2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
40 * Targeted preemption latency for CPU-bound tasks:
41 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
43 * NOTE: this latency value is not the same as the concept of
44 * 'timeslice length' - timeslices in CFS are of variable length
45 * and have no persistent notion like in traditional, time-slice
46 * based scheduling concepts.
48 * (to see the precise effective timeslice length of your workload,
49 * run vmstat and monitor the context-switches (cs) field)
51 unsigned int sysctl_sched_latency = 6000000ULL;
52 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
55 * The initial- and re-scaling of tunables is configurable
56 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
59 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
60 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
61 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
63 enum sched_tunable_scaling sysctl_sched_tunable_scaling
64 = SCHED_TUNABLESCALING_LOG;
67 * Minimal preemption granularity for CPU-bound tasks:
68 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
70 unsigned int sysctl_sched_min_granularity = 750000ULL;
71 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
74 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
76 static unsigned int sched_nr_latency = 8;
79 * After fork, child runs first. If set to 0 (default) then
80 * parent will (try to) run first.
82 unsigned int sysctl_sched_child_runs_first __read_mostly;
85 * SCHED_OTHER wake-up granularity.
86 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
88 * This option delays the preemption effects of decoupled workloads
89 * and reduces their over-scheduling. Synchronous workloads will still
90 * have immediate wakeup/sleep latencies.
92 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
93 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
95 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
98 * The exponential sliding window over which load is averaged for shares
102 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
104 #ifdef CONFIG_CFS_BANDWIDTH
106 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
107 * each time a cfs_rq requests quota.
109 * Note: in the case that the slice exceeds the runtime remaining (either due
110 * to consumption or the quota being specified to be smaller than the slice)
111 * we will always only issue the remaining available time.
113 * default: 5 msec, units: microseconds
115 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
118 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
124 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
130 static inline void update_load_set(struct load_weight *lw, unsigned long w)
137 * Increase the granularity value when there are more CPUs,
138 * because with more CPUs the 'effective latency' as visible
139 * to users decreases. But the relationship is not linear,
140 * so pick a second-best guess by going with the log2 of the
143 * This idea comes from the SD scheduler of Con Kolivas:
145 static unsigned int get_update_sysctl_factor(void)
147 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
150 switch (sysctl_sched_tunable_scaling) {
151 case SCHED_TUNABLESCALING_NONE:
154 case SCHED_TUNABLESCALING_LINEAR:
157 case SCHED_TUNABLESCALING_LOG:
159 factor = 1 + ilog2(cpus);
166 static void update_sysctl(void)
168 unsigned int factor = get_update_sysctl_factor();
170 #define SET_SYSCTL(name) \
171 (sysctl_##name = (factor) * normalized_sysctl_##name)
172 SET_SYSCTL(sched_min_granularity);
173 SET_SYSCTL(sched_latency);
174 SET_SYSCTL(sched_wakeup_granularity);
178 void sched_init_granularity(void)
183 #define WMULT_CONST (~0U)
184 #define WMULT_SHIFT 32
186 static void __update_inv_weight(struct load_weight *lw)
190 if (likely(lw->inv_weight))
193 w = scale_load_down(lw->weight);
195 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
197 else if (unlikely(!w))
198 lw->inv_weight = WMULT_CONST;
200 lw->inv_weight = WMULT_CONST / w;
204 * delta_exec * weight / lw.weight
206 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
208 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
209 * we're guaranteed shift stays positive because inv_weight is guaranteed to
210 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
212 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
213 * weight/lw.weight <= 1, and therefore our shift will also be positive.
215 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
217 u64 fact = scale_load_down(weight);
218 int shift = WMULT_SHIFT;
220 __update_inv_weight(lw);
222 if (unlikely(fact >> 32)) {
229 /* hint to use a 32x32->64 mul */
230 fact = (u64)(u32)fact * lw->inv_weight;
237 return mul_u64_u32_shr(delta_exec, fact, shift);
241 const struct sched_class fair_sched_class;
243 /**************************************************************
244 * CFS operations on generic schedulable entities:
247 #ifdef CONFIG_FAIR_GROUP_SCHED
249 /* cpu runqueue to which this cfs_rq is attached */
250 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
255 /* An entity is a task if it doesn't "own" a runqueue */
256 #define entity_is_task(se) (!se->my_q)
258 static inline struct task_struct *task_of(struct sched_entity *se)
260 #ifdef CONFIG_SCHED_DEBUG
261 WARN_ON_ONCE(!entity_is_task(se));
263 return container_of(se, struct task_struct, se);
266 /* Walk up scheduling entities hierarchy */
267 #define for_each_sched_entity(se) \
268 for (; se; se = se->parent)
270 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
275 /* runqueue on which this entity is (to be) queued */
276 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
281 /* runqueue "owned" by this group */
282 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
287 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
289 if (!cfs_rq->on_list) {
291 * Ensure we either appear before our parent (if already
292 * enqueued) or force our parent to appear after us when it is
293 * enqueued. The fact that we always enqueue bottom-up
294 * reduces this to two cases.
296 if (cfs_rq->tg->parent &&
297 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
298 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
299 &rq_of(cfs_rq)->leaf_cfs_rq_list);
301 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
302 &rq_of(cfs_rq)->leaf_cfs_rq_list);
309 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
311 if (cfs_rq->on_list) {
312 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
317 /* Iterate thr' all leaf cfs_rq's on a runqueue */
318 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
319 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
321 /* Do the two (enqueued) entities belong to the same group ? */
322 static inline struct cfs_rq *
323 is_same_group(struct sched_entity *se, struct sched_entity *pse)
325 if (se->cfs_rq == pse->cfs_rq)
331 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
339 int se_depth, pse_depth;
342 * preemption test can be made between sibling entities who are in the
343 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
344 * both tasks until we find their ancestors who are siblings of common
348 /* First walk up until both entities are at same depth */
349 se_depth = (*se)->depth;
350 pse_depth = (*pse)->depth;
352 while (se_depth > pse_depth) {
354 *se = parent_entity(*se);
357 while (pse_depth > se_depth) {
359 *pse = parent_entity(*pse);
362 while (!is_same_group(*se, *pse)) {
363 *se = parent_entity(*se);
364 *pse = parent_entity(*pse);
368 #else /* !CONFIG_FAIR_GROUP_SCHED */
370 static inline struct task_struct *task_of(struct sched_entity *se)
372 return container_of(se, struct task_struct, se);
375 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
377 return container_of(cfs_rq, struct rq, cfs);
380 #define entity_is_task(se) 1
382 #define for_each_sched_entity(se) \
383 for (; se; se = NULL)
385 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
387 return &task_rq(p)->cfs;
390 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
392 struct task_struct *p = task_of(se);
393 struct rq *rq = task_rq(p);
398 /* runqueue "owned" by this group */
399 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
404 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
408 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
412 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
413 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
415 static inline struct sched_entity *parent_entity(struct sched_entity *se)
421 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
425 #endif /* CONFIG_FAIR_GROUP_SCHED */
427 static __always_inline
428 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
430 /**************************************************************
431 * Scheduling class tree data structure manipulation methods:
434 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
436 s64 delta = (s64)(vruntime - max_vruntime);
438 max_vruntime = vruntime;
443 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
445 s64 delta = (s64)(vruntime - min_vruntime);
447 min_vruntime = vruntime;
452 static inline int entity_before(struct sched_entity *a,
453 struct sched_entity *b)
455 return (s64)(a->vruntime - b->vruntime) < 0;
458 static void update_min_vruntime(struct cfs_rq *cfs_rq)
460 u64 vruntime = cfs_rq->min_vruntime;
463 vruntime = cfs_rq->curr->vruntime;
465 if (cfs_rq->rb_leftmost) {
466 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
471 vruntime = se->vruntime;
473 vruntime = min_vruntime(vruntime, se->vruntime);
476 /* ensure we never gain time by being placed backwards. */
477 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
480 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
485 * Enqueue an entity into the rb-tree:
487 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
489 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
490 struct rb_node *parent = NULL;
491 struct sched_entity *entry;
495 * Find the right place in the rbtree:
499 entry = rb_entry(parent, struct sched_entity, run_node);
501 * We dont care about collisions. Nodes with
502 * the same key stay together.
504 if (entity_before(se, entry)) {
505 link = &parent->rb_left;
507 link = &parent->rb_right;
513 * Maintain a cache of leftmost tree entries (it is frequently
517 cfs_rq->rb_leftmost = &se->run_node;
519 rb_link_node(&se->run_node, parent, link);
520 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
523 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
525 if (cfs_rq->rb_leftmost == &se->run_node) {
526 struct rb_node *next_node;
528 next_node = rb_next(&se->run_node);
529 cfs_rq->rb_leftmost = next_node;
532 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
535 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
537 struct rb_node *left = cfs_rq->rb_leftmost;
542 return rb_entry(left, struct sched_entity, run_node);
545 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
547 struct rb_node *next = rb_next(&se->run_node);
552 return rb_entry(next, struct sched_entity, run_node);
555 #ifdef CONFIG_SCHED_DEBUG
556 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
558 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
563 return rb_entry(last, struct sched_entity, run_node);
566 /**************************************************************
567 * Scheduling class statistics methods:
570 int sched_proc_update_handler(struct ctl_table *table, int write,
571 void __user *buffer, size_t *lenp,
574 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
575 unsigned int factor = get_update_sysctl_factor();
580 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
581 sysctl_sched_min_granularity);
583 #define WRT_SYSCTL(name) \
584 (normalized_sysctl_##name = sysctl_##name / (factor))
585 WRT_SYSCTL(sched_min_granularity);
586 WRT_SYSCTL(sched_latency);
587 WRT_SYSCTL(sched_wakeup_granularity);
597 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
599 if (unlikely(se->load.weight != NICE_0_LOAD))
600 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
606 * The idea is to set a period in which each task runs once.
608 * When there are too many tasks (sched_nr_latency) we have to stretch
609 * this period because otherwise the slices get too small.
611 * p = (nr <= nl) ? l : l*nr/nl
613 static u64 __sched_period(unsigned long nr_running)
615 if (unlikely(nr_running > sched_nr_latency))
616 return nr_running * sysctl_sched_min_granularity;
618 return sysctl_sched_latency;
622 * We calculate the wall-time slice from the period by taking a part
623 * proportional to the weight.
627 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
629 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
631 for_each_sched_entity(se) {
632 struct load_weight *load;
633 struct load_weight lw;
635 cfs_rq = cfs_rq_of(se);
636 load = &cfs_rq->load;
638 if (unlikely(!se->on_rq)) {
641 update_load_add(&lw, se->load.weight);
644 slice = __calc_delta(slice, se->load.weight, load);
650 * We calculate the vruntime slice of a to-be-inserted task.
654 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
656 return calc_delta_fair(sched_slice(cfs_rq, se), se);
660 static int select_idle_sibling(struct task_struct *p, int cpu);
661 static unsigned long task_h_load(struct task_struct *p);
664 * We choose a half-life close to 1 scheduling period.
665 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
666 * dependent on this value.
668 #define LOAD_AVG_PERIOD 32
669 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
670 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
672 /* Give new sched_entity start runnable values to heavy its load in infant time */
673 void init_entity_runnable_average(struct sched_entity *se)
675 struct sched_avg *sa = &se->avg;
677 sa->last_update_time = 0;
679 * sched_avg's period_contrib should be strictly less then 1024, so
680 * we give it 1023 to make sure it is almost a period (1024us), and
681 * will definitely be update (after enqueue).
683 sa->period_contrib = 1023;
684 sa->load_avg = scale_load_down(se->load.weight);
685 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
686 sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
687 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
688 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
691 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
692 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
694 void init_entity_runnable_average(struct sched_entity *se)
700 * Update the current task's runtime statistics.
702 static void update_curr(struct cfs_rq *cfs_rq)
704 struct sched_entity *curr = cfs_rq->curr;
705 u64 now = rq_clock_task(rq_of(cfs_rq));
711 delta_exec = now - curr->exec_start;
712 if (unlikely((s64)delta_exec <= 0))
715 curr->exec_start = now;
717 schedstat_set(curr->statistics.exec_max,
718 max(delta_exec, curr->statistics.exec_max));
720 curr->sum_exec_runtime += delta_exec;
721 schedstat_add(cfs_rq, exec_clock, delta_exec);
723 curr->vruntime += calc_delta_fair(delta_exec, curr);
724 update_min_vruntime(cfs_rq);
726 if (entity_is_task(curr)) {
727 struct task_struct *curtask = task_of(curr);
729 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
730 cpuacct_charge(curtask, delta_exec);
731 account_group_exec_runtime(curtask, delta_exec);
734 account_cfs_rq_runtime(cfs_rq, delta_exec);
737 static void update_curr_fair(struct rq *rq)
739 update_curr(cfs_rq_of(&rq->curr->se));
743 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
745 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
749 * Task is being enqueued - update stats:
751 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
754 * Are we enqueueing a waiting task? (for current tasks
755 * a dequeue/enqueue event is a NOP)
757 if (se != cfs_rq->curr)
758 update_stats_wait_start(cfs_rq, se);
762 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
764 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
765 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
766 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
767 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
768 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
769 #ifdef CONFIG_SCHEDSTATS
770 if (entity_is_task(se)) {
771 trace_sched_stat_wait(task_of(se),
772 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
775 schedstat_set(se->statistics.wait_start, 0);
779 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
782 * Mark the end of the wait period if dequeueing a
785 if (se != cfs_rq->curr)
786 update_stats_wait_end(cfs_rq, se);
790 * We are picking a new current task - update its stats:
793 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
796 * We are starting a new run period:
798 se->exec_start = rq_clock_task(rq_of(cfs_rq));
801 /**************************************************
802 * Scheduling class queueing methods:
805 #ifdef CONFIG_NUMA_BALANCING
807 * Approximate time to scan a full NUMA task in ms. The task scan period is
808 * calculated based on the tasks virtual memory size and
809 * numa_balancing_scan_size.
811 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
812 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
814 /* Portion of address space to scan in MB */
815 unsigned int sysctl_numa_balancing_scan_size = 256;
817 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
818 unsigned int sysctl_numa_balancing_scan_delay = 1000;
820 static unsigned int task_nr_scan_windows(struct task_struct *p)
822 unsigned long rss = 0;
823 unsigned long nr_scan_pages;
826 * Calculations based on RSS as non-present and empty pages are skipped
827 * by the PTE scanner and NUMA hinting faults should be trapped based
830 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
831 rss = get_mm_rss(p->mm);
835 rss = round_up(rss, nr_scan_pages);
836 return rss / nr_scan_pages;
839 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
840 #define MAX_SCAN_WINDOW 2560
842 static unsigned int task_scan_min(struct task_struct *p)
844 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
845 unsigned int scan, floor;
846 unsigned int windows = 1;
848 if (scan_size < MAX_SCAN_WINDOW)
849 windows = MAX_SCAN_WINDOW / scan_size;
850 floor = 1000 / windows;
852 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
853 return max_t(unsigned int, floor, scan);
856 static unsigned int task_scan_max(struct task_struct *p)
858 unsigned int smin = task_scan_min(p);
861 /* Watch for min being lower than max due to floor calculations */
862 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
863 return max(smin, smax);
866 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
868 rq->nr_numa_running += (p->numa_preferred_nid != -1);
869 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
872 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
874 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
875 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
881 spinlock_t lock; /* nr_tasks, tasks */
886 nodemask_t active_nodes;
887 unsigned long total_faults;
889 * Faults_cpu is used to decide whether memory should move
890 * towards the CPU. As a consequence, these stats are weighted
891 * more by CPU use than by memory faults.
893 unsigned long *faults_cpu;
894 unsigned long faults[0];
897 /* Shared or private faults. */
898 #define NR_NUMA_HINT_FAULT_TYPES 2
900 /* Memory and CPU locality */
901 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
903 /* Averaged statistics, and temporary buffers. */
904 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
906 pid_t task_numa_group_id(struct task_struct *p)
908 return p->numa_group ? p->numa_group->gid : 0;
912 * The averaged statistics, shared & private, memory & cpu,
913 * occupy the first half of the array. The second half of the
914 * array is for current counters, which are averaged into the
915 * first set by task_numa_placement.
917 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
919 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
922 static inline unsigned long task_faults(struct task_struct *p, int nid)
927 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
928 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
931 static inline unsigned long group_faults(struct task_struct *p, int nid)
936 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
937 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
940 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
942 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
943 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
946 /* Handle placement on systems where not all nodes are directly connected. */
947 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
948 int maxdist, bool task)
950 unsigned long score = 0;
954 * All nodes are directly connected, and the same distance
955 * from each other. No need for fancy placement algorithms.
957 if (sched_numa_topology_type == NUMA_DIRECT)
961 * This code is called for each node, introducing N^2 complexity,
962 * which should be ok given the number of nodes rarely exceeds 8.
964 for_each_online_node(node) {
965 unsigned long faults;
966 int dist = node_distance(nid, node);
969 * The furthest away nodes in the system are not interesting
970 * for placement; nid was already counted.
972 if (dist == sched_max_numa_distance || node == nid)
976 * On systems with a backplane NUMA topology, compare groups
977 * of nodes, and move tasks towards the group with the most
978 * memory accesses. When comparing two nodes at distance
979 * "hoplimit", only nodes closer by than "hoplimit" are part
980 * of each group. Skip other nodes.
982 if (sched_numa_topology_type == NUMA_BACKPLANE &&
986 /* Add up the faults from nearby nodes. */
988 faults = task_faults(p, node);
990 faults = group_faults(p, node);
993 * On systems with a glueless mesh NUMA topology, there are
994 * no fixed "groups of nodes". Instead, nodes that are not
995 * directly connected bounce traffic through intermediate
996 * nodes; a numa_group can occupy any set of nodes.
997 * The further away a node is, the less the faults count.
998 * This seems to result in good task placement.
1000 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1001 faults *= (sched_max_numa_distance - dist);
1002 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1012 * These return the fraction of accesses done by a particular task, or
1013 * task group, on a particular numa node. The group weight is given a
1014 * larger multiplier, in order to group tasks together that are almost
1015 * evenly spread out between numa nodes.
1017 static inline unsigned long task_weight(struct task_struct *p, int nid,
1020 unsigned long faults, total_faults;
1022 if (!p->numa_faults)
1025 total_faults = p->total_numa_faults;
1030 faults = task_faults(p, nid);
1031 faults += score_nearby_nodes(p, nid, dist, true);
1033 return 1000 * faults / total_faults;
1036 static inline unsigned long group_weight(struct task_struct *p, int nid,
1039 unsigned long faults, total_faults;
1044 total_faults = p->numa_group->total_faults;
1049 faults = group_faults(p, nid);
1050 faults += score_nearby_nodes(p, nid, dist, false);
1052 return 1000 * faults / total_faults;
1055 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1056 int src_nid, int dst_cpu)
1058 struct numa_group *ng = p->numa_group;
1059 int dst_nid = cpu_to_node(dst_cpu);
1060 int last_cpupid, this_cpupid;
1062 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1065 * Multi-stage node selection is used in conjunction with a periodic
1066 * migration fault to build a temporal task<->page relation. By using
1067 * a two-stage filter we remove short/unlikely relations.
1069 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1070 * a task's usage of a particular page (n_p) per total usage of this
1071 * page (n_t) (in a given time-span) to a probability.
1073 * Our periodic faults will sample this probability and getting the
1074 * same result twice in a row, given these samples are fully
1075 * independent, is then given by P(n)^2, provided our sample period
1076 * is sufficiently short compared to the usage pattern.
1078 * This quadric squishes small probabilities, making it less likely we
1079 * act on an unlikely task<->page relation.
1081 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1082 if (!cpupid_pid_unset(last_cpupid) &&
1083 cpupid_to_nid(last_cpupid) != dst_nid)
1086 /* Always allow migrate on private faults */
1087 if (cpupid_match_pid(p, last_cpupid))
1090 /* A shared fault, but p->numa_group has not been set up yet. */
1095 * Do not migrate if the destination is not a node that
1096 * is actively used by this numa group.
1098 if (!node_isset(dst_nid, ng->active_nodes))
1102 * Source is a node that is not actively used by this
1103 * numa group, while the destination is. Migrate.
1105 if (!node_isset(src_nid, ng->active_nodes))
1109 * Both source and destination are nodes in active
1110 * use by this numa group. Maximize memory bandwidth
1111 * by migrating from more heavily used groups, to less
1112 * heavily used ones, spreading the load around.
1113 * Use a 1/4 hysteresis to avoid spurious page movement.
1115 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1118 static unsigned long weighted_cpuload(const int cpu);
1119 static unsigned long source_load(int cpu, int type);
1120 static unsigned long target_load(int cpu, int type);
1121 static unsigned long capacity_of(int cpu);
1122 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1124 /* Cached statistics for all CPUs within a node */
1126 unsigned long nr_running;
1129 /* Total compute capacity of CPUs on a node */
1130 unsigned long compute_capacity;
1132 /* Approximate capacity in terms of runnable tasks on a node */
1133 unsigned long task_capacity;
1134 int has_free_capacity;
1138 * XXX borrowed from update_sg_lb_stats
1140 static void update_numa_stats(struct numa_stats *ns, int nid)
1142 int smt, cpu, cpus = 0;
1143 unsigned long capacity;
1145 memset(ns, 0, sizeof(*ns));
1146 for_each_cpu(cpu, cpumask_of_node(nid)) {
1147 struct rq *rq = cpu_rq(cpu);
1149 ns->nr_running += rq->nr_running;
1150 ns->load += weighted_cpuload(cpu);
1151 ns->compute_capacity += capacity_of(cpu);
1157 * If we raced with hotplug and there are no CPUs left in our mask
1158 * the @ns structure is NULL'ed and task_numa_compare() will
1159 * not find this node attractive.
1161 * We'll either bail at !has_free_capacity, or we'll detect a huge
1162 * imbalance and bail there.
1167 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1168 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1169 capacity = cpus / smt; /* cores */
1171 ns->task_capacity = min_t(unsigned, capacity,
1172 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1173 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1176 struct task_numa_env {
1177 struct task_struct *p;
1179 int src_cpu, src_nid;
1180 int dst_cpu, dst_nid;
1182 struct numa_stats src_stats, dst_stats;
1187 struct task_struct *best_task;
1192 static void task_numa_assign(struct task_numa_env *env,
1193 struct task_struct *p, long imp)
1196 put_task_struct(env->best_task);
1201 env->best_imp = imp;
1202 env->best_cpu = env->dst_cpu;
1205 static bool load_too_imbalanced(long src_load, long dst_load,
1206 struct task_numa_env *env)
1209 long orig_src_load, orig_dst_load;
1210 long src_capacity, dst_capacity;
1213 * The load is corrected for the CPU capacity available on each node.
1216 * ------------ vs ---------
1217 * src_capacity dst_capacity
1219 src_capacity = env->src_stats.compute_capacity;
1220 dst_capacity = env->dst_stats.compute_capacity;
1222 /* We care about the slope of the imbalance, not the direction. */
1223 if (dst_load < src_load)
1224 swap(dst_load, src_load);
1226 /* Is the difference below the threshold? */
1227 imb = dst_load * src_capacity * 100 -
1228 src_load * dst_capacity * env->imbalance_pct;
1233 * The imbalance is above the allowed threshold.
1234 * Compare it with the old imbalance.
1236 orig_src_load = env->src_stats.load;
1237 orig_dst_load = env->dst_stats.load;
1239 if (orig_dst_load < orig_src_load)
1240 swap(orig_dst_load, orig_src_load);
1242 old_imb = orig_dst_load * src_capacity * 100 -
1243 orig_src_load * dst_capacity * env->imbalance_pct;
1245 /* Would this change make things worse? */
1246 return (imb > old_imb);
1250 * This checks if the overall compute and NUMA accesses of the system would
1251 * be improved if the source tasks was migrated to the target dst_cpu taking
1252 * into account that it might be best if task running on the dst_cpu should
1253 * be exchanged with the source task
1255 static void task_numa_compare(struct task_numa_env *env,
1256 long taskimp, long groupimp)
1258 struct rq *src_rq = cpu_rq(env->src_cpu);
1259 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1260 struct task_struct *cur;
1261 long src_load, dst_load;
1263 long imp = env->p->numa_group ? groupimp : taskimp;
1265 int dist = env->dist;
1269 raw_spin_lock_irq(&dst_rq->lock);
1272 * No need to move the exiting task, and this ensures that ->curr
1273 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1274 * is safe under RCU read lock.
1275 * Note that rcu_read_lock() itself can't protect from the final
1276 * put_task_struct() after the last schedule().
1278 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1280 raw_spin_unlock_irq(&dst_rq->lock);
1283 * Because we have preemption enabled we can get migrated around and
1284 * end try selecting ourselves (current == env->p) as a swap candidate.
1290 * "imp" is the fault differential for the source task between the
1291 * source and destination node. Calculate the total differential for
1292 * the source task and potential destination task. The more negative
1293 * the value is, the more rmeote accesses that would be expected to
1294 * be incurred if the tasks were swapped.
1297 /* Skip this swap candidate if cannot move to the source cpu */
1298 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1302 * If dst and source tasks are in the same NUMA group, or not
1303 * in any group then look only at task weights.
1305 if (cur->numa_group == env->p->numa_group) {
1306 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1307 task_weight(cur, env->dst_nid, dist);
1309 * Add some hysteresis to prevent swapping the
1310 * tasks within a group over tiny differences.
