rq(i);
9035 struct sched_entity *se;
9038 /* Propagate contribution to hierarchy */
9039 raw_spin_lock_irqsave(&rq->lock, flags);
9041 /* Possible calls to update_curr() need rq clock */
9042 update_rq_clock(rq);
9043 for_each_sched_entity(se)
9044 update_cfs_shares(group_cfs_rq(se));
9045 raw_spin_unlock_irqrestore(&rq->lock, flags);
9049 mutex_unlock(&shares_mutex);
9052 #else /* CONFIG_FAIR_GROUP_SCHED */
9054 void free_fair_sched_group(struct task_group *tg) { }
9056 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9061 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9063 #endif /* CONFIG_FAIR_GROUP_SCHED */
9066 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9068 struct sched_entity *se = &task->se;
9069 unsigned int rr_interval = 0;
9072 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9075 if (rq->cfs.load.weight)
9076 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9082 * All the scheduling class methods:
9084 const struct sched_class fair_sched_class = {
9085 .next = &idle_sched_class,
9086 .enqueue_task = enqueue_task_fair,
9087 .dequeue_task = dequeue_task_fair,
9088 .yield_task = yield_task_fair,
9089 .yield_to_task = yield_to_task_fair,
9091 .check_preempt_curr = check_preempt_wakeup,
9093 .pick_next_task = pick_next_task_fair,
9094 .put_prev_task = put_prev_task_fair,
9097 .select_task_rq = select_task_rq_fair,
9098 .migrate_task_rq = migrate_task_rq_fair,
9100 .rq_online = rq_online_fair,
9101 .rq_offline = rq_offline_fair,
9103 .task_waking = task_waking_fair,
9104 .task_dead = task_dead_fair,
9105 .set_cpus_allowed = set_cpus_allowed_common,
9108 .set_curr_task = set_curr_task_fair,
9109 .task_tick = task_tick_fair,
9110 .task_fork = task_fork_fair,
9112 .prio_changed = prio_changed_fair,
9113 .switched_from = switched_from_fair,
9114 .switched_to = switched_to_fair,
9116 .get_rr_interval = get_rr_interval_fair,
9118 .update_curr = update_curr_fair,
9120 #ifdef CONFIG_FAIR_GROUP_SCHED
9121 .task_move_group = task_move_group_fair,
9125 #ifdef CONFIG_SCHED_DEBUG
9126 void print_cfs_stats(struct seq_file *m, int cpu)
9128 struct cfs_rq *cfs_rq;
9131 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9132 print_cfs_rq(m, cpu, cfs_rq);
9136 #ifdef CONFIG_NUMA_BALANCING
9137 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9140 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9142 for_each_online_node(node) {
9143 if (p->numa_faults) {
9144 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9145 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9147 if (p->numa_group) {
9148 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9149 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9151 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9154 #endif /* CONFIG_NUMA_BALANCING */
9155 #endif /* CONFIG_SCHED_DEBUG */
9157 __init void init_sched_fair_class(void)
9160 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9162 #ifdef CONFIG_NO_HZ_COMMON
9163 nohz.next_balance = jiffies;
9164 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9165 cpu_notifier(sched_ilb_notifier, 0);