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;
4263 trace_sched_overutilized(true);
4267 * We want to potentially trigger a freq switch
4268 * request only for tasks that are waking up; this is
4269 * because we get here also during load balancing, but
4270 * in these cases it seems wise to trigger as single
4271 * request after load balancing is done.
4273 if (task_new || task_wakeup)
4274 update_capacity_of(cpu_of(rq));
4277 /* Update SchedTune accouting */
4278 schedtune_enqueue_task(p, cpu_of(rq));
4280 #endif /* CONFIG_SMP */
4284 static void set_next_buddy(struct sched_entity *se);
4287 * The dequeue_task method is called before nr_running is
4288 * decreased. We remove the task from the rbtree and
4289 * update the fair scheduling stats:
4291 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4293 struct cfs_rq *cfs_rq;
4294 struct sched_entity *se = &p->se;
4295 int task_sleep = flags & DEQUEUE_SLEEP;
4297 for_each_sched_entity(se) {
4298 cfs_rq = cfs_rq_of(se);
4299 dequeue_entity(cfs_rq, se, flags);
4302 * end evaluation on encountering a throttled cfs_rq
4304 * note: in the case of encountering a throttled cfs_rq we will
4305 * post the final h_nr_running decrement below.
4307 if (cfs_rq_throttled(cfs_rq))
4309 cfs_rq->h_nr_running--;
4310 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4312 /* Don't dequeue parent if it has other entities besides us */
4313 if (cfs_rq->load.weight) {
4315 * Bias pick_next to pick a task from this cfs_rq, as
4316 * p is sleeping when it is within its sched_slice.
4318 if (task_sleep && parent_entity(se))
4319 set_next_buddy(parent_entity(se));
4321 /* avoid re-evaluating load for this entity */
4322 se = parent_entity(se);
4325 flags |= DEQUEUE_SLEEP;
4328 for_each_sched_entity(se) {
4329 cfs_rq = cfs_rq_of(se);
4330 cfs_rq->h_nr_running--;
4331 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4333 if (cfs_rq_throttled(cfs_rq))
4336 update_load_avg(se, 1);
4337 update_cfs_shares(cfs_rq);
4341 sub_nr_running(rq, 1);
4346 walt_dec_cumulative_runnable_avg(rq, p);
4349 * We want to potentially trigger a freq switch
4350 * request only for tasks that are going to sleep;
4351 * this is because we get here also during load
4352 * balancing, but in these cases it seems wise to
4353 * trigger as single request after load balancing is
4357 if (rq->cfs.nr_running)
4358 update_capacity_of(cpu_of(rq));
4359 else if (sched_freq())
4360 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4364 /* Update SchedTune accouting */
4365 schedtune_dequeue_task(p, cpu_of(rq));
4367 #endif /* CONFIG_SMP */
4375 * per rq 'load' arrray crap; XXX kill this.
4379 * The exact cpuload at various idx values, calculated at every tick would be
4380 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4382 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4383 * on nth tick when cpu may be busy, then we have:
4384 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4385 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4387 * decay_load_missed() below does efficient calculation of
4388 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4389 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4391 * The calculation is approximated on a 128 point scale.
4392 * degrade_zero_ticks is the number of ticks after which load at any
4393 * particular idx is approximated to be zero.
4394 * degrade_factor is a precomputed table, a row for each load idx.
4395 * Each column corresponds to degradation factor for a power of two ticks,
4396 * based on 128 point scale.
4398 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4399 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4401 * With this power of 2 load factors, we can degrade the load n times
4402 * by looking at 1 bits in n and doing as many mult/shift instead of
4403 * n mult/shifts needed by the exact degradation.
4405 #define DEGRADE_SHIFT 7
4406 static const unsigned char
4407 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4408 static const unsigned char
4409 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4410 {0, 0, 0, 0, 0, 0, 0, 0},
4411 {64, 32, 8, 0, 0, 0, 0, 0},
4412 {96, 72, 40, 12, 1, 0, 0},
4413 {112, 98, 75, 43, 15, 1, 0},
4414 {120, 112, 98, 76, 45, 16, 2} };
4417 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4418 * would be when CPU is idle and so we just decay the old load without
4419 * adding any new load.
4421 static unsigned long
4422 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4426 if (!missed_updates)
4429 if (missed_updates >= degrade_zero_ticks[idx])
4433 return load >> missed_updates;
4435 while (missed_updates) {
4436 if (missed_updates % 2)
4437 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4439 missed_updates >>= 1;
4446 * Update rq->cpu_load[] statistics. This function is usually called every
4447 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4448 * every tick. We fix it up based on jiffies.
4450 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4451 unsigned long pending_updates)
4455 this_rq->nr_load_updates++;
4457 /* Update our load: */
4458 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4459 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4460 unsigned long old_load, new_load;
4462 /* scale is effectively 1 << i now, and >> i divides by scale */
4464 old_load = this_rq->cpu_load[i];
4465 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4466 new_load = this_load;
4468 * Round up the averaging division if load is increasing. This
4469 * prevents us from getting stuck on 9 if the load is 10, for
4472 if (new_load > old_load)
4473 new_load += scale - 1;
4475 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4478 sched_avg_update(this_rq);
4481 /* Used instead of source_load when we know the type == 0 */
4482 static unsigned long weighted_cpuload(const int cpu)
4484 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4487 #ifdef CONFIG_NO_HZ_COMMON
4489 * There is no sane way to deal with nohz on smp when using jiffies because the
4490 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4491 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4493 * Therefore we cannot use the delta approach from the regular tick since that
4494 * would seriously skew the load calculation. However we'll make do for those
4495 * updates happening while idle (nohz_idle_balance) or coming out of idle
4496 * (tick_nohz_idle_exit).
4498 * This means we might still be one tick off for nohz periods.
4502 * Called from nohz_idle_balance() to update the load ratings before doing the
4505 static void update_idle_cpu_load(struct rq *this_rq)
4507 unsigned long curr_jiffies = READ_ONCE(jiffies);
4508 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4509 unsigned long pending_updates;
4512 * bail if there's load or we're actually up-to-date.
4514 if (load || curr_jiffies == this_rq->last_load_update_tick)
4517 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4518 this_rq->last_load_update_tick = curr_jiffies;
4520 __update_cpu_load(this_rq, load, pending_updates);
4524 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4526 void update_cpu_load_nohz(void)
4528 struct rq *this_rq = this_rq();
4529 unsigned long curr_jiffies = READ_ONCE(jiffies);
4530 unsigned long pending_updates;
4532 if (curr_jiffies == this_rq->last_load_update_tick)
4535 raw_spin_lock(&this_rq->lock);
4536 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4537 if (pending_updates) {
4538 this_rq->last_load_update_tick = curr_jiffies;
4540 * We were idle, this means load 0, the current load might be
4541 * !0 due to remote wakeups and the sort.
4543 __update_cpu_load(this_rq, 0, pending_updates);
4545 raw_spin_unlock(&this_rq->lock);
4547 #endif /* CONFIG_NO_HZ */
4550 * Called from scheduler_tick()
4552 void update_cpu_load_active(struct rq *this_rq)
4554 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4556 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4558 this_rq->last_load_update_tick = jiffies;
4559 __update_cpu_load(this_rq, load, 1);
4563 * Return a low guess at the load of a migration-source cpu weighted
4564 * according to the scheduling class and "nice" value.
4566 * We want to under-estimate the load of migration sources, to
4567 * balance conservatively.
4569 static unsigned long source_load(int cpu, int type)
4571 struct rq *rq = cpu_rq(cpu);
4572 unsigned long total = weighted_cpuload(cpu);
4574 if (type == 0 || !sched_feat(LB_BIAS))
4577 return min(rq->cpu_load[type-1], total);
4581 * Return a high guess at the load of a migration-target cpu weighted
4582 * according to the scheduling class and "nice" value.
4584 static unsigned long target_load(int cpu, int type)
4586 struct rq *rq = cpu_rq(cpu);
4587 unsigned long total = weighted_cpuload(cpu);
4589 if (type == 0 || !sched_feat(LB_BIAS))
4592 return max(rq->cpu_load[type-1], total);
4596 static unsigned long cpu_avg_load_per_task(int cpu)
4598 struct rq *rq = cpu_rq(cpu);
4599 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4600 unsigned long load_avg = weighted_cpuload(cpu);
4603 return load_avg / nr_running;
4608 static void record_wakee(struct task_struct *p)
4611 * Rough decay (wiping) for cost saving, don't worry
4612 * about the boundary, really active task won't care
4615 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4616 current->wakee_flips >>= 1;
4617 current->wakee_flip_decay_ts = jiffies;
4620 if (current->last_wakee != p) {
4621 current->last_wakee = p;
4622 current->wakee_flips++;
4626 static void task_waking_fair(struct task_struct *p)
4628 struct sched_entity *se = &p->se;
4629 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4632 #ifndef CONFIG_64BIT
4633 u64 min_vruntime_copy;
4636 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4638 min_vruntime = cfs_rq->min_vruntime;
4639 } while (min_vruntime != min_vruntime_copy);
4641 min_vruntime = cfs_rq->min_vruntime;
4644 se->vruntime -= min_vruntime;
4648 #ifdef CONFIG_FAIR_GROUP_SCHED
4650 * effective_load() calculates the load change as seen from the root_task_group
4652 * Adding load to a group doesn't make a group heavier, but can cause movement
4653 * of group shares between cpus. Assuming the shares were perfectly aligned one
4654 * can calculate the shift in shares.
4656 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4657 * on this @cpu and results in a total addition (subtraction) of @wg to the
4658 * total group weight.
4660 * Given a runqueue weight distribution (rw_i) we can compute a shares
4661 * distribution (s_i) using:
4663 * s_i = rw_i / \Sum rw_j (1)
4665 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4666 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4667 * shares distribution (s_i):
4669 * rw_i = { 2, 4, 1, 0 }
4670 * s_i = { 2/7, 4/7, 1/7, 0 }
4672 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4673 * task used to run on and the CPU the waker is running on), we need to
4674 * compute the effect of waking a task on either CPU and, in case of a sync
4675 * wakeup, compute the effect of the current task going to sleep.
4677 * So for a change of @wl to the local @cpu with an overall group weight change
4678 * of @wl we can compute the new shares distribution (s'_i) using:
4680 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4682 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4683 * differences in waking a task to CPU 0. The additional task changes the
4684 * weight and shares distributions like:
4686 * rw'_i = { 3, 4, 1, 0 }
4687 * s'_i = { 3/8, 4/8, 1/8, 0 }
4689 * We can then compute the difference in effective weight by using:
4691 * dw_i = S * (s'_i - s_i) (3)
4693 * Where 'S' is the group weight as seen by its parent.
4695 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4696 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4697 * 4/7) times the weight of the group.
4699 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4701 struct sched_entity *se = tg->se[cpu];
4703 if (!tg->parent) /* the trivial, non-cgroup case */
4706 for_each_sched_entity(se) {
4707 struct cfs_rq *cfs_rq = se->my_q;
4708 long W, w = cfs_rq_load_avg(cfs_rq);
4713 * W = @wg + \Sum rw_j
4715 W = wg + atomic_long_read(&tg->load_avg);
4717 /* Ensure \Sum rw_j >= rw_i */
4718 W -= cfs_rq->tg_load_avg_contrib;
4727 * wl = S * s'_i; see (2)
4730 wl = (w * (long)tg->shares) / W;
4735 * Per the above, wl is the new se->load.weight value; since
4736 * those are clipped to [MIN_SHARES, ...) do so now. See
4737 * calc_cfs_shares().
4739 if (wl < MIN_SHARES)
4743 * wl = dw_i = S * (s'_i - s_i); see (3)
4745 wl -= se->avg.load_avg;
4748 * Recursively apply this logic to all parent groups to compute
4749 * the final effective load change on the root group. Since
4750 * only the @tg group gets extra weight, all parent groups can
4751 * only redistribute existing shares. @wl is the shift in shares
4752 * resulting from this level per the above.
4761 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4769 * Returns the current capacity of cpu after applying both
4770 * cpu and freq scaling.
4772 unsigned long capacity_curr_of(int cpu)
4774 return cpu_rq(cpu)->cpu_capacity_orig *
4775 arch_scale_freq_capacity(NULL, cpu)
4776 >> SCHED_CAPACITY_SHIFT;
4779 static inline bool energy_aware(void)
4781 return sched_feat(ENERGY_AWARE);
4785 struct sched_group *sg_top;
4786 struct sched_group *sg_cap;
4793 struct task_struct *task;
4808 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4809 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4810 * energy calculations. Using the scale-invariant util returned by
4811 * cpu_util() and approximating scale-invariant util by:
4813 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4815 * the normalized util can be found using the specific capacity.
4817 * capacity = capacity_orig * curr_freq/max_freq
4819 * norm_util = running_time/time ~ util/capacity
4821 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4823 int util = __cpu_util(cpu, delta);
4825 if (util >= capacity)
4826 return SCHED_CAPACITY_SCALE;
4828 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4831 static int calc_util_delta(struct energy_env *eenv, int cpu)
4833 if (cpu == eenv->src_cpu)
4834 return -eenv->util_delta;
4835 if (cpu == eenv->dst_cpu)
4836 return eenv->util_delta;
4841 unsigned long group_max_util(struct energy_env *eenv)
4844 unsigned long max_util = 0;
4846 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4847 delta = calc_util_delta(eenv, i);
4848 max_util = max(max_util, __cpu_util(i, delta));
4855 * group_norm_util() returns the approximated group util relative to it's
4856 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4857 * energy calculations. Since task executions may or may not overlap in time in
4858 * the group the true normalized util is between max(cpu_norm_util(i)) and
4859 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4860 * latter is used as the estimate as it leads to a more pessimistic energy
4861 * estimate (more busy).
4864 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4867 unsigned long util_sum = 0;
4868 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4870 for_each_cpu(i, sched_group_cpus(sg)) {
4871 delta = calc_util_delta(eenv, i);
4872 util_sum += __cpu_norm_util(i, capacity, delta);
4875 if (util_sum > SCHED_CAPACITY_SCALE)
4876 return SCHED_CAPACITY_SCALE;
4880 static int find_new_capacity(struct energy_env *eenv,
4881 const struct sched_group_energy const *sge)
4884 unsigned long util = group_max_util(eenv);
4886 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4887 if (sge->cap_states[idx].cap >= util)
4891 eenv->cap_idx = idx;
4896 static int group_idle_state(struct sched_group *sg)
4898 int i, state = INT_MAX;
4900 /* Find the shallowest idle state in the sched group. */
4901 for_each_cpu(i, sched_group_cpus(sg))
4902 state = min(state, idle_get_state_idx(cpu_rq(i)));
4904 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4911 * sched_group_energy(): Computes the absolute energy consumption of cpus
4912 * belonging to the sched_group including shared resources shared only by
4913 * members of the group. Iterates over all cpus in the hierarchy below the
4914 * sched_group starting from the bottom working it's way up before going to
4915 * the next cpu until all cpus are covered at all levels. The current
4916 * implementation is likely to gather the same util statistics multiple times.
