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);
2688 * Unsigned subtract and clamp on underflow.
2690 * Explicitly do a load-store to ensure the intermediate value never hits
2691 * memory. This allows lockless observations without ever seeing the negative
2694 #define sub_positive(_ptr, _val) do { \
2695 typeof(_ptr) ptr = (_ptr); \
2696 typeof(*ptr) val = (_val); \
2697 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2701 WRITE_ONCE(*ptr, res); \
2704 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2705 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2707 struct sched_avg *sa = &cfs_rq->avg;
2708 int decayed, removed = 0;
2710 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2711 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2712 sub_positive(&sa->load_avg, r);
2713 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2717 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2718 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2719 sub_positive(&sa->util_avg, r);
2720 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2723 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2724 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2726 #ifndef CONFIG_64BIT
2728 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2731 return decayed || removed;
2734 /* Update task and its cfs_rq load average */
2735 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2737 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2738 u64 now = cfs_rq_clock_task(cfs_rq);
2739 int cpu = cpu_of(rq_of(cfs_rq));
2742 * Track task load average for carrying it to new CPU after migrated, and
2743 * track group sched_entity load average for task_h_load calc in migration
2745 __update_load_avg(now, cpu, &se->avg,
2746 se->on_rq * scale_load_down(se->load.weight),
2747 cfs_rq->curr == se, NULL);
2749 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2750 update_tg_load_avg(cfs_rq, 0);
2752 if (entity_is_task(se))
2753 trace_sched_load_avg_task(task_of(se), &se->avg);
2754 trace_sched_load_avg_cpu(cpu, cfs_rq);
2757 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2759 if (!sched_feat(ATTACH_AGE_LOAD))
2763 * If we got migrated (either between CPUs or between cgroups) we'll
2764 * have aged the average right before clearing @last_update_time.
2766 if (se->avg.last_update_time) {
2767 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2768 &se->avg, 0, 0, NULL);
2771 * XXX: we could have just aged the entire load away if we've been
2772 * absent from the fair class for too long.
2777 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2778 cfs_rq->avg.load_avg += se->avg.load_avg;
2779 cfs_rq->avg.load_sum += se->avg.load_sum;
2780 cfs_rq->avg.util_avg += se->avg.util_avg;
2781 cfs_rq->avg.util_sum += se->avg.util_sum;
2784 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2786 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2787 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2788 cfs_rq->curr == se, NULL);
2790 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2791 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2792 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2793 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2796 /* Add the load generated by se into cfs_rq's load average */
2798 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2800 struct sched_avg *sa = &se->avg;
2801 u64 now = cfs_rq_clock_task(cfs_rq);
2802 int migrated, decayed;
2804 migrated = !sa->last_update_time;
2806 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2807 se->on_rq * scale_load_down(se->load.weight),
2808 cfs_rq->curr == se, NULL);
2811 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2813 cfs_rq->runnable_load_avg += sa->load_avg;
2814 cfs_rq->runnable_load_sum += sa->load_sum;
2817 attach_entity_load_avg(cfs_rq, se);
2819 if (decayed || migrated)
2820 update_tg_load_avg(cfs_rq, 0);
2823 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2825 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2827 update_load_avg(se, 1);
2829 cfs_rq->runnable_load_avg =
2830 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2831 cfs_rq->runnable_load_sum =
2832 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2835 #ifndef CONFIG_64BIT
2836 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2838 u64 last_update_time_copy;
2839 u64 last_update_time;
2842 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2844 last_update_time = cfs_rq->avg.last_update_time;
2845 } while (last_update_time != last_update_time_copy);
2847 return last_update_time;
2850 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2852 return cfs_rq->avg.last_update_time;
2857 * Task first catches up with cfs_rq, and then subtract
2858 * itself from the cfs_rq (task must be off the queue now).
2860 void remove_entity_load_avg(struct sched_entity *se)
2862 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2863 u64 last_update_time;
2866 * Newly created task or never used group entity should not be removed
2867 * from its (source) cfs_rq
2869 if (se->avg.last_update_time == 0)
2872 last_update_time = cfs_rq_last_update_time(cfs_rq);
2874 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2875 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2876 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2880 * Update the rq's load with the elapsed running time before entering
2881 * idle. if the last scheduled task is not a CFS task, idle_enter will
2882 * be the only way to update the runnable statistic.
2884 void idle_enter_fair(struct rq *this_rq)
2889 * Update the rq's load with the elapsed idle time before a task is
2890 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2891 * be the only way to update the runnable statistic.
2893 void idle_exit_fair(struct rq *this_rq)
2897 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2899 return cfs_rq->runnable_load_avg;
2902 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2904 return cfs_rq->avg.load_avg;
2907 static int idle_balance(struct rq *this_rq);
2909 #else /* CONFIG_SMP */
2911 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2913 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2915 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2916 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2919 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2921 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2923 static inline int idle_balance(struct rq *rq)
2928 #endif /* CONFIG_SMP */
2930 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2932 #ifdef CONFIG_SCHEDSTATS
2933 struct task_struct *tsk = NULL;
2935 if (entity_is_task(se))
2938 if (se->statistics.sleep_start) {
2939 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2944 if (unlikely(delta > se->statistics.sleep_max))
2945 se->statistics.sleep_max = delta;
2947 se->statistics.sleep_start = 0;
2948 se->statistics.sum_sleep_runtime += delta;
2951 account_scheduler_latency(tsk, delta >> 10, 1);
2952 trace_sched_stat_sleep(tsk, delta);
2955 if (se->statistics.block_start) {
2956 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2961 if (unlikely(delta > se->statistics.block_max))
2962 se->statistics.block_max = delta;
2964 se->statistics.block_start = 0;
2965 se->statistics.sum_sleep_runtime += delta;
2968 if (tsk->in_iowait) {
2969 se->statistics.iowait_sum += delta;
2970 se->statistics.iowait_count++;
2971 trace_sched_stat_iowait(tsk, delta);
2974 trace_sched_stat_blocked(tsk, delta);
2975 trace_sched_blocked_reason(tsk);
2978 * Blocking time is in units of nanosecs, so shift by
2979 * 20 to get a milliseconds-range estimation of the
2980 * amount of time that the task spent sleeping:
2982 if (unlikely(prof_on == SLEEP_PROFILING)) {
2983 profile_hits(SLEEP_PROFILING,
2984 (void *)get_wchan(tsk),
2987 account_scheduler_latency(tsk, delta >> 10, 0);
2993 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2995 #ifdef CONFIG_SCHED_DEBUG
2996 s64 d = se->vruntime - cfs_rq->min_vruntime;
3001 if (d > 3*sysctl_sched_latency)
3002 schedstat_inc(cfs_rq, nr_spread_over);
3007 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3009 u64 vruntime = cfs_rq->min_vruntime;
3012 * The 'current' period is already promised to the current tasks,
3013 * however the extra weight of the new task will slow them down a
3014 * little, place the new task so that it fits in the slot that
3015 * stays open at the end.
3017 if (initial && sched_feat(START_DEBIT))
3018 vruntime += sched_vslice(cfs_rq, se);
3020 /* sleeps up to a single latency don't count. */
3022 unsigned long thresh = sysctl_sched_latency;
3025 * Halve their sleep time's effect, to allow
3026 * for a gentler effect of sleepers:
3028 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3034 /* ensure we never gain time by being placed backwards. */
3035 se->vruntime = max_vruntime(se->vruntime, vruntime);
3038 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3041 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3044 * Update the normalized vruntime before updating min_vruntime
3045 * through calling update_curr().
3047 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3048 se->vruntime += cfs_rq->min_vruntime;
3051 * Update run-time statistics of the 'current'.
3053 update_curr(cfs_rq);
3054 enqueue_entity_load_avg(cfs_rq, se);
3055 account_entity_enqueue(cfs_rq, se);
3056 update_cfs_shares(cfs_rq);
3058 if (flags & ENQUEUE_WAKEUP) {
3059 place_entity(cfs_rq, se, 0);
3060 enqueue_sleeper(cfs_rq, se);
3063 update_stats_enqueue(cfs_rq, se);
3064 check_spread(cfs_rq, se);
3065 if (se != cfs_rq->curr)
3066 __enqueue_entity(cfs_rq, se);
3069 if (cfs_rq->nr_running == 1) {
3070 list_add_leaf_cfs_rq(cfs_rq);
3071 check_enqueue_throttle(cfs_rq);
3075 static void __clear_buddies_last(struct sched_entity *se)
3077 for_each_sched_entity(se) {
3078 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3079 if (cfs_rq->last != se)
3082 cfs_rq->last = NULL;
3086 static void __clear_buddies_next(struct sched_entity *se)
3088 for_each_sched_entity(se) {
3089 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3090 if (cfs_rq->next != se)
3093 cfs_rq->next = NULL;
3097 static void __clear_buddies_skip(struct sched_entity *se)
3099 for_each_sched_entity(se) {
3100 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3101 if (cfs_rq->skip != se)
3104 cfs_rq->skip = NULL;
3108 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3110 if (cfs_rq->last == se)
3111 __clear_buddies_last(se);
3113 if (cfs_rq->next == se)
3114 __clear_buddies_next(se);
3116 if (cfs_rq->skip == se)
3117 __clear_buddies_skip(se);
3120 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3123 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3126 * Update run-time statistics of the 'current'.
3128 update_curr(cfs_rq);
3129 dequeue_entity_load_avg(cfs_rq, se);
3131 update_stats_dequeue(cfs_rq, se);
3132 if (flags & DEQUEUE_SLEEP) {
3133 #ifdef CONFIG_SCHEDSTATS
3134 if (entity_is_task(se)) {
3135 struct task_struct *tsk = task_of(se);
3137 if (tsk->state & TASK_INTERRUPTIBLE)
3138 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3139 if (tsk->state & TASK_UNINTERRUPTIBLE)
3140 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3145 clear_buddies(cfs_rq, se);
3147 if (se != cfs_rq->curr)
3148 __dequeue_entity(cfs_rq, se);
3150 account_entity_dequeue(cfs_rq, se);
3153 * Normalize the entity after updating the min_vruntime because the
3154 * update can refer to the ->curr item and we need to reflect this
3155 * movement in our normalized position.
3157 if (!(flags & DEQUEUE_SLEEP))
3158 se->vruntime -= cfs_rq->min_vruntime;
3160 /* return excess runtime on last dequeue */
3161 return_cfs_rq_runtime(cfs_rq);
3163 update_min_vruntime(cfs_rq);
3164 update_cfs_shares(cfs_rq);
3168 * Preempt the current task with a newly woken task if needed:
3171 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3173 unsigned long ideal_runtime, delta_exec;
3174 struct sched_entity *se;
3177 ideal_runtime = sched_slice(cfs_rq, curr);
3178 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3179 if (delta_exec > ideal_runtime) {
3180 resched_curr(rq_of(cfs_rq));
3182 * The current task ran long enough, ensure it doesn't get
3183 * re-elected due to buddy favours.
3185 clear_buddies(cfs_rq, curr);
3190 * Ensure that a task that missed wakeup preemption by a
3191 * narrow margin doesn't have to wait for a full slice.
3192 * This also mitigates buddy induced latencies under load.
3194 if (delta_exec < sysctl_sched_min_granularity)
3197 se = __pick_first_entity(cfs_rq);
3198 delta = curr->vruntime - se->vruntime;
3203 if (delta > ideal_runtime)
3204 resched_curr(rq_of(cfs_rq));
3208 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3210 /* 'current' is not kept within the tree. */
3213 * Any task has to be enqueued before it get to execute on
3214 * a CPU. So account for the time it spent waiting on the
3217 update_stats_wait_end(cfs_rq, se);
3218 __dequeue_entity(cfs_rq, se);
3219 update_load_avg(se, 1);
3222 update_stats_curr_start(cfs_rq, se);
3224 #ifdef CONFIG_SCHEDSTATS
3226 * Track our maximum slice length, if the CPU's load is at
3227 * least twice that of our own weight (i.e. dont track it
3228 * when there are only lesser-weight tasks around):
3230 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3231 se->statistics.slice_max = max(se->statistics.slice_max,
3232 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3235 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3239 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3242 * Pick the next process, keeping these things in mind, in this order:
3243 * 1) keep things fair between processes/task groups
3244 * 2) pick the "next" process, since someone really wants that to run
3245 * 3) pick the "last" process, for cache locality
3246 * 4) do not run the "skip" process, if something else is available
3248 static struct sched_entity *
3249 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3251 struct sched_entity *left = __pick_first_entity(cfs_rq);
3252 struct sched_entity *se;
3255 * If curr is set we have to see if its left of the leftmost entity
3256 * still in the tree, provided there was anything in the tree at all.
3258 if (!left || (curr && entity_before(curr, left)))
3261 se = left; /* ideally we run the leftmost entity */
3264 * Avoid running the skip buddy, if running something else can
3265 * be done without getting too unfair.
3267 if (cfs_rq->skip == se) {
3268 struct sched_entity *second;
3271 second = __pick_first_entity(cfs_rq);
3273 second = __pick_next_entity(se);
3274 if (!second || (curr && entity_before(curr, second)))
3278 if (second && wakeup_preempt_entity(second, left) < 1)
3283 * Prefer last buddy, try to return the CPU to a preempted task.
3285 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3289 * Someone really wants this to run. If it's not unfair, run it.
3291 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3294 clear_buddies(cfs_rq, se);
3299 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3301 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3304 * If still on the runqueue then deactivate_task()
3305 * was not called and update_curr() has to be done:
3308 update_curr(cfs_rq);
3310 /* throttle cfs_rqs exceeding runtime */
3311 check_cfs_rq_runtime(cfs_rq);
3313 check_spread(cfs_rq, prev);
3315 update_stats_wait_start(cfs_rq, prev);
3316 /* Put 'current' back into the tree. */
3317 __enqueue_entity(cfs_rq, prev);
3318 /* in !on_rq case, update occurred at dequeue */
3319 update_load_avg(prev, 0);
3321 cfs_rq->curr = NULL;
3325 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3328 * Update run-time statistics of the 'current'.
3330 update_curr(cfs_rq);
3333 * Ensure that runnable average is periodically updated.
3335 update_load_avg(curr, 1);
3336 update_cfs_shares(cfs_rq);
3338 #ifdef CONFIG_SCHED_HRTICK
3340 * queued ticks are scheduled to match the slice, so don't bother
3341 * validating it and just reschedule.
3344 resched_curr(rq_of(cfs_rq));
3348 * don't let the period tick interfere with the hrtick preemption
3350 if (!sched_feat(DOUBLE_TICK) &&
3351 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3355 if (cfs_rq->nr_running > 1)
3356 check_preempt_tick(cfs_rq, curr);
3360 /**************************************************
3361 * CFS bandwidth control machinery
3364 #ifdef CONFIG_CFS_BANDWIDTH
3366 #ifdef HAVE_JUMP_LABEL
3367 static struct static_key __cfs_bandwidth_used;
3369 static inline bool cfs_bandwidth_used(void)
3371 return static_key_false(&__cfs_bandwidth_used);
3374 void cfs_bandwidth_usage_inc(void)
3376 static_key_slow_inc(&__cfs_bandwidth_used);
3379 void cfs_bandwidth_usage_dec(void)
3381 static_key_slow_dec(&__cfs_bandwidth_used);
3383 #else /* HAVE_JUMP_LABEL */
3384 static bool cfs_bandwidth_used(void)
3389 void cfs_bandwidth_usage_inc(void) {}
3390 void cfs_bandwidth_usage_dec(void) {}
3391 #endif /* HAVE_JUMP_LABEL */
3394 * default period for cfs group bandwidth.
3395 * default: 0.1s, units: nanoseconds
3397 static inline u64 default_cfs_period(void)
3399 return 100000000ULL;
3402 static inline u64 sched_cfs_bandwidth_slice(void)
3404 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3408 * Replenish runtime according to assigned quota and update expiration time.
3409 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3410 * additional synchronization around rq->lock.
3412 * requires cfs_b->lock
3414 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3418 if (cfs_b->quota == RUNTIME_INF)
3421 now = sched_clock_cpu(smp_processor_id());
3422 cfs_b->runtime = cfs_b->quota;
3423 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3426 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3428 return &tg->cfs_bandwidth;
3431 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3432 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3434 if (unlikely(cfs_rq->throttle_count))
3435 return cfs_rq->throttled_clock_task;
3437 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3440 /* returns 0 on failure to allocate runtime */
3441 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3443 struct task_group *tg = cfs_rq->tg;
3444 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3445 u64 amount = 0, min_amount, expires;
3447 /* note: this is a positive sum as runtime_remaining <= 0 */
3448 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3450 raw_spin_lock(&cfs_b->lock);
3451 if (cfs_b->quota == RUNTIME_INF)
3452 amount = min_amount;
3454 start_cfs_bandwidth(cfs_b);
3456 if (cfs_b->runtime > 0) {
3457 amount = min(cfs_b->runtime, min_amount);
3458 cfs_b->runtime -= amount;
3462 expires = cfs_b->runtime_expires;
3463 raw_spin_unlock(&cfs_b->lock);
3465 cfs_rq->runtime_remaining += amount;
3467 * we may have advanced our local expiration to account for allowed
3468 * spread between our sched_clock and the one on which runtime was
3471 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3472 cfs_rq->runtime_expires = expires;
3474 return cfs_rq->runtime_remaining > 0;
3478 * Note: This depends on the synchronization provided by sched_clock and the
3479 * fact that rq->clock snapshots this value.
3481 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3483 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3485 /* if the deadline is ahead of our clock, nothing to do */
3486 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3489 if (cfs_rq->runtime_remaining < 0)
3493 * If the local deadline has passed we have to consider the
3494 * possibility that our sched_clock is 'fast' and the global deadline
3495 * has not truly expired.
3497 * Fortunately we can check determine whether this the case by checking
3498 * whether the global deadline has advanced. It is valid to compare
3499 * cfs_b->runtime_expires without any locks since we only care about
3500 * exact equality, so a partial write will still work.
3503 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3504 /* extend local deadline, drift is bounded above by 2 ticks */
3505 cfs_rq->runtime_expires += TICK_NSEC;
3507 /* global deadline is ahead, expiration has passed */
3508 cfs_rq->runtime_remaining = 0;
3512 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3514 /* dock delta_exec before expiring quota (as it could span periods) */
3515 cfs_rq->runtime_remaining -= delta_exec;
3516 expire_cfs_rq_runtime(cfs_rq);
3518 if (likely(cfs_rq->runtime_remaining > 0))
3522 * if we're unable to extend our runtime we resched so that the active
3523 * hierarchy can be throttled
3525 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3526 resched_curr(rq_of(cfs_rq));
3529 static __always_inline
3530 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3532 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3535 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3538 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3540 return cfs_bandwidth_used() && cfs_rq->throttled;
3543 /* check whether cfs_rq, or any parent, is throttled */
3544 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3546 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3550 * Ensure that neither of the group entities corresponding to src_cpu or
3551 * dest_cpu are members of a throttled hierarchy when performing group
3552 * load-balance operations.