1312 if (cur->numa_group)
1316 * Compare the group weights. If a task is all by
1317 * itself (not part of a group), use the task weight
1320 if (cur->numa_group)
1321 imp += group_weight(cur, env->src_nid, dist) -
1322 group_weight(cur, env->dst_nid, dist);
1324 imp += task_weight(cur, env->src_nid, dist) -
1325 task_weight(cur, env->dst_nid, dist);
1329 if (imp <= env->best_imp && moveimp <= env->best_imp)
1333 /* Is there capacity at our destination? */
1334 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1335 !env->dst_stats.has_free_capacity)
1341 /* Balance doesn't matter much if we're running a task per cpu */
1342 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1343 dst_rq->nr_running == 1)
1347 * In the overloaded case, try and keep the load balanced.
1350 load = task_h_load(env->p);
1351 dst_load = env->dst_stats.load + load;
1352 src_load = env->src_stats.load - load;
1354 if (moveimp > imp && moveimp > env->best_imp) {
1356 * If the improvement from just moving env->p direction is
1357 * better than swapping tasks around, check if a move is
1358 * possible. Store a slightly smaller score than moveimp,
1359 * so an actually idle CPU will win.
1361 if (!load_too_imbalanced(src_load, dst_load, env)) {
1368 if (imp <= env->best_imp)
1372 load = task_h_load(cur);
1377 if (load_too_imbalanced(src_load, dst_load, env))
1381 * One idle CPU per node is evaluated for a task numa move.
1382 * Call select_idle_sibling to maybe find a better one.
1385 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1388 task_numa_assign(env, cur, imp);
1393 static void task_numa_find_cpu(struct task_numa_env *env,
1394 long taskimp, long groupimp)
1398 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1399 /* Skip this CPU if the source task cannot migrate */
1400 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1404 task_numa_compare(env, taskimp, groupimp);
1408 /* Only move tasks to a NUMA node less busy than the current node. */
1409 static bool numa_has_capacity(struct task_numa_env *env)
1411 struct numa_stats *src = &env->src_stats;
1412 struct numa_stats *dst = &env->dst_stats;
1414 if (src->has_free_capacity && !dst->has_free_capacity)
1418 * Only consider a task move if the source has a higher load
1419 * than the destination, corrected for CPU capacity on each node.
1421 * src->load dst->load
1422 * --------------------- vs ---------------------
1423 * src->compute_capacity dst->compute_capacity
1425 if (src->load * dst->compute_capacity * env->imbalance_pct >
1427 dst->load * src->compute_capacity * 100)
1433 static int task_numa_migrate(struct task_struct *p)
1435 struct task_numa_env env = {
1438 .src_cpu = task_cpu(p),
1439 .src_nid = task_node(p),
1441 .imbalance_pct = 112,
1447 struct sched_domain *sd;
1448 unsigned long taskweight, groupweight;
1450 long taskimp, groupimp;
1453 * Pick the lowest SD_NUMA domain, as that would have the smallest
1454 * imbalance and would be the first to start moving tasks about.
1456 * And we want to avoid any moving of tasks about, as that would create
1457 * random movement of tasks -- counter the numa conditions we're trying
1461 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1463 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1467 * Cpusets can break the scheduler domain tree into smaller
1468 * balance domains, some of which do not cross NUMA boundaries.
1469 * Tasks that are "trapped" in such domains cannot be migrated
1470 * elsewhere, so there is no point in (re)trying.
1472 if (unlikely(!sd)) {
1473 p->numa_preferred_nid = task_node(p);
1477 env.dst_nid = p->numa_preferred_nid;
1478 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1479 taskweight = task_weight(p, env.src_nid, dist);
1480 groupweight = group_weight(p, env.src_nid, dist);
1481 update_numa_stats(&env.src_stats, env.src_nid);
1482 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1483 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1484 update_numa_stats(&env.dst_stats, env.dst_nid);
1486 /* Try to find a spot on the preferred nid. */
1487 if (numa_has_capacity(&env))
1488 task_numa_find_cpu(&env, taskimp, groupimp);
1491 * Look at other nodes in these cases:
1492 * - there is no space available on the preferred_nid
1493 * - the task is part of a numa_group that is interleaved across
1494 * multiple NUMA nodes; in order to better consolidate the group,
1495 * we need to check other locations.
1497 if (env.best_cpu == -1 || (p->numa_group &&
1498 nodes_weight(p->numa_group->active_nodes) > 1)) {
1499 for_each_online_node(nid) {
1500 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1503 dist = node_distance(env.src_nid, env.dst_nid);
1504 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1506 taskweight = task_weight(p, env.src_nid, dist);
1507 groupweight = group_weight(p, env.src_nid, dist);
1510 /* Only consider nodes where both task and groups benefit */
1511 taskimp = task_weight(p, nid, dist) - taskweight;
1512 groupimp = group_weight(p, nid, dist) - groupweight;
1513 if (taskimp < 0 && groupimp < 0)
1518 update_numa_stats(&env.dst_stats, env.dst_nid);
1519 if (numa_has_capacity(&env))
1520 task_numa_find_cpu(&env, taskimp, groupimp);
1525 * If the task is part of a workload that spans multiple NUMA nodes,
1526 * and is migrating into one of the workload's active nodes, remember
1527 * this node as the task's preferred numa node, so the workload can
1529 * A task that migrated to a second choice node will be better off
1530 * trying for a better one later. Do not set the preferred node here.
1532 if (p->numa_group) {
1533 if (env.best_cpu == -1)
1538 if (node_isset(nid, p->numa_group->active_nodes))
1539 sched_setnuma(p, env.dst_nid);
1542 /* No better CPU than the current one was found. */
1543 if (env.best_cpu == -1)
1547 * Reset the scan period if the task is being rescheduled on an
1548 * alternative node to recheck if the tasks is now properly placed.
1550 p->numa_scan_period = task_scan_min(p);
1552 if (env.best_task == NULL) {
1553 ret = migrate_task_to(p, env.best_cpu);
1555 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1559 ret = migrate_swap(p, env.best_task);
1561 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1562 put_task_struct(env.best_task);
1566 /* Attempt to migrate a task to a CPU on the preferred node. */
1567 static void numa_migrate_preferred(struct task_struct *p)
1569 unsigned long interval = HZ;
1571 /* This task has no NUMA fault statistics yet */
1572 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1575 /* Periodically retry migrating the task to the preferred node */
1576 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1577 p->numa_migrate_retry = jiffies + interval;
1579 /* Success if task is already running on preferred CPU */
1580 if (task_node(p) == p->numa_preferred_nid)
1583 /* Otherwise, try migrate to a CPU on the preferred node */
1584 task_numa_migrate(p);
1588 * Find the nodes on which the workload is actively running. We do this by
1589 * tracking the nodes from which NUMA hinting faults are triggered. This can
1590 * be different from the set of nodes where the workload's memory is currently
1593 * The bitmask is used to make smarter decisions on when to do NUMA page
1594 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1595 * are added when they cause over 6/16 of the maximum number of faults, but
1596 * only removed when they drop below 3/16.
1598 static void update_numa_active_node_mask(struct numa_group *numa_group)
1600 unsigned long faults, max_faults = 0;
1603 for_each_online_node(nid) {
1604 faults = group_faults_cpu(numa_group, nid);
1605 if (faults > max_faults)
1606 max_faults = faults;
1609 for_each_online_node(nid) {
1610 faults = group_faults_cpu(numa_group, nid);
1611 if (!node_isset(nid, numa_group->active_nodes)) {
1612 if (faults > max_faults * 6 / 16)
1613 node_set(nid, numa_group->active_nodes);
1614 } else if (faults < max_faults * 3 / 16)
1615 node_clear(nid, numa_group->active_nodes);
1620 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1621 * increments. The more local the fault statistics are, the higher the scan
1622 * period will be for the next scan window. If local/(local+remote) ratio is
1623 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1624 * the scan period will decrease. Aim for 70% local accesses.
1626 #define NUMA_PERIOD_SLOTS 10
1627 #define NUMA_PERIOD_THRESHOLD 7
1630 * Increase the scan period (slow down scanning) if the majority of
1631 * our memory is already on our local node, or if the majority of
1632 * the page accesses are shared with other processes.
1633 * Otherwise, decrease the scan period.
1635 static void update_task_scan_period(struct task_struct *p,
1636 unsigned long shared, unsigned long private)
1638 unsigned int period_slot;
1642 unsigned long remote = p->numa_faults_locality[0];
1643 unsigned long local = p->numa_faults_locality[1];
1646 * If there were no record hinting faults then either the task is
1647 * completely idle or all activity is areas that are not of interest
1648 * to automatic numa balancing. Related to that, if there were failed
1649 * migration then it implies we are migrating too quickly or the local
1650 * node is overloaded. In either case, scan slower
1652 if (local + shared == 0 || p->numa_faults_locality[2]) {
1653 p->numa_scan_period = min(p->numa_scan_period_max,
1654 p->numa_scan_period << 1);
1656 p->mm->numa_next_scan = jiffies +
1657 msecs_to_jiffies(p->numa_scan_period);
1663 * Prepare to scale scan period relative to the current period.
1664 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1665 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1666 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1668 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1669 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1670 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1671 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1674 diff = slot * period_slot;
1676 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1679 * Scale scan rate increases based on sharing. There is an
1680 * inverse relationship between the degree of sharing and
1681 * the adjustment made to the scanning period. Broadly
1682 * speaking the intent is that there is little point
1683 * scanning faster if shared accesses dominate as it may
1684 * simply bounce migrations uselessly
1686 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1687 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1690 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1691 task_scan_min(p), task_scan_max(p));
1692 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1696 * Get the fraction of time the task has been running since the last
1697 * NUMA placement cycle. The scheduler keeps similar statistics, but
1698 * decays those on a 32ms period, which is orders of magnitude off
1699 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1700 * stats only if the task is so new there are no NUMA statistics yet.
1702 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1704 u64 runtime, delta, now;
1705 /* Use the start of this time slice to avoid calculations. */
1706 now = p->se.exec_start;
1707 runtime = p->se.sum_exec_runtime;
1709 if (p->last_task_numa_placement) {
1710 delta = runtime - p->last_sum_exec_runtime;
1711 *period = now - p->last_task_numa_placement;
1713 delta = p->se.avg.load_sum / p->se.load.weight;
1714 *period = LOAD_AVG_MAX;
1717 p->last_sum_exec_runtime = runtime;
1718 p->last_task_numa_placement = now;
1724 * Determine the preferred nid for a task in a numa_group. This needs to
1725 * be done in a way that produces consistent results with group_weight,
1726 * otherwise workloads might not converge.
1728 static int preferred_group_nid(struct task_struct *p, int nid)
1733 /* Direct connections between all NUMA nodes. */
1734 if (sched_numa_topology_type == NUMA_DIRECT)
1738 * On a system with glueless mesh NUMA topology, group_weight
1739 * scores nodes according to the number of NUMA hinting faults on
1740 * both the node itself, and on nearby nodes.
1742 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1743 unsigned long score, max_score = 0;
1744 int node, max_node = nid;
1746 dist = sched_max_numa_distance;
1748 for_each_online_node(node) {
1749 score = group_weight(p, node, dist);
1750 if (score > max_score) {
1759 * Finding the preferred nid in a system with NUMA backplane
1760 * interconnect topology is more involved. The goal is to locate
1761 * tasks from numa_groups near each other in the system, and
1762 * untangle workloads from different sides of the system. This requires
1763 * searching down the hierarchy of node groups, recursively searching
1764 * inside the highest scoring group of nodes. The nodemask tricks
1765 * keep the complexity of the search down.
1767 nodes = node_online_map;
1768 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1769 unsigned long max_faults = 0;
1770 nodemask_t max_group = NODE_MASK_NONE;
1773 /* Are there nodes at this distance from each other? */
1774 if (!find_numa_distance(dist))
1777 for_each_node_mask(a, nodes) {
1778 unsigned long faults = 0;
1779 nodemask_t this_group;
1780 nodes_clear(this_group);
1782 /* Sum group's NUMA faults; includes a==b case. */
1783 for_each_node_mask(b, nodes) {
1784 if (node_distance(a, b) < dist) {
1785 faults += group_faults(p, b);
1786 node_set(b, this_group);
1787 node_clear(b, nodes);
1791 /* Remember the top group. */
1792 if (faults > max_faults) {
1793 max_faults = faults;
1794 max_group = this_group;
1796 * subtle: at the smallest distance there is
1797 * just one node left in each "group", the
1798 * winner is the preferred nid.
1803 /* Next round, evaluate the nodes within max_group. */
1811 static void task_numa_placement(struct task_struct *p)
1813 int seq, nid, max_nid = -1, max_group_nid = -1;
1814 unsigned long max_faults = 0, max_group_faults = 0;
1815 unsigned long fault_types[2] = { 0, 0 };
1816 unsigned long total_faults;
1817 u64 runtime, period;
1818 spinlock_t *group_lock = NULL;
1821 * The p->mm->numa_scan_seq field gets updated without
1822 * exclusive access. Use READ_ONCE() here to ensure
1823 * that the field is read in a single access:
1825 seq = READ_ONCE(p->mm->numa_scan_seq);
1826 if (p->numa_scan_seq == seq)
1828 p->numa_scan_seq = seq;
1829 p->numa_scan_period_max = task_scan_max(p);
1831 total_faults = p->numa_faults_locality[0] +
1832 p->numa_faults_locality[1];
1833 runtime = numa_get_avg_runtime(p, &period);
1835 /* If the task is part of a group prevent parallel updates to group stats */
1836 if (p->numa_group) {
1837 group_lock = &p->numa_group->lock;
1838 spin_lock_irq(group_lock);
1841 /* Find the node with the highest number of faults */
1842 for_each_online_node(nid) {
1843 /* Keep track of the offsets in numa_faults array */
1844 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1845 unsigned long faults = 0, group_faults = 0;
1848 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1849 long diff, f_diff, f_weight;
1851 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1852 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1853 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1854 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1856 /* Decay existing window, copy faults since last scan */
1857 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1858 fault_types[priv] += p->numa_faults[membuf_idx];
1859 p->numa_faults[membuf_idx] = 0;
1862 * Normalize the faults_from, so all tasks in a group
1863 * count according to CPU use, instead of by the raw
1864 * number of faults. Tasks with little runtime have
1865 * little over-all impact on throughput, and thus their
1866 * faults are less important.
1868 f_weight = div64_u64(runtime << 16, period + 1);
1869 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1871 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1872 p->numa_faults[cpubuf_idx] = 0;
1874 p->numa_faults[mem_idx] += diff;
1875 p->numa_faults[cpu_idx] += f_diff;
1876 faults += p->numa_faults[mem_idx];
1877 p->total_numa_faults += diff;
1878 if (p->numa_group) {
1880 * safe because we can only change our own group
1882 * mem_idx represents the offset for a given
1883 * nid and priv in a specific region because it
1884 * is at the beginning of the numa_faults array.
1886 p->numa_group->faults[mem_idx] += diff;
1887 p->numa_group->faults_cpu[mem_idx] += f_diff;
1888 p->numa_group->total_faults += diff;
1889 group_faults += p->numa_group->faults[mem_idx];
1893 if (faults > max_faults) {
1894 max_faults = faults;
1898 if (group_faults > max_group_faults) {
1899 max_group_faults = group_faults;
1900 max_group_nid = nid;
1904 update_task_scan_period(p, fault_types[0], fault_types[1]);
1906 if (p->numa_group) {
1907 update_numa_active_node_mask(p->numa_group);
1908 spin_unlock_irq(group_lock);
1909 max_nid = preferred_group_nid(p, max_group_nid);
1913 /* Set the new preferred node */
1914 if (max_nid != p->numa_preferred_nid)
1915 sched_setnuma(p, max_nid);
1917 if (task_node(p) != p->numa_preferred_nid)
1918 numa_migrate_preferred(p);
1922 static inline int get_numa_group(struct numa_group *grp)
1924 return atomic_inc_not_zero(&grp->refcount);
1927 static inline void put_numa_group(struct numa_group *grp)
1929 if (atomic_dec_and_test(&grp->refcount))
1930 kfree_rcu(grp, rcu);
1933 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1936 struct numa_group *grp, *my_grp;
1937 struct task_struct *tsk;
1939 int cpu = cpupid_to_cpu(cpupid);
1942 if (unlikely(!p->numa_group)) {
1943 unsigned int size = sizeof(struct numa_group) +
1944 4*nr_node_ids*sizeof(unsigned long);
1946 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1950 atomic_set(&grp->refcount, 1);
1951 spin_lock_init(&grp->lock);
1953 /* Second half of the array tracks nids where faults happen */
1954 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1957 node_set(task_node(current), grp->active_nodes);
1959 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1960 grp->faults[i] = p->numa_faults[i];
1962 grp->total_faults = p->total_numa_faults;
1965 rcu_assign_pointer(p->numa_group, grp);
1969 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1971 if (!cpupid_match_pid(tsk, cpupid))
1974 grp = rcu_dereference(tsk->numa_group);
1978 my_grp = p->numa_group;
1983 * Only join the other group if its bigger; if we're the bigger group,
1984 * the other task will join us.
1986 if (my_grp->nr_tasks > grp->nr_tasks)
1990 * Tie-break on the grp address.
1992 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1995 /* Always join threads in the same process. */
1996 if (tsk->mm == current->mm)
1999 /* Simple filter to avoid false positives due to PID collisions */
2000 if (flags & TNF_SHARED)
2003 /* Update priv based on whether false sharing was detected */
2006 if (join && !get_numa_group(grp))
2014 BUG_ON(irqs_disabled());
2015 double_lock_irq(&my_grp->lock, &grp->lock);
2017 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2018 my_grp->faults[i] -= p->numa_faults[i];
2019 grp->faults[i] += p->numa_faults[i];
2021 my_grp->total_faults -= p->total_numa_faults;
2022 grp->total_faults += p->total_numa_faults;
2027 spin_unlock(&my_grp->lock);
2028 spin_unlock_irq(&grp->lock);
2030 rcu_assign_pointer(p->numa_group, grp);
2032 put_numa_group(my_grp);
2040 void task_numa_free(struct task_struct *p)
2042 struct numa_group *grp = p->numa_group;
2043 void *numa_faults = p->numa_faults;
2044 unsigned long flags;
2048 spin_lock_irqsave(&grp->lock, flags);
2049 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2050 grp->faults[i] -= p->numa_faults[i];
2051 grp->total_faults -= p->total_numa_faults;
2054 spin_unlock_irqrestore(&grp->lock, flags);
2055 RCU_INIT_POINTER(p->numa_group, NULL);
2056 put_numa_group(grp);
2059 p->numa_faults = NULL;
2064 * Got a PROT_NONE fault for a page on @node.
2066 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2068 struct task_struct *p = current;
2069 bool migrated = flags & TNF_MIGRATED;
2070 int cpu_node = task_node(current);
2071 int local = !!(flags & TNF_FAULT_LOCAL);
2074 if (!static_branch_likely(&sched_numa_balancing))
2077 /* for example, ksmd faulting in a user's mm */
2081 /* Allocate buffer to track faults on a per-node basis */
2082 if (unlikely(!p->numa_faults)) {
2083 int size = sizeof(*p->numa_faults) *
2084 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2086 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2087 if (!p->numa_faults)
2090 p->total_numa_faults = 0;
2091 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2095 * First accesses are treated as private, otherwise consider accesses
2096 * to be private if the accessing pid has not changed
2098 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2101 priv = cpupid_match_pid(p, last_cpupid);
2102 if (!priv && !(flags & TNF_NO_GROUP))
2103 task_numa_group(p, last_cpupid, flags, &priv);
2107 * If a workload spans multiple NUMA nodes, a shared fault that
2108 * occurs wholly within the set of nodes that the workload is
2109 * actively using should be counted as local. This allows the
2110 * scan rate to slow down when a workload has settled down.
2112 if (!priv && !local && p->numa_group &&
2113 node_isset(cpu_node, p->numa_group->active_nodes) &&
2114 node_isset(mem_node, p->numa_group->active_nodes))
2117 task_numa_placement(p);
2120 * Retry task to preferred node migration periodically, in case it
2121 * case it previously failed, or the scheduler moved us.
2123 if (time_after(jiffies, p->numa_migrate_retry))
2124 numa_migrate_preferred(p);
2127 p->numa_pages_migrated += pages;
2128 if (flags & TNF_MIGRATE_FAIL)
2129 p->numa_faults_locality[2] += pages;
2131 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2132 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2133 p->numa_faults_locality[local] += pages;
2136 static void reset_ptenuma_scan(struct task_struct *p)
2139 * We only did a read acquisition of the mmap sem, so
2140 * p->mm->numa_scan_seq is written to without exclusive access
2141 * and the update is not guaranteed to be atomic. That's not
2142 * much of an issue though, since this is just used for
2143 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2144 * expensive, to avoid any form of compiler optimizations:
2146 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2147 p->mm->numa_scan_offset = 0;
2151 * The expensive part of numa migration is done from task_work context.
2152 * Triggered from task_tick_numa().
2154 void task_numa_work(struct callback_head *work)
2156 unsigned long migrate, next_scan, now = jiffies;
2157 struct task_struct *p = current;
2158 struct mm_struct *mm = p->mm;
2159 struct vm_area_struct *vma;
2160 unsigned long start, end;
2161 unsigned long nr_pte_updates = 0;
2162 long pages, virtpages;
2164 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2166 work->next = work; /* protect against double add */
2168 * Who cares about NUMA placement when they're dying.
2170 * NOTE: make sure not to dereference p->mm before this check,
2171 * exit_task_work() happens _after_ exit_mm() so we could be called
2172 * without p->mm even though we still had it when we enqueued this
2175 if (p->flags & PF_EXITING)
2178 if (!mm->numa_next_scan) {
2179 mm->numa_next_scan = now +
2180 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2184 * Enforce maximal scan/migration frequency..
2186 migrate = mm->numa_next_scan;
2187 if (time_before(now, migrate))
2190 if (p->numa_scan_period == 0) {
2191 p->numa_scan_period_max = task_scan_max(p);
2192 p->numa_scan_period = task_scan_min(p);
2195 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2196 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2200 * Delay this task enough that another task of this mm will likely win
2201 * the next time around.
2203 p->node_stamp += 2 * TICK_NSEC;
2205 start = mm->numa_scan_offset;
2206 pages = sysctl_numa_balancing_scan_size;
2207 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2208 virtpages = pages * 8; /* Scan up to this much virtual space */
2213 down_read(&mm->mmap_sem);
2214 vma = find_vma(mm, start);
2216 reset_ptenuma_scan(p);
2220 for (; vma; vma = vma->vm_next) {
2221 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2222 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2227 * Shared library pages mapped by multiple processes are not
2228 * migrated as it is expected they are cache replicated. Avoid
2229 * hinting faults in read-only file-backed mappings or the vdso
2230 * as migrating the pages will be of marginal benefit.
2233 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2237 * Skip inaccessible VMAs to avoid any confusion between
2238 * PROT_NONE and NUMA hinting ptes
2240 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2244 start = max(start, vma->vm_start);
2245 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2246 end = min(end, vma->vm_end);
2247 nr_pte_updates = change_prot_numa(vma, start, end);
2250 * Try to scan sysctl_numa_balancing_size worth of
2251 * hpages that have at least one present PTE that
2252 * is not already pte-numa. If the VMA contains
2253 * areas that are unused or already full of prot_numa
2254 * PTEs, scan up to virtpages, to skip through those
2258 pages -= (end - start) >> PAGE_SHIFT;
2259 virtpages -= (end - start) >> PAGE_SHIFT;
2262 if (pages <= 0 || virtpages <= 0)
2266 } while (end != vma->vm_end);
2271 * It is possible to reach the end of the VMA list but the last few
2272 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2273 * would find the !migratable VMA on the next scan but not reset the
2274 * scanner to the start so check it now.
2277 mm->numa_scan_offset = start;
2279 reset_ptenuma_scan(p);
2280 up_read(&mm->mmap_sem);
2284 * Drive the periodic memory faults..
2286 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2288 struct callback_head *work = &curr->numa_work;
2292 * We don't care about NUMA placement if we don't have memory.
2294 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2298 * Using runtime rather than walltime has the dual advantage that
2299 * we (mostly) drive the selection from busy threads and that the
2300 * task needs to have done some actual work before we bother with
2303 now = curr->se.sum_exec_runtime;
2304 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2306 if (now > curr->node_stamp + period) {
2307 if (!curr->node_stamp)
2308 curr->numa_scan_period = task_scan_min(curr);
2309 curr->node_stamp += period;
2311 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2312 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2313 task_work_add(curr, work, true);
2318 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2322 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2326 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2329 #endif /* CONFIG_NUMA_BALANCING */
2332 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2334 update_load_add(&cfs_rq->load, se->load.weight);
2335 if (!parent_entity(se))
2336 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2338 if (entity_is_task(se)) {
2339 struct rq *rq = rq_of(cfs_rq);
2341 account_numa_enqueue(rq, task_of(se));
2342 list_add(&se->group_node, &rq->cfs_tasks);
2345 cfs_rq->nr_running++;
2349 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2351 update_load_sub(&cfs_rq->load, se->load.weight);
2352 if (!parent_entity(se))
2353 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2354 if (entity_is_task(se)) {
2355 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2356 list_del_init(&se->group_node);
2358 cfs_rq->nr_running--;
2361 #ifdef CONFIG_FAIR_GROUP_SCHED
2363 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2368 * Use this CPU's real-time load instead of the last load contribution
2369 * as the updating of the contribution is delayed, and we will use the
2370 * the real-time load to calc the share. See update_tg_load_avg().