4917 * This can probably be done in a faster but more complex way.
4918 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4920 static int sched_group_energy(struct energy_env *eenv)
4922 struct sched_domain *sd;
4923 int cpu, total_energy = 0;
4924 struct cpumask visit_cpus;
4925 struct sched_group *sg;
4927 WARN_ON(!eenv->sg_top->sge);
4929 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4931 while (!cpumask_empty(&visit_cpus)) {
4932 struct sched_group *sg_shared_cap = NULL;
4934 cpu = cpumask_first(&visit_cpus);
4937 * Is the group utilization affected by cpus outside this
4940 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4944 * We most probably raced with hotplug; returning a
4945 * wrong energy estimation is better than entering an
4951 sg_shared_cap = sd->parent->groups;
4953 for_each_domain(cpu, sd) {
4956 /* Has this sched_domain already been visited? */
4957 if (sd->child && group_first_cpu(sg) != cpu)
4961 unsigned long group_util;
4962 int sg_busy_energy, sg_idle_energy;
4963 int cap_idx, idle_idx;
4965 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
4966 eenv->sg_cap = sg_shared_cap;
4970 cap_idx = find_new_capacity(eenv, sg->sge);
4972 if (sg->group_weight == 1) {
4973 /* Remove capacity of src CPU (before task move) */
4974 if (eenv->util_delta == 0 &&
4975 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
4976 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
4977 eenv->cap.delta -= eenv->cap.before;
4979 /* Add capacity of dst CPU (after task move) */
4980 if (eenv->util_delta != 0 &&
4981 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
4982 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
4983 eenv->cap.delta += eenv->cap.after;
4987 idle_idx = group_idle_state(sg);
4988 group_util = group_norm_util(eenv, sg);
4989 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
4990 >> SCHED_CAPACITY_SHIFT;
4991 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
4992 * sg->sge->idle_states[idle_idx].power)
4993 >> SCHED_CAPACITY_SHIFT;
4995 total_energy += sg_busy_energy + sg_idle_energy;
4998 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5000 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5003 } while (sg = sg->next, sg != sd->groups);
5006 cpumask_clear_cpu(cpu, &visit_cpus);
5010 eenv->energy = total_energy;
5014 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5016 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5020 * energy_diff(): Estimate the energy impact of changing the utilization
5021 * distribution. eenv specifies the change: utilisation amount, source, and
5022 * destination cpu. Source or destination cpu may be -1 in which case the
5023 * utilization is removed from or added to the system (e.g. task wake-up). If
5024 * both are specified, the utilization is migrated.
5026 static inline int __energy_diff(struct energy_env *eenv)
5028 struct sched_domain *sd;
5029 struct sched_group *sg;
5030 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5032 struct energy_env eenv_before = {
5034 .src_cpu = eenv->src_cpu,
5035 .dst_cpu = eenv->dst_cpu,
5036 .nrg = { 0, 0, 0, 0},
5040 if (eenv->src_cpu == eenv->dst_cpu)
5043 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5044 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5047 return 0; /* Error */
5052 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5053 eenv_before.sg_top = eenv->sg_top = sg;
5055 if (sched_group_energy(&eenv_before))
5056 return 0; /* Invalid result abort */
5057 energy_before += eenv_before.energy;
5059 /* Keep track of SRC cpu (before) capacity */
5060 eenv->cap.before = eenv_before.cap.before;
5061 eenv->cap.delta = eenv_before.cap.delta;
5063 if (sched_group_energy(eenv))
5064 return 0; /* Invalid result abort */
5065 energy_after += eenv->energy;
5067 } while (sg = sg->next, sg != sd->groups);
5069 eenv->nrg.before = energy_before;
5070 eenv->nrg.after = energy_after;
5071 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5074 trace_sched_energy_diff(eenv->task,
5075 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5076 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5077 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5078 eenv->nrg.delta, eenv->payoff);
5080 return eenv->nrg.diff;
5083 #ifdef CONFIG_SCHED_TUNE
5085 struct target_nrg schedtune_target_nrg;
5088 * System energy normalization
5089 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5090 * corresponding to the specified energy variation.
5093 normalize_energy(int energy_diff)
5096 #ifdef CONFIG_SCHED_DEBUG
5099 /* Check for boundaries */
5100 max_delta = schedtune_target_nrg.max_power;
5101 max_delta -= schedtune_target_nrg.min_power;
5102 WARN_ON(abs(energy_diff) >= max_delta);
5105 /* Do scaling using positive numbers to increase the range */
5106 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5108 /* Scale by energy magnitude */
5109 normalized_nrg <<= SCHED_LOAD_SHIFT;
5111 /* Normalize on max energy for target platform */
5112 normalized_nrg = reciprocal_divide(
5113 normalized_nrg, schedtune_target_nrg.rdiv);
5115 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5119 energy_diff(struct energy_env *eenv)
5121 int boost = schedtune_task_boost(eenv->task);
5124 /* Conpute "absolute" energy diff */
5125 __energy_diff(eenv);
5127 /* Return energy diff when boost margin is 0 */
5129 return eenv->nrg.diff;
5131 /* Compute normalized energy diff */
5132 nrg_delta = normalize_energy(eenv->nrg.diff);
5133 eenv->nrg.delta = nrg_delta;
5135 eenv->payoff = schedtune_accept_deltas(
5141 * When SchedTune is enabled, the energy_diff() function will return
5142 * the computed energy payoff value. Since the energy_diff() return
5143 * value is expected to be negative by its callers, this evaluation
5144 * function return a negative value each time the evaluation return a
5145 * positive payoff, which is the condition for the acceptance of
5146 * a scheduling decision
5148 return -eenv->payoff;
5150 #else /* CONFIG_SCHED_TUNE */
5151 #define energy_diff(eenv) __energy_diff(eenv)
5155 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5156 * A waker of many should wake a different task than the one last awakened
5157 * at a frequency roughly N times higher than one of its wakees. In order
5158 * to determine whether we should let the load spread vs consolodating to
5159 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5160 * partner, and a factor of lls_size higher frequency in the other. With
5161 * both conditions met, we can be relatively sure that the relationship is
5162 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5163 * being client/server, worker/dispatcher, interrupt source or whatever is
5164 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5166 static int wake_wide(struct task_struct *p)
5168 unsigned int master = current->wakee_flips;
5169 unsigned int slave = p->wakee_flips;
5170 int factor = this_cpu_read(sd_llc_size);
5173 swap(master, slave);
5174 if (slave < factor || master < slave * factor)
5179 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5181 s64 this_load, load;
5182 s64 this_eff_load, prev_eff_load;
5183 int idx, this_cpu, prev_cpu;
5184 struct task_group *tg;
5185 unsigned long weight;
5189 this_cpu = smp_processor_id();
5190 prev_cpu = task_cpu(p);
5191 load = source_load(prev_cpu, idx);
5192 this_load = target_load(this_cpu, idx);
5195 * If sync wakeup then subtract the (maximum possible)
5196 * effect of the currently running task from the load
5197 * of the current CPU:
5200 tg = task_group(current);
5201 weight = current->se.avg.load_avg;
5203 this_load += effective_load(tg, this_cpu, -weight, -weight);
5204 load += effective_load(tg, prev_cpu, 0, -weight);
5208 weight = p->se.avg.load_avg;
5211 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5212 * due to the sync cause above having dropped this_load to 0, we'll
5213 * always have an imbalance, but there's really nothing you can do
5214 * about that, so that's good too.
5216 * Otherwise check if either cpus are near enough in load to allow this
5217 * task to be woken on this_cpu.
5219 this_eff_load = 100;
5220 this_eff_load *= capacity_of(prev_cpu);
5222 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5223 prev_eff_load *= capacity_of(this_cpu);
5225 if (this_load > 0) {
5226 this_eff_load *= this_load +
5227 effective_load(tg, this_cpu, weight, weight);
5229 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5232 balanced = this_eff_load <= prev_eff_load;
5234 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5239 schedstat_inc(sd, ttwu_move_affine);
5240 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5245 static inline unsigned long task_util(struct task_struct *p)
5247 #ifdef CONFIG_SCHED_WALT
5248 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5249 unsigned long demand = p->ravg.demand;
5250 return (demand << 10) / walt_ravg_window;
5253 return p->se.avg.util_avg;
5256 unsigned int capacity_margin = 1280; /* ~20% margin */
5258 static inline unsigned long boosted_task_util(struct task_struct *task);
5260 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5262 unsigned long capacity = capacity_of(cpu);
5264 util += boosted_task_util(p);
5266 return (capacity * 1024) > (util * capacity_margin);
5269 static inline bool task_fits_max(struct task_struct *p, int cpu)
5271 unsigned long capacity = capacity_of(cpu);
5272 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5274 if (capacity == max_capacity)
5277 if (capacity * capacity_margin > max_capacity * 1024)
5280 return __task_fits(p, cpu, 0);
5283 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5285 return __task_fits(p, cpu, cpu_util(cpu));
5288 static bool cpu_overutilized(int cpu)
5290 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5293 #ifdef CONFIG_SCHED_TUNE
5296 schedtune_margin(unsigned long signal, long boost)
5298 long long margin = 0;
5301 * Signal proportional compensation (SPC)
5303 * The Boost (B) value is used to compute a Margin (M) which is
5304 * proportional to the complement of the original Signal (S):
5305 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5306 * M = B * S, if B is negative
5307 * The obtained M could be used by the caller to "boost" S.
5310 margin = SCHED_LOAD_SCALE - signal;
5313 margin = -signal * boost;
5315 * Fast integer division by constant:
5316 * Constant : (C) = 100
5317 * Precision : 0.1% (P) = 0.1
5318 * Reference : C * 100 / P (R) = 100000
5321 * Shift bits : ceil(log(R,2)) (S) = 17
5322 * Mult const : round(2^S/C) (M) = 1311
5335 schedtune_cpu_margin(unsigned long util, int cpu)
5337 int boost = schedtune_cpu_boost(cpu);
5342 return schedtune_margin(util, boost);
5346 schedtune_task_margin(struct task_struct *task)
5348 int boost = schedtune_task_boost(task);
5355 util = task_util(task);
5356 margin = schedtune_margin(util, boost);
5361 #else /* CONFIG_SCHED_TUNE */
5364 schedtune_cpu_margin(unsigned long util, int cpu)
5370 schedtune_task_margin(struct task_struct *task)
5375 #endif /* CONFIG_SCHED_TUNE */
5377 static inline unsigned long
5378 boosted_cpu_util(int cpu)
5380 unsigned long util = cpu_util(cpu);
5381 long margin = schedtune_cpu_margin(util, cpu);
5383 trace_sched_boost_cpu(cpu, util, margin);
5385 return util + margin;
5388 static inline unsigned long
5389 boosted_task_util(struct task_struct *task)
5391 unsigned long util = task_util(task);
5392 long margin = schedtune_task_margin(task);
5394 trace_sched_boost_task(task, util, margin);
5396 return util + margin;
5400 * find_idlest_group finds and returns the least busy CPU group within the
5403 static struct sched_group *
5404 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5405 int this_cpu, int sd_flag)
5407 struct sched_group *idlest = NULL, *group = sd->groups;
5408 struct sched_group *fit_group = NULL, *spare_group = NULL;
5409 unsigned long min_load = ULONG_MAX, this_load = 0;
5410 unsigned long fit_capacity = ULONG_MAX;
5411 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5412 int load_idx = sd->forkexec_idx;
5413 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5415 if (sd_flag & SD_BALANCE_WAKE)
5416 load_idx = sd->wake_idx;
5419 unsigned long load, avg_load, spare_capacity;
5423 /* Skip over this group if it has no CPUs allowed */
5424 if (!cpumask_intersects(sched_group_cpus(group),
5425 tsk_cpus_allowed(p)))
5428 local_group = cpumask_test_cpu(this_cpu,
5429 sched_group_cpus(group));
5431 /* Tally up the load of all CPUs in the group */
5434 for_each_cpu(i, sched_group_cpus(group)) {
5435 /* Bias balancing toward cpus of our domain */
5437 load = source_load(i, load_idx);
5439 load = target_load(i, load_idx);
5444 * Look for most energy-efficient group that can fit
5445 * that can fit the task.
5447 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5448 fit_capacity = capacity_of(i);
5453 * Look for group which has most spare capacity on a
5456 spare_capacity = capacity_of(i) - cpu_util(i);
5457 if (spare_capacity > max_spare_capacity) {
5458 max_spare_capacity = spare_capacity;
5459 spare_group = group;
5463 /* Adjust by relative CPU capacity of the group */
5464 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5467 this_load = avg_load;
5468 } else if (avg_load < min_load) {
5469 min_load = avg_load;
5472 } while (group = group->next, group != sd->groups);
5480 if (!idlest || 100*this_load < imbalance*min_load)
5486 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5489 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5491 unsigned long load, min_load = ULONG_MAX;
5492 unsigned int min_exit_latency = UINT_MAX;
5493 u64 latest_idle_timestamp = 0;
5494 int least_loaded_cpu = this_cpu;
5495 int shallowest_idle_cpu = -1;
5498 /* Traverse only the allowed CPUs */
5499 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5500 if (task_fits_spare(p, i)) {
5501 struct rq *rq = cpu_rq(i);
5502 struct cpuidle_state *idle = idle_get_state(rq);
5503 if (idle && idle->exit_latency < min_exit_latency) {
5505 * We give priority to a CPU whose idle state
5506 * has the smallest exit latency irrespective
5507 * of any idle timestamp.
5509 min_exit_latency = idle->exit_latency;
5510 latest_idle_timestamp = rq->idle_stamp;
5511 shallowest_idle_cpu = i;
5512 } else if (idle_cpu(i) &&
5513 (!idle || idle->exit_latency == min_exit_latency) &&
5514 rq->idle_stamp > latest_idle_timestamp) {
5516 * If equal or no active idle state, then
5517 * the most recently idled CPU might have
5520 latest_idle_timestamp = rq->idle_stamp;
5521 shallowest_idle_cpu = i;
5522 } else if (shallowest_idle_cpu == -1) {
5524 * If we haven't found an idle CPU yet
5525 * pick a non-idle one that can fit the task as
5528 shallowest_idle_cpu = i;
5530 } else if (shallowest_idle_cpu == -1) {
5531 load = weighted_cpuload(i);
5532 if (load < min_load || (load == min_load && i == this_cpu)) {
5534 least_loaded_cpu = i;
5539 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5543 * Try and locate an idle CPU in the sched_domain.
5545 static int select_idle_sibling(struct task_struct *p, int target)
5547 struct sched_domain *sd;
5548 struct sched_group *sg;
5549 int i = task_cpu(p);
5551 int best_idle_cstate = -1;
5552 int best_idle_capacity = INT_MAX;
5554 if (!sysctl_sched_cstate_aware) {
5555 if (idle_cpu(target))
5559 * If the prevous cpu is cache affine and idle, don't be stupid.
5561 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5566 * Otherwise, iterate the domains and find an elegible idle cpu.
