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>
33 #include <linux/module.h>
35 #include <trace/events/sched.h>
42 * Targeted preemption latency for CPU-bound tasks:
43 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
45 * NOTE: this latency value is not the same as the concept of
46 * 'timeslice length' - timeslices in CFS are of variable length
47 * and have no persistent notion like in traditional, time-slice
48 * based scheduling concepts.
50 * (to see the precise effective timeslice length of your workload,
51 * run vmstat and monitor the context-switches (cs) field)
53 unsigned int sysctl_sched_latency = 6000000ULL;
54 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
56 unsigned int sysctl_sched_is_big_little = 0;
57 unsigned int sysctl_sched_sync_hint_enable = 1;
58 unsigned int sysctl_sched_initial_task_util = 0;
59 unsigned int sysctl_sched_cstate_aware = 1;
61 #ifdef CONFIG_SCHED_WALT
62 unsigned int sysctl_sched_use_walt_cpu_util = 1;
63 unsigned int sysctl_sched_use_walt_task_util = 1;
64 __read_mostly unsigned int sysctl_sched_walt_cpu_high_irqload =
68 * The initial- and re-scaling of tunables is configurable
69 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
72 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
73 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
74 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
76 enum sched_tunable_scaling sysctl_sched_tunable_scaling
77 = SCHED_TUNABLESCALING_LOG;
80 * Minimal preemption granularity for CPU-bound tasks:
81 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 unsigned int sysctl_sched_min_granularity = 750000ULL;
84 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
87 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
89 static unsigned int sched_nr_latency = 8;
92 * After fork, child runs first. If set to 0 (default) then
93 * parent will (try to) run first.
95 unsigned int sysctl_sched_child_runs_first __read_mostly;
98 * SCHED_OTHER wake-up granularity.
99 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
101 * This option delays the preemption effects of decoupled workloads
102 * and reduces their over-scheduling. Synchronous workloads will still
103 * have immediate wakeup/sleep latencies.
105 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
106 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
108 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
111 * The exponential sliding window over which load is averaged for shares
115 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
117 #ifdef CONFIG_CFS_BANDWIDTH
119 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
120 * each time a cfs_rq requests quota.
122 * Note: in the case that the slice exceeds the runtime remaining (either due
123 * to consumption or the quota being specified to be smaller than the slice)
124 * we will always only issue the remaining available time.
126 * default: 5 msec, units: microseconds
128 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
131 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
137 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
143 static inline void update_load_set(struct load_weight *lw, unsigned long w)
150 * Increase the granularity value when there are more CPUs,
151 * because with more CPUs the 'effective latency' as visible
152 * to users decreases. But the relationship is not linear,
153 * so pick a second-best guess by going with the log2 of the
156 * This idea comes from the SD scheduler of Con Kolivas:
158 static unsigned int get_update_sysctl_factor(void)
160 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
163 switch (sysctl_sched_tunable_scaling) {
164 case SCHED_TUNABLESCALING_NONE:
167 case SCHED_TUNABLESCALING_LINEAR:
170 case SCHED_TUNABLESCALING_LOG:
172 factor = 1 + ilog2(cpus);
179 static void update_sysctl(void)
181 unsigned int factor = get_update_sysctl_factor();
183 #define SET_SYSCTL(name) \
184 (sysctl_##name = (factor) * normalized_sysctl_##name)
185 SET_SYSCTL(sched_min_granularity);
186 SET_SYSCTL(sched_latency);
187 SET_SYSCTL(sched_wakeup_granularity);
191 void sched_init_granularity(void)
196 #define WMULT_CONST (~0U)
197 #define WMULT_SHIFT 32
199 static void __update_inv_weight(struct load_weight *lw)
203 if (likely(lw->inv_weight))
206 w = scale_load_down(lw->weight);
208 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
210 else if (unlikely(!w))
211 lw->inv_weight = WMULT_CONST;
213 lw->inv_weight = WMULT_CONST / w;
217 * delta_exec * weight / lw.weight
219 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
221 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
222 * we're guaranteed shift stays positive because inv_weight is guaranteed to
223 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
225 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
226 * weight/lw.weight <= 1, and therefore our shift will also be positive.
228 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
230 u64 fact = scale_load_down(weight);
231 int shift = WMULT_SHIFT;
233 __update_inv_weight(lw);
235 if (unlikely(fact >> 32)) {
242 /* hint to use a 32x32->64 mul */
243 fact = (u64)(u32)fact * lw->inv_weight;
250 return mul_u64_u32_shr(delta_exec, fact, shift);
254 const struct sched_class fair_sched_class;
256 /**************************************************************
257 * CFS operations on generic schedulable entities:
260 #ifdef CONFIG_FAIR_GROUP_SCHED
262 /* cpu runqueue to which this cfs_rq is attached */
263 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
268 /* An entity is a task if it doesn't "own" a runqueue */
269 #define entity_is_task(se) (!se->my_q)
271 static inline struct task_struct *task_of(struct sched_entity *se)
273 #ifdef CONFIG_SCHED_DEBUG
274 WARN_ON_ONCE(!entity_is_task(se));
276 return container_of(se, struct task_struct, se);
279 /* Walk up scheduling entities hierarchy */
280 #define for_each_sched_entity(se) \
281 for (; se; se = se->parent)
283 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
288 /* runqueue on which this entity is (to be) queued */
289 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
294 /* runqueue "owned" by this group */
295 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
300 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
302 if (!cfs_rq->on_list) {
304 * Ensure we either appear before our parent (if already
305 * enqueued) or force our parent to appear after us when it is
306 * enqueued. The fact that we always enqueue bottom-up
307 * reduces this to two cases.
309 if (cfs_rq->tg->parent &&
310 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
311 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
312 &rq_of(cfs_rq)->leaf_cfs_rq_list);
314 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
315 &rq_of(cfs_rq)->leaf_cfs_rq_list);
322 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
324 if (cfs_rq->on_list) {
325 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
330 /* Iterate thr' all leaf cfs_rq's on a runqueue */
331 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
332 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
334 /* Do the two (enqueued) entities belong to the same group ? */
335 static inline struct cfs_rq *
336 is_same_group(struct sched_entity *se, struct sched_entity *pse)
338 if (se->cfs_rq == pse->cfs_rq)
344 static inline struct sched_entity *parent_entity(struct sched_entity *se)
350 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
352 int se_depth, pse_depth;
355 * preemption test can be made between sibling entities who are in the
356 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
357 * both tasks until we find their ancestors who are siblings of common
361 /* First walk up until both entities are at same depth */
362 se_depth = (*se)->depth;
363 pse_depth = (*pse)->depth;
365 while (se_depth > pse_depth) {
367 *se = parent_entity(*se);
370 while (pse_depth > se_depth) {
372 *pse = parent_entity(*pse);
375 while (!is_same_group(*se, *pse)) {
376 *se = parent_entity(*se);
377 *pse = parent_entity(*pse);
381 #else /* !CONFIG_FAIR_GROUP_SCHED */
383 static inline struct task_struct *task_of(struct sched_entity *se)
385 return container_of(se, struct task_struct, se);
388 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
390 return container_of(cfs_rq, struct rq, cfs);
393 #define entity_is_task(se) 1
395 #define for_each_sched_entity(se) \
396 for (; se; se = NULL)
398 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
400 return &task_rq(p)->cfs;
403 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
405 struct task_struct *p = task_of(se);
406 struct rq *rq = task_rq(p);
411 /* runqueue "owned" by this group */
412 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
417 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
421 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
425 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
426 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 static inline struct sched_entity *parent_entity(struct sched_entity *se)
434 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
438 #endif /* CONFIG_FAIR_GROUP_SCHED */
440 static __always_inline
441 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
443 /**************************************************************
444 * Scheduling class tree data structure manipulation methods:
447 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
449 s64 delta = (s64)(vruntime - max_vruntime);
451 max_vruntime = vruntime;
456 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
458 s64 delta = (s64)(vruntime - min_vruntime);
460 min_vruntime = vruntime;
465 static inline int entity_before(struct sched_entity *a,
466 struct sched_entity *b)
468 return (s64)(a->vruntime - b->vruntime) < 0;
471 static void update_min_vruntime(struct cfs_rq *cfs_rq)
473 u64 vruntime = cfs_rq->min_vruntime;
476 vruntime = cfs_rq->curr->vruntime;
478 if (cfs_rq->rb_leftmost) {
479 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
484 vruntime = se->vruntime;
486 vruntime = min_vruntime(vruntime, se->vruntime);
489 /* ensure we never gain time by being placed backwards. */
490 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
493 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
498 * Enqueue an entity into the rb-tree:
500 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
502 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
503 struct rb_node *parent = NULL;
504 struct sched_entity *entry;
508 * Find the right place in the rbtree:
512 entry = rb_entry(parent, struct sched_entity, run_node);
514 * We dont care about collisions. Nodes with
515 * the same key stay together.
517 if (entity_before(se, entry)) {
518 link = &parent->rb_left;
520 link = &parent->rb_right;
526 * Maintain a cache of leftmost tree entries (it is frequently
530 cfs_rq->rb_leftmost = &se->run_node;
532 rb_link_node(&se->run_node, parent, link);
533 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
536 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
538 if (cfs_rq->rb_leftmost == &se->run_node) {
539 struct rb_node *next_node;
541 next_node = rb_next(&se->run_node);
542 cfs_rq->rb_leftmost = next_node;
545 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
548 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
550 struct rb_node *left = cfs_rq->rb_leftmost;
555 return rb_entry(left, struct sched_entity, run_node);
558 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
560 struct rb_node *next = rb_next(&se->run_node);
565 return rb_entry(next, struct sched_entity, run_node);
568 #ifdef CONFIG_SCHED_DEBUG
569 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
571 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
576 return rb_entry(last, struct sched_entity, run_node);
579 /**************************************************************
580 * Scheduling class statistics methods:
583 int sched_proc_update_handler(struct ctl_table *table, int write,
584 void __user *buffer, size_t *lenp,
587 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
588 unsigned int factor = get_update_sysctl_factor();
593 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
594 sysctl_sched_min_granularity);
596 #define WRT_SYSCTL(name) \
597 (normalized_sysctl_##name = sysctl_##name / (factor))
598 WRT_SYSCTL(sched_min_granularity);
599 WRT_SYSCTL(sched_latency);
600 WRT_SYSCTL(sched_wakeup_granularity);
610 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
612 if (unlikely(se->load.weight != NICE_0_LOAD))
613 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
619 * The idea is to set a period in which each task runs once.
621 * When there are too many tasks (sched_nr_latency) we have to stretch
622 * this period because otherwise the slices get too small.
624 * p = (nr <= nl) ? l : l*nr/nl
626 static u64 __sched_period(unsigned long nr_running)
628 if (unlikely(nr_running > sched_nr_latency))
629 return nr_running * sysctl_sched_min_granularity;
631 return sysctl_sched_latency;
635 * We calculate the wall-time slice from the period by taking a part
636 * proportional to the weight.
640 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
642 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
644 for_each_sched_entity(se) {
645 struct load_weight *load;
646 struct load_weight lw;
648 cfs_rq = cfs_rq_of(se);
649 load = &cfs_rq->load;
651 if (unlikely(!se->on_rq)) {
654 update_load_add(&lw, se->load.weight);
657 slice = __calc_delta(slice, se->load.weight, load);
663 * We calculate the vruntime slice of a to-be-inserted task.
667 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
669 return calc_delta_fair(sched_slice(cfs_rq, se), se);
673 static int select_idle_sibling(struct task_struct *p, int cpu);
674 static unsigned long task_h_load(struct task_struct *p);
677 * We choose a half-life close to 1 scheduling period.
678 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
679 * dependent on this value.
681 #define LOAD_AVG_PERIOD 32
682 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
683 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
685 /* Give new sched_entity start runnable values to heavy its load in infant time */
686 void init_entity_runnable_average(struct sched_entity *se)
688 struct sched_avg *sa = &se->avg;
690 sa->last_update_time = 0;
692 * sched_avg's period_contrib should be strictly less then 1024, so
693 * we give it 1023 to make sure it is almost a period (1024us), and
694 * will definitely be update (after enqueue).
696 sa->period_contrib = 1023;
697 sa->load_avg = scale_load_down(se->load.weight);
698 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
699 sa->util_avg = sched_freq() ?
700 sysctl_sched_initial_task_util :
701 scale_load_down(SCHED_LOAD_SCALE);
702 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
703 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
707 void init_entity_runnable_average(struct sched_entity *se)
713 * Update the current task's runtime statistics.
715 static void update_curr(struct cfs_rq *cfs_rq)
717 struct sched_entity *curr = cfs_rq->curr;
718 u64 now = rq_clock_task(rq_of(cfs_rq));
724 delta_exec = now - curr->exec_start;
725 if (unlikely((s64)delta_exec <= 0))
728 curr->exec_start = now;
730 schedstat_set(curr->statistics.exec_max,
731 max(delta_exec, curr->statistics.exec_max));
733 curr->sum_exec_runtime += delta_exec;
734 schedstat_add(cfs_rq, exec_clock, delta_exec);
736 curr->vruntime += calc_delta_fair(delta_exec, curr);
737 update_min_vruntime(cfs_rq);
739 if (entity_is_task(curr)) {
740 struct task_struct *curtask = task_of(curr);
742 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
743 cpuacct_charge(curtask, delta_exec);
744 account_group_exec_runtime(curtask, delta_exec);
747 account_cfs_rq_runtime(cfs_rq, delta_exec);
750 static void update_curr_fair(struct rq *rq)
752 update_curr(cfs_rq_of(&rq->curr->se));
756 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
758 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
762 * Task is being enqueued - update stats:
764 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
767 * Are we enqueueing a waiting task? (for current tasks
768 * a dequeue/enqueue event is a NOP)
770 if (se != cfs_rq->curr)
771 update_stats_wait_start(cfs_rq, se);
775 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
777 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
778 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
779 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
780 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
781 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
782 #ifdef CONFIG_SCHEDSTATS
783 if (entity_is_task(se)) {
784 trace_sched_stat_wait(task_of(se),
785 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
788 schedstat_set(se->statistics.wait_start, 0);
792 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
795 * Mark the end of the wait period if dequeueing a
798 if (se != cfs_rq->curr)
799 update_stats_wait_end(cfs_rq, se);
803 * We are picking a new current task - update its stats:
806 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
809 * We are starting a new run period:
811 se->exec_start = rq_clock_task(rq_of(cfs_rq));
814 /**************************************************
815 * Scheduling class queueing methods:
818 #ifdef CONFIG_NUMA_BALANCING
820 * Approximate time to scan a full NUMA task in ms. The task scan period is
821 * calculated based on the tasks virtual memory size and
822 * numa_balancing_scan_size.
824 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
825 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
827 /* Portion of address space to scan in MB */
828 unsigned int sysctl_numa_balancing_scan_size = 256;
830 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
831 unsigned int sysctl_numa_balancing_scan_delay = 1000;
833 static unsigned int task_nr_scan_windows(struct task_struct *p)
835 unsigned long rss = 0;
836 unsigned long nr_scan_pages;
839 * Calculations based on RSS as non-present and empty pages are skipped
840 * by the PTE scanner and NUMA hinting faults should be trapped based
843 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
844 rss = get_mm_rss(p->mm);
848 rss = round_up(rss, nr_scan_pages);
849 return rss / nr_scan_pages;
852 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
853 #define MAX_SCAN_WINDOW 2560
855 static unsigned int task_scan_min(struct task_struct *p)
857 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
858 unsigned int scan, floor;
859 unsigned int windows = 1;
861 if (scan_size < MAX_SCAN_WINDOW)
862 windows = MAX_SCAN_WINDOW / scan_size;
863 floor = 1000 / windows;
865 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
866 return max_t(unsigned int, floor, scan);
869 static unsigned int task_scan_max(struct task_struct *p)
871 unsigned int smin = task_scan_min(p);
874 /* Watch for min being lower than max due to floor calculations */
875 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
876 return max(smin, smax);
879 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
881 rq->nr_numa_running += (p->numa_preferred_nid != -1);
882 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
885 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
887 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
888 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
894 spinlock_t lock; /* nr_tasks, tasks */
899 nodemask_t active_nodes;
900 unsigned long total_faults;
902 * Faults_cpu is used to decide whether memory should move
903 * towards the CPU. As a consequence, these stats are weighted
904 * more by CPU use than by memory faults.
906 unsigned long *faults_cpu;
907 unsigned long faults[0];
910 /* Shared or private faults. */
911 #define NR_NUMA_HINT_FAULT_TYPES 2
913 /* Memory and CPU locality */
914 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
916 /* Averaged statistics, and temporary buffers. */
917 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
919 pid_t task_numa_group_id(struct task_struct *p)
921 return p->numa_group ? p->numa_group->gid : 0;
925 * The averaged statistics, shared & private, memory & cpu,
926 * occupy the first half of the array. The second half of the
927 * array is for current counters, which are averaged into the
928 * first set by task_numa_placement.
930 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
932 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
935 static inline unsigned long task_faults(struct task_struct *p, int nid)
940 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
941 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
944 static inline unsigned long group_faults(struct task_struct *p, int nid)
949 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
950 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
953 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
955 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
956 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
959 /* Handle placement on systems where not all nodes are directly connected. */
960 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
961 int maxdist, bool task)
963 unsigned long score = 0;
967 * All nodes are directly connected, and the same distance
968 * from each other. No need for fancy placement algorithms.
970 if (sched_numa_topology_type == NUMA_DIRECT)
974 * This code is called for each node, introducing N^2 complexity,
975 * which should be ok given the number of nodes rarely exceeds 8.
977 for_each_online_node(node) {
978 unsigned long faults;
979 int dist = node_distance(nid, node);
982 * The furthest away nodes in the system are not interesting
983 * for placement; nid was already counted.
985 if (dist == sched_max_numa_distance || node == nid)
989 * On systems with a backplane NUMA topology, compare groups
990 * of nodes, and move tasks towards the group with the most
991 * memory accesses. When comparing two nodes at distance
992 * "hoplimit", only nodes closer by than "hoplimit" are part
993 * of each group. Skip other nodes.
995 if (sched_numa_topology_type == NUMA_BACKPLANE &&
999 /* Add up the faults from nearby nodes. */
1001 faults = task_faults(p, node);
1003 faults = group_faults(p, node);
1006 * On systems with a glueless mesh NUMA topology, there are
1007 * no fixed "groups of nodes". Instead, nodes that are not
1008 * directly connected bounce traffic through intermediate
1009 * nodes; a numa_group can occupy any set of nodes.
1010 * The further away a node is, the less the faults count.
1011 * This seems to result in good task placement.
1013 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1014 faults *= (sched_max_numa_distance - dist);
1015 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1025 * These return the fraction of accesses done by a particular task, or
1026 * task group, on a particular numa node. The group weight is given a
1027 * larger multiplier, in order to group tasks together that are almost
1028 * evenly spread out between numa nodes.
1030 static inline unsigned long task_weight(struct task_struct *p, int nid,
1033 unsigned long faults, total_faults;
1035 if (!p->numa_faults)
1038 total_faults = p->total_numa_faults;
1043 faults = task_faults(p, nid);
1044 faults += score_nearby_nodes(p, nid, dist, true);
1046 return 1000 * faults / total_faults;
1049 static inline unsigned long group_weight(struct task_struct *p, int nid,
1052 unsigned long faults, total_faults;
1057 total_faults = p->numa_group->total_faults;
1062 faults = group_faults(p, nid);
1063 faults += score_nearby_nodes(p, nid, dist, false);
1065 return 1000 * faults / total_faults;
1068 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1069 int src_nid, int dst_cpu)
1071 struct numa_group *ng = p->numa_group;
1072 int dst_nid = cpu_to_node(dst_cpu);
1073 int last_cpupid, this_cpupid;
1075 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1078 * Multi-stage node selection is used in conjunction with a periodic
1079 * migration fault to build a temporal task<->page relation. By using
1080 * a two-stage filter we remove short/unlikely relations.
1082 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1083 * a task's usage of a particular page (n_p) per total usage of this
1084 * page (n_t) (in a given time-span) to a probability.
1086 * Our periodic faults will sample this probability and getting the
1087 * same result twice in a row, given these samples are fully
1088 * independent, is then given by P(n)^2, provided our sample period
1089 * is sufficiently short compared to the usage pattern.
1091 * This quadric squishes small probabilities, making it less likely we
1092 * act on an unlikely task<->page relation.
1094 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1095 if (!cpupid_pid_unset(last_cpupid) &&
1096 cpupid_to_nid(last_cpupid) != dst_nid)
1099 /* Always allow migrate on private faults */
1100 if (cpupid_match_pid(p, last_cpupid))
1103 /* A shared fault, but p->numa_group has not been set up yet. */
1108 * Do not migrate if the destination is not a node that
1109 * is actively used by this numa group.
1111 if (!node_isset(dst_nid, ng->active_nodes))
1115 * Source is a node that is not actively used by this
1116 * numa group, while the destination is. Migrate.
1118 if (!node_isset(src_nid, ng->active_nodes))
1122 * Both source and destination are nodes in active
1123 * use by this numa group. Maximize memory bandwidth
1124 * by migrating from more heavily used groups, to less
1125 * heavily used ones, spreading the load around.
1126 * Use a 1/4 hysteresis to avoid spurious page movement.
1128 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1131 static unsigned long weighted_cpuload(const int cpu);
1132 static unsigned long source_load(int cpu, int type);
1133 static unsigned long target_load(int cpu, int type);
1134 static unsigned long capacity_of(int cpu);
1135 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1137 /* Cached statistics for all CPUs within a node */
1139 unsigned long nr_running;
1142 /* Total compute capacity of CPUs on a node */
1143 unsigned long compute_capacity;
1145 /* Approximate capacity in terms of runnable tasks on a node */
1146 unsigned long task_capacity;
1147 int has_free_capacity;
1151 * XXX borrowed from update_sg_lb_stats
1153 static void update_numa_stats(struct numa_stats *ns, int nid)
1155 int smt, cpu, cpus = 0;
1156 unsigned long capacity;
1158 memset(ns, 0, sizeof(*ns));
1159 for_each_cpu(cpu, cpumask_of_node(nid)) {
1160 struct rq *rq = cpu_rq(cpu);
1162 ns->nr_running += rq->nr_running;
1163 ns->load += weighted_cpuload(cpu);
1164 ns->compute_capacity += capacity_of(cpu);
1170 * If we raced with hotplug and there are no CPUs left in our mask
1171 * the @ns structure is NULL'ed and task_numa_compare() will
1172 * not find this node attractive.
1174 * We'll either bail at !has_free_capacity, or we'll detect a huge
1175 * imbalance and bail there.
1180 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1181 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1182 capacity = cpus / smt; /* cores */
1184 ns->task_capacity = min_t(unsigned, capacity,
1185 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1186 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1189 struct task_numa_env {
1190 struct task_struct *p;
1192 int src_cpu, src_nid;
1193 int dst_cpu, dst_nid;
1195 struct numa_stats src_stats, dst_stats;
1200 struct task_struct *best_task;
1205 static void task_numa_assign(struct task_numa_env *env,
1206 struct task_struct *p, long imp)
1209 put_task_struct(env->best_task);
1214 env->best_imp = imp;
1215 env->best_cpu = env->dst_cpu;
1218 static bool load_too_imbalanced(long src_load, long dst_load,
1219 struct task_numa_env *env)
1222 long orig_src_load, orig_dst_load;
1223 long src_capacity, dst_capacity;
1226 * The load is corrected for the CPU capacity available on each node.
1229 * ------------ vs ---------
1230 * src_capacity dst_capacity
1232 src_capacity = env->src_stats.compute_capacity;
1233 dst_capacity = env->dst_stats.compute_capacity;
1235 /* We care about the slope of the imbalance, not the direction. */
1236 if (dst_load < src_load)
1237 swap(dst_load, src_load);
1239 /* Is the difference below the threshold? */
1240 imb = dst_load * src_capacity * 100 -
1241 src_load * dst_capacity * env->imbalance_pct;
1246 * The imbalance is above the allowed threshold.
1247 * Compare it with the old imbalance.
1249 orig_src_load = env->src_stats.load;
1250 orig_dst_load = env->dst_stats.load;
1252 if (orig_dst_load < orig_src_load)
1253 swap(orig_dst_load, orig_src_load);
1255 old_imb = orig_dst_load * src_capacity * 100 -
1256 orig_src_load * dst_capacity * env->imbalance_pct;
1258 /* Would this change make things worse? */
1259 return (imb > old_imb);
1263 * This checks if the overall compute and NUMA accesses of the system would
1264 * be improved if the source tasks was migrated to the target dst_cpu taking
1265 * into account that it might be best if task running on the dst_cpu should
1266 * be exchanged with the source task
1268 static void task_numa_compare(struct task_numa_env *env,
1269 long taskimp, long groupimp)
1271 struct rq *src_rq = cpu_rq(env->src_cpu);
1272 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1273 struct task_struct *cur;
1274 long src_load, dst_load;
1276 long imp = env->p->numa_group ? groupimp : taskimp;
1278 int dist = env->dist;
1282 raw_spin_lock_irq(&dst_rq->lock);
1285 * No need to move the exiting task, and this ensures that ->curr
1286 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1287 * is safe under RCU read lock.
1288 * Note that rcu_read_lock() itself can't protect from the final
1289 * put_task_struct() after the last schedule().
1291 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1293 raw_spin_unlock_irq(&dst_rq->lock);
1296 * Because we have preemption enabled we can get migrated around and
1297 * end try selecting ourselves (current == env->p) as a swap candidate.
1303 * "imp" is the fault differential for the source task between the
1304 * source and destination node. Calculate the total differential for
1305 * the source task and potential destination task. The more negative
1306 * the value is, the more rmeote accesses that would be expected to
1307 * be incurred if the tasks were swapped.
1310 /* Skip this swap candidate if cannot move to the source cpu */
1311 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1315 * If dst and source tasks are in the same NUMA group, or not
1316 * in any group then look only at task weights.
1318 if (cur->numa_group == env->p->numa_group) {
1319 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1320 task_weight(cur, env->dst_nid, dist);
1322 * Add some hysteresis to prevent swapping the
1323 * tasks within a group over tiny differences.