2372 tg_weight = atomic_long_read(&tg->load_avg);
2373 tg_weight -= cfs_rq->tg_load_avg_contrib;
2374 tg_weight += cfs_rq->load.weight;
2379 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2381 long tg_weight, load, shares;
2383 tg_weight = calc_tg_weight(tg, cfs_rq);
2384 load = cfs_rq->load.weight;
2386 shares = (tg->shares * load);
2388 shares /= tg_weight;
2390 if (shares < MIN_SHARES)
2391 shares = MIN_SHARES;
2392 if (shares > tg->shares)
2393 shares = tg->shares;
2397 # else /* CONFIG_SMP */
2398 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2402 # endif /* CONFIG_SMP */
2403 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2404 unsigned long weight)
2407 /* commit outstanding execution time */
2408 if (cfs_rq->curr == se)
2409 update_curr(cfs_rq);
2410 account_entity_dequeue(cfs_rq, se);
2413 update_load_set(&se->load, weight);
2416 account_entity_enqueue(cfs_rq, se);
2419 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2421 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2423 struct task_group *tg;
2424 struct sched_entity *se;
2428 se = tg->se[cpu_of(rq_of(cfs_rq))];
2429 if (!se || throttled_hierarchy(cfs_rq))
2432 if (likely(se->load.weight == tg->shares))
2435 shares = calc_cfs_shares(cfs_rq, tg);
2437 reweight_entity(cfs_rq_of(se), se, shares);
2439 #else /* CONFIG_FAIR_GROUP_SCHED */
2440 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2443 #endif /* CONFIG_FAIR_GROUP_SCHED */
2446 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2447 static const u32 runnable_avg_yN_inv[] = {
2448 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2449 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2450 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2451 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2452 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2453 0x85aac367, 0x82cd8698,
2457 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2458 * over-estimates when re-combining.
2460 static const u32 runnable_avg_yN_sum[] = {
2461 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2462 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2463 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2468 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2470 static __always_inline u64 decay_load(u64 val, u64 n)
2472 unsigned int local_n;
2476 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2479 /* after bounds checking we can collapse to 32-bit */
2483 * As y^PERIOD = 1/2, we can combine
2484 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2485 * With a look-up table which covers y^n (n<PERIOD)
2487 * To achieve constant time decay_load.
2489 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2490 val >>= local_n / LOAD_AVG_PERIOD;
2491 local_n %= LOAD_AVG_PERIOD;
2494 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2499 * For updates fully spanning n periods, the contribution to runnable
2500 * average will be: \Sum 1024*y^n
2502 * We can compute this reasonably efficiently by combining:
2503 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2505 static u32 __compute_runnable_contrib(u64 n)
2509 if (likely(n <= LOAD_AVG_PERIOD))
2510 return runnable_avg_yN_sum[n];
2511 else if (unlikely(n >= LOAD_AVG_MAX_N))
2512 return LOAD_AVG_MAX;
2514 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2516 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2517 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2519 n -= LOAD_AVG_PERIOD;
2520 } while (n > LOAD_AVG_PERIOD);
2522 contrib = decay_load(contrib, n);
2523 return contrib + runnable_avg_yN_sum[n];
2526 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2527 #error "load tracking assumes 2^10 as unit"
2530 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2533 * We can represent the historical contribution to runnable average as the
2534 * coefficients of a geometric series. To do this we sub-divide our runnable
2535 * history into segments of approximately 1ms (1024us); label the segment that
2536 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2538 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2540 * (now) (~1ms ago) (~2ms ago)
2542 * Let u_i denote the fraction of p_i that the entity was runnable.
2544 * We then designate the fractions u_i as our co-efficients, yielding the
2545 * following representation of historical load:
2546 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2548 * We choose y based on the with of a reasonably scheduling period, fixing:
2551 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2552 * approximately half as much as the contribution to load within the last ms
2555 * When a period "rolls over" and we have new u_0`, multiplying the previous
2556 * sum again by y is sufficient to update:
2557 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2558 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2560 static __always_inline int
2561 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2562 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2564 u64 delta, scaled_delta, periods;
2566 unsigned int delta_w, scaled_delta_w, decayed = 0;
2567 unsigned long scale_freq, scale_cpu;
2569 delta = now - sa->last_update_time;
2571 * This should only happen when time goes backwards, which it
2572 * unfortunately does during sched clock init when we swap over to TSC.
2574 if ((s64)delta < 0) {
2575 sa->last_update_time = now;
2580 * Use 1024ns as the unit of measurement since it's a reasonable
2581 * approximation of 1us and fast to compute.
2586 sa->last_update_time = now;
2588 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2589 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2590 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2592 /* delta_w is the amount already accumulated against our next period */
2593 delta_w = sa->period_contrib;
2594 if (delta + delta_w >= 1024) {
2597 /* how much left for next period will start over, we don't know yet */
2598 sa->period_contrib = 0;
2601 * Now that we know we're crossing a period boundary, figure
2602 * out how much from delta we need to complete the current
2603 * period and accrue it.
2605 delta_w = 1024 - delta_w;
2606 scaled_delta_w = cap_scale(delta_w, scale_freq);
2608 sa->load_sum += weight * scaled_delta_w;
2610 cfs_rq->runnable_load_sum +=
2611 weight * scaled_delta_w;
2615 sa->util_sum += scaled_delta_w * scale_cpu;
2619 /* Figure out how many additional periods this update spans */
2620 periods = delta / 1024;
2623 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2625 cfs_rq->runnable_load_sum =
2626 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2628 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2630 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2631 contrib = __compute_runnable_contrib(periods);
2632 contrib = cap_scale(contrib, scale_freq);
2634 sa->load_sum += weight * contrib;
2636 cfs_rq->runnable_load_sum += weight * contrib;
2639 sa->util_sum += contrib * scale_cpu;
2642 /* Remainder of delta accrued against u_0` */
2643 scaled_delta = cap_scale(delta, scale_freq);
2645 sa->load_sum += weight * scaled_delta;
2647 cfs_rq->runnable_load_sum += weight * scaled_delta;
2650 sa->util_sum += scaled_delta * scale_cpu;
2652 sa->period_contrib += delta;
2655 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2657 cfs_rq->runnable_load_avg =
2658 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2660 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2666 #ifdef CONFIG_FAIR_GROUP_SCHED
2668 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2669 * and effective_load (which is not done because it is too costly).
2671 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2673 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2675 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2676 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2677 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2681 #else /* CONFIG_FAIR_GROUP_SCHED */
2682 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2683 #endif /* CONFIG_FAIR_GROUP_SCHED */
2685 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2687 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2688 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2690 struct sched_avg *sa = &cfs_rq->avg;
2691 int decayed, removed = 0;
2693 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2694 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2695 sa->load_avg = max_t(long, sa->load_avg - r, 0);
2696 sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
2700 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2701 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2702 sa->util_avg = max_t(long, sa->util_avg - r, 0);
2703 sa->util_sum = max_t(s32, sa->util_sum - r * LOAD_AVG_MAX, 0);
2706 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2707 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2709 #ifndef CONFIG_64BIT
2711 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2714 return decayed || removed;
2717 /* Update task and its cfs_rq load average */
2718 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2720 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2721 u64 now = cfs_rq_clock_task(cfs_rq);
2722 int cpu = cpu_of(rq_of(cfs_rq));
2725 * Track task load average for carrying it to new CPU after migrated, and
2726 * track group sched_entity load average for task_h_load calc in migration
2728 __update_load_avg(now, cpu, &se->avg,
2729 se->on_rq * scale_load_down(se->load.weight),
2730 cfs_rq->curr == se, NULL);
2732 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2733 update_tg_load_avg(cfs_rq, 0);
2735 if (entity_is_task(se))
2736 trace_sched_load_avg_task(task_of(se), &se->avg);
2737 trace_sched_load_avg_cpu(cpu, cfs_rq);
2740 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2742 if (!sched_feat(ATTACH_AGE_LOAD))
2746 * If we got migrated (either between CPUs or between cgroups) we'll
2747 * have aged the average right before clearing @last_update_time.
2749 if (se->avg.last_update_time) {
2750 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2751 &se->avg, 0, 0, NULL);
2754 * XXX: we could have just aged the entire load away if we've been
2755 * absent from the fair class for too long.
2760 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2761 cfs_rq->avg.load_avg += se->avg.load_avg;
2762 cfs_rq->avg.load_sum += se->avg.load_sum;
2763 cfs_rq->avg.util_avg += se->avg.util_avg;
2764 cfs_rq->avg.util_sum += se->avg.util_sum;
2767 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2769 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2770 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2771 cfs_rq->curr == se, NULL);
2773 cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
2774 cfs_rq->avg.load_sum = max_t(s64, cfs_rq->avg.load_sum - se->avg.load_sum, 0);
2775 cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
2776 cfs_rq->avg.util_sum = max_t(s32, cfs_rq->avg.util_sum - se->avg.util_sum, 0);
2779 /* Add the load generated by se into cfs_rq's load average */
2781 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2783 struct sched_avg *sa = &se->avg;
2784 u64 now = cfs_rq_clock_task(cfs_rq);
2785 int migrated, decayed;
2787 migrated = !sa->last_update_time;
2789 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2790 se->on_rq * scale_load_down(se->load.weight),
2791 cfs_rq->curr == se, NULL);
2794 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2796 cfs_rq->runnable_load_avg += sa->load_avg;
2797 cfs_rq->runnable_load_sum += sa->load_sum;
2800 attach_entity_load_avg(cfs_rq, se);
2802 if (decayed || migrated)
2803 update_tg_load_avg(cfs_rq, 0);
2806 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2808 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2810 update_load_avg(se, 1);
2812 cfs_rq->runnable_load_avg =
2813 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2814 cfs_rq->runnable_load_sum =
2815 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2818 #ifndef CONFIG_64BIT
2819 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2821 u64 last_update_time_copy;
2822 u64 last_update_time;
2825 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2827 last_update_time = cfs_rq->avg.last_update_time;
2828 } while (last_update_time != last_update_time_copy);
2830 return last_update_time;
2833 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2835 return cfs_rq->avg.last_update_time;
2840 * Task first catches up with cfs_rq, and then subtract
2841 * itself from the cfs_rq (task must be off the queue now).
2843 void remove_entity_load_avg(struct sched_entity *se)
2845 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2846 u64 last_update_time;
2849 * Newly created task or never used group entity should not be removed
2850 * from its (source) cfs_rq
2852 if (se->avg.last_update_time == 0)
2855 last_update_time = cfs_rq_last_update_time(cfs_rq);
2857 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2858 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2859 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2863 * Update the rq's load with the elapsed running time before entering
2864 * idle. if the last scheduled task is not a CFS task, idle_enter will
2865 * be the only way to update the runnable statistic.
2867 void idle_enter_fair(struct rq *this_rq)
2872 * Update the rq's load with the elapsed idle time before a task is
2873 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2874 * be the only way to update the runnable statistic.
2876 void idle_exit_fair(struct rq *this_rq)
2880 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2882 return cfs_rq->runnable_load_avg;
2885 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2887 return cfs_rq->avg.load_avg;
2890 static int idle_balance(struct rq *this_rq);
2892 #else /* CONFIG_SMP */
2894 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2896 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2898 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2899 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2902 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2904 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2906 static inline int idle_balance(struct rq *rq)
2911 #endif /* CONFIG_SMP */
2913 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2915 #ifdef CONFIG_SCHEDSTATS
2916 struct task_struct *tsk = NULL;
2918 if (entity_is_task(se))
2921 if (se->statistics.sleep_start) {
2922 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2927 if (unlikely(delta > se->statistics.sleep_max))
2928 se->statistics.sleep_max = delta;
2930 se->statistics.sleep_start = 0;
2931 se->statistics.sum_sleep_runtime += delta;
2934 account_scheduler_latency(tsk, delta >> 10, 1);
2935 trace_sched_stat_sleep(tsk, delta);
2938 if (se->statistics.block_start) {
2939 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2944 if (unlikely(delta > se->statistics.block_max))
2945 se->statistics.block_max = delta;
2947 se->statistics.block_start = 0;
2948 se->statistics.sum_sleep_runtime += delta;
2951 if (tsk->in_iowait) {
2952 se->statistics.iowait_sum += delta;
2953 se->statistics.iowait_count++;
2954 trace_sched_stat_iowait(tsk, delta);
2957 trace_sched_stat_blocked(tsk, delta);
2958 trace_sched_blocked_reason(tsk);
2961 * Blocking time is in units of nanosecs, so shift by
2962 * 20 to get a milliseconds-range estimation of the
2963 * amount of time that the task spent sleeping:
2965 if (unlikely(prof_on == SLEEP_PROFILING)) {
2966 profile_hits(SLEEP_PROFILING,
2967 (void *)get_wchan(tsk),
2970 account_scheduler_latency(tsk, delta >> 10, 0);
2976 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2978 #ifdef CONFIG_SCHED_DEBUG
2979 s64 d = se->vruntime - cfs_rq->min_vruntime;
2984 if (d > 3*sysctl_sched_latency)
2985 schedstat_inc(cfs_rq, nr_spread_over);
2990 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2992 u64 vruntime = cfs_rq->min_vruntime;
2995 * The 'current' period is already promised to the current tasks,
2996 * however the extra weight of the new task will slow them down a
2997 * little, place the new task so that it fits in the slot that
2998 * stays open at the end.
3000 if (initial && sched_feat(START_DEBIT))
3001 vruntime += sched_vslice(cfs_rq, se);
3003 /* sleeps up to a single latency don't count. */
3005 unsigned long thresh = sysctl_sched_latency;
3008 * Halve their sleep time's effect, to allow
3009 * for a gentler effect of sleepers:
3011 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3017 /* ensure we never gain time by being placed backwards. */
3018 se->vruntime = max_vruntime(se->vruntime, vruntime);
3021 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3024 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3027 * Update the normalized vruntime before updating min_vruntime
3028 * through calling update_curr().
3030 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3031 se->vruntime += cfs_rq->min_vruntime;
3034 * Update run-time statistics of the 'current'.
3036 update_curr(cfs_rq);
3037 enqueue_entity_load_avg(cfs_rq, se);
3038 account_entity_enqueue(cfs_rq, se);
3039 update_cfs_shares(cfs_rq);
3041 if (flags & ENQUEUE_WAKEUP) {
3042 place_entity(cfs_rq, se, 0);
3043 enqueue_sleeper(cfs_rq, se);
3046 update_stats_enqueue(cfs_rq, se);
3047 check_spread(cfs_rq, se);
3048 if (se != cfs_rq->curr)
3049 __enqueue_entity(cfs_rq, se);
3052 if (cfs_rq->nr_running == 1) {
3053 list_add_leaf_cfs_rq(cfs_rq);
3054 check_enqueue_throttle(cfs_rq);
3058 static void __clear_buddies_last(struct sched_entity *se)
3060 for_each_sched_entity(se) {
3061 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3062 if (cfs_rq->last != se)
3065 cfs_rq->last = NULL;
3069 static void __clear_buddies_next(struct sched_entity *se)
3071 for_each_sched_entity(se) {
3072 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3073 if (cfs_rq->next != se)
3076 cfs_rq->next = NULL;
3080 static void __clear_buddies_skip(struct sched_entity *se)
3082 for_each_sched_entity(se) {
3083 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3084 if (cfs_rq->skip != se)
3087 cfs_rq->skip = NULL;
3091 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3093 if (cfs_rq->last == se)
3094 __clear_buddies_last(se);
3096 if (cfs_rq->next == se)
3097 __clear_buddies_next(se);
3099 if (cfs_rq->skip == se)
3100 __clear_buddies_skip(se);
3103 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3106 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3109 * Update run-time statistics of the 'current'.
3111 update_curr(cfs_rq);
3112 dequeue_entity_load_avg(cfs_rq, se);
3114 update_stats_dequeue(cfs_rq, se);
3115 if (flags & DEQUEUE_SLEEP) {
3116 #ifdef CONFIG_SCHEDSTATS
3117 if (entity_is_task(se)) {
3118 struct task_struct *tsk = task_of(se);
3120 if (tsk->state & TASK_INTERRUPTIBLE)
3121 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3122 if (tsk->state & TASK_UNINTERRUPTIBLE)
3123 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3128 clear_buddies(cfs_rq, se);
3130 if (se != cfs_rq->curr)
3131 __dequeue_entity(cfs_rq, se);
3133 account_entity_dequeue(cfs_rq, se);
3136 * Normalize the entity after updating the min_vruntime because the
3137 * update can refer to the ->curr item and we need to reflect this
3138 * movement in our normalized position.
3140 if (!(flags & DEQUEUE_SLEEP))
3141 se->vruntime -= cfs_rq->min_vruntime;
3143 /* return excess runtime on last dequeue */
3144 return_cfs_rq_runtime(cfs_rq);
3146 update_min_vruntime(cfs_rq);
3147 update_cfs_shares(cfs_rq);
3151 * Preempt the current task with a newly woken task if needed:
3154 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3156 unsigned long ideal_runtime, delta_exec;
3157 struct sched_entity *se;
3160 ideal_runtime = sched_slice(cfs_rq, curr);
3161 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3162 if (delta_exec > ideal_runtime) {
3163 resched_curr(rq_of(cfs_rq));
3165 * The current task ran long enough, ensure it doesn't get
3166 * re-elected due to buddy favours.
3168 clear_buddies(cfs_rq, curr);
3173 * Ensure that a task that missed wakeup preemption by a
3174 * narrow margin doesn't have to wait for a full slice.
3175 * This also mitigates buddy induced latencies under load.
3177 if (delta_exec < sysctl_sched_min_granularity)
3180 se = __pick_first_entity(cfs_rq);
3181 delta = curr->vruntime - se->vruntime;
3186 if (delta > ideal_runtime)
3187 resched_curr(rq_of(cfs_rq));
3191 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3193 /* 'current' is not kept within the tree. */
3196 * Any task has to be enqueued before it get to execute on
3197 * a CPU. So account for the time it spent waiting on the
3200 update_stats_wait_end(cfs_rq, se);
3201 __dequeue_entity(cfs_rq, se);
3202 update_load_avg(se, 1);
3205 update_stats_curr_start(cfs_rq, se);
3207 #ifdef CONFIG_SCHEDSTATS
3209 * Track our maximum slice length, if the CPU's load is at
3210 * least twice that of our own weight (i.e. dont track it
3211 * when there are only lesser-weight tasks around):
3213 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3214 se->statistics.slice_max = max(se->statistics.slice_max,
3215 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3218 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3222 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3225 * Pick the next process, keeping these things in mind, in this order:
3226 * 1) keep things fair between processes/task groups
3227 * 2) pick the "next" process, since someone really wants that to run
3228 * 3) pick the "last" process, for cache locality
3229 * 4) do not run the "skip" process, if something else is available
3231 static struct sched_entity *
3232 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3234 struct sched_entity *left = __pick_first_entity(cfs_rq);
3235 struct sched_entity *se;
3238 * If curr is set we have to see if its left of the leftmost entity
3239 * still in the tree, provided there was anything in the tree at all.
3241 if (!left || (curr && entity_before(curr, left)))
3244 se = left; /* ideally we run the leftmost entity */
3247 * Avoid running the skip buddy, if running something else can
3248 * be done without getting too unfair.
3250 if (cfs_rq->skip == se) {
3251 struct sched_entity *second;
3254 second = __pick_first_entity(cfs_rq);
3256 second = __pick_next_entity(se);
3257 if (!second || (curr && entity_before(curr, second)))
3261 if (second && wakeup_preempt_entity(second, left) < 1)
3266 * Prefer last buddy, try to return the CPU to a preempted task.
3268 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3272 * Someone really wants this to run. If it's not unfair, run it.
3274 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3277 clear_buddies(cfs_rq, se);
3282 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3284 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3287 * If still on the runqueue then deactivate_task()
3288 * was not called and update_curr() has to be done:
3291 update_curr(cfs_rq);
3293 /* throttle cfs_rqs exceeding runtime */
3294 check_cfs_rq_runtime(cfs_rq);
3296 check_spread(cfs_rq, prev);
3298 update_stats_wait_start(cfs_rq, prev);
3299 /* Put 'current' back into the tree. */
3300 __enqueue_entity(cfs_rq, prev);
3301 /* in !on_rq case, update occurred at dequeue */
3302 update_load_avg(prev, 0);
3304 cfs_rq->curr = NULL;
3308 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3311 * Update run-time statistics of the 'current'.
3313 update_curr(cfs_rq);
3316 * Ensure that runnable average is periodically updated.
3318 update_load_avg(curr, 1);
3319 update_cfs_shares(cfs_rq);
3321 #ifdef CONFIG_SCHED_HRTICK
3323 * queued ticks are scheduled to match the slice, so don't bother
3324 * validating it and just reschedule.
3327 resched_curr(rq_of(cfs_rq));
3331 * don't let the period tick interfere with the hrtick preemption
3333 if (!sched_feat(DOUBLE_TICK) &&
3334 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3338 if (cfs_rq->nr_running > 1)
3339 check_preempt_tick(cfs_rq, curr);
3343 /**************************************************
3344 * CFS bandwidth control machinery
3347 #ifdef CONFIG_CFS_BANDWIDTH
3349 #ifdef HAVE_JUMP_LABEL
3350 static struct static_key __cfs_bandwidth_used;
3352 static inline bool cfs_bandwidth_used(void)
3354 return static_key_false(&__cfs_bandwidth_used);
3357 void cfs_bandwidth_usage_inc(void)
3359 static_key_slow_inc(&__cfs_bandwidth_used);
3362 void cfs_bandwidth_usage_dec(void)
3364 static_key_slow_dec(&__cfs_bandwidth_used);
3366 #else /* HAVE_JUMP_LABEL */
3367 static bool cfs_bandwidth_used(void)
3372 void cfs_bandwidth_usage_inc(void) {}
3373 void cfs_bandwidth_usage_dec(void) {}
3374 #endif /* HAVE_JUMP_LABEL */
3377 * default period for cfs group bandwidth.
3378 * default: 0.1s, units: nanoseconds
3380 static inline u64 default_cfs_period(void)
3382 return 100000000ULL;
3385 static inline u64 sched_cfs_bandwidth_slice(void)
3387 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3391 * Replenish runtime according to assigned quota and update expiration time.
3392 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3393 * additional synchronization around rq->lock.
3395 * requires cfs_b->lock
3397 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3401 if (cfs_b->quota == RUNTIME_INF)
3404 now = sched_clock_cpu(smp_processor_id());
3405 cfs_b->runtime = cfs_b->quota;
3406 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3409 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3411 return &tg->cfs_bandwidth;
3414 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3415 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3417 if (unlikely(cfs_rq->throttle_count))
3418 return cfs_rq->throttled_clock_task;
3420 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3423 /* returns 0 on failure to allocate runtime */
3424 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3426 struct task_group *tg = cfs_rq->tg;
3427 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3428 u64 amount = 0, min_amount, expires;
3430 /* note: this is a positive sum as runtime_remaining <= 0 */
3431 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3433 raw_spin_lock(&cfs_b->lock);
3434 if (cfs_b->quota == RUNTIME_INF)
3435 amount = min_amount;
3437 start_cfs_bandwidth(cfs_b);
3439 if (cfs_b->runtime > 0) {
3440 amount = min(cfs_b->runtime, min_amount);
3441 cfs_b->runtime -= amount;
3445 expires = cfs_b->runtime_expires;
3446 raw_spin_unlock(&cfs_b->lock);
3448 cfs_rq->runtime_remaining += amount;
3450 * we may have advanced our local expiration to account for allowed
3451 * spread between our sched_clock and the one on which runtime was
3454 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3455 cfs_rq->runtime_expires = expires;
3457 return cfs_rq->runtime_remaining > 0;
3461 * Note: This depends on the synchronization provided by sched_clock and the
3462 * fact that rq->clock snapshots this value.
3464 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3466 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3468 /* if the deadline is ahead of our clock, nothing to do */
3469 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3472 if (cfs_rq->runtime_remaining < 0)
3476 * If the local deadline has passed we have to consider the
3477 * possibility that our sched_clock is 'fast' and the global deadline
3478 * has not truly expired.
3480 * Fortunately we can check determine whether this the case by checking
3481 * whether the global deadline has advanced. It is valid to compare
3482 * cfs_b->runtime_expires without any locks since we only care about
3483 * exact equality, so a partial write will still work.
3486 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3487 /* extend local deadline, drift is bounded above by 2 ticks */
3488 cfs_rq->runtime_expires += TICK_NSEC;
3490 /* global deadline is ahead, expiration has passed */
3491 cfs_rq->runtime_remaining = 0;
3495 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3497 /* dock delta_exec before expiring quota (as it could span periods) */
3498 cfs_rq->runtime_remaining -= delta_exec;
3499 expire_cfs_rq_runtime(cfs_rq);
3501 if (likely(cfs_rq->runtime_remaining > 0))
3505 * if we're unable to extend our runtime we resched so that the active
3506 * hierarchy can be throttled
3508 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3509 resched_curr(rq_of(cfs_rq));
3512 static __always_inline
3513 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3515 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3518 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3521 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3523 return cfs_bandwidth_used() && cfs_rq->throttled;
3526 /* check whether cfs_rq, or any parent, is throttled */
3527 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3529 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3533 * Ensure that neither of the group entities corresponding to src_cpu or
3534 * dest_cpu are members of a throttled hierarchy when performing group
3535 * load-balance operations.