3554 static inline int throttled_lb_pair(struct task_group *tg,
3555 int src_cpu, int dest_cpu)
3557 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3559 src_cfs_rq = tg->cfs_rq[src_cpu];
3560 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3562 return throttled_hierarchy(src_cfs_rq) ||
3563 throttled_hierarchy(dest_cfs_rq);
3566 /* updated child weight may affect parent so we have to do this bottom up */
3567 static int tg_unthrottle_up(struct task_group *tg, void *data)
3569 struct rq *rq = data;
3570 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3572 cfs_rq->throttle_count--;
3574 if (!cfs_rq->throttle_count) {
3575 /* adjust cfs_rq_clock_task() */
3576 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3577 cfs_rq->throttled_clock_task;
3584 static int tg_throttle_down(struct task_group *tg, void *data)
3586 struct rq *rq = data;
3587 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3589 /* group is entering throttled state, stop time */
3590 if (!cfs_rq->throttle_count)
3591 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3592 cfs_rq->throttle_count++;
3597 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3599 struct rq *rq = rq_of(cfs_rq);
3600 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3601 struct sched_entity *se;
3602 long task_delta, dequeue = 1;
3605 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3607 /* freeze hierarchy runnable averages while throttled */
3609 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3612 task_delta = cfs_rq->h_nr_running;
3613 for_each_sched_entity(se) {
3614 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3615 /* throttled entity or throttle-on-deactivate */
3620 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3621 qcfs_rq->h_nr_running -= task_delta;
3623 if (qcfs_rq->load.weight)
3628 sub_nr_running(rq, task_delta);
3630 cfs_rq->throttled = 1;
3631 cfs_rq->throttled_clock = rq_clock(rq);
3632 raw_spin_lock(&cfs_b->lock);
3633 empty = list_empty(&cfs_b->throttled_cfs_rq);
3636 * Add to the _head_ of the list, so that an already-started
3637 * distribute_cfs_runtime will not see us
3639 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3642 * If we're the first throttled task, make sure the bandwidth
3646 start_cfs_bandwidth(cfs_b);
3648 raw_spin_unlock(&cfs_b->lock);
3651 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3653 struct rq *rq = rq_of(cfs_rq);
3654 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3655 struct sched_entity *se;
3659 se = cfs_rq->tg->se[cpu_of(rq)];
3661 cfs_rq->throttled = 0;
3663 update_rq_clock(rq);
3665 raw_spin_lock(&cfs_b->lock);
3666 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3667 list_del_rcu(&cfs_rq->throttled_list);
3668 raw_spin_unlock(&cfs_b->lock);
3670 /* update hierarchical throttle state */
3671 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3673 if (!cfs_rq->load.weight)
3676 task_delta = cfs_rq->h_nr_running;
3677 for_each_sched_entity(se) {
3681 cfs_rq = cfs_rq_of(se);
3683 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3684 cfs_rq->h_nr_running += task_delta;
3686 if (cfs_rq_throttled(cfs_rq))
3691 add_nr_running(rq, task_delta);
3693 /* determine whether we need to wake up potentially idle cpu */
3694 if (rq->curr == rq->idle && rq->cfs.nr_running)
3698 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3699 u64 remaining, u64 expires)
3701 struct cfs_rq *cfs_rq;
3703 u64 starting_runtime = remaining;
3706 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3708 struct rq *rq = rq_of(cfs_rq);
3710 raw_spin_lock(&rq->lock);
3711 if (!cfs_rq_throttled(cfs_rq))
3714 runtime = -cfs_rq->runtime_remaining + 1;
3715 if (runtime > remaining)
3716 runtime = remaining;
3717 remaining -= runtime;
3719 cfs_rq->runtime_remaining += runtime;
3720 cfs_rq->runtime_expires = expires;
3722 /* we check whether we're throttled above */
3723 if (cfs_rq->runtime_remaining > 0)
3724 unthrottle_cfs_rq(cfs_rq);
3727 raw_spin_unlock(&rq->lock);
3734 return starting_runtime - remaining;
3738 * Responsible for refilling a task_group's bandwidth and unthrottling its
3739 * cfs_rqs as appropriate. If there has been no activity within the last
3740 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3741 * used to track this state.
3743 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3745 u64 runtime, runtime_expires;
3748 /* no need to continue the timer with no bandwidth constraint */
3749 if (cfs_b->quota == RUNTIME_INF)
3750 goto out_deactivate;
3752 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3753 cfs_b->nr_periods += overrun;
3756 * idle depends on !throttled (for the case of a large deficit), and if
3757 * we're going inactive then everything else can be deferred
3759 if (cfs_b->idle && !throttled)
3760 goto out_deactivate;
3762 __refill_cfs_bandwidth_runtime(cfs_b);
3765 /* mark as potentially idle for the upcoming period */
3770 /* account preceding periods in which throttling occurred */
3771 cfs_b->nr_throttled += overrun;
3773 runtime_expires = cfs_b->runtime_expires;
3776 * This check is repeated as we are holding onto the new bandwidth while
3777 * we unthrottle. This can potentially race with an unthrottled group
3778 * trying to acquire new bandwidth from the global pool. This can result
3779 * in us over-using our runtime if it is all used during this loop, but
3780 * only by limited amounts in that extreme case.
3782 while (throttled && cfs_b->runtime > 0) {
3783 runtime = cfs_b->runtime;
3784 raw_spin_unlock(&cfs_b->lock);
3785 /* we can't nest cfs_b->lock while distributing bandwidth */
3786 runtime = distribute_cfs_runtime(cfs_b, runtime,
3788 raw_spin_lock(&cfs_b->lock);
3790 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3792 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3796 * While we are ensured activity in the period following an
3797 * unthrottle, this also covers the case in which the new bandwidth is
3798 * insufficient to cover the existing bandwidth deficit. (Forcing the
3799 * timer to remain active while there are any throttled entities.)
3809 /* a cfs_rq won't donate quota below this amount */
3810 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3811 /* minimum remaining period time to redistribute slack quota */
3812 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3813 /* how long we wait to gather additional slack before distributing */
3814 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3817 * Are we near the end of the current quota period?
3819 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3820 * hrtimer base being cleared by hrtimer_start. In the case of
3821 * migrate_hrtimers, base is never cleared, so we are fine.
3823 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3825 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3828 /* if the call-back is running a quota refresh is already occurring */
3829 if (hrtimer_callback_running(refresh_timer))
3832 /* is a quota refresh about to occur? */
3833 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3834 if (remaining < min_expire)
3840 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3842 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3844 /* if there's a quota refresh soon don't bother with slack */
3845 if (runtime_refresh_within(cfs_b, min_left))
3848 hrtimer_start(&cfs_b->slack_timer,
3849 ns_to_ktime(cfs_bandwidth_slack_period),
3853 /* we know any runtime found here is valid as update_curr() precedes return */
3854 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3856 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3857 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3859 if (slack_runtime <= 0)
3862 raw_spin_lock(&cfs_b->lock);
3863 if (cfs_b->quota != RUNTIME_INF &&
3864 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3865 cfs_b->runtime += slack_runtime;
3867 /* we are under rq->lock, defer unthrottling using a timer */
3868 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3869 !list_empty(&cfs_b->throttled_cfs_rq))
3870 start_cfs_slack_bandwidth(cfs_b);
3872 raw_spin_unlock(&cfs_b->lock);
3874 /* even if it's not valid for return we don't want to try again */
3875 cfs_rq->runtime_remaining -= slack_runtime;
3878 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3880 if (!cfs_bandwidth_used())
3883 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3886 __return_cfs_rq_runtime(cfs_rq);
3890 * This is done with a timer (instead of inline with bandwidth return) since
3891 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3893 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3895 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3898 /* confirm we're still not at a refresh boundary */
3899 raw_spin_lock(&cfs_b->lock);
3900 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3901 raw_spin_unlock(&cfs_b->lock);
3905 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3906 runtime = cfs_b->runtime;
3908 expires = cfs_b->runtime_expires;
3909 raw_spin_unlock(&cfs_b->lock);
3914 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3916 raw_spin_lock(&cfs_b->lock);
3917 if (expires == cfs_b->runtime_expires)
3918 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3919 raw_spin_unlock(&cfs_b->lock);
3923 * When a group wakes up we want to make sure that its quota is not already
3924 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3925 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3927 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3929 if (!cfs_bandwidth_used())
3932 /* an active group must be handled by the update_curr()->put() path */
3933 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3936 /* ensure the group is not already throttled */
3937 if (cfs_rq_throttled(cfs_rq))
3940 /* update runtime allocation */
3941 account_cfs_rq_runtime(cfs_rq, 0);
3942 if (cfs_rq->runtime_remaining <= 0)
3943 throttle_cfs_rq(cfs_rq);
3946 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3947 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3949 if (!cfs_bandwidth_used())
3952 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3956 * it's possible for a throttled entity to be forced into a running
3957 * state (e.g. set_curr_task), in this case we're finished.
3959 if (cfs_rq_throttled(cfs_rq))
3962 throttle_cfs_rq(cfs_rq);
3966 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3968 struct cfs_bandwidth *cfs_b =
3969 container_of(timer, struct cfs_bandwidth, slack_timer);
3971 do_sched_cfs_slack_timer(cfs_b);
3973 return HRTIMER_NORESTART;
3976 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3978 struct cfs_bandwidth *cfs_b =
3979 container_of(timer, struct cfs_bandwidth, period_timer);
3983 raw_spin_lock(&cfs_b->lock);
3985 overrun = hrtimer_forward_now(timer, cfs_b->period);
3989 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3992 cfs_b->period_active = 0;
3993 raw_spin_unlock(&cfs_b->lock);
3995 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3998 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4000 raw_spin_lock_init(&cfs_b->lock);
4002 cfs_b->quota = RUNTIME_INF;
4003 cfs_b->period = ns_to_ktime(default_cfs_period());
4005 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4006 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4007 cfs_b->period_timer.function = sched_cfs_period_timer;
4008 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4009 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4012 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4014 cfs_rq->runtime_enabled = 0;
4015 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4018 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4020 lockdep_assert_held(&cfs_b->lock);
4022 if (!cfs_b->period_active) {
4023 cfs_b->period_active = 1;
4024 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4025 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4029 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4031 /* init_cfs_bandwidth() was not called */
4032 if (!cfs_b->throttled_cfs_rq.next)
4035 hrtimer_cancel(&cfs_b->period_timer);
4036 hrtimer_cancel(&cfs_b->slack_timer);
4039 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4041 struct cfs_rq *cfs_rq;
4043 for_each_leaf_cfs_rq(rq, cfs_rq) {
4044 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4046 raw_spin_lock(&cfs_b->lock);
4047 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4048 raw_spin_unlock(&cfs_b->lock);
4052 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4054 struct cfs_rq *cfs_rq;
4056 for_each_leaf_cfs_rq(rq, cfs_rq) {
4057 if (!cfs_rq->runtime_enabled)
4061 * clock_task is not advancing so we just need to make sure
4062 * there's some valid quota amount
4064 cfs_rq->runtime_remaining = 1;
4066 * Offline rq is schedulable till cpu is completely disabled
4067 * in take_cpu_down(), so we prevent new cfs throttling here.
4069 cfs_rq->runtime_enabled = 0;
4071 if (cfs_rq_throttled(cfs_rq))
4072 unthrottle_cfs_rq(cfs_rq);
4076 #else /* CONFIG_CFS_BANDWIDTH */
4077 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4079 return rq_clock_task(rq_of(cfs_rq));
4082 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4083 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4084 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4085 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4087 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4092 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4097 static inline int throttled_lb_pair(struct task_group *tg,
4098 int src_cpu, int dest_cpu)
4103 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4105 #ifdef CONFIG_FAIR_GROUP_SCHED
4106 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4109 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4113 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4114 static inline void update_runtime_enabled(struct rq *rq) {}
4115 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4117 #endif /* CONFIG_CFS_BANDWIDTH */
4119 /**************************************************
4120 * CFS operations on tasks:
4123 #ifdef CONFIG_SCHED_HRTICK
4124 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4126 struct sched_entity *se = &p->se;
4127 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4129 WARN_ON(task_rq(p) != rq);
4131 if (cfs_rq->nr_running > 1) {
4132 u64 slice = sched_slice(cfs_rq, se);
4133 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4134 s64 delta = slice - ran;
4141 hrtick_start(rq, delta);
4146 * called from enqueue/dequeue and updates the hrtick when the
4147 * current task is from our class and nr_running is low enough
4150 static void hrtick_update(struct rq *rq)
4152 struct task_struct *curr = rq->curr;
4154 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4157 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4158 hrtick_start_fair(rq, curr);
4160 #else /* !CONFIG_SCHED_HRTICK */
4162 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4166 static inline void hrtick_update(struct rq *rq)
4171 static inline unsigned long boosted_cpu_util(int cpu);
4173 static void update_capacity_of(int cpu)
4175 unsigned long req_cap;
4180 /* Convert scale-invariant capacity to cpu. */
4181 req_cap = boosted_cpu_util(cpu);
4182 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4183 set_cfs_cpu_capacity(cpu, true, req_cap);
4186 static bool cpu_overutilized(int cpu);
4189 * The enqueue_task method is called before nr_running is
4190 * increased. Here we update the fair scheduling stats and
4191 * then put the task into the rbtree:
4194 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4196 struct cfs_rq *cfs_rq;
4197 struct sched_entity *se = &p->se;
4198 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4199 int task_wakeup = flags & ENQUEUE_WAKEUP;
4201 for_each_sched_entity(se) {
4204 cfs_rq = cfs_rq_of(se);
4205 enqueue_entity(cfs_rq, se, flags);
4208 * end evaluation on encountering a throttled cfs_rq
4210 * note: in the case of encountering a throttled cfs_rq we will
4211 * post the final h_nr_running increment below.
4213 if (cfs_rq_throttled(cfs_rq))
4215 cfs_rq->h_nr_running++;
4217 flags = ENQUEUE_WAKEUP;
4220 for_each_sched_entity(se) {
4221 cfs_rq = cfs_rq_of(se);
4222 cfs_rq->h_nr_running++;
4224 if (cfs_rq_throttled(cfs_rq))
4227 update_load_avg(se, 1);
4228 update_cfs_shares(cfs_rq);
4232 add_nr_running(rq, 1);
4233 if (!task_new && !rq->rd->overutilized &&
4234 cpu_overutilized(rq->cpu))
4235 rq->rd->overutilized = true;
4237 schedtune_enqueue_task(p, cpu_of(rq));
4240 * We want to potentially trigger a freq switch
4241 * request only for tasks that are waking up; this is
4242 * because we get here also during load balancing, but
4243 * in these cases it seems wise to trigger as single
4244 * request after load balancing is done.
4246 if (task_new || task_wakeup)
4247 update_capacity_of(cpu_of(rq));
4252 static void set_next_buddy(struct sched_entity *se);
4255 * The dequeue_task method is called before nr_running is
4256 * decreased. We remove the task from the rbtree and
4257 * update the fair scheduling stats:
4259 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4261 struct cfs_rq *cfs_rq;
4262 struct sched_entity *se = &p->se;
4263 int task_sleep = flags & DEQUEUE_SLEEP;
4265 for_each_sched_entity(se) {
4266 cfs_rq = cfs_rq_of(se);
4267 dequeue_entity(cfs_rq, se, flags);
4270 * end evaluation on encountering a throttled cfs_rq
4272 * note: in the case of encountering a throttled cfs_rq we will
4273 * post the final h_nr_running decrement below.
4275 if (cfs_rq_throttled(cfs_rq))
4277 cfs_rq->h_nr_running--;
4279 /* Don't dequeue parent if it has other entities besides us */
4280 if (cfs_rq->load.weight) {
4282 * Bias pick_next to pick a task from this cfs_rq, as
4283 * p is sleeping when it is within its sched_slice.
4285 if (task_sleep && parent_entity(se))
4286 set_next_buddy(parent_entity(se));
4288 /* avoid re-evaluating load for this entity */
4289 se = parent_entity(se);
4292 flags |= DEQUEUE_SLEEP;
4295 for_each_sched_entity(se) {
4296 cfs_rq = cfs_rq_of(se);
4297 cfs_rq->h_nr_running--;
4299 if (cfs_rq_throttled(cfs_rq))
4302 update_load_avg(se, 1);
4303 update_cfs_shares(cfs_rq);
4307 sub_nr_running(rq, 1);
4308 schedtune_dequeue_task(p, cpu_of(rq));
4311 * We want to potentially trigger a freq switch
4312 * request only for tasks that are going to sleep;
4313 * this is because we get here also during load
4314 * balancing, but in these cases it seems wise to
4315 * trigger as single request after load balancing is
4319 if (rq->cfs.nr_running)
4320 update_capacity_of(cpu_of(rq));
4321 else if (sched_freq())
4322 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4331 * per rq 'load' arrray crap; XXX kill this.
4335 * The exact cpuload at various idx values, calculated at every tick would be
4336 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4338 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4339 * on nth tick when cpu may be busy, then we have:
4340 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4341 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4343 * decay_load_missed() below does efficient calculation of
4344 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4345 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4347 * The calculation is approximated on a 128 point scale.
4348 * degrade_zero_ticks is the number of ticks after which load at any
4349 * particular idx is approximated to be zero.
4350 * degrade_factor is a precomputed table, a row for each load idx.
4351 * Each column corresponds to degradation factor for a power of two ticks,
4352 * based on 128 point scale.
4354 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4355 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4357 * With this power of 2 load factors, we can degrade the load n times
4358 * by looking at 1 bits in n and doing as many mult/shift instead of
4359 * n mult/shifts needed by the exact degradation.
4361 #define DEGRADE_SHIFT 7
4362 static const unsigned char
4363 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4364 static const unsigned char
4365 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4366 {0, 0, 0, 0, 0, 0, 0, 0},
4367 {64, 32, 8, 0, 0, 0, 0, 0},
4368 {96, 72, 40, 12, 1, 0, 0},
4369 {112, 98, 75, 43, 15, 1, 0},
4370 {120, 112, 98, 76, 45, 16, 2} };
4373 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4374 * would be when CPU is idle and so we just decay the old load without
4375 * adding any new load.