1325 if (cur->numa_group)
1329 * Compare the group weights. If a task is all by
1330 * itself (not part of a group), use the task weight
1333 if (cur->numa_group)
1334 imp += group_weight(cur, env->src_nid, dist) -
1335 group_weight(cur, env->dst_nid, dist);
1337 imp += task_weight(cur, env->src_nid, dist) -
1338 task_weight(cur, env->dst_nid, dist);
1342 if (imp <= env->best_imp && moveimp <= env->best_imp)
1346 /* Is there capacity at our destination? */
1347 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1348 !env->dst_stats.has_free_capacity)
1354 /* Balance doesn't matter much if we're running a task per cpu */
1355 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1356 dst_rq->nr_running == 1)
1360 * In the overloaded case, try and keep the load balanced.
1363 load = task_h_load(env->p);
1364 dst_load = env->dst_stats.load + load;
1365 src_load = env->src_stats.load - load;
1367 if (moveimp > imp && moveimp > env->best_imp) {
1369 * If the improvement from just moving env->p direction is
1370 * better than swapping tasks around, check if a move is
1371 * possible. Store a slightly smaller score than moveimp,
1372 * so an actually idle CPU will win.
1374 if (!load_too_imbalanced(src_load, dst_load, env)) {
1381 if (imp <= env->best_imp)
1385 load = task_h_load(cur);
1390 if (load_too_imbalanced(src_load, dst_load, env))
1394 * One idle CPU per node is evaluated for a task numa move.
1395 * Call select_idle_sibling to maybe find a better one.
1398 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1401 task_numa_assign(env, cur, imp);
1406 static void task_numa_find_cpu(struct task_numa_env *env,
1407 long taskimp, long groupimp)
1411 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1412 /* Skip this CPU if the source task cannot migrate */
1413 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1417 task_numa_compare(env, taskimp, groupimp);
1421 /* Only move tasks to a NUMA node less busy than the current node. */
1422 static bool numa_has_capacity(struct task_numa_env *env)
1424 struct numa_stats *src = &env->src_stats;
1425 struct numa_stats *dst = &env->dst_stats;
1427 if (src->has_free_capacity && !dst->has_free_capacity)
1431 * Only consider a task move if the source has a higher load
1432 * than the destination, corrected for CPU capacity on each node.
1434 * src->load dst->load
1435 * --------------------- vs ---------------------
1436 * src->compute_capacity dst->compute_capacity
1438 if (src->load * dst->compute_capacity * env->imbalance_pct >
1440 dst->load * src->compute_capacity * 100)
1446 static int task_numa_migrate(struct task_struct *p)
1448 struct task_numa_env env = {
1451 .src_cpu = task_cpu(p),
1452 .src_nid = task_node(p),
1454 .imbalance_pct = 112,
1460 struct sched_domain *sd;
1461 unsigned long taskweight, groupweight;
1463 long taskimp, groupimp;
1466 * Pick the lowest SD_NUMA domain, as that would have the smallest
1467 * imbalance and would be the first to start moving tasks about.
1469 * And we want to avoid any moving of tasks about, as that would create
1470 * random movement of tasks -- counter the numa conditions we're trying
1474 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1476 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1480 * Cpusets can break the scheduler domain tree into smaller
1481 * balance domains, some of which do not cross NUMA boundaries.
1482 * Tasks that are "trapped" in such domains cannot be migrated
1483 * elsewhere, so there is no point in (re)trying.
1485 if (unlikely(!sd)) {
1486 p->numa_preferred_nid = task_node(p);
1490 env.dst_nid = p->numa_preferred_nid;
1491 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1492 taskweight = task_weight(p, env.src_nid, dist);
1493 groupweight = group_weight(p, env.src_nid, dist);
1494 update_numa_stats(&env.src_stats, env.src_nid);
1495 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1496 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1497 update_numa_stats(&env.dst_stats, env.dst_nid);
1499 /* Try to find a spot on the preferred nid. */
1500 if (numa_has_capacity(&env))
1501 task_numa_find_cpu(&env, taskimp, groupimp);
1504 * Look at other nodes in these cases:
1505 * - there is no space available on the preferred_nid
1506 * - the task is part of a numa_group that is interleaved across
1507 * multiple NUMA nodes; in order to better consolidate the group,
1508 * we need to check other locations.
1510 if (env.best_cpu == -1 || (p->numa_group &&
1511 nodes_weight(p->numa_group->active_nodes) > 1)) {
1512 for_each_online_node(nid) {
1513 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1516 dist = node_distance(env.src_nid, env.dst_nid);
1517 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1519 taskweight = task_weight(p, env.src_nid, dist);
1520 groupweight = group_weight(p, env.src_nid, dist);
1523 /* Only consider nodes where both task and groups benefit */
1524 taskimp = task_weight(p, nid, dist) - taskweight;
1525 groupimp = group_weight(p, nid, dist) - groupweight;
1526 if (taskimp < 0 && groupimp < 0)
1531 update_numa_stats(&env.dst_stats, env.dst_nid);
1532 if (numa_has_capacity(&env))
1533 task_numa_find_cpu(&env, taskimp, groupimp);
1538 * If the task is part of a workload that spans multiple NUMA nodes,
1539 * and is migrating into one of the workload's active nodes, remember
1540 * this node as the task's preferred numa node, so the workload can
1542 * A task that migrated to a second choice node will be better off
1543 * trying for a better one later. Do not set the preferred node here.
1545 if (p->numa_group) {
1546 if (env.best_cpu == -1)
1551 if (node_isset(nid, p->numa_group->active_nodes))
1552 sched_setnuma(p, env.dst_nid);
1555 /* No better CPU than the current one was found. */
1556 if (env.best_cpu == -1)
1560 * Reset the scan period if the task is being rescheduled on an
1561 * alternative node to recheck if the tasks is now properly placed.
1563 p->numa_scan_period = task_scan_min(p);
1565 if (env.best_task == NULL) {
1566 ret = migrate_task_to(p, env.best_cpu);
1568 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1572 ret = migrate_swap(p, env.best_task);
1574 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1575 put_task_struct(env.best_task);
1579 /* Attempt to migrate a task to a CPU on the preferred node. */
1580 static void numa_migrate_preferred(struct task_struct *p)
1582 unsigned long interval = HZ;
1584 /* This task has no NUMA fault statistics yet */
1585 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1588 /* Periodically retry migrating the task to the preferred node */
1589 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1590 p->numa_migrate_retry = jiffies + interval;
1592 /* Success if task is already running on preferred CPU */
1593 if (task_node(p) == p->numa_preferred_nid)
1596 /* Otherwise, try migrate to a CPU on the preferred node */
1597 task_numa_migrate(p);
1601 * Find the nodes on which the workload is actively running. We do this by
1602 * tracking the nodes from which NUMA hinting faults are triggered. This can
1603 * be different from the set of nodes where the workload's memory is currently
1606 * The bitmask is used to make smarter decisions on when to do NUMA page
1607 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1608 * are added when they cause over 6/16 of the maximum number of faults, but
1609 * only removed when they drop below 3/16.
1611 static void update_numa_active_node_mask(struct numa_group *numa_group)
1613 unsigned long faults, max_faults = 0;
1616 for_each_online_node(nid) {
1617 faults = group_faults_cpu(numa_group, nid);
1618 if (faults > max_faults)
1619 max_faults = faults;
1622 for_each_online_node(nid) {
1623 faults = group_faults_cpu(numa_group, nid);
1624 if (!node_isset(nid, numa_group->active_nodes)) {
1625 if (faults > max_faults * 6 / 16)
1626 node_set(nid, numa_group->active_nodes);
1627 } else if (faults < max_faults * 3 / 16)
1628 node_clear(nid, numa_group->active_nodes);
1633 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1634 * increments. The more local the fault statistics are, the higher the scan
1635 * period will be for the next scan window. If local/(local+remote) ratio is
1636 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1637 * the scan period will decrease. Aim for 70% local accesses.
1639 #define NUMA_PERIOD_SLOTS 10
1640 #define NUMA_PERIOD_THRESHOLD 7
1643 * Increase the scan period (slow down scanning) if the majority of
1644 * our memory is already on our local node, or if the majority of
1645 * the page accesses are shared with other processes.
1646 * Otherwise, decrease the scan period.
1648 static void update_task_scan_period(struct task_struct *p,
1649 unsigned long shared, unsigned long private)
1651 unsigned int period_slot;
1655 unsigned long remote = p->numa_faults_locality[0];
1656 unsigned long local = p->numa_faults_locality[1];
1659 * If there were no record hinting faults then either the task is
1660 * completely idle or all activity is areas that are not of interest
1661 * to automatic numa balancing. Related to that, if there were failed
1662 * migration then it implies we are migrating too quickly or the local
1663 * node is overloaded. In either case, scan slower
1665 if (local + shared == 0 || p->numa_faults_locality[2]) {
1666 p->numa_scan_period = min(p->numa_scan_period_max,
1667 p->numa_scan_period << 1);
1669 p->mm->numa_next_scan = jiffies +
1670 msecs_to_jiffies(p->numa_scan_period);
1676 * Prepare to scale scan period relative to the current period.
1677 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1678 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1679 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1681 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1682 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1683 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1684 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1687 diff = slot * period_slot;
1689 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1692 * Scale scan rate increases based on sharing. There is an
1693 * inverse relationship between the degree of sharing and
1694 * the adjustment made to the scanning period. Broadly
1695 * speaking the intent is that there is little point
1696 * scanning faster if shared accesses dominate as it may
1697 * simply bounce migrations uselessly
1699 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1700 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1703 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1704 task_scan_min(p), task_scan_max(p));
1705 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1709 * Get the fraction of time the task has been running since the last
1710 * NUMA placement cycle. The scheduler keeps similar statistics, but
1711 * decays those on a 32ms period, which is orders of magnitude off
1712 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1713 * stats only if the task is so new there are no NUMA statistics yet.
1715 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1717 u64 runtime, delta, now;
1718 /* Use the start of this time slice to avoid calculations. */
1719 now = p->se.exec_start;
1720 runtime = p->se.sum_exec_runtime;
1722 if (p->last_task_numa_placement) {
1723 delta = runtime - p->last_sum_exec_runtime;
1724 *period = now - p->last_task_numa_placement;
1726 delta = p->se.avg.load_sum / p->se.load.weight;
1727 *period = LOAD_AVG_MAX;
1730 p->last_sum_exec_runtime = runtime;
1731 p->last_task_numa_placement = now;
1737 * Determine the preferred nid for a task in a numa_group. This needs to
1738 * be done in a way that produces consistent results with group_weight,
1739 * otherwise workloads might not converge.
1741 static int preferred_group_nid(struct task_struct *p, int nid)
1746 /* Direct connections between all NUMA nodes. */
1747 if (sched_numa_topology_type == NUMA_DIRECT)
1751 * On a system with glueless mesh NUMA topology, group_weight
1752 * scores nodes according to the number of NUMA hinting faults on
1753 * both the node itself, and on nearby nodes.
1755 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1756 unsigned long score, max_score = 0;
1757 int node, max_node = nid;
1759 dist = sched_max_numa_distance;
1761 for_each_online_node(node) {
1762 score = group_weight(p, node, dist);
1763 if (score > max_score) {
1772 * Finding the preferred nid in a system with NUMA backplane
1773 * interconnect topology is more involved. The goal is to locate
1774 * tasks from numa_groups near each other in the system, and
1775 * untangle workloads from different sides of the system. This requires
1776 * searching down the hierarchy of node groups, recursively searching
1777 * inside the highest scoring group of nodes. The nodemask tricks
1778 * keep the complexity of the search down.
1780 nodes = node_online_map;
1781 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1782 unsigned long max_faults = 0;
1783 nodemask_t max_group = NODE_MASK_NONE;
1786 /* Are there nodes at this distance from each other? */
1787 if (!find_numa_distance(dist))
1790 for_each_node_mask(a, nodes) {
1791 unsigned long faults = 0;
1792 nodemask_t this_group;
1793 nodes_clear(this_group);
1795 /* Sum group's NUMA faults; includes a==b case. */
1796 for_each_node_mask(b, nodes) {
1797 if (node_distance(a, b) < dist) {
1798 faults += group_faults(p, b);
1799 node_set(b, this_group);
1800 node_clear(b, nodes);
1804 /* Remember the top group. */
1805 if (faults > max_faults) {
1806 max_faults = faults;
1807 max_group = this_group;
1809 * subtle: at the smallest distance there is
1810 * just one node left in each "group", the
1811 * winner is the preferred nid.
1816 /* Next round, evaluate the nodes within max_group. */
1824 static void task_numa_placement(struct task_struct *p)
1826 int seq, nid, max_nid = -1, max_group_nid = -1;
1827 unsigned long max_faults = 0, max_group_faults = 0;
1828 unsigned long fault_types[2] = { 0, 0 };
1829 unsigned long total_faults;
1830 u64 runtime, period;
1831 spinlock_t *group_lock = NULL;
1834 * The p->mm->numa_scan_seq field gets updated without
1835 * exclusive access. Use READ_ONCE() here to ensure
1836 * that the field is read in a single access:
1838 seq = READ_ONCE(p->mm->numa_scan_seq);
1839 if (p->numa_scan_seq == seq)
1841 p->numa_scan_seq = seq;
1842 p->numa_scan_period_max = task_scan_max(p);
1844 total_faults = p->numa_faults_locality[0] +
1845 p->numa_faults_locality[1];
1846 runtime = numa_get_avg_runtime(p, &period);
1848 /* If the task is part of a group prevent parallel updates to group stats */
1849 if (p->numa_group) {
1850 group_lock = &p->numa_group->lock;
1851 spin_lock_irq(group_lock);
1854 /* Find the node with the highest number of faults */
1855 for_each_online_node(nid) {
1856 /* Keep track of the offsets in numa_faults array */
1857 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1858 unsigned long faults = 0, group_faults = 0;
1861 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1862 long diff, f_diff, f_weight;
1864 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1865 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1866 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1867 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1869 /* Decay existing window, copy faults since last scan */
1870 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1871 fault_types[priv] += p->numa_faults[membuf_idx];
1872 p->numa_faults[membuf_idx] = 0;
1875 * Normalize the faults_from, so all tasks in a group
1876 * count according to CPU use, instead of by the raw
1877 * number of faults. Tasks with little runtime have
1878 * little over-all impact on throughput, and thus their
1879 * faults are less important.
1881 f_weight = div64_u64(runtime << 16, period + 1);
1882 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1884 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1885 p->numa_faults[cpubuf_idx] = 0;
1887 p->numa_faults[mem_idx] += diff;
1888 p->numa_faults[cpu_idx] += f_diff;
1889 faults += p->numa_faults[mem_idx];
1890 p->total_numa_faults += diff;
1891 if (p->numa_group) {
1893 * safe because we can only change our own group
1895 * mem_idx represents the offset for a given
1896 * nid and priv in a specific region because it
1897 * is at the beginning of the numa_faults array.
1899 p->numa_group->faults[mem_idx] += diff;
1900 p->numa_group->faults_cpu[mem_idx] += f_diff;
1901 p->numa_group->total_faults += diff;
1902 group_faults += p->numa_group->faults[mem_idx];
1906 if (faults > max_faults) {
1907 max_faults = faults;
1911 if (group_faults > max_group_faults) {
1912 max_group_faults = group_faults;
1913 max_group_nid = nid;
1917 update_task_scan_period(p, fault_types[0], fault_types[1]);
1919 if (p->numa_group) {
1920 update_numa_active_node_mask(p->numa_group);
1921 spin_unlock_irq(group_lock);
1922 max_nid = preferred_group_nid(p, max_group_nid);
1926 /* Set the new preferred node */
1927 if (max_nid != p->numa_preferred_nid)
1928 sched_setnuma(p, max_nid);
1930 if (task_node(p) != p->numa_preferred_nid)
1931 numa_migrate_preferred(p);
1935 static inline int get_numa_group(struct numa_group *grp)
1937 return atomic_inc_not_zero(&grp->refcount);
1940 static inline void put_numa_group(struct numa_group *grp)
1942 if (atomic_dec_and_test(&grp->refcount))
1943 kfree_rcu(grp, rcu);
1946 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1949 struct numa_group *grp, *my_grp;
1950 struct task_struct *tsk;
1952 int cpu = cpupid_to_cpu(cpupid);
1955 if (unlikely(!p->numa_group)) {
1956 unsigned int size = sizeof(struct numa_group) +
1957 4*nr_node_ids*sizeof(unsigned long);
1959 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1963 atomic_set(&grp->refcount, 1);
1964 spin_lock_init(&grp->lock);
1966 /* Second half of the array tracks nids where faults happen */
1967 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1970 node_set(task_node(current), grp->active_nodes);
1972 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1973 grp->faults[i] = p->numa_faults[i];
1975 grp->total_faults = p->total_numa_faults;
1978 rcu_assign_pointer(p->numa_group, grp);
1982 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1984 if (!cpupid_match_pid(tsk, cpupid))
1987 grp = rcu_dereference(tsk->numa_group);
1991 my_grp = p->numa_group;
1996 * Only join the other group if its bigger; if we're the bigger group,
1997 * the other task will join us.
1999 if (my_grp->nr_tasks > grp->nr_tasks)
2003 * Tie-break on the grp address.
2005 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2008 /* Always join threads in the same process. */
2009 if (tsk->mm == current->mm)
2012 /* Simple filter to avoid false positives due to PID collisions */
2013 if (flags & TNF_SHARED)
2016 /* Update priv based on whether false sharing was detected */
2019 if (join && !get_numa_group(grp))
2027 BUG_ON(irqs_disabled());
2028 double_lock_irq(&my_grp->lock, &grp->lock);
2030 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2031 my_grp->faults[i] -= p->numa_faults[i];
2032 grp->faults[i] += p->numa_faults[i];
2034 my_grp->total_faults -= p->total_numa_faults;
2035 grp->total_faults += p->total_numa_faults;
2040 spin_unlock(&my_grp->lock);
2041 spin_unlock_irq(&grp->lock);
2043 rcu_assign_pointer(p->numa_group, grp);
2045 put_numa_group(my_grp);
2053 void task_numa_free(struct task_struct *p)
2055 struct numa_group *grp = p->numa_group;
2056 void *numa_faults = p->numa_faults;
2057 unsigned long flags;
2061 spin_lock_irqsave(&grp->lock, flags);
2062 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2063 grp->faults[i] -= p->numa_faults[i];
2064 grp->total_faults -= p->total_numa_faults;
2067 spin_unlock_irqrestore(&grp->lock, flags);
2068 RCU_INIT_POINTER(p->numa_group, NULL);
2069 put_numa_group(grp);
2072 p->numa_faults = NULL;
2077 * Got a PROT_NONE fault for a page on @node.
2079 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2081 struct task_struct *p = current;
2082 bool migrated = flags & TNF_MIGRATED;
2083 int cpu_node = task_node(current);
2084 int local = !!(flags & TNF_FAULT_LOCAL);
2087 if (!static_branch_likely(&sched_numa_balancing))
2090 /* for example, ksmd faulting in a user's mm */
2094 /* Allocate buffer to track faults on a per-node basis */
2095 if (unlikely(!p->numa_faults)) {
2096 int size = sizeof(*p->numa_faults) *
2097 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2099 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2100 if (!p->numa_faults)
2103 p->total_numa_faults = 0;
2104 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2108 * First accesses are treated as private, otherwise consider accesses
2109 * to be private if the accessing pid has not changed
2111 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2114 priv = cpupid_match_pid(p, last_cpupid);
2115 if (!priv && !(flags & TNF_NO_GROUP))
2116 task_numa_group(p, last_cpupid, flags, &priv);
2120 * If a workload spans multiple NUMA nodes, a shared fault that
2121 * occurs wholly within the set of nodes that the workload is
2122 * actively using should be counted as local. This allows the
2123 * scan rate to slow down when a workload has settled down.
2125 if (!priv && !local && p->numa_group &&
2126 node_isset(cpu_node, p->numa_group->active_nodes) &&
2127 node_isset(mem_node, p->numa_group->active_nodes))
2130 task_numa_placement(p);
2133 * Retry task to preferred node migration periodically, in case it
2134 * case it previously failed, or the scheduler moved us.
2136 if (time_after(jiffies, p->numa_migrate_retry))
2137 numa_migrate_preferred(p);
2140 p->numa_pages_migrated += pages;
2141 if (flags & TNF_MIGRATE_FAIL)
2142 p->numa_faults_locality[2] += pages;
2144 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2145 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2146 p->numa_faults_locality[local] += pages;
2149 static void reset_ptenuma_scan(struct task_struct *p)
2152 * We only did a read acquisition of the mmap sem, so
2153 * p->mm->numa_scan_seq is written to without exclusive access
2154 * and the update is not guaranteed to be atomic. That's not
2155 * much of an issue though, since this is just used for
2156 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2157 * expensive, to avoid any form of compiler optimizations:
2159 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2160 p->mm->numa_scan_offset = 0;
2164 * The expensive part of numa migration is done from task_work context.
2165 * Triggered from task_tick_numa().
2167 void task_numa_work(struct callback_head *work)
2169 unsigned long migrate, next_scan, now = jiffies;
2170 struct task_struct *p = current;
2171 struct mm_struct *mm = p->mm;
2172 struct vm_area_struct *vma;
2173 unsigned long start, end;
2174 unsigned long nr_pte_updates = 0;
2175 long pages, virtpages;
2177 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2179 work->next = work; /* protect against double add */
2181 * Who cares about NUMA placement when they're dying.
2183 * NOTE: make sure not to dereference p->mm before this check,
2184 * exit_task_work() happens _after_ exit_mm() so we could be called
2185 * without p->mm even though we still had it when we enqueued this
2188 if (p->flags & PF_EXITING)
2191 if (!mm->numa_next_scan) {
2192 mm->numa_next_scan = now +
2193 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2197 * Enforce maximal scan/migration frequency..
2199 migrate = mm->numa_next_scan;
2200 if (time_before(now, migrate))
2203 if (p->numa_scan_period == 0) {
2204 p->numa_scan_period_max = task_scan_max(p);
2205 p->numa_scan_period = task_scan_min(p);
2208 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2209 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2213 * Delay this task enough that another task of this mm will likely win
2214 * the next time around.
2216 p->node_stamp += 2 * TICK_NSEC;
2218 start = mm->numa_scan_offset;
2219 pages = sysctl_numa_balancing_scan_size;
2220 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2221 virtpages = pages * 8; /* Scan up to this much virtual space */
2226 down_read(&mm->mmap_sem);
2227 vma = find_vma(mm, start);
2229 reset_ptenuma_scan(p);
2233 for (; vma; vma = vma->vm_next) {
2234 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2235 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2240 * Shared library pages mapped by multiple processes are not
2241 * migrated as it is expected they are cache replicated. Avoid
2242 * hinting faults in read-only file-backed mappings or the vdso
2243 * as migrating the pages will be of marginal benefit.
2246 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2250 * Skip inaccessible VMAs to avoid any confusion between
2251 * PROT_NONE and NUMA hinting ptes
2253 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2257 start = max(start, vma->vm_start);
2258 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2259 end = min(end, vma->vm_end);
2260 nr_pte_updates = change_prot_numa(vma, start, end);
2263 * Try to scan sysctl_numa_balancing_size worth of
2264 * hpages that have at least one present PTE that
2265 * is not already pte-numa. If the VMA contains
2266 * areas that are unused or already full of prot_numa
2267 * PTEs, scan up to virtpages, to skip through those
2271 pages -= (end - start) >> PAGE_SHIFT;
2272 virtpages -= (end - start) >> PAGE_SHIFT;
2275 if (pages <= 0 || virtpages <= 0)
2279 } while (end != vma->vm_end);
2284 * It is possible to reach the end of the VMA list but the last few
2285 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2286 * would find the !migratable VMA on the next scan but not reset the
2287 * scanner to the start so check it now.
2290 mm->numa_scan_offset = start;
2292 reset_ptenuma_scan(p);
2293 up_read(&mm->mmap_sem);
2297 * Drive the periodic memory faults..
2299 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2301 struct callback_head *work = &curr->numa_work;
2305 * We don't care about NUMA placement if we don't have memory.
2307 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2311 * Using runtime rather than walltime has the dual advantage that
2312 * we (mostly) drive the selection from busy threads and that the
2313 * task needs to have done some actual work before we bother with
2316 now = curr->se.sum_exec_runtime;
2317 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2319 if (now > curr->node_stamp + period) {
2320 if (!curr->node_stamp)
2321 curr->numa_scan_period = task_scan_min(curr);
2322 curr->node_stamp += period;
2324 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2325 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2326 task_work_add(curr, work, true);
2331 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2335 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2339 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2342 #endif /* CONFIG_NUMA_BALANCING */
2345 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2347 update_load_add(&cfs_rq->load, se->load.weight);
2348 if (!parent_entity(se))
2349 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2351 if (entity_is_task(se)) {
2352 struct rq *rq = rq_of(cfs_rq);
2354 account_numa_enqueue(rq, task_of(se));
2355 list_add(&se->group_node, &rq->cfs_tasks);
2358 cfs_rq->nr_running++;
2362 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2364 update_load_sub(&cfs_rq->load, se->load.weight);
2365 if (!parent_entity(se))
2366 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2367 if (entity_is_task(se)) {
2368 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2369 list_del_init(&se->group_node);
2371 cfs_rq->nr_running--;
2374 #ifdef CONFIG_FAIR_GROUP_SCHED
2376 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2381 * Use this CPU's real-time load instead of the last load contribution
2382 * as the updating of the contribution is delayed, and we will use the
2383 * the real-time load to calc the share. See update_tg_load_avg().
2385 tg_weight = atomic_long_read(&tg->load_avg);
2386 tg_weight -= cfs_rq->tg_load_avg_contrib;
2387 tg_weight += cfs_rq->load.weight;
2392 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2394 long tg_weight, load, shares;
2396 tg_weight = calc_tg_weight(tg, cfs_rq);
2397 load = cfs_rq->load.weight;
2399 shares = (tg->shares * load);
2401 shares /= tg_weight;
2403 if (shares < MIN_SHARES)
2404 shares = MIN_SHARES;
2405 if (shares > tg->shares)
2406 shares = tg->shares;
2410 # else /* CONFIG_SMP */
2411 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2415 # endif /* CONFIG_SMP */
2416 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2417 unsigned long weight)
2420 /* commit outstanding execution time */
2421 if (cfs_rq->curr == se)
2422 update_curr(cfs_rq);
2423 account_entity_dequeue(cfs_rq, se);
2426 update_load_set(&se->load, weight);
2429 account_entity_enqueue(cfs_rq, se);
2432 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2434 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2436 struct task_group *tg;
2437 struct sched_entity *se;
2441 se = tg->se[cpu_of(rq_of(cfs_rq))];
2442 if (!se || throttled_hierarchy(cfs_rq))
2445 if (likely(se->load.weight == tg->shares))
2448 shares = calc_cfs_shares(cfs_rq, tg);
2450 reweight_entity(cfs_rq_of(se), se, shares);
2452 #else /* CONFIG_FAIR_GROUP_SCHED */
2453 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2456 #endif /* CONFIG_FAIR_GROUP_SCHED */
2459 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2460 static const u32 runnable_avg_yN_inv[] = {
2461 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2462 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2463 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2464 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2465 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2466 0x85aac367, 0x82cd8698,
2470 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2471 * over-estimates when re-combining.