5568 sd = rcu_dereference(per_cpu(sd_llc, target));
5569 for_each_lower_domain(sd) {
5572 if (!cpumask_intersects(sched_group_cpus(sg),
5573 tsk_cpus_allowed(p)))
5576 if (sysctl_sched_cstate_aware) {
5577 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5578 struct rq *rq = cpu_rq(i);
5579 int idle_idx = idle_get_state_idx(rq);
5580 unsigned long new_usage = boosted_task_util(p);
5581 unsigned long capacity_orig = capacity_orig_of(i);
5582 if (new_usage > capacity_orig || !idle_cpu(i))
5585 if (i == target && new_usage <= capacity_curr_of(target))
5588 if (best_idle < 0 || (idle_idx < best_idle_cstate && capacity_orig <= best_idle_capacity)) {
5590 best_idle_cstate = idle_idx;
5591 best_idle_capacity = capacity_orig;
5595 for_each_cpu(i, sched_group_cpus(sg)) {
5596 if (i == target || !idle_cpu(i))
5600 target = cpumask_first_and(sched_group_cpus(sg),
5601 tsk_cpus_allowed(p));
5606 } while (sg != sd->groups);
5615 static inline int find_best_target(struct task_struct *p, bool prefer_idle)
5618 int target_cpu = -1;
5619 int target_util = 0;
5620 int backup_capacity = 0;
5621 int best_idle_cpu = -1;
5622 int best_idle_cstate = INT_MAX;
5623 int backup_cpu = -1;
5624 unsigned long task_util_boosted, new_util;
5626 task_util_boosted = boosted_task_util(p);
5627 for (iter_cpu = 0; iter_cpu < NR_CPUS; iter_cpu++) {
5633 * favor higher cpus for tasks that prefer idle cores
5635 int i = prefer_idle ? NR_CPUS-iter_cpu-1 : iter_cpu;
5637 if (!cpu_online(i) || !cpumask_test_cpu(i, tsk_cpus_allowed(p)))
5641 * p's blocked utilization is still accounted for on prev_cpu
5642 * so prev_cpu will receive a negative bias due to the double
5643 * accounting. However, the blocked utilization may be zero.
5645 new_util = cpu_util(i) + task_util_boosted;
5648 * Ensure minimum capacity to grant the required boost.
5649 * The target CPU can be already at a capacity level higher
5650 * than the one required to boost the task.
5652 if (new_util > capacity_orig_of(i))
5655 #ifdef CONFIG_SCHED_WALT
5656 if (walt_cpu_high_irqload(i))
5660 * Unconditionally favoring tasks that prefer idle cpus to
5663 if (idle_cpu(i) && prefer_idle) {
5664 if (best_idle_cpu < 0)
5669 cur_capacity = capacity_curr_of(i);
5671 idle_idx = idle_get_state_idx(rq);
5673 if (new_util < cur_capacity) {
5674 if (cpu_rq(i)->nr_running) {
5675 if (target_util == 0 ||
5676 target_util > new_util) {
5678 target_util = new_util;
5680 } else if (!prefer_idle) {
5681 if (best_idle_cpu < 0 ||
5682 (sysctl_sched_cstate_aware &&
5683 best_idle_cstate > idle_idx)) {
5684 best_idle_cstate = idle_idx;
5688 } else if (backup_capacity == 0 ||
5689 backup_capacity > cur_capacity) {
5690 backup_capacity = cur_capacity;
5695 if (prefer_idle && best_idle_cpu >= 0)
5696 target_cpu = best_idle_cpu;
5697 else if (target_cpu < 0)
5698 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5703 static int energy_aware_wake_cpu(struct task_struct *p, int target, int sync)
5705 struct sched_domain *sd;
5706 struct sched_group *sg, *sg_target;
5707 int target_max_cap = INT_MAX;
5708 int target_cpu = task_cpu(p);
5709 unsigned long task_util_boosted, new_util;
5712 if (sysctl_sched_sync_hint_enable && sync) {
5713 int cpu = smp_processor_id();
5714 cpumask_t search_cpus;
5715 cpumask_and(&search_cpus, tsk_cpus_allowed(p), cpu_online_mask);
5716 if (cpumask_test_cpu(cpu, &search_cpus))
5720 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5728 if (sysctl_sched_is_big_little) {
5731 * Find group with sufficient capacity. We only get here if no cpu is
5732 * overutilized. We may end up overutilizing a cpu by adding the task,
5733 * but that should not be any worse than select_idle_sibling().
5734 * load_balance() should sort it out later as we get above the tipping
5738 /* Assuming all cpus are the same in group */
5739 int max_cap_cpu = group_first_cpu(sg);
5742 * Assume smaller max capacity means more energy-efficient.
5743 * Ideally we should query the energy model for the right
5744 * answer but it easily ends up in an exhaustive search.
5746 if (capacity_of(max_cap_cpu) < target_max_cap &&
5747 task_fits_max(p, max_cap_cpu)) {
5749 target_max_cap = capacity_of(max_cap_cpu);
5751 } while (sg = sg->next, sg != sd->groups);
5753 task_util_boosted = boosted_task_util(p);
5754 /* Find cpu with sufficient capacity */
5755 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5757 * p's blocked utilization is still accounted for on prev_cpu
5758 * so prev_cpu will receive a negative bias due to the double
5759 * accounting. However, the blocked utilization may be zero.
5761 new_util = cpu_util(i) + task_util_boosted;
5764 * Ensure minimum capacity to grant the required boost.
5765 * The target CPU can be already at a capacity level higher
5766 * than the one required to boost the task.
5768 if (new_util > capacity_orig_of(i))
5771 if (new_util < capacity_curr_of(i)) {
5773 if (cpu_rq(i)->nr_running)
5777 /* cpu has capacity at higher OPP, keep it as fallback */
5778 if (target_cpu == task_cpu(p))
5783 * Find a cpu with sufficient capacity
5785 #ifdef CONFIG_CGROUP_SCHEDTUNE
5786 bool boosted = schedtune_task_boost(p) > 0;
5787 bool prefer_idle = schedtune_prefer_idle(p) > 0;
5790 bool prefer_idle = 0;
5792 int tmp_target = find_best_target(p, boosted || prefer_idle);
5793 if (tmp_target >= 0) {
5794 target_cpu = tmp_target;
5795 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
5800 if (target_cpu != task_cpu(p)) {
5801 struct energy_env eenv = {
5802 .util_delta = task_util(p),
5803 .src_cpu = task_cpu(p),
5804 .dst_cpu = target_cpu,
5808 /* Not enough spare capacity on previous cpu */
5809 if (cpu_overutilized(task_cpu(p)))
5812 if (energy_diff(&eenv) >= 0)
5820 * select_task_rq_fair: Select target runqueue for the waking task in domains
5821 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5822 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5824 * Balances load by selecting the idlest cpu in the idlest group, or under
5825 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5827 * Returns the target cpu number.
5829 * preempt must be disabled.
5832 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5834 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5835 int cpu = smp_processor_id();
5836 int new_cpu = prev_cpu;
5837 int want_affine = 0;
5838 int sync = wake_flags & WF_SYNC;
5840 if (sd_flag & SD_BALANCE_WAKE)
5841 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5842 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5846 for_each_domain(cpu, tmp) {
5847 if (!(tmp->flags & SD_LOAD_BALANCE))
5851 * If both cpu and prev_cpu are part of this domain,
5852 * cpu is a valid SD_WAKE_AFFINE target.
5854 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5855 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5860 if (tmp->flags & sd_flag)
5862 else if (!want_affine)
5867 sd = NULL; /* Prefer wake_affine over balance flags */
5868 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5873 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5874 new_cpu = energy_aware_wake_cpu(p, prev_cpu, sync);
5875 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5876 new_cpu = select_idle_sibling(p, new_cpu);
5879 struct sched_group *group;
5882 if (!(sd->flags & sd_flag)) {
5887 group = find_idlest_group(sd, p, cpu, sd_flag);
5893 new_cpu = find_idlest_cpu(group, p, cpu);
5894 if (new_cpu == -1 || new_cpu == cpu) {
5895 /* Now try balancing at a lower domain level of cpu */
5900 /* Now try balancing at a lower domain level of new_cpu */
5902 weight = sd->span_weight;
5904 for_each_domain(cpu, tmp) {
5905 if (weight <= tmp->span_weight)
5907 if (tmp->flags & sd_flag)
5910 /* while loop will break here if sd == NULL */
5918 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5919 * cfs_rq_of(p) references at time of call are still valid and identify the
5920 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5921 * other assumptions, including the state of rq->lock, should be made.
5923 static void migrate_task_rq_fair(struct task_struct *p)
5926 * We are supposed to update the task to "current" time, then its up to date
5927 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5928 * what current time is, so simply throw away the out-of-date time. This
5929 * will result in the wakee task is less decayed, but giving the wakee more
5930 * load sounds not bad.
5932 remove_entity_load_avg(&p->se);
5934 /* Tell new CPU we are migrated */
5935 p->se.avg.last_update_time = 0;
5937 /* We have migrated, no longer consider this task hot */
5938 p->se.exec_start = 0;
5941 static void task_dead_fair(struct task_struct *p)
5943 remove_entity_load_avg(&p->se);
5946 #define task_fits_max(p, cpu) true
5947 #endif /* CONFIG_SMP */
5949 static unsigned long
5950 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5952 unsigned long gran = sysctl_sched_wakeup_granularity;
5955 * Since its curr running now, convert the gran from real-time
5956 * to virtual-time in his units.
5958 * By using 'se' instead of 'curr' we penalize light tasks, so
5959 * they get preempted easier. That is, if 'se' < 'curr' then
5960 * the resulting gran will be larger, therefore penalizing the
5961 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5962 * be smaller, again penalizing the lighter task.
5964 * This is especially important for buddies when the leftmost
5965 * task is higher priority than the buddy.
5967 return calc_delta_fair(gran, se);
5971 * Should 'se' preempt 'curr'.
5985 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5987 s64 gran, vdiff = curr->vruntime - se->vruntime;
5992 gran = wakeup_gran(curr, se);
5999 static void set_last_buddy(struct sched_entity *se)
6001 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6004 for_each_sched_entity(se)
6005 cfs_rq_of(se)->last = se;
6008 static void set_next_buddy(struct sched_entity *se)
6010 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6013 for_each_sched_entity(se)
6014 cfs_rq_of(se)->next = se;
6017 static void set_skip_buddy(struct sched_entity *se)
6019 for_each_sched_entity(se)
6020 cfs_rq_of(se)->skip = se;
6024 * Preempt the current task with a newly woken task if needed:
6026 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6028 struct task_struct *curr = rq->curr;
6029 struct sched_entity *se = &curr->se, *pse = &p->se;
6030 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6031 int scale = cfs_rq->nr_running >= sched_nr_latency;
6032 int next_buddy_marked = 0;
6034 if (unlikely(se == pse))
6038 * This is possible from callers such as attach_tasks(), in which we
6039 * unconditionally check_prempt_curr() after an enqueue (which may have
6040 * lead to a throttle). This both saves work and prevents false
6041 * next-buddy nomination below.
6043 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6046 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6047 set_next_buddy(pse);
6048 next_buddy_marked = 1;
6052 * We can come here with TIF_NEED_RESCHED already set from new task
6055 * Note: this also catches the edge-case of curr being in a throttled
6056 * group (e.g. via set_curr_task), since update_curr() (in the
6057 * enqueue of curr) will have resulted in resched being set. This
6058 * prevents us from potentially nominating it as a false LAST_BUDDY
6061 if (test_tsk_need_resched(curr))
6064 /* Idle tasks are by definition preempted by non-idle tasks. */
6065 if (unlikely(curr->policy == SCHED_IDLE) &&
6066 likely(p->policy != SCHED_IDLE))
6070 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6071 * is driven by the tick):
6073 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6076 find_matching_se(&se, &pse);
6077 update_curr(cfs_rq_of(se));
6079 if (wakeup_preempt_entity(se, pse) == 1) {
6081 * Bias pick_next to pick the sched entity that is
6082 * triggering this preemption.
6084 if (!next_buddy_marked)
6085 set_next_buddy(pse);
6094 * Only set the backward buddy when the current task is still
6095 * on the rq. This can happen when a wakeup gets interleaved
6096 * with schedule on the ->pre_schedule() or idle_balance()
6097 * point, either of which can * drop the rq lock.
6099 * Also, during early boot the idle thread is in the fair class,
6100 * for obvious reasons its a bad idea to schedule back to it.
6102 if (unlikely(!se->on_rq || curr == rq->idle))
6105 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6109 static struct task_struct *
6110 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6112 struct cfs_rq *cfs_rq = &rq->cfs;
6113 struct sched_entity *se;
6114 struct task_struct *p;
6118 #ifdef CONFIG_FAIR_GROUP_SCHED
6119 if (!cfs_rq->nr_running)
6122 if (prev->sched_class != &fair_sched_class)
6126 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6127 * likely that a next task is from the same cgroup as the current.
6129 * Therefore attempt to avoid putting and setting the entire cgroup
6130 * hierarchy, only change the part that actually changes.
6134 struct sched_entity *curr = cfs_rq->curr;
6137 * Since we got here without doing put_prev_entity() we also
6138 * have to consider cfs_rq->curr. If it is still a runnable
6139 * entity, update_curr() will update its vruntime, otherwise
6140 * forget we've ever seen it.
6144 update_curr(cfs_rq);
6149 * This call to check_cfs_rq_runtime() will do the
6150 * throttle and dequeue its entity in the parent(s).
6151 * Therefore the 'simple' nr_running test will indeed
6154 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6158 se = pick_next_entity(cfs_rq, curr);
6159 cfs_rq = group_cfs_rq(se);
6165 * Since we haven't yet done put_prev_entity and if the selected task
6166 * is a different task than we started out with, try and touch the
6167 * least amount of cfs_rqs.
6170 struct sched_entity *pse = &prev->se;
6172 while (!(cfs_rq = is_same_group(se, pse))) {
6173 int se_depth = se->depth;
6174 int pse_depth = pse->depth;
6176 if (se_depth <= pse_depth) {
6177 put_prev_entity(cfs_rq_of(pse), pse);
6178 pse = parent_entity(pse);
6180 if (se_depth >= pse_depth) {
6181 set_next_entity(cfs_rq_of(se), se);
6182 se = parent_entity(se);
6186 put_prev_entity(cfs_rq, pse);
6187 set_next_entity(cfs_rq, se);
6190 if (hrtick_enabled(rq))
6191 hrtick_start_fair(rq, p);
6193 rq->misfit_task = !task_fits_max(p, rq->cpu);
6200 if (!cfs_rq->nr_running)
6203 put_prev_task(rq, prev);
6206 se = pick_next_entity(cfs_rq, NULL);
6207 set_next_entity(cfs_rq, se);
6208 cfs_rq = group_cfs_rq(se);
6213 if (hrtick_enabled(rq))
6214 hrtick_start_fair(rq, p);
6216 rq->misfit_task = !task_fits_max(p, rq->cpu);
6221 rq->misfit_task = 0;
6223 * This is OK, because current is on_cpu, which avoids it being picked
6224 * for load-balance and preemption/IRQs are still disabled avoiding
6225 * further scheduler activity on it and we're being very careful to
6226 * re-start the picking loop.