4377 static unsigned long
4378 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4382 if (!missed_updates)
4385 if (missed_updates >= degrade_zero_ticks[idx])
4389 return load >> missed_updates;
4391 while (missed_updates) {
4392 if (missed_updates % 2)
4393 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4395 missed_updates >>= 1;
4402 * Update rq->cpu_load[] statistics. This function is usually called every
4403 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4404 * every tick. We fix it up based on jiffies.
4406 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4407 unsigned long pending_updates)
4411 this_rq->nr_load_updates++;
4413 /* Update our load: */
4414 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4415 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4416 unsigned long old_load, new_load;
4418 /* scale is effectively 1 << i now, and >> i divides by scale */
4420 old_load = this_rq->cpu_load[i];
4421 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4422 new_load = this_load;
4424 * Round up the averaging division if load is increasing. This
4425 * prevents us from getting stuck on 9 if the load is 10, for
4428 if (new_load > old_load)
4429 new_load += scale - 1;
4431 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4434 sched_avg_update(this_rq);
4437 /* Used instead of source_load when we know the type == 0 */
4438 static unsigned long weighted_cpuload(const int cpu)
4440 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4443 #ifdef CONFIG_NO_HZ_COMMON
4445 * There is no sane way to deal with nohz on smp when using jiffies because the
4446 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4447 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4449 * Therefore we cannot use the delta approach from the regular tick since that
4450 * would seriously skew the load calculation. However we'll make do for those
4451 * updates happening while idle (nohz_idle_balance) or coming out of idle
4452 * (tick_nohz_idle_exit).
4454 * This means we might still be one tick off for nohz periods.
4458 * Called from nohz_idle_balance() to update the load ratings before doing the
4461 static void update_idle_cpu_load(struct rq *this_rq)
4463 unsigned long curr_jiffies = READ_ONCE(jiffies);
4464 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4465 unsigned long pending_updates;
4468 * bail if there's load or we're actually up-to-date.
4470 if (load || curr_jiffies == this_rq->last_load_update_tick)
4473 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4474 this_rq->last_load_update_tick = curr_jiffies;
4476 __update_cpu_load(this_rq, load, pending_updates);
4480 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4482 void update_cpu_load_nohz(void)
4484 struct rq *this_rq = this_rq();
4485 unsigned long curr_jiffies = READ_ONCE(jiffies);
4486 unsigned long pending_updates;
4488 if (curr_jiffies == this_rq->last_load_update_tick)
4491 raw_spin_lock(&this_rq->lock);
4492 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4493 if (pending_updates) {
4494 this_rq->last_load_update_tick = curr_jiffies;
4496 * We were idle, this means load 0, the current load might be
4497 * !0 due to remote wakeups and the sort.
4499 __update_cpu_load(this_rq, 0, pending_updates);
4501 raw_spin_unlock(&this_rq->lock);
4503 #endif /* CONFIG_NO_HZ */
4506 * Called from scheduler_tick()
4508 void update_cpu_load_active(struct rq *this_rq)
4510 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4512 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4514 this_rq->last_load_update_tick = jiffies;
4515 __update_cpu_load(this_rq, load, 1);
4519 * Return a low guess at the load of a migration-source cpu weighted
4520 * according to the scheduling class and "nice" value.
4522 * We want to under-estimate the load of migration sources, to
4523 * balance conservatively.
4525 static unsigned long source_load(int cpu, int type)
4527 struct rq *rq = cpu_rq(cpu);
4528 unsigned long total = weighted_cpuload(cpu);
4530 if (type == 0 || !sched_feat(LB_BIAS))
4533 return min(rq->cpu_load[type-1], total);
4537 * Return a high guess at the load of a migration-target cpu weighted
4538 * according to the scheduling class and "nice" value.
4540 static unsigned long target_load(int cpu, int type)
4542 struct rq *rq = cpu_rq(cpu);
4543 unsigned long total = weighted_cpuload(cpu);
4545 if (type == 0 || !sched_feat(LB_BIAS))
4548 return max(rq->cpu_load[type-1], total);
4552 static unsigned long cpu_avg_load_per_task(int cpu)
4554 struct rq *rq = cpu_rq(cpu);
4555 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4556 unsigned long load_avg = weighted_cpuload(cpu);
4559 return load_avg / nr_running;
4564 static void record_wakee(struct task_struct *p)
4567 * Rough decay (wiping) for cost saving, don't worry
4568 * about the boundary, really active task won't care
4571 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4572 current->wakee_flips >>= 1;
4573 current->wakee_flip_decay_ts = jiffies;
4576 if (current->last_wakee != p) {
4577 current->last_wakee = p;
4578 current->wakee_flips++;
4582 static void task_waking_fair(struct task_struct *p)
4584 struct sched_entity *se = &p->se;
4585 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4588 #ifndef CONFIG_64BIT
4589 u64 min_vruntime_copy;
4592 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4594 min_vruntime = cfs_rq->min_vruntime;
4595 } while (min_vruntime != min_vruntime_copy);
4597 min_vruntime = cfs_rq->min_vruntime;
4600 se->vruntime -= min_vruntime;
4604 #ifdef CONFIG_FAIR_GROUP_SCHED
4606 * effective_load() calculates the load change as seen from the root_task_group
4608 * Adding load to a group doesn't make a group heavier, but can cause movement
4609 * of group shares between cpus. Assuming the shares were perfectly aligned one
4610 * can calculate the shift in shares.
4612 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4613 * on this @cpu and results in a total addition (subtraction) of @wg to the
4614 * total group weight.
4616 * Given a runqueue weight distribution (rw_i) we can compute a shares
4617 * distribution (s_i) using:
4619 * s_i = rw_i / \Sum rw_j (1)
4621 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4622 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4623 * shares distribution (s_i):
4625 * rw_i = { 2, 4, 1, 0 }
4626 * s_i = { 2/7, 4/7, 1/7, 0 }
4628 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4629 * task used to run on and the CPU the waker is running on), we need to
4630 * compute the effect of waking a task on either CPU and, in case of a sync
4631 * wakeup, compute the effect of the current task going to sleep.
4633 * So for a change of @wl to the local @cpu with an overall group weight change
4634 * of @wl we can compute the new shares distribution (s'_i) using:
4636 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4638 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4639 * differences in waking a task to CPU 0. The additional task changes the
4640 * weight and shares distributions like:
4642 * rw'_i = { 3, 4, 1, 0 }
4643 * s'_i = { 3/8, 4/8, 1/8, 0 }
4645 * We can then compute the difference in effective weight by using:
4647 * dw_i = S * (s'_i - s_i) (3)
4649 * Where 'S' is the group weight as seen by its parent.
4651 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4652 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4653 * 4/7) times the weight of the group.
4655 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4657 struct sched_entity *se = tg->se[cpu];
4659 if (!tg->parent) /* the trivial, non-cgroup case */
4662 for_each_sched_entity(se) {
4668 * W = @wg + \Sum rw_j
4670 W = wg + calc_tg_weight(tg, se->my_q);
4675 w = cfs_rq_load_avg(se->my_q) + wl;
4678 * wl = S * s'_i; see (2)
4681 wl = (w * (long)tg->shares) / W;
4686 * Per the above, wl is the new se->load.weight value; since
4687 * those are clipped to [MIN_SHARES, ...) do so now. See
4688 * calc_cfs_shares().
4690 if (wl < MIN_SHARES)
4694 * wl = dw_i = S * (s'_i - s_i); see (3)
4696 wl -= se->avg.load_avg;
4699 * Recursively apply this logic to all parent groups to compute
4700 * the final effective load change on the root group. Since
4701 * only the @tg group gets extra weight, all parent groups can
4702 * only redistribute existing shares. @wl is the shift in shares
4703 * resulting from this level per the above.
4712 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4720 * Returns the current capacity of cpu after applying both
4721 * cpu and freq scaling.
4723 unsigned long capacity_curr_of(int cpu)
4725 return cpu_rq(cpu)->cpu_capacity_orig *
4726 arch_scale_freq_capacity(NULL, cpu)
4727 >> SCHED_CAPACITY_SHIFT;
4730 static inline bool energy_aware(void)
4732 return sched_feat(ENERGY_AWARE);
4736 struct sched_group *sg_top;
4737 struct sched_group *sg_cap;
4744 struct task_struct *task;
4759 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4760 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4761 * energy calculations. Using the scale-invariant util returned by
4762 * cpu_util() and approximating scale-invariant util by:
4764 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4766 * the normalized util can be found using the specific capacity.
4768 * capacity = capacity_orig * curr_freq/max_freq
4770 * norm_util = running_time/time ~ util/capacity
4772 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4774 int util = __cpu_util(cpu, delta);
4776 if (util >= capacity)
4777 return SCHED_CAPACITY_SCALE;
4779 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4782 static int calc_util_delta(struct energy_env *eenv, int cpu)
4784 if (cpu == eenv->src_cpu)
4785 return -eenv->util_delta;
4786 if (cpu == eenv->dst_cpu)
4787 return eenv->util_delta;
4792 unsigned long group_max_util(struct energy_env *eenv)
4795 unsigned long max_util = 0;
4797 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4798 delta = calc_util_delta(eenv, i);
4799 max_util = max(max_util, __cpu_util(i, delta));
4806 * group_norm_util() returns the approximated group util relative to it's
4807 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4808 * energy calculations. Since task executions may or may not overlap in time in
4809 * the group the true normalized util is between max(cpu_norm_util(i)) and
4810 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4811 * latter is used as the estimate as it leads to a more pessimistic energy
4812 * estimate (more busy).
4815 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4818 unsigned long util_sum = 0;
4819 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4821 for_each_cpu(i, sched_group_cpus(sg)) {
4822 delta = calc_util_delta(eenv, i);
4823 util_sum += __cpu_norm_util(i, capacity, delta);
4826 if (util_sum > SCHED_CAPACITY_SCALE)
4827 return SCHED_CAPACITY_SCALE;
4831 static int find_new_capacity(struct energy_env *eenv,
4832 const struct sched_group_energy const *sge)
4835 unsigned long util = group_max_util(eenv);
4837 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4838 if (sge->cap_states[idx].cap >= util)
4842 eenv->cap_idx = idx;
4847 static int group_idle_state(struct sched_group *sg)
4849 int i, state = INT_MAX;
4851 /* Find the shallowest idle state in the sched group. */
4852 for_each_cpu(i, sched_group_cpus(sg))
4853 state = min(state, idle_get_state_idx(cpu_rq(i)));
4855 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4862 * sched_group_energy(): Computes the absolute energy consumption of cpus
4863 * belonging to the sched_group including shared resources shared only by
4864 * members of the group. Iterates over all cpus in the hierarchy below the
4865 * sched_group starting from the bottom working it's way up before going to
4866 * the next cpu until all cpus are covered at all levels. The current
4867 * implementation is likely to gather the same util statistics multiple times.
4868 * This can probably be done in a faster but more complex way.
4869 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4871 static int sched_group_energy(struct energy_env *eenv)
4873 struct sched_domain *sd;
4874 int cpu, total_energy = 0;
4875 struct cpumask visit_cpus;
4876 struct sched_group *sg;
4878 WARN_ON(!eenv->sg_top->sge);
4880 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4882 while (!cpumask_empty(&visit_cpus)) {
4883 struct sched_group *sg_shared_cap = NULL;
4885 cpu = cpumask_first(&visit_cpus);
4888 * Is the group utilization affected by cpus outside this
4891 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4895 * We most probably raced with hotplug; returning a
4896 * wrong energy estimation is better than entering an
4902 sg_shared_cap = sd->parent->groups;
4904 for_each_domain(cpu, sd) {
4907 /* Has this sched_domain already been visited? */
4908 if (sd->child && group_first_cpu(sg) != cpu)
4912 unsigned long group_util;
4913 int sg_busy_energy, sg_idle_energy;
4914 int cap_idx, idle_idx;
4916 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
4917 eenv->sg_cap = sg_shared_cap;
4921 cap_idx = find_new_capacity(eenv, sg->sge);
4923 if (sg->group_weight == 1) {
4924 /* Remove capacity of src CPU (before task move) */
4925 if (eenv->util_delta == 0 &&
4926 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
4927 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
4928 eenv->cap.delta -= eenv->cap.before;
4930 /* Add capacity of dst CPU (after task move) */
4931 if (eenv->util_delta != 0 &&
4932 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
4933 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
4934 eenv->cap.delta += eenv->cap.after;
4938 idle_idx = group_idle_state(sg);
4939 group_util = group_norm_util(eenv, sg);
4940 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
4941 >> SCHED_CAPACITY_SHIFT;
4942 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
4943 * sg->sge->idle_states[idle_idx].power)
4944 >> SCHED_CAPACITY_SHIFT;
4946 total_energy += sg_busy_energy + sg_idle_energy;
4949 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
4951 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
4954 } while (sg = sg->next, sg != sd->groups);
4960 eenv->energy = total_energy;
4964 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
4966 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
4969 #ifdef CONFIG_SCHED_TUNE
4970 static int energy_diff_evaluate(struct energy_env *eenv)
4975 /* Return energy diff when boost margin is 0 */
4976 #ifdef CONFIG_CGROUP_SCHEDTUNE
4977 boost = schedtune_task_boost(eenv->task);
4979 boost = get_sysctl_sched_cfs_boost();
4982 return eenv->nrg.diff;
4984 /* Compute normalized energy diff */
4985 nrg_delta = schedtune_normalize_energy(eenv->nrg.diff);
4986 eenv->nrg.delta = nrg_delta;
4988 eenv->payoff = schedtune_accept_deltas(
4994 * When SchedTune is enabled, the energy_diff() function will return
4995 * the computed energy payoff value. Since the energy_diff() return
4996 * value is expected to be negative by its callers, this evaluation
4997 * function return a negative value each time the evaluation return a
4998 * positive payoff, which is the condition for the acceptance of
4999 * a scheduling decision
5001 return -eenv->payoff;
5003 #else /* CONFIG_SCHED_TUNE */
5004 #define energy_diff_evaluate(eenv) eenv->nrg.diff
5008 * energy_diff(): Estimate the energy impact of changing the utilization
5009 * distribution. eenv specifies the change: utilisation amount, source, and
5010 * destination cpu. Source or destination cpu may be -1 in which case the
5011 * utilization is removed from or added to the system (e.g. task wake-up). If
5012 * both are specified, the utilization is migrated.
5014 static int energy_diff(struct energy_env *eenv)
5016 struct sched_domain *sd;
5017 struct sched_group *sg;
5018 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5020 struct energy_env eenv_before = {
5022 .src_cpu = eenv->src_cpu,
5023 .dst_cpu = eenv->dst_cpu,
5024 .nrg = { 0, 0, 0, 0},
5028 if (eenv->src_cpu == eenv->dst_cpu)
5031 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5032 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5035 return 0; /* Error */
5040 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5041 eenv_before.sg_top = eenv->sg_top = sg;
5043 if (sched_group_energy(&eenv_before))
5044 return 0; /* Invalid result abort */
5045 energy_before += eenv_before.energy;
5047 /* Keep track of SRC cpu (before) capacity */
5048 eenv->cap.before = eenv_before.cap.before;
5049 eenv->cap.delta = eenv_before.cap.delta;
5051 if (sched_group_energy(eenv))
5052 return 0; /* Invalid result abort */
5053 energy_after += eenv->energy;
5055 } while (sg = sg->next, sg != sd->groups);
5057 eenv->nrg.before = energy_before;
5058 eenv->nrg.after = energy_after;
5059 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5062 return energy_diff_evaluate(eenv);
5066 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5067 * A waker of many should wake a different task than the one last awakened
5068 * at a frequency roughly N times higher than one of its wakees. In order
5069 * to determine whether we should let the load spread vs consolodating to
5070 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5071 * partner, and a factor of lls_size higher frequency in the other. With
5072 * both conditions met, we can be relatively sure that the relationship is
5073 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5074 * being client/server, worker/dispatcher, interrupt source or whatever is
5075 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5077 static int wake_wide(struct task_struct *p)
5079 unsigned int master = current->wakee_flips;
5080 unsigned int slave = p->wakee_flips;
5081 int factor = this_cpu_read(sd_llc_size);
5084 swap(master, slave);
5085 if (slave < factor || master < slave * factor)
5090 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5092 s64 this_load, load;
5093 s64 this_eff_load, prev_eff_load;
5094 int idx, this_cpu, prev_cpu;
5095 struct task_group *tg;
5096 unsigned long weight;
5100 this_cpu = smp_processor_id();
5101 prev_cpu = task_cpu(p);
5102 load = source_load(prev_cpu, idx);
5103 this_load = target_load(this_cpu, idx);
5106 * If sync wakeup then subtract the (maximum possible)
5107 * effect of the currently running task from the load
5108 * of the current CPU:
5111 tg = task_group(current);
5112 weight = current->se.avg.load_avg;
5114 this_load += effective_load(tg, this_cpu, -weight, -weight);
5115 load += effective_load(tg, prev_cpu, 0, -weight);
5119 weight = p->se.avg.load_avg;
5122 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5123 * due to the sync cause above having dropped this_load to 0, we'll
5124 * always have an imbalance, but there's really nothing you can do
5125 * about that, so that's good too.
5127 * Otherwise check if either cpus are near enough in load to allow this
5128 * task to be woken on this_cpu.
5130 this_eff_load = 100;
5131 this_eff_load *= capacity_of(prev_cpu);
5133 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5134 prev_eff_load *= capacity_of(this_cpu);
5136 if (this_load > 0) {
5137 this_eff_load *= this_load +
5138 effective_load(tg, this_cpu, weight, weight);
5140 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5143 balanced = this_eff_load <= prev_eff_load;
5145 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5150 schedstat_inc(sd, ttwu_move_affine);
5151 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5156 static inline unsigned long task_util(struct task_struct *p)
5158 return p->se.avg.util_avg;
5161 unsigned int capacity_margin = 1280; /* ~20% margin */
5163 static inline unsigned long boosted_task_util(struct task_struct *task);
5165 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5167 unsigned long capacity = capacity_of(cpu);
5169 util += boosted_task_util(p);
5171 return (capacity * 1024) > (util * capacity_margin);
5174 static inline bool task_fits_max(struct task_struct *p, int cpu)
5176 unsigned long capacity = capacity_of(cpu);
5177 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5179 if (capacity == max_capacity)
5182 if (capacity * capacity_margin > max_capacity * 1024)
5185 return __task_fits(p, cpu, 0);
5188 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5190 return __task_fits(p, cpu, cpu_util(cpu));
5193 static bool cpu_overutilized(int cpu)
5195 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5198 #ifdef CONFIG_SCHED_TUNE
5200 static unsigned long
5201 schedtune_margin(unsigned long signal, unsigned long boost)
5203 unsigned long long margin = 0;
5206 * Signal proportional compensation (SPC)
5208 * The Boost (B) value is used to compute a Margin (M) which is
5209 * proportional to the complement of the original Signal (S):
5210 * M = B * (SCHED_LOAD_SCALE - S)
5211 * The obtained M could be used by the caller to "boost" S.