2473 static const u32 runnable_avg_yN_sum[] = {
2474 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2475 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2476 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2481 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2483 static __always_inline u64 decay_load(u64 val, u64 n)
2485 unsigned int local_n;
2489 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2492 /* after bounds checking we can collapse to 32-bit */
2496 * As y^PERIOD = 1/2, we can combine
2497 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2498 * With a look-up table which covers y^n (n<PERIOD)
2500 * To achieve constant time decay_load.
2502 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2503 val >>= local_n / LOAD_AVG_PERIOD;
2504 local_n %= LOAD_AVG_PERIOD;
2507 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2512 * For updates fully spanning n periods, the contribution to runnable
2513 * average will be: \Sum 1024*y^n
2515 * We can compute this reasonably efficiently by combining:
2516 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2518 static u32 __compute_runnable_contrib(u64 n)
2522 if (likely(n <= LOAD_AVG_PERIOD))
2523 return runnable_avg_yN_sum[n];
2524 else if (unlikely(n >= LOAD_AVG_MAX_N))
2525 return LOAD_AVG_MAX;
2527 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2529 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2530 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2532 n -= LOAD_AVG_PERIOD;
2533 } while (n > LOAD_AVG_PERIOD);
2535 contrib = decay_load(contrib, n);
2536 return contrib + runnable_avg_yN_sum[n];
2539 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2540 #error "load tracking assumes 2^10 as unit"
2543 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2546 * We can represent the historical contribution to runnable average as the
2547 * coefficients of a geometric series. To do this we sub-divide our runnable
2548 * history into segments of approximately 1ms (1024us); label the segment that
2549 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2551 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2553 * (now) (~1ms ago) (~2ms ago)
2555 * Let u_i denote the fraction of p_i that the entity was runnable.
2557 * We then designate the fractions u_i as our co-efficients, yielding the
2558 * following representation of historical load:
2559 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2561 * We choose y based on the with of a reasonably scheduling period, fixing:
2564 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2565 * approximately half as much as the contribution to load within the last ms
2568 * When a period "rolls over" and we have new u_0`, multiplying the previous
2569 * sum again by y is sufficient to update:
2570 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2571 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2573 static __always_inline int
2574 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2575 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2577 u64 delta, scaled_delta, periods;
2579 unsigned int delta_w, scaled_delta_w, decayed = 0;
2580 unsigned long scale_freq, scale_cpu;
2582 delta = now - sa->last_update_time;
2584 * This should only happen when time goes backwards, which it
2585 * unfortunately does during sched clock init when we swap over to TSC.
2587 if ((s64)delta < 0) {
2588 sa->last_update_time = now;
2593 * Use 1024ns as the unit of measurement since it's a reasonable
2594 * approximation of 1us and fast to compute.
2599 sa->last_update_time = now;
2601 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2602 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2603 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2605 /* delta_w is the amount already accumulated against our next period */
2606 delta_w = sa->period_contrib;
2607 if (delta + delta_w >= 1024) {
2610 /* how much left for next period will start over, we don't know yet */
2611 sa->period_contrib = 0;
2614 * Now that we know we're crossing a period boundary, figure
2615 * out how much from delta we need to complete the current
2616 * period and accrue it.
2618 delta_w = 1024 - delta_w;
2619 scaled_delta_w = cap_scale(delta_w, scale_freq);
2621 sa->load_sum += weight * scaled_delta_w;
2623 cfs_rq->runnable_load_sum +=
2624 weight * scaled_delta_w;
2628 sa->util_sum += scaled_delta_w * scale_cpu;
2632 /* Figure out how many additional periods this update spans */
2633 periods = delta / 1024;
2636 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2638 cfs_rq->runnable_load_sum =
2639 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2641 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2643 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2644 contrib = __compute_runnable_contrib(periods);
2645 contrib = cap_scale(contrib, scale_freq);
2647 sa->load_sum += weight * contrib;
2649 cfs_rq->runnable_load_sum += weight * contrib;
2652 sa->util_sum += contrib * scale_cpu;
2655 /* Remainder of delta accrued against u_0` */
2656 scaled_delta = cap_scale(delta, scale_freq);
2658 sa->load_sum += weight * scaled_delta;
2660 cfs_rq->runnable_load_sum += weight * scaled_delta;
2663 sa->util_sum += scaled_delta * scale_cpu;
2665 sa->period_contrib += delta;
2668 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2670 cfs_rq->runnable_load_avg =
2671 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2673 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2679 #ifdef CONFIG_FAIR_GROUP_SCHED
2681 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2682 * and effective_load (which is not done because it is too costly).
2684 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2686 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2688 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2689 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2690 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2694 #else /* CONFIG_FAIR_GROUP_SCHED */
2695 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2696 #endif /* CONFIG_FAIR_GROUP_SCHED */
2698 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2701 * Unsigned subtract and clamp on underflow.
2703 * Explicitly do a load-store to ensure the intermediate value never hits
2704 * memory. This allows lockless observations without ever seeing the negative
2707 #define sub_positive(_ptr, _val) do { \
2708 typeof(_ptr) ptr = (_ptr); \
2709 typeof(*ptr) val = (_val); \
2710 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2714 WRITE_ONCE(*ptr, res); \
2717 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2718 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2720 struct sched_avg *sa = &cfs_rq->avg;
2721 int decayed, removed = 0;
2723 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2724 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2725 sub_positive(&sa->load_avg, r);
2726 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2730 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2731 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2732 sub_positive(&sa->util_avg, r);
2733 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2736 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2737 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2739 #ifndef CONFIG_64BIT
2741 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2744 return decayed || removed;
2747 /* Update task and its cfs_rq load average */
2748 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2750 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2751 u64 now = cfs_rq_clock_task(cfs_rq);
2752 int cpu = cpu_of(rq_of(cfs_rq));
2755 * Track task load average for carrying it to new CPU after migrated, and
2756 * track group sched_entity load average for task_h_load calc in migration
2758 __update_load_avg(now, cpu, &se->avg,
2759 se->on_rq * scale_load_down(se->load.weight),
2760 cfs_rq->curr == se, NULL);
2762 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2763 update_tg_load_avg(cfs_rq, 0);
2765 if (entity_is_task(se))
2766 trace_sched_load_avg_task(task_of(se), &se->avg);
2767 trace_sched_load_avg_cpu(cpu, cfs_rq);
2770 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2772 if (!sched_feat(ATTACH_AGE_LOAD))
2776 * If we got migrated (either between CPUs or between cgroups) we'll
2777 * have aged the average right before clearing @last_update_time.
2779 if (se->avg.last_update_time) {
2780 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2781 &se->avg, 0, 0, NULL);
2784 * XXX: we could have just aged the entire load away if we've been
2785 * absent from the fair class for too long.
2790 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2791 cfs_rq->avg.load_avg += se->avg.load_avg;
2792 cfs_rq->avg.load_sum += se->avg.load_sum;
2793 cfs_rq->avg.util_avg += se->avg.util_avg;
2794 cfs_rq->avg.util_sum += se->avg.util_sum;
2797 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2799 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2800 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2801 cfs_rq->curr == se, NULL);
2803 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2804 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2805 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2806 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2809 /* Add the load generated by se into cfs_rq's load average */
2811 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2813 struct sched_avg *sa = &se->avg;
2814 u64 now = cfs_rq_clock_task(cfs_rq);
2815 int migrated, decayed;
2817 migrated = !sa->last_update_time;
2819 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2820 se->on_rq * scale_load_down(se->load.weight),
2821 cfs_rq->curr == se, NULL);
2824 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2826 cfs_rq->runnable_load_avg += sa->load_avg;
2827 cfs_rq->runnable_load_sum += sa->load_sum;
2830 attach_entity_load_avg(cfs_rq, se);
2832 if (decayed || migrated)
2833 update_tg_load_avg(cfs_rq, 0);
2836 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2838 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2840 update_load_avg(se, 1);
2842 cfs_rq->runnable_load_avg =
2843 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2844 cfs_rq->runnable_load_sum =
2845 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2848 #ifndef CONFIG_64BIT
2849 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2851 u64 last_update_time_copy;
2852 u64 last_update_time;
2855 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2857 last_update_time = cfs_rq->avg.last_update_time;
2858 } while (last_update_time != last_update_time_copy);
2860 return last_update_time;
2863 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2865 return cfs_rq->avg.last_update_time;
2870 * Task first catches up with cfs_rq, and then subtract
2871 * itself from the cfs_rq (task must be off the queue now).
2873 void remove_entity_load_avg(struct sched_entity *se)
2875 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2876 u64 last_update_time;
2879 * Newly created task or never used group entity should not be removed
2880 * from its (source) cfs_rq
2882 if (se->avg.last_update_time == 0)
2885 last_update_time = cfs_rq_last_update_time(cfs_rq);
2887 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2888 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2889 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2893 * Update the rq's load with the elapsed running time before entering
2894 * idle. if the last scheduled task is not a CFS task, idle_enter will
2895 * be the only way to update the runnable statistic.
2897 void idle_enter_fair(struct rq *this_rq)
2902 * Update the rq's load with the elapsed idle time before a task is
2903 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2904 * be the only way to update the runnable statistic.
2906 void idle_exit_fair(struct rq *this_rq)
2910 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2912 return cfs_rq->runnable_load_avg;
2915 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2917 return cfs_rq->avg.load_avg;
2920 static int idle_balance(struct rq *this_rq);
2922 #else /* CONFIG_SMP */
2924 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2926 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2928 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2929 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2932 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2934 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2936 static inline int idle_balance(struct rq *rq)
2941 #endif /* CONFIG_SMP */
2943 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2945 #ifdef CONFIG_SCHEDSTATS
2946 struct task_struct *tsk = NULL;
2948 if (entity_is_task(se))
2951 if (se->statistics.sleep_start) {
2952 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2957 if (unlikely(delta > se->statistics.sleep_max))
2958 se->statistics.sleep_max = delta;
2960 se->statistics.sleep_start = 0;
2961 se->statistics.sum_sleep_runtime += delta;
2964 account_scheduler_latency(tsk, delta >> 10, 1);
2965 trace_sched_stat_sleep(tsk, delta);
2968 if (se->statistics.block_start) {
2969 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2974 if (unlikely(delta > se->statistics.block_max))
2975 se->statistics.block_max = delta;
2977 se->statistics.block_start = 0;
2978 se->statistics.sum_sleep_runtime += delta;
2981 if (tsk->in_iowait) {
2982 se->statistics.iowait_sum += delta;
2983 se->statistics.iowait_count++;
2984 trace_sched_stat_iowait(tsk, delta);
2987 trace_sched_stat_blocked(tsk, delta);
2988 trace_sched_blocked_reason(tsk);
2991 * Blocking time is in units of nanosecs, so shift by
2992 * 20 to get a milliseconds-range estimation of the
2993 * amount of time that the task spent sleeping:
2995 if (unlikely(prof_on == SLEEP_PROFILING)) {
2996 profile_hits(SLEEP_PROFILING,
2997 (void *)get_wchan(tsk),
3000 account_scheduler_latency(tsk, delta >> 10, 0);
3006 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3008 #ifdef CONFIG_SCHED_DEBUG
3009 s64 d = se->vruntime - cfs_rq->min_vruntime;
3014 if (d > 3*sysctl_sched_latency)
3015 schedstat_inc(cfs_rq, nr_spread_over);
3020 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3022 u64 vruntime = cfs_rq->min_vruntime;
3025 * The 'current' period is already promised to the current tasks,
3026 * however the extra weight of the new task will slow them down a
3027 * little, place the new task so that it fits in the slot that
3028 * stays open at the end.
3030 if (initial && sched_feat(START_DEBIT))
3031 vruntime += sched_vslice(cfs_rq, se);
3033 /* sleeps up to a single latency don't count. */
3035 unsigned long thresh = sysctl_sched_latency;
3038 * Halve their sleep time's effect, to allow
3039 * for a gentler effect of sleepers:
3041 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3047 /* ensure we never gain time by being placed backwards. */
3048 se->vruntime = max_vruntime(se->vruntime, vruntime);
3051 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3054 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3057 * Update the normalized vruntime before updating min_vruntime
3058 * through calling update_curr().
3060 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3061 se->vruntime += cfs_rq->min_vruntime;
3064 * Update run-time statistics of the 'current'.
3066 update_curr(cfs_rq);
3067 enqueue_entity_load_avg(cfs_rq, se);
3068 account_entity_enqueue(cfs_rq, se);
3069 update_cfs_shares(cfs_rq);
3071 if (flags & ENQUEUE_WAKEUP) {
3072 place_entity(cfs_rq, se, 0);
3073 enqueue_sleeper(cfs_rq, se);
3076 update_stats_enqueue(cfs_rq, se);
3077 check_spread(cfs_rq, se);
3078 if (se != cfs_rq->curr)
3079 __enqueue_entity(cfs_rq, se);
3082 if (cfs_rq->nr_running == 1) {
3083 list_add_leaf_cfs_rq(cfs_rq);
3084 check_enqueue_throttle(cfs_rq);
3088 static void __clear_buddies_last(struct sched_entity *se)
3090 for_each_sched_entity(se) {
3091 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3092 if (cfs_rq->last != se)
3095 cfs_rq->last = NULL;
3099 static void __clear_buddies_next(struct sched_entity *se)
3101 for_each_sched_entity(se) {
3102 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3103 if (cfs_rq->next != se)
3106 cfs_rq->next = NULL;
3110 static void __clear_buddies_skip(struct sched_entity *se)
3112 for_each_sched_entity(se) {
3113 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3114 if (cfs_rq->skip != se)
3117 cfs_rq->skip = NULL;
3121 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3123 if (cfs_rq->last == se)
3124 __clear_buddies_last(se);
3126 if (cfs_rq->next == se)
3127 __clear_buddies_next(se);
3129 if (cfs_rq->skip == se)
3130 __clear_buddies_skip(se);
3133 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3136 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3139 * Update run-time statistics of the 'current'.
3141 update_curr(cfs_rq);
3142 dequeue_entity_load_avg(cfs_rq, se);
3144 update_stats_dequeue(cfs_rq, se);
3145 if (flags & DEQUEUE_SLEEP) {
3146 #ifdef CONFIG_SCHEDSTATS
3147 if (entity_is_task(se)) {
3148 struct task_struct *tsk = task_of(se);
3150 if (tsk->state & TASK_INTERRUPTIBLE)
3151 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3152 if (tsk->state & TASK_UNINTERRUPTIBLE)
3153 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3158 clear_buddies(cfs_rq, se);
3160 if (se != cfs_rq->curr)
3161 __dequeue_entity(cfs_rq, se);
3163 account_entity_dequeue(cfs_rq, se);
3166 * Normalize the entity after updating the min_vruntime because the
3167 * update can refer to the ->curr item and we need to reflect this
3168 * movement in our normalized position.
3170 if (!(flags & DEQUEUE_SLEEP))
3171 se->vruntime -= cfs_rq->min_vruntime;
3173 /* return excess runtime on last dequeue */
3174 return_cfs_rq_runtime(cfs_rq);
3176 update_min_vruntime(cfs_rq);
3177 update_cfs_shares(cfs_rq);
3181 * Preempt the current task with a newly woken task if needed:
3184 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3186 unsigned long ideal_runtime, delta_exec;
3187 struct sched_entity *se;
3190 ideal_runtime = sched_slice(cfs_rq, curr);
3191 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3192 if (delta_exec > ideal_runtime) {
3193 resched_curr(rq_of(cfs_rq));
3195 * The current task ran long enough, ensure it doesn't get
3196 * re-elected due to buddy favours.
3198 clear_buddies(cfs_rq, curr);
3203 * Ensure that a task that missed wakeup preemption by a
3204 * narrow margin doesn't have to wait for a full slice.
3205 * This also mitigates buddy induced latencies under load.
3207 if (delta_exec < sysctl_sched_min_granularity)
3210 se = __pick_first_entity(cfs_rq);
3211 delta = curr->vruntime - se->vruntime;
3216 if (delta > ideal_runtime)
3217 resched_curr(rq_of(cfs_rq));
3221 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3223 /* 'current' is not kept within the tree. */
3226 * Any task has to be enqueued before it get to execute on
3227 * a CPU. So account for the time it spent waiting on the
3230 update_stats_wait_end(cfs_rq, se);
3231 __dequeue_entity(cfs_rq, se);
3232 update_load_avg(se, 1);
3235 update_stats_curr_start(cfs_rq, se);
3237 #ifdef CONFIG_SCHEDSTATS
3239 * Track our maximum slice length, if the CPU's load is at
3240 * least twice that of our own weight (i.e. dont track it
3241 * when there are only lesser-weight tasks around):
3243 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3244 se->statistics.slice_max = max(se->statistics.slice_max,
3245 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3248 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3252 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3255 * Pick the next process, keeping these things in mind, in this order:
3256 * 1) keep things fair between processes/task groups
3257 * 2) pick the "next" process, since someone really wants that to run
3258 * 3) pick the "last" process, for cache locality
3259 * 4) do not run the "skip" process, if something else is available
3261 static struct sched_entity *
3262 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3264 struct sched_entity *left = __pick_first_entity(cfs_rq);
3265 struct sched_entity *se;
3268 * If curr is set we have to see if its left of the leftmost entity
3269 * still in the tree, provided there was anything in the tree at all.
3271 if (!left || (curr && entity_before(curr, left)))
3274 se = left; /* ideally we run the leftmost entity */
3277 * Avoid running the skip buddy, if running something else can
3278 * be done without getting too unfair.
3280 if (cfs_rq->skip == se) {
3281 struct sched_entity *second;
3284 second = __pick_first_entity(cfs_rq);
3286 second = __pick_next_entity(se);
3287 if (!second || (curr && entity_before(curr, second)))
3291 if (second && wakeup_preempt_entity(second, left) < 1)
3296 * Prefer last buddy, try to return the CPU to a preempted task.
3298 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3302 * Someone really wants this to run. If it's not unfair, run it.
3304 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3307 clear_buddies(cfs_rq, se);
3312 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3314 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3317 * If still on the runqueue then deactivate_task()
3318 * was not called and update_curr() has to be done:
3321 update_curr(cfs_rq);
3323 /* throttle cfs_rqs exceeding runtime */
3324 check_cfs_rq_runtime(cfs_rq);
3326 check_spread(cfs_rq, prev);
3328 update_stats_wait_start(cfs_rq, prev);
3329 /* Put 'current' back into the tree. */
3330 __enqueue_entity(cfs_rq, prev);
3331 /* in !on_rq case, update occurred at dequeue */
3332 update_load_avg(prev, 0);
3334 cfs_rq->curr = NULL;
3338 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3341 * Update run-time statistics of the 'current'.
3343 update_curr(cfs_rq);
3346 * Ensure that runnable average is periodically updated.
3348 update_load_avg(curr, 1);
3349 update_cfs_shares(cfs_rq);
3351 #ifdef CONFIG_SCHED_HRTICK
3353 * queued ticks are scheduled to match the slice, so don't bother
3354 * validating it and just reschedule.
3357 resched_curr(rq_of(cfs_rq));
3361 * don't let the period tick interfere with the hrtick preemption
3363 if (!sched_feat(DOUBLE_TICK) &&
3364 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3368 if (cfs_rq->nr_running > 1)
3369 check_preempt_tick(cfs_rq, curr);
3373 /**************************************************
3374 * CFS bandwidth control machinery
3377 #ifdef CONFIG_CFS_BANDWIDTH
3379 #ifdef HAVE_JUMP_LABEL
3380 static struct static_key __cfs_bandwidth_used;
3382 static inline bool cfs_bandwidth_used(void)
3384 return static_key_false(&__cfs_bandwidth_used);
3387 void cfs_bandwidth_usage_inc(void)
3389 static_key_slow_inc(&__cfs_bandwidth_used);
3392 void cfs_bandwidth_usage_dec(void)
3394 static_key_slow_dec(&__cfs_bandwidth_used);
3396 #else /* HAVE_JUMP_LABEL */
3397 static bool cfs_bandwidth_used(void)
3402 void cfs_bandwidth_usage_inc(void) {}
3403 void cfs_bandwidth_usage_dec(void) {}
3404 #endif /* HAVE_JUMP_LABEL */
3407 * default period for cfs group bandwidth.
3408 * default: 0.1s, units: nanoseconds
3410 static inline u64 default_cfs_period(void)
3412 return 100000000ULL;
3415 static inline u64 sched_cfs_bandwidth_slice(void)
3417 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3421 * Replenish runtime according to assigned quota and update expiration time.
3422 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3423 * additional synchronization around rq->lock.
3425 * requires cfs_b->lock
3427 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3431 if (cfs_b->quota == RUNTIME_INF)
3434 now = sched_clock_cpu(smp_processor_id());
3435 cfs_b->runtime = cfs_b->quota;
3436 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3439 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3441 return &tg->cfs_bandwidth;
3444 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3445 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3447 if (unlikely(cfs_rq->throttle_count))
3448 return cfs_rq->throttled_clock_task;
3450 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3453 /* returns 0 on failure to allocate runtime */
3454 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3456 struct task_group *tg = cfs_rq->tg;
3457 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3458 u64 amount = 0, min_amount, expires;
3460 /* note: this is a positive sum as runtime_remaining <= 0 */
3461 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3463 raw_spin_lock(&cfs_b->lock);
3464 if (cfs_b->quota == RUNTIME_INF)
3465 amount = min_amount;
3467 start_cfs_bandwidth(cfs_b);
3469 if (cfs_b->runtime > 0) {
3470 amount = min(cfs_b->runtime, min_amount);
3471 cfs_b->runtime -= amount;
3475 expires = cfs_b->runtime_expires;
3476 raw_spin_unlock(&cfs_b->lock);
3478 cfs_rq->runtime_remaining += amount;
3480 * we may have advanced our local expiration to account for allowed
3481 * spread between our sched_clock and the one on which runtime was
3484 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3485 cfs_rq->runtime_expires = expires;
3487 return cfs_rq->runtime_remaining > 0;
3491 * Note: This depends on the synchronization provided by sched_clock and the
3492 * fact that rq->clock snapshots this value.
3494 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3496 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3498 /* if the deadline is ahead of our clock, nothing to do */
3499 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3502 if (cfs_rq->runtime_remaining < 0)
3506 * If the local deadline has passed we have to consider the
3507 * possibility that our sched_clock is 'fast' and the global deadline
3508 * has not truly expired.
3510 * Fortunately we can check determine whether this the case by checking
3511 * whether the global deadline has advanced. It is valid to compare
3512 * cfs_b->runtime_expires without any locks since we only care about
3513 * exact equality, so a partial write will still work.
3516 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3517 /* extend local deadline, drift is bounded above by 2 ticks */
3518 cfs_rq->runtime_expires += TICK_NSEC;
3520 /* global deadline is ahead, expiration has passed */
3521 cfs_rq->runtime_remaining = 0;
3525 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3527 /* dock delta_exec before expiring quota (as it could span periods) */
3528 cfs_rq->runtime_remaining -= delta_exec;
3529 expire_cfs_rq_runtime(cfs_rq);
3531 if (likely(cfs_rq->runtime_remaining > 0))
3535 * if we're unable to extend our runtime we resched so that the active
3536 * hierarchy can be throttled
3538 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3539 resched_curr(rq_of(cfs_rq));
3542 static __always_inline
3543 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3545 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3548 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3551 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3553 return cfs_bandwidth_used() && cfs_rq->throttled;
3556 /* check whether cfs_rq, or any parent, is throttled */
3557 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3559 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3563 * Ensure that neither of the group entities corresponding to src_cpu or
3564 * dest_cpu are members of a throttled hierarchy when performing group
3565 * load-balance operations.