6228 lockdep_unpin_lock(&rq->lock);
6229 new_tasks = idle_balance(rq);
6230 lockdep_pin_lock(&rq->lock);
6232 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6233 * possible for any higher priority task to appear. In that case we
6234 * must re-start the pick_next_entity() loop.
6246 * Account for a descheduled task:
6248 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6250 struct sched_entity *se = &prev->se;
6251 struct cfs_rq *cfs_rq;
6253 for_each_sched_entity(se) {
6254 cfs_rq = cfs_rq_of(se);
6255 put_prev_entity(cfs_rq, se);
6260 * sched_yield() is very simple
6262 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6264 static void yield_task_fair(struct rq *rq)
6266 struct task_struct *curr = rq->curr;
6267 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6268 struct sched_entity *se = &curr->se;
6271 * Are we the only task in the tree?
6273 if (unlikely(rq->nr_running == 1))
6276 clear_buddies(cfs_rq, se);
6278 if (curr->policy != SCHED_BATCH) {
6279 update_rq_clock(rq);
6281 * Update run-time statistics of the 'current'.
6283 update_curr(cfs_rq);
6285 * Tell update_rq_clock() that we've just updated,
6286 * so we don't do microscopic update in schedule()
6287 * and double the fastpath cost.
6289 rq_clock_skip_update(rq, true);
6295 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6297 struct sched_entity *se = &p->se;
6299 /* throttled hierarchies are not runnable */
6300 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6303 /* Tell the scheduler that we'd really like pse to run next. */
6306 yield_task_fair(rq);
6312 /**************************************************
6313 * Fair scheduling class load-balancing methods.
6317 * The purpose of load-balancing is to achieve the same basic fairness the
6318 * per-cpu scheduler provides, namely provide a proportional amount of compute
6319 * time to each task. This is expressed in the following equation:
6321 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6323 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6324 * W_i,0 is defined as:
6326 * W_i,0 = \Sum_j w_i,j (2)
6328 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6329 * is derived from the nice value as per prio_to_weight[].
6331 * The weight average is an exponential decay average of the instantaneous
6334 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6336 * C_i is the compute capacity of cpu i, typically it is the
6337 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6338 * can also include other factors [XXX].
6340 * To achieve this balance we define a measure of imbalance which follows
6341 * directly from (1):
6343 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6345 * We them move tasks around to minimize the imbalance. In the continuous
6346 * function space it is obvious this converges, in the discrete case we get
6347 * a few fun cases generally called infeasible weight scenarios.
6350 * - infeasible weights;
6351 * - local vs global optima in the discrete case. ]
6356 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6357 * for all i,j solution, we create a tree of cpus that follows the hardware
6358 * topology where each level pairs two lower groups (or better). This results
6359 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6360 * tree to only the first of the previous level and we decrease the frequency
6361 * of load-balance at each level inv. proportional to the number of cpus in
6367 * \Sum { --- * --- * 2^i } = O(n) (5)
6369 * `- size of each group
6370 * | | `- number of cpus doing load-balance
6372 * `- sum over all levels
6374 * Coupled with a limit on how many tasks we can migrate every balance pass,
6375 * this makes (5) the runtime complexity of the balancer.
6377 * An important property here is that each CPU is still (indirectly) connected
6378 * to every other cpu in at most O(log n) steps:
6380 * The adjacency matrix of the resulting graph is given by:
6383 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6386 * And you'll find that:
6388 * A^(log_2 n)_i,j != 0 for all i,j (7)
6390 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6391 * The task movement gives a factor of O(m), giving a convergence complexity
6394 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6399 * In order to avoid CPUs going idle while there's still work to do, new idle
6400 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6401 * tree itself instead of relying on other CPUs to bring it work.
6403 * This adds some complexity to both (5) and (8) but it reduces the total idle
6411 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6414 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6419 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6421 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6423 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6426 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6427 * rewrite all of this once again.]
6430 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6432 enum fbq_type { regular, remote, all };
6441 #define LBF_ALL_PINNED 0x01
6442 #define LBF_NEED_BREAK 0x02
6443 #define LBF_DST_PINNED 0x04
6444 #define LBF_SOME_PINNED 0x08
6447 struct sched_domain *sd;
6455 struct cpumask *dst_grpmask;
6457 enum cpu_idle_type idle;
6459 unsigned int src_grp_nr_running;
6460 /* The set of CPUs under consideration for load-balancing */
6461 struct cpumask *cpus;
6466 unsigned int loop_break;
6467 unsigned int loop_max;
6469 enum fbq_type fbq_type;
6470 enum group_type busiest_group_type;
6471 struct list_head tasks;
6475 * Is this task likely cache-hot:
6477 static int task_hot(struct task_struct *p, struct lb_env *env)
6481 lockdep_assert_held(&env->src_rq->lock);
6483 if (p->sched_class != &fair_sched_class)
6486 if (unlikely(p->policy == SCHED_IDLE))
6490 * Buddy candidates are cache hot:
6492 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6493 (&p->se == cfs_rq_of(&p->se)->next ||
6494 &p->se == cfs_rq_of(&p->se)->last))
6497 if (sysctl_sched_migration_cost == -1)
6499 if (sysctl_sched_migration_cost == 0)
6502 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6504 return delta < (s64)sysctl_sched_migration_cost;
6507 #ifdef CONFIG_NUMA_BALANCING
6509 * Returns 1, if task migration degrades locality
6510 * Returns 0, if task migration improves locality i.e migration preferred.
6511 * Returns -1, if task migration is not affected by locality.
6513 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6515 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6516 unsigned long src_faults, dst_faults;
6517 int src_nid, dst_nid;
6519 if (!static_branch_likely(&sched_numa_balancing))
6522 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6525 src_nid = cpu_to_node(env->src_cpu);
6526 dst_nid = cpu_to_node(env->dst_cpu);
6528 if (src_nid == dst_nid)
6531 /* Migrating away from the preferred node is always bad. */
6532 if (src_nid == p->numa_preferred_nid) {
6533 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6539 /* Encourage migration to the preferred node. */
6540 if (dst_nid == p->numa_preferred_nid)
6544 src_faults = group_faults(p, src_nid);
6545 dst_faults = group_faults(p, dst_nid);
6547 src_faults = task_faults(p, src_nid);
6548 dst_faults = task_faults(p, dst_nid);
6551 return dst_faults < src_faults;
6555 static inline int migrate_degrades_locality(struct task_struct *p,
6563 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6566 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6570 lockdep_assert_held(&env->src_rq->lock);
6573 * We do not migrate tasks that are:
6574 * 1) throttled_lb_pair, or
6575 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6576 * 3) running (obviously), or
6577 * 4) are cache-hot on their current CPU.
6579 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6582 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6585 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6587 env->flags |= LBF_SOME_PINNED;
6590 * Remember if this task can be migrated to any other cpu in
6591 * our sched_group. We may want to revisit it if we couldn't
6592 * meet load balance goals by pulling other tasks on src_cpu.
6594 * Also avoid computing new_dst_cpu if we have already computed
6595 * one in current iteration.
6597 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6600 /* Prevent to re-select dst_cpu via env's cpus */
6601 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6602 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6603 env->flags |= LBF_DST_PINNED;
6604 env->new_dst_cpu = cpu;
6612 /* Record that we found atleast one task that could run on dst_cpu */
6613 env->flags &= ~LBF_ALL_PINNED;
6615 if (task_running(env->src_rq, p)) {
6616 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6621 * Aggressive migration if:
6622 * 1) destination numa is preferred
6623 * 2) task is cache cold, or
6624 * 3) too many balance attempts have failed.
6626 tsk_cache_hot = migrate_degrades_locality(p, env);
6627 if (tsk_cache_hot == -1)
6628 tsk_cache_hot = task_hot(p, env);
6630 if (tsk_cache_hot <= 0 ||
6631 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6632 if (tsk_cache_hot == 1) {
6633 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6634 schedstat_inc(p, se.statistics.nr_forced_migrations);
6639 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6644 * detach_task() -- detach the task for the migration specified in env
6646 static void detach_task(struct task_struct *p, struct lb_env *env)
6648 lockdep_assert_held(&env->src_rq->lock);
6650 deactivate_task(env->src_rq, p, 0);
6651 p->on_rq = TASK_ON_RQ_MIGRATING;
6652 double_lock_balance(env->src_rq, env->dst_rq);
6653 set_task_cpu(p, env->dst_cpu);
6654 double_unlock_balance(env->src_rq, env->dst_rq);
6658 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6659 * part of active balancing operations within "domain".
6661 * Returns a task if successful and NULL otherwise.
6663 static struct task_struct *detach_one_task(struct lb_env *env)
6665 struct task_struct *p, *n;
6667 lockdep_assert_held(&env->src_rq->lock);
6669 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6670 if (!can_migrate_task(p, env))
6673 detach_task(p, env);
6676 * Right now, this is only the second place where
6677 * lb_gained[env->idle] is updated (other is detach_tasks)
6678 * so we can safely collect stats here rather than
6679 * inside detach_tasks().
6681 schedstat_inc(env->sd, lb_gained[env->idle]);
6687 static const unsigned int sched_nr_migrate_break = 32;
6690 * detach_tasks() -- tries to detach up to imbalance weighted load from
6691 * busiest_rq, as part of a balancing operation within domain "sd".
6693 * Returns number of detached tasks if successful and 0 otherwise.
6695 static int detach_tasks(struct lb_env *env)
6697 struct list_head *tasks = &env->src_rq->cfs_tasks;
6698 struct task_struct *p;
6702 lockdep_assert_held(&env->src_rq->lock);
6704 if (env->imbalance <= 0)
6707 while (!list_empty(tasks)) {
6709 * We don't want to steal all, otherwise we may be treated likewise,
6710 * which could at worst lead to a livelock crash.
6712 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6715 p = list_first_entry(tasks, struct task_struct, se.group_node);
6718 /* We've more or less seen every task there is, call it quits */
6719 if (env->loop > env->loop_max)
6722 /* take a breather every nr_migrate tasks */
6723 if (env->loop > env->loop_break) {
6724 env->loop_break += sched_nr_migrate_break;
6725 env->flags |= LBF_NEED_BREAK;
6729 if (!can_migrate_task(p, env))
6732 load = task_h_load(p);
6734 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6737 if ((load / 2) > env->imbalance)
6740 detach_task(p, env);
6741 list_add(&p->se.group_node, &env->tasks);
6744 env->imbalance -= load;
6746 #ifdef CONFIG_PREEMPT
6748 * NEWIDLE balancing is a source of latency, so preemptible
6749 * kernels will stop after the first task is detached to minimize
6750 * the critical section.
6752 if (env->idle == CPU_NEWLY_IDLE)
6757 * We only want to steal up to the prescribed amount of
6760 if (env->imbalance <= 0)
6765 list_move_tail(&p->se.group_node, tasks);
6769 * Right now, this is one of only two places we collect this stat
6770 * so we can safely collect detach_one_task() stats here rather
6771 * than inside detach_one_task().
6773 schedstat_add(env->sd, lb_gained[env->idle], detached);
6779 * attach_task() -- attach the task detached by detach_task() to its new rq.
6781 static void attach_task(struct rq *rq, struct task_struct *p)
6783 lockdep_assert_held(&rq->lock);
6785 BUG_ON(task_rq(p) != rq);
6786 p->on_rq = TASK_ON_RQ_QUEUED;
6787 activate_task(rq, p, 0);
6788 check_preempt_curr(rq, p, 0);
6792 * attach_one_task() -- attaches the task returned from detach_one_task() to
6795 static void attach_one_task(struct rq *rq, struct task_struct *p)
6797 raw_spin_lock(&rq->lock);
6800 * We want to potentially raise target_cpu's OPP.
6802 update_capacity_of(cpu_of(rq));
6803 raw_spin_unlock(&rq->lock);
6807 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6810 static void attach_tasks(struct lb_env *env)
6812 struct list_head *tasks = &env->tasks;
6813 struct task_struct *p;
6815 raw_spin_lock(&env->dst_rq->lock);
6817 while (!list_empty(tasks)) {
6818 p = list_first_entry(tasks, struct task_struct, se.group_node);
6819 list_del_init(&p->se.group_node);
6821 attach_task(env->dst_rq, p);
6825 * We want to potentially raise env.dst_cpu's OPP.
6827 update_capacity_of(env->dst_cpu);
6829 raw_spin_unlock(&env->dst_rq->lock);
6832 #ifdef CONFIG_FAIR_GROUP_SCHED
6833 static void update_blocked_averages(int cpu)
6835 struct rq *rq = cpu_rq(cpu);
6836 struct cfs_rq *cfs_rq;
6837 unsigned long flags;
6839 raw_spin_lock_irqsave(&rq->lock, flags);
6840 update_rq_clock(rq);
6843 * Iterates the task_group tree in a bottom up fashion, see
6844 * list_add_leaf_cfs_rq() for details.
6846 for_each_leaf_cfs_rq(rq, cfs_rq) {
6847 /* throttled entities do not contribute to load */
6848 if (throttled_hierarchy(cfs_rq))
6851 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6852 update_tg_load_avg(cfs_rq, 0);
6854 raw_spin_unlock_irqrestore(&rq->lock, flags);
6858 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6859 * This needs to be done in a top-down fashion because the load of a child
6860 * group is a fraction of its parents load.
6862 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6864 struct rq *rq = rq_of(cfs_rq);
6865 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6866 unsigned long now = jiffies;
6869 if (cfs_rq->last_h_load_update == now)
6872 cfs_rq->h_load_next = NULL;
6873 for_each_sched_entity(se) {
6874 cfs_rq = cfs_rq_of(se);
6875 cfs_rq->h_load_next = se;
6876 if (cfs_rq->last_h_load_update == now)
6881 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6882 cfs_rq->last_h_load_update = now;
6885 while ((se = cfs_rq->h_load_next) != NULL) {
6886 load = cfs_rq->h_load;
6887 load = div64_ul(load * se->avg.load_avg,
6888 cfs_rq_load_avg(cfs_rq) + 1);
6889 cfs_rq = group_cfs_rq(se);
6890 cfs_rq->h_load = load;
6891 cfs_rq->last_h_load_update = now;
6895 static unsigned long task_h_load(struct task_struct *p)
6897 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6899 update_cfs_rq_h_load(cfs_rq);
6900 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6901 cfs_rq_load_avg(cfs_rq) + 1);
6904 static inline void update_blocked_averages(int cpu)
6906 struct rq *rq = cpu_rq(cpu);
6907 struct cfs_rq *cfs_rq = &rq->cfs;
6908 unsigned long flags;
6910 raw_spin_lock_irqsave(&rq->lock, flags);
6911 update_rq_clock(rq);
6912 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6913 raw_spin_unlock_irqrestore(&rq->lock, flags);
6916 static unsigned long task_h_load(struct task_struct *p)
6918 return p->se.avg.load_avg;
6922 /********** Helpers for find_busiest_group ************************/
6925 * sg_lb_stats - stats of a sched_group required for load_balancing
6927 struct sg_lb_stats {
6928 unsigned long avg_load; /*Avg load across the CPUs of the group */
6929 unsigned long group_load; /* Total load over the CPUs of the group */
6930 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6931 unsigned long load_per_task;
6932 unsigned long group_capacity;
6933 unsigned long group_util; /* Total utilization of the group */
6934 unsigned int sum_nr_running; /* Nr tasks running in the group */
6935 unsigned int idle_cpus;
6936 unsigned int group_weight;
6937 enum group_type group_type;
6938 int group_no_capacity;
6939 int group_misfit_task; /* A cpu has a task too big for its capacity */
6940 #ifdef CONFIG_NUMA_BALANCING
6941 unsigned int nr_numa_running;
6942 unsigned int nr_preferred_running;
6947 * sd_lb_stats - Structure to store the statistics of a sched_domain
6948 * during load balancing.