5213 margin = SCHED_LOAD_SCALE - signal;
5217 * Fast integer division by constant:
5218 * Constant : (C) = 100
5219 * Precision : 0.1% (P) = 0.1
5220 * Reference : C * 100 / P (R) = 100000
5223 * Shift bits : ceil(log(R,2)) (S) = 17
5224 * Mult const : round(2^S/C) (M) = 1311
5234 static inline unsigned int
5235 schedtune_cpu_margin(unsigned long util, int cpu)
5239 #ifdef CONFIG_CGROUP_SCHEDTUNE
5240 boost = schedtune_cpu_boost(cpu);
5242 boost = get_sysctl_sched_cfs_boost();
5247 return schedtune_margin(util, boost);
5250 static inline unsigned long
5251 schedtune_task_margin(struct task_struct *task)
5255 unsigned long margin;
5257 #ifdef CONFIG_CGROUP_SCHEDTUNE
5258 boost = schedtune_task_boost(task);
5260 boost = get_sysctl_sched_cfs_boost();
5265 util = task_util(task);
5266 margin = schedtune_margin(util, boost);
5271 #else /* CONFIG_SCHED_TUNE */
5273 static inline unsigned int
5274 schedtune_cpu_margin(unsigned long util, int cpu)
5279 static inline unsigned int
5280 schedtune_task_margin(struct task_struct *task)
5285 #endif /* CONFIG_SCHED_TUNE */
5287 static inline unsigned long
5288 boosted_cpu_util(int cpu)
5290 unsigned long util = cpu_util(cpu);
5291 unsigned long margin = schedtune_cpu_margin(util, cpu);
5293 trace_sched_boost_cpu(cpu, util, margin);
5295 return util + margin;
5298 static inline unsigned long
5299 boosted_task_util(struct task_struct *task)
5301 unsigned long util = task_util(task);
5302 unsigned long margin = schedtune_task_margin(task);
5304 return util + margin;
5308 * find_idlest_group finds and returns the least busy CPU group within the
5311 static struct sched_group *
5312 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5313 int this_cpu, int sd_flag)
5315 struct sched_group *idlest = NULL, *group = sd->groups;
5316 struct sched_group *fit_group = NULL, *spare_group = NULL;
5317 unsigned long min_load = ULONG_MAX, this_load = 0;
5318 unsigned long fit_capacity = ULONG_MAX;
5319 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5320 int load_idx = sd->forkexec_idx;
5321 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5323 if (sd_flag & SD_BALANCE_WAKE)
5324 load_idx = sd->wake_idx;
5327 unsigned long load, avg_load, spare_capacity;
5331 /* Skip over this group if it has no CPUs allowed */
5332 if (!cpumask_intersects(sched_group_cpus(group),
5333 tsk_cpus_allowed(p)))
5336 local_group = cpumask_test_cpu(this_cpu,
5337 sched_group_cpus(group));
5339 /* Tally up the load of all CPUs in the group */
5342 for_each_cpu(i, sched_group_cpus(group)) {
5343 /* Bias balancing toward cpus of our domain */
5345 load = source_load(i, load_idx);
5347 load = target_load(i, load_idx);
5352 * Look for most energy-efficient group that can fit
5353 * that can fit the task.
5355 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5356 fit_capacity = capacity_of(i);
5361 * Look for group which has most spare capacity on a
5364 spare_capacity = capacity_of(i) - cpu_util(i);
5365 if (spare_capacity > max_spare_capacity) {
5366 max_spare_capacity = spare_capacity;
5367 spare_group = group;
5371 /* Adjust by relative CPU capacity of the group */
5372 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5375 this_load = avg_load;
5376 } else if (avg_load < min_load) {
5377 min_load = avg_load;
5380 } while (group = group->next, group != sd->groups);
5388 if (!idlest || 100*this_load < imbalance*min_load)
5394 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5397 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5399 unsigned long load, min_load = ULONG_MAX;
5400 unsigned int min_exit_latency = UINT_MAX;
5401 u64 latest_idle_timestamp = 0;
5402 int least_loaded_cpu = this_cpu;
5403 int shallowest_idle_cpu = -1;
5406 /* Traverse only the allowed CPUs */
5407 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5408 if (task_fits_spare(p, i)) {
5409 struct rq *rq = cpu_rq(i);
5410 struct cpuidle_state *idle = idle_get_state(rq);
5411 if (idle && idle->exit_latency < min_exit_latency) {
5413 * We give priority to a CPU whose idle state
5414 * has the smallest exit latency irrespective
5415 * of any idle timestamp.
5417 min_exit_latency = idle->exit_latency;
5418 latest_idle_timestamp = rq->idle_stamp;
5419 shallowest_idle_cpu = i;
5420 } else if (idle_cpu(i) &&
5421 (!idle || idle->exit_latency == min_exit_latency) &&
5422 rq->idle_stamp > latest_idle_timestamp) {
5424 * If equal or no active idle state, then
5425 * the most recently idled CPU might have
5428 latest_idle_timestamp = rq->idle_stamp;
5429 shallowest_idle_cpu = i;
5430 } else if (shallowest_idle_cpu == -1) {
5432 * If we haven't found an idle CPU yet
5433 * pick a non-idle one that can fit the task as
5436 shallowest_idle_cpu = i;
5438 } else if (shallowest_idle_cpu == -1) {
5439 load = weighted_cpuload(i);
5440 if (load < min_load || (load == min_load && i == this_cpu)) {
5442 least_loaded_cpu = i;
5447 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5451 * Try and locate an idle CPU in the sched_domain.
5453 static int select_idle_sibling(struct task_struct *p, int target)
5455 struct sched_domain *sd;
5456 struct sched_group *sg;
5457 int i = task_cpu(p);
5459 if (idle_cpu(target))
5463 * If the prevous cpu is cache affine and idle, don't be stupid.
5465 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5469 * Otherwise, iterate the domains and find an elegible idle cpu.
5471 sd = rcu_dereference(per_cpu(sd_llc, target));
5472 for_each_lower_domain(sd) {
5475 if (!cpumask_intersects(sched_group_cpus(sg),
5476 tsk_cpus_allowed(p)))
5479 for_each_cpu(i, sched_group_cpus(sg)) {
5480 if (i == target || !idle_cpu(i))
5484 target = cpumask_first_and(sched_group_cpus(sg),
5485 tsk_cpus_allowed(p));
5489 } while (sg != sd->groups);
5495 static int energy_aware_wake_cpu(struct task_struct *p, int target)
5497 struct sched_domain *sd;
5498 struct sched_group *sg, *sg_target;
5499 int target_max_cap = INT_MAX;
5500 int target_cpu = task_cpu(p);
5503 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5512 * Find group with sufficient capacity. We only get here if no cpu is
5513 * overutilized. We may end up overutilizing a cpu by adding the task,
5514 * but that should not be any worse than select_idle_sibling().
5515 * load_balance() should sort it out later as we get above the tipping
5519 /* Assuming all cpus are the same in group */
5520 int max_cap_cpu = group_first_cpu(sg);
5523 * Assume smaller max capacity means more energy-efficient.
5524 * Ideally we should query the energy model for the right
5525 * answer but it easily ends up in an exhaustive search.
5527 if (capacity_of(max_cap_cpu) < target_max_cap &&
5528 task_fits_max(p, max_cap_cpu)) {
5530 target_max_cap = capacity_of(max_cap_cpu);
5532 } while (sg = sg->next, sg != sd->groups);
5534 /* Find cpu with sufficient capacity */
5535 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5537 * p's blocked utilization is still accounted for on prev_cpu
5538 * so prev_cpu will receive a negative bias due to the double
5539 * accounting. However, the blocked utilization may be zero.
5541 int new_util = cpu_util(i) + boosted_task_util(p);
5543 if (new_util > capacity_orig_of(i))
5546 if (new_util < capacity_curr_of(i)) {
5548 if (cpu_rq(i)->nr_running)
5552 /* cpu has capacity at higher OPP, keep it as fallback */
5553 if (target_cpu == task_cpu(p))
5557 if (target_cpu != task_cpu(p)) {
5558 struct energy_env eenv = {
5559 .util_delta = task_util(p),
5560 .src_cpu = task_cpu(p),
5561 .dst_cpu = target_cpu,
5565 /* Not enough spare capacity on previous cpu */
5566 if (cpu_overutilized(task_cpu(p)))
5569 if (energy_diff(&eenv) >= 0)
5577 * select_task_rq_fair: Select target runqueue for the waking task in domains
5578 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5579 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5581 * Balances load by selecting the idlest cpu in the idlest group, or under
5582 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5584 * Returns the target cpu number.
5586 * preempt must be disabled.
5589 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5591 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5592 int cpu = smp_processor_id();
5593 int new_cpu = prev_cpu;
5594 int want_affine = 0;
5595 int sync = wake_flags & WF_SYNC;
5597 if (sd_flag & SD_BALANCE_WAKE)
5598 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5599 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5603 for_each_domain(cpu, tmp) {
5604 if (!(tmp->flags & SD_LOAD_BALANCE))
5608 * If both cpu and prev_cpu are part of this domain,
5609 * cpu is a valid SD_WAKE_AFFINE target.
5611 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5612 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5617 if (tmp->flags & sd_flag)
5619 else if (!want_affine)
5624 sd = NULL; /* Prefer wake_affine over balance flags */
5625 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5630 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5631 new_cpu = energy_aware_wake_cpu(p, prev_cpu);
5632 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5633 new_cpu = select_idle_sibling(p, new_cpu);
5636 struct sched_group *group;
5639 if (!(sd->flags & sd_flag)) {
5644 group = find_idlest_group(sd, p, cpu, sd_flag);
5650 new_cpu = find_idlest_cpu(group, p, cpu);
5651 if (new_cpu == -1 || new_cpu == cpu) {
5652 /* Now try balancing at a lower domain level of cpu */
5657 /* Now try balancing at a lower domain level of new_cpu */
5659 weight = sd->span_weight;
5661 for_each_domain(cpu, tmp) {
5662 if (weight <= tmp->span_weight)
5664 if (tmp->flags & sd_flag)
5667 /* while loop will break here if sd == NULL */
5675 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5676 * cfs_rq_of(p) references at time of call are still valid and identify the
5677 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5678 * other assumptions, including the state of rq->lock, should be made.
5680 static void migrate_task_rq_fair(struct task_struct *p)
5683 * We are supposed to update the task to "current" time, then its up to date
5684 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5685 * what current time is, so simply throw away the out-of-date time. This
5686 * will result in the wakee task is less decayed, but giving the wakee more
5687 * load sounds not bad.
5689 remove_entity_load_avg(&p->se);
5691 /* Tell new CPU we are migrated */
5692 p->se.avg.last_update_time = 0;
5694 /* We have migrated, no longer consider this task hot */
5695 p->se.exec_start = 0;
5698 static void task_dead_fair(struct task_struct *p)
5700 remove_entity_load_avg(&p->se);
5702 #endif /* CONFIG_SMP */
5704 static unsigned long
5705 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5707 unsigned long gran = sysctl_sched_wakeup_granularity;
5710 * Since its curr running now, convert the gran from real-time
5711 * to virtual-time in his units.
5713 * By using 'se' instead of 'curr' we penalize light tasks, so
5714 * they get preempted easier. That is, if 'se' < 'curr' then
5715 * the resulting gran will be larger, therefore penalizing the
5716 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5717 * be smaller, again penalizing the lighter task.
5719 * This is especially important for buddies when the leftmost
5720 * task is higher priority than the buddy.
5722 return calc_delta_fair(gran, se);
5726 * Should 'se' preempt 'curr'.
5740 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5742 s64 gran, vdiff = curr->vruntime - se->vruntime;
5747 gran = wakeup_gran(curr, se);
5754 static void set_last_buddy(struct sched_entity *se)
5756 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5759 for_each_sched_entity(se)
5760 cfs_rq_of(se)->last = se;
5763 static void set_next_buddy(struct sched_entity *se)
5765 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5768 for_each_sched_entity(se)
5769 cfs_rq_of(se)->next = se;
5772 static void set_skip_buddy(struct sched_entity *se)
5774 for_each_sched_entity(se)
5775 cfs_rq_of(se)->skip = se;
5779 * Preempt the current task with a newly woken task if needed:
5781 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5783 struct task_struct *curr = rq->curr;
5784 struct sched_entity *se = &curr->se, *pse = &p->se;
5785 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5786 int scale = cfs_rq->nr_running >= sched_nr_latency;
5787 int next_buddy_marked = 0;
5789 if (unlikely(se == pse))
5793 * This is possible from callers such as attach_tasks(), in which we
5794 * unconditionally check_prempt_curr() after an enqueue (which may have
5795 * lead to a throttle). This both saves work and prevents false
5796 * next-buddy nomination below.
5798 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5801 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5802 set_next_buddy(pse);
5803 next_buddy_marked = 1;
5807 * We can come here with TIF_NEED_RESCHED already set from new task
5810 * Note: this also catches the edge-case of curr being in a throttled
5811 * group (e.g. via set_curr_task), since update_curr() (in the
5812 * enqueue of curr) will have resulted in resched being set. This
5813 * prevents us from potentially nominating it as a false LAST_BUDDY
5816 if (test_tsk_need_resched(curr))
5819 /* Idle tasks are by definition preempted by non-idle tasks. */
5820 if (unlikely(curr->policy == SCHED_IDLE) &&
5821 likely(p->policy != SCHED_IDLE))
5825 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5826 * is driven by the tick):
5828 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5831 find_matching_se(&se, &pse);
5832 update_curr(cfs_rq_of(se));
5834 if (wakeup_preempt_entity(se, pse) == 1) {
5836 * Bias pick_next to pick the sched entity that is
5837 * triggering this preemption.
5839 if (!next_buddy_marked)
5840 set_next_buddy(pse);
5849 * Only set the backward buddy when the current task is still
5850 * on the rq. This can happen when a wakeup gets interleaved
5851 * with schedule on the ->pre_schedule() or idle_balance()
5852 * point, either of which can * drop the rq lock.
5854 * Also, during early boot the idle thread is in the fair class,
5855 * for obvious reasons its a bad idea to schedule back to it.
5857 if (unlikely(!se->on_rq || curr == rq->idle))
5860 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5864 static struct task_struct *
5865 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5867 struct cfs_rq *cfs_rq = &rq->cfs;
5868 struct sched_entity *se;
5869 struct task_struct *p;
5873 #ifdef CONFIG_FAIR_GROUP_SCHED
5874 if (!cfs_rq->nr_running)
5877 if (prev->sched_class != &fair_sched_class)
5881 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5882 * likely that a next task is from the same cgroup as the current.
5884 * Therefore attempt to avoid putting and setting the entire cgroup
5885 * hierarchy, only change the part that actually changes.
5889 struct sched_entity *curr = cfs_rq->curr;
5892 * Since we got here without doing put_prev_entity() we also
5893 * have to consider cfs_rq->curr. If it is still a runnable
5894 * entity, update_curr() will update its vruntime, otherwise
5895 * forget we've ever seen it.
5899 update_curr(cfs_rq);
5904 * This call to check_cfs_rq_runtime() will do the
5905 * throttle and dequeue its entity in the parent(s).
5906 * Therefore the 'simple' nr_running test will indeed
5909 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5913 se = pick_next_entity(cfs_rq, curr);
5914 cfs_rq = group_cfs_rq(se);
5920 * Since we haven't yet done put_prev_entity and if the selected task
5921 * is a different task than we started out with, try and touch the
5922 * least amount of cfs_rqs.
5925 struct sched_entity *pse = &prev->se;
5927 while (!(cfs_rq = is_same_group(se, pse))) {
5928 int se_depth = se->depth;
5929 int pse_depth = pse->depth;
5931 if (se_depth <= pse_depth) {
5932 put_prev_entity(cfs_rq_of(pse), pse);
5933 pse = parent_entity(pse);
5935 if (se_depth >= pse_depth) {
5936 set_next_entity(cfs_rq_of(se), se);
5937 se = parent_entity(se);
5941 put_prev_entity(cfs_rq, pse);
5942 set_next_entity(cfs_rq, se);
5945 if (hrtick_enabled(rq))
5946 hrtick_start_fair(rq, p);
5948 rq->misfit_task = !task_fits_max(p, rq->cpu);
5955 if (!cfs_rq->nr_running)
5958 put_prev_task(rq, prev);
5961 se = pick_next_entity(cfs_rq, NULL);
5962 set_next_entity(cfs_rq, se);
5963 cfs_rq = group_cfs_rq(se);
5968 if (hrtick_enabled(rq))
5969 hrtick_start_fair(rq, p);
5971 rq->misfit_task = !task_fits_max(p, rq->cpu);
5976 rq->misfit_task = 0;
5978 * This is OK, because current is on_cpu, which avoids it being picked
5979 * for load-balance and preemption/IRQs are still disabled avoiding
5980 * further scheduler activity on it and we're being very careful to
5981 * re-start the picking loop.
5983 lockdep_unpin_lock(&rq->lock);
5984 new_tasks = idle_balance(rq);
5985 lockdep_pin_lock(&rq->lock);
5987 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5988 * possible for any higher priority task to appear. In that case we
5989 * must re-start the pick_next_entity() loop.
6001 * Account for a descheduled task:
6003 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6005 struct sched_entity *se = &prev->se;
6006 struct cfs_rq *cfs_rq;
6008 for_each_sched_entity(se) {
6009 cfs_rq = cfs_rq_of(se);
6010 put_prev_entity(cfs_rq, se);
6015 * sched_yield() is very simple
6017 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6019 static void yield_task_fair(struct rq *rq)
6021 struct task_struct *curr = rq->curr;
6022 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6023 struct sched_entity *se = &curr->se;
6026 * Are we the only task in the tree?