3567 static inline int throttled_lb_pair(struct task_group *tg,
3568 int src_cpu, int dest_cpu)
3570 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3572 src_cfs_rq = tg->cfs_rq[src_cpu];
3573 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3575 return throttled_hierarchy(src_cfs_rq) ||
3576 throttled_hierarchy(dest_cfs_rq);
3579 /* updated child weight may affect parent so we have to do this bottom up */
3580 static int tg_unthrottle_up(struct task_group *tg, void *data)
3582 struct rq *rq = data;
3583 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3585 cfs_rq->throttle_count--;
3587 if (!cfs_rq->throttle_count) {
3588 /* adjust cfs_rq_clock_task() */
3589 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3590 cfs_rq->throttled_clock_task;
3597 static int tg_throttle_down(struct task_group *tg, void *data)
3599 struct rq *rq = data;
3600 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3602 /* group is entering throttled state, stop time */
3603 if (!cfs_rq->throttle_count)
3604 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3605 cfs_rq->throttle_count++;
3610 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3612 struct rq *rq = rq_of(cfs_rq);
3613 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3614 struct sched_entity *se;
3615 long task_delta, dequeue = 1;
3618 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3620 /* freeze hierarchy runnable averages while throttled */
3622 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3625 task_delta = cfs_rq->h_nr_running;
3626 for_each_sched_entity(se) {
3627 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3628 /* throttled entity or throttle-on-deactivate */
3633 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3634 qcfs_rq->h_nr_running -= task_delta;
3636 if (qcfs_rq->load.weight)
3641 sub_nr_running(rq, task_delta);
3643 cfs_rq->throttled = 1;
3644 cfs_rq->throttled_clock = rq_clock(rq);
3645 raw_spin_lock(&cfs_b->lock);
3646 empty = list_empty(&cfs_b->throttled_cfs_rq);
3649 * Add to the _head_ of the list, so that an already-started
3650 * distribute_cfs_runtime will not see us
3652 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3655 * If we're the first throttled task, make sure the bandwidth
3659 start_cfs_bandwidth(cfs_b);
3661 raw_spin_unlock(&cfs_b->lock);
3664 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3666 struct rq *rq = rq_of(cfs_rq);
3667 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3668 struct sched_entity *se;
3672 se = cfs_rq->tg->se[cpu_of(rq)];
3674 cfs_rq->throttled = 0;
3676 update_rq_clock(rq);
3678 raw_spin_lock(&cfs_b->lock);
3679 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3680 list_del_rcu(&cfs_rq->throttled_list);
3681 raw_spin_unlock(&cfs_b->lock);
3683 /* update hierarchical throttle state */
3684 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3686 if (!cfs_rq->load.weight)
3689 task_delta = cfs_rq->h_nr_running;
3690 for_each_sched_entity(se) {
3694 cfs_rq = cfs_rq_of(se);
3696 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3697 cfs_rq->h_nr_running += task_delta;
3699 if (cfs_rq_throttled(cfs_rq))
3704 add_nr_running(rq, task_delta);
3706 /* determine whether we need to wake up potentially idle cpu */
3707 if (rq->curr == rq->idle && rq->cfs.nr_running)
3711 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3712 u64 remaining, u64 expires)
3714 struct cfs_rq *cfs_rq;
3716 u64 starting_runtime = remaining;
3719 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3721 struct rq *rq = rq_of(cfs_rq);
3723 raw_spin_lock(&rq->lock);
3724 if (!cfs_rq_throttled(cfs_rq))
3727 runtime = -cfs_rq->runtime_remaining + 1;
3728 if (runtime > remaining)
3729 runtime = remaining;
3730 remaining -= runtime;
3732 cfs_rq->runtime_remaining += runtime;
3733 cfs_rq->runtime_expires = expires;
3735 /* we check whether we're throttled above */
3736 if (cfs_rq->runtime_remaining > 0)
3737 unthrottle_cfs_rq(cfs_rq);
3740 raw_spin_unlock(&rq->lock);
3747 return starting_runtime - remaining;
3751 * Responsible for refilling a task_group's bandwidth and unthrottling its
3752 * cfs_rqs as appropriate. If there has been no activity within the last
3753 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3754 * used to track this state.
3756 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3758 u64 runtime, runtime_expires;
3761 /* no need to continue the timer with no bandwidth constraint */
3762 if (cfs_b->quota == RUNTIME_INF)
3763 goto out_deactivate;
3765 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3766 cfs_b->nr_periods += overrun;
3769 * idle depends on !throttled (for the case of a large deficit), and if
3770 * we're going inactive then everything else can be deferred
3772 if (cfs_b->idle && !throttled)
3773 goto out_deactivate;
3775 __refill_cfs_bandwidth_runtime(cfs_b);
3778 /* mark as potentially idle for the upcoming period */
3783 /* account preceding periods in which throttling occurred */
3784 cfs_b->nr_throttled += overrun;
3786 runtime_expires = cfs_b->runtime_expires;
3789 * This check is repeated as we are holding onto the new bandwidth while
3790 * we unthrottle. This can potentially race with an unthrottled group
3791 * trying to acquire new bandwidth from the global pool. This can result
3792 * in us over-using our runtime if it is all used during this loop, but
3793 * only by limited amounts in that extreme case.
3795 while (throttled && cfs_b->runtime > 0) {
3796 runtime = cfs_b->runtime;
3797 raw_spin_unlock(&cfs_b->lock);
3798 /* we can't nest cfs_b->lock while distributing bandwidth */
3799 runtime = distribute_cfs_runtime(cfs_b, runtime,
3801 raw_spin_lock(&cfs_b->lock);
3803 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3805 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3809 * While we are ensured activity in the period following an
3810 * unthrottle, this also covers the case in which the new bandwidth is
3811 * insufficient to cover the existing bandwidth deficit. (Forcing the
3812 * timer to remain active while there are any throttled entities.)
3822 /* a cfs_rq won't donate quota below this amount */
3823 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3824 /* minimum remaining period time to redistribute slack quota */
3825 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3826 /* how long we wait to gather additional slack before distributing */
3827 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3830 * Are we near the end of the current quota period?
3832 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3833 * hrtimer base being cleared by hrtimer_start. In the case of
3834 * migrate_hrtimers, base is never cleared, so we are fine.
3836 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3838 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3841 /* if the call-back is running a quota refresh is already occurring */
3842 if (hrtimer_callback_running(refresh_timer))
3845 /* is a quota refresh about to occur? */
3846 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3847 if (remaining < min_expire)
3853 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3855 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3857 /* if there's a quota refresh soon don't bother with slack */
3858 if (runtime_refresh_within(cfs_b, min_left))
3861 hrtimer_start(&cfs_b->slack_timer,
3862 ns_to_ktime(cfs_bandwidth_slack_period),
3866 /* we know any runtime found here is valid as update_curr() precedes return */
3867 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3869 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3870 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3872 if (slack_runtime <= 0)
3875 raw_spin_lock(&cfs_b->lock);
3876 if (cfs_b->quota != RUNTIME_INF &&
3877 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3878 cfs_b->runtime += slack_runtime;
3880 /* we are under rq->lock, defer unthrottling using a timer */
3881 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3882 !list_empty(&cfs_b->throttled_cfs_rq))
3883 start_cfs_slack_bandwidth(cfs_b);
3885 raw_spin_unlock(&cfs_b->lock);
3887 /* even if it's not valid for return we don't want to try again */
3888 cfs_rq->runtime_remaining -= slack_runtime;
3891 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3893 if (!cfs_bandwidth_used())
3896 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3899 __return_cfs_rq_runtime(cfs_rq);
3903 * This is done with a timer (instead of inline with bandwidth return) since
3904 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3906 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3908 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3911 /* confirm we're still not at a refresh boundary */
3912 raw_spin_lock(&cfs_b->lock);
3913 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3914 raw_spin_unlock(&cfs_b->lock);
3918 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3919 runtime = cfs_b->runtime;
3921 expires = cfs_b->runtime_expires;
3922 raw_spin_unlock(&cfs_b->lock);
3927 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3929 raw_spin_lock(&cfs_b->lock);
3930 if (expires == cfs_b->runtime_expires)
3931 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3932 raw_spin_unlock(&cfs_b->lock);
3936 * When a group wakes up we want to make sure that its quota is not already
3937 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3938 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3940 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3942 if (!cfs_bandwidth_used())
3945 /* an active group must be handled by the update_curr()->put() path */
3946 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3949 /* ensure the group is not already throttled */
3950 if (cfs_rq_throttled(cfs_rq))
3953 /* update runtime allocation */
3954 account_cfs_rq_runtime(cfs_rq, 0);
3955 if (cfs_rq->runtime_remaining <= 0)
3956 throttle_cfs_rq(cfs_rq);
3959 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3960 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3962 if (!cfs_bandwidth_used())
3965 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3969 * it's possible for a throttled entity to be forced into a running
3970 * state (e.g. set_curr_task), in this case we're finished.
3972 if (cfs_rq_throttled(cfs_rq))
3975 throttle_cfs_rq(cfs_rq);
3979 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3981 struct cfs_bandwidth *cfs_b =
3982 container_of(timer, struct cfs_bandwidth, slack_timer);
3984 do_sched_cfs_slack_timer(cfs_b);
3986 return HRTIMER_NORESTART;
3989 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3991 struct cfs_bandwidth *cfs_b =
3992 container_of(timer, struct cfs_bandwidth, period_timer);
3996 raw_spin_lock(&cfs_b->lock);
3998 overrun = hrtimer_forward_now(timer, cfs_b->period);
4002 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4005 cfs_b->period_active = 0;
4006 raw_spin_unlock(&cfs_b->lock);
4008 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4011 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4013 raw_spin_lock_init(&cfs_b->lock);
4015 cfs_b->quota = RUNTIME_INF;
4016 cfs_b->period = ns_to_ktime(default_cfs_period());
4018 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4019 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4020 cfs_b->period_timer.function = sched_cfs_period_timer;
4021 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4022 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4025 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4027 cfs_rq->runtime_enabled = 0;
4028 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4031 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4033 lockdep_assert_held(&cfs_b->lock);
4035 if (!cfs_b->period_active) {
4036 cfs_b->period_active = 1;
4037 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4038 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4042 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4044 /* init_cfs_bandwidth() was not called */
4045 if (!cfs_b->throttled_cfs_rq.next)
4048 hrtimer_cancel(&cfs_b->period_timer);
4049 hrtimer_cancel(&cfs_b->slack_timer);
4052 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4054 struct cfs_rq *cfs_rq;
4056 for_each_leaf_cfs_rq(rq, cfs_rq) {
4057 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4059 raw_spin_lock(&cfs_b->lock);
4060 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4061 raw_spin_unlock(&cfs_b->lock);
4065 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4067 struct cfs_rq *cfs_rq;
4069 for_each_leaf_cfs_rq(rq, cfs_rq) {
4070 if (!cfs_rq->runtime_enabled)
4074 * clock_task is not advancing so we just need to make sure
4075 * there's some valid quota amount
4077 cfs_rq->runtime_remaining = 1;
4079 * Offline rq is schedulable till cpu is completely disabled
4080 * in take_cpu_down(), so we prevent new cfs throttling here.
4082 cfs_rq->runtime_enabled = 0;
4084 if (cfs_rq_throttled(cfs_rq))
4085 unthrottle_cfs_rq(cfs_rq);
4089 #else /* CONFIG_CFS_BANDWIDTH */
4090 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4092 return rq_clock_task(rq_of(cfs_rq));
4095 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4096 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4097 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4098 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4100 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4105 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4110 static inline int throttled_lb_pair(struct task_group *tg,
4111 int src_cpu, int dest_cpu)
4116 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4118 #ifdef CONFIG_FAIR_GROUP_SCHED
4119 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4122 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4126 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4127 static inline void update_runtime_enabled(struct rq *rq) {}
4128 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4130 #endif /* CONFIG_CFS_BANDWIDTH */
4132 /**************************************************
4133 * CFS operations on tasks:
4136 #ifdef CONFIG_SCHED_HRTICK
4137 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4139 struct sched_entity *se = &p->se;
4140 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4142 WARN_ON(task_rq(p) != rq);
4144 if (cfs_rq->nr_running > 1) {
4145 u64 slice = sched_slice(cfs_rq, se);
4146 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4147 s64 delta = slice - ran;
4154 hrtick_start(rq, delta);
4159 * called from enqueue/dequeue and updates the hrtick when the
4160 * current task is from our class and nr_running is low enough
4163 static void hrtick_update(struct rq *rq)
4165 struct task_struct *curr = rq->curr;
4167 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4170 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4171 hrtick_start_fair(rq, curr);
4173 #else /* !CONFIG_SCHED_HRTICK */
4175 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4179 static inline void hrtick_update(struct rq *rq)
4185 static bool cpu_overutilized(int cpu);
4186 static inline unsigned long boosted_cpu_util(int cpu);
4188 #define boosted_cpu_util(cpu) cpu_util(cpu)
4192 static void update_capacity_of(int cpu)
4194 unsigned long req_cap;
4199 /* Convert scale-invariant capacity to cpu. */
4200 req_cap = boosted_cpu_util(cpu);
4201 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4202 set_cfs_cpu_capacity(cpu, true, req_cap);
4207 * The enqueue_task method is called before nr_running is
4208 * increased. Here we update the fair scheduling stats and
4209 * then put the task into the rbtree:
4212 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4214 struct cfs_rq *cfs_rq;
4215 struct sched_entity *se = &p->se;
4217 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4218 int task_wakeup = flags & ENQUEUE_WAKEUP;
4221 for_each_sched_entity(se) {
4224 cfs_rq = cfs_rq_of(se);
4225 enqueue_entity(cfs_rq, se, flags);
4228 * end evaluation on encountering a throttled cfs_rq
4230 * note: in the case of encountering a throttled cfs_rq we will
4231 * post the final h_nr_running increment below.
4233 if (cfs_rq_throttled(cfs_rq))
4235 cfs_rq->h_nr_running++;
4236 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4238 flags = ENQUEUE_WAKEUP;
4241 for_each_sched_entity(se) {
4242 cfs_rq = cfs_rq_of(se);
4243 cfs_rq->h_nr_running++;
4244 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4246 if (cfs_rq_throttled(cfs_rq))
4249 update_load_avg(se, 1);
4250 update_cfs_shares(cfs_rq);
4254 add_nr_running(rq, 1);
4259 walt_inc_cumulative_runnable_avg(rq, p);
4260 if (!task_new && !rq->rd->overutilized &&
4261 cpu_overutilized(rq->cpu))
4262 rq->rd->overutilized = true;
4265 * We want to potentially trigger a freq switch
4266 * request only for tasks that are waking up; this is
4267 * because we get here also during load balancing, but
4268 * in these cases it seems wise to trigger as single
4269 * request after load balancing is done.
4271 if (task_new || task_wakeup)
4272 update_capacity_of(cpu_of(rq));
4275 /* Update SchedTune accouting */
4276 schedtune_enqueue_task(p, cpu_of(rq));
4278 #endif /* CONFIG_SMP */
4282 static void set_next_buddy(struct sched_entity *se);
4285 * The dequeue_task method is called before nr_running is
4286 * decreased. We remove the task from the rbtree and
4287 * update the fair scheduling stats:
4289 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4291 struct cfs_rq *cfs_rq;
4292 struct sched_entity *se = &p->se;
4293 int task_sleep = flags & DEQUEUE_SLEEP;
4295 for_each_sched_entity(se) {
4296 cfs_rq = cfs_rq_of(se);
4297 dequeue_entity(cfs_rq, se, flags);
4300 * end evaluation on encountering a throttled cfs_rq
4302 * note: in the case of encountering a throttled cfs_rq we will
4303 * post the final h_nr_running decrement below.
4305 if (cfs_rq_throttled(cfs_rq))
4307 cfs_rq->h_nr_running--;
4308 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4310 /* Don't dequeue parent if it has other entities besides us */
4311 if (cfs_rq->load.weight) {
4313 * Bias pick_next to pick a task from this cfs_rq, as
4314 * p is sleeping when it is within its sched_slice.
4316 if (task_sleep && parent_entity(se))
4317 set_next_buddy(parent_entity(se));
4319 /* avoid re-evaluating load for this entity */
4320 se = parent_entity(se);
4323 flags |= DEQUEUE_SLEEP;
4326 for_each_sched_entity(se) {
4327 cfs_rq = cfs_rq_of(se);
4328 cfs_rq->h_nr_running--;
4329 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4331 if (cfs_rq_throttled(cfs_rq))
4334 update_load_avg(se, 1);
4335 update_cfs_shares(cfs_rq);
4339 sub_nr_running(rq, 1);
4344 walt_dec_cumulative_runnable_avg(rq, p);
4347 * We want to potentially trigger a freq switch
4348 * request only for tasks that are going to sleep;
4349 * this is because we get here also during load
4350 * balancing, but in these cases it seems wise to
4351 * trigger as single request after load balancing is
4355 if (rq->cfs.nr_running)
4356 update_capacity_of(cpu_of(rq));
4357 else if (sched_freq())
4358 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4362 /* Update SchedTune accouting */
4363 schedtune_dequeue_task(p, cpu_of(rq));
4365 #endif /* CONFIG_SMP */
4373 * per rq 'load' arrray crap; XXX kill this.
4377 * The exact cpuload at various idx values, calculated at every tick would be
4378 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4380 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4381 * on nth tick when cpu may be busy, then we have:
4382 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4383 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4385 * decay_load_missed() below does efficient calculation of
4386 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4387 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4389 * The calculation is approximated on a 128 point scale.
4390 * degrade_zero_ticks is the number of ticks after which load at any
4391 * particular idx is approximated to be zero.
4392 * degrade_factor is a precomputed table, a row for each load idx.
4393 * Each column corresponds to degradation factor for a power of two ticks,
4394 * based on 128 point scale.
4396 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4397 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4399 * With this power of 2 load factors, we can degrade the load n times
4400 * by looking at 1 bits in n and doing as many mult/shift instead of
4401 * n mult/shifts needed by the exact degradation.
4403 #define DEGRADE_SHIFT 7
4404 static const unsigned char
4405 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4406 static const unsigned char
4407 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4408 {0, 0, 0, 0, 0, 0, 0, 0},
4409 {64, 32, 8, 0, 0, 0, 0, 0},
4410 {96, 72, 40, 12, 1, 0, 0},
4411 {112, 98, 75, 43, 15, 1, 0},
4412 {120, 112, 98, 76, 45, 16, 2} };
4415 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4416 * would be when CPU is idle and so we just decay the old load without
4417 * adding any new load.
4419 static unsigned long
4420 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4424 if (!missed_updates)
4427 if (missed_updates >= degrade_zero_ticks[idx])
4431 return load >> missed_updates;
4433 while (missed_updates) {
4434 if (missed_updates % 2)
4435 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4437 missed_updates >>= 1;
4444 * Update rq->cpu_load[] statistics. This function is usually called every
4445 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4446 * every tick. We fix it up based on jiffies.
4448 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4449 unsigned long pending_updates)
4453 this_rq->nr_load_updates++;
4455 /* Update our load: */
4456 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4457 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4458 unsigned long old_load, new_load;
4460 /* scale is effectively 1 << i now, and >> i divides by scale */
4462 old_load = this_rq->cpu_load[i];
4463 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4464 new_load = this_load;
4466 * Round up the averaging division if load is increasing. This
4467 * prevents us from getting stuck on 9 if the load is 10, for
4470 if (new_load > old_load)
4471 new_load += scale - 1;
4473 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4476 sched_avg_update(this_rq);
4479 /* Used instead of source_load when we know the type == 0 */
4480 static unsigned long weighted_cpuload(const int cpu)
4482 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4485 #ifdef CONFIG_NO_HZ_COMMON
4487 * There is no sane way to deal with nohz on smp when using jiffies because the
4488 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4489 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4491 * Therefore we cannot use the delta approach from the regular tick since that
4492 * would seriously skew the load calculation. However we'll make do for those
4493 * updates happening while idle (nohz_idle_balance) or coming out of idle
4494 * (tick_nohz_idle_exit).
4496 * This means we might still be one tick off for nohz periods.
4500 * Called from nohz_idle_balance() to update the load ratings before doing the
4503 static void update_idle_cpu_load(struct rq *this_rq)
4505 unsigned long curr_jiffies = READ_ONCE(jiffies);
4506 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4507 unsigned long pending_updates;
4510 * bail if there's load or we're actually up-to-date.
4512 if (load || curr_jiffies == this_rq->last_load_update_tick)
4515 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4516 this_rq->last_load_update_tick = curr_jiffies;
4518 __update_cpu_load(this_rq, load, pending_updates);
4522 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4524 void update_cpu_load_nohz(void)
4526 struct rq *this_rq = this_rq();
4527 unsigned long curr_jiffies = READ_ONCE(jiffies);
4528 unsigned long pending_updates;
4530 if (curr_jiffies == this_rq->last_load_update_tick)
4533 raw_spin_lock(&this_rq->lock);
4534 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4535 if (pending_updates) {
4536 this_rq->last_load_update_tick = curr_jiffies;
4538 * We were idle, this means load 0, the current load might be
4539 * !0 due to remote wakeups and the sort.
4541 __update_cpu_load(this_rq, 0, pending_updates);
4543 raw_spin_unlock(&this_rq->lock);
4545 #endif /* CONFIG_NO_HZ */
4548 * Called from scheduler_tick()
4550 void update_cpu_load_active(struct rq *this_rq)
4552 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4554 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4556 this_rq->last_load_update_tick = jiffies;
4557 __update_cpu_load(this_rq, load, 1);
4561 * Return a low guess at the load of a migration-source cpu weighted
4562 * according to the scheduling class and "nice" value.
4564 * We want to under-estimate the load of migration sources, to
4565 * balance conservatively.
4567 static unsigned long source_load(int cpu, int type)
4569 struct rq *rq = cpu_rq(cpu);
4570 unsigned long total = weighted_cpuload(cpu);
4572 if (type == 0 || !sched_feat(LB_BIAS))
4575 return min(rq->cpu_load[type-1], total);
4579 * Return a high guess at the load of a migration-target cpu weighted
4580 * according to the scheduling class and "nice" value.
4582 static unsigned long target_load(int cpu, int type)
4584 struct rq *rq = cpu_rq(cpu);
4585 unsigned long total = weighted_cpuload(cpu);
4587 if (type == 0 || !sched_feat(LB_BIAS))
4590 return max(rq->cpu_load[type-1], total);
4594 static unsigned long cpu_avg_load_per_task(int cpu)
4596 struct rq *rq = cpu_rq(cpu);
4597 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4598 unsigned long load_avg = weighted_cpuload(cpu);
4601 return load_avg / nr_running;
4606 static void record_wakee(struct task_struct *p)
4609 * Rough decay (wiping) for cost saving, don't worry
4610 * about the boundary, really active task won't care
4613 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4614 current->wakee_flips >>= 1;
4615 current->wakee_flip_decay_ts = jiffies;
4618 if (current->last_wakee != p) {
4619 current->last_wakee = p;
4620 current->wakee_flips++;
4624 static void task_waking_fair(struct task_struct *p)
4626 struct sched_entity *se = &p->se;
4627 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4630 #ifndef CONFIG_64BIT
4631 u64 min_vruntime_copy;
4634 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4636 min_vruntime = cfs_rq->min_vruntime;
4637 } while (min_vruntime != min_vruntime_copy);
4639 min_vruntime = cfs_rq->min_vruntime;
4642 se->vruntime -= min_vruntime;
4646 #ifdef CONFIG_FAIR_GROUP_SCHED
4648 * effective_load() calculates the load change as seen from the root_task_group
4650 * Adding load to a group doesn't make a group heavier, but can cause movement
4651 * of group shares between cpus. Assuming the shares were perfectly aligned one
4652 * can calculate the shift in shares.
4654 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4655 * on this @cpu and results in a total addition (subtraction) of @wg to the
4656 * total group weight.
4658 * Given a runqueue weight distribution (rw_i) we can compute a shares
4659 * distribution (s_i) using:
4661 * s_i = rw_i / \Sum rw_j (1)
4663 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4664 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4665 * shares distribution (s_i):
4667 * rw_i = { 2, 4, 1, 0 }
4668 * s_i = { 2/7, 4/7, 1/7, 0 }
4670 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4671 * task used to run on and the CPU the waker is running on), we need to
4672 * compute the effect of waking a task on either CPU and, in case of a sync
4673 * wakeup, compute the effect of the current task going to sleep.
4675 * So for a change of @wl to the local @cpu with an overall group weight change
4676 * of @wl we can compute the new shares distribution (s'_i) using:
4678 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4680 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4681 * differences in waking a task to CPU 0. The additional task changes the
4682 * weight and shares distributions like:
4684 * rw'_i = { 3, 4, 1, 0 }
4685 * s'_i = { 3/8, 4/8, 1/8, 0 }
4687 * We can then compute the difference in effective weight by using:
4689 * dw_i = S * (s'_i - s_i) (3)
4691 * Where 'S' is the group weight as seen by its parent.
4693 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4694 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4695 * 4/7) times the weight of the group.
4697 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4699 struct sched_entity *se = tg->se[cpu];
4701 if (!tg->parent) /* the trivial, non-cgroup case */
4704 for_each_sched_entity(se) {
4705 struct cfs_rq *cfs_rq = se->my_q;
4706 long W, w = cfs_rq_load_avg(cfs_rq);
4711 * W = @wg + \Sum rw_j
4713 W = wg + atomic_long_read(&tg->load_avg);
4715 /* Ensure \Sum rw_j >= rw_i */
4716 W -= cfs_rq->tg_load_avg_contrib;
4725 * wl = S * s'_i; see (2)
4728 wl = (w * (long)tg->shares) / W;
4733 * Per the above, wl is the new se->load.weight value; since
4734 * those are clipped to [MIN_SHARES, ...) do so now. See
4735 * calc_cfs_shares().
4737 if (wl < MIN_SHARES)
4741 * wl = dw_i = S * (s'_i - s_i); see (3)
4743 wl -= se->avg.load_avg;
4746 * Recursively apply this logic to all parent groups to compute
4747 * the final effective load change on the root group. Since
4748 * only the @tg group gets extra weight, all parent groups can
4749 * only redistribute existing shares. @wl is the shift in shares
4750 * resulting from this level per the above.
4759 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4767 * Returns the current capacity of cpu after applying both
4768 * cpu and freq scaling.
4770 unsigned long capacity_curr_of(int cpu)
4772 return cpu_rq(cpu)->cpu_capacity_orig *
4773 arch_scale_freq_capacity(NULL, cpu)
4774 >> SCHED_CAPACITY_SHIFT;
4777 static inline bool energy_aware(void)
4779 return sched_feat(ENERGY_AWARE);
4783 struct sched_group *sg_top;
4784 struct sched_group *sg_cap;
4791 struct task_struct *task;
4806 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4807 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4808 * energy calculations. Using the scale-invariant util returned by
4809 * cpu_util() and approximating scale-invariant util by:
4811 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4813 * the normalized util can be found using the specific capacity.
4815 * capacity = capacity_orig * curr_freq/max_freq
4817 * norm_util = running_time/time ~ util/capacity
4819 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4821 int util = __cpu_util(cpu, delta);
4823 if (util >= capacity)
4824 return SCHED_CAPACITY_SCALE;
4826 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4829 static int calc_util_delta(struct energy_env *eenv, int cpu)
4831 if (cpu == eenv->src_cpu)
4832 return -eenv->util_delta;
4833 if (cpu == eenv->dst_cpu)
4834 return eenv->util_delta;
4839 unsigned long group_max_util(struct energy_env *eenv)
4842 unsigned long max_util = 0;
4844 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4845 delta = calc_util_delta(eenv, i);
4846 max_util = max(max_util, __cpu_util(i, delta));
4853 * group_norm_util() returns the approximated group util relative to it's
4854 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4855 * energy calculations. Since task executions may or may not overlap in time in
4856 * the group the true normalized util is between max(cpu_norm_util(i)) and
4857 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4858 * latter is used as the estimate as it leads to a more pessimistic energy
4859 * estimate (more busy).