6950 struct sd_lb_stats {
6951 struct sched_group *busiest; /* Busiest group in this sd */
6952 struct sched_group *local; /* Local group in this sd */
6953 unsigned long total_load; /* Total load of all groups in sd */
6954 unsigned long total_capacity; /* Total capacity of all groups in sd */
6955 unsigned long avg_load; /* Average load across all groups in sd */
6957 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6958 struct sg_lb_stats local_stat; /* Statistics of the local group */
6961 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6964 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6965 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6966 * We must however clear busiest_stat::avg_load because
6967 * update_sd_pick_busiest() reads this before assignment.
6969 *sds = (struct sd_lb_stats){
6973 .total_capacity = 0UL,
6976 .sum_nr_running = 0,
6977 .group_type = group_other,
6983 * get_sd_load_idx - Obtain the load index for a given sched domain.
6984 * @sd: The sched_domain whose load_idx is to be obtained.
6985 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6987 * Return: The load index.
6989 static inline int get_sd_load_idx(struct sched_domain *sd,
6990 enum cpu_idle_type idle)
6996 load_idx = sd->busy_idx;
6999 case CPU_NEWLY_IDLE:
7000 load_idx = sd->newidle_idx;
7003 load_idx = sd->idle_idx;
7010 static unsigned long scale_rt_capacity(int cpu)
7012 struct rq *rq = cpu_rq(cpu);
7013 u64 total, used, age_stamp, avg;
7017 * Since we're reading these variables without serialization make sure
7018 * we read them once before doing sanity checks on them.
7020 age_stamp = READ_ONCE(rq->age_stamp);
7021 avg = READ_ONCE(rq->rt_avg);
7022 delta = __rq_clock_broken(rq) - age_stamp;
7024 if (unlikely(delta < 0))
7027 total = sched_avg_period() + delta;
7029 used = div_u64(avg, total);
7032 * deadline bandwidth is defined at system level so we must
7033 * weight this bandwidth with the max capacity of the system.
7034 * As a reminder, avg_bw is 20bits width and
7035 * scale_cpu_capacity is 10 bits width
7037 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7039 if (likely(used < SCHED_CAPACITY_SCALE))
7040 return SCHED_CAPACITY_SCALE - used;
7045 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7047 raw_spin_lock_init(&mcc->lock);
7052 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7054 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7055 struct sched_group *sdg = sd->groups;
7056 struct max_cpu_capacity *mcc;
7057 unsigned long max_capacity;
7059 unsigned long flags;
7061 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7063 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7065 raw_spin_lock_irqsave(&mcc->lock, flags);
7066 max_capacity = mcc->val;
7067 max_cap_cpu = mcc->cpu;
7069 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7070 (max_capacity < capacity)) {
7071 mcc->val = capacity;
7073 #ifdef CONFIG_SCHED_DEBUG
7074 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7075 pr_info("CPU%d: update max cpu_capacity %lu\n", cpu, capacity);
7079 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7081 skip_unlock: __attribute__ ((unused));
7082 capacity *= scale_rt_capacity(cpu);
7083 capacity >>= SCHED_CAPACITY_SHIFT;
7088 cpu_rq(cpu)->cpu_capacity = capacity;
7089 sdg->sgc->capacity = capacity;
7090 sdg->sgc->max_capacity = capacity;
7093 void update_group_capacity(struct sched_domain *sd, int cpu)
7095 struct sched_domain *child = sd->child;
7096 struct sched_group *group, *sdg = sd->groups;
7097 unsigned long capacity, max_capacity;
7098 unsigned long interval;
7100 interval = msecs_to_jiffies(sd->balance_interval);
7101 interval = clamp(interval, 1UL, max_load_balance_interval);
7102 sdg->sgc->next_update = jiffies + interval;
7105 update_cpu_capacity(sd, cpu);
7112 if (child->flags & SD_OVERLAP) {
7114 * SD_OVERLAP domains cannot assume that child groups
7115 * span the current group.
7118 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7119 struct sched_group_capacity *sgc;
7120 struct rq *rq = cpu_rq(cpu);
7123 * build_sched_domains() -> init_sched_groups_capacity()
7124 * gets here before we've attached the domains to the
7127 * Use capacity_of(), which is set irrespective of domains
7128 * in update_cpu_capacity().
7130 * This avoids capacity from being 0 and
7131 * causing divide-by-zero issues on boot.
7133 if (unlikely(!rq->sd)) {
7134 capacity += capacity_of(cpu);
7136 sgc = rq->sd->groups->sgc;
7137 capacity += sgc->capacity;
7140 max_capacity = max(capacity, max_capacity);
7144 * !SD_OVERLAP domains can assume that child groups
7145 * span the current group.
7148 group = child->groups;
7150 struct sched_group_capacity *sgc = group->sgc;
7152 capacity += sgc->capacity;
7153 max_capacity = max(sgc->max_capacity, max_capacity);
7154 group = group->next;
7155 } while (group != child->groups);
7158 sdg->sgc->capacity = capacity;
7159 sdg->sgc->max_capacity = max_capacity;
7163 * Check whether the capacity of the rq has been noticeably reduced by side
7164 * activity. The imbalance_pct is used for the threshold.
7165 * Return true is the capacity is reduced
7168 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7170 return ((rq->cpu_capacity * sd->imbalance_pct) <
7171 (rq->cpu_capacity_orig * 100));
7175 * Group imbalance indicates (and tries to solve) the problem where balancing
7176 * groups is inadequate due to tsk_cpus_allowed() constraints.
7178 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7179 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7182 * { 0 1 2 3 } { 4 5 6 7 }
7185 * If we were to balance group-wise we'd place two tasks in the first group and
7186 * two tasks in the second group. Clearly this is undesired as it will overload
7187 * cpu 3 and leave one of the cpus in the second group unused.
7189 * The current solution to this issue is detecting the skew in the first group
7190 * by noticing the lower domain failed to reach balance and had difficulty
7191 * moving tasks due to affinity constraints.
7193 * When this is so detected; this group becomes a candidate for busiest; see
7194 * update_sd_pick_busiest(). And calculate_imbalance() and
7195 * find_busiest_group() avoid some of the usual balance conditions to allow it
7196 * to create an effective group imbalance.
7198 * This is a somewhat tricky proposition since the next run might not find the
7199 * group imbalance and decide the groups need to be balanced again. A most
7200 * subtle and fragile situation.
7203 static inline int sg_imbalanced(struct sched_group *group)
7205 return group->sgc->imbalance;
7209 * group_has_capacity returns true if the group has spare capacity that could
7210 * be used by some tasks.
7211 * We consider that a group has spare capacity if the * number of task is
7212 * smaller than the number of CPUs or if the utilization is lower than the
7213 * available capacity for CFS tasks.
7214 * For the latter, we use a threshold to stabilize the state, to take into
7215 * account the variance of the tasks' load and to return true if the available
7216 * capacity in meaningful for the load balancer.
7217 * As an example, an available capacity of 1% can appear but it doesn't make
7218 * any benefit for the load balance.
7221 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7223 if (sgs->sum_nr_running < sgs->group_weight)
7226 if ((sgs->group_capacity * 100) >
7227 (sgs->group_util * env->sd->imbalance_pct))
7234 * group_is_overloaded returns true if the group has more tasks than it can
7236 * group_is_overloaded is not equals to !group_has_capacity because a group
7237 * with the exact right number of tasks, has no more spare capacity but is not
7238 * overloaded so both group_has_capacity and group_is_overloaded return
7242 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7244 if (sgs->sum_nr_running <= sgs->group_weight)
7247 if ((sgs->group_capacity * 100) <
7248 (sgs->group_util * env->sd->imbalance_pct))
7256 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7257 * per-cpu capacity than sched_group ref.
7260 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7262 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7263 ref->sgc->max_capacity;
7267 group_type group_classify(struct sched_group *group,
7268 struct sg_lb_stats *sgs)
7270 if (sgs->group_no_capacity)
7271 return group_overloaded;
7273 if (sg_imbalanced(group))
7274 return group_imbalanced;
7276 if (sgs->group_misfit_task)
7277 return group_misfit_task;
7283 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7284 * @env: The load balancing environment.
7285 * @group: sched_group whose statistics are to be updated.
7286 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7287 * @local_group: Does group contain this_cpu.
7288 * @sgs: variable to hold the statistics for this group.
7289 * @overload: Indicate more than one runnable task for any CPU.
7290 * @overutilized: Indicate overutilization for any CPU.
7292 static inline void update_sg_lb_stats(struct lb_env *env,
7293 struct sched_group *group, int load_idx,
7294 int local_group, struct sg_lb_stats *sgs,
7295 bool *overload, bool *overutilized)
7300 memset(sgs, 0, sizeof(*sgs));
7302 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7303 struct rq *rq = cpu_rq(i);
7305 /* Bias balancing toward cpus of our domain */
7307 load = target_load(i, load_idx);
7309 load = source_load(i, load_idx);
7311 sgs->group_load += load;
7312 sgs->group_util += cpu_util(i);
7313 sgs->sum_nr_running += rq->cfs.h_nr_running;
7315 if (rq->nr_running > 1)
7318 #ifdef CONFIG_NUMA_BALANCING
7319 sgs->nr_numa_running += rq->nr_numa_running;
7320 sgs->nr_preferred_running += rq->nr_preferred_running;
7322 sgs->sum_weighted_load += weighted_cpuload(i);
7326 if (cpu_overutilized(i)) {
7327 *overutilized = true;
7328 if (!sgs->group_misfit_task && rq->misfit_task)
7329 sgs->group_misfit_task = capacity_of(i);
7333 /* Adjust by relative CPU capacity of the group */
7334 sgs->group_capacity = group->sgc->capacity;
7335 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7337 if (sgs->sum_nr_running)
7338 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7340 sgs->group_weight = group->group_weight;
7342 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7343 sgs->group_type = group_classify(group, sgs);
7347 * update_sd_pick_busiest - return 1 on busiest group
7348 * @env: The load balancing environment.
7349 * @sds: sched_domain statistics
7350 * @sg: sched_group candidate to be checked for being the busiest
7351 * @sgs: sched_group statistics
7353 * Determine if @sg is a busier group than the previously selected
7356 * Return: %true if @sg is a busier group than the previously selected
7357 * busiest group. %false otherwise.
7359 static bool update_sd_pick_busiest(struct lb_env *env,
7360 struct sd_lb_stats *sds,
7361 struct sched_group *sg,
7362 struct sg_lb_stats *sgs)
7364 struct sg_lb_stats *busiest = &sds->busiest_stat;
7366 if (sgs->group_type > busiest->group_type)
7369 if (sgs->group_type < busiest->group_type)
7373 * Candidate sg doesn't face any serious load-balance problems
7374 * so don't pick it if the local sg is already filled up.
7376 if (sgs->group_type == group_other &&
7377 !group_has_capacity(env, &sds->local_stat))
7380 if (sgs->avg_load <= busiest->avg_load)
7384 * Candiate sg has no more than one task per cpu and has higher
7385 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7387 if (sgs->sum_nr_running <= sgs->group_weight &&
7388 group_smaller_cpu_capacity(sds->local, sg))
7391 /* This is the busiest node in its class. */
7392 if (!(env->sd->flags & SD_ASYM_PACKING))
7396 * ASYM_PACKING needs to move all the work to the lowest
7397 * numbered CPUs in the group, therefore mark all groups
7398 * higher than ourself as busy.
7400 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7404 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7411 #ifdef CONFIG_NUMA_BALANCING
7412 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7414 if (sgs->sum_nr_running > sgs->nr_numa_running)
7416 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7421 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7423 if (rq->nr_running > rq->nr_numa_running)
7425 if (rq->nr_running > rq->nr_preferred_running)
7430 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7435 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7439 #endif /* CONFIG_NUMA_BALANCING */
7442 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7443 * @env: The load balancing environment.
7444 * @sds: variable to hold the statistics for this sched_domain.
7446 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7448 struct sched_domain *child = env->sd->child;
7449 struct sched_group *sg = env->sd->groups;
7450 struct sg_lb_stats tmp_sgs;
7451 int load_idx, prefer_sibling = 0;
7452 bool overload = false, overutilized = false;
7454 if (child && child->flags & SD_PREFER_SIBLING)
7457 load_idx = get_sd_load_idx(env->sd, env->idle);
7460 struct sg_lb_stats *sgs = &tmp_sgs;
7463 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7466 sgs = &sds->local_stat;
7468 if (env->idle != CPU_NEWLY_IDLE ||
7469 time_after_eq(jiffies, sg->sgc->next_update))
7470 update_group_capacity(env->sd, env->dst_cpu);
7473 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7474 &overload, &overutilized);
7480 * In case the child domain prefers tasks go to siblings
7481 * first, lower the sg capacity so that we'll try
7482 * and move all the excess tasks away. We lower the capacity
7483 * of a group only if the local group has the capacity to fit
7484 * these excess tasks. The extra check prevents the case where
7485 * you always pull from the heaviest group when it is already
7486 * under-utilized (possible with a large weight task outweighs
7487 * the tasks on the system).
7489 if (prefer_sibling && sds->local &&
7490 group_has_capacity(env, &sds->local_stat) &&
7491 (sgs->sum_nr_running > 1)) {
7492 sgs->group_no_capacity = 1;
7493 sgs->group_type = group_classify(sg, sgs);
7497 * Ignore task groups with misfit tasks if local group has no
7498 * capacity or if per-cpu capacity isn't higher.
7500 if (sgs->group_type == group_misfit_task &&
7501 (!group_has_capacity(env, &sds->local_stat) ||
7502 !group_smaller_cpu_capacity(sg, sds->local)))
7503 sgs->group_type = group_other;
7505 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7507 sds->busiest_stat = *sgs;
7511 /* Now, start updating sd_lb_stats */
7512 sds->total_load += sgs->group_load;
7513 sds->total_capacity += sgs->group_capacity;
7516 } while (sg != env->sd->groups);
7518 if (env->sd->flags & SD_NUMA)
7519 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7521 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7523 if (!env->sd->parent) {
7524 /* update overload indicator if we are at root domain */
7525 if (env->dst_rq->rd->overload != overload)
7526 env->dst_rq->rd->overload = overload;
7528 /* Update over-utilization (tipping point, U >= 0) indicator */
7529 if (env->dst_rq->rd->overutilized != overutilized) {
7530 env->dst_rq->rd->overutilized = overutilized;
7531 trace_sched_overutilized(overutilized);
7534 if (!env->dst_rq->rd->overutilized && overutilized) {
7535 env->dst_rq->rd->overutilized = true;
7536 trace_sched_overutilized(true);
7543 * check_asym_packing - Check to see if the group is packed into the
7546 * This is primarily intended to used at the sibling level. Some
7547 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7548 * case of POWER7, it can move to lower SMT modes only when higher
7549 * threads are idle. When in lower SMT modes, the threads will
7550 * perform better since they share less core resources. Hence when we
7551 * have idle threads, we want them to be the higher ones.