6028 if (unlikely(rq->nr_running == 1))
6031 clear_buddies(cfs_rq, se);
6033 if (curr->policy != SCHED_BATCH) {
6034 update_rq_clock(rq);
6036 * Update run-time statistics of the 'current'.
6038 update_curr(cfs_rq);
6040 * Tell update_rq_clock() that we've just updated,
6041 * so we don't do microscopic update in schedule()
6042 * and double the fastpath cost.
6044 rq_clock_skip_update(rq, true);
6050 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6052 struct sched_entity *se = &p->se;
6054 /* throttled hierarchies are not runnable */
6055 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6058 /* Tell the scheduler that we'd really like pse to run next. */
6061 yield_task_fair(rq);
6067 /**************************************************
6068 * Fair scheduling class load-balancing methods.
6072 * The purpose of load-balancing is to achieve the same basic fairness the
6073 * per-cpu scheduler provides, namely provide a proportional amount of compute
6074 * time to each task. This is expressed in the following equation:
6076 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6078 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6079 * W_i,0 is defined as:
6081 * W_i,0 = \Sum_j w_i,j (2)
6083 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6084 * is derived from the nice value as per prio_to_weight[].
6086 * The weight average is an exponential decay average of the instantaneous
6089 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6091 * C_i is the compute capacity of cpu i, typically it is the
6092 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6093 * can also include other factors [XXX].
6095 * To achieve this balance we define a measure of imbalance which follows
6096 * directly from (1):
6098 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6100 * We them move tasks around to minimize the imbalance. In the continuous
6101 * function space it is obvious this converges, in the discrete case we get
6102 * a few fun cases generally called infeasible weight scenarios.
6105 * - infeasible weights;
6106 * - local vs global optima in the discrete case. ]
6111 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6112 * for all i,j solution, we create a tree of cpus that follows the hardware
6113 * topology where each level pairs two lower groups (or better). This results
6114 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6115 * tree to only the first of the previous level and we decrease the frequency
6116 * of load-balance at each level inv. proportional to the number of cpus in
6122 * \Sum { --- * --- * 2^i } = O(n) (5)
6124 * `- size of each group
6125 * | | `- number of cpus doing load-balance
6127 * `- sum over all levels
6129 * Coupled with a limit on how many tasks we can migrate every balance pass,
6130 * this makes (5) the runtime complexity of the balancer.
6132 * An important property here is that each CPU is still (indirectly) connected
6133 * to every other cpu in at most O(log n) steps:
6135 * The adjacency matrix of the resulting graph is given by:
6138 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6141 * And you'll find that:
6143 * A^(log_2 n)_i,j != 0 for all i,j (7)
6145 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6146 * The task movement gives a factor of O(m), giving a convergence complexity
6149 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6154 * In order to avoid CPUs going idle while there's still work to do, new idle
6155 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6156 * tree itself instead of relying on other CPUs to bring it work.
6158 * This adds some complexity to both (5) and (8) but it reduces the total idle
6166 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6169 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6174 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6176 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6178 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6181 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6182 * rewrite all of this once again.]
6185 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6187 enum fbq_type { regular, remote, all };
6196 #define LBF_ALL_PINNED 0x01
6197 #define LBF_NEED_BREAK 0x02
6198 #define LBF_DST_PINNED 0x04
6199 #define LBF_SOME_PINNED 0x08
6202 struct sched_domain *sd;
6210 struct cpumask *dst_grpmask;
6212 enum cpu_idle_type idle;
6214 unsigned int src_grp_nr_running;
6215 /* The set of CPUs under consideration for load-balancing */
6216 struct cpumask *cpus;
6221 unsigned int loop_break;
6222 unsigned int loop_max;
6224 enum fbq_type fbq_type;
6225 enum group_type busiest_group_type;
6226 struct list_head tasks;
6230 * Is this task likely cache-hot:
6232 static int task_hot(struct task_struct *p, struct lb_env *env)
6236 lockdep_assert_held(&env->src_rq->lock);
6238 if (p->sched_class != &fair_sched_class)
6241 if (unlikely(p->policy == SCHED_IDLE))
6245 * Buddy candidates are cache hot:
6247 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6248 (&p->se == cfs_rq_of(&p->se)->next ||
6249 &p->se == cfs_rq_of(&p->se)->last))
6252 if (sysctl_sched_migration_cost == -1)
6254 if (sysctl_sched_migration_cost == 0)
6257 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6259 return delta < (s64)sysctl_sched_migration_cost;
6262 #ifdef CONFIG_NUMA_BALANCING
6264 * Returns 1, if task migration degrades locality
6265 * Returns 0, if task migration improves locality i.e migration preferred.
6266 * Returns -1, if task migration is not affected by locality.
6268 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6270 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6271 unsigned long src_faults, dst_faults;
6272 int src_nid, dst_nid;
6274 if (!static_branch_likely(&sched_numa_balancing))
6277 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6280 src_nid = cpu_to_node(env->src_cpu);
6281 dst_nid = cpu_to_node(env->dst_cpu);
6283 if (src_nid == dst_nid)
6286 /* Migrating away from the preferred node is always bad. */
6287 if (src_nid == p->numa_preferred_nid) {
6288 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6294 /* Encourage migration to the preferred node. */
6295 if (dst_nid == p->numa_preferred_nid)
6299 src_faults = group_faults(p, src_nid);
6300 dst_faults = group_faults(p, dst_nid);
6302 src_faults = task_faults(p, src_nid);
6303 dst_faults = task_faults(p, dst_nid);
6306 return dst_faults < src_faults;
6310 static inline int migrate_degrades_locality(struct task_struct *p,
6318 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6321 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6325 lockdep_assert_held(&env->src_rq->lock);
6328 * We do not migrate tasks that are:
6329 * 1) throttled_lb_pair, or
6330 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6331 * 3) running (obviously), or
6332 * 4) are cache-hot on their current CPU.
6334 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6337 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6340 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6342 env->flags |= LBF_SOME_PINNED;
6345 * Remember if this task can be migrated to any other cpu in
6346 * our sched_group. We may want to revisit it if we couldn't
6347 * meet load balance goals by pulling other tasks on src_cpu.
6349 * Also avoid computing new_dst_cpu if we have already computed
6350 * one in current iteration.
6352 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6355 /* Prevent to re-select dst_cpu via env's cpus */
6356 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6357 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6358 env->flags |= LBF_DST_PINNED;
6359 env->new_dst_cpu = cpu;
6367 /* Record that we found atleast one task that could run on dst_cpu */
6368 env->flags &= ~LBF_ALL_PINNED;
6370 if (task_running(env->src_rq, p)) {
6371 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6376 * Aggressive migration if:
6377 * 1) destination numa is preferred
6378 * 2) task is cache cold, or
6379 * 3) too many balance attempts have failed.
6381 tsk_cache_hot = migrate_degrades_locality(p, env);
6382 if (tsk_cache_hot == -1)
6383 tsk_cache_hot = task_hot(p, env);
6385 if (tsk_cache_hot <= 0 ||
6386 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6387 if (tsk_cache_hot == 1) {
6388 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6389 schedstat_inc(p, se.statistics.nr_forced_migrations);
6394 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6399 * detach_task() -- detach the task for the migration specified in env
6401 static void detach_task(struct task_struct *p, struct lb_env *env)
6403 lockdep_assert_held(&env->src_rq->lock);
6405 deactivate_task(env->src_rq, p, 0);
6406 p->on_rq = TASK_ON_RQ_MIGRATING;
6407 set_task_cpu(p, env->dst_cpu);
6411 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6412 * part of active balancing operations within "domain".
6414 * Returns a task if successful and NULL otherwise.
6416 static struct task_struct *detach_one_task(struct lb_env *env)
6418 struct task_struct *p, *n;
6420 lockdep_assert_held(&env->src_rq->lock);
6422 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6423 if (!can_migrate_task(p, env))
6426 detach_task(p, env);
6429 * Right now, this is only the second place where
6430 * lb_gained[env->idle] is updated (other is detach_tasks)
6431 * so we can safely collect stats here rather than
6432 * inside detach_tasks().
6434 schedstat_inc(env->sd, lb_gained[env->idle]);
6440 static const unsigned int sched_nr_migrate_break = 32;
6443 * detach_tasks() -- tries to detach up to imbalance weighted load from
6444 * busiest_rq, as part of a balancing operation within domain "sd".
6446 * Returns number of detached tasks if successful and 0 otherwise.
6448 static int detach_tasks(struct lb_env *env)
6450 struct list_head *tasks = &env->src_rq->cfs_tasks;
6451 struct task_struct *p;
6455 lockdep_assert_held(&env->src_rq->lock);
6457 if (env->imbalance <= 0)
6460 while (!list_empty(tasks)) {
6462 * We don't want to steal all, otherwise we may be treated likewise,
6463 * which could at worst lead to a livelock crash.
6465 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6468 p = list_first_entry(tasks, struct task_struct, se.group_node);
6471 /* We've more or less seen every task there is, call it quits */
6472 if (env->loop > env->loop_max)
6475 /* take a breather every nr_migrate tasks */
6476 if (env->loop > env->loop_break) {
6477 env->loop_break += sched_nr_migrate_break;
6478 env->flags |= LBF_NEED_BREAK;
6482 if (!can_migrate_task(p, env))
6485 load = task_h_load(p);
6487 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6490 if ((load / 2) > env->imbalance)
6493 detach_task(p, env);
6494 list_add(&p->se.group_node, &env->tasks);
6497 env->imbalance -= load;
6499 #ifdef CONFIG_PREEMPT
6501 * NEWIDLE balancing is a source of latency, so preemptible
6502 * kernels will stop after the first task is detached to minimize
6503 * the critical section.
6505 if (env->idle == CPU_NEWLY_IDLE)
6510 * We only want to steal up to the prescribed amount of
6513 if (env->imbalance <= 0)
6518 list_move_tail(&p->se.group_node, tasks);
6522 * Right now, this is one of only two places we collect this stat
6523 * so we can safely collect detach_one_task() stats here rather
6524 * than inside detach_one_task().
6526 schedstat_add(env->sd, lb_gained[env->idle], detached);
6532 * attach_task() -- attach the task detached by detach_task() to its new rq.
6534 static void attach_task(struct rq *rq, struct task_struct *p)
6536 lockdep_assert_held(&rq->lock);
6538 BUG_ON(task_rq(p) != rq);
6539 p->on_rq = TASK_ON_RQ_QUEUED;
6540 activate_task(rq, p, 0);
6541 check_preempt_curr(rq, p, 0);
6545 * attach_one_task() -- attaches the task returned from detach_one_task() to
6548 static void attach_one_task(struct rq *rq, struct task_struct *p)
6550 raw_spin_lock(&rq->lock);
6553 * We want to potentially raise target_cpu's OPP.
6555 update_capacity_of(cpu_of(rq));
6556 raw_spin_unlock(&rq->lock);
6560 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6563 static void attach_tasks(struct lb_env *env)
6565 struct list_head *tasks = &env->tasks;
6566 struct task_struct *p;
6568 raw_spin_lock(&env->dst_rq->lock);
6570 while (!list_empty(tasks)) {
6571 p = list_first_entry(tasks, struct task_struct, se.group_node);
6572 list_del_init(&p->se.group_node);
6574 attach_task(env->dst_rq, p);
6578 * We want to potentially raise env.dst_cpu's OPP.
6580 update_capacity_of(env->dst_cpu);
6582 raw_spin_unlock(&env->dst_rq->lock);
6585 #ifdef CONFIG_FAIR_GROUP_SCHED
6586 static void update_blocked_averages(int cpu)
6588 struct rq *rq = cpu_rq(cpu);
6589 struct cfs_rq *cfs_rq;
6590 unsigned long flags;
6592 raw_spin_lock_irqsave(&rq->lock, flags);
6593 update_rq_clock(rq);
6596 * Iterates the task_group tree in a bottom up fashion, see
6597 * list_add_leaf_cfs_rq() for details.
6599 for_each_leaf_cfs_rq(rq, cfs_rq) {
6600 /* throttled entities do not contribute to load */
6601 if (throttled_hierarchy(cfs_rq))
6604 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6605 update_tg_load_avg(cfs_rq, 0);
6607 raw_spin_unlock_irqrestore(&rq->lock, flags);
6611 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6612 * This needs to be done in a top-down fashion because the load of a child
6613 * group is a fraction of its parents load.
6615 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6617 struct rq *rq = rq_of(cfs_rq);
6618 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6619 unsigned long now = jiffies;
6622 if (cfs_rq->last_h_load_update == now)
6625 cfs_rq->h_load_next = NULL;
6626 for_each_sched_entity(se) {
6627 cfs_rq = cfs_rq_of(se);
6628 cfs_rq->h_load_next = se;
6629 if (cfs_rq->last_h_load_update == now)
6634 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6635 cfs_rq->last_h_load_update = now;
6638 while ((se = cfs_rq->h_load_next) != NULL) {
6639 load = cfs_rq->h_load;
6640 load = div64_ul(load * se->avg.load_avg,
6641 cfs_rq_load_avg(cfs_rq) + 1);
6642 cfs_rq = group_cfs_rq(se);
6643 cfs_rq->h_load = load;
6644 cfs_rq->last_h_load_update = now;
6648 static unsigned long task_h_load(struct task_struct *p)
6650 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6652 update_cfs_rq_h_load(cfs_rq);
6653 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6654 cfs_rq_load_avg(cfs_rq) + 1);
6657 static inline void update_blocked_averages(int cpu)
6659 struct rq *rq = cpu_rq(cpu);
6660 struct cfs_rq *cfs_rq = &rq->cfs;
6661 unsigned long flags;
6663 raw_spin_lock_irqsave(&rq->lock, flags);
6664 update_rq_clock(rq);
6665 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6666 raw_spin_unlock_irqrestore(&rq->lock, flags);
6669 static unsigned long task_h_load(struct task_struct *p)
6671 return p->se.avg.load_avg;
6675 /********** Helpers for find_busiest_group ************************/
6678 * sg_lb_stats - stats of a sched_group required for load_balancing
6680 struct sg_lb_stats {
6681 unsigned long avg_load; /*Avg load across the CPUs of the group */
6682 unsigned long group_load; /* Total load over the CPUs of the group */
6683 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6684 unsigned long load_per_task;
6685 unsigned long group_capacity;
6686 unsigned long group_util; /* Total utilization of the group */
6687 unsigned int sum_nr_running; /* Nr tasks running in the group */
6688 unsigned int idle_cpus;
6689 unsigned int group_weight;
6690 enum group_type group_type;
6691 int group_no_capacity;
6692 int group_misfit_task; /* A cpu has a task too big for its capacity */
6693 #ifdef CONFIG_NUMA_BALANCING
6694 unsigned int nr_numa_running;
6695 unsigned int nr_preferred_running;
6700 * sd_lb_stats - Structure to store the statistics of a sched_domain
6701 * during load balancing.
6703 struct sd_lb_stats {
6704 struct sched_group *busiest; /* Busiest group in this sd */
6705 struct sched_group *local; /* Local group in this sd */
6706 unsigned long total_load; /* Total load of all groups in sd */
6707 unsigned long total_capacity; /* Total capacity of all groups in sd */
6708 unsigned long avg_load; /* Average load across all groups in sd */
6710 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6711 struct sg_lb_stats local_stat; /* Statistics of the local group */
6714 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6717 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6718 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6719 * We must however clear busiest_stat::avg_load because
6720 * update_sd_pick_busiest() reads this before assignment.
6722 *sds = (struct sd_lb_stats){
6726 .total_capacity = 0UL,
6729 .sum_nr_running = 0,
6730 .group_type = group_other,
6736 * get_sd_load_idx - Obtain the load index for a given sched domain.
6737 * @sd: The sched_domain whose load_idx is to be obtained.
6738 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6740 * Return: The load index.
6742 static inline int get_sd_load_idx(struct sched_domain *sd,
6743 enum cpu_idle_type idle)
6749 load_idx = sd->busy_idx;
6752 case CPU_NEWLY_IDLE:
6753 load_idx = sd->newidle_idx;
6756 load_idx = sd->idle_idx;
6763 static unsigned long scale_rt_capacity(int cpu)
6765 struct rq *rq = cpu_rq(cpu);
6766 u64 total, used, age_stamp, avg;
6770 * Since we're reading these variables without serialization make sure
6771 * we read them once before doing sanity checks on them.
6773 age_stamp = READ_ONCE(rq->age_stamp);
6774 avg = READ_ONCE(rq->rt_avg);
6775 delta = __rq_clock_broken(rq) - age_stamp;
6777 if (unlikely(delta < 0))
6780 total = sched_avg_period() + delta;
6782 used = div_u64(avg, total);
6785 * deadline bandwidth is defined at system level so we must
6786 * weight this bandwidth with the max capacity of the system.
6787 * As a reminder, avg_bw is 20bits width and
6788 * scale_cpu_capacity is 10 bits width
6790 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
6792 if (likely(used < SCHED_CAPACITY_SCALE))
6793 return SCHED_CAPACITY_SCALE - used;
6798 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
6800 raw_spin_lock_init(&mcc->lock);
6805 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6807 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6808 struct sched_group *sdg = sd->groups;
6809 struct max_cpu_capacity *mcc;
6810 unsigned long max_capacity;
6812 unsigned long flags;
6814 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6816 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
6818 raw_spin_lock_irqsave(&mcc->lock, flags);
6819 max_capacity = mcc->val;
6820 max_cap_cpu = mcc->cpu;
6822 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
6823 (max_capacity < capacity)) {
6824 mcc->val = capacity;
6826 #ifdef CONFIG_SCHED_DEBUG
6827 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6828 //pr_info("CPU%d: update max cpu_capacity %lu\n", cpu, capacity);
6832 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6834 skip_unlock: __attribute__ ((unused));
6835 capacity *= scale_rt_capacity(cpu);
6836 capacity >>= SCHED_CAPACITY_SHIFT;
6841 cpu_rq(cpu)->cpu_capacity = capacity;
6842 sdg->sgc->capacity = capacity;
6843 sdg->sgc->max_capacity = capacity;
6846 void update_group_capacity(struct sched_domain *sd, int cpu)
6848 struct sched_domain *child = sd->child;
6849 struct sched_group *group, *sdg = sd->groups;
6850 unsigned long capacity, max_capacity;
6851 unsigned long interval;
6853 interval = msecs_to_jiffies(sd->balance_interval);
6854 interval = clamp(interval, 1UL, max_load_balance_interval);
6855 sdg->sgc->next_update = jiffies + interval;
6858 update_cpu_capacity(sd, cpu);
6865 if (child->flags & SD_OVERLAP) {
6867 * SD_OVERLAP domains cannot assume that child groups
6868 * span the current group.