4862 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4865 unsigned long util_sum = 0;
4866 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4868 for_each_cpu(i, sched_group_cpus(sg)) {
4869 delta = calc_util_delta(eenv, i);
4870 util_sum += __cpu_norm_util(i, capacity, delta);
4873 if (util_sum > SCHED_CAPACITY_SCALE)
4874 return SCHED_CAPACITY_SCALE;
4878 static int find_new_capacity(struct energy_env *eenv,
4879 const struct sched_group_energy const *sge)
4882 unsigned long util = group_max_util(eenv);
4884 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4885 if (sge->cap_states[idx].cap >= util)
4889 eenv->cap_idx = idx;
4894 static int group_idle_state(struct sched_group *sg)
4896 int i, state = INT_MAX;
4898 /* Find the shallowest idle state in the sched group. */
4899 for_each_cpu(i, sched_group_cpus(sg))
4900 state = min(state, idle_get_state_idx(cpu_rq(i)));
4902 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4909 * sched_group_energy(): Computes the absolute energy consumption of cpus
4910 * belonging to the sched_group including shared resources shared only by
4911 * members of the group. Iterates over all cpus in the hierarchy below the
4912 * sched_group starting from the bottom working it's way up before going to
4913 * the next cpu until all cpus are covered at all levels. The current
4914 * implementation is likely to gather the same util statistics multiple times.
4915 * This can probably be done in a faster but more complex way.
4916 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4918 static int sched_group_energy(struct energy_env *eenv)
4920 struct sched_domain *sd;
4921 int cpu, total_energy = 0;
4922 struct cpumask visit_cpus;
4923 struct sched_group *sg;
4925 WARN_ON(!eenv->sg_top->sge);
4927 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4929 while (!cpumask_empty(&visit_cpus)) {
4930 struct sched_group *sg_shared_cap = NULL;
4932 cpu = cpumask_first(&visit_cpus);
4935 * Is the group utilization affected by cpus outside this
4938 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4942 * We most probably raced with hotplug; returning a
4943 * wrong energy estimation is better than entering an
4949 sg_shared_cap = sd->parent->groups;
4951 for_each_domain(cpu, sd) {
4954 /* Has this sched_domain already been visited? */
4955 if (sd->child && group_first_cpu(sg) != cpu)
4959 unsigned long group_util;
4960 int sg_busy_energy, sg_idle_energy;
4961 int cap_idx, idle_idx;
4963 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
4964 eenv->sg_cap = sg_shared_cap;
4968 cap_idx = find_new_capacity(eenv, sg->sge);
4970 if (sg->group_weight == 1) {
4971 /* Remove capacity of src CPU (before task move) */
4972 if (eenv->util_delta == 0 &&
4973 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
4974 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
4975 eenv->cap.delta -= eenv->cap.before;
4977 /* Add capacity of dst CPU (after task move) */
4978 if (eenv->util_delta != 0 &&
4979 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
4980 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
4981 eenv->cap.delta += eenv->cap.after;
4985 idle_idx = group_idle_state(sg);
4986 group_util = group_norm_util(eenv, sg);
4987 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
4988 >> SCHED_CAPACITY_SHIFT;
4989 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
4990 * sg->sge->idle_states[idle_idx].power)
4991 >> SCHED_CAPACITY_SHIFT;
4993 total_energy += sg_busy_energy + sg_idle_energy;
4996 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
4998 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5001 } while (sg = sg->next, sg != sd->groups);
5004 cpumask_clear_cpu(cpu, &visit_cpus);
5008 eenv->energy = total_energy;
5012 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5014 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5018 * energy_diff(): Estimate the energy impact of changing the utilization
5019 * distribution. eenv specifies the change: utilisation amount, source, and
5020 * destination cpu. Source or destination cpu may be -1 in which case the
5021 * utilization is removed from or added to the system (e.g. task wake-up). If
5022 * both are specified, the utilization is migrated.
5024 static inline int __energy_diff(struct energy_env *eenv)
5026 struct sched_domain *sd;
5027 struct sched_group *sg;
5028 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5030 struct energy_env eenv_before = {
5032 .src_cpu = eenv->src_cpu,
5033 .dst_cpu = eenv->dst_cpu,
5034 .nrg = { 0, 0, 0, 0},
5038 if (eenv->src_cpu == eenv->dst_cpu)
5041 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5042 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5045 return 0; /* Error */
5050 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5051 eenv_before.sg_top = eenv->sg_top = sg;
5053 if (sched_group_energy(&eenv_before))
5054 return 0; /* Invalid result abort */
5055 energy_before += eenv_before.energy;
5057 /* Keep track of SRC cpu (before) capacity */
5058 eenv->cap.before = eenv_before.cap.before;
5059 eenv->cap.delta = eenv_before.cap.delta;
5061 if (sched_group_energy(eenv))
5062 return 0; /* Invalid result abort */
5063 energy_after += eenv->energy;
5065 } while (sg = sg->next, sg != sd->groups);
5067 eenv->nrg.before = energy_before;
5068 eenv->nrg.after = energy_after;
5069 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5072 trace_sched_energy_diff(eenv->task,
5073 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5074 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5075 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5076 eenv->nrg.delta, eenv->payoff);
5078 return eenv->nrg.diff;
5081 #ifdef CONFIG_SCHED_TUNE
5083 struct target_nrg schedtune_target_nrg;
5086 * System energy normalization
5087 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5088 * corresponding to the specified energy variation.
5091 normalize_energy(int energy_diff)
5094 #ifdef CONFIG_SCHED_DEBUG
5097 /* Check for boundaries */
5098 max_delta = schedtune_target_nrg.max_power;
5099 max_delta -= schedtune_target_nrg.min_power;
5100 WARN_ON(abs(energy_diff) >= max_delta);
5103 /* Do scaling using positive numbers to increase the range */
5104 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5106 /* Scale by energy magnitude */
5107 normalized_nrg <<= SCHED_LOAD_SHIFT;
5109 /* Normalize on max energy for target platform */
5110 normalized_nrg = reciprocal_divide(
5111 normalized_nrg, schedtune_target_nrg.rdiv);
5113 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5117 energy_diff(struct energy_env *eenv)
5119 int boost = schedtune_task_boost(eenv->task);
5122 /* Conpute "absolute" energy diff */
5123 __energy_diff(eenv);
5125 /* Return energy diff when boost margin is 0 */
5127 return eenv->nrg.diff;
5129 /* Compute normalized energy diff */
5130 nrg_delta = normalize_energy(eenv->nrg.diff);
5131 eenv->nrg.delta = nrg_delta;
5133 eenv->payoff = schedtune_accept_deltas(
5139 * When SchedTune is enabled, the energy_diff() function will return
5140 * the computed energy payoff value. Since the energy_diff() return
5141 * value is expected to be negative by its callers, this evaluation
5142 * function return a negative value each time the evaluation return a
5143 * positive payoff, which is the condition for the acceptance of
5144 * a scheduling decision
5146 return -eenv->payoff;
5148 #else /* CONFIG_SCHED_TUNE */
5149 #define energy_diff(eenv) __energy_diff(eenv)
5153 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5154 * A waker of many should wake a different task than the one last awakened
5155 * at a frequency roughly N times higher than one of its wakees. In order
5156 * to determine whether we should let the load spread vs consolodating to
5157 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5158 * partner, and a factor of lls_size higher frequency in the other. With
5159 * both conditions met, we can be relatively sure that the relationship is
5160 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5161 * being client/server, worker/dispatcher, interrupt source or whatever is
5162 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5164 static int wake_wide(struct task_struct *p)
5166 unsigned int master = current->wakee_flips;
5167 unsigned int slave = p->wakee_flips;
5168 int factor = this_cpu_read(sd_llc_size);
5171 swap(master, slave);
5172 if (slave < factor || master < slave * factor)
5177 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5179 s64 this_load, load;
5180 s64 this_eff_load, prev_eff_load;
5181 int idx, this_cpu, prev_cpu;
5182 struct task_group *tg;
5183 unsigned long weight;
5187 this_cpu = smp_processor_id();
5188 prev_cpu = task_cpu(p);
5189 load = source_load(prev_cpu, idx);
5190 this_load = target_load(this_cpu, idx);
5193 * If sync wakeup then subtract the (maximum possible)
5194 * effect of the currently running task from the load
5195 * of the current CPU:
5198 tg = task_group(current);
5199 weight = current->se.avg.load_avg;
5201 this_load += effective_load(tg, this_cpu, -weight, -weight);
5202 load += effective_load(tg, prev_cpu, 0, -weight);
5206 weight = p->se.avg.load_avg;
5209 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5210 * due to the sync cause above having dropped this_load to 0, we'll
5211 * always have an imbalance, but there's really nothing you can do
5212 * about that, so that's good too.
5214 * Otherwise check if either cpus are near enough in load to allow this
5215 * task to be woken on this_cpu.
5217 this_eff_load = 100;
5218 this_eff_load *= capacity_of(prev_cpu);
5220 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5221 prev_eff_load *= capacity_of(this_cpu);
5223 if (this_load > 0) {
5224 this_eff_load *= this_load +
5225 effective_load(tg, this_cpu, weight, weight);
5227 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5230 balanced = this_eff_load <= prev_eff_load;
5232 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5237 schedstat_inc(sd, ttwu_move_affine);
5238 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5243 static inline unsigned long task_util(struct task_struct *p)
5245 #ifdef CONFIG_SCHED_WALT
5246 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5247 unsigned long demand = p->ravg.demand;
5248 return (demand << 10) / walt_ravg_window;
5251 return p->se.avg.util_avg;
5254 unsigned int capacity_margin = 1280; /* ~20% margin */
5256 static inline unsigned long boosted_task_util(struct task_struct *task);
5258 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5260 unsigned long capacity = capacity_of(cpu);
5262 util += boosted_task_util(p);
5264 return (capacity * 1024) > (util * capacity_margin);
5267 static inline bool task_fits_max(struct task_struct *p, int cpu)
5269 unsigned long capacity = capacity_of(cpu);
5270 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5272 if (capacity == max_capacity)
5275 if (capacity * capacity_margin > max_capacity * 1024)
5278 return __task_fits(p, cpu, 0);
5281 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5283 return __task_fits(p, cpu, cpu_util(cpu));
5286 static bool cpu_overutilized(int cpu)
5288 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5291 #ifdef CONFIG_SCHED_TUNE
5294 schedtune_margin(unsigned long signal, long boost)
5296 long long margin = 0;
5299 * Signal proportional compensation (SPC)
5301 * The Boost (B) value is used to compute a Margin (M) which is
5302 * proportional to the complement of the original Signal (S):
5303 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5304 * M = B * S, if B is negative
5305 * The obtained M could be used by the caller to "boost" S.
5308 margin = SCHED_LOAD_SCALE - signal;
5311 margin = -signal * boost;
5313 * Fast integer division by constant:
5314 * Constant : (C) = 100
5315 * Precision : 0.1% (P) = 0.1
5316 * Reference : C * 100 / P (R) = 100000
5319 * Shift bits : ceil(log(R,2)) (S) = 17
5320 * Mult const : round(2^S/C) (M) = 1311
5333 schedtune_cpu_margin(unsigned long util, int cpu)
5335 int boost = schedtune_cpu_boost(cpu);
5340 return schedtune_margin(util, boost);
5344 schedtune_task_margin(struct task_struct *task)
5346 int boost = schedtune_task_boost(task);
5353 util = task_util(task);
5354 margin = schedtune_margin(util, boost);
5359 #else /* CONFIG_SCHED_TUNE */
5362 schedtune_cpu_margin(unsigned long util, int cpu)
5368 schedtune_task_margin(struct task_struct *task)
5373 #endif /* CONFIG_SCHED_TUNE */
5375 static inline unsigned long
5376 boosted_cpu_util(int cpu)
5378 unsigned long util = cpu_util(cpu);
5379 long margin = schedtune_cpu_margin(util, cpu);
5381 trace_sched_boost_cpu(cpu, util, margin);
5383 return util + margin;
5386 static inline unsigned long
5387 boosted_task_util(struct task_struct *task)
5389 unsigned long util = task_util(task);
5390 long margin = schedtune_task_margin(task);
5392 trace_sched_boost_task(task, util, margin);
5394 return util + margin;
5398 * find_idlest_group finds and returns the least busy CPU group within the
5401 static struct sched_group *
5402 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5403 int this_cpu, int sd_flag)
5405 struct sched_group *idlest = NULL, *group = sd->groups;
5406 struct sched_group *fit_group = NULL, *spare_group = NULL;
5407 unsigned long min_load = ULONG_MAX, this_load = 0;
5408 unsigned long fit_capacity = ULONG_MAX;
5409 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5410 int load_idx = sd->forkexec_idx;
5411 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5413 if (sd_flag & SD_BALANCE_WAKE)
5414 load_idx = sd->wake_idx;
5417 unsigned long load, avg_load, spare_capacity;
5421 /* Skip over this group if it has no CPUs allowed */
5422 if (!cpumask_intersects(sched_group_cpus(group),
5423 tsk_cpus_allowed(p)))
5426 local_group = cpumask_test_cpu(this_cpu,
5427 sched_group_cpus(group));
5429 /* Tally up the load of all CPUs in the group */
5432 for_each_cpu(i, sched_group_cpus(group)) {
5433 /* Bias balancing toward cpus of our domain */
5435 load = source_load(i, load_idx);
5437 load = target_load(i, load_idx);
5442 * Look for most energy-efficient group that can fit
5443 * that can fit the task.
5445 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5446 fit_capacity = capacity_of(i);
5451 * Look for group which has most spare capacity on a
5454 spare_capacity = capacity_of(i) - cpu_util(i);
5455 if (spare_capacity > max_spare_capacity) {
5456 max_spare_capacity = spare_capacity;
5457 spare_group = group;
5461 /* Adjust by relative CPU capacity of the group */
5462 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5465 this_load = avg_load;
5466 } else if (avg_load < min_load) {
5467 min_load = avg_load;
5470 } while (group = group->next, group != sd->groups);
5478 if (!idlest || 100*this_load < imbalance*min_load)
5484 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5487 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5489 unsigned long load, min_load = ULONG_MAX;
5490 unsigned int min_exit_latency = UINT_MAX;
5491 u64 latest_idle_timestamp = 0;
5492 int least_loaded_cpu = this_cpu;
5493 int shallowest_idle_cpu = -1;
5496 /* Traverse only the allowed CPUs */
5497 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5498 if (task_fits_spare(p, i)) {
5499 struct rq *rq = cpu_rq(i);
5500 struct cpuidle_state *idle = idle_get_state(rq);
5501 if (idle && idle->exit_latency < min_exit_latency) {
5503 * We give priority to a CPU whose idle state
5504 * has the smallest exit latency irrespective
5505 * of any idle timestamp.
5507 min_exit_latency = idle->exit_latency;
5508 latest_idle_timestamp = rq->idle_stamp;
5509 shallowest_idle_cpu = i;
5510 } else if (idle_cpu(i) &&
5511 (!idle || idle->exit_latency == min_exit_latency) &&
5512 rq->idle_stamp > latest_idle_timestamp) {
5514 * If equal or no active idle state, then
5515 * the most recently idled CPU might have
5518 latest_idle_timestamp = rq->idle_stamp;
5519 shallowest_idle_cpu = i;
5520 } else if (shallowest_idle_cpu == -1) {
5522 * If we haven't found an idle CPU yet
5523 * pick a non-idle one that can fit the task as
5526 shallowest_idle_cpu = i;
5528 } else if (shallowest_idle_cpu == -1) {
5529 load = weighted_cpuload(i);
5530 if (load < min_load || (load == min_load && i == this_cpu)) {
5532 least_loaded_cpu = i;
5537 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5541 * Try and locate an idle CPU in the sched_domain.
5543 static int select_idle_sibling(struct task_struct *p, int target)
5545 struct sched_domain *sd;
5546 struct sched_group *sg;
5547 int i = task_cpu(p);
5549 int best_idle_cstate = -1;
5550 int best_idle_capacity = INT_MAX;
5552 if (!sysctl_sched_cstate_aware) {
5553 if (idle_cpu(target))
5557 * If the prevous cpu is cache affine and idle, don't be stupid.
5559 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5564 * Otherwise, iterate the domains and find an elegible idle cpu.
5566 sd = rcu_dereference(per_cpu(sd_llc, target));
5567 for_each_lower_domain(sd) {
5570 if (!cpumask_intersects(sched_group_cpus(sg),
5571 tsk_cpus_allowed(p)))
5574 if (sysctl_sched_cstate_aware) {
5575 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5576 struct rq *rq = cpu_rq(i);
5577 int idle_idx = idle_get_state_idx(rq);
5578 unsigned long new_usage = boosted_task_util(p);
5579 unsigned long capacity_orig = capacity_orig_of(i);
5580 if (new_usage > capacity_orig || !idle_cpu(i))
5583 if (i == target && new_usage <= capacity_curr_of(target))
5586 if (best_idle < 0 || (idle_idx < best_idle_cstate && capacity_orig <= best_idle_capacity)) {
5588 best_idle_cstate = idle_idx;
5589 best_idle_capacity = capacity_orig;
5593 for_each_cpu(i, sched_group_cpus(sg)) {
5594 if (i == target || !idle_cpu(i))
5598 target = cpumask_first_and(sched_group_cpus(sg),
5599 tsk_cpus_allowed(p));
5604 } while (sg != sd->groups);
5613 static inline int find_best_target(struct task_struct *p, bool boosted)
5616 int target_cpu = -1;
5617 int target_util = 0;
5618 int backup_capacity = 0;
5619 int best_idle_cpu = -1;
5620 int best_idle_cstate = INT_MAX;
5621 int backup_cpu = -1;
5622 unsigned long task_util_boosted, new_util;
5624 task_util_boosted = boosted_task_util(p);
5625 for (iter_cpu = 0; iter_cpu < NR_CPUS; iter_cpu++) {
5631 * favor higher cpus for boosted tasks
5633 int i = boosted ? NR_CPUS-iter_cpu-1 : iter_cpu;
5635 if (!cpu_online(i) || !cpumask_test_cpu(i, tsk_cpus_allowed(p)))
5639 * p's blocked utilization is still accounted for on prev_cpu
5640 * so prev_cpu will receive a negative bias due to the double
5641 * accounting. However, the blocked utilization may be zero.
5643 new_util = cpu_util(i) + task_util_boosted;
5646 * Ensure minimum capacity to grant the required boost.
5647 * The target CPU can be already at a capacity level higher
5648 * than the one required to boost the task.
5650 if (new_util > capacity_orig_of(i))
5653 #ifdef CONFIG_SCHED_WALT
5654 if (walt_cpu_high_irqload(i))
5658 * For boosted tasks we favor idle cpus unconditionally to
5661 if (idle_cpu(i) && boosted) {
5662 if (best_idle_cpu < 0)
5667 cur_capacity = capacity_curr_of(i);
5669 idle_idx = idle_get_state_idx(rq);
5671 if (new_util < cur_capacity) {
5672 if (cpu_rq(i)->nr_running) {
5673 if (target_util == 0 ||
5674 target_util > new_util) {
5676 target_util = new_util;
5678 } else if (!boosted) {
5679 if (best_idle_cpu < 0 ||
5680 (sysctl_sched_cstate_aware &&
5681 best_idle_cstate > idle_idx)) {
5682 best_idle_cstate = idle_idx;
5686 } else if (backup_capacity == 0 ||
5687 backup_capacity > cur_capacity) {
5688 backup_capacity = cur_capacity;
5693 if (boosted && best_idle_cpu >= 0)
5694 target_cpu = best_idle_cpu;
5695 else if (target_cpu < 0)
5696 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5701 static int energy_aware_wake_cpu(struct task_struct *p, int target, int sync)
5703 struct sched_domain *sd;
5704 struct sched_group *sg, *sg_target;
5705 int target_max_cap = INT_MAX;
5706 int target_cpu = task_cpu(p);
5707 unsigned long task_util_boosted, new_util;
5710 if (sysctl_sched_sync_hint_enable && sync) {
5711 int cpu = smp_processor_id();
5712 cpumask_t search_cpus;
5713 cpumask_and(&search_cpus, tsk_cpus_allowed(p), cpu_online_mask);
5714 if (cpumask_test_cpu(cpu, &search_cpus))
5718 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5726 if (sysctl_sched_is_big_little) {
5729 * Find group with sufficient capacity. We only get here if no cpu is
5730 * overutilized. We may end up overutilizing a cpu by adding the task,
5731 * but that should not be any worse than select_idle_sibling().
5732 * load_balance() should sort it out later as we get above the tipping
5736 /* Assuming all cpus are the same in group */
5737 int max_cap_cpu = group_first_cpu(sg);
5740 * Assume smaller max capacity means more energy-efficient.
5741 * Ideally we should query the energy model for the right
5742 * answer but it easily ends up in an exhaustive search.
5744 if (capacity_of(max_cap_cpu) < target_max_cap &&
5745 task_fits_max(p, max_cap_cpu)) {
5747 target_max_cap = capacity_of(max_cap_cpu);
5749 } while (sg = sg->next, sg != sd->groups);
5751 task_util_boosted = boosted_task_util(p);
5752 /* Find cpu with sufficient capacity */
5753 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5755 * p's blocked utilization is still accounted for on prev_cpu
5756 * so prev_cpu will receive a negative bias due to the double
5757 * accounting. However, the blocked utilization may be zero.
5759 new_util = cpu_util(i) + task_util_boosted;
5762 * Ensure minimum capacity to grant the required boost.
5763 * The target CPU can be already at a capacity level higher
5764 * than the one required to boost the task.
5766 if (new_util > capacity_orig_of(i))
5769 if (new_util < capacity_curr_of(i)) {
5771 if (cpu_rq(i)->nr_running)
5775 /* cpu has capacity at higher OPP, keep it as fallback */
5776 if (target_cpu == task_cpu(p))
5781 * Find a cpu with sufficient capacity
5783 #ifdef CONFIG_CGROUP_SCHEDTUNE
5784 bool boosted = schedtune_task_boost(p) > 0;
5788 int tmp_target = find_best_target(p, boosted);
5789 if (tmp_target >= 0)
5790 target_cpu = tmp_target;
5791 if (boosted && idle_cpu(target_cpu))
5795 if (target_cpu != task_cpu(p)) {
5796 struct energy_env eenv = {
5797 .util_delta = task_util(p),
5798 .src_cpu = task_cpu(p),
5799 .dst_cpu = target_cpu,
5803 /* Not enough spare capacity on previous cpu */
5804 if (cpu_overutilized(task_cpu(p)))
5807 if (energy_diff(&eenv) >= 0)
5815 * select_task_rq_fair: Select target runqueue for the waking task in domains
5816 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5817 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5819 * Balances load by selecting the idlest cpu in the idlest group, or under
5820 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5822 * Returns the target cpu number.
5824 * preempt must be disabled.
5827 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5829 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5830 int cpu = smp_processor_id();
5831 int new_cpu = prev_cpu;
5832 int want_affine = 0;
5833 int sync = wake_flags & WF_SYNC;
5835 if (sd_flag & SD_BALANCE_WAKE)
5836 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5837 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5841 for_each_domain(cpu, tmp) {
5842 if (!(tmp->flags & SD_LOAD_BALANCE))
5846 * If both cpu and prev_cpu are part of this domain,
5847 * cpu is a valid SD_WAKE_AFFINE target.
5849 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5850 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5855 if (tmp->flags & sd_flag)
5857 else if (!want_affine)
5862 sd = NULL; /* Prefer wake_affine over balance flags */
5863 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5868 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5869 new_cpu = energy_aware_wake_cpu(p, prev_cpu, sync);
5870 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5871 new_cpu = select_idle_sibling(p, new_cpu);
5874 struct sched_group *group;
5877 if (!(sd->flags & sd_flag)) {
5882 group = find_idlest_group(sd, p, cpu, sd_flag);
5888 new_cpu = find_idlest_cpu(group, p, cpu);
5889 if (new_cpu == -1 || new_cpu == cpu) {
5890 /* Now try balancing at a lower domain level of cpu */
5895 /* Now try balancing at a lower domain level of new_cpu */
5897 weight = sd->span_weight;
5899 for_each_domain(cpu, tmp) {
5900 if (weight <= tmp->span_weight)
5902 if (tmp->flags & sd_flag)
5905 /* while loop will break here if sd == NULL */
5913 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5914 * cfs_rq_of(p) references at time of call are still valid and identify the
5915 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5916 * other assumptions, including the state of rq->lock, should be made.
5918 static void migrate_task_rq_fair(struct task_struct *p)
5921 * We are supposed to update the task to "current" time, then its up to date
5922 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5923 * what current time is, so simply throw away the out-of-date time. This
5924 * will result in the wakee task is less decayed, but giving the wakee more
5925 * load sounds not bad.
5927 remove_entity_load_avg(&p->se);
5929 /* Tell new CPU we are migrated */
5930 p->se.avg.last_update_time = 0;
5932 /* We have migrated, no longer consider this task hot */
5933 p->se.exec_start = 0;
5936 static void task_dead_fair(struct task_struct *p)
5938 remove_entity_load_avg(&p->se);
5941 #define task_fits_max(p, cpu) true
5942 #endif /* CONFIG_SMP */
5944 static unsigned long
5945 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5947 unsigned long gran = sysctl_sched_wakeup_granularity;
5950 * Since its curr running now, convert the gran from real-time
5951 * to virtual-time in his units.
5953 * By using 'se' instead of 'curr' we penalize light tasks, so
5954 * they get preempted easier. That is, if 'se' < 'curr' then
5955 * the resulting gran will be larger, therefore penalizing the
5956 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5957 * be smaller, again penalizing the lighter task.
5959 * This is especially important for buddies when the leftmost
5960 * task is higher priority than the buddy.
5962 return calc_delta_fair(gran, se);
5966 * Should 'se' preempt 'curr'.
5980 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5982 s64 gran, vdiff = curr->vruntime - se->vruntime;
5987 gran = wakeup_gran(curr, se);
5994 static void set_last_buddy(struct sched_entity *se)
5996 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5999 for_each_sched_entity(se)
6000 cfs_rq_of(se)->last = se;
6003 static void set_next_buddy(struct sched_entity *se)
6005 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6008 for_each_sched_entity(se)
6009 cfs_rq_of(se)->next = se;
6012 static void set_skip_buddy(struct sched_entity *se)
6014 for_each_sched_entity(se)
6015 cfs_rq_of(se)->skip = se;
6019 * Preempt the current task with a newly woken task if needed:
6021 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6023 struct task_struct *curr = rq->curr;
6024 struct sched_entity *se = &curr->se, *pse = &p->se;
6025 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6026 int scale = cfs_rq->nr_running >= sched_nr_latency;
6027 int next_buddy_marked = 0;
6029 if (unlikely(se == pse))
6033 * This is possible from callers such as attach_tasks(), in which we
6034 * unconditionally check_prempt_curr() after an enqueue (which may have
6035 * lead to a throttle). This both saves work and prevents false
6036 * next-buddy nomination below.