3537 static inline int throttled_lb_pair(struct task_group *tg,
3538 int src_cpu, int dest_cpu)
3540 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3542 src_cfs_rq = tg->cfs_rq[src_cpu];
3543 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3545 return throttled_hierarchy(src_cfs_rq) ||
3546 throttled_hierarchy(dest_cfs_rq);
3549 /* updated child weight may affect parent so we have to do this bottom up */
3550 static int tg_unthrottle_up(struct task_group *tg, void *data)
3552 struct rq *rq = data;
3553 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3555 cfs_rq->throttle_count--;
3557 if (!cfs_rq->throttle_count) {
3558 /* adjust cfs_rq_clock_task() */
3559 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3560 cfs_rq->throttled_clock_task;
3567 static int tg_throttle_down(struct task_group *tg, void *data)
3569 struct rq *rq = data;
3570 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3572 /* group is entering throttled state, stop time */
3573 if (!cfs_rq->throttle_count)
3574 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3575 cfs_rq->throttle_count++;
3580 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3582 struct rq *rq = rq_of(cfs_rq);
3583 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3584 struct sched_entity *se;
3585 long task_delta, dequeue = 1;
3588 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3590 /* freeze hierarchy runnable averages while throttled */
3592 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3595 task_delta = cfs_rq->h_nr_running;
3596 for_each_sched_entity(se) {
3597 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3598 /* throttled entity or throttle-on-deactivate */
3603 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3604 qcfs_rq->h_nr_running -= task_delta;
3606 if (qcfs_rq->load.weight)
3611 sub_nr_running(rq, task_delta);
3613 cfs_rq->throttled = 1;
3614 cfs_rq->throttled_clock = rq_clock(rq);
3615 raw_spin_lock(&cfs_b->lock);
3616 empty = list_empty(&cfs_b->throttled_cfs_rq);
3619 * Add to the _head_ of the list, so that an already-started
3620 * distribute_cfs_runtime will not see us
3622 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3625 * If we're the first throttled task, make sure the bandwidth
3629 start_cfs_bandwidth(cfs_b);
3631 raw_spin_unlock(&cfs_b->lock);
3634 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3636 struct rq *rq = rq_of(cfs_rq);
3637 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3638 struct sched_entity *se;
3642 se = cfs_rq->tg->se[cpu_of(rq)];
3644 cfs_rq->throttled = 0;
3646 update_rq_clock(rq);
3648 raw_spin_lock(&cfs_b->lock);
3649 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3650 list_del_rcu(&cfs_rq->throttled_list);
3651 raw_spin_unlock(&cfs_b->lock);
3653 /* update hierarchical throttle state */
3654 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3656 if (!cfs_rq->load.weight)
3659 task_delta = cfs_rq->h_nr_running;
3660 for_each_sched_entity(se) {
3664 cfs_rq = cfs_rq_of(se);
3666 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3667 cfs_rq->h_nr_running += task_delta;
3669 if (cfs_rq_throttled(cfs_rq))
3674 add_nr_running(rq, task_delta);
3676 /* determine whether we need to wake up potentially idle cpu */
3677 if (rq->curr == rq->idle && rq->cfs.nr_running)
3681 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3682 u64 remaining, u64 expires)
3684 struct cfs_rq *cfs_rq;
3686 u64 starting_runtime = remaining;
3689 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3691 struct rq *rq = rq_of(cfs_rq);
3693 raw_spin_lock(&rq->lock);
3694 if (!cfs_rq_throttled(cfs_rq))
3697 runtime = -cfs_rq->runtime_remaining + 1;
3698 if (runtime > remaining)
3699 runtime = remaining;
3700 remaining -= runtime;
3702 cfs_rq->runtime_remaining += runtime;
3703 cfs_rq->runtime_expires = expires;
3705 /* we check whether we're throttled above */
3706 if (cfs_rq->runtime_remaining > 0)
3707 unthrottle_cfs_rq(cfs_rq);
3710 raw_spin_unlock(&rq->lock);
3717 return starting_runtime - remaining;
3721 * Responsible for refilling a task_group's bandwidth and unthrottling its
3722 * cfs_rqs as appropriate. If there has been no activity within the last
3723 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3724 * used to track this state.
3726 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3728 u64 runtime, runtime_expires;
3731 /* no need to continue the timer with no bandwidth constraint */
3732 if (cfs_b->quota == RUNTIME_INF)
3733 goto out_deactivate;
3735 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3736 cfs_b->nr_periods += overrun;
3739 * idle depends on !throttled (for the case of a large deficit), and if
3740 * we're going inactive then everything else can be deferred
3742 if (cfs_b->idle && !throttled)
3743 goto out_deactivate;
3745 __refill_cfs_bandwidth_runtime(cfs_b);
3748 /* mark as potentially idle for the upcoming period */
3753 /* account preceding periods in which throttling occurred */
3754 cfs_b->nr_throttled += overrun;
3756 runtime_expires = cfs_b->runtime_expires;
3759 * This check is repeated as we are holding onto the new bandwidth while
3760 * we unthrottle. This can potentially race with an unthrottled group
3761 * trying to acquire new bandwidth from the global pool. This can result
3762 * in us over-using our runtime if it is all used during this loop, but
3763 * only by limited amounts in that extreme case.
3765 while (throttled && cfs_b->runtime > 0) {
3766 runtime = cfs_b->runtime;
3767 raw_spin_unlock(&cfs_b->lock);
3768 /* we can't nest cfs_b->lock while distributing bandwidth */
3769 runtime = distribute_cfs_runtime(cfs_b, runtime,
3771 raw_spin_lock(&cfs_b->lock);
3773 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3775 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3779 * While we are ensured activity in the period following an
3780 * unthrottle, this also covers the case in which the new bandwidth is
3781 * insufficient to cover the existing bandwidth deficit. (Forcing the
3782 * timer to remain active while there are any throttled entities.)
3792 /* a cfs_rq won't donate quota below this amount */
3793 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3794 /* minimum remaining period time to redistribute slack quota */
3795 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3796 /* how long we wait to gather additional slack before distributing */
3797 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3800 * Are we near the end of the current quota period?
3802 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3803 * hrtimer base being cleared by hrtimer_start. In the case of
3804 * migrate_hrtimers, base is never cleared, so we are fine.
3806 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3808 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3811 /* if the call-back is running a quota refresh is already occurring */
3812 if (hrtimer_callback_running(refresh_timer))
3815 /* is a quota refresh about to occur? */
3816 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3817 if (remaining < min_expire)
3823 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3825 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3827 /* if there's a quota refresh soon don't bother with slack */
3828 if (runtime_refresh_within(cfs_b, min_left))
3831 hrtimer_start(&cfs_b->slack_timer,
3832 ns_to_ktime(cfs_bandwidth_slack_period),
3836 /* we know any runtime found here is valid as update_curr() precedes return */
3837 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3839 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3840 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3842 if (slack_runtime <= 0)
3845 raw_spin_lock(&cfs_b->lock);
3846 if (cfs_b->quota != RUNTIME_INF &&
3847 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3848 cfs_b->runtime += slack_runtime;
3850 /* we are under rq->lock, defer unthrottling using a timer */
3851 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3852 !list_empty(&cfs_b->throttled_cfs_rq))
3853 start_cfs_slack_bandwidth(cfs_b);
3855 raw_spin_unlock(&cfs_b->lock);
3857 /* even if it's not valid for return we don't want to try again */
3858 cfs_rq->runtime_remaining -= slack_runtime;
3861 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3863 if (!cfs_bandwidth_used())
3866 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3869 __return_cfs_rq_runtime(cfs_rq);
3873 * This is done with a timer (instead of inline with bandwidth return) since
3874 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3876 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3878 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3881 /* confirm we're still not at a refresh boundary */
3882 raw_spin_lock(&cfs_b->lock);
3883 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3884 raw_spin_unlock(&cfs_b->lock);
3888 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3889 runtime = cfs_b->runtime;
3891 expires = cfs_b->runtime_expires;
3892 raw_spin_unlock(&cfs_b->lock);
3897 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3899 raw_spin_lock(&cfs_b->lock);
3900 if (expires == cfs_b->runtime_expires)
3901 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3902 raw_spin_unlock(&cfs_b->lock);
3906 * When a group wakes up we want to make sure that its quota is not already
3907 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3908 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3910 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3912 if (!cfs_bandwidth_used())
3915 /* an active group must be handled by the update_curr()->put() path */
3916 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3919 /* ensure the group is not already throttled */
3920 if (cfs_rq_throttled(cfs_rq))
3923 /* update runtime allocation */
3924 account_cfs_rq_runtime(cfs_rq, 0);
3925 if (cfs_rq->runtime_remaining <= 0)
3926 throttle_cfs_rq(cfs_rq);
3929 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3930 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3932 if (!cfs_bandwidth_used())
3935 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3939 * it's possible for a throttled entity to be forced into a running
3940 * state (e.g. set_curr_task), in this case we're finished.
3942 if (cfs_rq_throttled(cfs_rq))
3945 throttle_cfs_rq(cfs_rq);
3949 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3951 struct cfs_bandwidth *cfs_b =
3952 container_of(timer, struct cfs_bandwidth, slack_timer);
3954 do_sched_cfs_slack_timer(cfs_b);
3956 return HRTIMER_NORESTART;
3959 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3961 struct cfs_bandwidth *cfs_b =
3962 container_of(timer, struct cfs_bandwidth, period_timer);
3966 raw_spin_lock(&cfs_b->lock);
3968 overrun = hrtimer_forward_now(timer, cfs_b->period);
3972 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3975 cfs_b->period_active = 0;
3976 raw_spin_unlock(&cfs_b->lock);
3978 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3981 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3983 raw_spin_lock_init(&cfs_b->lock);
3985 cfs_b->quota = RUNTIME_INF;
3986 cfs_b->period = ns_to_ktime(default_cfs_period());
3988 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3989 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
3990 cfs_b->period_timer.function = sched_cfs_period_timer;
3991 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3992 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3995 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3997 cfs_rq->runtime_enabled = 0;
3998 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4001 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4003 lockdep_assert_held(&cfs_b->lock);
4005 if (!cfs_b->period_active) {
4006 cfs_b->period_active = 1;
4007 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4008 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4012 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4014 /* init_cfs_bandwidth() was not called */
4015 if (!cfs_b->throttled_cfs_rq.next)
4018 hrtimer_cancel(&cfs_b->period_timer);
4019 hrtimer_cancel(&cfs_b->slack_timer);
4022 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4024 struct cfs_rq *cfs_rq;
4026 for_each_leaf_cfs_rq(rq, cfs_rq) {
4027 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4029 raw_spin_lock(&cfs_b->lock);
4030 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4031 raw_spin_unlock(&cfs_b->lock);
4035 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4037 struct cfs_rq *cfs_rq;
4039 for_each_leaf_cfs_rq(rq, cfs_rq) {
4040 if (!cfs_rq->runtime_enabled)
4044 * clock_task is not advancing so we just need to make sure
4045 * there's some valid quota amount
4047 cfs_rq->runtime_remaining = 1;
4049 * Offline rq is schedulable till cpu is completely disabled
4050 * in take_cpu_down(), so we prevent new cfs throttling here.
4052 cfs_rq->runtime_enabled = 0;
4054 if (cfs_rq_throttled(cfs_rq))
4055 unthrottle_cfs_rq(cfs_rq);
4059 #else /* CONFIG_CFS_BANDWIDTH */
4060 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4062 return rq_clock_task(rq_of(cfs_rq));
4065 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4066 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4067 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4068 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4070 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4075 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4080 static inline int throttled_lb_pair(struct task_group *tg,
4081 int src_cpu, int dest_cpu)
4086 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4088 #ifdef CONFIG_FAIR_GROUP_SCHED
4089 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4092 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4096 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4097 static inline void update_runtime_enabled(struct rq *rq) {}
4098 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4100 #endif /* CONFIG_CFS_BANDWIDTH */
4102 /**************************************************
4103 * CFS operations on tasks:
4106 #ifdef CONFIG_SCHED_HRTICK
4107 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4109 struct sched_entity *se = &p->se;
4110 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4112 WARN_ON(task_rq(p) != rq);
4114 if (cfs_rq->nr_running > 1) {
4115 u64 slice = sched_slice(cfs_rq, se);
4116 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4117 s64 delta = slice - ran;
4124 hrtick_start(rq, delta);
4129 * called from enqueue/dequeue and updates the hrtick when the
4130 * current task is from our class and nr_running is low enough
4133 static void hrtick_update(struct rq *rq)
4135 struct task_struct *curr = rq->curr;
4137 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4140 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4141 hrtick_start_fair(rq, curr);
4143 #else /* !CONFIG_SCHED_HRTICK */
4145 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4149 static inline void hrtick_update(struct rq *rq)
4154 static inline unsigned long boosted_cpu_util(int cpu);
4156 static void update_capacity_of(int cpu)
4158 unsigned long req_cap;
4163 /* Convert scale-invariant capacity to cpu. */
4164 req_cap = boosted_cpu_util(cpu);
4165 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4166 set_cfs_cpu_capacity(cpu, true, req_cap);
4169 static bool cpu_overutilized(int cpu);
4172 * The enqueue_task method is called before nr_running is
4173 * increased. Here we update the fair scheduling stats and
4174 * then put the task into the rbtree:
4177 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4179 struct cfs_rq *cfs_rq;
4180 struct sched_entity *se = &p->se;
4181 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4182 int task_wakeup = flags & ENQUEUE_WAKEUP;
4184 for_each_sched_entity(se) {
4187 cfs_rq = cfs_rq_of(se);
4188 enqueue_entity(cfs_rq, se, flags);
4191 * end evaluation on encountering a throttled cfs_rq
4193 * note: in the case of encountering a throttled cfs_rq we will
4194 * post the final h_nr_running increment below.
4196 if (cfs_rq_throttled(cfs_rq))
4198 cfs_rq->h_nr_running++;
4200 flags = ENQUEUE_WAKEUP;
4203 for_each_sched_entity(se) {
4204 cfs_rq = cfs_rq_of(se);
4205 cfs_rq->h_nr_running++;
4207 if (cfs_rq_throttled(cfs_rq))
4210 update_load_avg(se, 1);
4211 update_cfs_shares(cfs_rq);
4215 add_nr_running(rq, 1);
4216 if (!task_new && !rq->rd->overutilized &&
4217 cpu_overutilized(rq->cpu))
4218 rq->rd->overutilized = true;
4220 schedtune_enqueue_task(p, cpu_of(rq));
4223 * We want to potentially trigger a freq switch
4224 * request only for tasks that are waking up; this is
4225 * because we get here also during load balancing, but
4226 * in these cases it seems wise to trigger as single
4227 * request after load balancing is done.
4229 if (task_new || task_wakeup)
4230 update_capacity_of(cpu_of(rq));
4235 static void set_next_buddy(struct sched_entity *se);
4238 * The dequeue_task method is called before nr_running is
4239 * decreased. We remove the task from the rbtree and
4240 * update the fair scheduling stats:
4242 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4244 struct cfs_rq *cfs_rq;
4245 struct sched_entity *se = &p->se;
4246 int task_sleep = flags & DEQUEUE_SLEEP;
4248 for_each_sched_entity(se) {
4249 cfs_rq = cfs_rq_of(se);
4250 dequeue_entity(cfs_rq, se, flags);
4253 * end evaluation on encountering a throttled cfs_rq
4255 * note: in the case of encountering a throttled cfs_rq we will
4256 * post the final h_nr_running decrement below.
4258 if (cfs_rq_throttled(cfs_rq))
4260 cfs_rq->h_nr_running--;
4262 /* Don't dequeue parent if it has other entities besides us */
4263 if (cfs_rq->load.weight) {
4265 * Bias pick_next to pick a task from this cfs_rq, as
4266 * p is sleeping when it is within its sched_slice.
4268 if (task_sleep && parent_entity(se))
4269 set_next_buddy(parent_entity(se));
4271 /* avoid re-evaluating load for this entity */
4272 se = parent_entity(se);
4275 flags |= DEQUEUE_SLEEP;
4278 for_each_sched_entity(se) {
4279 cfs_rq = cfs_rq_of(se);
4280 cfs_rq->h_nr_running--;
4282 if (cfs_rq_throttled(cfs_rq))
4285 update_load_avg(se, 1);
4286 update_cfs_shares(cfs_rq);
4290 sub_nr_running(rq, 1);
4291 schedtune_dequeue_task(p, cpu_of(rq));
4294 * We want to potentially trigger a freq switch
4295 * request only for tasks that are going to sleep;
4296 * this is because we get here also during load
4297 * balancing, but in these cases it seems wise to
4298 * trigger as single request after load balancing is
4302 if (rq->cfs.nr_running)
4303 update_capacity_of(cpu_of(rq));
4304 else if (sched_freq())
4305 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4314 * per rq 'load' arrray crap; XXX kill this.
4318 * The exact cpuload at various idx values, calculated at every tick would be
4319 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4321 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4322 * on nth tick when cpu may be busy, then we have:
4323 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4324 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4326 * decay_load_missed() below does efficient calculation of
4327 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4328 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4330 * The calculation is approximated on a 128 point scale.
4331 * degrade_zero_ticks is the number of ticks after which load at any
4332 * particular idx is approximated to be zero.
4333 * degrade_factor is a precomputed table, a row for each load idx.
4334 * Each column corresponds to degradation factor for a power of two ticks,
4335 * based on 128 point scale.
4337 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4338 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4340 * With this power of 2 load factors, we can degrade the load n times
4341 * by looking at 1 bits in n and doing as many mult/shift instead of
4342 * n mult/shifts needed by the exact degradation.
4344 #define DEGRADE_SHIFT 7
4345 static const unsigned char
4346 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4347 static const unsigned char
4348 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4349 {0, 0, 0, 0, 0, 0, 0, 0},
4350 {64, 32, 8, 0, 0, 0, 0, 0},
4351 {96, 72, 40, 12, 1, 0, 0},
4352 {112, 98, 75, 43, 15, 1, 0},
4353 {120, 112, 98, 76, 45, 16, 2} };
4356 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4357 * would be when CPU is idle and so we just decay the old load without
4358 * adding any new load.
4360 static unsigned long
4361 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4365 if (!missed_updates)
4368 if (missed_updates >= degrade_zero_ticks[idx])
4372 return load >> missed_updates;
4374 while (missed_updates) {
4375 if (missed_updates % 2)
4376 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4378 missed_updates >>= 1;
4385 * Update rq->cpu_load[] statistics. This function is usually called every
4386 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4387 * every tick. We fix it up based on jiffies.
4389 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4390 unsigned long pending_updates)
4394 this_rq->nr_load_updates++;
4396 /* Update our load: */
4397 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4398 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4399 unsigned long old_load, new_load;
4401 /* scale is effectively 1 << i now, and >> i divides by scale */
4403 old_load = this_rq->cpu_load[i];
4404 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4405 new_load = this_load;
4407 * Round up the averaging division if load is increasing. This
4408 * prevents us from getting stuck on 9 if the load is 10, for
4411 if (new_load > old_load)
4412 new_load += scale - 1;
4414 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4417 sched_avg_update(this_rq);
4420 /* Used instead of source_load when we know the type == 0 */
4421 static unsigned long weighted_cpuload(const int cpu)
4423 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4426 #ifdef CONFIG_NO_HZ_COMMON
4428 * There is no sane way to deal with nohz on smp when using jiffies because the
4429 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4430 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4432 * Therefore we cannot use the delta approach from the regular tick since that
4433 * would seriously skew the load calculation. However we'll make do for those
4434 * updates happening while idle (nohz_idle_balance) or coming out of idle
4435 * (tick_nohz_idle_exit).
4437 * This means we might still be one tick off for nohz periods.
4441 * Called from nohz_idle_balance() to update the load ratings before doing the
4444 static void update_idle_cpu_load(struct rq *this_rq)
4446 unsigned long curr_jiffies = READ_ONCE(jiffies);
4447 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4448 unsigned long pending_updates;
4451 * bail if there's load or we're actually up-to-date.
4453 if (load || curr_jiffies == this_rq->last_load_update_tick)
4456 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4457 this_rq->last_load_update_tick = curr_jiffies;
4459 __update_cpu_load(this_rq, load, pending_updates);
4463 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4465 void update_cpu_load_nohz(void)
4467 struct rq *this_rq = this_rq();
4468 unsigned long curr_jiffies = READ_ONCE(jiffies);
4469 unsigned long pending_updates;
4471 if (curr_jiffies == this_rq->last_load_update_tick)
4474 raw_spin_lock(&this_rq->lock);
4475 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4476 if (pending_updates) {
4477 this_rq->last_load_update_tick = curr_jiffies;
4479 * We were idle, this means load 0, the current load might be
4480 * !0 due to remote wakeups and the sort.
4482 __update_cpu_load(this_rq, 0, pending_updates);
4484 raw_spin_unlock(&this_rq->lock);
4486 #endif /* CONFIG_NO_HZ */
4489 * Called from scheduler_tick()
4491 void update_cpu_load_active(struct rq *this_rq)
4493 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4495 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4497 this_rq->last_load_update_tick = jiffies;
4498 __update_cpu_load(this_rq, load, 1);
4502 * Return a low guess at the load of a migration-source cpu weighted
4503 * according to the scheduling class and "nice" value.
4505 * We want to under-estimate the load of migration sources, to
4506 * balance conservatively.
4508 static unsigned long source_load(int cpu, int type)
4510 struct rq *rq = cpu_rq(cpu);
4511 unsigned long total = weighted_cpuload(cpu);
4513 if (type == 0 || !sched_feat(LB_BIAS))
4516 return min(rq->cpu_load[type-1], total);
4520 * Return a high guess at the load of a migration-target cpu weighted
4521 * according to the scheduling class and "nice" value.
4523 static unsigned long target_load(int cpu, int type)
4525 struct rq *rq = cpu_rq(cpu);
4526 unsigned long total = weighted_cpuload(cpu);
4528 if (type == 0 || !sched_feat(LB_BIAS))
4531 return max(rq->cpu_load[type-1], total);
4535 static unsigned long cpu_avg_load_per_task(int cpu)
4537 struct rq *rq = cpu_rq(cpu);
4538 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4539 unsigned long load_avg = weighted_cpuload(cpu);
4542 return load_avg / nr_running;
4547 static void record_wakee(struct task_struct *p)
4550 * Rough decay (wiping) for cost saving, don't worry
4551 * about the boundary, really active task won't care
4554 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4555 current->wakee_flips >>= 1;
4556 current->wakee_flip_decay_ts = jiffies;
4559 if (current->last_wakee != p) {
4560 current->last_wakee = p;
4561 current->wakee_flips++;
4565 static void task_waking_fair(struct task_struct *p)
4567 struct sched_entity *se = &p->se;
4568 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4571 #ifndef CONFIG_64BIT
4572 u64 min_vruntime_copy;
4575 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4577 min_vruntime = cfs_rq->min_vruntime;
4578 } while (min_vruntime != min_vruntime_copy);
4580 min_vruntime = cfs_rq->min_vruntime;
4583 se->vruntime -= min_vruntime;
4587 #ifdef CONFIG_FAIR_GROUP_SCHED
4589 * effective_load() calculates the load change as seen from the root_task_group
4591 * Adding load to a group doesn't make a group heavier, but can cause movement
4592 * of group shares between cpus. Assuming the shares were perfectly aligned one
4593 * can calculate the shift in shares.
4595 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4596 * on this @cpu and results in a total addition (subtraction) of @wg to the
4597 * total group weight.
4599 * Given a runqueue weight distribution (rw_i) we can compute a shares
4600 * distribution (s_i) using:
4602 * s_i = rw_i / \Sum rw_j (1)
4604 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4605 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4606 * shares distribution (s_i):
4608 * rw_i = { 2, 4, 1, 0 }
4609 * s_i = { 2/7, 4/7, 1/7, 0 }
4611 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4612 * task used to run on and the CPU the waker is running on), we need to
4613 * compute the effect of waking a task on either CPU and, in case of a sync
4614 * wakeup, compute the effect of the current task going to sleep.
4616 * So for a change of @wl to the local @cpu with an overall group weight change
4617 * of @wl we can compute the new shares distribution (s'_i) using:
4619 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4621 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4622 * differences in waking a task to CPU 0. The additional task changes the
4623 * weight and shares distributions like:
4625 * rw'_i = { 3, 4, 1, 0 }
4626 * s'_i = { 3/8, 4/8, 1/8, 0 }
4628 * We can then compute the difference in effective weight by using:
4630 * dw_i = S * (s'_i - s_i) (3)
4632 * Where 'S' is the group weight as seen by its parent.
4634 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4635 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4636 * 4/7) times the weight of the group.
4638 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4640 struct sched_entity *se = tg->se[cpu];
4642 if (!tg->parent) /* the trivial, non-cgroup case */
4645 for_each_sched_entity(se) {
4651 * W = @wg + \Sum rw_j
4653 W = wg + calc_tg_weight(tg, se->my_q);
4658 w = cfs_rq_load_avg(se->my_q) + wl;
4661 * wl = S * s'_i; see (2)
4664 wl = (w * (long)tg->shares) / W;
4669 * Per the above, wl is the new se->load.weight value; since
4670 * those are clipped to [MIN_SHARES, ...) do so now. See
4671 * calc_cfs_shares().
4673 if (wl < MIN_SHARES)
4677 * wl = dw_i = S * (s'_i - s_i); see (3)
4679 wl -= se->avg.load_avg;
4682 * Recursively apply this logic to all parent groups to compute
4683 * the final effective load change on the root group. Since
4684 * only the @tg group gets extra weight, all parent groups can
4685 * only redistribute existing shares. @wl is the shift in shares
4686 * resulting from this level per the above.
4695 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4703 * Returns the current capacity of cpu after applying both
4704 * cpu and freq scaling.
4706 unsigned long capacity_curr_of(int cpu)
4708 return cpu_rq(cpu)->cpu_capacity_orig *
4709 arch_scale_freq_capacity(NULL, cpu)
4710 >> SCHED_CAPACITY_SHIFT;
4713 static inline bool energy_aware(void)
4715 return sched_feat(ENERGY_AWARE);
4719 struct sched_group *sg_top;
4720 struct sched_group *sg_cap;
4727 struct task_struct *task;
4742 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4743 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4744 * energy calculations. Using the scale-invariant util returned by
4745 * cpu_util() and approximating scale-invariant util by:
4747 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4749 * the normalized util can be found using the specific capacity.