7553 * This packing function is run on idle threads. It checks to see if
7554 * the busiest CPU in this domain (core in the P7 case) has a higher
7555 * CPU number than the packing function is being run on. Here we are
7556 * assuming lower CPU number will be equivalent to lower a SMT thread
7559 * Return: 1 when packing is required and a task should be moved to
7560 * this CPU. The amount of the imbalance is returned in *imbalance.
7562 * @env: The load balancing environment.
7563 * @sds: Statistics of the sched_domain which is to be packed
7565 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7569 if (!(env->sd->flags & SD_ASYM_PACKING))
7575 busiest_cpu = group_first_cpu(sds->busiest);
7576 if (env->dst_cpu > busiest_cpu)
7579 env->imbalance = DIV_ROUND_CLOSEST(
7580 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7581 SCHED_CAPACITY_SCALE);
7587 * fix_small_imbalance - Calculate the minor imbalance that exists
7588 * amongst the groups of a sched_domain, during
7590 * @env: The load balancing environment.
7591 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7594 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7596 unsigned long tmp, capa_now = 0, capa_move = 0;
7597 unsigned int imbn = 2;
7598 unsigned long scaled_busy_load_per_task;
7599 struct sg_lb_stats *local, *busiest;
7601 local = &sds->local_stat;
7602 busiest = &sds->busiest_stat;
7604 if (!local->sum_nr_running)
7605 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7606 else if (busiest->load_per_task > local->load_per_task)
7609 scaled_busy_load_per_task =
7610 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7611 busiest->group_capacity;
7613 if (busiest->avg_load + scaled_busy_load_per_task >=
7614 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7615 env->imbalance = busiest->load_per_task;
7620 * OK, we don't have enough imbalance to justify moving tasks,
7621 * however we may be able to increase total CPU capacity used by
7625 capa_now += busiest->group_capacity *
7626 min(busiest->load_per_task, busiest->avg_load);
7627 capa_now += local->group_capacity *
7628 min(local->load_per_task, local->avg_load);
7629 capa_now /= SCHED_CAPACITY_SCALE;
7631 /* Amount of load we'd subtract */
7632 if (busiest->avg_load > scaled_busy_load_per_task) {
7633 capa_move += busiest->group_capacity *
7634 min(busiest->load_per_task,
7635 busiest->avg_load - scaled_busy_load_per_task);
7638 /* Amount of load we'd add */
7639 if (busiest->avg_load * busiest->group_capacity <
7640 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7641 tmp = (busiest->avg_load * busiest->group_capacity) /
7642 local->group_capacity;
7644 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7645 local->group_capacity;
7647 capa_move += local->group_capacity *
7648 min(local->load_per_task, local->avg_load + tmp);
7649 capa_move /= SCHED_CAPACITY_SCALE;
7651 /* Move if we gain throughput */
7652 if (capa_move > capa_now)
7653 env->imbalance = busiest->load_per_task;
7657 * calculate_imbalance - Calculate the amount of imbalance present within the
7658 * groups of a given sched_domain during load balance.
7659 * @env: load balance environment
7660 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7662 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7664 unsigned long max_pull, load_above_capacity = ~0UL;
7665 struct sg_lb_stats *local, *busiest;
7667 local = &sds->local_stat;
7668 busiest = &sds->busiest_stat;
7670 if (busiest->group_type == group_imbalanced) {
7672 * In the group_imb case we cannot rely on group-wide averages
7673 * to ensure cpu-load equilibrium, look at wider averages. XXX
7675 busiest->load_per_task =
7676 min(busiest->load_per_task, sds->avg_load);
7680 * In the presence of smp nice balancing, certain scenarios can have
7681 * max load less than avg load(as we skip the groups at or below
7682 * its cpu_capacity, while calculating max_load..)
7684 if (busiest->avg_load <= sds->avg_load ||
7685 local->avg_load >= sds->avg_load) {
7686 /* Misfitting tasks should be migrated in any case */
7687 if (busiest->group_type == group_misfit_task) {
7688 env->imbalance = busiest->group_misfit_task;
7693 * Busiest group is overloaded, local is not, use the spare
7694 * cycles to maximize throughput
7696 if (busiest->group_type == group_overloaded &&
7697 local->group_type <= group_misfit_task) {
7698 env->imbalance = busiest->load_per_task;
7703 return fix_small_imbalance(env, sds);
7707 * If there aren't any idle cpus, avoid creating some.
7709 if (busiest->group_type == group_overloaded &&
7710 local->group_type == group_overloaded) {
7711 load_above_capacity = busiest->sum_nr_running *
7713 if (load_above_capacity > busiest->group_capacity)
7714 load_above_capacity -= busiest->group_capacity;
7716 load_above_capacity = ~0UL;
7720 * We're trying to get all the cpus to the average_load, so we don't
7721 * want to push ourselves above the average load, nor do we wish to
7722 * reduce the max loaded cpu below the average load. At the same time,
7723 * we also don't want to reduce the group load below the group capacity
7724 * (so that we can implement power-savings policies etc). Thus we look
7725 * for the minimum possible imbalance.
7727 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7729 /* How much load to actually move to equalise the imbalance */
7730 env->imbalance = min(
7731 max_pull * busiest->group_capacity,
7732 (sds->avg_load - local->avg_load) * local->group_capacity
7733 ) / SCHED_CAPACITY_SCALE;
7735 /* Boost imbalance to allow misfit task to be balanced. */
7736 if (busiest->group_type == group_misfit_task)
7737 env->imbalance = max_t(long, env->imbalance,
7738 busiest->group_misfit_task);
7741 * if *imbalance is less than the average load per runnable task
7742 * there is no guarantee that any tasks will be moved so we'll have
7743 * a think about bumping its value to force at least one task to be
7746 if (env->imbalance < busiest->load_per_task)
7747 return fix_small_imbalance(env, sds);
7750 /******* find_busiest_group() helpers end here *********************/
7753 * find_busiest_group - Returns the busiest group within the sched_domain
7754 * if there is an imbalance. If there isn't an imbalance, and
7755 * the user has opted for power-savings, it returns a group whose
7756 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7757 * such a group exists.
7759 * Also calculates the amount of weighted load which should be moved
7760 * to restore balance.
7762 * @env: The load balancing environment.
7764 * Return: - The busiest group if imbalance exists.
7765 * - If no imbalance and user has opted for power-savings balance,
7766 * return the least loaded group whose CPUs can be
7767 * put to idle by rebalancing its tasks onto our group.
7769 static struct sched_group *find_busiest_group(struct lb_env *env)
7771 struct sg_lb_stats *local, *busiest;
7772 struct sd_lb_stats sds;
7774 init_sd_lb_stats(&sds);
7777 * Compute the various statistics relavent for load balancing at
7780 update_sd_lb_stats(env, &sds);
7782 if (energy_aware() && !env->dst_rq->rd->overutilized)
7785 local = &sds.local_stat;
7786 busiest = &sds.busiest_stat;
7788 /* ASYM feature bypasses nice load balance check */
7789 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7790 check_asym_packing(env, &sds))
7793 /* There is no busy sibling group to pull tasks from */
7794 if (!sds.busiest || busiest->sum_nr_running == 0)
7797 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7798 / sds.total_capacity;
7801 * If the busiest group is imbalanced the below checks don't
7802 * work because they assume all things are equal, which typically
7803 * isn't true due to cpus_allowed constraints and the like.
7805 if (busiest->group_type == group_imbalanced)
7808 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7809 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7810 busiest->group_no_capacity)
7813 /* Misfitting tasks should be dealt with regardless of the avg load */
7814 if (busiest->group_type == group_misfit_task) {
7819 * If the local group is busier than the selected busiest group
7820 * don't try and pull any tasks.
7822 if (local->avg_load >= busiest->avg_load)
7826 * Don't pull any tasks if this group is already above the domain
7829 if (local->avg_load >= sds.avg_load)
7832 if (env->idle == CPU_IDLE) {
7834 * This cpu is idle. If the busiest group is not overloaded
7835 * and there is no imbalance between this and busiest group
7836 * wrt idle cpus, it is balanced. The imbalance becomes
7837 * significant if the diff is greater than 1 otherwise we
7838 * might end up to just move the imbalance on another group
7840 if ((busiest->group_type != group_overloaded) &&
7841 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7842 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7846 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7847 * imbalance_pct to be conservative.
7849 if (100 * busiest->avg_load <=
7850 env->sd->imbalance_pct * local->avg_load)
7855 env->busiest_group_type = busiest->group_type;
7856 /* Looks like there is an imbalance. Compute it */
7857 calculate_imbalance(env, &sds);
7866 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7868 static struct rq *find_busiest_queue(struct lb_env *env,
7869 struct sched_group *group)
7871 struct rq *busiest = NULL, *rq;
7872 unsigned long busiest_load = 0, busiest_capacity = 1;
7875 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7876 unsigned long capacity, wl;
7880 rt = fbq_classify_rq(rq);
7883 * We classify groups/runqueues into three groups:
7884 * - regular: there are !numa tasks
7885 * - remote: there are numa tasks that run on the 'wrong' node
7886 * - all: there is no distinction
7888 * In order to avoid migrating ideally placed numa tasks,
7889 * ignore those when there's better options.
7891 * If we ignore the actual busiest queue to migrate another
7892 * task, the next balance pass can still reduce the busiest
7893 * queue by moving tasks around inside the node.
7895 * If we cannot move enough load due to this classification
7896 * the next pass will adjust the group classification and
7897 * allow migration of more tasks.
7899 * Both cases only affect the total convergence complexity.
7901 if (rt > env->fbq_type)
7904 capacity = capacity_of(i);
7906 wl = weighted_cpuload(i);
7909 * When comparing with imbalance, use weighted_cpuload()
7910 * which is not scaled with the cpu capacity.
7913 if (rq->nr_running == 1 && wl > env->imbalance &&
7914 !check_cpu_capacity(rq, env->sd) &&
7915 env->busiest_group_type != group_misfit_task)
7919 * For the load comparisons with the other cpu's, consider
7920 * the weighted_cpuload() scaled with the cpu capacity, so
7921 * that the load can be moved away from the cpu that is
7922 * potentially running at a lower capacity.
7924 * Thus we're looking for max(wl_i / capacity_i), crosswise
7925 * multiplication to rid ourselves of the division works out
7926 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7927 * our previous maximum.
7929 if (wl * busiest_capacity > busiest_load * capacity) {
7931 busiest_capacity = capacity;
7940 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7941 * so long as it is large enough.
7943 #define MAX_PINNED_INTERVAL 512
7945 /* Working cpumask for load_balance and load_balance_newidle. */
7946 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7948 static int need_active_balance(struct lb_env *env)
7950 struct sched_domain *sd = env->sd;
7952 if (env->idle == CPU_NEWLY_IDLE) {
7955 * ASYM_PACKING needs to force migrate tasks from busy but
7956 * higher numbered CPUs in order to pack all tasks in the
7957 * lowest numbered CPUs.
7959 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7964 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7965 * It's worth migrating the task if the src_cpu's capacity is reduced
7966 * because of other sched_class or IRQs if more capacity stays
7967 * available on dst_cpu.
7969 if ((env->idle != CPU_NOT_IDLE) &&
7970 (env->src_rq->cfs.h_nr_running == 1)) {
7971 if ((check_cpu_capacity(env->src_rq, sd)) &&
7972 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7976 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
7977 env->src_rq->cfs.h_nr_running == 1 &&
7978 cpu_overutilized(env->src_cpu) &&
7979 !cpu_overutilized(env->dst_cpu)) {
7983 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7986 static int active_load_balance_cpu_stop(void *data);
7988 static int should_we_balance(struct lb_env *env)
7990 struct sched_group *sg = env->sd->groups;
7991 struct cpumask *sg_cpus, *sg_mask;
7992 int cpu, balance_cpu = -1;
7995 * In the newly idle case, we will allow all the cpu's
7996 * to do the newly idle load balance.
7998 if (env->idle == CPU_NEWLY_IDLE)
8001 sg_cpus = sched_group_cpus(sg);
8002 sg_mask = sched_group_mask(sg);
8003 /* Try to find first idle cpu */
8004 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8005 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8012 if (balance_cpu == -1)
8013 balance_cpu = group_balance_cpu(sg);
8016 * First idle cpu or the first cpu(busiest) in this sched group
8017 * is eligible for doing load balancing at this and above domains.
8019 return balance_cpu == env->dst_cpu;
8023 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8024 * tasks if there is an imbalance.
8026 static int load_balance(int this_cpu, struct rq *this_rq,
8027 struct sched_domain *sd, enum cpu_idle_type idle,
8028 int *continue_balancing)
8030 int ld_moved, cur_ld_moved, active_balance = 0;
8031 struct sched_domain *sd_parent = sd->parent;
8032 struct sched_group *group;
8034 unsigned long flags;
8035 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8037 struct lb_env env = {
8039 .dst_cpu = this_cpu,
8041 .dst_grpmask = sched_group_cpus(sd->groups),
8043 .loop_break = sched_nr_migrate_break,
8046 .tasks = LIST_HEAD_INIT(env.tasks),
8050 * For NEWLY_IDLE load_balancing, we don't need to consider
8051 * other cpus in our group
8053 if (idle == CPU_NEWLY_IDLE)
8054 env.dst_grpmask = NULL;
8056 cpumask_copy(cpus, cpu_active_mask);
8058 schedstat_inc(sd, lb_count[idle]);
8061 if (!should_we_balance(&env)) {
8062 *continue_balancing = 0;
8066 group = find_busiest_group(&env);
8068 schedstat_inc(sd, lb_nobusyg[idle]);
8072 busiest = find_busiest_queue(&env, group);
8074 schedstat_inc(sd, lb_nobusyq[idle]);
8078 BUG_ON(busiest == env.dst_rq);
8080 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8082 env.src_cpu = busiest->cpu;
8083 env.src_rq = busiest;
8086 if (busiest->nr_running > 1) {
8088 * Attempt to move tasks. If find_busiest_group has found
8089 * an imbalance but busiest->nr_running <= 1, the group is
8090 * still unbalanced. ld_moved simply stays zero, so it is
8091 * correctly treated as an imbalance.
8093 env.flags |= LBF_ALL_PINNED;
8094 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8097 raw_spin_lock_irqsave(&busiest->lock, flags);
8100 * cur_ld_moved - load moved in current iteration
8101 * ld_moved - cumulative load moved across iterations
8103 cur_ld_moved = detach_tasks(&env);
8105 * We want to potentially lower env.src_cpu's OPP.