6038 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6041 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6042 set_next_buddy(pse);
6043 next_buddy_marked = 1;
6047 * We can come here with TIF_NEED_RESCHED already set from new task
6050 * Note: this also catches the edge-case of curr being in a throttled
6051 * group (e.g. via set_curr_task), since update_curr() (in the
6052 * enqueue of curr) will have resulted in resched being set. This
6053 * prevents us from potentially nominating it as a false LAST_BUDDY
6056 if (test_tsk_need_resched(curr))
6059 /* Idle tasks are by definition preempted by non-idle tasks. */
6060 if (unlikely(curr->policy == SCHED_IDLE) &&
6061 likely(p->policy != SCHED_IDLE))
6065 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6066 * is driven by the tick):
6068 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6071 find_matching_se(&se, &pse);
6072 update_curr(cfs_rq_of(se));
6074 if (wakeup_preempt_entity(se, pse) == 1) {
6076 * Bias pick_next to pick the sched entity that is
6077 * triggering this preemption.
6079 if (!next_buddy_marked)
6080 set_next_buddy(pse);
6089 * Only set the backward buddy when the current task is still
6090 * on the rq. This can happen when a wakeup gets interleaved
6091 * with schedule on the ->pre_schedule() or idle_balance()
6092 * point, either of which can * drop the rq lock.
6094 * Also, during early boot the idle thread is in the fair class,
6095 * for obvious reasons its a bad idea to schedule back to it.
6097 if (unlikely(!se->on_rq || curr == rq->idle))
6100 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6104 static struct task_struct *
6105 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6107 struct cfs_rq *cfs_rq = &rq->cfs;
6108 struct sched_entity *se;
6109 struct task_struct *p;
6113 #ifdef CONFIG_FAIR_GROUP_SCHED
6114 if (!cfs_rq->nr_running)
6117 if (prev->sched_class != &fair_sched_class)
6121 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6122 * likely that a next task is from the same cgroup as the current.
6124 * Therefore attempt to avoid putting and setting the entire cgroup
6125 * hierarchy, only change the part that actually changes.
6129 struct sched_entity *curr = cfs_rq->curr;
6132 * Since we got here without doing put_prev_entity() we also
6133 * have to consider cfs_rq->curr. If it is still a runnable
6134 * entity, update_curr() will update its vruntime, otherwise
6135 * forget we've ever seen it.
6139 update_curr(cfs_rq);
6144 * This call to check_cfs_rq_runtime() will do the
6145 * throttle and dequeue its entity in the parent(s).
6146 * Therefore the 'simple' nr_running test will indeed
6149 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6153 se = pick_next_entity(cfs_rq, curr);
6154 cfs_rq = group_cfs_rq(se);
6160 * Since we haven't yet done put_prev_entity and if the selected task
6161 * is a different task than we started out with, try and touch the
6162 * least amount of cfs_rqs.
6165 struct sched_entity *pse = &prev->se;
6167 while (!(cfs_rq = is_same_group(se, pse))) {
6168 int se_depth = se->depth;
6169 int pse_depth = pse->depth;
6171 if (se_depth <= pse_depth) {
6172 put_prev_entity(cfs_rq_of(pse), pse);
6173 pse = parent_entity(pse);
6175 if (se_depth >= pse_depth) {
6176 set_next_entity(cfs_rq_of(se), se);
6177 se = parent_entity(se);
6181 put_prev_entity(cfs_rq, pse);
6182 set_next_entity(cfs_rq, se);
6185 if (hrtick_enabled(rq))
6186 hrtick_start_fair(rq, p);
6188 rq->misfit_task = !task_fits_max(p, rq->cpu);
6195 if (!cfs_rq->nr_running)
6198 put_prev_task(rq, prev);
6201 se = pick_next_entity(cfs_rq, NULL);
6202 set_next_entity(cfs_rq, se);
6203 cfs_rq = group_cfs_rq(se);
6208 if (hrtick_enabled(rq))
6209 hrtick_start_fair(rq, p);
6211 rq->misfit_task = !task_fits_max(p, rq->cpu);
6216 rq->misfit_task = 0;
6218 * This is OK, because current is on_cpu, which avoids it being picked
6219 * for load-balance and preemption/IRQs are still disabled avoiding
6220 * further scheduler activity on it and we're being very careful to
6221 * re-start the picking loop.
6223 lockdep_unpin_lock(&rq->lock);
6224 new_tasks = idle_balance(rq);
6225 lockdep_pin_lock(&rq->lock);
6227 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6228 * possible for any higher priority task to appear. In that case we
6229 * must re-start the pick_next_entity() loop.
6241 * Account for a descheduled task:
6243 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6245 struct sched_entity *se = &prev->se;
6246 struct cfs_rq *cfs_rq;
6248 for_each_sched_entity(se) {
6249 cfs_rq = cfs_rq_of(se);
6250 put_prev_entity(cfs_rq, se);
6255 * sched_yield() is very simple
6257 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6259 static void yield_task_fair(struct rq *rq)
6261 struct task_struct *curr = rq->curr;
6262 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6263 struct sched_entity *se = &curr->se;
6266 * Are we the only task in the tree?
6268 if (unlikely(rq->nr_running == 1))
6271 clear_buddies(cfs_rq, se);
6273 if (curr->policy != SCHED_BATCH) {
6274 update_rq_clock(rq);
6276 * Update run-time statistics of the 'current'.
6278 update_curr(cfs_rq);
6280 * Tell update_rq_clock() that we've just updated,
6281 * so we don't do microscopic update in schedule()
6282 * and double the fastpath cost.
6284 rq_clock_skip_update(rq, true);
6290 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6292 struct sched_entity *se = &p->se;
6294 /* throttled hierarchies are not runnable */
6295 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6298 /* Tell the scheduler that we'd really like pse to run next. */
6301 yield_task_fair(rq);
6307 /**************************************************
6308 * Fair scheduling class load-balancing methods.
6312 * The purpose of load-balancing is to achieve the same basic fairness the
6313 * per-cpu scheduler provides, namely provide a proportional amount of compute
6314 * time to each task. This is expressed in the following equation:
6316 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6318 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6319 * W_i,0 is defined as:
6321 * W_i,0 = \Sum_j w_i,j (2)
6323 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6324 * is derived from the nice value as per prio_to_weight[].
6326 * The weight average is an exponential decay average of the instantaneous
6329 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6331 * C_i is the compute capacity of cpu i, typically it is the
6332 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6333 * can also include other factors [XXX].
6335 * To achieve this balance we define a measure of imbalance which follows
6336 * directly from (1):
6338 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6340 * We them move tasks around to minimize the imbalance. In the continuous
6341 * function space it is obvious this converges, in the discrete case we get
6342 * a few fun cases generally called infeasible weight scenarios.
6345 * - infeasible weights;
6346 * - local vs global optima in the discrete case. ]
6351 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6352 * for all i,j solution, we create a tree of cpus that follows the hardware
6353 * topology where each level pairs two lower groups (or better). This results
6354 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6355 * tree to only the first of the previous level and we decrease the frequency
6356 * of load-balance at each level inv. proportional to the number of cpus in
6362 * \Sum { --- * --- * 2^i } = O(n) (5)
6364 * `- size of each group
6365 * | | `- number of cpus doing load-balance
6367 * `- sum over all levels
6369 * Coupled with a limit on how many tasks we can migrate every balance pass,
6370 * this makes (5) the runtime complexity of the balancer.
6372 * An important property here is that each CPU is still (indirectly) connected
6373 * to every other cpu in at most O(log n) steps:
6375 * The adjacency matrix of the resulting graph is given by:
6378 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6381 * And you'll find that:
6383 * A^(log_2 n)_i,j != 0 for all i,j (7)
6385 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6386 * The task movement gives a factor of O(m), giving a convergence complexity
6389 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6394 * In order to avoid CPUs going idle while there's still work to do, new idle
6395 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6396 * tree itself instead of relying on other CPUs to bring it work.
6398 * This adds some complexity to both (5) and (8) but it reduces the total idle
6406 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6409 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6414 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6416 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6418 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6421 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6422 * rewrite all of this once again.]
6425 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6427 enum fbq_type { regular, remote, all };
6436 #define LBF_ALL_PINNED 0x01
6437 #define LBF_NEED_BREAK 0x02
6438 #define LBF_DST_PINNED 0x04
6439 #define LBF_SOME_PINNED 0x08
6442 struct sched_domain *sd;
6450 struct cpumask *dst_grpmask;
6452 enum cpu_idle_type idle;
6454 unsigned int src_grp_nr_running;
6455 /* The set of CPUs under consideration for load-balancing */
6456 struct cpumask *cpus;
6461 unsigned int loop_break;
6462 unsigned int loop_max;
6464 enum fbq_type fbq_type;
6465 enum group_type busiest_group_type;
6466 struct list_head tasks;
6470 * Is this task likely cache-hot:
6472 static int task_hot(struct task_struct *p, struct lb_env *env)
6476 lockdep_assert_held(&env->src_rq->lock);
6478 if (p->sched_class != &fair_sched_class)
6481 if (unlikely(p->policy == SCHED_IDLE))
6485 * Buddy candidates are cache hot:
6487 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6488 (&p->se == cfs_rq_of(&p->se)->next ||
6489 &p->se == cfs_rq_of(&p->se)->last))
6492 if (sysctl_sched_migration_cost == -1)
6494 if (sysctl_sched_migration_cost == 0)
6497 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6499 return delta < (s64)sysctl_sched_migration_cost;
6502 #ifdef CONFIG_NUMA_BALANCING
6504 * Returns 1, if task migration degrades locality
6505 * Returns 0, if task migration improves locality i.e migration preferred.
6506 * Returns -1, if task migration is not affected by locality.
6508 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6510 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6511 unsigned long src_faults, dst_faults;
6512 int src_nid, dst_nid;
6514 if (!static_branch_likely(&sched_numa_balancing))
6517 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6520 src_nid = cpu_to_node(env->src_cpu);
6521 dst_nid = cpu_to_node(env->dst_cpu);
6523 if (src_nid == dst_nid)
6526 /* Migrating away from the preferred node is always bad. */
6527 if (src_nid == p->numa_preferred_nid) {
6528 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6534 /* Encourage migration to the preferred node. */
6535 if (dst_nid == p->numa_preferred_nid)
6539 src_faults = group_faults(p, src_nid);
6540 dst_faults = group_faults(p, dst_nid);
6542 src_faults = task_faults(p, src_nid);
6543 dst_faults = task_faults(p, dst_nid);
6546 return dst_faults < src_faults;
6550 static inline int migrate_degrades_locality(struct task_struct *p,
6558 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6561 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6565 lockdep_assert_held(&env->src_rq->lock);
6568 * We do not migrate tasks that are:
6569 * 1) throttled_lb_pair, or
6570 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6571 * 3) running (obviously), or
6572 * 4) are cache-hot on their current CPU.
6574 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6577 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6580 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6582 env->flags |= LBF_SOME_PINNED;
6585 * Remember if this task can be migrated to any other cpu in
6586 * our sched_group. We may want to revisit it if we couldn't
6587 * meet load balance goals by pulling other tasks on src_cpu.
6589 * Also avoid computing new_dst_cpu if we have already computed
6590 * one in current iteration.
6592 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6595 /* Prevent to re-select dst_cpu via env's cpus */
6596 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6597 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6598 env->flags |= LBF_DST_PINNED;
6599 env->new_dst_cpu = cpu;
6607 /* Record that we found atleast one task that could run on dst_cpu */
6608 env->flags &= ~LBF_ALL_PINNED;
6610 if (task_running(env->src_rq, p)) {
6611 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6616 * Aggressive migration if:
6617 * 1) destination numa is preferred
6618 * 2) task is cache cold, or
6619 * 3) too many balance attempts have failed.
6621 tsk_cache_hot = migrate_degrades_locality(p, env);
6622 if (tsk_cache_hot == -1)
6623 tsk_cache_hot = task_hot(p, env);
6625 if (tsk_cache_hot <= 0 ||
6626 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6627 if (tsk_cache_hot == 1) {
6628 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6629 schedstat_inc(p, se.statistics.nr_forced_migrations);
6634 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6639 * detach_task() -- detach the task for the migration specified in env
6641 static void detach_task(struct task_struct *p, struct lb_env *env)
6643 lockdep_assert_held(&env->src_rq->lock);
6645 deactivate_task(env->src_rq, p, 0);
6646 p->on_rq = TASK_ON_RQ_MIGRATING;
6647 double_lock_balance(env->src_rq, env->dst_rq);
6648 set_task_cpu(p, env->dst_cpu);
6649 double_unlock_balance(env->src_rq, env->dst_rq);
6653 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6654 * part of active balancing operations within "domain".
6656 * Returns a task if successful and NULL otherwise.
6658 static struct task_struct *detach_one_task(struct lb_env *env)
6660 struct task_struct *p, *n;
6662 lockdep_assert_held(&env->src_rq->lock);
6664 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6665 if (!can_migrate_task(p, env))
6668 detach_task(p, env);
6671 * Right now, this is only the second place where
6672 * lb_gained[env->idle] is updated (other is detach_tasks)
6673 * so we can safely collect stats here rather than
6674 * inside detach_tasks().
6676 schedstat_inc(env->sd, lb_gained[env->idle]);
6682 static const unsigned int sched_nr_migrate_break = 32;
6685 * detach_tasks() -- tries to detach up to imbalance weighted load from
6686 * busiest_rq, as part of a balancing operation within domain "sd".
6688 * Returns number of detached tasks if successful and 0 otherwise.
6690 static int detach_tasks(struct lb_env *env)
6692 struct list_head *tasks = &env->src_rq->cfs_tasks;
6693 struct task_struct *p;
6697 lockdep_assert_held(&env->src_rq->lock);
6699 if (env->imbalance <= 0)
6702 while (!list_empty(tasks)) {
6704 * We don't want to steal all, otherwise we may be treated likewise,
6705 * which could at worst lead to a livelock crash.
6707 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6710 p = list_first_entry(tasks, struct task_struct, se.group_node);
6713 /* We've more or less seen every task there is, call it quits */
6714 if (env->loop > env->loop_max)
6717 /* take a breather every nr_migrate tasks */
6718 if (env->loop > env->loop_break) {
6719 env->loop_break += sched_nr_migrate_break;
6720 env->flags |= LBF_NEED_BREAK;
6724 if (!can_migrate_task(p, env))
6727 load = task_h_load(p);
6729 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6732 if ((load / 2) > env->imbalance)
6735 detach_task(p, env);
6736 list_add(&p->se.group_node, &env->tasks);
6739 env->imbalance -= load;
6741 #ifdef CONFIG_PREEMPT
6743 * NEWIDLE balancing is a source of latency, so preemptible
6744 * kernels will stop after the first task is detached to minimize
6745 * the critical section.
6747 if (env->idle == CPU_NEWLY_IDLE)
6752 * We only want to steal up to the prescribed amount of
6755 if (env->imbalance <= 0)
6760 list_move_tail(&p->se.group_node, tasks);
6764 * Right now, this is one of only two places we collect this stat
6765 * so we can safely collect detach_one_task() stats here rather
6766 * than inside detach_one_task().
6768 schedstat_add(env->sd, lb_gained[env->idle], detached);
6774 * attach_task() -- attach the task detached by detach_task() to its new rq.
6776 static void attach_task(struct rq *rq, struct task_struct *p)
6778 lockdep_assert_held(&rq->lock);
6780 BUG_ON(task_rq(p) != rq);
6781 p->on_rq = TASK_ON_RQ_QUEUED;
6782 activate_task(rq, p, 0);
6783 check_preempt_curr(rq, p, 0);
6787 * attach_one_task() -- attaches the task returned from detach_one_task() to
6790 static void attach_one_task(struct rq *rq, struct task_struct *p)
6792 raw_spin_lock(&rq->lock);
6795 * We want to potentially raise target_cpu's OPP.
6797 update_capacity_of(cpu_of(rq));
6798 raw_spin_unlock(&rq->lock);
6802 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6805 static void attach_tasks(struct lb_env *env)
6807 struct list_head *tasks = &env->tasks;
6808 struct task_struct *p;
6810 raw_spin_lock(&env->dst_rq->lock);
6812 while (!list_empty(tasks)) {
6813 p = list_first_entry(tasks, struct task_struct, se.group_node);
6814 list_del_init(&p->se.group_node);
6816 attach_task(env->dst_rq, p);
6820 * We want to potentially raise env.dst_cpu's OPP.
6822 update_capacity_of(env->dst_cpu);
6824 raw_spin_unlock(&env->dst_rq->lock);
6827 #ifdef CONFIG_FAIR_GROUP_SCHED
6828 static void update_blocked_averages(int cpu)
6830 struct rq *rq = cpu_rq(cpu);
6831 struct cfs_rq *cfs_rq;
6832 unsigned long flags;
6834 raw_spin_lock_irqsave(&rq->lock, flags);
6835 update_rq_clock(rq);
6838 * Iterates the task_group tree in a bottom up fashion, see
6839 * list_add_leaf_cfs_rq() for details.
6841 for_each_leaf_cfs_rq(rq, cfs_rq) {
6842 /* throttled entities do not contribute to load */
6843 if (throttled_hierarchy(cfs_rq))
6846 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6847 update_tg_load_avg(cfs_rq, 0);
6849 raw_spin_unlock_irqrestore(&rq->lock, flags);
6853 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6854 * This needs to be done in a top-down fashion because the load of a child
6855 * group is a fraction of its parents load.
6857 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6859 struct rq *rq = rq_of(cfs_rq);
6860 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6861 unsigned long now = jiffies;
6864 if (cfs_rq->last_h_load_update == now)
6867 cfs_rq->h_load_next = NULL;
6868 for_each_sched_entity(se) {
6869 cfs_rq = cfs_rq_of(se);
6870 cfs_rq->h_load_next = se;
6871 if (cfs_rq->last_h_load_update == now)
6876 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6877 cfs_rq->last_h_load_update = now;
6880 while ((se = cfs_rq->h_load_next) != NULL) {
6881 load = cfs_rq->h_load;
6882 load = div64_ul(load * se->avg.load_avg,
6883 cfs_rq_load_avg(cfs_rq) + 1);
6884 cfs_rq = group_cfs_rq(se);
6885 cfs_rq->h_load = load;
6886 cfs_rq->last_h_load_update = now;
6890 static unsigned long task_h_load(struct task_struct *p)
6892 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6894 update_cfs_rq_h_load(cfs_rq);
6895 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6896 cfs_rq_load_avg(cfs_rq) + 1);
6899 static inline void update_blocked_averages(int cpu)
6901 struct rq *rq = cpu_rq(cpu);
6902 struct cfs_rq *cfs_rq = &rq->cfs;
6903 unsigned long flags;
6905 raw_spin_lock_irqsave(&rq->lock, flags);
6906 update_rq_clock(rq);
6907 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6908 raw_spin_unlock_irqrestore(&rq->lock, flags);
6911 static unsigned long task_h_load(struct task_struct *p)
6913 return p->se.avg.load_avg;
6917 /********** Helpers for find_busiest_group ************************/
6920 * sg_lb_stats - stats of a sched_group required for load_balancing
6922 struct sg_lb_stats {
6923 unsigned long avg_load; /*Avg load across the CPUs of the group */
6924 unsigned long group_load; /* Total load over the CPUs of the group */
6925 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6926 unsigned long load_per_task;
6927 unsigned long group_capacity;
6928 unsigned long group_util; /* Total utilization of the group */
6929 unsigned int sum_nr_running; /* Nr tasks running in the group */
6930 unsigned int idle_cpus;
6931 unsigned int group_weight;
6932 enum group_type group_type;
6933 int group_no_capacity;
6934 int group_misfit_task; /* A cpu has a task too big for its capacity */
6935 #ifdef CONFIG_NUMA_BALANCING
6936 unsigned int nr_numa_running;
6937 unsigned int nr_preferred_running;
6942 * sd_lb_stats - Structure to store the statistics of a sched_domain
6943 * during load balancing.
6945 struct sd_lb_stats {
6946 struct sched_group *busiest; /* Busiest group in this sd */
6947 struct sched_group *local; /* Local group in this sd */
6948 unsigned long total_load; /* Total load of all groups in sd */
6949 unsigned long total_capacity; /* Total capacity of all groups in sd */
6950 unsigned long avg_load; /* Average load across all groups in sd */
6952 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6953 struct sg_lb_stats local_stat; /* Statistics of the local group */
6956 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6959 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6960 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6961 * We must however clear busiest_stat::avg_load because
6962 * update_sd_pick_busiest() reads this before assignment.
6964 *sds = (struct sd_lb_stats){
6968 .total_capacity = 0UL,
6971 .sum_nr_running = 0,
6972 .group_type = group_other,
6978 * get_sd_load_idx - Obtain the load index for a given sched domain.
6979 * @sd: The sched_domain whose load_idx is to be obtained.
6980 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6982 * Return: The load index.
6984 static inline int get_sd_load_idx(struct sched_domain *sd,
6985 enum cpu_idle_type idle)
6991 load_idx = sd->busy_idx;
6994 case CPU_NEWLY_IDLE:
6995 load_idx = sd->newidle_idx;
6998 load_idx = sd->idle_idx;
7005 static unsigned long scale_rt_capacity(int cpu)
7007 struct rq *rq = cpu_rq(cpu);
7008 u64 total, used, age_stamp, avg;
7012 * Since we're reading these variables without serialization make sure
7013 * we read them once before doing sanity checks on them.
7015 age_stamp = READ_ONCE(rq->age_stamp);
7016 avg = READ_ONCE(rq->rt_avg);
7017 delta = __rq_clock_broken(rq) - age_stamp;
7019 if (unlikely(delta < 0))
7022 total = sched_avg_period() + delta;
7024 used = div_u64(avg, total);
7027 * deadline bandwidth is defined at system level so we must
7028 * weight this bandwidth with the max capacity of the system.
7029 * As a reminder, avg_bw is 20bits width and
7030 * scale_cpu_capacity is 10 bits width
7032 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7034 if (likely(used < SCHED_CAPACITY_SCALE))
7035 return SCHED_CAPACITY_SCALE - used;
7040 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7042 raw_spin_lock_init(&mcc->lock);
7047 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7049 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7050 struct sched_group *sdg = sd->groups;
7051 struct max_cpu_capacity *mcc;
7052 unsigned long max_capacity;
7054 unsigned long flags;
7056 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7058 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7060 raw_spin_lock_irqsave(&mcc->lock, flags);
7061 max_capacity = mcc->val;
7062 max_cap_cpu = mcc->cpu;
7064 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7065 (max_capacity < capacity)) {
7066 mcc->val = capacity;
7068 #ifdef CONFIG_SCHED_DEBUG
7069 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7070 pr_info("CPU%d: update max cpu_capacity %lu\n", cpu, capacity);
7074 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7076 skip_unlock: __attribute__ ((unused));
7077 capacity *= scale_rt_capacity(cpu);
7078 capacity >>= SCHED_CAPACITY_SHIFT;
7083 cpu_rq(cpu)->cpu_capacity = capacity;
7084 sdg->sgc->capacity = capacity;
7085 sdg->sgc->max_capacity = capacity;
7088 void update_group_capacity(struct sched_domain *sd, int cpu)
7090 struct sched_domain *child = sd->child;
7091 struct sched_group *group, *sdg = sd->groups;
7092 unsigned long capacity, max_capacity;
7093 unsigned long interval;
7095 interval = msecs_to_jiffies(sd->balance_interval);
7096 interval = clamp(interval, 1UL, max_load_balance_interval);
7097 sdg->sgc->next_update = jiffies + interval;
7100 update_cpu_capacity(sd, cpu);
7107 if (child->flags & SD_OVERLAP) {
7109 * SD_OVERLAP domains cannot assume that child groups
7110 * span the current group.
7113 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7114 struct sched_group_capacity *sgc;
7115 struct rq *rq = cpu_rq(cpu);
7118 * build_sched_domains() -> init_sched_groups_capacity()
7119 * gets here before we've attached the domains to the
7122 * Use capacity_of(), which is set irrespective of domains
7123 * in update_cpu_capacity().
7125 * This avoids capacity from being 0 and
7126 * causing divide-by-zero issues on boot.
7128 if (unlikely(!rq->sd)) {
7129 capacity += capacity_of(cpu);
7131 sgc = rq->sd->groups->sgc;
7132 capacity += sgc->capacity;
7135 max_capacity = max(capacity, max_capacity);
7139 * !SD_OVERLAP domains can assume that child groups
7140 * span the current group.
7143 group = child->groups;
7145 struct sched_group_capacity *sgc = group->sgc;
7147 capacity += sgc->capacity;
7148 max_capacity = max(sgc->max_capacity, max_capacity);
7149 group = group->next;
7150 } while (group != child->groups);
7153 sdg->sgc->capacity = capacity;
7154 sdg->sgc->max_capacity = max_capacity;
7158 * Check whether the capacity of the rq has been noticeably reduced by side
7159 * activity. The imbalance_pct is used for the threshold.
7160 * Return true is the capacity is reduced
7163 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7165 return ((rq->cpu_capacity * sd->imbalance_pct) <
7166 (rq->cpu_capacity_orig * 100));
7170 * Group imbalance indicates (and tries to solve) the problem where balancing
7171 * groups is inadequate due to tsk_cpus_allowed() constraints.
7173 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7174 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7177 * { 0 1 2 3 } { 4 5 6 7 }
7180 * If we were to balance group-wise we'd place two tasks in the first group and
7181 * two tasks in the second group. Clearly this is undesired as it will overload
7182 * cpu 3 and leave one of the cpus in the second group unused.
7184 * The current solution to this issue is detecting the skew in the first group
7185 * by noticing the lower domain failed to reach balance and had difficulty
7186 * moving tasks due to affinity constraints.
7188 * When this is so detected; this group becomes a candidate for busiest; see
7189 * update_sd_pick_busiest(). And calculate_imbalance() and
7190 * find_busiest_group() avoid some of the usual balance conditions to allow it
7191 * to create an effective group imbalance.
7193 * This is a somewhat tricky proposition since the next run might not find the
7194 * group imbalance and decide the groups need to be balanced again. A most
7195 * subtle and fragile situation.