4751 * capacity = capacity_orig * curr_freq/max_freq
4753 * norm_util = running_time/time ~ util/capacity
4755 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4757 int util = __cpu_util(cpu, delta);
4759 if (util >= capacity)
4760 return SCHED_CAPACITY_SCALE;
4762 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4765 static int calc_util_delta(struct energy_env *eenv, int cpu)
4767 if (cpu == eenv->src_cpu)
4768 return -eenv->util_delta;
4769 if (cpu == eenv->dst_cpu)
4770 return eenv->util_delta;
4775 unsigned long group_max_util(struct energy_env *eenv)
4778 unsigned long max_util = 0;
4780 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4781 delta = calc_util_delta(eenv, i);
4782 max_util = max(max_util, __cpu_util(i, delta));
4789 * group_norm_util() returns the approximated group util relative to it's
4790 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4791 * energy calculations. Since task executions may or may not overlap in time in
4792 * the group the true normalized util is between max(cpu_norm_util(i)) and
4793 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4794 * latter is used as the estimate as it leads to a more pessimistic energy
4795 * estimate (more busy).
4798 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4801 unsigned long util_sum = 0;
4802 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4804 for_each_cpu(i, sched_group_cpus(sg)) {
4805 delta = calc_util_delta(eenv, i);
4806 util_sum += __cpu_norm_util(i, capacity, delta);
4809 if (util_sum > SCHED_CAPACITY_SCALE)
4810 return SCHED_CAPACITY_SCALE;
4814 static int find_new_capacity(struct energy_env *eenv,
4815 const struct sched_group_energy const *sge)
4818 unsigned long util = group_max_util(eenv);
4820 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4821 if (sge->cap_states[idx].cap >= util)
4825 eenv->cap_idx = idx;
4830 static int group_idle_state(struct sched_group *sg)
4832 int i, state = INT_MAX;
4834 /* Find the shallowest idle state in the sched group. */
4835 for_each_cpu(i, sched_group_cpus(sg))
4836 state = min(state, idle_get_state_idx(cpu_rq(i)));
4838 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4845 * sched_group_energy(): Computes the absolute energy consumption of cpus
4846 * belonging to the sched_group including shared resources shared only by
4847 * members of the group. Iterates over all cpus in the hierarchy below the
4848 * sched_group starting from the bottom working it's way up before going to
4849 * the next cpu until all cpus are covered at all levels. The current
4850 * implementation is likely to gather the same util statistics multiple times.
4851 * This can probably be done in a faster but more complex way.
4852 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4854 static int sched_group_energy(struct energy_env *eenv)
4856 struct sched_domain *sd;
4857 int cpu, total_energy = 0;
4858 struct cpumask visit_cpus;
4859 struct sched_group *sg;
4861 WARN_ON(!eenv->sg_top->sge);
4863 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4865 while (!cpumask_empty(&visit_cpus)) {
4866 struct sched_group *sg_shared_cap = NULL;
4868 cpu = cpumask_first(&visit_cpus);
4869 cpumask_clear_cpu(cpu, &visit_cpus);
4872 * Is the group utilization affected by cpus outside this
4875 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4879 * We most probably raced with hotplug; returning a
4880 * wrong energy estimation is better than entering an
4886 sg_shared_cap = sd->parent->groups;
4888 for_each_domain(cpu, sd) {
4891 /* Has this sched_domain already been visited? */
4892 if (sd->child && group_first_cpu(sg) != cpu)
4896 unsigned long group_util;
4897 int sg_busy_energy, sg_idle_energy;
4898 int cap_idx, idle_idx;
4900 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
4901 eenv->sg_cap = sg_shared_cap;
4905 cap_idx = find_new_capacity(eenv, sg->sge);
4907 if (sg->group_weight == 1) {
4908 /* Remove capacity of src CPU (before task move) */
4909 if (eenv->util_delta == 0 &&
4910 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
4911 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
4912 eenv->cap.delta -= eenv->cap.before;
4914 /* Add capacity of dst CPU (after task move) */
4915 if (eenv->util_delta != 0 &&
4916 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
4917 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
4918 eenv->cap.delta += eenv->cap.after;
4922 idle_idx = group_idle_state(sg);
4923 group_util = group_norm_util(eenv, sg);
4924 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
4925 >> SCHED_CAPACITY_SHIFT;
4926 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
4927 * sg->sge->idle_states[idle_idx].power)
4928 >> SCHED_CAPACITY_SHIFT;
4930 total_energy += sg_busy_energy + sg_idle_energy;
4935 for_each_cpu(i, sched_group_cpus(sg))
4936 cpumask_clear_cpu(i, &visit_cpus);
4939 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
4942 } while (sg = sg->next, sg != sd->groups);
4948 eenv->energy = total_energy;
4952 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
4954 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
4957 #ifdef CONFIG_SCHED_TUNE
4958 static int energy_diff_evaluate(struct energy_env *eenv)
4963 /* Return energy diff when boost margin is 0 */
4964 #ifdef CONFIG_CGROUP_SCHEDTUNE
4965 boost = schedtune_task_boost(eenv->task);
4967 boost = get_sysctl_sched_cfs_boost();
4970 return eenv->nrg.diff;
4972 /* Compute normalized energy diff */
4973 nrg_delta = schedtune_normalize_energy(eenv->nrg.diff);
4974 eenv->nrg.delta = nrg_delta;
4976 eenv->payoff = schedtune_accept_deltas(
4982 * When SchedTune is enabled, the energy_diff() function will return
4983 * the computed energy payoff value. Since the energy_diff() return
4984 * value is expected to be negative by its callers, this evaluation
4985 * function return a negative value each time the evaluation return a
4986 * positive payoff, which is the condition for the acceptance of
4987 * a scheduling decision
4989 return -eenv->payoff;
4991 #else /* CONFIG_SCHED_TUNE */
4992 #define energy_diff_evaluate(eenv) eenv->nrg.diff
4996 * energy_diff(): Estimate the energy impact of changing the utilization
4997 * distribution. eenv specifies the change: utilisation amount, source, and
4998 * destination cpu. Source or destination cpu may be -1 in which case the
4999 * utilization is removed from or added to the system (e.g. task wake-up). If
5000 * both are specified, the utilization is migrated.
5002 static int energy_diff(struct energy_env *eenv)
5004 struct sched_domain *sd;
5005 struct sched_group *sg;
5006 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5008 struct energy_env eenv_before = {
5010 .src_cpu = eenv->src_cpu,
5011 .dst_cpu = eenv->dst_cpu,
5012 .nrg = { 0, 0, 0, 0},
5016 if (eenv->src_cpu == eenv->dst_cpu)
5019 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5020 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5023 return 0; /* Error */
5028 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5029 eenv_before.sg_top = eenv->sg_top = sg;
5031 if (sched_group_energy(&eenv_before))
5032 return 0; /* Invalid result abort */
5033 energy_before += eenv_before.energy;
5035 /* Keep track of SRC cpu (before) capacity */
5036 eenv->cap.before = eenv_before.cap.before;
5037 eenv->cap.delta = eenv_before.cap.delta;
5039 if (sched_group_energy(eenv))
5040 return 0; /* Invalid result abort */
5041 energy_after += eenv->energy;
5043 } while (sg = sg->next, sg != sd->groups);
5045 eenv->nrg.before = energy_before;
5046 eenv->nrg.after = energy_after;
5047 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5050 return energy_diff_evaluate(eenv);
5054 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5055 * A waker of many should wake a different task than the one last awakened
5056 * at a frequency roughly N times higher than one of its wakees. In order
5057 * to determine whether we should let the load spread vs consolodating to
5058 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5059 * partner, and a factor of lls_size higher frequency in the other. With
5060 * both conditions met, we can be relatively sure that the relationship is
5061 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5062 * being client/server, worker/dispatcher, interrupt source or whatever is
5063 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5065 static int wake_wide(struct task_struct *p)
5067 unsigned int master = current->wakee_flips;
5068 unsigned int slave = p->wakee_flips;
5069 int factor = this_cpu_read(sd_llc_size);
5072 swap(master, slave);
5073 if (slave < factor || master < slave * factor)
5078 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5080 s64 this_load, load;
5081 s64 this_eff_load, prev_eff_load;
5082 int idx, this_cpu, prev_cpu;
5083 struct task_group *tg;
5084 unsigned long weight;
5088 this_cpu = smp_processor_id();
5089 prev_cpu = task_cpu(p);
5090 load = source_load(prev_cpu, idx);
5091 this_load = target_load(this_cpu, idx);
5094 * If sync wakeup then subtract the (maximum possible)
5095 * effect of the currently running task from the load
5096 * of the current CPU:
5099 tg = task_group(current);
5100 weight = current->se.avg.load_avg;
5102 this_load += effective_load(tg, this_cpu, -weight, -weight);
5103 load += effective_load(tg, prev_cpu, 0, -weight);
5107 weight = p->se.avg.load_avg;
5110 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5111 * due to the sync cause above having dropped this_load to 0, we'll
5112 * always have an imbalance, but there's really nothing you can do
5113 * about that, so that's good too.
5115 * Otherwise check if either cpus are near enough in load to allow this
5116 * task to be woken on this_cpu.
5118 this_eff_load = 100;
5119 this_eff_load *= capacity_of(prev_cpu);
5121 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5122 prev_eff_load *= capacity_of(this_cpu);
5124 if (this_load > 0) {
5125 this_eff_load *= this_load +
5126 effective_load(tg, this_cpu, weight, weight);
5128 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5131 balanced = this_eff_load <= prev_eff_load;
5133 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5138 schedstat_inc(sd, ttwu_move_affine);
5139 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5144 static inline unsigned long task_util(struct task_struct *p)
5146 return p->se.avg.util_avg;
5149 unsigned int capacity_margin = 1280; /* ~20% margin */
5151 static inline unsigned long boosted_task_util(struct task_struct *task);
5153 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5155 unsigned long capacity = capacity_of(cpu);
5157 util += boosted_task_util(p);
5159 return (capacity * 1024) > (util * capacity_margin);
5162 static inline bool task_fits_max(struct task_struct *p, int cpu)
5164 unsigned long capacity = capacity_of(cpu);
5165 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5167 if (capacity == max_capacity)
5170 if (capacity * capacity_margin > max_capacity * 1024)
5173 return __task_fits(p, cpu, 0);
5176 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5178 return __task_fits(p, cpu, cpu_util(cpu));
5181 static bool cpu_overutilized(int cpu)
5183 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5186 #ifdef CONFIG_SCHED_TUNE
5188 static unsigned long
5189 schedtune_margin(unsigned long signal, unsigned long boost)
5191 unsigned long long margin = 0;
5194 * Signal proportional compensation (SPC)
5196 * The Boost (B) value is used to compute a Margin (M) which is
5197 * proportional to the complement of the original Signal (S):
5198 * M = B * (SCHED_LOAD_SCALE - S)
5199 * The obtained M could be used by the caller to "boost" S.
5201 margin = SCHED_LOAD_SCALE - signal;
5205 * Fast integer division by constant:
5206 * Constant : (C) = 100
5207 * Precision : 0.1% (P) = 0.1
5208 * Reference : C * 100 / P (R) = 100000
5211 * Shift bits : ceil(log(R,2)) (S) = 17
5212 * Mult const : round(2^S/C) (M) = 1311
5222 static inline unsigned int
5223 schedtune_cpu_margin(unsigned long util, int cpu)
5227 #ifdef CONFIG_CGROUP_SCHEDTUNE
5228 boost = schedtune_cpu_boost(cpu);
5230 boost = get_sysctl_sched_cfs_boost();
5235 return schedtune_margin(util, boost);
5238 static inline unsigned long
5239 schedtune_task_margin(struct task_struct *task)
5243 unsigned long margin;
5245 #ifdef CONFIG_CGROUP_SCHEDTUNE
5246 boost = schedtune_task_boost(task);
5248 boost = get_sysctl_sched_cfs_boost();
5253 util = task_util(task);
5254 margin = schedtune_margin(util, boost);
5259 #else /* CONFIG_SCHED_TUNE */
5261 static inline unsigned int
5262 schedtune_cpu_margin(unsigned long util, int cpu)
5267 static inline unsigned int
5268 schedtune_task_margin(struct task_struct *task)
5273 #endif /* CONFIG_SCHED_TUNE */
5275 static inline unsigned long
5276 boosted_cpu_util(int cpu)
5278 unsigned long util = cpu_util(cpu);
5279 unsigned long margin = schedtune_cpu_margin(util, cpu);
5281 trace_sched_boost_cpu(cpu, util, margin);
5283 return util + margin;
5286 static inline unsigned long
5287 boosted_task_util(struct task_struct *task)
5289 unsigned long util = task_util(task);
5290 unsigned long margin = schedtune_task_margin(task);
5292 return util + margin;
5296 * find_idlest_group finds and returns the least busy CPU group within the
5299 static struct sched_group *
5300 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5301 int this_cpu, int sd_flag)
5303 struct sched_group *idlest = NULL, *group = sd->groups;
5304 struct sched_group *fit_group = NULL, *spare_group = NULL;
5305 unsigned long min_load = ULONG_MAX, this_load = 0;
5306 unsigned long fit_capacity = ULONG_MAX;
5307 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5308 int load_idx = sd->forkexec_idx;
5309 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5311 if (sd_flag & SD_BALANCE_WAKE)
5312 load_idx = sd->wake_idx;
5315 unsigned long load, avg_load, spare_capacity;
5319 /* Skip over this group if it has no CPUs allowed */
5320 if (!cpumask_intersects(sched_group_cpus(group),
5321 tsk_cpus_allowed(p)))
5324 local_group = cpumask_test_cpu(this_cpu,
5325 sched_group_cpus(group));
5327 /* Tally up the load of all CPUs in the group */
5330 for_each_cpu(i, sched_group_cpus(group)) {
5331 /* Bias balancing toward cpus of our domain */
5333 load = source_load(i, load_idx);
5335 load = target_load(i, load_idx);
5340 * Look for most energy-efficient group that can fit
5341 * that can fit the task.
5343 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5344 fit_capacity = capacity_of(i);
5349 * Look for group which has most spare capacity on a
5352 spare_capacity = capacity_of(i) - cpu_util(i);
5353 if (spare_capacity > max_spare_capacity) {
5354 max_spare_capacity = spare_capacity;
5355 spare_group = group;
5359 /* Adjust by relative CPU capacity of the group */
5360 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5363 this_load = avg_load;
5364 } else if (avg_load < min_load) {
5365 min_load = avg_load;
5368 } while (group = group->next, group != sd->groups);
5376 if (!idlest || 100*this_load < imbalance*min_load)
5382 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5385 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5387 unsigned long load, min_load = ULONG_MAX;
5388 unsigned int min_exit_latency = UINT_MAX;
5389 u64 latest_idle_timestamp = 0;
5390 int least_loaded_cpu = this_cpu;
5391 int shallowest_idle_cpu = -1;
5394 /* Traverse only the allowed CPUs */
5395 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5396 if (task_fits_spare(p, i)) {
5397 struct rq *rq = cpu_rq(i);
5398 struct cpuidle_state *idle = idle_get_state(rq);
5399 if (idle && idle->exit_latency < min_exit_latency) {
5401 * We give priority to a CPU whose idle state
5402 * has the smallest exit latency irrespective
5403 * of any idle timestamp.
5405 min_exit_latency = idle->exit_latency;
5406 latest_idle_timestamp = rq->idle_stamp;
5407 shallowest_idle_cpu = i;
5408 } else if (idle_cpu(i) &&
5409 (!idle || idle->exit_latency == min_exit_latency) &&
5410 rq->idle_stamp > latest_idle_timestamp) {
5412 * If equal or no active idle state, then
5413 * the most recently idled CPU might have
5416 latest_idle_timestamp = rq->idle_stamp;
5417 shallowest_idle_cpu = i;
5418 } else if (shallowest_idle_cpu == -1) {
5420 * If we haven't found an idle CPU yet
5421 * pick a non-idle one that can fit the task as
5424 shallowest_idle_cpu = i;
5426 } else if (shallowest_idle_cpu == -1) {
5427 load = weighted_cpuload(i);
5428 if (load < min_load || (load == min_load && i == this_cpu)) {
5430 least_loaded_cpu = i;
5435 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5439 * Try and locate an idle CPU in the sched_domain.
5441 static int select_idle_sibling(struct task_struct *p, int target)
5443 struct sched_domain *sd;
5444 struct sched_group *sg;
5445 int i = task_cpu(p);
5447 if (idle_cpu(target))
5451 * If the prevous cpu is cache affine and idle, don't be stupid.
5453 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5457 * Otherwise, iterate the domains and find an elegible idle cpu.
5459 sd = rcu_dereference(per_cpu(sd_llc, target));
5460 for_each_lower_domain(sd) {
5463 if (!cpumask_intersects(sched_group_cpus(sg),
5464 tsk_cpus_allowed(p)))
5467 for_each_cpu(i, sched_group_cpus(sg)) {
5468 if (i == target || !idle_cpu(i))
5472 target = cpumask_first_and(sched_group_cpus(sg),
5473 tsk_cpus_allowed(p));
5477 } while (sg != sd->groups);
5483 static int energy_aware_wake_cpu(struct task_struct *p, int target)
5485 struct sched_domain *sd;
5486 struct sched_group *sg, *sg_target;
5487 int target_max_cap = INT_MAX;
5488 int target_cpu = task_cpu(p);
5491 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5500 * Find group with sufficient capacity. We only get here if no cpu is
5501 * overutilized. We may end up overutilizing a cpu by adding the task,
5502 * but that should not be any worse than select_idle_sibling().
5503 * load_balance() should sort it out later as we get above the tipping
5507 /* Assuming all cpus are the same in group */
5508 int max_cap_cpu = group_first_cpu(sg);
5511 * Assume smaller max capacity means more energy-efficient.
5512 * Ideally we should query the energy model for the right
5513 * answer but it easily ends up in an exhaustive search.
5515 if (capacity_of(max_cap_cpu) < target_max_cap &&
5516 task_fits_max(p, max_cap_cpu)) {
5518 target_max_cap = capacity_of(max_cap_cpu);
5520 } while (sg = sg->next, sg != sd->groups);
5522 /* Find cpu with sufficient capacity */
5523 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5525 * p's blocked utilization is still accounted for on prev_cpu
5526 * so prev_cpu will receive a negative bias due to the double
5527 * accounting. However, the blocked utilization may be zero.
5529 int new_util = cpu_util(i) + boosted_task_util(p);
5531 if (new_util > capacity_orig_of(i))
5534 if (new_util < capacity_curr_of(i)) {
5536 if (cpu_rq(i)->nr_running)
5540 /* cpu has capacity at higher OPP, keep it as fallback */
5541 if (target_cpu == task_cpu(p))
5545 if (target_cpu != task_cpu(p)) {
5546 struct energy_env eenv = {
5547 .util_delta = task_util(p),
5548 .src_cpu = task_cpu(p),
5549 .dst_cpu = target_cpu,
5553 /* Not enough spare capacity on previous cpu */
5554 if (cpu_overutilized(task_cpu(p)))
5557 if (energy_diff(&eenv) >= 0)
5565 * select_task_rq_fair: Select target runqueue for the waking task in domains
5566 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5567 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5569 * Balances load by selecting the idlest cpu in the idlest group, or under
5570 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5572 * Returns the target cpu number.
5574 * preempt must be disabled.
5577 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5579 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5580 int cpu = smp_processor_id();
5581 int new_cpu = prev_cpu;
5582 int want_affine = 0;
5583 int sync = wake_flags & WF_SYNC;
5585 if (sd_flag & SD_BALANCE_WAKE)
5586 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5587 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5591 for_each_domain(cpu, tmp) {
5592 if (!(tmp->flags & SD_LOAD_BALANCE))
5596 * If both cpu and prev_cpu are part of this domain,
5597 * cpu is a valid SD_WAKE_AFFINE target.
5599 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5600 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5605 if (tmp->flags & sd_flag)
5607 else if (!want_affine)
5612 sd = NULL; /* Prefer wake_affine over balance flags */
5613 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5618 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5619 new_cpu = energy_aware_wake_cpu(p, prev_cpu);
5620 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5621 new_cpu = select_idle_sibling(p, new_cpu);
5624 struct sched_group *group;
5627 if (!(sd->flags & sd_flag)) {
5632 group = find_idlest_group(sd, p, cpu, sd_flag);
5638 new_cpu = find_idlest_cpu(group, p, cpu);
5639 if (new_cpu == -1 || new_cpu == cpu) {
5640 /* Now try balancing at a lower domain level of cpu */
5645 /* Now try balancing at a lower domain level of new_cpu */
5647 weight = sd->span_weight;
5649 for_each_domain(cpu, tmp) {
5650 if (weight <= tmp->span_weight)
5652 if (tmp->flags & sd_flag)
5655 /* while loop will break here if sd == NULL */
5663 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5664 * cfs_rq_of(p) references at time of call are still valid and identify the
5665 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5666 * other assumptions, including the state of rq->lock, should be made.
5668 static void migrate_task_rq_fair(struct task_struct *p)
5671 * We are supposed to update the task to "current" time, then its up to date
5672 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5673 * what current time is, so simply throw away the out-of-date time. This
5674 * will result in the wakee task is less decayed, but giving the wakee more
5675 * load sounds not bad.
5677 remove_entity_load_avg(&p->se);
5679 /* Tell new CPU we are migrated */
5680 p->se.avg.last_update_time = 0;
5682 /* We have migrated, no longer consider this task hot */
5683 p->se.exec_start = 0;
5686 static void task_dead_fair(struct task_struct *p)
5688 remove_entity_load_avg(&p->se);
5690 #endif /* CONFIG_SMP */
5692 static unsigned long
5693 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5695 unsigned long gran = sysctl_sched_wakeup_granularity;
5698 * Since its curr running now, convert the gran from real-time
5699 * to virtual-time in his units.
5701 * By using 'se' instead of 'curr' we penalize light tasks, so
5702 * they get preempted easier. That is, if 'se' < 'curr' then
5703 * the resulting gran will be larger, therefore penalizing the
5704 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5705 * be smaller, again penalizing the lighter task.
5707 * This is especially important for buddies when the leftmost
5708 * task is higher priority than the buddy.
5710 return calc_delta_fair(gran, se);
5714 * Should 'se' preempt 'curr'.
5728 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5730 s64 gran, vdiff = curr->vruntime - se->vruntime;
5735 gran = wakeup_gran(curr, se);
5742 static void set_last_buddy(struct sched_entity *se)
5744 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5747 for_each_sched_entity(se)
5748 cfs_rq_of(se)->last = se;
5751 static void set_next_buddy(struct sched_entity *se)
5753 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5756 for_each_sched_entity(se)
5757 cfs_rq_of(se)->next = se;
5760 static void set_skip_buddy(struct sched_entity *se)
5762 for_each_sched_entity(se)
5763 cfs_rq_of(se)->skip = se;
5767 * Preempt the current task with a newly woken task if needed:
5769 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5771 struct task_struct *curr = rq->curr;
5772 struct sched_entity *se = &curr->se, *pse = &p->se;
5773 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5774 int scale = cfs_rq->nr_running >= sched_nr_latency;
5775 int next_buddy_marked = 0;
5777 if (unlikely(se == pse))
5781 * This is possible from callers such as attach_tasks(), in which we
5782 * unconditionally check_prempt_curr() after an enqueue (which may have
5783 * lead to a throttle). This both saves work and prevents false
5784 * next-buddy nomination below.
5786 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5789 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5790 set_next_buddy(pse);
5791 next_buddy_marked = 1;
5795 * We can come here with TIF_NEED_RESCHED already set from new task
5798 * Note: this also catches the edge-case of curr being in a throttled
5799 * group (e.g. via set_curr_task), since update_curr() (in the
5800 * enqueue of curr) will have resulted in resched being set. This
5801 * prevents us from potentially nominating it as a false LAST_BUDDY
5804 if (test_tsk_need_resched(curr))
5807 /* Idle tasks are by definition preempted by non-idle tasks. */
5808 if (unlikely(curr->policy == SCHED_IDLE) &&
5809 likely(p->policy != SCHED_IDLE))
5813 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5814 * is driven by the tick):
5816 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5819 find_matching_se(&se, &pse);
5820 update_curr(cfs_rq_of(se));
5822 if (wakeup_preempt_entity(se, pse) == 1) {
5824 * Bias pick_next to pick the sched entity that is
5825 * triggering this preemption.
5827 if (!next_buddy_marked)
5828 set_next_buddy(pse);
5837 * Only set the backward buddy when the current task is still
5838 * on the rq. This can happen when a wakeup gets interleaved
5839 * with schedule on the ->pre_schedule() or idle_balance()
5840 * point, either of which can * drop the rq lock.
5842 * Also, during early boot the idle thread is in the fair class,
5843 * for obvious reasons its a bad idea to schedule back to it.
5845 if (unlikely(!se->on_rq || curr == rq->idle))
5848 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5852 static struct task_struct *
5853 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5855 struct cfs_rq *cfs_rq = &rq->cfs;
5856 struct sched_entity *se;
5857 struct task_struct *p;
5861 #ifdef CONFIG_FAIR_GROUP_SCHED
5862 if (!cfs_rq->nr_running)
5865 if (prev->sched_class != &fair_sched_class)
5869 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5870 * likely that a next task is from the same cgroup as the current.
5872 * Therefore attempt to avoid putting and setting the entire cgroup
5873 * hierarchy, only change the part that actually changes.