6871 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6872 struct sched_group_capacity *sgc;
6873 struct rq *rq = cpu_rq(cpu);
6876 * build_sched_domains() -> init_sched_groups_capacity()
6877 * gets here before we've attached the domains to the
6880 * Use capacity_of(), which is set irrespective of domains
6881 * in update_cpu_capacity().
6883 * This avoids capacity from being 0 and
6884 * causing divide-by-zero issues on boot.
6886 if (unlikely(!rq->sd)) {
6887 capacity += capacity_of(cpu);
6889 sgc = rq->sd->groups->sgc;
6890 capacity += sgc->capacity;
6893 max_capacity = max(capacity, max_capacity);
6897 * !SD_OVERLAP domains can assume that child groups
6898 * span the current group.
6901 group = child->groups;
6903 struct sched_group_capacity *sgc = group->sgc;
6905 capacity += sgc->capacity;
6906 max_capacity = max(sgc->max_capacity, max_capacity);
6907 group = group->next;
6908 } while (group != child->groups);
6911 sdg->sgc->capacity = capacity;
6912 sdg->sgc->max_capacity = max_capacity;
6916 * Check whether the capacity of the rq has been noticeably reduced by side
6917 * activity. The imbalance_pct is used for the threshold.
6918 * Return true is the capacity is reduced
6921 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6923 return ((rq->cpu_capacity * sd->imbalance_pct) <
6924 (rq->cpu_capacity_orig * 100));
6928 * Group imbalance indicates (and tries to solve) the problem where balancing
6929 * groups is inadequate due to tsk_cpus_allowed() constraints.
6931 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6932 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6935 * { 0 1 2 3 } { 4 5 6 7 }
6938 * If we were to balance group-wise we'd place two tasks in the first group and
6939 * two tasks in the second group. Clearly this is undesired as it will overload
6940 * cpu 3 and leave one of the cpus in the second group unused.
6942 * The current solution to this issue is detecting the skew in the first group
6943 * by noticing the lower domain failed to reach balance and had difficulty
6944 * moving tasks due to affinity constraints.
6946 * When this is so detected; this group becomes a candidate for busiest; see
6947 * update_sd_pick_busiest(). And calculate_imbalance() and
6948 * find_busiest_group() avoid some of the usual balance conditions to allow it
6949 * to create an effective group imbalance.
6951 * This is a somewhat tricky proposition since the next run might not find the
6952 * group imbalance and decide the groups need to be balanced again. A most
6953 * subtle and fragile situation.
6956 static inline int sg_imbalanced(struct sched_group *group)
6958 return group->sgc->imbalance;
6962 * group_has_capacity returns true if the group has spare capacity that could
6963 * be used by some tasks.
6964 * We consider that a group has spare capacity if the * number of task is
6965 * smaller than the number of CPUs or if the utilization is lower than the
6966 * available capacity for CFS tasks.
6967 * For the latter, we use a threshold to stabilize the state, to take into
6968 * account the variance of the tasks' load and to return true if the available
6969 * capacity in meaningful for the load balancer.
6970 * As an example, an available capacity of 1% can appear but it doesn't make
6971 * any benefit for the load balance.
6974 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6976 if (sgs->sum_nr_running < sgs->group_weight)
6979 if ((sgs->group_capacity * 100) >
6980 (sgs->group_util * env->sd->imbalance_pct))
6987 * group_is_overloaded returns true if the group has more tasks than it can
6989 * group_is_overloaded is not equals to !group_has_capacity because a group
6990 * with the exact right number of tasks, has no more spare capacity but is not
6991 * overloaded so both group_has_capacity and group_is_overloaded return
6995 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6997 if (sgs->sum_nr_running <= sgs->group_weight)
7000 if ((sgs->group_capacity * 100) <
7001 (sgs->group_util * env->sd->imbalance_pct))
7009 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7010 * per-cpu capacity than sched_group ref.
7013 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7015 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7016 ref->sgc->max_capacity;
7020 group_type group_classify(struct sched_group *group,
7021 struct sg_lb_stats *sgs)
7023 if (sgs->group_no_capacity)
7024 return group_overloaded;
7026 if (sg_imbalanced(group))
7027 return group_imbalanced;
7029 if (sgs->group_misfit_task)
7030 return group_misfit_task;
7036 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7037 * @env: The load balancing environment.
7038 * @group: sched_group whose statistics are to be updated.
7039 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7040 * @local_group: Does group contain this_cpu.
7041 * @sgs: variable to hold the statistics for this group.
7042 * @overload: Indicate more than one runnable task for any CPU.
7043 * @overutilized: Indicate overutilization for any CPU.
7045 static inline void update_sg_lb_stats(struct lb_env *env,
7046 struct sched_group *group, int load_idx,
7047 int local_group, struct sg_lb_stats *sgs,
7048 bool *overload, bool *overutilized)
7053 memset(sgs, 0, sizeof(*sgs));
7055 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7056 struct rq *rq = cpu_rq(i);
7058 /* Bias balancing toward cpus of our domain */
7060 load = target_load(i, load_idx);
7062 load = source_load(i, load_idx);
7064 sgs->group_load += load;
7065 sgs->group_util += cpu_util(i);
7066 sgs->sum_nr_running += rq->cfs.h_nr_running;
7068 if (rq->nr_running > 1)
7071 #ifdef CONFIG_NUMA_BALANCING
7072 sgs->nr_numa_running += rq->nr_numa_running;
7073 sgs->nr_preferred_running += rq->nr_preferred_running;
7075 sgs->sum_weighted_load += weighted_cpuload(i);
7079 if (cpu_overutilized(i)) {
7080 *overutilized = true;
7081 if (!sgs->group_misfit_task && rq->misfit_task)
7082 sgs->group_misfit_task = capacity_of(i);
7086 /* Adjust by relative CPU capacity of the group */
7087 sgs->group_capacity = group->sgc->capacity;
7088 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7090 if (sgs->sum_nr_running)
7091 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7093 sgs->group_weight = group->group_weight;
7095 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7096 sgs->group_type = group_classify(group, sgs);
7100 * update_sd_pick_busiest - return 1 on busiest group
7101 * @env: The load balancing environment.
7102 * @sds: sched_domain statistics
7103 * @sg: sched_group candidate to be checked for being the busiest
7104 * @sgs: sched_group statistics
7106 * Determine if @sg is a busier group than the previously selected
7109 * Return: %true if @sg is a busier group than the previously selected
7110 * busiest group. %false otherwise.
7112 static bool update_sd_pick_busiest(struct lb_env *env,
7113 struct sd_lb_stats *sds,
7114 struct sched_group *sg,
7115 struct sg_lb_stats *sgs)
7117 struct sg_lb_stats *busiest = &sds->busiest_stat;
7119 if (sgs->group_type > busiest->group_type)
7122 if (sgs->group_type < busiest->group_type)
7126 * Candidate sg doesn't face any serious load-balance problems
7127 * so don't pick it if the local sg is already filled up.
7129 if (sgs->group_type == group_other &&
7130 !group_has_capacity(env, &sds->local_stat))
7133 if (sgs->avg_load <= busiest->avg_load)
7137 * Candiate sg has no more than one task per cpu and has higher
7138 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7140 if (sgs->sum_nr_running <= sgs->group_weight &&
7141 group_smaller_cpu_capacity(sds->local, sg))
7144 /* This is the busiest node in its class. */
7145 if (!(env->sd->flags & SD_ASYM_PACKING))
7149 * ASYM_PACKING needs to move all the work to the lowest
7150 * numbered CPUs in the group, therefore mark all groups
7151 * higher than ourself as busy.
7153 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7157 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7164 #ifdef CONFIG_NUMA_BALANCING
7165 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7167 if (sgs->sum_nr_running > sgs->nr_numa_running)
7169 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7174 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7176 if (rq->nr_running > rq->nr_numa_running)
7178 if (rq->nr_running > rq->nr_preferred_running)
7183 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7188 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7192 #endif /* CONFIG_NUMA_BALANCING */
7195 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7196 * @env: The load balancing environment.
7197 * @sds: variable to hold the statistics for this sched_domain.
7199 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7201 struct sched_domain *child = env->sd->child;
7202 struct sched_group *sg = env->sd->groups;
7203 struct sg_lb_stats tmp_sgs;
7204 int load_idx, prefer_sibling = 0;
7205 bool overload = false, overutilized = false;
7207 if (child && child->flags & SD_PREFER_SIBLING)
7210 load_idx = get_sd_load_idx(env->sd, env->idle);
7213 struct sg_lb_stats *sgs = &tmp_sgs;
7216 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7219 sgs = &sds->local_stat;
7221 if (env->idle != CPU_NEWLY_IDLE ||
7222 time_after_eq(jiffies, sg->sgc->next_update))
7223 update_group_capacity(env->sd, env->dst_cpu);
7226 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7227 &overload, &overutilized);
7233 * In case the child domain prefers tasks go to siblings
7234 * first, lower the sg capacity so that we'll try
7235 * and move all the excess tasks away. We lower the capacity
7236 * of a group only if the local group has the capacity to fit
7237 * these excess tasks. The extra check prevents the case where
7238 * you always pull from the heaviest group when it is already
7239 * under-utilized (possible with a large weight task outweighs
7240 * the tasks on the system).
7242 if (prefer_sibling && sds->local &&
7243 group_has_capacity(env, &sds->local_stat) &&
7244 (sgs->sum_nr_running > 1)) {
7245 sgs->group_no_capacity = 1;
7246 sgs->group_type = group_classify(sg, sgs);
7250 * Ignore task groups with misfit tasks if local group has no
7251 * capacity or if per-cpu capacity isn't higher.
7253 if (sgs->group_type == group_misfit_task &&
7254 (!group_has_capacity(env, &sds->local_stat) ||
7255 !group_smaller_cpu_capacity(sg, sds->local)))
7256 sgs->group_type = group_other;
7258 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7260 sds->busiest_stat = *sgs;
7264 /* Now, start updating sd_lb_stats */
7265 sds->total_load += sgs->group_load;
7266 sds->total_capacity += sgs->group_capacity;
7269 } while (sg != env->sd->groups);
7271 if (env->sd->flags & SD_NUMA)
7272 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7274 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7276 if (!env->sd->parent) {
7277 /* update overload indicator if we are at root domain */
7278 if (env->dst_rq->rd->overload != overload)
7279 env->dst_rq->rd->overload = overload;
7281 /* Update over-utilization (tipping point, U >= 0) indicator */
7282 if (env->dst_rq->rd->overutilized != overutilized)
7283 env->dst_rq->rd->overutilized = overutilized;
7285 if (!env->dst_rq->rd->overutilized && overutilized)
7286 env->dst_rq->rd->overutilized = true;
7291 * check_asym_packing - Check to see if the group is packed into the
7294 * This is primarily intended to used at the sibling level. Some
7295 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7296 * case of POWER7, it can move to lower SMT modes only when higher
7297 * threads are idle. When in lower SMT modes, the threads will
7298 * perform better since they share less core resources. Hence when we
7299 * have idle threads, we want them to be the higher ones.
7301 * This packing function is run on idle threads. It checks to see if
7302 * the busiest CPU in this domain (core in the P7 case) has a higher
7303 * CPU number than the packing function is being run on. Here we are
7304 * assuming lower CPU number will be equivalent to lower a SMT thread
7307 * Return: 1 when packing is required and a task should be moved to
7308 * this CPU. The amount of the imbalance is returned in *imbalance.
7310 * @env: The load balancing environment.
7311 * @sds: Statistics of the sched_domain which is to be packed
7313 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7317 if (!(env->sd->flags & SD_ASYM_PACKING))
7323 busiest_cpu = group_first_cpu(sds->busiest);
7324 if (env->dst_cpu > busiest_cpu)
7327 env->imbalance = DIV_ROUND_CLOSEST(
7328 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7329 SCHED_CAPACITY_SCALE);
7335 * fix_small_imbalance - Calculate the minor imbalance that exists
7336 * amongst the groups of a sched_domain, during
7338 * @env: The load balancing environment.
7339 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7342 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7344 unsigned long tmp, capa_now = 0, capa_move = 0;
7345 unsigned int imbn = 2;
7346 unsigned long scaled_busy_load_per_task;
7347 struct sg_lb_stats *local, *busiest;
7349 local = &sds->local_stat;
7350 busiest = &sds->busiest_stat;
7352 if (!local->sum_nr_running)
7353 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7354 else if (busiest->load_per_task > local->load_per_task)
7357 scaled_busy_load_per_task =
7358 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7359 busiest->group_capacity;
7361 if (busiest->avg_load + scaled_busy_load_per_task >=
7362 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7363 env->imbalance = busiest->load_per_task;
7368 * OK, we don't have enough imbalance to justify moving tasks,
7369 * however we may be able to increase total CPU capacity used by
7373 capa_now += busiest->group_capacity *
7374 min(busiest->load_per_task, busiest->avg_load);
7375 capa_now += local->group_capacity *
7376 min(local->load_per_task, local->avg_load);
7377 capa_now /= SCHED_CAPACITY_SCALE;
7379 /* Amount of load we'd subtract */
7380 if (busiest->avg_load > scaled_busy_load_per_task) {
7381 capa_move += busiest->group_capacity *
7382 min(busiest->load_per_task,
7383 busiest->avg_load - scaled_busy_load_per_task);
7386 /* Amount of load we'd add */
7387 if (busiest->avg_load * busiest->group_capacity <
7388 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7389 tmp = (busiest->avg_load * busiest->group_capacity) /
7390 local->group_capacity;
7392 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7393 local->group_capacity;
7395 capa_move += local->group_capacity *
7396 min(local->load_per_task, local->avg_load + tmp);
7397 capa_move /= SCHED_CAPACITY_SCALE;
7399 /* Move if we gain throughput */
7400 if (capa_move > capa_now)
7401 env->imbalance = busiest->load_per_task;
7405 * calculate_imbalance - Calculate the amount of imbalance present within the
7406 * groups of a given sched_domain during load balance.
7407 * @env: load balance environment
7408 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7410 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7412 unsigned long max_pull, load_above_capacity = ~0UL;
7413 struct sg_lb_stats *local, *busiest;
7415 local = &sds->local_stat;
7416 busiest = &sds->busiest_stat;
7418 if (busiest->group_type == group_imbalanced) {
7420 * In the group_imb case we cannot rely on group-wide averages
7421 * to ensure cpu-load equilibrium, look at wider averages. XXX
7423 busiest->load_per_task =
7424 min(busiest->load_per_task, sds->avg_load);
7428 * In the presence of smp nice balancing, certain scenarios can have
7429 * max load less than avg load(as we skip the groups at or below
7430 * its cpu_capacity, while calculating max_load..)
7432 if (busiest->avg_load <= sds->avg_load ||
7433 local->avg_load >= sds->avg_load) {
7434 /* Misfitting tasks should be migrated in any case */
7435 if (busiest->group_type == group_misfit_task) {
7436 env->imbalance = busiest->group_misfit_task;
7441 * Busiest group is overloaded, local is not, use the spare
7442 * cycles to maximize throughput
7444 if (busiest->group_type == group_overloaded &&
7445 local->group_type <= group_misfit_task) {
7446 env->imbalance = busiest->load_per_task;
7451 return fix_small_imbalance(env, sds);
7455 * If there aren't any idle cpus, avoid creating some.
7457 if (busiest->group_type == group_overloaded &&
7458 local->group_type == group_overloaded) {
7459 load_above_capacity = busiest->sum_nr_running *
7461 if (load_above_capacity > busiest->group_capacity)
7462 load_above_capacity -= busiest->group_capacity;
7464 load_above_capacity = ~0UL;
7468 * We're trying to get all the cpus to the average_load, so we don't
7469 * want to push ourselves above the average load, nor do we wish to
7470 * reduce the max loaded cpu below the average load. At the same time,
7471 * we also don't want to reduce the group load below the group capacity
7472 * (so that we can implement power-savings policies etc). Thus we look
7473 * for the minimum possible imbalance.
7475 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7477 /* How much load to actually move to equalise the imbalance */
7478 env->imbalance = min(
7479 max_pull * busiest->group_capacity,
7480 (sds->avg_load - local->avg_load) * local->group_capacity
7481 ) / SCHED_CAPACITY_SCALE;
7483 /* Boost imbalance to allow misfit task to be balanced. */
7484 if (busiest->group_type == group_misfit_task)
7485 env->imbalance = max_t(long, env->imbalance,
7486 busiest->group_misfit_task);
7489 * if *imbalance is less than the average load per runnable task
7490 * there is no guarantee that any tasks will be moved so we'll have
7491 * a think about bumping its value to force at least one task to be
7494 if (env->imbalance < busiest->load_per_task)
7495 return fix_small_imbalance(env, sds);
7498 /******* find_busiest_group() helpers end here *********************/
7501 * find_busiest_group - Returns the busiest group within the sched_domain
7502 * if there is an imbalance. If there isn't an imbalance, and
7503 * the user has opted for power-savings, it returns a group whose
7504 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7505 * such a group exists.
7507 * Also calculates the amount of weighted load which should be moved
7508 * to restore balance.
7510 * @env: The load balancing environment.
7512 * Return: - The busiest group if imbalance exists.
7513 * - If no imbalance and user has opted for power-savings balance,
7514 * return the least loaded group whose CPUs can be
7515 * put to idle by rebalancing its tasks onto our group.
7517 static struct sched_group *find_busiest_group(struct lb_env *env)
7519 struct sg_lb_stats *local, *busiest;
7520 struct sd_lb_stats sds;
7522 init_sd_lb_stats(&sds);
7525 * Compute the various statistics relavent for load balancing at
7528 update_sd_lb_stats(env, &sds);
7530 if (energy_aware() && !env->dst_rq->rd->overutilized)
7533 local = &sds.local_stat;
7534 busiest = &sds.busiest_stat;
7536 /* ASYM feature bypasses nice load balance check */
7537 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7538 check_asym_packing(env, &sds))
7541 /* There is no busy sibling group to pull tasks from */
7542 if (!sds.busiest || busiest->sum_nr_running == 0)
7545 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7546 / sds.total_capacity;
7549 * If the busiest group is imbalanced the below checks don't
7550 * work because they assume all things are equal, which typically
7551 * isn't true due to cpus_allowed constraints and the like.
7553 if (busiest->group_type == group_imbalanced)
7556 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7557 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7558 busiest->group_no_capacity)
7561 /* Misfitting tasks should be dealt with regardless of the avg load */
7562 if (busiest->group_type == group_misfit_task) {
7567 * If the local group is busier than the selected busiest group
7568 * don't try and pull any tasks.