7198 static inline int sg_imbalanced(struct sched_group *group)
7200 return group->sgc->imbalance;
7204 * group_has_capacity returns true if the group has spare capacity that could
7205 * be used by some tasks.
7206 * We consider that a group has spare capacity if the * number of task is
7207 * smaller than the number of CPUs or if the utilization is lower than the
7208 * available capacity for CFS tasks.
7209 * For the latter, we use a threshold to stabilize the state, to take into
7210 * account the variance of the tasks' load and to return true if the available
7211 * capacity in meaningful for the load balancer.
7212 * As an example, an available capacity of 1% can appear but it doesn't make
7213 * any benefit for the load balance.
7216 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7218 if (sgs->sum_nr_running < sgs->group_weight)
7221 if ((sgs->group_capacity * 100) >
7222 (sgs->group_util * env->sd->imbalance_pct))
7229 * group_is_overloaded returns true if the group has more tasks than it can
7231 * group_is_overloaded is not equals to !group_has_capacity because a group
7232 * with the exact right number of tasks, has no more spare capacity but is not
7233 * overloaded so both group_has_capacity and group_is_overloaded return
7237 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7239 if (sgs->sum_nr_running <= sgs->group_weight)
7242 if ((sgs->group_capacity * 100) <
7243 (sgs->group_util * env->sd->imbalance_pct))
7251 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7252 * per-cpu capacity than sched_group ref.
7255 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7257 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7258 ref->sgc->max_capacity;
7262 group_type group_classify(struct sched_group *group,
7263 struct sg_lb_stats *sgs)
7265 if (sgs->group_no_capacity)
7266 return group_overloaded;
7268 if (sg_imbalanced(group))
7269 return group_imbalanced;
7271 if (sgs->group_misfit_task)
7272 return group_misfit_task;
7278 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7279 * @env: The load balancing environment.
7280 * @group: sched_group whose statistics are to be updated.
7281 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7282 * @local_group: Does group contain this_cpu.
7283 * @sgs: variable to hold the statistics for this group.
7284 * @overload: Indicate more than one runnable task for any CPU.
7285 * @overutilized: Indicate overutilization for any CPU.
7287 static inline void update_sg_lb_stats(struct lb_env *env,
7288 struct sched_group *group, int load_idx,
7289 int local_group, struct sg_lb_stats *sgs,
7290 bool *overload, bool *overutilized)
7295 memset(sgs, 0, sizeof(*sgs));
7297 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7298 struct rq *rq = cpu_rq(i);
7300 /* Bias balancing toward cpus of our domain */
7302 load = target_load(i, load_idx);
7304 load = source_load(i, load_idx);
7306 sgs->group_load += load;
7307 sgs->group_util += cpu_util(i);
7308 sgs->sum_nr_running += rq->cfs.h_nr_running;
7310 if (rq->nr_running > 1)
7313 #ifdef CONFIG_NUMA_BALANCING
7314 sgs->nr_numa_running += rq->nr_numa_running;
7315 sgs->nr_preferred_running += rq->nr_preferred_running;
7317 sgs->sum_weighted_load += weighted_cpuload(i);
7321 if (cpu_overutilized(i)) {
7322 *overutilized = true;
7323 if (!sgs->group_misfit_task && rq->misfit_task)
7324 sgs->group_misfit_task = capacity_of(i);
7328 /* Adjust by relative CPU capacity of the group */
7329 sgs->group_capacity = group->sgc->capacity;
7330 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7332 if (sgs->sum_nr_running)
7333 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7335 sgs->group_weight = group->group_weight;
7337 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7338 sgs->group_type = group_classify(group, sgs);
7342 * update_sd_pick_busiest - return 1 on busiest group
7343 * @env: The load balancing environment.
7344 * @sds: sched_domain statistics
7345 * @sg: sched_group candidate to be checked for being the busiest
7346 * @sgs: sched_group statistics
7348 * Determine if @sg is a busier group than the previously selected
7351 * Return: %true if @sg is a busier group than the previously selected
7352 * busiest group. %false otherwise.
7354 static bool update_sd_pick_busiest(struct lb_env *env,
7355 struct sd_lb_stats *sds,
7356 struct sched_group *sg,
7357 struct sg_lb_stats *sgs)
7359 struct sg_lb_stats *busiest = &sds->busiest_stat;
7361 if (sgs->group_type > busiest->group_type)
7364 if (sgs->group_type < busiest->group_type)
7368 * Candidate sg doesn't face any serious load-balance problems
7369 * so don't pick it if the local sg is already filled up.
7371 if (sgs->group_type == group_other &&
7372 !group_has_capacity(env, &sds->local_stat))
7375 if (sgs->avg_load <= busiest->avg_load)
7379 * Candiate sg has no more than one task per cpu and has higher
7380 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7382 if (sgs->sum_nr_running <= sgs->group_weight &&
7383 group_smaller_cpu_capacity(sds->local, sg))
7386 /* This is the busiest node in its class. */
7387 if (!(env->sd->flags & SD_ASYM_PACKING))
7391 * ASYM_PACKING needs to move all the work to the lowest
7392 * numbered CPUs in the group, therefore mark all groups
7393 * higher than ourself as busy.
7395 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7399 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7406 #ifdef CONFIG_NUMA_BALANCING
7407 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7409 if (sgs->sum_nr_running > sgs->nr_numa_running)
7411 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7416 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7418 if (rq->nr_running > rq->nr_numa_running)
7420 if (rq->nr_running > rq->nr_preferred_running)
7425 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7430 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7434 #endif /* CONFIG_NUMA_BALANCING */
7437 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7438 * @env: The load balancing environment.
7439 * @sds: variable to hold the statistics for this sched_domain.
7441 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7443 struct sched_domain *child = env->sd->child;
7444 struct sched_group *sg = env->sd->groups;
7445 struct sg_lb_stats tmp_sgs;
7446 int load_idx, prefer_sibling = 0;
7447 bool overload = false, overutilized = false;
7449 if (child && child->flags & SD_PREFER_SIBLING)
7452 load_idx = get_sd_load_idx(env->sd, env->idle);
7455 struct sg_lb_stats *sgs = &tmp_sgs;
7458 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7461 sgs = &sds->local_stat;
7463 if (env->idle != CPU_NEWLY_IDLE ||
7464 time_after_eq(jiffies, sg->sgc->next_update))
7465 update_group_capacity(env->sd, env->dst_cpu);
7468 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7469 &overload, &overutilized);
7475 * In case the child domain prefers tasks go to siblings
7476 * first, lower the sg capacity so that we'll try
7477 * and move all the excess tasks away. We lower the capacity
7478 * of a group only if the local group has the capacity to fit
7479 * these excess tasks. The extra check prevents the case where
7480 * you always pull from the heaviest group when it is already
7481 * under-utilized (possible with a large weight task outweighs
7482 * the tasks on the system).
7484 if (prefer_sibling && sds->local &&
7485 group_has_capacity(env, &sds->local_stat) &&
7486 (sgs->sum_nr_running > 1)) {
7487 sgs->group_no_capacity = 1;
7488 sgs->group_type = group_classify(sg, sgs);
7492 * Ignore task groups with misfit tasks if local group has no
7493 * capacity or if per-cpu capacity isn't higher.
7495 if (sgs->group_type == group_misfit_task &&
7496 (!group_has_capacity(env, &sds->local_stat) ||
7497 !group_smaller_cpu_capacity(sg, sds->local)))
7498 sgs->group_type = group_other;
7500 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7502 sds->busiest_stat = *sgs;
7506 /* Now, start updating sd_lb_stats */
7507 sds->total_load += sgs->group_load;
7508 sds->total_capacity += sgs->group_capacity;
7511 } while (sg != env->sd->groups);
7513 if (env->sd->flags & SD_NUMA)
7514 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7516 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7518 if (!env->sd->parent) {
7519 /* update overload indicator if we are at root domain */
7520 if (env->dst_rq->rd->overload != overload)
7521 env->dst_rq->rd->overload = overload;
7523 /* Update over-utilization (tipping point, U >= 0) indicator */
7524 if (env->dst_rq->rd->overutilized != overutilized)
7525 env->dst_rq->rd->overutilized = overutilized;
7527 if (!env->dst_rq->rd->overutilized && overutilized)
7528 env->dst_rq->rd->overutilized = true;
7533 * check_asym_packing - Check to see if the group is packed into the
7536 * This is primarily intended to used at the sibling level. Some
7537 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7538 * case of POWER7, it can move to lower SMT modes only when higher
7539 * threads are idle. When in lower SMT modes, the threads will
7540 * perform better since they share less core resources. Hence when we
7541 * have idle threads, we want them to be the higher ones.
7543 * This packing function is run on idle threads. It checks to see if
7544 * the busiest CPU in this domain (core in the P7 case) has a higher
7545 * CPU number than the packing function is being run on. Here we are
7546 * assuming lower CPU number will be equivalent to lower a SMT thread
7549 * Return: 1 when packing is required and a task should be moved to
7550 * this CPU. The amount of the imbalance is returned in *imbalance.
7552 * @env: The load balancing environment.
7553 * @sds: Statistics of the sched_domain which is to be packed
7555 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7559 if (!(env->sd->flags & SD_ASYM_PACKING))
7565 busiest_cpu = group_first_cpu(sds->busiest);
7566 if (env->dst_cpu > busiest_cpu)
7569 env->imbalance = DIV_ROUND_CLOSEST(
7570 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7571 SCHED_CAPACITY_SCALE);
7577 * fix_small_imbalance - Calculate the minor imbalance that exists
7578 * amongst the groups of a sched_domain, during
7580 * @env: The load balancing environment.
7581 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7584 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7586 unsigned long tmp, capa_now = 0, capa_move = 0;
7587 unsigned int imbn = 2;
7588 unsigned long scaled_busy_load_per_task;
7589 struct sg_lb_stats *local, *busiest;
7591 local = &sds->local_stat;
7592 busiest = &sds->busiest_stat;
7594 if (!local->sum_nr_running)
7595 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7596 else if (busiest->load_per_task > local->load_per_task)
7599 scaled_busy_load_per_task =
7600 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7601 busiest->group_capacity;
7603 if (busiest->avg_load + scaled_busy_load_per_task >=
7604 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7605 env->imbalance = busiest->load_per_task;
7610 * OK, we don't have enough imbalance to justify moving tasks,
7611 * however we may be able to increase total CPU capacity used by
7615 capa_now += busiest->group_capacity *
7616 min(busiest->load_per_task, busiest->avg_load);
7617 capa_now += local->group_capacity *
7618 min(local->load_per_task, local->avg_load);
7619 capa_now /= SCHED_CAPACITY_SCALE;
7621 /* Amount of load we'd subtract */
7622 if (busiest->avg_load > scaled_busy_load_per_task) {
7623 capa_move += busiest->group_capacity *
7624 min(busiest->load_per_task,
7625 busiest->avg_load - scaled_busy_load_per_task);
7628 /* Amount of load we'd add */
7629 if (busiest->avg_load * busiest->group_capacity <
7630 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7631 tmp = (busiest->avg_load * busiest->group_capacity) /
7632 local->group_capacity;
7634 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7635 local->group_capacity;
7637 capa_move += local->group_capacity *
7638 min(local->load_per_task, local->avg_load + tmp);
7639 capa_move /= SCHED_CAPACITY_SCALE;
7641 /* Move if we gain throughput */
7642 if (capa_move > capa_now)
7643 env->imbalance = busiest->load_per_task;
7647 * calculate_imbalance - Calculate the amount of imbalance present within the
7648 * groups of a given sched_domain during load balance.
7649 * @env: load balance environment
7650 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7652 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7654 unsigned long max_pull, load_above_capacity = ~0UL;
7655 struct sg_lb_stats *local, *busiest;
7657 local = &sds->local_stat;
7658 busiest = &sds->busiest_stat;
7660 if (busiest->group_type == group_imbalanced) {
7662 * In the group_imb case we cannot rely on group-wide averages
7663 * to ensure cpu-load equilibrium, look at wider averages. XXX
7665 busiest->load_per_task =
7666 min(busiest->load_per_task, sds->avg_load);
7670 * In the presence of smp nice balancing, certain scenarios can have
7671 * max load less than avg load(as we skip the groups at or below
7672 * its cpu_capacity, while calculating max_load..)
7674 if (busiest->avg_load <= sds->avg_load ||
7675 local->avg_load >= sds->avg_load) {
7676 /* Misfitting tasks should be migrated in any case */
7677 if (busiest->group_type == group_misfit_task) {
7678 env->imbalance = busiest->group_misfit_task;
7683 * Busiest group is overloaded, local is not, use the spare
7684 * cycles to maximize throughput
7686 if (busiest->group_type == group_overloaded &&
7687 local->group_type <= group_misfit_task) {
7688 env->imbalance = busiest->load_per_task;
7693 return fix_small_imbalance(env, sds);
7697 * If there aren't any idle cpus, avoid creating some.
7699 if (busiest->group_type == group_overloaded &&
7700 local->group_type == group_overloaded) {
7701 load_above_capacity = busiest->sum_nr_running *
7703 if (load_above_capacity > busiest->group_capacity)
7704 load_above_capacity -= busiest->group_capacity;
7706 load_above_capacity = ~0UL;
7710 * We're trying to get all the cpus to the average_load, so we don't
7711 * want to push ourselves above the average load, nor do we wish to
7712 * reduce the max loaded cpu below the average load. At the same time,
7713 * we also don't want to reduce the group load below the group capacity
7714 * (so that we can implement power-savings policies etc). Thus we look
7715 * for the minimum possible imbalance.
7717 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7719 /* How much load to actually move to equalise the imbalance */
7720 env->imbalance = min(
7721 max_pull * busiest->group_capacity,
7722 (sds->avg_load - local->avg_load) * local->group_capacity
7723 ) / SCHED_CAPACITY_SCALE;
7725 /* Boost imbalance to allow misfit task to be balanced. */
7726 if (busiest->group_type == group_misfit_task)
7727 env->imbalance = max_t(long, env->imbalance,
7728 busiest->group_misfit_task);
7731 * if *imbalance is less than the average load per runnable task
7732 * there is no guarantee that any tasks will be moved so we'll have
7733 * a think about bumping its value to force at least one task to be
7736 if (env->imbalance < busiest->load_per_task)
7737 return fix_small_imbalance(env, sds);
7740 /******* find_busiest_group() helpers end here *********************/
7743 * find_busiest_group - Returns the busiest group within the sched_domain
7744 * if there is an imbalance. If there isn't an imbalance, and
7745 * the user has opted for power-savings, it returns a group whose
7746 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7747 * such a group exists.
7749 * Also calculates the amount of weighted load which should be moved
7750 * to restore balance.
7752 * @env: The load balancing environment.
7754 * Return: - The busiest group if imbalance exists.
7755 * - If no imbalance and user has opted for power-savings balance,
7756 * return the least loaded group whose CPUs can be
7757 * put to idle by rebalancing its tasks onto our group.
7759 static struct sched_group *find_busiest_group(struct lb_env *env)
7761 struct sg_lb_stats *local, *busiest;
7762 struct sd_lb_stats sds;
7764 init_sd_lb_stats(&sds);
7767 * Compute the various statistics relavent for load balancing at
7770 update_sd_lb_stats(env, &sds);
7772 if (energy_aware() && !env->dst_rq->rd->overutilized)
7775 local = &sds.local_stat;
7776 busiest = &sds.busiest_stat;
7778 /* ASYM feature bypasses nice load balance check */
7779 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7780 check_asym_packing(env, &sds))
7783 /* There is no busy sibling group to pull tasks from */
7784 if (!sds.busiest || busiest->sum_nr_running == 0)
7787 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7788 / sds.total_capacity;
7791 * If the busiest group is imbalanced the below checks don't
7792 * work because they assume all things are equal, which typically
7793 * isn't true due to cpus_allowed constraints and the like.
7795 if (busiest->group_type == group_imbalanced)
7798 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7799 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7800 busiest->group_no_capacity)
7803 /* Misfitting tasks should be dealt with regardless of the avg load */
7804 if (busiest->group_type == group_misfit_task) {
7809 * If the local group is busier than the selected busiest group
7810 * don't try and pull any tasks.
7812 if (local->avg_load >= busiest->avg_load)
7816 * Don't pull any tasks if this group is already above the domain
7819 if (local->avg_load >= sds.avg_load)
7822 if (env->idle == CPU_IDLE) {
7824 * This cpu is idle. If the busiest group is not overloaded
7825 * and there is no imbalance between this and busiest group
7826 * wrt idle cpus, it is balanced. The imbalance becomes
7827 * significant if the diff is greater than 1 otherwise we
7828 * might end up to just move the imbalance on another group
7830 if ((busiest->group_type != group_overloaded) &&
7831 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7832 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7836 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7837 * imbalance_pct to be conservative.
7839 if (100 * busiest->avg_load <=
7840 env->sd->imbalance_pct * local->avg_load)
7845 env->busiest_group_type = busiest->group_type;
7846 /* Looks like there is an imbalance. Compute it */
7847 calculate_imbalance(env, &sds);
7856 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7858 static struct rq *find_busiest_queue(struct lb_env *env,
7859 struct sched_group *group)
7861 struct rq *busiest = NULL, *rq;
7862 unsigned long busiest_load = 0, busiest_capacity = 1;
7865 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7866 unsigned long capacity, wl;
7870 rt = fbq_classify_rq(rq);
7873 * We classify groups/runqueues into three groups:
7874 * - regular: there are !numa tasks
7875 * - remote: there are numa tasks that run on the 'wrong' node
7876 * - all: there is no distinction
7878 * In order to avoid migrating ideally placed numa tasks,
7879 * ignore those when there's better options.
7881 * If we ignore the actual busiest queue to migrate another
7882 * task, the next balance pass can still reduce the busiest
7883 * queue by moving tasks around inside the node.
7885 * If we cannot move enough load due to this classification
7886 * the next pass will adjust the group classification and
7887 * allow migration of more tasks.
7889 * Both cases only affect the total convergence complexity.
7891 if (rt > env->fbq_type)
7894 capacity = capacity_of(i);
7896 wl = weighted_cpuload(i);
7899 * When comparing with imbalance, use weighted_cpuload()
7900 * which is not scaled with the cpu capacity.
7903 if (rq->nr_running == 1 && wl > env->imbalance &&
7904 !check_cpu_capacity(rq, env->sd) &&
7905 env->busiest_group_type != group_misfit_task)
7909 * For the load comparisons with the other cpu's, consider
7910 * the weighted_cpuload() scaled with the cpu capacity, so
7911 * that the load can be moved away from the cpu that is
7912 * potentially running at a lower capacity.
7914 * Thus we're looking for max(wl_i / capacity_i), crosswise
7915 * multiplication to rid ourselves of the division works out
7916 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7917 * our previous maximum.
7919 if (wl * busiest_capacity > busiest_load * capacity) {
7921 busiest_capacity = capacity;
7930 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7931 * so long as it is large enough.
7933 #define MAX_PINNED_INTERVAL 512
7935 /* Working cpumask for load_balance and load_balance_newidle. */
7936 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7938 static int need_active_balance(struct lb_env *env)
7940 struct sched_domain *sd = env->sd;
7942 if (env->idle == CPU_NEWLY_IDLE) {
7945 * ASYM_PACKING needs to force migrate tasks from busy but
7946 * higher numbered CPUs in order to pack all tasks in the
7947 * lowest numbered CPUs.
7949 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7954 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7955 * It's worth migrating the task if the src_cpu's capacity is reduced
7956 * because of other sched_class or IRQs if more capacity stays
7957 * available on dst_cpu.
7959 if ((env->idle != CPU_NOT_IDLE) &&
7960 (env->src_rq->cfs.h_nr_running == 1)) {
7961 if ((check_cpu_capacity(env->src_rq, sd)) &&
7962 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7966 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
7967 env->src_rq->cfs.h_nr_running == 1 &&
7968 cpu_overutilized(env->src_cpu) &&
7969 !cpu_overutilized(env->dst_cpu)) {
7973 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7976 static int active_load_balance_cpu_stop(void *data);
7978 static int should_we_balance(struct lb_env *env)
7980 struct sched_group *sg = env->sd->groups;
7981 struct cpumask *sg_cpus, *sg_mask;
7982 int cpu, balance_cpu = -1;
7985 * In the newly idle case, we will allow all the cpu's
7986 * to do the newly idle load balance.
7988 if (env->idle == CPU_NEWLY_IDLE)
7991 sg_cpus = sched_group_cpus(sg);
7992 sg_mask = sched_group_mask(sg);
7993 /* Try to find first idle cpu */
7994 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7995 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8002 if (balance_cpu == -1)
8003 balance_cpu = group_balance_cpu(sg);
8006 * First idle cpu or the first cpu(busiest) in this sched group
8007 * is eligible for doing load balancing at this and above domains.
8009 return balance_cpu == env->dst_cpu;
8013 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8014 * tasks if there is an imbalance.
8016 static int load_balance(int this_cpu, struct rq *this_rq,
8017 struct sched_domain *sd, enum cpu_idle_type idle,
8018 int *continue_balancing)
8020 int ld_moved, cur_ld_moved, active_balance = 0;
8021 struct sched_domain *sd_parent = sd->parent;
8022 struct sched_group *group;
8024 unsigned long flags;
8025 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8027 struct lb_env env = {
8029 .dst_cpu = this_cpu,
8031 .dst_grpmask = sched_group_cpus(sd->groups),
8033 .loop_break = sched_nr_migrate_break,
8036 .tasks = LIST_HEAD_INIT(env.tasks),
8040 * For NEWLY_IDLE load_balancing, we don't need to consider
8041 * other cpus in our group
8043 if (idle == CPU_NEWLY_IDLE)
8044 env.dst_grpmask = NULL;
8046 cpumask_copy(cpus, cpu_active_mask);
8048 schedstat_inc(sd, lb_count[idle]);
8051 if (!should_we_balance(&env)) {
8052 *continue_balancing = 0;
8056 group = find_busiest_group(&env);
8058 schedstat_inc(sd, lb_nobusyg[idle]);
8062 busiest = find_busiest_queue(&env, group);
8064 schedstat_inc(sd, lb_nobusyq[idle]);
8068 BUG_ON(busiest == env.dst_rq);
8070 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8072 env.src_cpu = busiest->cpu;
8073 env.src_rq = busiest;
8076 if (busiest->nr_running > 1) {
8078 * Attempt to move tasks. If find_busiest_group has found
8079 * an imbalance but busiest->nr_running <= 1, the group is
8080 * still unbalanced. ld_moved simply stays zero, so it is
8081 * correctly treated as an imbalance.
8083 env.flags |= LBF_ALL_PINNED;
8084 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8087 raw_spin_lock_irqsave(&busiest->lock, flags);
8090 * cur_ld_moved - load moved in current iteration
8091 * ld_moved - cumulative load moved across iterations
8093 cur_ld_moved = detach_tasks(&env);
8095 * We want to potentially lower env.src_cpu's OPP.
8098 update_capacity_of(env.src_cpu);
8101 * We've detached some tasks from busiest_rq. Every
8102 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8103 * unlock busiest->lock, and we are able to be sure
8104 * that nobody can manipulate the tasks in parallel.
8105 * See task_rq_lock() family for the details.
8108 raw_spin_unlock(&busiest->lock);
8112 ld_moved += cur_ld_moved;
8115 local_irq_restore(flags);
8117 if (env.flags & LBF_NEED_BREAK) {
8118 env.flags &= ~LBF_NEED_BREAK;
8123 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8124 * us and move them to an alternate dst_cpu in our sched_group
8125 * where they can run. The upper limit on how many times we
8126 * iterate on same src_cpu is dependent on number of cpus in our
8129 * This changes load balance semantics a bit on who can move
8130 * load to a given_cpu. In addition to the given_cpu itself
8131 * (or a ilb_cpu acting on its behalf where given_cpu is
8132 * nohz-idle), we now have balance_cpu in a position to move
8133 * load to given_cpu. In rare situations, this may cause
8134 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8135 * _independently_ and at _same_ time to move some load to
8136 * given_cpu) causing exceess load to be moved to given_cpu.
8137 * This however should not happen so much in practice and
8138 * moreover subsequent load balance cycles should correct the
8139 * excess load moved.
8141 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8143 /* Prevent to re-select dst_cpu via env's cpus */
8144 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8146 env.dst_rq = cpu_rq(env.new_dst_cpu);
8147 env.dst_cpu = env.new_dst_cpu;
8148 env.flags &= ~LBF_DST_PINNED;
8150 env.loop_break = sched_nr_migrate_break;
8153 * Go back to "more_balance" rather than "redo" since we
8154 * need to continue with same src_cpu.
8160 * We failed to reach balance because of affinity.
8163 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8165 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8166 *group_imbalance = 1;
8169 /* All tasks on this runqueue were pinned by CPU affinity */
8170 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8171 cpumask_clear_cpu(cpu_of(busiest), cpus);
8172 if (!cpumask_empty(cpus)) {
8174 env.loop_break = sched_nr_migrate_break;
8177 goto out_all_pinned;
8182 schedstat_inc(sd, lb_failed[idle]);
8184 * Increment the failure counter only on periodic balance.
8185 * We do not want newidle balance, which can be very
8186 * frequent, pollute the failure counter causing
8187 * excessive cache_hot migrations and active balances.
8189 if (idle != CPU_NEWLY_IDLE)
8190 if (env.src_grp_nr_running > 1)
8191 sd->nr_balance_failed++;
8193 if (need_active_balance(&env)) {
8194 raw_spin_lock_irqsave(&busiest->lock, flags);
8196 /* don't kick the active_load_balance_cpu_stop,
8197 * if the curr task on busiest cpu can't be
8200 if (!cpumask_test_cpu(this_cpu,
8201 tsk_cpus_allowed(busiest->curr))) {
8202 raw_spin_unlock_irqrestore(&busiest->lock,
8204 env.flags |= LBF_ALL_PINNED;
8205 goto out_one_pinned;
8209 * ->active_balance synchronizes accesses to
8210 * ->active_balance_work. Once set, it's cleared
8211 * only after active load balance is finished.