8108 update_capacity_of(env.src_cpu);
8111 * We've detached some tasks from busiest_rq. Every
8112 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8113 * unlock busiest->lock, and we are able to be sure
8114 * that nobody can manipulate the tasks in parallel.
8115 * See task_rq_lock() family for the details.
8118 raw_spin_unlock(&busiest->lock);
8122 ld_moved += cur_ld_moved;
8125 local_irq_restore(flags);
8127 if (env.flags & LBF_NEED_BREAK) {
8128 env.flags &= ~LBF_NEED_BREAK;
8133 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8134 * us and move them to an alternate dst_cpu in our sched_group
8135 * where they can run. The upper limit on how many times we
8136 * iterate on same src_cpu is dependent on number of cpus in our
8139 * This changes load balance semantics a bit on who can move
8140 * load to a given_cpu. In addition to the given_cpu itself
8141 * (or a ilb_cpu acting on its behalf where given_cpu is
8142 * nohz-idle), we now have balance_cpu in a position to move
8143 * load to given_cpu. In rare situations, this may cause
8144 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8145 * _independently_ and at _same_ time to move some load to
8146 * given_cpu) causing exceess load to be moved to given_cpu.
8147 * This however should not happen so much in practice and
8148 * moreover subsequent load balance cycles should correct the
8149 * excess load moved.
8151 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8153 /* Prevent to re-select dst_cpu via env's cpus */
8154 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8156 env.dst_rq = cpu_rq(env.new_dst_cpu);
8157 env.dst_cpu = env.new_dst_cpu;
8158 env.flags &= ~LBF_DST_PINNED;
8160 env.loop_break = sched_nr_migrate_break;
8163 * Go back to "more_balance" rather than "redo" since we
8164 * need to continue with same src_cpu.
8170 * We failed to reach balance because of affinity.
8173 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8175 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8176 *group_imbalance = 1;
8179 /* All tasks on this runqueue were pinned by CPU affinity */
8180 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8181 cpumask_clear_cpu(cpu_of(busiest), cpus);
8182 if (!cpumask_empty(cpus)) {
8184 env.loop_break = sched_nr_migrate_break;
8187 goto out_all_pinned;
8192 schedstat_inc(sd, lb_failed[idle]);
8194 * Increment the failure counter only on periodic balance.
8195 * We do not want newidle balance, which can be very
8196 * frequent, pollute the failure counter causing
8197 * excessive cache_hot migrations and active balances.
8199 if (idle != CPU_NEWLY_IDLE)
8200 if (env.src_grp_nr_running > 1)
8201 sd->nr_balance_failed++;
8203 if (need_active_balance(&env)) {
8204 raw_spin_lock_irqsave(&busiest->lock, flags);
8206 /* don't kick the active_load_balance_cpu_stop,
8207 * if the curr task on busiest cpu can't be
8210 if (!cpumask_test_cpu(this_cpu,
8211 tsk_cpus_allowed(busiest->curr))) {
8212 raw_spin_unlock_irqrestore(&busiest->lock,
8214 env.flags |= LBF_ALL_PINNED;
8215 goto out_one_pinned;
8219 * ->active_balance synchronizes accesses to
8220 * ->active_balance_work. Once set, it's cleared
8221 * only after active load balance is finished.
8223 if (!busiest->active_balance) {
8224 busiest->active_balance = 1;
8225 busiest->push_cpu = this_cpu;
8228 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8230 if (active_balance) {
8231 stop_one_cpu_nowait(cpu_of(busiest),
8232 active_load_balance_cpu_stop, busiest,
8233 &busiest->active_balance_work);
8237 * We've kicked active balancing, reset the failure
8240 sd->nr_balance_failed = sd->cache_nice_tries+1;
8243 sd->nr_balance_failed = 0;
8245 if (likely(!active_balance)) {
8246 /* We were unbalanced, so reset the balancing interval */
8247 sd->balance_interval = sd->min_interval;
8250 * If we've begun active balancing, start to back off. This
8251 * case may not be covered by the all_pinned logic if there
8252 * is only 1 task on the busy runqueue (because we don't call
8255 if (sd->balance_interval < sd->max_interval)
8256 sd->balance_interval *= 2;
8263 * We reach balance although we may have faced some affinity
8264 * constraints. Clear the imbalance flag if it was set.
8267 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8269 if (*group_imbalance)
8270 *group_imbalance = 0;
8275 * We reach balance because all tasks are pinned at this level so
8276 * we can't migrate them. Let the imbalance flag set so parent level
8277 * can try to migrate them.
8279 schedstat_inc(sd, lb_balanced[idle]);
8281 sd->nr_balance_failed = 0;
8284 /* tune up the balancing interval */
8285 if (((env.flags & LBF_ALL_PINNED) &&
8286 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8287 (sd->balance_interval < sd->max_interval))
8288 sd->balance_interval *= 2;
8295 static inline unsigned long
8296 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8298 unsigned long interval = sd->balance_interval;
8301 interval *= sd->busy_factor;
8303 /* scale ms to jiffies */
8304 interval = msecs_to_jiffies(interval);
8305 interval = clamp(interval, 1UL, max_load_balance_interval);
8311 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8313 unsigned long interval, next;
8315 interval = get_sd_balance_interval(sd, cpu_busy);
8316 next = sd->last_balance + interval;
8318 if (time_after(*next_balance, next))
8319 *next_balance = next;
8323 * idle_balance is called by schedule() if this_cpu is about to become
8324 * idle. Attempts to pull tasks from other CPUs.
8326 static int idle_balance(struct rq *this_rq)
8328 unsigned long next_balance = jiffies + HZ;
8329 int this_cpu = this_rq->cpu;
8330 struct sched_domain *sd;
8331 int pulled_task = 0;
8334 idle_enter_fair(this_rq);
8337 * We must set idle_stamp _before_ calling idle_balance(), such that we
8338 * measure the duration of idle_balance() as idle time.
8340 this_rq->idle_stamp = rq_clock(this_rq);
8342 if (!energy_aware() &&
8343 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8344 !this_rq->rd->overload)) {
8346 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8348 update_next_balance(sd, 0, &next_balance);
8354 raw_spin_unlock(&this_rq->lock);
8356 update_blocked_averages(this_cpu);
8358 for_each_domain(this_cpu, sd) {
8359 int continue_balancing = 1;
8360 u64 t0, domain_cost;
8362 if (!(sd->flags & SD_LOAD_BALANCE))
8365 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8366 update_next_balance(sd, 0, &next_balance);
8370 if (sd->flags & SD_BALANCE_NEWIDLE) {
8371 t0 = sched_clock_cpu(this_cpu);
8373 pulled_task = load_balance(this_cpu, this_rq,
8375 &continue_balancing);
8377 domain_cost = sched_clock_cpu(this_cpu) - t0;
8378 if (domain_cost > sd->max_newidle_lb_cost)
8379 sd->max_newidle_lb_cost = domain_cost;
8381 curr_cost += domain_cost;
8384 update_next_balance(sd, 0, &next_balance);
8387 * Stop searching for tasks to pull if there are
8388 * now runnable tasks on this rq.
8390 if (pulled_task || this_rq->nr_running > 0)
8395 raw_spin_lock(&this_rq->lock);
8397 if (curr_cost > this_rq->max_idle_balance_cost)
8398 this_rq->max_idle_balance_cost = curr_cost;
8401 * While browsing the domains, we released the rq lock, a task could
8402 * have been enqueued in the meantime. Since we're not going idle,
8403 * pretend we pulled a task.
8405 if (this_rq->cfs.h_nr_running && !pulled_task)
8409 /* Move the next balance forward */
8410 if (time_after(this_rq->next_balance, next_balance))
8411 this_rq->next_balance = next_balance;
8413 /* Is there a task of a high priority class? */
8414 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8418 idle_exit_fair(this_rq);
8419 this_rq->idle_stamp = 0;
8426 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8427 * running tasks off the busiest CPU onto idle CPUs. It requires at
8428 * least 1 task to be running on each physical CPU where possible, and
8429 * avoids physical / logical imbalances.
8431 static int active_load_balance_cpu_stop(void *data)
8433 struct rq *busiest_rq = data;
8434 int busiest_cpu = cpu_of(busiest_rq);
8435 int target_cpu = busiest_rq->push_cpu;
8436 struct rq *target_rq = cpu_rq(target_cpu);
8437 struct sched_domain *sd;
8438 struct task_struct *p = NULL;
8440 raw_spin_lock_irq(&busiest_rq->lock);
8442 /* make sure the requested cpu hasn't gone down in the meantime */
8443 if (unlikely(busiest_cpu != smp_processor_id() ||
8444 !busiest_rq->active_balance))
8447 /* Is there any task to move? */
8448 if (busiest_rq->nr_running <= 1)
8452 * This condition is "impossible", if it occurs
8453 * we need to fix it. Originally reported by
8454 * Bjorn Helgaas on a 128-cpu setup.
8456 BUG_ON(busiest_rq == target_rq);
8458 /* Search for an sd spanning us and the target CPU. */
8460 for_each_domain(target_cpu, sd) {
8461 if ((sd->flags & SD_LOAD_BALANCE) &&
8462 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8467 struct lb_env env = {
8469 .dst_cpu = target_cpu,
8470 .dst_rq = target_rq,
8471 .src_cpu = busiest_rq->cpu,
8472 .src_rq = busiest_rq,
8476 schedstat_inc(sd, alb_count);
8478 p = detach_one_task(&env);
8480 schedstat_inc(sd, alb_pushed);
8482 * We want to potentially lower env.src_cpu's OPP.
8484 update_capacity_of(env.src_cpu);
8487 schedstat_inc(sd, alb_failed);
8491 busiest_rq->active_balance = 0;
8492 raw_spin_unlock(&busiest_rq->lock);
8495 attach_one_task(target_rq, p);
8502 static inline int on_null_domain(struct rq *rq)
8504 return unlikely(!rcu_dereference_sched(rq->sd));
8507 #ifdef CONFIG_NO_HZ_COMMON
8509 * idle load balancing details
8510 * - When one of the busy CPUs notice that there may be an idle rebalancing
8511 * needed, they will kick the idle load balancer, which then does idle
8512 * load balancing for all the idle CPUs.
8515 cpumask_var_t idle_cpus_mask;
8517 unsigned long next_balance; /* in jiffy units */
8518 } nohz ____cacheline_aligned;
8520 static inline int find_new_ilb(void)
8522 int ilb = cpumask_first(nohz.idle_cpus_mask);
8524 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8531 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8532 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8533 * CPU (if there is one).
8535 static void nohz_balancer_kick(void)
8539 nohz.next_balance++;
8541 ilb_cpu = find_new_ilb();
8543 if (ilb_cpu >= nr_cpu_ids)
8546 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8549 * Use smp_send_reschedule() instead of resched_cpu().
8550 * This way we generate a sched IPI on the target cpu which
8551 * is idle. And the softirq performing nohz idle load balance
8552 * will be run before returning from the IPI.
8554 smp_send_reschedule(ilb_cpu);
8558 static inline void nohz_balance_exit_idle(int cpu)
8560 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8562 * Completely isolated CPUs don't ever set, so we must test.
8564 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8565 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8566 atomic_dec(&nohz.nr_cpus);
8568 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8572 static inline void set_cpu_sd_state_busy(void)
8574 struct sched_domain *sd;
8575 int cpu = smp_processor_id();
8578 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8580 if (!sd || !sd->nohz_idle)
8584 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8589 void set_cpu_sd_state_idle(void)
8591 struct sched_domain *sd;
8592 int cpu = smp_processor_id();
8595 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8597 if (!sd || sd->nohz_idle)
8601 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8607 * This routine will record that the cpu is going idle with tick stopped.
8608 * This info will be used in performing idle load balancing in the future.
8610 void nohz_balance_enter_idle(int cpu)
8613 * If this cpu is going down, then nothing needs to be done.
8615 if (!cpu_active(cpu))
8618 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8622 * If we're a completely isolated CPU, we don't play.
8624 if (on_null_domain(cpu_rq(cpu)))
8627 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8628 atomic_inc(&nohz.nr_cpus);
8629 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8632 static int sched_ilb_notifier(struct notifier_block *nfb,
8633 unsigned long action, void *hcpu)
8635 switch (action & ~CPU_TASKS_FROZEN) {
8637 nohz_balance_exit_idle(smp_processor_id());
8645 static DEFINE_SPINLOCK(balancing);
8648 * Scale the max load_balance interval with the number of CPUs in the system.
8649 * This trades load-balance latency on larger machines for less cross talk.
8651 void update_max_interval(void)
8653 max_load_balance_interval = HZ*num_online_cpus()/10;
8657 * It checks each scheduling domain to see if it is due to be balanced,
8658 * and initiates a balancing operation if so.
8660 * Balancing parameters are set up in init_sched_domains.
8662 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8664 int continue_balancing = 1;
8666 unsigned long interval;
8667 struct sched_domain *sd;
8668 /* Earliest time when we have to do rebalance again */
8669 unsigned long next_balance = jiffies + 60*HZ;
8670 int update_next_balance = 0;
8671 int need_serialize, need_decay = 0;
8674 update_blocked_averages(cpu);
8677 for_each_domain(cpu, sd) {
8679 * Decay the newidle max times here because this is a regular
8680 * visit to all the domains. Decay ~1% per second.
8682 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8683 sd->max_newidle_lb_cost =
8684 (sd->max_newidle_lb_cost * 253) / 256;
8685 sd->next_decay_max_lb_cost = jiffies + HZ;
8688 max_cost += sd->max_newidle_lb_cost;
8690 if (!(sd->flags & SD_LOAD_BALANCE))
8694 * Stop the load balance at this level. There is another
8695 * CPU in our sched group which is doing load balancing more
8698 if (!continue_balancing) {
8704 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8706 need_serialize = sd->flags & SD_SERIALIZE;
8707 if (need_serialize) {
8708 if (!spin_trylock(&balancing))
8712 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8713 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8715 * The LBF_DST_PINNED logic could have changed
8716 * env->dst_cpu, so we can't know our idle
8717 * state even if we migrated tasks. Update it.
8719 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8721 sd->last_balance = jiffies;
8722 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8725 spin_unlock(&balancing);
8727 if (time_after(next_balance, sd->last_balance + interval)) {
8728 next_balance = sd->last_balance + interval;
8729 update_next_balance = 1;
8734 * Ensure the rq-wide value also decays but keep it at a
8735 * reasonable floor to avoid funnies with rq->avg_idle.
8737 rq->max_idle_balance_cost =
8738 max((u64)sysctl_sched_migration_cost, max_cost);
8743 * next_balance will be updated only when there is a need.
8744 * When the cpu is attached to null domain for ex, it will not be
8747 if (likely(update_next_balance)) {
8748 rq->next_balance = next_balance;
8750 #ifdef CONFIG_NO_HZ_COMMON
8752 * If this CPU has been elected to perform the nohz idle
8753 * balance. Other idle CPUs have already rebalanced with
8754 * nohz_idle_balance() and nohz.next_balance has been
8755 * updated accordingly. This CPU is now running the idle load
8756 * balance for itself and we need to update the
8757 * nohz.next_balance accordingly.