5877 struct sched_entity *curr = cfs_rq->curr;
5880 * Since we got here without doing put_prev_entity() we also
5881 * have to consider cfs_rq->curr. If it is still a runnable
5882 * entity, update_curr() will update its vruntime, otherwise
5883 * forget we've ever seen it.
5887 update_curr(cfs_rq);
5892 * This call to check_cfs_rq_runtime() will do the
5893 * throttle and dequeue its entity in the parent(s).
5894 * Therefore the 'simple' nr_running test will indeed
5897 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5901 se = pick_next_entity(cfs_rq, curr);
5902 cfs_rq = group_cfs_rq(se);
5908 * Since we haven't yet done put_prev_entity and if the selected task
5909 * is a different task than we started out with, try and touch the
5910 * least amount of cfs_rqs.
5913 struct sched_entity *pse = &prev->se;
5915 while (!(cfs_rq = is_same_group(se, pse))) {
5916 int se_depth = se->depth;
5917 int pse_depth = pse->depth;
5919 if (se_depth <= pse_depth) {
5920 put_prev_entity(cfs_rq_of(pse), pse);
5921 pse = parent_entity(pse);
5923 if (se_depth >= pse_depth) {
5924 set_next_entity(cfs_rq_of(se), se);
5925 se = parent_entity(se);
5929 put_prev_entity(cfs_rq, pse);
5930 set_next_entity(cfs_rq, se);
5933 if (hrtick_enabled(rq))
5934 hrtick_start_fair(rq, p);
5936 rq->misfit_task = !task_fits_max(p, rq->cpu);
5943 if (!cfs_rq->nr_running)
5946 put_prev_task(rq, prev);
5949 se = pick_next_entity(cfs_rq, NULL);
5950 set_next_entity(cfs_rq, se);
5951 cfs_rq = group_cfs_rq(se);
5956 if (hrtick_enabled(rq))
5957 hrtick_start_fair(rq, p);
5959 rq->misfit_task = !task_fits_max(p, rq->cpu);
5964 rq->misfit_task = 0;
5966 * This is OK, because current is on_cpu, which avoids it being picked
5967 * for load-balance and preemption/IRQs are still disabled avoiding
5968 * further scheduler activity on it and we're being very careful to
5969 * re-start the picking loop.
5971 lockdep_unpin_lock(&rq->lock);
5972 new_tasks = idle_balance(rq);
5973 lockdep_pin_lock(&rq->lock);
5975 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5976 * possible for any higher priority task to appear. In that case we
5977 * must re-start the pick_next_entity() loop.
5989 * Account for a descheduled task:
5991 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5993 struct sched_entity *se = &prev->se;
5994 struct cfs_rq *cfs_rq;
5996 for_each_sched_entity(se) {
5997 cfs_rq = cfs_rq_of(se);
5998 put_prev_entity(cfs_rq, se);
6003 * sched_yield() is very simple
6005 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6007 static void yield_task_fair(struct rq *rq)
6009 struct task_struct *curr = rq->curr;
6010 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6011 struct sched_entity *se = &curr->se;
6014 * Are we the only task in the tree?
6016 if (unlikely(rq->nr_running == 1))
6019 clear_buddies(cfs_rq, se);
6021 if (curr->policy != SCHED_BATCH) {
6022 update_rq_clock(rq);
6024 * Update run-time statistics of the 'current'.
6026 update_curr(cfs_rq);
6028 * Tell update_rq_clock() that we've just updated,
6029 * so we don't do microscopic update in schedule()
6030 * and double the fastpath cost.
6032 rq_clock_skip_update(rq, true);
6038 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6040 struct sched_entity *se = &p->se;
6042 /* throttled hierarchies are not runnable */
6043 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6046 /* Tell the scheduler that we'd really like pse to run next. */
6049 yield_task_fair(rq);
6055 /**************************************************
6056 * Fair scheduling class load-balancing methods.
6060 * The purpose of load-balancing is to achieve the same basic fairness the
6061 * per-cpu scheduler provides, namely provide a proportional amount of compute
6062 * time to each task. This is expressed in the following equation:
6064 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6066 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6067 * W_i,0 is defined as:
6069 * W_i,0 = \Sum_j w_i,j (2)
6071 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6072 * is derived from the nice value as per prio_to_weight[].
6074 * The weight average is an exponential decay average of the instantaneous
6077 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6079 * C_i is the compute capacity of cpu i, typically it is the
6080 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6081 * can also include other factors [XXX].
6083 * To achieve this balance we define a measure of imbalance which follows
6084 * directly from (1):
6086 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6088 * We them move tasks around to minimize the imbalance. In the continuous
6089 * function space it is obvious this converges, in the discrete case we get
6090 * a few fun cases generally called infeasible weight scenarios.
6093 * - infeasible weights;
6094 * - local vs global optima in the discrete case. ]
6099 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6100 * for all i,j solution, we create a tree of cpus that follows the hardware
6101 * topology where each level pairs two lower groups (or better). This results
6102 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6103 * tree to only the first of the previous level and we decrease the frequency
6104 * of load-balance at each level inv. proportional to the number of cpus in
6110 * \Sum { --- * --- * 2^i } = O(n) (5)
6112 * `- size of each group
6113 * | | `- number of cpus doing load-balance
6115 * `- sum over all levels
6117 * Coupled with a limit on how many tasks we can migrate every balance pass,
6118 * this makes (5) the runtime complexity of the balancer.
6120 * An important property here is that each CPU is still (indirectly) connected
6121 * to every other cpu in at most O(log n) steps:
6123 * The adjacency matrix of the resulting graph is given by:
6126 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6129 * And you'll find that:
6131 * A^(log_2 n)_i,j != 0 for all i,j (7)
6133 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6134 * The task movement gives a factor of O(m), giving a convergence complexity
6137 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6142 * In order to avoid CPUs going idle while there's still work to do, new idle
6143 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6144 * tree itself instead of relying on other CPUs to bring it work.
6146 * This adds some complexity to both (5) and (8) but it reduces the total idle
6154 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6157 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6162 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6164 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6166 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6169 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6170 * rewrite all of this once again.]
6173 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6175 enum fbq_type { regular, remote, all };
6184 #define LBF_ALL_PINNED 0x01
6185 #define LBF_NEED_BREAK 0x02
6186 #define LBF_DST_PINNED 0x04
6187 #define LBF_SOME_PINNED 0x08
6190 struct sched_domain *sd;
6198 struct cpumask *dst_grpmask;
6200 enum cpu_idle_type idle;
6202 unsigned int src_grp_nr_running;
6203 /* The set of CPUs under consideration for load-balancing */
6204 struct cpumask *cpus;
6209 unsigned int loop_break;
6210 unsigned int loop_max;
6212 enum fbq_type fbq_type;
6213 enum group_type busiest_group_type;
6214 struct list_head tasks;
6218 * Is this task likely cache-hot:
6220 static int task_hot(struct task_struct *p, struct lb_env *env)
6224 lockdep_assert_held(&env->src_rq->lock);
6226 if (p->sched_class != &fair_sched_class)
6229 if (unlikely(p->policy == SCHED_IDLE))
6233 * Buddy candidates are cache hot:
6235 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6236 (&p->se == cfs_rq_of(&p->se)->next ||
6237 &p->se == cfs_rq_of(&p->se)->last))
6240 if (sysctl_sched_migration_cost == -1)
6242 if (sysctl_sched_migration_cost == 0)
6245 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6247 return delta < (s64)sysctl_sched_migration_cost;
6250 #ifdef CONFIG_NUMA_BALANCING
6252 * Returns 1, if task migration degrades locality
6253 * Returns 0, if task migration improves locality i.e migration preferred.
6254 * Returns -1, if task migration is not affected by locality.
6256 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6258 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6259 unsigned long src_faults, dst_faults;
6260 int src_nid, dst_nid;
6262 if (!static_branch_likely(&sched_numa_balancing))
6265 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6268 src_nid = cpu_to_node(env->src_cpu);
6269 dst_nid = cpu_to_node(env->dst_cpu);
6271 if (src_nid == dst_nid)
6274 /* Migrating away from the preferred node is always bad. */
6275 if (src_nid == p->numa_preferred_nid) {
6276 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6282 /* Encourage migration to the preferred node. */
6283 if (dst_nid == p->numa_preferred_nid)
6287 src_faults = group_faults(p, src_nid);
6288 dst_faults = group_faults(p, dst_nid);
6290 src_faults = task_faults(p, src_nid);
6291 dst_faults = task_faults(p, dst_nid);
6294 return dst_faults < src_faults;
6298 static inline int migrate_degrades_locality(struct task_struct *p,
6306 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6309 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6313 lockdep_assert_held(&env->src_rq->lock);
6316 * We do not migrate tasks that are:
6317 * 1) throttled_lb_pair, or
6318 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6319 * 3) running (obviously), or
6320 * 4) are cache-hot on their current CPU.
6322 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6325 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6328 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6330 env->flags |= LBF_SOME_PINNED;
6333 * Remember if this task can be migrated to any other cpu in
6334 * our sched_group. We may want to revisit it if we couldn't
6335 * meet load balance goals by pulling other tasks on src_cpu.
6337 * Also avoid computing new_dst_cpu if we have already computed
6338 * one in current iteration.
6340 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6343 /* Prevent to re-select dst_cpu via env's cpus */
6344 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6345 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6346 env->flags |= LBF_DST_PINNED;
6347 env->new_dst_cpu = cpu;
6355 /* Record that we found atleast one task that could run on dst_cpu */
6356 env->flags &= ~LBF_ALL_PINNED;
6358 if (task_running(env->src_rq, p)) {
6359 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6364 * Aggressive migration if:
6365 * 1) destination numa is preferred
6366 * 2) task is cache cold, or
6367 * 3) too many balance attempts have failed.
6369 tsk_cache_hot = migrate_degrades_locality(p, env);
6370 if (tsk_cache_hot == -1)
6371 tsk_cache_hot = task_hot(p, env);
6373 if (tsk_cache_hot <= 0 ||
6374 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6375 if (tsk_cache_hot == 1) {
6376 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6377 schedstat_inc(p, se.statistics.nr_forced_migrations);
6382 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6387 * detach_task() -- detach the task for the migration specified in env
6389 static void detach_task(struct task_struct *p, struct lb_env *env)
6391 lockdep_assert_held(&env->src_rq->lock);
6393 deactivate_task(env->src_rq, p, 0);
6394 p->on_rq = TASK_ON_RQ_MIGRATING;
6395 set_task_cpu(p, env->dst_cpu);
6399 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6400 * part of active balancing operations within "domain".
6402 * Returns a task if successful and NULL otherwise.
6404 static struct task_struct *detach_one_task(struct lb_env *env)
6406 struct task_struct *p, *n;
6408 lockdep_assert_held(&env->src_rq->lock);
6410 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6411 if (!can_migrate_task(p, env))
6414 detach_task(p, env);
6417 * Right now, this is only the second place where
6418 * lb_gained[env->idle] is updated (other is detach_tasks)
6419 * so we can safely collect stats here rather than
6420 * inside detach_tasks().
6422 schedstat_inc(env->sd, lb_gained[env->idle]);
6428 static const unsigned int sched_nr_migrate_break = 32;
6431 * detach_tasks() -- tries to detach up to imbalance weighted load from
6432 * busiest_rq, as part of a balancing operation within domain "sd".
6434 * Returns number of detached tasks if successful and 0 otherwise.
6436 static int detach_tasks(struct lb_env *env)
6438 struct list_head *tasks = &env->src_rq->cfs_tasks;
6439 struct task_struct *p;
6443 lockdep_assert_held(&env->src_rq->lock);
6445 if (env->imbalance <= 0)
6448 while (!list_empty(tasks)) {
6450 * We don't want to steal all, otherwise we may be treated likewise,
6451 * which could at worst lead to a livelock crash.
6453 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6456 p = list_first_entry(tasks, struct task_struct, se.group_node);
6459 /* We've more or less seen every task there is, call it quits */
6460 if (env->loop > env->loop_max)
6463 /* take a breather every nr_migrate tasks */
6464 if (env->loop > env->loop_break) {
6465 env->loop_break += sched_nr_migrate_break;
6466 env->flags |= LBF_NEED_BREAK;
6470 if (!can_migrate_task(p, env))
6473 load = task_h_load(p);
6475 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6478 if ((load / 2) > env->imbalance)
6481 detach_task(p, env);
6482 list_add(&p->se.group_node, &env->tasks);
6485 env->imbalance -= load;
6487 #ifdef CONFIG_PREEMPT
6489 * NEWIDLE balancing is a source of latency, so preemptible
6490 * kernels will stop after the first task is detached to minimize
6491 * the critical section.
6493 if (env->idle == CPU_NEWLY_IDLE)
6498 * We only want to steal up to the prescribed amount of
6501 if (env->imbalance <= 0)
6506 list_move_tail(&p->se.group_node, tasks);
6510 * Right now, this is one of only two places we collect this stat
6511 * so we can safely collect detach_one_task() stats here rather
6512 * than inside detach_one_task().
6514 schedstat_add(env->sd, lb_gained[env->idle], detached);
6520 * attach_task() -- attach the task detached by detach_task() to its new rq.
6522 static void attach_task(struct rq *rq, struct task_struct *p)
6524 lockdep_assert_held(&rq->lock);
6526 BUG_ON(task_rq(p) != rq);
6527 p->on_rq = TASK_ON_RQ_QUEUED;
6528 activate_task(rq, p, 0);
6529 check_preempt_curr(rq, p, 0);
6533 * attach_one_task() -- attaches the task returned from detach_one_task() to
6536 static void attach_one_task(struct rq *rq, struct task_struct *p)
6538 raw_spin_lock(&rq->lock);
6541 * We want to potentially raise target_cpu's OPP.
6543 update_capacity_of(cpu_of(rq));
6544 raw_spin_unlock(&rq->lock);
6548 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6551 static void attach_tasks(struct lb_env *env)
6553 struct list_head *tasks = &env->tasks;
6554 struct task_struct *p;
6556 raw_spin_lock(&env->dst_rq->lock);
6558 while (!list_empty(tasks)) {
6559 p = list_first_entry(tasks, struct task_struct, se.group_node);
6560 list_del_init(&p->se.group_node);
6562 attach_task(env->dst_rq, p);
6566 * We want to potentially raise env.dst_cpu's OPP.
6568 update_capacity_of(env->dst_cpu);
6570 raw_spin_unlock(&env->dst_rq->lock);
6573 #ifdef CONFIG_FAIR_GROUP_SCHED
6574 static void update_blocked_averages(int cpu)
6576 struct rq *rq = cpu_rq(cpu);
6577 struct cfs_rq *cfs_rq;
6578 unsigned long flags;
6580 raw_spin_lock_irqsave(&rq->lock, flags);
6581 update_rq_clock(rq);
6584 * Iterates the task_group tree in a bottom up fashion, see
6585 * list_add_leaf_cfs_rq() for details.
6587 for_each_leaf_cfs_rq(rq, cfs_rq) {
6588 /* throttled entities do not contribute to load */
6589 if (throttled_hierarchy(cfs_rq))
6592 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6593 update_tg_load_avg(cfs_rq, 0);
6595 raw_spin_unlock_irqrestore(&rq->lock, flags);
6599 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6600 * This needs to be done in a top-down fashion because the load of a child
6601 * group is a fraction of its parents load.
6603 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6605 struct rq *rq = rq_of(cfs_rq);
6606 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6607 unsigned long now = jiffies;
6610 if (cfs_rq->last_h_load_update == now)
6613 cfs_rq->h_load_next = NULL;
6614 for_each_sched_entity(se) {
6615 cfs_rq = cfs_rq_of(se);
6616 cfs_rq->h_load_next = se;
6617 if (cfs_rq->last_h_load_update == now)
6622 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6623 cfs_rq->last_h_load_update = now;
6626 while ((se = cfs_rq->h_load_next) != NULL) {
6627 load = cfs_rq->h_load;
6628 load = div64_ul(load * se->avg.load_avg,
6629 cfs_rq_load_avg(cfs_rq) + 1);
6630 cfs_rq = group_cfs_rq(se);
6631 cfs_rq->h_load = load;
6632 cfs_rq->last_h_load_update = now;
6636 static unsigned long task_h_load(struct task_struct *p)
6638 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6640 update_cfs_rq_h_load(cfs_rq);
6641 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6642 cfs_rq_load_avg(cfs_rq) + 1);
6645 static inline void update_blocked_averages(int cpu)
6647 struct rq *rq = cpu_rq(cpu);
6648 struct cfs_rq *cfs_rq = &rq->cfs;
6649 unsigned long flags;
6651 raw_spin_lock_irqsave(&rq->lock, flags);
6652 update_rq_clock(rq);
6653 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6654 raw_spin_unlock_irqrestore(&rq->lock, flags);
6657 static unsigned long task_h_load(struct task_struct *p)
6659 return p->se.avg.load_avg;
6663 /********** Helpers for find_busiest_group ************************/
6666 * sg_lb_stats - stats of a sched_group required for load_balancing
6668 struct sg_lb_stats {
6669 unsigned long avg_load; /*Avg load across the CPUs of the group */
6670 unsigned long group_load; /* Total load over the CPUs of the group */
6671 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6672 unsigned long load_per_task;
6673 unsigned long group_capacity;
6674 unsigned long group_util; /* Total utilization of the group */
6675 unsigned int sum_nr_running; /* Nr tasks running in the group */
6676 unsigned int idle_cpus;
6677 unsigned int group_weight;
6678 enum group_type group_type;
6679 int group_no_capacity;
6680 int group_misfit_task; /* A cpu has a task too big for its capacity */
6681 #ifdef CONFIG_NUMA_BALANCING
6682 unsigned int nr_numa_running;
6683 unsigned int nr_preferred_running;
6688 * sd_lb_stats - Structure to store the statistics of a sched_domain
6689 * during load balancing.
6691 struct sd_lb_stats {
6692 struct sched_group *busiest; /* Busiest group in this sd */
6693 struct sched_group *local; /* Local group in this sd */
6694 unsigned long total_load; /* Total load of all groups in sd */
6695 unsigned long total_capacity; /* Total capacity of all groups in sd */
6696 unsigned long avg_load; /* Average load across all groups in sd */
6698 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6699 struct sg_lb_stats local_stat; /* Statistics of the local group */
6702 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6705 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6706 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6707 * We must however clear busiest_stat::avg_load because
6708 * update_sd_pick_busiest() reads this before assignment.
6710 *sds = (struct sd_lb_stats){
6714 .total_capacity = 0UL,
6717 .sum_nr_running = 0,
6718 .group_type = group_other,
6724 * get_sd_load_idx - Obtain the load index for a given sched domain.
6725 * @sd: The sched_domain whose load_idx is to be obtained.
6726 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6728 * Return: The load index.
6730 static inline int get_sd_load_idx(struct sched_domain *sd,
6731 enum cpu_idle_type idle)
6737 load_idx = sd->busy_idx;
6740 case CPU_NEWLY_IDLE:
6741 load_idx = sd->newidle_idx;
6744 load_idx = sd->idle_idx;
6751 static unsigned long scale_rt_capacity(int cpu)
6753 struct rq *rq = cpu_rq(cpu);
6754 u64 total, used, age_stamp, avg;
6758 * Since we're reading these variables without serialization make sure
6759 * we read them once before doing sanity checks on them.
6761 age_stamp = READ_ONCE(rq->age_stamp);
6762 avg = READ_ONCE(rq->rt_avg);
6763 delta = __rq_clock_broken(rq) - age_stamp;
6765 if (unlikely(delta < 0))
6768 total = sched_avg_period() + delta;
6770 used = div_u64(avg, total);
6773 * deadline bandwidth is defined at system level so we must
6774 * weight this bandwidth with the max capacity of the system.
6775 * As a reminder, avg_bw is 20bits width and
6776 * scale_cpu_capacity is 10 bits width
6778 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
6780 if (likely(used < SCHED_CAPACITY_SCALE))
6781 return SCHED_CAPACITY_SCALE - used;
6786 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
6788 raw_spin_lock_init(&mcc->lock);
6793 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6795 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6796 struct sched_group *sdg = sd->groups;
6797 struct max_cpu_capacity *mcc;
6798 unsigned long max_capacity;
6800 unsigned long flags;
6802 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6804 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
6806 raw_spin_lock_irqsave(&mcc->lock, flags);
6807 max_capacity = mcc->val;
6808 max_cap_cpu = mcc->cpu;
6810 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
6811 (max_capacity < capacity)) {
6812 mcc->val = capacity;
6814 #ifdef CONFIG_SCHED_DEBUG
6815 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6816 //pr_info("CPU%d: update max cpu_capacity %lu\n", cpu, capacity);
6820 raw_spin_unlock_irqrestore(&mcc->lock, flags);
6822 skip_unlock: __attribute__ ((unused));
6823 capacity *= scale_rt_capacity(cpu);
6824 capacity >>= SCHED_CAPACITY_SHIFT;
6829 cpu_rq(cpu)->cpu_capacity = capacity;
6830 sdg->sgc->capacity = capacity;
6831 sdg->sgc->max_capacity = capacity;
6834 void update_group_capacity(struct sched_domain *sd, int cpu)
6836 struct sched_domain *child = sd->child;
6837 struct sched_group *group, *sdg = sd->groups;
6838 unsigned long capacity, max_capacity;
6839 unsigned long interval;
6841 interval = msecs_to_jiffies(sd->balance_interval);
6842 interval = clamp(interval, 1UL, max_load_balance_interval);
6843 sdg->sgc->next_update = jiffies + interval;
6846 update_cpu_capacity(sd, cpu);
6853 if (child->flags & SD_OVERLAP) {
6855 * SD_OVERLAP domains cannot assume that child groups
6856 * span the current group.
6859 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6860 struct sched_group_capacity *sgc;
6861 struct rq *rq = cpu_rq(cpu);
6864 * build_sched_domains() -> init_sched_groups_capacity()
6865 * gets here before we've attached the domains to the
6868 * Use capacity_of(), which is set irrespective of domains
6869 * in update_cpu_capacity().
6871 * This avoids capacity from being 0 and
6872 * causing divide-by-zero issues on boot.
6874 if (unlikely(!rq->sd)) {
6875 capacity += capacity_of(cpu);
6877 sgc = rq->sd->groups->sgc;
6878 capacity += sgc->capacity;
6881 max_capacity = max(capacity, max_capacity);
6885 * !SD_OVERLAP domains can assume that child groups
6886 * span the current group.
6889 group = child->groups;
6891 struct sched_group_capacity *sgc = group->sgc;
6893 capacity += sgc->capacity;
6894 max_capacity = max(sgc->max_capacity, max_capacity);
6895 group = group->next;
6896 } while (group != child->groups);
6899 sdg->sgc->capacity = capacity;
6900 sdg->sgc->max_capacity = max_capacity;
6904 * Check whether the capacity of the rq has been noticeably reduced by side
6905 * activity. The imbalance_pct is used for the threshold.
6906 * Return true is the capacity is reduced
6909 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6911 return ((rq->cpu_capacity * sd->imbalance_pct) <
6912 (rq->cpu_capacity_orig * 100));
6916 * Group imbalance indicates (and tries to solve) the problem where balancing
6917 * groups is inadequate due to tsk_cpus_allowed() constraints.
6919 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6920 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6923 * { 0 1 2 3 } { 4 5 6 7 }
6926 * If we were to balance group-wise we'd place two tasks in the first group and
6927 * two tasks in the second group. Clearly this is undesired as it will overload
6928 * cpu 3 and leave one of the cpus in the second group unused.
6930 * The current solution to this issue is detecting the skew in the first group
6931 * by noticing the lower domain failed to reach balance and had difficulty
6932 * moving tasks due to affinity constraints.
6934 * When this is so detected; this group becomes a candidate for busiest; see
6935 * update_sd_pick_busiest(). And calculate_imbalance() and
6936 * find_busiest_group() avoid some of the usual balance conditions to allow it
6937 * to create an effective group imbalance.
6939 * This is a somewhat tricky proposition since the next run might not find the
6940 * group imbalance and decide the groups need to be balanced again. A most
6941 * subtle and fragile situation.
6944 static inline int sg_imbalanced(struct sched_group *group)
6946 return group->sgc->imbalance;
6950 * group_has_capacity returns true if the group has spare capacity that could
6951 * be used by some tasks.
6952 * We consider that a group has spare capacity if the * number of task is
6953 * smaller than the number of CPUs or if the utilization is lower than the
6954 * available capacity for CFS tasks.
6955 * For the latter, we use a threshold to stabilize the state, to take into
6956 * account the variance of the tasks' load and to return true if the available
6957 * capacity in meaningful for the load balancer.
6958 * As an example, an available capacity of 1% can appear but it doesn't make
6959 * any benefit for the load balance.
6962 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6964 if (sgs->sum_nr_running < sgs->group_weight)
6967 if ((sgs->group_capacity * 100) >
6968 (sgs->group_util * env->sd->imbalance_pct))
6975 * group_is_overloaded returns true if the group has more tasks than it can
6977 * group_is_overloaded is not equals to !group_has_capacity because a group
6978 * with the exact right number of tasks, has no more spare capacity but is not
6979 * overloaded so both group_has_capacity and group_is_overloaded return
6983 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6985 if (sgs->sum_nr_running <= sgs->group_weight)
6988 if ((sgs->group_capacity * 100) <
6989 (sgs->group_util * env->sd->imbalance_pct))
6997 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
6998 * per-cpu capacity than sched_group ref.