7570 if (local->avg_load >= busiest->avg_load)
7574 * Don't pull any tasks if this group is already above the domain
7577 if (local->avg_load >= sds.avg_load)
7580 if (env->idle == CPU_IDLE) {
7582 * This cpu is idle. If the busiest group is not overloaded
7583 * and there is no imbalance between this and busiest group
7584 * wrt idle cpus, it is balanced. The imbalance becomes
7585 * significant if the diff is greater than 1 otherwise we
7586 * might end up to just move the imbalance on another group
7588 if ((busiest->group_type != group_overloaded) &&
7589 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7590 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7594 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7595 * imbalance_pct to be conservative.
7597 if (100 * busiest->avg_load <=
7598 env->sd->imbalance_pct * local->avg_load)
7603 env->busiest_group_type = busiest->group_type;
7604 /* Looks like there is an imbalance. Compute it */
7605 calculate_imbalance(env, &sds);
7614 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7616 static struct rq *find_busiest_queue(struct lb_env *env,
7617 struct sched_group *group)
7619 struct rq *busiest = NULL, *rq;
7620 unsigned long busiest_load = 0, busiest_capacity = 1;
7623 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7624 unsigned long capacity, wl;
7628 rt = fbq_classify_rq(rq);
7631 * We classify groups/runqueues into three groups:
7632 * - regular: there are !numa tasks
7633 * - remote: there are numa tasks that run on the 'wrong' node
7634 * - all: there is no distinction
7636 * In order to avoid migrating ideally placed numa tasks,
7637 * ignore those when there's better options.
7639 * If we ignore the actual busiest queue to migrate another
7640 * task, the next balance pass can still reduce the busiest
7641 * queue by moving tasks around inside the node.
7643 * If we cannot move enough load due to this classification
7644 * the next pass will adjust the group classification and
7645 * allow migration of more tasks.
7647 * Both cases only affect the total convergence complexity.
7649 if (rt > env->fbq_type)
7652 capacity = capacity_of(i);
7654 wl = weighted_cpuload(i);
7657 * When comparing with imbalance, use weighted_cpuload()
7658 * which is not scaled with the cpu capacity.
7661 if (rq->nr_running == 1 && wl > env->imbalance &&
7662 !check_cpu_capacity(rq, env->sd) &&
7663 env->busiest_group_type != group_misfit_task)
7667 * For the load comparisons with the other cpu's, consider
7668 * the weighted_cpuload() scaled with the cpu capacity, so
7669 * that the load can be moved away from the cpu that is
7670 * potentially running at a lower capacity.
7672 * Thus we're looking for max(wl_i / capacity_i), crosswise
7673 * multiplication to rid ourselves of the division works out
7674 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7675 * our previous maximum.
7677 if (wl * busiest_capacity > busiest_load * capacity) {
7679 busiest_capacity = capacity;
7688 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7689 * so long as it is large enough.
7691 #define MAX_PINNED_INTERVAL 512
7693 /* Working cpumask for load_balance and load_balance_newidle. */
7694 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7696 static int need_active_balance(struct lb_env *env)
7698 struct sched_domain *sd = env->sd;
7700 if (env->idle == CPU_NEWLY_IDLE) {
7703 * ASYM_PACKING needs to force migrate tasks from busy but
7704 * higher numbered CPUs in order to pack all tasks in the
7705 * lowest numbered CPUs.
7707 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7712 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7713 * It's worth migrating the task if the src_cpu's capacity is reduced
7714 * because of other sched_class or IRQs if more capacity stays
7715 * available on dst_cpu.
7717 if ((env->idle != CPU_NOT_IDLE) &&
7718 (env->src_rq->cfs.h_nr_running == 1)) {
7719 if ((check_cpu_capacity(env->src_rq, sd)) &&
7720 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7724 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
7725 env->src_rq->cfs.h_nr_running == 1 &&
7726 cpu_overutilized(env->src_cpu) &&
7727 !cpu_overutilized(env->dst_cpu)) {
7731 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7734 static int active_load_balance_cpu_stop(void *data);
7736 static int should_we_balance(struct lb_env *env)
7738 struct sched_group *sg = env->sd->groups;
7739 struct cpumask *sg_cpus, *sg_mask;
7740 int cpu, balance_cpu = -1;
7743 * In the newly idle case, we will allow all the cpu's
7744 * to do the newly idle load balance.
7746 if (env->idle == CPU_NEWLY_IDLE)
7749 sg_cpus = sched_group_cpus(sg);
7750 sg_mask = sched_group_mask(sg);
7751 /* Try to find first idle cpu */
7752 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7753 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7760 if (balance_cpu == -1)
7761 balance_cpu = group_balance_cpu(sg);
7764 * First idle cpu or the first cpu(busiest) in this sched group
7765 * is eligible for doing load balancing at this and above domains.
7767 return balance_cpu == env->dst_cpu;
7771 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7772 * tasks if there is an imbalance.
7774 static int load_balance(int this_cpu, struct rq *this_rq,
7775 struct sched_domain *sd, enum cpu_idle_type idle,
7776 int *continue_balancing)
7778 int ld_moved, cur_ld_moved, active_balance = 0;
7779 struct sched_domain *sd_parent = sd->parent;
7780 struct sched_group *group;
7782 unsigned long flags;
7783 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7785 struct lb_env env = {
7787 .dst_cpu = this_cpu,
7789 .dst_grpmask = sched_group_cpus(sd->groups),
7791 .loop_break = sched_nr_migrate_break,
7794 .tasks = LIST_HEAD_INIT(env.tasks),
7798 * For NEWLY_IDLE load_balancing, we don't need to consider
7799 * other cpus in our group
7801 if (idle == CPU_NEWLY_IDLE)
7802 env.dst_grpmask = NULL;
7804 cpumask_copy(cpus, cpu_active_mask);
7806 schedstat_inc(sd, lb_count[idle]);
7809 if (!should_we_balance(&env)) {
7810 *continue_balancing = 0;
7814 group = find_busiest_group(&env);
7816 schedstat_inc(sd, lb_nobusyg[idle]);
7820 busiest = find_busiest_queue(&env, group);
7822 schedstat_inc(sd, lb_nobusyq[idle]);
7826 BUG_ON(busiest == env.dst_rq);
7828 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7830 env.src_cpu = busiest->cpu;
7831 env.src_rq = busiest;
7834 if (busiest->nr_running > 1) {
7836 * Attempt to move tasks. If find_busiest_group has found
7837 * an imbalance but busiest->nr_running <= 1, the group is
7838 * still unbalanced. ld_moved simply stays zero, so it is
7839 * correctly treated as an imbalance.
7841 env.flags |= LBF_ALL_PINNED;
7842 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7845 raw_spin_lock_irqsave(&busiest->lock, flags);
7848 * cur_ld_moved - load moved in current iteration
7849 * ld_moved - cumulative load moved across iterations
7851 cur_ld_moved = detach_tasks(&env);
7853 * We want to potentially lower env.src_cpu's OPP.
7856 update_capacity_of(env.src_cpu);
7859 * We've detached some tasks from busiest_rq. Every
7860 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7861 * unlock busiest->lock, and we are able to be sure
7862 * that nobody can manipulate the tasks in parallel.
7863 * See task_rq_lock() family for the details.
7866 raw_spin_unlock(&busiest->lock);
7870 ld_moved += cur_ld_moved;
7873 local_irq_restore(flags);
7875 if (env.flags & LBF_NEED_BREAK) {
7876 env.flags &= ~LBF_NEED_BREAK;
7881 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7882 * us and move them to an alternate dst_cpu in our sched_group
7883 * where they can run. The upper limit on how many times we
7884 * iterate on same src_cpu is dependent on number of cpus in our
7887 * This changes load balance semantics a bit on who can move
7888 * load to a given_cpu. In addition to the given_cpu itself
7889 * (or a ilb_cpu acting on its behalf where given_cpu is
7890 * nohz-idle), we now have balance_cpu in a position to move
7891 * load to given_cpu. In rare situations, this may cause
7892 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7893 * _independently_ and at _same_ time to move some load to
7894 * given_cpu) causing exceess load to be moved to given_cpu.
7895 * This however should not happen so much in practice and
7896 * moreover subsequent load balance cycles should correct the
7897 * excess load moved.
7899 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7901 /* Prevent to re-select dst_cpu via env's cpus */
7902 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7904 env.dst_rq = cpu_rq(env.new_dst_cpu);
7905 env.dst_cpu = env.new_dst_cpu;
7906 env.flags &= ~LBF_DST_PINNED;
7908 env.loop_break = sched_nr_migrate_break;
7911 * Go back to "more_balance" rather than "redo" since we
7912 * need to continue with same src_cpu.
7918 * We failed to reach balance because of affinity.
7921 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7923 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7924 *group_imbalance = 1;
7927 /* All tasks on this runqueue were pinned by CPU affinity */
7928 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7929 cpumask_clear_cpu(cpu_of(busiest), cpus);
7930 if (!cpumask_empty(cpus)) {
7932 env.loop_break = sched_nr_migrate_break;
7935 goto out_all_pinned;
7940 schedstat_inc(sd, lb_failed[idle]);
7942 * Increment the failure counter only on periodic balance.
7943 * We do not want newidle balance, which can be very
7944 * frequent, pollute the failure counter causing
7945 * excessive cache_hot migrations and active balances.
7947 if (idle != CPU_NEWLY_IDLE)
7948 if (env.src_grp_nr_running > 1)
7949 sd->nr_balance_failed++;
7951 if (need_active_balance(&env)) {
7952 raw_spin_lock_irqsave(&busiest->lock, flags);
7954 /* don't kick the active_load_balance_cpu_stop,
7955 * if the curr task on busiest cpu can't be
7958 if (!cpumask_test_cpu(this_cpu,
7959 tsk_cpus_allowed(busiest->curr))) {
7960 raw_spin_unlock_irqrestore(&busiest->lock,
7962 env.flags |= LBF_ALL_PINNED;
7963 goto out_one_pinned;
7967 * ->active_balance synchronizes accesses to
7968 * ->active_balance_work. Once set, it's cleared
7969 * only after active load balance is finished.
7971 if (!busiest->active_balance) {
7972 busiest->active_balance = 1;
7973 busiest->push_cpu = this_cpu;
7976 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7978 if (active_balance) {
7979 stop_one_cpu_nowait(cpu_of(busiest),
7980 active_load_balance_cpu_stop, busiest,
7981 &busiest->active_balance_work);
7985 * We've kicked active balancing, reset the failure
7988 sd->nr_balance_failed = sd->cache_nice_tries+1;
7991 sd->nr_balance_failed = 0;
7993 if (likely(!active_balance)) {
7994 /* We were unbalanced, so reset the balancing interval */
7995 sd->balance_interval = sd->min_interval;
7998 * If we've begun active balancing, start to back off. This
7999 * case may not be covered by the all_pinned logic if there
8000 * is only 1 task on the busy runqueue (because we don't call
8003 if (sd->balance_interval < sd->max_interval)
8004 sd->balance_interval *= 2;
8011 * We reach balance although we may have faced some affinity
8012 * constraints. Clear the imbalance flag if it was set.
8015 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8017 if (*group_imbalance)
8018 *group_imbalance = 0;
8023 * We reach balance because all tasks are pinned at this level so
8024 * we can't migrate them. Let the imbalance flag set so parent level
8025 * can try to migrate them.
8027 schedstat_inc(sd, lb_balanced[idle]);
8029 sd->nr_balance_failed = 0;
8032 /* tune up the balancing interval */
8033 if (((env.flags & LBF_ALL_PINNED) &&
8034 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8035 (sd->balance_interval < sd->max_interval))
8036 sd->balance_interval *= 2;
8043 static inline unsigned long
8044 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8046 unsigned long interval = sd->balance_interval;
8049 interval *= sd->busy_factor;
8051 /* scale ms to jiffies */
8052 interval = msecs_to_jiffies(interval);
8053 interval = clamp(interval, 1UL, max_load_balance_interval);
8059 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8061 unsigned long interval, next;
8063 interval = get_sd_balance_interval(sd, cpu_busy);
8064 next = sd->last_balance + interval;
8066 if (time_after(*next_balance, next))
8067 *next_balance = next;
8071 * idle_balance is called by schedule() if this_cpu is about to become
8072 * idle. Attempts to pull tasks from other CPUs.
8074 static int idle_balance(struct rq *this_rq)
8076 unsigned long next_balance = jiffies + HZ;
8077 int this_cpu = this_rq->cpu;
8078 struct sched_domain *sd;
8079 int pulled_task = 0;
8082 idle_enter_fair(this_rq);
8085 * We must set idle_stamp _before_ calling idle_balance(), such that we
8086 * measure the duration of idle_balance() as idle time.
8088 this_rq->idle_stamp = rq_clock(this_rq);
8090 if (!energy_aware() &&
8091 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8092 !this_rq->rd->overload)) {
8094 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8096 update_next_balance(sd, 0, &next_balance);
8102 raw_spin_unlock(&this_rq->lock);
8104 update_blocked_averages(this_cpu);
8106 for_each_domain(this_cpu, sd) {
8107 int continue_balancing = 1;
8108 u64 t0, domain_cost;
8110 if (!(sd->flags & SD_LOAD_BALANCE))
8113 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8114 update_next_balance(sd, 0, &next_balance);
8118 if (sd->flags & SD_BALANCE_NEWIDLE) {
8119 t0 = sched_clock_cpu(this_cpu);
8121 pulled_task = load_balance(this_cpu, this_rq,
8123 &continue_balancing);
8125 domain_cost = sched_clock_cpu(this_cpu) - t0;
8126 if (domain_cost > sd->max_newidle_lb_cost)
8127 sd->max_newidle_lb_cost = domain_cost;
8129 curr_cost += domain_cost;
8132 update_next_balance(sd, 0, &next_balance);
8135 * Stop searching for tasks to pull if there are
8136 * now runnable tasks on this rq.
8138 if (pulled_task || this_rq->nr_running > 0)
8143 raw_spin_lock(&this_rq->lock);
8145 if (curr_cost > this_rq->max_idle_balance_cost)
8146 this_rq->max_idle_balance_cost = curr_cost;
8149 * While browsing the domains, we released the rq lock, a task could
8150 * have been enqueued in the meantime. Since we're not going idle,
8151 * pretend we pulled a task.
8153 if (this_rq->cfs.h_nr_running && !pulled_task)
8157 /* Move the next balance forward */
8158 if (time_after(this_rq->next_balance, next_balance))
8159 this_rq->next_balance = next_balance;
8161 /* Is there a task of a high priority class? */
8162 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8166 idle_exit_fair(this_rq);
8167 this_rq->idle_stamp = 0;
8174 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8175 * running tasks off the busiest CPU onto idle CPUs. It requires at
8176 * least 1 task to be running on each physical CPU where possible, and
8177 * avoids physical / logical imbalances.
8179 static int active_load_balance_cpu_stop(void *data)
8181 struct rq *busiest_rq = data;
8182 int busiest_cpu = cpu_of(busiest_rq);
8183 int target_cpu = busiest_rq->push_cpu;
8184 struct rq *target_rq = cpu_rq(target_cpu);
8185 struct sched_domain *sd;
8186 struct task_struct *p = NULL;
8188 raw_spin_lock_irq(&busiest_rq->lock);
8190 /* make sure the requested cpu hasn't gone down in the meantime */
8191 if (unlikely(busiest_cpu != smp_processor_id() ||
8192 !busiest_rq->active_balance))
8195 /* Is there any task to move? */
8196 if (busiest_rq->nr_running <= 1)
8200 * This condition is "impossible", if it occurs
8201 * we need to fix it. Originally reported by
8202 * Bjorn Helgaas on a 128-cpu setup.
8204 BUG_ON(busiest_rq == target_rq);
8206 /* Search for an sd spanning us and the target CPU. */
8208 for_each_domain(target_cpu, sd) {
8209 if ((sd->flags & SD_LOAD_BALANCE) &&
8210 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8215 struct lb_env env = {
8217 .dst_cpu = target_cpu,
8218 .dst_rq = target_rq,
8219 .src_cpu = busiest_rq->cpu,
8220 .src_rq = busiest_rq,
8224 schedstat_inc(sd, alb_count);
8226 p = detach_one_task(&env);
8228 schedstat_inc(sd, alb_pushed);
8230 * We want to potentially lower env.src_cpu's OPP.
8232 update_capacity_of(env.src_cpu);
8235 schedstat_inc(sd, alb_failed);
8239 busiest_rq->active_balance = 0;
8240 raw_spin_unlock(&busiest_rq->lock);
8243 attach_one_task(target_rq, p);
8250 static inline int on_null_domain(struct rq *rq)
8252 return unlikely(!rcu_dereference_sched(rq->sd));
8255 #ifdef CONFIG_NO_HZ_COMMON
8257 * idle load balancing details
8258 * - When one of the busy CPUs notice that there may be an idle rebalancing
8259 * needed, they will kick the idle load balancer, which then does idle
8260 * load balancing for all the idle CPUs.
8263 cpumask_var_t idle_cpus_mask;
8265 unsigned long next_balance; /* in jiffy units */
8266 } nohz ____cacheline_aligned;
8268 static inline int find_new_ilb(void)
8270 int ilb = cpumask_first(nohz.idle_cpus_mask);
8272 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8279 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8280 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8281 * CPU (if there is one).
8283 static void nohz_balancer_kick(void)
8287 nohz.next_balance++;
8289 ilb_cpu = find_new_ilb();
8291 if (ilb_cpu >= nr_cpu_ids)
8294 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8297 * Use smp_send_reschedule() instead of resched_cpu().
8298 * This way we generate a sched IPI on the target cpu which
8299 * is idle. And the softirq performing nohz idle load balance
8300 * will be run before returning from the IPI.
8302 smp_send_reschedule(ilb_cpu);
8306 static inline void nohz_balance_exit_idle(int cpu)
8308 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8310 * Completely isolated CPUs don't ever set, so we must test.
8312 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8313 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8314 atomic_dec(&nohz.nr_cpus);
8316 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8320 static inline void set_cpu_sd_state_busy(void)
8322 struct sched_domain *sd;
8323 int cpu = smp_processor_id();
8326 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8328 if (!sd || !sd->nohz_idle)
8332 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8337 void set_cpu_sd_state_idle(void)
8339 struct sched_domain *sd;
8340 int cpu = smp_processor_id();
8343 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8345 if (!sd || sd->nohz_idle)
8349 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8355 * This routine will record that the cpu is going idle with tick stopped.