8213 if (!busiest->active_balance) {
8214 busiest->active_balance = 1;
8215 busiest->push_cpu = this_cpu;
8218 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8220 if (active_balance) {
8221 stop_one_cpu_nowait(cpu_of(busiest),
8222 active_load_balance_cpu_stop, busiest,
8223 &busiest->active_balance_work);
8227 * We've kicked active balancing, reset the failure
8230 sd->nr_balance_failed = sd->cache_nice_tries+1;
8233 sd->nr_balance_failed = 0;
8235 if (likely(!active_balance)) {
8236 /* We were unbalanced, so reset the balancing interval */
8237 sd->balance_interval = sd->min_interval;
8240 * If we've begun active balancing, start to back off. This
8241 * case may not be covered by the all_pinned logic if there
8242 * is only 1 task on the busy runqueue (because we don't call
8245 if (sd->balance_interval < sd->max_interval)
8246 sd->balance_interval *= 2;
8253 * We reach balance although we may have faced some affinity
8254 * constraints. Clear the imbalance flag if it was set.
8257 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8259 if (*group_imbalance)
8260 *group_imbalance = 0;
8265 * We reach balance because all tasks are pinned at this level so
8266 * we can't migrate them. Let the imbalance flag set so parent level
8267 * can try to migrate them.
8269 schedstat_inc(sd, lb_balanced[idle]);
8271 sd->nr_balance_failed = 0;
8274 /* tune up the balancing interval */
8275 if (((env.flags & LBF_ALL_PINNED) &&
8276 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8277 (sd->balance_interval < sd->max_interval))
8278 sd->balance_interval *= 2;
8285 static inline unsigned long
8286 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8288 unsigned long interval = sd->balance_interval;
8291 interval *= sd->busy_factor;
8293 /* scale ms to jiffies */
8294 interval = msecs_to_jiffies(interval);
8295 interval = clamp(interval, 1UL, max_load_balance_interval);
8301 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8303 unsigned long interval, next;
8305 interval = get_sd_balance_interval(sd, cpu_busy);
8306 next = sd->last_balance + interval;
8308 if (time_after(*next_balance, next))
8309 *next_balance = next;
8313 * idle_balance is called by schedule() if this_cpu is about to become
8314 * idle. Attempts to pull tasks from other CPUs.
8316 static int idle_balance(struct rq *this_rq)
8318 unsigned long next_balance = jiffies + HZ;
8319 int this_cpu = this_rq->cpu;
8320 struct sched_domain *sd;
8321 int pulled_task = 0;
8324 idle_enter_fair(this_rq);
8327 * We must set idle_stamp _before_ calling idle_balance(), such that we
8328 * measure the duration of idle_balance() as idle time.
8330 this_rq->idle_stamp = rq_clock(this_rq);
8332 if (!energy_aware() &&
8333 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8334 !this_rq->rd->overload)) {
8336 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8338 update_next_balance(sd, 0, &next_balance);
8344 raw_spin_unlock(&this_rq->lock);
8346 update_blocked_averages(this_cpu);
8348 for_each_domain(this_cpu, sd) {
8349 int continue_balancing = 1;
8350 u64 t0, domain_cost;
8352 if (!(sd->flags & SD_LOAD_BALANCE))
8355 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8356 update_next_balance(sd, 0, &next_balance);
8360 if (sd->flags & SD_BALANCE_NEWIDLE) {
8361 t0 = sched_clock_cpu(this_cpu);
8363 pulled_task = load_balance(this_cpu, this_rq,
8365 &continue_balancing);
8367 domain_cost = sched_clock_cpu(this_cpu) - t0;
8368 if (domain_cost > sd->max_newidle_lb_cost)
8369 sd->max_newidle_lb_cost = domain_cost;
8371 curr_cost += domain_cost;
8374 update_next_balance(sd, 0, &next_balance);
8377 * Stop searching for tasks to pull if there are
8378 * now runnable tasks on this rq.
8380 if (pulled_task || this_rq->nr_running > 0)
8385 raw_spin_lock(&this_rq->lock);
8387 if (curr_cost > this_rq->max_idle_balance_cost)
8388 this_rq->max_idle_balance_cost = curr_cost;
8391 * While browsing the domains, we released the rq lock, a task could
8392 * have been enqueued in the meantime. Since we're not going idle,
8393 * pretend we pulled a task.
8395 if (this_rq->cfs.h_nr_running && !pulled_task)
8399 /* Move the next balance forward */
8400 if (time_after(this_rq->next_balance, next_balance))
8401 this_rq->next_balance = next_balance;
8403 /* Is there a task of a high priority class? */
8404 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8408 idle_exit_fair(this_rq);
8409 this_rq->idle_stamp = 0;
8416 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8417 * running tasks off the busiest CPU onto idle CPUs. It requires at
8418 * least 1 task to be running on each physical CPU where possible, and
8419 * avoids physical / logical imbalances.
8421 static int active_load_balance_cpu_stop(void *data)
8423 struct rq *busiest_rq = data;
8424 int busiest_cpu = cpu_of(busiest_rq);
8425 int target_cpu = busiest_rq->push_cpu;
8426 struct rq *target_rq = cpu_rq(target_cpu);
8427 struct sched_domain *sd;
8428 struct task_struct *p = NULL;
8430 raw_spin_lock_irq(&busiest_rq->lock);
8432 /* make sure the requested cpu hasn't gone down in the meantime */
8433 if (unlikely(busiest_cpu != smp_processor_id() ||
8434 !busiest_rq->active_balance))
8437 /* Is there any task to move? */
8438 if (busiest_rq->nr_running <= 1)
8442 * This condition is "impossible", if it occurs
8443 * we need to fix it. Originally reported by
8444 * Bjorn Helgaas on a 128-cpu setup.
8446 BUG_ON(busiest_rq == target_rq);
8448 /* Search for an sd spanning us and the target CPU. */
8450 for_each_domain(target_cpu, sd) {
8451 if ((sd->flags & SD_LOAD_BALANCE) &&
8452 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8457 struct lb_env env = {
8459 .dst_cpu = target_cpu,
8460 .dst_rq = target_rq,
8461 .src_cpu = busiest_rq->cpu,
8462 .src_rq = busiest_rq,
8466 schedstat_inc(sd, alb_count);
8468 p = detach_one_task(&env);
8470 schedstat_inc(sd, alb_pushed);
8472 * We want to potentially lower env.src_cpu's OPP.
8474 update_capacity_of(env.src_cpu);
8477 schedstat_inc(sd, alb_failed);
8481 busiest_rq->active_balance = 0;
8482 raw_spin_unlock(&busiest_rq->lock);
8485 attach_one_task(target_rq, p);
8492 static inline int on_null_domain(struct rq *rq)
8494 return unlikely(!rcu_dereference_sched(rq->sd));
8497 #ifdef CONFIG_NO_HZ_COMMON
8499 * idle load balancing details
8500 * - When one of the busy CPUs notice that there may be an idle rebalancing
8501 * needed, they will kick the idle load balancer, which then does idle
8502 * load balancing for all the idle CPUs.
8505 cpumask_var_t idle_cpus_mask;
8507 unsigned long next_balance; /* in jiffy units */
8508 } nohz ____cacheline_aligned;
8510 static inline int find_new_ilb(void)
8512 int ilb = cpumask_first(nohz.idle_cpus_mask);
8514 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8521 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8522 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8523 * CPU (if there is one).
8525 static void nohz_balancer_kick(void)
8529 nohz.next_balance++;
8531 ilb_cpu = find_new_ilb();
8533 if (ilb_cpu >= nr_cpu_ids)
8536 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8539 * Use smp_send_reschedule() instead of resched_cpu().
8540 * This way we generate a sched IPI on the target cpu which
8541 * is idle. And the softirq performing nohz idle load balance
8542 * will be run before returning from the IPI.
8544 smp_send_reschedule(ilb_cpu);
8548 static inline void nohz_balance_exit_idle(int cpu)
8550 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8552 * Completely isolated CPUs don't ever set, so we must test.
8554 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8555 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8556 atomic_dec(&nohz.nr_cpus);
8558 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8562 static inline void set_cpu_sd_state_busy(void)
8564 struct sched_domain *sd;
8565 int cpu = smp_processor_id();
8568 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8570 if (!sd || !sd->nohz_idle)
8574 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8579 void set_cpu_sd_state_idle(void)
8581 struct sched_domain *sd;
8582 int cpu = smp_processor_id();
8585 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8587 if (!sd || sd->nohz_idle)
8591 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8597 * This routine will record that the cpu is going idle with tick stopped.
8598 * This info will be used in performing idle load balancing in the future.
8600 void nohz_balance_enter_idle(int cpu)
8603 * If this cpu is going down, then nothing needs to be done.
8605 if (!cpu_active(cpu))
8608 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8612 * If we're a completely isolated CPU, we don't play.
8614 if (on_null_domain(cpu_rq(cpu)))
8617 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8618 atomic_inc(&nohz.nr_cpus);
8619 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8622 static int sched_ilb_notifier(struct notifier_block *nfb,
8623 unsigned long action, void *hcpu)
8625 switch (action & ~CPU_TASKS_FROZEN) {
8627 nohz_balance_exit_idle(smp_processor_id());
8635 static DEFINE_SPINLOCK(balancing);
8638 * Scale the max load_balance interval with the number of CPUs in the system.
8639 * This trades load-balance latency on larger machines for less cross talk.
8641 void update_max_interval(void)
8643 max_load_balance_interval = HZ*num_online_cpus()/10;
8647 * It checks each scheduling domain to see if it is due to be balanced,
8648 * and initiates a balancing operation if so.
8650 * Balancing parameters are set up in init_sched_domains.
8652 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8654 int continue_balancing = 1;
8656 unsigned long interval;
8657 struct sched_domain *sd;
8658 /* Earliest time when we have to do rebalance again */
8659 unsigned long next_balance = jiffies + 60*HZ;
8660 int update_next_balance = 0;
8661 int need_serialize, need_decay = 0;
8664 update_blocked_averages(cpu);
8667 for_each_domain(cpu, sd) {
8669 * Decay the newidle max times here because this is a regular
8670 * visit to all the domains. Decay ~1% per second.
8672 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8673 sd->max_newidle_lb_cost =
8674 (sd->max_newidle_lb_cost * 253) / 256;
8675 sd->next_decay_max_lb_cost = jiffies + HZ;
8678 max_cost += sd->max_newidle_lb_cost;
8680 if (!(sd->flags & SD_LOAD_BALANCE))
8684 * Stop the load balance at this level. There is another
8685 * CPU in our sched group which is doing load balancing more
8688 if (!continue_balancing) {
8694 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8696 need_serialize = sd->flags & SD_SERIALIZE;
8697 if (need_serialize) {
8698 if (!spin_trylock(&balancing))
8702 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8703 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8705 * The LBF_DST_PINNED logic could have changed
8706 * env->dst_cpu, so we can't know our idle
8707 * state even if we migrated tasks. Update it.
8709 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8711 sd->last_balance = jiffies;
8712 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8715 spin_unlock(&balancing);
8717 if (time_after(next_balance, sd->last_balance + interval)) {
8718 next_balance = sd->last_balance + interval;
8719 update_next_balance = 1;
8724 * Ensure the rq-wide value also decays but keep it at a
8725 * reasonable floor to avoid funnies with rq->avg_idle.
8727 rq->max_idle_balance_cost =
8728 max((u64)sysctl_sched_migration_cost, max_cost);
8733 * next_balance will be updated only when there is a need.
8734 * When the cpu is attached to null domain for ex, it will not be
8737 if (likely(update_next_balance)) {
8738 rq->next_balance = next_balance;
8740 #ifdef CONFIG_NO_HZ_COMMON
8742 * If this CPU has been elected to perform the nohz idle
8743 * balance. Other idle CPUs have already rebalanced with
8744 * nohz_idle_balance() and nohz.next_balance has been
8745 * updated accordingly. This CPU is now running the idle load
8746 * balance for itself and we need to update the
8747 * nohz.next_balance accordingly.
8749 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8750 nohz.next_balance = rq->next_balance;
8755 #ifdef CONFIG_NO_HZ_COMMON
8757 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8758 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8760 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8762 int this_cpu = this_rq->cpu;
8765 /* Earliest time when we have to do rebalance again */
8766 unsigned long next_balance = jiffies + 60*HZ;
8767 int update_next_balance = 0;
8769 if (idle != CPU_IDLE ||
8770 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8773 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8774 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8778 * If this cpu gets work to do, stop the load balancing
8779 * work being done for other cpus. Next load
8780 * balancing owner will pick it up.
8785 rq = cpu_rq(balance_cpu);
8788 * If time for next balance is due,
8791 if (time_after_eq(jiffies, rq->next_balance)) {
8792 raw_spin_lock_irq(&rq->lock);
8793 update_rq_clock(rq);
8794 update_idle_cpu_load(rq);
8795 raw_spin_unlock_irq(&rq->lock);
8796 rebalance_domains(rq, CPU_IDLE);
8799 if (time_after(next_balance, rq->next_balance)) {
8800 next_balance = rq->next_balance;
8801 update_next_balance = 1;
8806 * next_balance will be updated only when there is a need.
8807 * When the CPU is attached to null domain for ex, it will not be
8810 if (likely(update_next_balance))
8811 nohz.next_balance = next_balance;
8813 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8817 * Current heuristic for kicking the idle load balancer in the presence
8818 * of an idle cpu in the system.
8819 * - This rq has more than one task.
8820 * - This rq has at least one CFS task and the capacity of the CPU is
8821 * significantly reduced because of RT tasks or IRQs.
8822 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8823 * multiple busy cpu.
8824 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8825 * domain span are idle.
8827 static inline bool nohz_kick_needed(struct rq *rq)
8829 unsigned long now = jiffies;
8830 struct sched_domain *sd;
8831 struct sched_group_capacity *sgc;
8832 int nr_busy, cpu = rq->cpu;
8835 if (unlikely(rq->idle_balance))
8839 * We may be recently in ticked or tickless idle mode. At the first
8840 * busy tick after returning from idle, we will update the busy stats.
8842 set_cpu_sd_state_busy();
8843 nohz_balance_exit_idle(cpu);
8846 * None are in tickless mode and hence no need for NOHZ idle load
8849 if (likely(!atomic_read(&nohz.nr_cpus)))
8852 if (time_before(now, nohz.next_balance))
8855 if (rq->nr_running >= 2 &&
8856 (!energy_aware() || cpu_overutilized(cpu)))
8860 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8861 if (sd && !energy_aware()) {
8862 sgc = sd->groups->sgc;
8863 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8872 sd = rcu_dereference(rq->sd);
8874 if ((rq->cfs.h_nr_running >= 1) &&
8875 check_cpu_capacity(rq, sd)) {
8881 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8882 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8883 sched_domain_span(sd)) < cpu)) {
8893 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8897 * run_rebalance_domains is triggered when needed from the scheduler tick.
8898 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8900 static void run_rebalance_domains(struct softirq_action *h)
8902 struct rq *this_rq = this_rq();
8903 enum cpu_idle_type idle = this_rq->idle_balance ?
8904 CPU_IDLE : CPU_NOT_IDLE;
8907 * If this cpu has a pending nohz_balance_kick, then do the
8908 * balancing on behalf of the other idle cpus whose ticks are
8909 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8910 * give the idle cpus a chance to load balance. Else we may
8911 * load balance only within the local sched_domain hierarchy
8912 * and abort nohz_idle_balance altogether if we pull some load.
8914 nohz_idle_balance(this_rq, idle);
8915 rebalance_domains(this_rq, idle);
8919 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8921 void trigger_load_balance(struct rq *rq)
8923 /* Don't need to rebalance while attached to NULL domain */
8924 if (unlikely(on_null_domain(rq)))
8927 if (time_after_eq(jiffies, rq->next_balance))
8928 raise_softirq(SCHED_SOFTIRQ);
8929 #ifdef CONFIG_NO_HZ_COMMON
8930 if (nohz_kick_needed(rq))
8931 nohz_balancer_kick();
8935 static void rq_online_fair(struct rq *rq)
8939 update_runtime_enabled(rq);
8942 static void rq_offline_fair(struct rq *rq)
8946 /* Ensure any throttled groups are reachable by pick_next_task */
8947 unthrottle_offline_cfs_rqs(rq);
8950 #endif /* CONFIG_SMP */
8953 * scheduler tick hitting a task of our scheduling class:
8955 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8957 struct cfs_rq *cfs_rq;
8958 struct sched_entity *se = &curr->se;
8960 for_each_sched_entity(se) {
8961 cfs_rq = cfs_rq_of(se);
8962 entity_tick(cfs_rq, se, queued);
8965 if (static_branch_unlikely(&sched_numa_balancing))
8966 task_tick_numa(rq, curr);
8969 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr)))
8970 rq->rd->overutilized = true;
8972 rq->misfit_task = !task_fits_max(curr, rq->cpu);
8978 * called on fork with the child task as argument from the parent's context
8979 * - child not yet on the tasklist
8980 * - preemption disabled
8982 static void task_fork_fair(struct task_struct *p)
8984 struct cfs_rq *cfs_rq;
8985 struct sched_entity *se = &p->se, *curr;
8986 int this_cpu = smp_processor_id();
8987 struct rq *rq = this_rq();
8988 unsigned long flags;
8990 raw_spin_lock_irqsave(&rq->lock, flags);
8992 update_rq_clock(rq);
8994 cfs_rq = task_cfs_rq(current);
8995 curr = cfs_rq->curr;
8998 * Not only the cpu but also the task_group of the parent might have
8999 * been changed after parent->se.parent,cfs_rq were copied to
9000 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9001 * of child point to valid ones.
9004 __set_task_cpu(p, this_cpu);
9007 update_curr(cfs_rq);
9010 se->vruntime = curr->vruntime;
9011 place_entity(cfs_rq, se, 1);
9013 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9015 * Upon rescheduling, sched_class::put_prev_task() will place
9016 * 'current' within the tree based on its new key value.
9018 swap(curr->vruntime, se->vruntime);
9022 se->vruntime -= cfs_rq->min_vruntime;
9024 raw_spin_unlock_irqrestore(&rq->lock, flags);
9028 * Priority of the task has changed. Check to see if we preempt
9032 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9034 if (!task_on_rq_queued(p))
9038 * Reschedule if we are currently running on this runqueue and
9039 * our priority decreased, or if we are not currently running on
9040 * this runqueue and our priority is higher than the current's
9042 if (rq->curr == p) {
9043 if (p->prio > oldprio)
9046 check_preempt_curr(rq, p, 0);
9049 static inline bool vruntime_normalized(struct task_struct *p)
9051 struct sched_entity *se = &p->se;
9054 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9055 * the dequeue_entity(.flags=0) will already have normalized the
9062 * When !on_rq, vruntime of the task has usually NOT been normalized.
9063 * But there are some cases where it has already been normalized:
9065 * - A forked child which is waiting for being woken up by
9066 * wake_up_new_task().
9067 * - A task which has been woken up by try_to_wake_up() and
9068 * waiting for actually being woken up by sched_ttwu_pending().
9070 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9076 static void detach_task_cfs_rq(struct task_struct *p)
9078 struct sched_entity *se = &p->se;
9079 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9081 if (!vruntime_normalized(p)) {
9083 * Fix up our vruntime so that the current sleep doesn't
9084 * cause 'unlimited' sleep bonus.
9086 place_entity(cfs_rq, se, 0);
9087 se->vruntime -= cfs_rq->min_vruntime;
9090 /* Catch up with the cfs_rq and remove our load when we leave */
9091 detach_entity_load_avg(cfs_rq, se);
9094 static void attach_task_cfs_rq(struct task_struct *p)
9096 struct sched_entity *se = &p->se;
9097 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9099 #ifdef CONFIG_FAIR_GROUP_SCHED
9101 * Since the real-depth could have been changed (only FAIR
9102 * class maintain depth value), reset depth properly.
9104 se->depth = se->parent ? se->parent->depth + 1 : 0;
9107 /* Synchronize task with its cfs_rq */
9108 attach_entity_load_avg(cfs_rq, se);
9110 if (!vruntime_normalized(p))
9111 se->vruntime += cfs_rq->min_vruntime;
9114 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9116 detach_task_cfs_rq(p);
9119 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9121 attach_task_cfs_rq(p);
9123 if (task_on_rq_queued(p)) {
9125 * We were most likely switched from sched_rt, so
9126 * kick off the schedule if running, otherwise just see
9127 * if we can still preempt the current task.
9132 check_preempt_curr(rq, p, 0);
9136 /* Account for a task changing its policy or group.
9138 * This routine is mostly called to set cfs_rq->curr field when a task
9139 * migrates between groups/classes.
9141 static void set_curr_task_fair(struct rq *rq)
9143 struct sched_entity *se = &rq->curr->se;
9145 for_each_sched_entity(se) {
9146 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9148 set_next_entity(cfs_rq, se);
9149 /* ensure bandwidth has been allocated on our new cfs_rq */
9150 account_cfs_rq_runtime(cfs_rq, 0);
9154 void init_cfs_rq(struct cfs_rq *cfs_rq)
9156 cfs_rq->tasks_timeline = RB_ROOT;
9157 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9158 #ifndef CONFIG_64BIT
9159 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9162 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9163 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9167 #ifdef CONFIG_FAIR_GROUP_SCHED
9168 static void task_move_group_fair(struct task_struct *p)
9170 detach_task_cfs_rq(p);
9171 set_task_rq(p, task_cpu(p));
9174 /* Tell se's cfs_rq has been changed -- migrated */
9175 p->se.avg.last_update_time = 0;
9177 attach_task_cfs_rq(p);
9180 void free_fair_sched_group(struct task_group *tg)
9184 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9186 for_each_possible_cpu(i) {
9188 kfree(tg->cfs_rq[i]);
9191 remove_entity_load_avg(tg->se[i]);
9200 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9202 struct cfs_rq *cfs_rq;
9203 struct sched_entity *se;
9206 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9209 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9213 tg->shares = NICE_0_LOAD;
9215 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9217 for_each_possible_cpu(i) {
9218 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9219 GFP_KERNEL, cpu_to_node(i));
9223 se = kzalloc_node(sizeof(struct sched_entity),
9224 GFP_KERNEL, cpu_to_node(i));
9228 init_cfs_rq(cfs_rq);
9229 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9230 init_entity_runnable_average(se);
9241 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9243 struct rq *rq = cpu_rq(cpu);
9244 unsigned long flags;
9247 * Only empty task groups can be destroyed; so we can speculatively
9248 * check on_list without danger of it being re-added.
9250 if (!tg->cfs_rq[cpu]->on_list)
9253 raw_spin_lock_irqsave(&rq->lock, flags);
9254 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9255 raw_spin_unlock_irqrestore(&rq->lock, flags);
9258 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9259 struct sched_entity *se, int cpu,
9260 struct sched_entity *parent)
9262 struct rq *rq = cpu_rq(cpu);
9266 init_cfs_rq_runtime(cfs_rq);
9268 tg->cfs_rq[cpu] = cfs_rq;
9271 /* se could be NULL for root_task_group */
9276 se->cfs_rq = &rq->cfs;
9279 se->cfs_rq = parent->my_q;
9280 se->depth = parent->depth + 1;
9284 /* guarantee group entities always have weight */
9285 update_load_set(&se->load, NICE_0_LOAD);
9286 se->parent = parent;
9289 static DEFINE_MUTEX(shares_mutex);
9291 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9294 unsigned long flags;
9297 * We can't change the weight of the root cgroup.
9302 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9304 mutex_lock(&shares_mutex);
9305 if (tg->shares == shares)
9308 tg->shares = shares;
9309 for_each_possible_cpu(i) {
9310 struct rq *rq = cpu_rq(i);
9311 struct sched_entity *se;
9314 /* Propagate contribution to hierarchy */
9315 raw_spin_lock_irqsave(&rq->lock, flags);
9317 /* Possible calls to update_curr() need rq clock */
9318 update_rq_clock(rq);
9319 for_each_sched_entity(se)
9320 update_cfs_shares(group_cfs_rq(se));
9321 raw_spin_unlock_irqrestore(&rq->lock, flags);
9325 mutex_unlock(&shares_mutex);
9328 #else /* CONFIG_FAIR_GROUP_SCHED */
9330 void free_fair_sched_group(struct task_group *tg) { }
9332 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9337 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9339 #endif /* CONFIG_FAIR_GROUP_SCHED */
9342 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9344 struct sched_entity *se = &task->se;
9345 unsigned int rr_interval = 0;
9348 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9351 if (rq->cfs.load.weight)
9352 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9358 * All the scheduling class methods:
9360 const struct sched_class fair_sched_class = {
9361 .next = &idle_sched_class,
9362 .enqueue_task = enqueue_task_fair,
9363 .dequeue_task = dequeue_task_fair,
9364 .yield_task = yield_task_fair,
9365 .yield_to_task = yield_to_task_fair,
9367 .check_preempt_curr = check_preempt_wakeup,
9369 .pick_next_task = pick_next_task_fair,
9370 .put_prev_task = put_prev_task_fair,
9373 .select_task_rq = select_task_rq_fair,
9374 .migrate_task_rq = migrate_task_rq_fair,
9376 .rq_online = rq_online_fair,
9377 .rq_offline = rq_offline_fair,
9379 .task_waking = task_waking_fair,
9380 .task_dead = task_dead_fair,
9381 .set_cpus_allowed = set_cpus_allowed_common,
9384 .set_curr_task = set_curr_task_fair,
9385 .task_tick = task_tick_fair,
9386 .task_fork = task_fork_fair,
9388 .prio_changed = prio_changed_fair,
9389 .switched_from = switched_from_fair,
9390 .switched_to = switched_to_fair,
9392 .get_rr_interval = get_rr_interval_fair,
9394 .update_curr = update_curr_fair,
9396 #ifdef CONFIG_FAIR_GROUP_SCHED
9397 .task_move_group = task_move_group_fair,
9401 #ifdef CONFIG_SCHED_DEBUG
9402 void print_cfs_stats(struct seq_file *m, int cpu)
9404 struct cfs_rq *cfs_rq;
9407 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9408 print_cfs_rq(m, cpu, cfs_rq);
9412 #ifdef CONFIG_NUMA_BALANCING
9413 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9416 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9418 for_each_online_node(node) {
9419 if (p->numa_faults) {
9420 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9421 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9423 if (p->numa_group) {
9424 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9425 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9427 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9430 #endif /* CONFIG_NUMA_BALANCING */
9431 #endif /* CONFIG_SCHED_DEBUG */
9433 __init void init_sched_fair_class(void)
9436 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9438 #ifdef CONFIG_NO_HZ_COMMON
9439 nohz.next_balance = jiffies;
9440 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9441 cpu_notifier(sched_ilb_notifier, 0);