7001 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7003 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7004 ref->sgc->max_capacity;
7008 group_type group_classify(struct sched_group *group,
7009 struct sg_lb_stats *sgs)
7011 if (sgs->group_no_capacity)
7012 return group_overloaded;
7014 if (sg_imbalanced(group))
7015 return group_imbalanced;
7017 if (sgs->group_misfit_task)
7018 return group_misfit_task;
7024 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7025 * @env: The load balancing environment.
7026 * @group: sched_group whose statistics are to be updated.
7027 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7028 * @local_group: Does group contain this_cpu.
7029 * @sgs: variable to hold the statistics for this group.
7030 * @overload: Indicate more than one runnable task for any CPU.
7031 * @overutilized: Indicate overutilization for any CPU.
7033 static inline void update_sg_lb_stats(struct lb_env *env,
7034 struct sched_group *group, int load_idx,
7035 int local_group, struct sg_lb_stats *sgs,
7036 bool *overload, bool *overutilized)
7041 memset(sgs, 0, sizeof(*sgs));
7043 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7044 struct rq *rq = cpu_rq(i);
7046 /* Bias balancing toward cpus of our domain */
7048 load = target_load(i, load_idx);
7050 load = source_load(i, load_idx);
7052 sgs->group_load += load;
7053 sgs->group_util += cpu_util(i);
7054 sgs->sum_nr_running += rq->cfs.h_nr_running;
7056 if (rq->nr_running > 1)
7059 #ifdef CONFIG_NUMA_BALANCING
7060 sgs->nr_numa_running += rq->nr_numa_running;
7061 sgs->nr_preferred_running += rq->nr_preferred_running;
7063 sgs->sum_weighted_load += weighted_cpuload(i);
7067 if (cpu_overutilized(i)) {
7068 *overutilized = true;
7069 if (!sgs->group_misfit_task && rq->misfit_task)
7070 sgs->group_misfit_task = capacity_of(i);
7074 /* Adjust by relative CPU capacity of the group */
7075 sgs->group_capacity = group->sgc->capacity;
7076 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7078 if (sgs->sum_nr_running)
7079 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7081 sgs->group_weight = group->group_weight;
7083 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7084 sgs->group_type = group_classify(group, sgs);
7088 * update_sd_pick_busiest - return 1 on busiest group
7089 * @env: The load balancing environment.
7090 * @sds: sched_domain statistics
7091 * @sg: sched_group candidate to be checked for being the busiest
7092 * @sgs: sched_group statistics
7094 * Determine if @sg is a busier group than the previously selected
7097 * Return: %true if @sg is a busier group than the previously selected
7098 * busiest group. %false otherwise.
7100 static bool update_sd_pick_busiest(struct lb_env *env,
7101 struct sd_lb_stats *sds,
7102 struct sched_group *sg,
7103 struct sg_lb_stats *sgs)
7105 struct sg_lb_stats *busiest = &sds->busiest_stat;
7107 if (sgs->group_type > busiest->group_type)
7110 if (sgs->group_type < busiest->group_type)
7114 * Candidate sg doesn't face any serious load-balance problems
7115 * so don't pick it if the local sg is already filled up.
7117 if (sgs->group_type == group_other &&
7118 !group_has_capacity(env, &sds->local_stat))
7121 if (sgs->avg_load <= busiest->avg_load)
7125 * Candiate sg has no more than one task per cpu and has higher
7126 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7128 if (sgs->sum_nr_running <= sgs->group_weight &&
7129 group_smaller_cpu_capacity(sds->local, sg))
7132 /* This is the busiest node in its class. */
7133 if (!(env->sd->flags & SD_ASYM_PACKING))
7137 * ASYM_PACKING needs to move all the work to the lowest
7138 * numbered CPUs in the group, therefore mark all groups
7139 * higher than ourself as busy.
7141 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7145 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7152 #ifdef CONFIG_NUMA_BALANCING
7153 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7155 if (sgs->sum_nr_running > sgs->nr_numa_running)
7157 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7162 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7164 if (rq->nr_running > rq->nr_numa_running)
7166 if (rq->nr_running > rq->nr_preferred_running)
7171 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7176 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7180 #endif /* CONFIG_NUMA_BALANCING */
7183 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7184 * @env: The load balancing environment.
7185 * @sds: variable to hold the statistics for this sched_domain.
7187 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7189 struct sched_domain *child = env->sd->child;
7190 struct sched_group *sg = env->sd->groups;
7191 struct sg_lb_stats tmp_sgs;
7192 int load_idx, prefer_sibling = 0;
7193 bool overload = false, overutilized = false;
7195 if (child && child->flags & SD_PREFER_SIBLING)
7198 load_idx = get_sd_load_idx(env->sd, env->idle);
7201 struct sg_lb_stats *sgs = &tmp_sgs;
7204 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7207 sgs = &sds->local_stat;
7209 if (env->idle != CPU_NEWLY_IDLE ||
7210 time_after_eq(jiffies, sg->sgc->next_update))
7211 update_group_capacity(env->sd, env->dst_cpu);
7214 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7215 &overload, &overutilized);
7221 * In case the child domain prefers tasks go to siblings
7222 * first, lower the sg capacity so that we'll try
7223 * and move all the excess tasks away. We lower the capacity
7224 * of a group only if the local group has the capacity to fit
7225 * these excess tasks. The extra check prevents the case where
7226 * you always pull from the heaviest group when it is already
7227 * under-utilized (possible with a large weight task outweighs
7228 * the tasks on the system).
7230 if (prefer_sibling && sds->local &&
7231 group_has_capacity(env, &sds->local_stat) &&
7232 (sgs->sum_nr_running > 1)) {
7233 sgs->group_no_capacity = 1;
7234 sgs->group_type = group_classify(sg, sgs);
7238 * Ignore task groups with misfit tasks if local group has no
7239 * capacity or if per-cpu capacity isn't higher.
7241 if (sgs->group_type == group_misfit_task &&
7242 (!group_has_capacity(env, &sds->local_stat) ||
7243 !group_smaller_cpu_capacity(sg, sds->local)))
7244 sgs->group_type = group_other;
7246 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7248 sds->busiest_stat = *sgs;
7252 /* Now, start updating sd_lb_stats */
7253 sds->total_load += sgs->group_load;
7254 sds->total_capacity += sgs->group_capacity;
7257 } while (sg != env->sd->groups);
7259 if (env->sd->flags & SD_NUMA)
7260 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7262 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7264 if (!env->sd->parent) {
7265 /* update overload indicator if we are at root domain */
7266 if (env->dst_rq->rd->overload != overload)
7267 env->dst_rq->rd->overload = overload;
7269 /* Update over-utilization (tipping point, U >= 0) indicator */
7270 if (env->dst_rq->rd->overutilized != overutilized)
7271 env->dst_rq->rd->overutilized = overutilized;
7273 if (!env->dst_rq->rd->overutilized && overutilized)
7274 env->dst_rq->rd->overutilized = true;
7279 * check_asym_packing - Check to see if the group is packed into the
7282 * This is primarily intended to used at the sibling level. Some
7283 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7284 * case of POWER7, it can move to lower SMT modes only when higher
7285 * threads are idle. When in lower SMT modes, the threads will
7286 * perform better since they share less core resources. Hence when we
7287 * have idle threads, we want them to be the higher ones.
7289 * This packing function is run on idle threads. It checks to see if
7290 * the busiest CPU in this domain (core in the P7 case) has a higher
7291 * CPU number than the packing function is being run on. Here we are
7292 * assuming lower CPU number will be equivalent to lower a SMT thread
7295 * Return: 1 when packing is required and a task should be moved to
7296 * this CPU. The amount of the imbalance is returned in *imbalance.
7298 * @env: The load balancing environment.
7299 * @sds: Statistics of the sched_domain which is to be packed
7301 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7305 if (!(env->sd->flags & SD_ASYM_PACKING))
7311 busiest_cpu = group_first_cpu(sds->busiest);
7312 if (env->dst_cpu > busiest_cpu)
7315 env->imbalance = DIV_ROUND_CLOSEST(
7316 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7317 SCHED_CAPACITY_SCALE);
7323 * fix_small_imbalance - Calculate the minor imbalance that exists
7324 * amongst the groups of a sched_domain, during
7326 * @env: The load balancing environment.
7327 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7330 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7332 unsigned long tmp, capa_now = 0, capa_move = 0;
7333 unsigned int imbn = 2;
7334 unsigned long scaled_busy_load_per_task;
7335 struct sg_lb_stats *local, *busiest;
7337 local = &sds->local_stat;
7338 busiest = &sds->busiest_stat;
7340 if (!local->sum_nr_running)
7341 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7342 else if (busiest->load_per_task > local->load_per_task)
7345 scaled_busy_load_per_task =
7346 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7347 busiest->group_capacity;
7349 if (busiest->avg_load + scaled_busy_load_per_task >=
7350 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7351 env->imbalance = busiest->load_per_task;
7356 * OK, we don't have enough imbalance to justify moving tasks,
7357 * however we may be able to increase total CPU capacity used by
7361 capa_now += busiest->group_capacity *
7362 min(busiest->load_per_task, busiest->avg_load);
7363 capa_now += local->group_capacity *
7364 min(local->load_per_task, local->avg_load);
7365 capa_now /= SCHED_CAPACITY_SCALE;
7367 /* Amount of load we'd subtract */
7368 if (busiest->avg_load > scaled_busy_load_per_task) {
7369 capa_move += busiest->group_capacity *
7370 min(busiest->load_per_task,
7371 busiest->avg_load - scaled_busy_load_per_task);
7374 /* Amount of load we'd add */
7375 if (busiest->avg_load * busiest->group_capacity <
7376 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7377 tmp = (busiest->avg_load * busiest->group_capacity) /
7378 local->group_capacity;
7380 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7381 local->group_capacity;
7383 capa_move += local->group_capacity *
7384 min(local->load_per_task, local->avg_load + tmp);
7385 capa_move /= SCHED_CAPACITY_SCALE;
7387 /* Move if we gain throughput */
7388 if (capa_move > capa_now)
7389 env->imbalance = busiest->load_per_task;
7393 * calculate_imbalance - Calculate the amount of imbalance present within the
7394 * groups of a given sched_domain during load balance.
7395 * @env: load balance environment
7396 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7398 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7400 unsigned long max_pull, load_above_capacity = ~0UL;
7401 struct sg_lb_stats *local, *busiest;
7403 local = &sds->local_stat;
7404 busiest = &sds->busiest_stat;
7406 if (busiest->group_type == group_imbalanced) {
7408 * In the group_imb case we cannot rely on group-wide averages
7409 * to ensure cpu-load equilibrium, look at wider averages. XXX
7411 busiest->load_per_task =
7412 min(busiest->load_per_task, sds->avg_load);
7416 * In the presence of smp nice balancing, certain scenarios can have
7417 * max load less than avg load(as we skip the groups at or below
7418 * its cpu_capacity, while calculating max_load..)
7420 if (busiest->avg_load <= sds->avg_load ||
7421 local->avg_load >= sds->avg_load) {
7422 /* Misfitting tasks should be migrated in any case */
7423 if (busiest->group_type == group_misfit_task) {
7424 env->imbalance = busiest->group_misfit_task;
7429 * Busiest group is overloaded, local is not, use the spare
7430 * cycles to maximize throughput
7432 if (busiest->group_type == group_overloaded &&
7433 local->group_type <= group_misfit_task) {
7434 env->imbalance = busiest->load_per_task;
7439 return fix_small_imbalance(env, sds);
7443 * If there aren't any idle cpus, avoid creating some.
7445 if (busiest->group_type == group_overloaded &&
7446 local->group_type == group_overloaded) {
7447 load_above_capacity = busiest->sum_nr_running *
7449 if (load_above_capacity > busiest->group_capacity)
7450 load_above_capacity -= busiest->group_capacity;
7452 load_above_capacity = ~0UL;
7456 * We're trying to get all the cpus to the average_load, so we don't
7457 * want to push ourselves above the average load, nor do we wish to
7458 * reduce the max loaded cpu below the average load. At the same time,
7459 * we also don't want to reduce the group load below the group capacity
7460 * (so that we can implement power-savings policies etc). Thus we look
7461 * for the minimum possible imbalance.
7463 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7465 /* How much load to actually move to equalise the imbalance */
7466 env->imbalance = min(
7467 max_pull * busiest->group_capacity,
7468 (sds->avg_load - local->avg_load) * local->group_capacity
7469 ) / SCHED_CAPACITY_SCALE;
7471 /* Boost imbalance to allow misfit task to be balanced. */
7472 if (busiest->group_type == group_misfit_task)
7473 env->imbalance = max_t(long, env->imbalance,
7474 busiest->group_misfit_task);
7477 * if *imbalance is less than the average load per runnable task
7478 * there is no guarantee that any tasks will be moved so we'll have
7479 * a think about bumping its value to force at least one task to be
7482 if (env->imbalance < busiest->load_per_task)
7483 return fix_small_imbalance(env, sds);
7486 /******* find_busiest_group() helpers end here *********************/
7489 * find_busiest_group - Returns the busiest group within the sched_domain
7490 * if there is an imbalance. If there isn't an imbalance, and
7491 * the user has opted for power-savings, it returns a group whose
7492 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7493 * such a group exists.
7495 * Also calculates the amount of weighted load which should be moved
7496 * to restore balance.
7498 * @env: The load balancing environment.
7500 * Return: - The busiest group if imbalance exists.
7501 * - If no imbalance and user has opted for power-savings balance,
7502 * return the least loaded group whose CPUs can be
7503 * put to idle by rebalancing its tasks onto our group.
7505 static struct sched_group *find_busiest_group(struct lb_env *env)
7507 struct sg_lb_stats *local, *busiest;
7508 struct sd_lb_stats sds;
7510 init_sd_lb_stats(&sds);
7513 * Compute the various statistics relavent for load balancing at
7516 update_sd_lb_stats(env, &sds);
7518 if (energy_aware() && !env->dst_rq->rd->overutilized)
7521 local = &sds.local_stat;
7522 busiest = &sds.busiest_stat;
7524 /* ASYM feature bypasses nice load balance check */
7525 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7526 check_asym_packing(env, &sds))
7529 /* There is no busy sibling group to pull tasks from */
7530 if (!sds.busiest || busiest->sum_nr_running == 0)
7533 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7534 / sds.total_capacity;
7537 * If the busiest group is imbalanced the below checks don't
7538 * work because they assume all things are equal, which typically
7539 * isn't true due to cpus_allowed constraints and the like.
7541 if (busiest->group_type == group_imbalanced)
7544 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7545 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7546 busiest->group_no_capacity)
7549 /* Misfitting tasks should be dealt with regardless of the avg load */
7550 if (busiest->group_type == group_misfit_task) {
7555 * If the local group is busier than the selected busiest group
7556 * don't try and pull any tasks.
7558 if (local->avg_load >= busiest->avg_load)
7562 * Don't pull any tasks if this group is already above the domain
7565 if (local->avg_load >= sds.avg_load)
7568 if (env->idle == CPU_IDLE) {
7570 * This cpu is idle. If the busiest group is not overloaded
7571 * and there is no imbalance between this and busiest group
7572 * wrt idle cpus, it is balanced. The imbalance becomes
7573 * significant if the diff is greater than 1 otherwise we
7574 * might end up to just move the imbalance on another group
7576 if ((busiest->group_type != group_overloaded) &&
7577 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7578 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7582 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7583 * imbalance_pct to be conservative.
7585 if (100 * busiest->avg_load <=
7586 env->sd->imbalance_pct * local->avg_load)
7591 env->busiest_group_type = busiest->group_type;
7592 /* Looks like there is an imbalance. Compute it */
7593 calculate_imbalance(env, &sds);
7602 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7604 static struct rq *find_busiest_queue(struct lb_env *env,
7605 struct sched_group *group)
7607 struct rq *busiest = NULL, *rq;
7608 unsigned long busiest_load = 0, busiest_capacity = 1;
7611 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7612 unsigned long capacity, wl;
7616 rt = fbq_classify_rq(rq);
7619 * We classify groups/runqueues into three groups:
7620 * - regular: there are !numa tasks
7621 * - remote: there are numa tasks that run on the 'wrong' node
7622 * - all: there is no distinction
7624 * In order to avoid migrating ideally placed numa tasks,
7625 * ignore those when there's better options.
7627 * If we ignore the actual busiest queue to migrate another
7628 * task, the next balance pass can still reduce the busiest
7629 * queue by moving tasks around inside the node.
7631 * If we cannot move enough load due to this classification
7632 * the next pass will adjust the group classification and
7633 * allow migration of more tasks.
7635 * Both cases only affect the total convergence complexity.
7637 if (rt > env->fbq_type)
7640 capacity = capacity_of(i);
7642 wl = weighted_cpuload(i);
7645 * When comparing with imbalance, use weighted_cpuload()
7646 * which is not scaled with the cpu capacity.
7649 if (rq->nr_running == 1 && wl > env->imbalance &&
7650 !check_cpu_capacity(rq, env->sd) &&
7651 env->busiest_group_type != group_misfit_task)
7655 * For the load comparisons with the other cpu's, consider
7656 * the weighted_cpuload() scaled with the cpu capacity, so
7657 * that the load can be moved away from the cpu that is
7658 * potentially running at a lower capacity.
7660 * Thus we're looking for max(wl_i / capacity_i), crosswise
7661 * multiplication to rid ourselves of the division works out
7662 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7663 * our previous maximum.
7665 if (wl * busiest_capacity > busiest_load * capacity) {
7667 busiest_capacity = capacity;
7676 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7677 * so long as it is large enough.
7679 #define MAX_PINNED_INTERVAL 512
7681 /* Working cpumask for load_balance and load_balance_newidle. */
7682 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7684 static int need_active_balance(struct lb_env *env)
7686 struct sched_domain *sd = env->sd;
7688 if (env->idle == CPU_NEWLY_IDLE) {
7691 * ASYM_PACKING needs to force migrate tasks from busy but
7692 * higher numbered CPUs in order to pack all tasks in the
7693 * lowest numbered CPUs.
7695 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7700 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7701 * It's worth migrating the task if the src_cpu's capacity is reduced
7702 * because of other sched_class or IRQs if more capacity stays
7703 * available on dst_cpu.
7705 if ((env->idle != CPU_NOT_IDLE) &&
7706 (env->src_rq->cfs.h_nr_running == 1)) {
7707 if ((check_cpu_capacity(env->src_rq, sd)) &&
7708 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7712 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
7713 env->src_rq->cfs.h_nr_running == 1 &&
7714 cpu_overutilized(env->src_cpu) &&
7715 !cpu_overutilized(env->dst_cpu)) {
7719 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7722 static int active_load_balance_cpu_stop(void *data);
7724 static int should_we_balance(struct lb_env *env)
7726 struct sched_group *sg = env->sd->groups;
7727 struct cpumask *sg_cpus, *sg_mask;
7728 int cpu, balance_cpu = -1;
7731 * In the newly idle case, we will allow all the cpu's
7732 * to do the newly idle load balance.
7734 if (env->idle == CPU_NEWLY_IDLE)
7737 sg_cpus = sched_group_cpus(sg);
7738 sg_mask = sched_group_mask(sg);
7739 /* Try to find first idle cpu */
7740 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7741 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7748 if (balance_cpu == -1)
7749 balance_cpu = group_balance_cpu(sg);
7752 * First idle cpu or the first cpu(busiest) in this sched group
7753 * is eligible for doing load balancing at this and above domains.
7755 return balance_cpu == env->dst_cpu;
7759 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7760 * tasks if there is an imbalance.
7762 static int load_balance(int this_cpu, struct rq *this_rq,
7763 struct sched_domain *sd, enum cpu_idle_type idle,
7764 int *continue_balancing)
7766 int ld_moved, cur_ld_moved, active_balance = 0;
7767 struct sched_domain *sd_parent = sd->parent;
7768 struct sched_group *group;
7770 unsigned long flags;
7771 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7773 struct lb_env env = {
7775 .dst_cpu = this_cpu,
7777 .dst_grpmask = sched_group_cpus(sd->groups),
7779 .loop_break = sched_nr_migrate_break,
7782 .tasks = LIST_HEAD_INIT(env.tasks),
7786 * For NEWLY_IDLE load_balancing, we don't need to consider
7787 * other cpus in our group
7789 if (idle == CPU_NEWLY_IDLE)
7790 env.dst_grpmask = NULL;
7792 cpumask_copy(cpus, cpu_active_mask);
7794 schedstat_inc(sd, lb_count[idle]);
7797 if (!should_we_balance(&env)) {
7798 *continue_balancing = 0;
7802 group = find_busiest_group(&env);
7804 schedstat_inc(sd, lb_nobusyg[idle]);
7808 busiest = find_busiest_queue(&env, group);
7810 schedstat_inc(sd, lb_nobusyq[idle]);
7814 BUG_ON(busiest == env.dst_rq);
7816 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7818 env.src_cpu = busiest->cpu;
7819 env.src_rq = busiest;
7822 if (busiest->nr_running > 1) {
7824 * Attempt to move tasks. If find_busiest_group has found
7825 * an imbalance but busiest->nr_running <= 1, the group is
7826 * still unbalanced. ld_moved simply stays zero, so it is
7827 * correctly treated as an imbalance.
7829 env.flags |= LBF_ALL_PINNED;
7830 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7833 raw_spin_lock_irqsave(&busiest->lock, flags);
7836 * cur_ld_moved - load moved in current iteration
7837 * ld_moved - cumulative load moved across iterations
7839 cur_ld_moved = detach_tasks(&env);
7841 * We want to potentially lower env.src_cpu's OPP.
7844 update_capacity_of(env.src_cpu);
7847 * We've detached some tasks from busiest_rq. Every
7848 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7849 * unlock busiest->lock, and we are able to be sure
7850 * that nobody can manipulate the tasks in parallel.
7851 * See task_rq_lock() family for the details.
7854 raw_spin_unlock(&busiest->lock);
7858 ld_moved += cur_ld_moved;
7861 local_irq_restore(flags);
7863 if (env.flags & LBF_NEED_BREAK) {
7864 env.flags &= ~LBF_NEED_BREAK;
7869 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7870 * us and move them to an alternate dst_cpu in our sched_group
7871 * where they can run. The upper limit on how many times we
7872 * iterate on same src_cpu is dependent on number of cpus in our
7875 * This changes load balance semantics a bit on who can move
7876 * load to a given_cpu. In addition to the given_cpu itself
7877 * (or a ilb_cpu acting on its behalf where given_cpu is
7878 * nohz-idle), we now have balance_cpu in a position to move
7879 * load to given_cpu. In rare situations, this may cause
7880 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7881 * _independently_ and at _same_ time to move some load to
7882 * given_cpu) causing exceess load to be moved to given_cpu.
7883 * This however should not happen so much in practice and
7884 * moreover subsequent load balance cycles should correct the
7885 * excess load moved.
7887 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7889 /* Prevent to re-select dst_cpu via env's cpus */
7890 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7892 env.dst_rq = cpu_rq(env.new_dst_cpu);
7893 env.dst_cpu = env.new_dst_cpu;
7894 env.flags &= ~LBF_DST_PINNED;
7896 env.loop_break = sched_nr_migrate_break;
7899 * Go back to "more_balance" rather than "redo" since we
7900 * need to continue with same src_cpu.
7906 * We failed to reach balance because of affinity.
7909 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7911 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7912 *group_imbalance = 1;
7915 /* All tasks on this runqueue were pinned by CPU affinity */
7916 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7917 cpumask_clear_cpu(cpu_of(busiest), cpus);
7918 if (!cpumask_empty(cpus)) {
7920 env.loop_break = sched_nr_migrate_break;
7923 goto out_all_pinned;
7928 schedstat_inc(sd, lb_failed[idle]);
7930 * Increment the failure counter only on periodic balance.
7931 * We do not want newidle balance, which can be very
7932 * frequent, pollute the failure counter causing
7933 * excessive cache_hot migrations and active balances.
7935 if (idle != CPU_NEWLY_IDLE)
7936 if (env.src_grp_nr_running > 1)
7937 sd->nr_balance_failed++;
7939 if (need_active_balance(&env)) {
7940 raw_spin_lock_irqsave(&busiest->lock, flags);
7942 /* don't kick the active_load_balance_cpu_stop,
7943 * if the curr task on busiest cpu can't be
7946 if (!cpumask_test_cpu(this_cpu,
7947 tsk_cpus_allowed(busiest->curr))) {
7948 raw_spin_unlock_irqrestore(&busiest->lock,
7950 env.flags |= LBF_ALL_PINNED;
7951 goto out_one_pinned;
7955 * ->active_balance synchronizes accesses to
7956 * ->active_balance_work. Once set, it's cleared
7957 * only after active load balance is finished.
7959 if (!busiest->active_balance) {
7960 busiest->active_balance = 1;
7961 busiest->push_cpu = this_cpu;
7964 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7966 if (active_balance) {
7967 stop_one_cpu_nowait(cpu_of(busiest),
7968 active_load_balance_cpu_stop, busiest,
7969 &busiest->active_balance_work);
7973 * We've kicked active balancing, reset the failure
7976 sd->nr_balance_failed = sd->cache_nice_tries+1;
7979 sd->nr_balance_failed = 0;
7981 if (likely(!active_balance)) {
7982 /* We were unbalanced, so reset the balancing interval */
7983 sd->balance_interval = sd->min_interval;
7986 * If we've begun active balancing, start to back off. This
7987 * case may not be covered by the all_pinned logic if there
7988 * is only 1 task on the busy runqueue (because we don't call
7991 if (sd->balance_interval < sd->max_interval)
7992 sd->balance_interval *= 2;
7999 * We reach balance although we may have faced some affinity
8000 * constraints. Clear the imbalance flag if it was set.