8356 * This info will be used in performing idle load balancing in the future.
8358 void nohz_balance_enter_idle(int cpu)
8361 * If this cpu is going down, then nothing needs to be done.
8363 if (!cpu_active(cpu))
8366 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8370 * If we're a completely isolated CPU, we don't play.
8372 if (on_null_domain(cpu_rq(cpu)))
8375 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8376 atomic_inc(&nohz.nr_cpus);
8377 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8380 static int sched_ilb_notifier(struct notifier_block *nfb,
8381 unsigned long action, void *hcpu)
8383 switch (action & ~CPU_TASKS_FROZEN) {
8385 nohz_balance_exit_idle(smp_processor_id());
8393 static DEFINE_SPINLOCK(balancing);
8396 * Scale the max load_balance interval with the number of CPUs in the system.
8397 * This trades load-balance latency on larger machines for less cross talk.
8399 void update_max_interval(void)
8401 max_load_balance_interval = HZ*num_online_cpus()/10;
8405 * It checks each scheduling domain to see if it is due to be balanced,
8406 * and initiates a balancing operation if so.
8408 * Balancing parameters are set up in init_sched_domains.
8410 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8412 int continue_balancing = 1;
8414 unsigned long interval;
8415 struct sched_domain *sd;
8416 /* Earliest time when we have to do rebalance again */
8417 unsigned long next_balance = jiffies + 60*HZ;
8418 int update_next_balance = 0;
8419 int need_serialize, need_decay = 0;
8422 update_blocked_averages(cpu);
8425 for_each_domain(cpu, sd) {
8427 * Decay the newidle max times here because this is a regular
8428 * visit to all the domains. Decay ~1% per second.
8430 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8431 sd->max_newidle_lb_cost =
8432 (sd->max_newidle_lb_cost * 253) / 256;
8433 sd->next_decay_max_lb_cost = jiffies + HZ;
8436 max_cost += sd->max_newidle_lb_cost;
8438 if (!(sd->flags & SD_LOAD_BALANCE))
8442 * Stop the load balance at this level. There is another
8443 * CPU in our sched group which is doing load balancing more
8446 if (!continue_balancing) {
8452 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8454 need_serialize = sd->flags & SD_SERIALIZE;
8455 if (need_serialize) {
8456 if (!spin_trylock(&balancing))
8460 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8461 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8463 * The LBF_DST_PINNED logic could have changed
8464 * env->dst_cpu, so we can't know our idle
8465 * state even if we migrated tasks. Update it.
8467 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8469 sd->last_balance = jiffies;
8470 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8473 spin_unlock(&balancing);
8475 if (time_after(next_balance, sd->last_balance + interval)) {
8476 next_balance = sd->last_balance + interval;
8477 update_next_balance = 1;
8482 * Ensure the rq-wide value also decays but keep it at a
8483 * reasonable floor to avoid funnies with rq->avg_idle.
8485 rq->max_idle_balance_cost =
8486 max((u64)sysctl_sched_migration_cost, max_cost);
8491 * next_balance will be updated only when there is a need.
8492 * When the cpu is attached to null domain for ex, it will not be
8495 if (likely(update_next_balance)) {
8496 rq->next_balance = next_balance;
8498 #ifdef CONFIG_NO_HZ_COMMON
8500 * If this CPU has been elected to perform the nohz idle
8501 * balance. Other idle CPUs have already rebalanced with
8502 * nohz_idle_balance() and nohz.next_balance has been
8503 * updated accordingly. This CPU is now running the idle load
8504 * balance for itself and we need to update the
8505 * nohz.next_balance accordingly.
8507 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8508 nohz.next_balance = rq->next_balance;
8513 #ifdef CONFIG_NO_HZ_COMMON
8515 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8516 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8518 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8520 int this_cpu = this_rq->cpu;
8523 /* Earliest time when we have to do rebalance again */
8524 unsigned long next_balance = jiffies + 60*HZ;
8525 int update_next_balance = 0;
8527 if (idle != CPU_IDLE ||
8528 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8531 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8532 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8536 * If this cpu gets work to do, stop the load balancing
8537 * work being done for other cpus. Next load
8538 * balancing owner will pick it up.
8543 rq = cpu_rq(balance_cpu);
8546 * If time for next balance is due,
8549 if (time_after_eq(jiffies, rq->next_balance)) {
8550 raw_spin_lock_irq(&rq->lock);
8551 update_rq_clock(rq);
8552 update_idle_cpu_load(rq);
8553 raw_spin_unlock_irq(&rq->lock);
8554 rebalance_domains(rq, CPU_IDLE);
8557 if (time_after(next_balance, rq->next_balance)) {
8558 next_balance = rq->next_balance;
8559 update_next_balance = 1;
8564 * next_balance will be updated only when there is a need.
8565 * When the CPU is attached to null domain for ex, it will not be
8568 if (likely(update_next_balance))
8569 nohz.next_balance = next_balance;
8571 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8575 * Current heuristic for kicking the idle load balancer in the presence
8576 * of an idle cpu in the system.
8577 * - This rq has more than one task.
8578 * - This rq has at least one CFS task and the capacity of the CPU is
8579 * significantly reduced because of RT tasks or IRQs.
8580 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8581 * multiple busy cpu.
8582 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8583 * domain span are idle.
8585 static inline bool nohz_kick_needed(struct rq *rq)
8587 unsigned long now = jiffies;
8588 struct sched_domain *sd;
8589 struct sched_group_capacity *sgc;
8590 int nr_busy, cpu = rq->cpu;
8593 if (unlikely(rq->idle_balance))
8597 * We may be recently in ticked or tickless idle mode. At the first
8598 * busy tick after returning from idle, we will update the busy stats.
8600 set_cpu_sd_state_busy();
8601 nohz_balance_exit_idle(cpu);
8604 * None are in tickless mode and hence no need for NOHZ idle load
8607 if (likely(!atomic_read(&nohz.nr_cpus)))
8610 if (time_before(now, nohz.next_balance))
8613 if (rq->nr_running >= 2 &&
8614 (!energy_aware() || cpu_overutilized(cpu)))
8618 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8619 if (sd && !energy_aware()) {
8620 sgc = sd->groups->sgc;
8621 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8630 sd = rcu_dereference(rq->sd);
8632 if ((rq->cfs.h_nr_running >= 1) &&
8633 check_cpu_capacity(rq, sd)) {
8639 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8640 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8641 sched_domain_span(sd)) < cpu)) {
8651 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8655 * run_rebalance_domains is triggered when needed from the scheduler tick.
8656 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8658 static void run_rebalance_domains(struct softirq_action *h)
8660 struct rq *this_rq = this_rq();
8661 enum cpu_idle_type idle = this_rq->idle_balance ?
8662 CPU_IDLE : CPU_NOT_IDLE;
8665 * If this cpu has a pending nohz_balance_kick, then do the
8666 * balancing on behalf of the other idle cpus whose ticks are
8667 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8668 * give the idle cpus a chance to load balance. Else we may
8669 * load balance only within the local sched_domain hierarchy
8670 * and abort nohz_idle_balance altogether if we pull some load.
8672 nohz_idle_balance(this_rq, idle);
8673 rebalance_domains(this_rq, idle);
8677 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8679 void trigger_load_balance(struct rq *rq)
8681 /* Don't need to rebalance while attached to NULL domain */
8682 if (unlikely(on_null_domain(rq)))
8685 if (time_after_eq(jiffies, rq->next_balance))
8686 raise_softirq(SCHED_SOFTIRQ);
8687 #ifdef CONFIG_NO_HZ_COMMON
8688 if (nohz_kick_needed(rq))
8689 nohz_balancer_kick();
8693 static void rq_online_fair(struct rq *rq)
8697 update_runtime_enabled(rq);
8700 static void rq_offline_fair(struct rq *rq)
8704 /* Ensure any throttled groups are reachable by pick_next_task */
8705 unthrottle_offline_cfs_rqs(rq);
8708 #endif /* CONFIG_SMP */
8711 * scheduler tick hitting a task of our scheduling class:
8713 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8715 struct cfs_rq *cfs_rq;
8716 struct sched_entity *se = &curr->se;
8718 for_each_sched_entity(se) {
8719 cfs_rq = cfs_rq_of(se);
8720 entity_tick(cfs_rq, se, queued);
8723 if (static_branch_unlikely(&sched_numa_balancing))
8724 task_tick_numa(rq, curr);
8726 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr)))
8727 rq->rd->overutilized = true;
8729 rq->misfit_task = !task_fits_max(curr, rq->cpu);
8733 * called on fork with the child task as argument from the parent's context
8734 * - child not yet on the tasklist
8735 * - preemption disabled
8737 static void task_fork_fair(struct task_struct *p)
8739 struct cfs_rq *cfs_rq;
8740 struct sched_entity *se = &p->se, *curr;
8741 int this_cpu = smp_processor_id();
8742 struct rq *rq = this_rq();
8743 unsigned long flags;
8745 raw_spin_lock_irqsave(&rq->lock, flags);
8747 update_rq_clock(rq);
8749 cfs_rq = task_cfs_rq(current);
8750 curr = cfs_rq->curr;
8753 * Not only the cpu but also the task_group of the parent might have
8754 * been changed after parent->se.parent,cfs_rq were copied to
8755 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8756 * of child point to valid ones.
8759 __set_task_cpu(p, this_cpu);
8762 update_curr(cfs_rq);
8765 se->vruntime = curr->vruntime;
8766 place_entity(cfs_rq, se, 1);
8768 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8770 * Upon rescheduling, sched_class::put_prev_task() will place
8771 * 'current' within the tree based on its new key value.
8773 swap(curr->vruntime, se->vruntime);
8777 se->vruntime -= cfs_rq->min_vruntime;
8779 raw_spin_unlock_irqrestore(&rq->lock, flags);
8783 * Priority of the task has changed. Check to see if we preempt
8787 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8789 if (!task_on_rq_queued(p))
8793 * Reschedule if we are currently running on this runqueue and
8794 * our priority decreased, or if we are not currently running on
8795 * this runqueue and our priority is higher than the current's
8797 if (rq->curr == p) {
8798 if (p->prio > oldprio)
8801 check_preempt_curr(rq, p, 0);
8804 static inline bool vruntime_normalized(struct task_struct *p)
8806 struct sched_entity *se = &p->se;
8809 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8810 * the dequeue_entity(.flags=0) will already have normalized the
8817 * When !on_rq, vruntime of the task has usually NOT been normalized.
8818 * But there are some cases where it has already been normalized:
8820 * - A forked child which is waiting for being woken up by
8821 * wake_up_new_task().
8822 * - A task which has been woken up by try_to_wake_up() and
8823 * waiting for actually being woken up by sched_ttwu_pending().
8825 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8831 static void detach_task_cfs_rq(struct task_struct *p)
8833 struct sched_entity *se = &p->se;
8834 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8836 if (!vruntime_normalized(p)) {
8838 * Fix up our vruntime so that the current sleep doesn't
8839 * cause 'unlimited' sleep bonus.
8841 place_entity(cfs_rq, se, 0);
8842 se->vruntime -= cfs_rq->min_vruntime;
8845 /* Catch up with the cfs_rq and remove our load when we leave */
8846 detach_entity_load_avg(cfs_rq, se);
8849 static void attach_task_cfs_rq(struct task_struct *p)
8851 struct sched_entity *se = &p->se;
8852 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8854 #ifdef CONFIG_FAIR_GROUP_SCHED
8856 * Since the real-depth could have been changed (only FAIR
8857 * class maintain depth value), reset depth properly.
8859 se->depth = se->parent ? se->parent->depth + 1 : 0;
8862 /* Synchronize task with its cfs_rq */
8863 attach_entity_load_avg(cfs_rq, se);
8865 if (!vruntime_normalized(p))
8866 se->vruntime += cfs_rq->min_vruntime;
8869 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8871 detach_task_cfs_rq(p);
8874 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8876 attach_task_cfs_rq(p);
8878 if (task_on_rq_queued(p)) {
8880 * We were most likely switched from sched_rt, so
8881 * kick off the schedule if running, otherwise just see
8882 * if we can still preempt the current task.
8887 check_preempt_curr(rq, p, 0);
8891 /* Account for a task changing its policy or group.
8893 * This routine is mostly called to set cfs_rq->curr field when a task
8894 * migrates between groups/classes.
8896 static void set_curr_task_fair(struct rq *rq)
8898 struct sched_entity *se = &rq->curr->se;
8900 for_each_sched_entity(se) {
8901 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8903 set_next_entity(cfs_rq, se);
8904 /* ensure bandwidth has been allocated on our new cfs_rq */
8905 account_cfs_rq_runtime(cfs_rq, 0);
8909 void init_cfs_rq(struct cfs_rq *cfs_rq)
8911 cfs_rq->tasks_timeline = RB_ROOT;
8912 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8913 #ifndef CONFIG_64BIT
8914 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8917 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8918 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8922 #ifdef CONFIG_FAIR_GROUP_SCHED
8923 static void task_move_group_fair(struct task_struct *p)
8925 detach_task_cfs_rq(p);
8926 set_task_rq(p, task_cpu(p));
8929 /* Tell se's cfs_rq has been changed -- migrated */
8930 p->se.avg.last_update_time = 0;
8932 attach_task_cfs_rq(p);
8935 void free_fair_sched_group(struct task_group *tg)
8939 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8941 for_each_possible_cpu(i) {
8943 kfree(tg->cfs_rq[i]);
8946 remove_entity_load_avg(tg->se[i]);
8955 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8957 struct cfs_rq *cfs_rq;
8958 struct sched_entity *se;
8961 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8964 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8968 tg->shares = NICE_0_LOAD;
8970 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8972 for_each_possible_cpu(i) {
8973 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8974 GFP_KERNEL, cpu_to_node(i));
8978 se = kzalloc_node(sizeof(struct sched_entity),
8979 GFP_KERNEL, cpu_to_node(i));
8983 init_cfs_rq(cfs_rq);
8984 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8985 init_entity_runnable_average(se);
8996 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8998 struct rq *rq = cpu_rq(cpu);
8999 unsigned long flags;
9002 * Only empty task groups can be destroyed; so we can speculatively
9003 * check on_list without danger of it being re-added.
9005 if (!tg->cfs_rq[cpu]->on_list)
9008 raw_spin_lock_irqsave(&rq->lock, flags);
9009 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9010 raw_spin_unlock_irqrestore(&rq->lock, flags);
9013 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9014 struct sched_entity *se, int cpu,
9015 struct sched_entity *parent)
9017 struct rq *rq = cpu_rq(cpu);
9021 init_cfs_rq_runtime(cfs_rq);
9023 tg->cfs_rq[cpu] = cfs_rq;
9026 /* se could be NULL for root_task_group */
9031 se->cfs_rq = &rq->cfs;
9034 se->cfs_rq = parent->my_q;
9035 se->depth = parent->depth + 1;
9039 /* guarantee group entities always have weight */
9040 update_load_set(&se->load, NICE_0_LOAD);
9041 se->parent = parent;
9044 static DEFINE_MUTEX(shares_mutex);
9046 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9049 unsigned long flags;
9052 * We can't change the weight of the root cgroup.
9057 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9059 mutex_lock(&shares_mutex);
9060 if (tg->shares == shares)
9063 tg->shares = shares;
9064 for_each_possible_cpu(i) {
9065 struct rq *rq = cpu_rq(i);
9066 struct sched_entity *se;
9069 /* Propagate contribution to hierarchy */
9070 raw_spin_lock_irqsave(&rq->lock, flags);
9072 /* Possible calls to update_curr() need rq clock */
9073 update_rq_clock(rq);
9074 for_each_sched_entity(se)
9075 update_cfs_shares(group_cfs_rq(se));
9076 raw_spin_unlock_irqrestore(&rq->lock, flags);
9080 mutex_unlock(&shares_mutex);
9083 #else /* CONFIG_FAIR_GROUP_SCHED */
9085 void free_fair_sched_group(struct task_group *tg) { }
9087 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9092 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9094 #endif /* CONFIG_FAIR_GROUP_SCHED */
9097 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9099 struct sched_entity *se = &task->se;
9100 unsigned int rr_interval = 0;
9103 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9106 if (rq->cfs.load.weight)
9107 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9113 * All the scheduling class methods:
9115 const struct sched_class fair_sched_class = {
9116 .next = &idle_sched_class,
9117 .enqueue_task = enqueue_task_fair,
9118 .dequeue_task = dequeue_task_fair,
9119 .yield_task = yield_task_fair,
9120 .yield_to_task = yield_to_task_fair,
9122 .check_preempt_curr = check_preempt_wakeup,
9124 .pick_next_task = pick_next_task_fair,
9125 .put_prev_task = put_prev_task_fair,
9128 .select_task_rq = select_task_rq_fair,
9129 .migrate_task_rq = migrate_task_rq_fair,
9131 .rq_online = rq_online_fair,
9132 .rq_offline = rq_offline_fair,
9134 .task_waking = task_waking_fair,
9135 .task_dead = task_dead_fair,
9136 .set_cpus_allowed = set_cpus_allowed_common,
9139 .set_curr_task = set_curr_task_fair,
9140 .task_tick = task_tick_fair,
9141 .task_fork = task_fork_fair,
9143 .prio_changed = prio_changed_fair,
9144 .switched_from = switched_from_fair,
9145 .switched_to = switched_to_fair,
9147 .get_rr_interval = get_rr_interval_fair,
9149 .update_curr = update_curr_fair,
9151 #ifdef CONFIG_FAIR_GROUP_SCHED
9152 .task_move_group = task_move_group_fair,
9156 #ifdef CONFIG_SCHED_DEBUG
9157 void print_cfs_stats(struct seq_file *m, int cpu)
9159 struct cfs_rq *cfs_rq;
9162 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9163 print_cfs_rq(m, cpu, cfs_rq);
9167 #ifdef CONFIG_NUMA_BALANCING
9168 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9171 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9173 for_each_online_node(node) {
9174 if (p->numa_faults) {
9175 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9176 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9178 if (p->numa_group) {
9179 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9180 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9182 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9185 #endif /* CONFIG_NUMA_BALANCING */
9186 #endif /* CONFIG_SCHED_DEBUG */
9188 __init void init_sched_fair_class(void)
9191 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9193 #ifdef CONFIG_NO_HZ_COMMON
9194 nohz.next_balance = jiffies;
9195 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9196 cpu_notifier(sched_ilb_notifier, 0);