8759 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8760 nohz.next_balance = rq->next_balance;
8765 #ifdef CONFIG_NO_HZ_COMMON
8767 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8768 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8770 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8772 int this_cpu = this_rq->cpu;
8775 /* Earliest time when we have to do rebalance again */
8776 unsigned long next_balance = jiffies + 60*HZ;
8777 int update_next_balance = 0;
8779 if (idle != CPU_IDLE ||
8780 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8783 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8784 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8788 * If this cpu gets work to do, stop the load balancing
8789 * work being done for other cpus. Next load
8790 * balancing owner will pick it up.
8795 rq = cpu_rq(balance_cpu);
8798 * If time for next balance is due,
8801 if (time_after_eq(jiffies, rq->next_balance)) {
8802 raw_spin_lock_irq(&rq->lock);
8803 update_rq_clock(rq);
8804 update_idle_cpu_load(rq);
8805 raw_spin_unlock_irq(&rq->lock);
8806 rebalance_domains(rq, CPU_IDLE);
8809 if (time_after(next_balance, rq->next_balance)) {
8810 next_balance = rq->next_balance;
8811 update_next_balance = 1;
8816 * next_balance will be updated only when there is a need.
8817 * When the CPU is attached to null domain for ex, it will not be
8820 if (likely(update_next_balance))
8821 nohz.next_balance = next_balance;
8823 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8827 * Current heuristic for kicking the idle load balancer in the presence
8828 * of an idle cpu in the system.
8829 * - This rq has more than one task.
8830 * - This rq has at least one CFS task and the capacity of the CPU is
8831 * significantly reduced because of RT tasks or IRQs.
8832 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8833 * multiple busy cpu.
8834 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8835 * domain span are idle.
8837 static inline bool nohz_kick_needed(struct rq *rq)
8839 unsigned long now = jiffies;
8840 struct sched_domain *sd;
8841 struct sched_group_capacity *sgc;
8842 int nr_busy, cpu = rq->cpu;
8845 if (unlikely(rq->idle_balance))
8849 * We may be recently in ticked or tickless idle mode. At the first
8850 * busy tick after returning from idle, we will update the busy stats.
8852 set_cpu_sd_state_busy();
8853 nohz_balance_exit_idle(cpu);
8856 * None are in tickless mode and hence no need for NOHZ idle load
8859 if (likely(!atomic_read(&nohz.nr_cpus)))
8862 if (time_before(now, nohz.next_balance))
8865 if (rq->nr_running >= 2 &&
8866 (!energy_aware() || cpu_overutilized(cpu)))
8870 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8871 if (sd && !energy_aware()) {
8872 sgc = sd->groups->sgc;
8873 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8882 sd = rcu_dereference(rq->sd);
8884 if ((rq->cfs.h_nr_running >= 1) &&
8885 check_cpu_capacity(rq, sd)) {
8891 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8892 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8893 sched_domain_span(sd)) < cpu)) {
8903 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8907 * run_rebalance_domains is triggered when needed from the scheduler tick.
8908 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8910 static void run_rebalance_domains(struct softirq_action *h)
8912 struct rq *this_rq = this_rq();
8913 enum cpu_idle_type idle = this_rq->idle_balance ?
8914 CPU_IDLE : CPU_NOT_IDLE;
8917 * If this cpu has a pending nohz_balance_kick, then do the
8918 * balancing on behalf of the other idle cpus whose ticks are
8919 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8920 * give the idle cpus a chance to load balance. Else we may
8921 * load balance only within the local sched_domain hierarchy
8922 * and abort nohz_idle_balance altogether if we pull some load.
8924 nohz_idle_balance(this_rq, idle);
8925 rebalance_domains(this_rq, idle);
8929 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8931 void trigger_load_balance(struct rq *rq)
8933 /* Don't need to rebalance while attached to NULL domain */
8934 if (unlikely(on_null_domain(rq)))
8937 if (time_after_eq(jiffies, rq->next_balance))
8938 raise_softirq(SCHED_SOFTIRQ);
8939 #ifdef CONFIG_NO_HZ_COMMON
8940 if (nohz_kick_needed(rq))
8941 nohz_balancer_kick();
8945 static void rq_online_fair(struct rq *rq)
8949 update_runtime_enabled(rq);
8952 static void rq_offline_fair(struct rq *rq)
8956 /* Ensure any throttled groups are reachable by pick_next_task */
8957 unthrottle_offline_cfs_rqs(rq);
8960 #endif /* CONFIG_SMP */
8963 * scheduler tick hitting a task of our scheduling class:
8965 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8967 struct cfs_rq *cfs_rq;
8968 struct sched_entity *se = &curr->se;
8970 for_each_sched_entity(se) {
8971 cfs_rq = cfs_rq_of(se);
8972 entity_tick(cfs_rq, se, queued);
8975 if (static_branch_unlikely(&sched_numa_balancing))
8976 task_tick_numa(rq, curr);
8979 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
8980 rq->rd->overutilized = true;
8981 trace_sched_overutilized(true);
8984 rq->misfit_task = !task_fits_max(curr, rq->cpu);
8990 * called on fork with the child task as argument from the parent's context
8991 * - child not yet on the tasklist
8992 * - preemption disabled
8994 static void task_fork_fair(struct task_struct *p)
8996 struct cfs_rq *cfs_rq;
8997 struct sched_entity *se = &p->se, *curr;
8998 int this_cpu = smp_processor_id();
8999 struct rq *rq = this_rq();
9000 unsigned long flags;
9002 raw_spin_lock_irqsave(&rq->lock, flags);
9004 update_rq_clock(rq);
9006 cfs_rq = task_cfs_rq(current);
9007 curr = cfs_rq->curr;
9010 * Not only the cpu but also the task_group of the parent might have
9011 * been changed after parent->se.parent,cfs_rq were copied to
9012 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9013 * of child point to valid ones.
9016 __set_task_cpu(p, this_cpu);
9019 update_curr(cfs_rq);
9022 se->vruntime = curr->vruntime;
9023 place_entity(cfs_rq, se, 1);
9025 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9027 * Upon rescheduling, sched_class::put_prev_task() will place
9028 * 'current' within the tree based on its new key value.
9030 swap(curr->vruntime, se->vruntime);
9034 se->vruntime -= cfs_rq->min_vruntime;
9036 raw_spin_unlock_irqrestore(&rq->lock, flags);
9040 * Priority of the task has changed. Check to see if we preempt
9044 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9046 if (!task_on_rq_queued(p))
9050 * Reschedule if we are currently running on this runqueue and
9051 * our priority decreased, or if we are not currently running on
9052 * this runqueue and our priority is higher than the current's
9054 if (rq->curr == p) {
9055 if (p->prio > oldprio)
9058 check_preempt_curr(rq, p, 0);
9061 static inline bool vruntime_normalized(struct task_struct *p)
9063 struct sched_entity *se = &p->se;
9066 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9067 * the dequeue_entity(.flags=0) will already have normalized the
9074 * When !on_rq, vruntime of the task has usually NOT been normalized.
9075 * But there are some cases where it has already been normalized:
9077 * - A forked child which is waiting for being woken up by
9078 * wake_up_new_task().
9079 * - A task which has been woken up by try_to_wake_up() and
9080 * waiting for actually being woken up by sched_ttwu_pending().
9082 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9088 static void detach_task_cfs_rq(struct task_struct *p)
9090 struct sched_entity *se = &p->se;
9091 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9093 if (!vruntime_normalized(p)) {
9095 * Fix up our vruntime so that the current sleep doesn't
9096 * cause 'unlimited' sleep bonus.
9098 place_entity(cfs_rq, se, 0);
9099 se->vruntime -= cfs_rq->min_vruntime;
9102 /* Catch up with the cfs_rq and remove our load when we leave */
9103 detach_entity_load_avg(cfs_rq, se);
9106 static void attach_task_cfs_rq(struct task_struct *p)
9108 struct sched_entity *se = &p->se;
9109 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9111 #ifdef CONFIG_FAIR_GROUP_SCHED
9113 * Since the real-depth could have been changed (only FAIR
9114 * class maintain depth value), reset depth properly.
9116 se->depth = se->parent ? se->parent->depth + 1 : 0;
9119 /* Synchronize task with its cfs_rq */
9120 attach_entity_load_avg(cfs_rq, se);
9122 if (!vruntime_normalized(p))
9123 se->vruntime += cfs_rq->min_vruntime;
9126 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9128 detach_task_cfs_rq(p);
9131 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9133 attach_task_cfs_rq(p);
9135 if (task_on_rq_queued(p)) {
9137 * We were most likely switched from sched_rt, so
9138 * kick off the schedule if running, otherwise just see
9139 * if we can still preempt the current task.
9144 check_preempt_curr(rq, p, 0);
9148 /* Account for a task changing its policy or group.
9150 * This routine is mostly called to set cfs_rq->curr field when a task
9151 * migrates between groups/classes.
9153 static void set_curr_task_fair(struct rq *rq)
9155 struct sched_entity *se = &rq->curr->se;
9157 for_each_sched_entity(se) {
9158 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9160 set_next_entity(cfs_rq, se);
9161 /* ensure bandwidth has been allocated on our new cfs_rq */
9162 account_cfs_rq_runtime(cfs_rq, 0);
9166 void init_cfs_rq(struct cfs_rq *cfs_rq)
9168 cfs_rq->tasks_timeline = RB_ROOT;
9169 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9170 #ifndef CONFIG_64BIT
9171 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9174 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9175 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9179 #ifdef CONFIG_FAIR_GROUP_SCHED
9180 static void task_move_group_fair(struct task_struct *p)
9182 detach_task_cfs_rq(p);
9183 set_task_rq(p, task_cpu(p));
9186 /* Tell se's cfs_rq has been changed -- migrated */
9187 p->se.avg.last_update_time = 0;
9189 attach_task_cfs_rq(p);
9192 void free_fair_sched_group(struct task_group *tg)
9196 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9198 for_each_possible_cpu(i) {
9200 kfree(tg->cfs_rq[i]);
9203 remove_entity_load_avg(tg->se[i]);
9212 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9214 struct cfs_rq *cfs_rq;
9215 struct sched_entity *se;
9218 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9221 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9225 tg->shares = NICE_0_LOAD;
9227 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9229 for_each_possible_cpu(i) {
9230 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9231 GFP_KERNEL, cpu_to_node(i));
9235 se = kzalloc_node(sizeof(struct sched_entity),
9236 GFP_KERNEL, cpu_to_node(i));
9240 init_cfs_rq(cfs_rq);
9241 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9242 init_entity_runnable_average(se);
9253 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9255 struct rq *rq = cpu_rq(cpu);
9256 unsigned long flags;
9259 * Only empty task groups can be destroyed; so we can speculatively
9260 * check on_list without danger of it being re-added.
9262 if (!tg->cfs_rq[cpu]->on_list)
9265 raw_spin_lock_irqsave(&rq->lock, flags);
9266 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9267 raw_spin_unlock_irqrestore(&rq->lock, flags);
9270 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9271 struct sched_entity *se, int cpu,
9272 struct sched_entity *parent)
9274 struct rq *rq = cpu_rq(cpu);
9278 init_cfs_rq_runtime(cfs_rq);
9280 tg->cfs_rq[cpu] = cfs_rq;
9283 /* se could be NULL for root_task_group */
9288 se->cfs_rq = &rq->cfs;
9291 se->cfs_rq = parent->my_q;
9292 se->depth = parent->depth + 1;
9296 /* guarantee group entities always have weight */
9297 update_load_set(&se->load, NICE_0_LOAD);
9298 se->parent = parent;
9301 static DEFINE_MUTEX(shares_mutex);
9303 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9306 unsigned long flags;
9309 * We can't change the weight of the root cgroup.
9314 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9316 mutex_lock(&shares_mutex);
9317 if (tg->shares == shares)
9320 tg->shares = shares;
9321 for_each_possible_cpu(i) {
9322 struct rq *rq = cpu_rq(i);
9323 struct sched_entity *se;
9326 /* Propagate contribution to hierarchy */
9327 raw_spin_lock_irqsave(&rq->lock, flags);
9329 /* Possible calls to update_curr() need rq clock */
9330 update_rq_clock(rq);
9331 for_each_sched_entity(se)
9332 update_cfs_shares(group_cfs_rq(se));
9333 raw_spin_unlock_irqrestore(&rq->lock, flags);
9337 mutex_unlock(&shares_mutex);
9340 #else /* CONFIG_FAIR_GROUP_SCHED */
9342 void free_fair_sched_group(struct task_group *tg) { }
9344 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9349 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9351 #endif /* CONFIG_FAIR_GROUP_SCHED */
9354 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9356 struct sched_entity *se = &task->se;
9357 unsigned int rr_interval = 0;
9360 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9363 if (rq->cfs.load.weight)
9364 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9370 * All the scheduling class methods:
9372 const struct sched_class fair_sched_class = {
9373 .next = &idle_sched_class,
9374 .enqueue_task = enqueue_task_fair,
9375 .dequeue_task = dequeue_task_fair,
9376 .yield_task = yield_task_fair,
9377 .yield_to_task = yield_to_task_fair,
9379 .check_preempt_curr = check_preempt_wakeup,
9381 .pick_next_task = pick_next_task_fair,
9382 .put_prev_task = put_prev_task_fair,
9385 .select_task_rq = select_task_rq_fair,
9386 .migrate_task_rq = migrate_task_rq_fair,
9388 .rq_online = rq_online_fair,
9389 .rq_offline = rq_offline_fair,
9391 .task_waking = task_waking_fair,
9392 .task_dead = task_dead_fair,
9393 .set_cpus_allowed = set_cpus_allowed_common,
9396 .set_curr_task = set_curr_task_fair,
9397 .task_tick = task_tick_fair,
9398 .task_fork = task_fork_fair,
9400 .prio_changed = prio_changed_fair,
9401 .switched_from = switched_from_fair,
9402 .switched_to = switched_to_fair,
9404 .get_rr_interval = get_rr_interval_fair,
9406 .update_curr = update_curr_fair,
9408 #ifdef CONFIG_FAIR_GROUP_SCHED
9409 .task_move_group = task_move_group_fair,
9413 #ifdef CONFIG_SCHED_DEBUG
9414 void print_cfs_stats(struct seq_file *m, int cpu)
9416 struct cfs_rq *cfs_rq;
9419 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9420 print_cfs_rq(m, cpu, cfs_rq);
9424 #ifdef CONFIG_NUMA_BALANCING
9425 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9428 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9430 for_each_online_node(node) {
9431 if (p->numa_faults) {
9432 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9433 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9435 if (p->numa_group) {
9436 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9437 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9439 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9442 #endif /* CONFIG_NUMA_BALANCING */
9443 #endif /* CONFIG_SCHED_DEBUG */
9445 __init void init_sched_fair_class(void)
9448 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9450 #ifdef CONFIG_NO_HZ_COMMON
9451 nohz.next_balance = jiffies;
9452 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9453 cpu_notifier(sched_ilb_notifier, 0);