8003 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8005 if (*group_imbalance)
8006 *group_imbalance = 0;
8011 * We reach balance because all tasks are pinned at this level so
8012 * we can't migrate them. Let the imbalance flag set so parent level
8013 * can try to migrate them.
8015 schedstat_inc(sd, lb_balanced[idle]);
8017 sd->nr_balance_failed = 0;
8020 /* tune up the balancing interval */
8021 if (((env.flags & LBF_ALL_PINNED) &&
8022 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8023 (sd->balance_interval < sd->max_interval))
8024 sd->balance_interval *= 2;
8031 static inline unsigned long
8032 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8034 unsigned long interval = sd->balance_interval;
8037 interval *= sd->busy_factor;
8039 /* scale ms to jiffies */
8040 interval = msecs_to_jiffies(interval);
8041 interval = clamp(interval, 1UL, max_load_balance_interval);
8047 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8049 unsigned long interval, next;
8051 interval = get_sd_balance_interval(sd, cpu_busy);
8052 next = sd->last_balance + interval;
8054 if (time_after(*next_balance, next))
8055 *next_balance = next;
8059 * idle_balance is called by schedule() if this_cpu is about to become
8060 * idle. Attempts to pull tasks from other CPUs.
8062 static int idle_balance(struct rq *this_rq)
8064 unsigned long next_balance = jiffies + HZ;
8065 int this_cpu = this_rq->cpu;
8066 struct sched_domain *sd;
8067 int pulled_task = 0;
8070 idle_enter_fair(this_rq);
8073 * We must set idle_stamp _before_ calling idle_balance(), such that we
8074 * measure the duration of idle_balance() as idle time.
8076 this_rq->idle_stamp = rq_clock(this_rq);
8078 if (!energy_aware() &&
8079 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8080 !this_rq->rd->overload)) {
8082 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8084 update_next_balance(sd, 0, &next_balance);
8090 raw_spin_unlock(&this_rq->lock);
8092 update_blocked_averages(this_cpu);
8094 for_each_domain(this_cpu, sd) {
8095 int continue_balancing = 1;
8096 u64 t0, domain_cost;
8098 if (!(sd->flags & SD_LOAD_BALANCE))
8101 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8102 update_next_balance(sd, 0, &next_balance);
8106 if (sd->flags & SD_BALANCE_NEWIDLE) {
8107 t0 = sched_clock_cpu(this_cpu);
8109 pulled_task = load_balance(this_cpu, this_rq,
8111 &continue_balancing);
8113 domain_cost = sched_clock_cpu(this_cpu) - t0;
8114 if (domain_cost > sd->max_newidle_lb_cost)
8115 sd->max_newidle_lb_cost = domain_cost;
8117 curr_cost += domain_cost;
8120 update_next_balance(sd, 0, &next_balance);
8123 * Stop searching for tasks to pull if there are
8124 * now runnable tasks on this rq.
8126 if (pulled_task || this_rq->nr_running > 0)
8131 raw_spin_lock(&this_rq->lock);
8133 if (curr_cost > this_rq->max_idle_balance_cost)
8134 this_rq->max_idle_balance_cost = curr_cost;
8137 * While browsing the domains, we released the rq lock, a task could
8138 * have been enqueued in the meantime. Since we're not going idle,
8139 * pretend we pulled a task.
8141 if (this_rq->cfs.h_nr_running && !pulled_task)
8145 /* Move the next balance forward */
8146 if (time_after(this_rq->next_balance, next_balance))
8147 this_rq->next_balance = next_balance;
8149 /* Is there a task of a high priority class? */
8150 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8154 idle_exit_fair(this_rq);
8155 this_rq->idle_stamp = 0;
8162 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8163 * running tasks off the busiest CPU onto idle CPUs. It requires at
8164 * least 1 task to be running on each physical CPU where possible, and
8165 * avoids physical / logical imbalances.
8167 static int active_load_balance_cpu_stop(void *data)
8169 struct rq *busiest_rq = data;
8170 int busiest_cpu = cpu_of(busiest_rq);
8171 int target_cpu = busiest_rq->push_cpu;
8172 struct rq *target_rq = cpu_rq(target_cpu);
8173 struct sched_domain *sd;
8174 struct task_struct *p = NULL;
8176 raw_spin_lock_irq(&busiest_rq->lock);
8178 /* make sure the requested cpu hasn't gone down in the meantime */
8179 if (unlikely(busiest_cpu != smp_processor_id() ||
8180 !busiest_rq->active_balance))
8183 /* Is there any task to move? */
8184 if (busiest_rq->nr_running <= 1)
8188 * This condition is "impossible", if it occurs
8189 * we need to fix it. Originally reported by
8190 * Bjorn Helgaas on a 128-cpu setup.
8192 BUG_ON(busiest_rq == target_rq);
8194 /* Search for an sd spanning us and the target CPU. */
8196 for_each_domain(target_cpu, sd) {
8197 if ((sd->flags & SD_LOAD_BALANCE) &&
8198 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8203 struct lb_env env = {
8205 .dst_cpu = target_cpu,
8206 .dst_rq = target_rq,
8207 .src_cpu = busiest_rq->cpu,
8208 .src_rq = busiest_rq,
8212 schedstat_inc(sd, alb_count);
8214 p = detach_one_task(&env);
8216 schedstat_inc(sd, alb_pushed);
8218 * We want to potentially lower env.src_cpu's OPP.
8220 update_capacity_of(env.src_cpu);
8223 schedstat_inc(sd, alb_failed);
8227 busiest_rq->active_balance = 0;
8228 raw_spin_unlock(&busiest_rq->lock);
8231 attach_one_task(target_rq, p);
8238 static inline int on_null_domain(struct rq *rq)
8240 return unlikely(!rcu_dereference_sched(rq->sd));
8243 #ifdef CONFIG_NO_HZ_COMMON
8245 * idle load balancing details
8246 * - When one of the busy CPUs notice that there may be an idle rebalancing
8247 * needed, they will kick the idle load balancer, which then does idle
8248 * load balancing for all the idle CPUs.
8251 cpumask_var_t idle_cpus_mask;
8253 unsigned long next_balance; /* in jiffy units */
8254 } nohz ____cacheline_aligned;
8256 static inline int find_new_ilb(void)
8258 int ilb = cpumask_first(nohz.idle_cpus_mask);
8260 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8267 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8268 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8269 * CPU (if there is one).
8271 static void nohz_balancer_kick(void)
8275 nohz.next_balance++;
8277 ilb_cpu = find_new_ilb();
8279 if (ilb_cpu >= nr_cpu_ids)
8282 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8285 * Use smp_send_reschedule() instead of resched_cpu().
8286 * This way we generate a sched IPI on the target cpu which
8287 * is idle. And the softirq performing nohz idle load balance
8288 * will be run before returning from the IPI.
8290 smp_send_reschedule(ilb_cpu);
8294 static inline void nohz_balance_exit_idle(int cpu)
8296 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8298 * Completely isolated CPUs don't ever set, so we must test.
8300 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8301 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8302 atomic_dec(&nohz.nr_cpus);
8304 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8308 static inline void set_cpu_sd_state_busy(void)
8310 struct sched_domain *sd;
8311 int cpu = smp_processor_id();
8314 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8316 if (!sd || !sd->nohz_idle)
8320 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8325 void set_cpu_sd_state_idle(void)
8327 struct sched_domain *sd;
8328 int cpu = smp_processor_id();
8331 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8333 if (!sd || sd->nohz_idle)
8337 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8343 * This routine will record that the cpu is going idle with tick stopped.
8344 * This info will be used in performing idle load balancing in the future.
8346 void nohz_balance_enter_idle(int cpu)
8349 * If this cpu is going down, then nothing needs to be done.
8351 if (!cpu_active(cpu))
8354 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8358 * If we're a completely isolated CPU, we don't play.
8360 if (on_null_domain(cpu_rq(cpu)))
8363 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8364 atomic_inc(&nohz.nr_cpus);
8365 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8368 static int sched_ilb_notifier(struct notifier_block *nfb,
8369 unsigned long action, void *hcpu)
8371 switch (action & ~CPU_TASKS_FROZEN) {
8373 nohz_balance_exit_idle(smp_processor_id());
8381 static DEFINE_SPINLOCK(balancing);
8384 * Scale the max load_balance interval with the number of CPUs in the system.
8385 * This trades load-balance latency on larger machines for less cross talk.
8387 void update_max_interval(void)
8389 max_load_balance_interval = HZ*num_online_cpus()/10;
8393 * It checks each scheduling domain to see if it is due to be balanced,
8394 * and initiates a balancing operation if so.
8396 * Balancing parameters are set up in init_sched_domains.
8398 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8400 int continue_balancing = 1;
8402 unsigned long interval;
8403 struct sched_domain *sd;
8404 /* Earliest time when we have to do rebalance again */
8405 unsigned long next_balance = jiffies + 60*HZ;
8406 int update_next_balance = 0;
8407 int need_serialize, need_decay = 0;
8410 update_blocked_averages(cpu);
8413 for_each_domain(cpu, sd) {
8415 * Decay the newidle max times here because this is a regular
8416 * visit to all the domains. Decay ~1% per second.
8418 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8419 sd->max_newidle_lb_cost =
8420 (sd->max_newidle_lb_cost * 253) / 256;
8421 sd->next_decay_max_lb_cost = jiffies + HZ;
8424 max_cost += sd->max_newidle_lb_cost;
8426 if (!(sd->flags & SD_LOAD_BALANCE))
8430 * Stop the load balance at this level. There is another
8431 * CPU in our sched group which is doing load balancing more
8434 if (!continue_balancing) {
8440 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8442 need_serialize = sd->flags & SD_SERIALIZE;
8443 if (need_serialize) {
8444 if (!spin_trylock(&balancing))
8448 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8449 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8451 * The LBF_DST_PINNED logic could have changed
8452 * env->dst_cpu, so we can't know our idle
8453 * state even if we migrated tasks. Update it.
8455 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8457 sd->last_balance = jiffies;
8458 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8461 spin_unlock(&balancing);
8463 if (time_after(next_balance, sd->last_balance + interval)) {
8464 next_balance = sd->last_balance + interval;
8465 update_next_balance = 1;
8470 * Ensure the rq-wide value also decays but keep it at a
8471 * reasonable floor to avoid funnies with rq->avg_idle.
8473 rq->max_idle_balance_cost =
8474 max((u64)sysctl_sched_migration_cost, max_cost);
8479 * next_balance will be updated only when there is a need.
8480 * When the cpu is attached to null domain for ex, it will not be
8483 if (likely(update_next_balance)) {
8484 rq->next_balance = next_balance;
8486 #ifdef CONFIG_NO_HZ_COMMON
8488 * If this CPU has been elected to perform the nohz idle
8489 * balance. Other idle CPUs have already rebalanced with
8490 * nohz_idle_balance() and nohz.next_balance has been
8491 * updated accordingly. This CPU is now running the idle load
8492 * balance for itself and we need to update the
8493 * nohz.next_balance accordingly.
8495 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8496 nohz.next_balance = rq->next_balance;
8501 #ifdef CONFIG_NO_HZ_COMMON
8503 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8504 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8506 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8508 int this_cpu = this_rq->cpu;
8511 /* Earliest time when we have to do rebalance again */
8512 unsigned long next_balance = jiffies + 60*HZ;
8513 int update_next_balance = 0;
8515 if (idle != CPU_IDLE ||
8516 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8519 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8520 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8524 * If this cpu gets work to do, stop the load balancing
8525 * work being done for other cpus. Next load
8526 * balancing owner will pick it up.
8531 rq = cpu_rq(balance_cpu);
8534 * If time for next balance is due,
8537 if (time_after_eq(jiffies, rq->next_balance)) {
8538 raw_spin_lock_irq(&rq->lock);
8539 update_rq_clock(rq);
8540 update_idle_cpu_load(rq);
8541 raw_spin_unlock_irq(&rq->lock);
8542 rebalance_domains(rq, CPU_IDLE);
8545 if (time_after(next_balance, rq->next_balance)) {
8546 next_balance = rq->next_balance;
8547 update_next_balance = 1;
8552 * next_balance will be updated only when there is a need.
8553 * When the CPU is attached to null domain for ex, it will not be
8556 if (likely(update_next_balance))
8557 nohz.next_balance = next_balance;
8559 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8563 * Current heuristic for kicking the idle load balancer in the presence
8564 * of an idle cpu in the system.
8565 * - This rq has more than one task.
8566 * - This rq has at least one CFS task and the capacity of the CPU is
8567 * significantly reduced because of RT tasks or IRQs.
8568 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8569 * multiple busy cpu.
8570 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8571 * domain span are idle.
8573 static inline bool nohz_kick_needed(struct rq *rq)
8575 unsigned long now = jiffies;
8576 struct sched_domain *sd;
8577 struct sched_group_capacity *sgc;
8578 int nr_busy, cpu = rq->cpu;
8581 if (unlikely(rq->idle_balance))
8585 * We may be recently in ticked or tickless idle mode. At the first
8586 * busy tick after returning from idle, we will update the busy stats.
8588 set_cpu_sd_state_busy();
8589 nohz_balance_exit_idle(cpu);
8592 * None are in tickless mode and hence no need for NOHZ idle load
8595 if (likely(!atomic_read(&nohz.nr_cpus)))
8598 if (time_before(now, nohz.next_balance))
8601 if (rq->nr_running >= 2 &&
8602 (!energy_aware() || cpu_overutilized(cpu)))
8606 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8607 if (sd && !energy_aware()) {
8608 sgc = sd->groups->sgc;
8609 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8618 sd = rcu_dereference(rq->sd);
8620 if ((rq->cfs.h_nr_running >= 1) &&
8621 check_cpu_capacity(rq, sd)) {
8627 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8628 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8629 sched_domain_span(sd)) < cpu)) {
8639 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8643 * run_rebalance_domains is triggered when needed from the scheduler tick.
8644 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8646 static void run_rebalance_domains(struct softirq_action *h)
8648 struct rq *this_rq = this_rq();
8649 enum cpu_idle_type idle = this_rq->idle_balance ?
8650 CPU_IDLE : CPU_NOT_IDLE;
8653 * If this cpu has a pending nohz_balance_kick, then do the
8654 * balancing on behalf of the other idle cpus whose ticks are
8655 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8656 * give the idle cpus a chance to load balance. Else we may
8657 * load balance only within the local sched_domain hierarchy
8658 * and abort nohz_idle_balance altogether if we pull some load.
8660 nohz_idle_balance(this_rq, idle);
8661 rebalance_domains(this_rq, idle);
8665 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8667 void trigger_load_balance(struct rq *rq)
8669 /* Don't need to rebalance while attached to NULL domain */
8670 if (unlikely(on_null_domain(rq)))
8673 if (time_after_eq(jiffies, rq->next_balance))
8674 raise_softirq(SCHED_SOFTIRQ);
8675 #ifdef CONFIG_NO_HZ_COMMON
8676 if (nohz_kick_needed(rq))
8677 nohz_balancer_kick();
8681 static void rq_online_fair(struct rq *rq)
8685 update_runtime_enabled(rq);
8688 static void rq_offline_fair(struct rq *rq)
8692 /* Ensure any throttled groups are reachable by pick_next_task */
8693 unthrottle_offline_cfs_rqs(rq);
8696 #endif /* CONFIG_SMP */
8699 * scheduler tick hitting a task of our scheduling class:
8701 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8703 struct cfs_rq *cfs_rq;
8704 struct sched_entity *se = &curr->se;
8706 for_each_sched_entity(se) {
8707 cfs_rq = cfs_rq_of(se);
8708 entity_tick(cfs_rq, se, queued);
8711 if (static_branch_unlikely(&sched_numa_balancing))
8712 task_tick_numa(rq, curr);
8714 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr)))
8715 rq->rd->overutilized = true;
8717 rq->misfit_task = !task_fits_max(curr, rq->cpu);
8721 * called on fork with the child task as argument from the parent's context
8722 * - child not yet on the tasklist
8723 * - preemption disabled
8725 static void task_fork_fair(struct task_struct *p)
8727 struct cfs_rq *cfs_rq;
8728 struct sched_entity *se = &p->se, *curr;
8729 int this_cpu = smp_processor_id();
8730 struct rq *rq = this_rq();
8731 unsigned long flags;
8733 raw_spin_lock_irqsave(&rq->lock, flags);
8735 update_rq_clock(rq);
8737 cfs_rq = task_cfs_rq(current);
8738 curr = cfs_rq->curr;
8741 * Not only the cpu but also the task_group of the parent might have
8742 * been changed after parent->se.parent,cfs_rq were copied to
8743 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8744 * of child point to valid ones.
8747 __set_task_cpu(p, this_cpu);
8750 update_curr(cfs_rq);
8753 se->vruntime = curr->vruntime;
8754 place_entity(cfs_rq, se, 1);
8756 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8758 * Upon rescheduling, sched_class::put_prev_task() will place
8759 * 'current' within the tree based on its new key value.
8761 swap(curr->vruntime, se->vruntime);
8765 se->vruntime -= cfs_rq->min_vruntime;
8767 raw_spin_unlock_irqrestore(&rq->lock, flags);
8771 * Priority of the task has changed. Check to see if we preempt
8775 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8777 if (!task_on_rq_queued(p))
8781 * Reschedule if we are currently running on this runqueue and
8782 * our priority decreased, or if we are not currently running on
8783 * this runqueue and our priority is higher than the current's
8785 if (rq->curr == p) {
8786 if (p->prio > oldprio)
8789 check_preempt_curr(rq, p, 0);
8792 static inline bool vruntime_normalized(struct task_struct *p)
8794 struct sched_entity *se = &p->se;
8797 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8798 * the dequeue_entity(.flags=0) will already have normalized the
8805 * When !on_rq, vruntime of the task has usually NOT been normalized.
8806 * But there are some cases where it has already been normalized:
8808 * - A forked child which is waiting for being woken up by
8809 * wake_up_new_task().
8810 * - A task which has been woken up by try_to_wake_up() and
8811 * waiting for actually being woken up by sched_ttwu_pending().
8813 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8819 static void detach_task_cfs_rq(struct task_struct *p)
8821 struct sched_entity *se = &p->se;
8822 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8824 if (!vruntime_normalized(p)) {
8826 * Fix up our vruntime so that the current sleep doesn't
8827 * cause 'unlimited' sleep bonus.
8829 place_entity(cfs_rq, se, 0);
8830 se->vruntime -= cfs_rq->min_vruntime;
8833 /* Catch up with the cfs_rq and remove our load when we leave */
8834 detach_entity_load_avg(cfs_rq, se);
8837 static void attach_task_cfs_rq(struct task_struct *p)
8839 struct sched_entity *se = &p->se;
8840 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8842 #ifdef CONFIG_FAIR_GROUP_SCHED
8844 * Since the real-depth could have been changed (only FAIR
8845 * class maintain depth value), reset depth properly.
8847 se->depth = se->parent ? se->parent->depth + 1 : 0;
8850 /* Synchronize task with its cfs_rq */
8851 attach_entity_load_avg(cfs_rq, se);
8853 if (!vruntime_normalized(p))
8854 se->vruntime += cfs_rq->min_vruntime;
8857 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8859 detach_task_cfs_rq(p);
8862 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8864 attach_task_cfs_rq(p);
8866 if (task_on_rq_queued(p)) {
8868 * We were most likely switched from sched_rt, so
8869 * kick off the schedule if running, otherwise just see
8870 * if we can still preempt the current task.
8875 check_preempt_curr(rq, p, 0);
8879 /* Account for a task changing its policy or group.
8881 * This routine is mostly called to set cfs_rq->curr field when a task
8882 * migrates between groups/classes.
8884 static void set_curr_task_fair(struct rq *rq)
8886 struct sched_entity *se = &rq->curr->se;
8888 for_each_sched_entity(se) {
8889 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8891 set_next_entity(cfs_rq, se);
8892 /* ensure bandwidth has been allocated on our new cfs_rq */
8893 account_cfs_rq_runtime(cfs_rq, 0);
8897 void init_cfs_rq(struct cfs_rq *cfs_rq)
8899 cfs_rq->tasks_timeline = RB_ROOT;
8900 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8901 #ifndef CONFIG_64BIT
8902 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8905 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8906 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8910 #ifdef CONFIG_FAIR_GROUP_SCHED
8911 static void task_move_group_fair(struct task_struct *p)
8913 detach_task_cfs_rq(p);
8914 set_task_rq(p, task_cpu(p));
8917 /* Tell se's cfs_rq has been changed -- migrated */
8918 p->se.avg.last_update_time = 0;
8920 attach_task_cfs_rq(p);
8923 void free_fair_sched_group(struct task_group *tg)
8927 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8929 for_each_possible_cpu(i) {
8931 kfree(tg->cfs_rq[i]);
8934 remove_entity_load_avg(tg->se[i]);
8943 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8945 struct cfs_rq *cfs_rq;
8946 struct sched_entity *se;
8949 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8952 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8956 tg->shares = NICE_0_LOAD;
8958 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8960 for_each_possible_cpu(i) {
8961 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8962 GFP_KERNEL, cpu_to_node(i));
8966 se = kzalloc_node(sizeof(struct sched_entity),
8967 GFP_KERNEL, cpu_to_node(i));
8971 init_cfs_rq(cfs_rq);
8972 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8973 init_entity_runnable_average(se);
8984 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8986 struct rq *rq = cpu_rq(cpu);
8987 unsigned long flags;
8990 * Only empty task groups can be destroyed; so we can speculatively
8991 * check on_list without danger of it being re-added.
8993 if (!tg->cfs_rq[cpu]->on_list)
8996 raw_spin_lock_irqsave(&rq->lock, flags);
8997 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8998 raw_spin_unlock_irqrestore(&rq->lock, flags);
9001 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9002 struct sched_entity *se, int cpu,
9003 struct sched_entity *parent)
9005 struct rq *rq = cpu_rq(cpu);
9009 init_cfs_rq_runtime(cfs_rq);
9011 tg->cfs_rq[cpu] = cfs_rq;
9014 /* se could be NULL for root_task_group */
9019 se->cfs_rq = &rq->cfs;
9022 se->cfs_rq = parent->my_q;
9023 se->depth = parent->depth + 1;
9027 /* guarantee group entities always have weight */
9028 update_load_set(&se->load, NICE_0_LOAD);
9029 se->parent = parent;
9032 static DEFINE_MUTEX(shares_mutex);
9034 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9037 unsigned long flags;
9040 * We can't change the weight of the root cgroup.
9045 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9047 mutex_lock(&shares_mutex);
9048 if (tg->shares == shares)
9051 tg->shares = shares;
9052 for_each_possible_cpu(i) {
9053 struct rq *rq = cpu_rq(i);
9054 struct sched_entity *se;
9057 /* Propagate contribution to hierarchy */
9058 raw_spin_lock_irqsave(&rq->lock, flags);
9060 /* Possible calls to update_curr() need rq clock */
9061 update_rq_clock(rq);
9062 for_each_sched_entity(se)
9063 update_cfs_shares(group_cfs_rq(se));
9064 raw_spin_unlock_irqrestore(&rq->lock, flags);
9068 mutex_unlock(&shares_mutex);
9071 #else /* CONFIG_FAIR_GROUP_SCHED */
9073 void free_fair_sched_group(struct task_group *tg) { }
9075 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9080 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9082 #endif /* CONFIG_FAIR_GROUP_SCHED */
9085 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9087 struct sched_entity *se = &task->se;
9088 unsigned int rr_interval = 0;
9091 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9094 if (rq->cfs.load.weight)
9095 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9101 * All the scheduling class methods:
9103 const struct sched_class fair_sched_class = {
9104 .next = &idle_sched_class,
9105 .enqueue_task = enqueue_task_fair,
9106 .dequeue_task = dequeue_task_fair,
9107 .yield_task = yield_task_fair,
9108 .yield_to_task = yield_to_task_fair,
9110 .check_preempt_curr = check_preempt_wakeup,
9112 .pick_next_task = pick_next_task_fair,
9113 .put_prev_task = put_prev_task_fair,
9116 .select_task_rq = select_task_rq_fair,
9117 .migrate_task_rq = migrate_task_rq_fair,
9119 .rq_online = rq_online_fair,
9120 .rq_offline = rq_offline_fair,
9122 .task_waking = task_waking_fair,
9123 .task_dead = task_dead_fair,
9124 .set_cpus_allowed = set_cpus_allowed_common,
9127 .set_curr_task = set_curr_task_fair,
9128 .task_tick = task_tick_fair,
9129 .task_fork = task_fork_fair,
9131 .prio_changed = prio_changed_fair,
9132 .switched_from = switched_from_fair,
9133 .switched_to = switched_to_fair,
9135 .get_rr_interval = get_rr_interval_fair,
9137 .update_curr = update_curr_fair,
9139 #ifdef CONFIG_FAIR_GROUP_SCHED
9140 .task_move_group = task_move_group_fair,
9144 #ifdef CONFIG_SCHED_DEBUG
9145 void print_cfs_stats(struct seq_file *m, int cpu)
9147 struct cfs_rq *cfs_rq;
9150 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9151 print_cfs_rq(m, cpu, cfs_rq);
9155 #ifdef CONFIG_NUMA_BALANCING
9156 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9159 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9161 for_each_online_node(node) {
9162 if (p->numa_faults) {
9163 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9164 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9166 if (p->numa_group) {
9167 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9168 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9170 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9173 #endif /* CONFIG_NUMA_BALANCING */
9174 #endif /* CONFIG_SCHED_DEBUG */
9176 __init void init_sched_fair_class(void)
9179 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9181 #ifdef CONFIG_NO_HZ_COMMON
9182 nohz.next_balance = jiffies;
9183 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9184 cpu_notifier(sched_ilb_notifier, 0);