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);
1212 env->best_imp = imp;
1213 env->best_cpu = env->dst_cpu;
1216 static bool load_too_imbalanced(long src_load, long dst_load,
1217 struct task_numa_env *env)
1220 long orig_src_load, orig_dst_load;
1221 long src_capacity, dst_capacity;
1224 * The load is corrected for the CPU capacity available on each node.
1227 * ------------ vs ---------
1228 * src_capacity dst_capacity
1230 src_capacity = env->src_stats.compute_capacity;
1231 dst_capacity = env->dst_stats.compute_capacity;
1233 /* We care about the slope of the imbalance, not the direction. */
1234 if (dst_load < src_load)
1235 swap(dst_load, src_load);
1237 /* Is the difference below the threshold? */
1238 imb = dst_load * src_capacity * 100 -
1239 src_load * dst_capacity * env->imbalance_pct;
1244 * The imbalance is above the allowed threshold.
1245 * Compare it with the old imbalance.
1247 orig_src_load = env->src_stats.load;
1248 orig_dst_load = env->dst_stats.load;
1250 if (orig_dst_load < orig_src_load)
1251 swap(orig_dst_load, orig_src_load);
1253 old_imb = orig_dst_load * src_capacity * 100 -
1254 orig_src_load * dst_capacity * env->imbalance_pct;
1256 /* Would this change make things worse? */
1257 return (imb > old_imb);
1261 * This checks if the overall compute and NUMA accesses of the system would
1262 * be improved if the source tasks was migrated to the target dst_cpu taking
1263 * into account that it might be best if task running on the dst_cpu should
1264 * be exchanged with the source task
1266 static void task_numa_compare(struct task_numa_env *env,
1267 long taskimp, long groupimp)
1269 struct rq *src_rq = cpu_rq(env->src_cpu);
1270 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1271 struct task_struct *cur;
1272 long src_load, dst_load;
1274 long imp = env->p->numa_group ? groupimp : taskimp;
1276 int dist = env->dist;
1277 bool assigned = false;
1281 raw_spin_lock_irq(&dst_rq->lock);
1284 * No need to move the exiting task or idle task.
1286 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1290 * The task_struct must be protected here to protect the
1291 * p->numa_faults access in the task_weight since the
1292 * numa_faults could already be freed in the following path:
1293 * finish_task_switch()
1294 * --> put_task_struct()
1295 * --> __put_task_struct()
1296 * --> task_numa_free()
1298 get_task_struct(cur);
1301 raw_spin_unlock_irq(&dst_rq->lock);
1304 * Because we have preemption enabled we can get migrated around and
1305 * end try selecting ourselves (current == env->p) as a swap candidate.
1311 * "imp" is the fault differential for the source task between the
1312 * source and destination node. Calculate the total differential for
1313 * the source task and potential destination task. The more negative
1314 * the value is, the more rmeote accesses that would be expected to
1315 * be incurred if the tasks were swapped.
1318 /* Skip this swap candidate if cannot move to the source cpu */
1319 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1323 * If dst and source tasks are in the same NUMA group, or not
1324 * in any group then look only at task weights.
1326 if (cur->numa_group == env->p->numa_group) {
1327 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1328 task_weight(cur, env->dst_nid, dist);
1330 * Add some hysteresis to prevent swapping the
1331 * tasks within a group over tiny differences.
1333 if (cur->numa_group)
1337 * Compare the group weights. If a task is all by
1338 * itself (not part of a group), use the task weight
1341 if (cur->numa_group)
1342 imp += group_weight(cur, env->src_nid, dist) -
1343 group_weight(cur, env->dst_nid, dist);
1345 imp += task_weight(cur, env->src_nid, dist) -
1346 task_weight(cur, env->dst_nid, dist);
1350 if (imp <= env->best_imp && moveimp <= env->best_imp)
1354 /* Is there capacity at our destination? */
1355 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1356 !env->dst_stats.has_free_capacity)
1362 /* Balance doesn't matter much if we're running a task per cpu */
1363 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1364 dst_rq->nr_running == 1)
1368 * In the overloaded case, try and keep the load balanced.
1371 load = task_h_load(env->p);
1372 dst_load = env->dst_stats.load + load;
1373 src_load = env->src_stats.load - load;
1375 if (moveimp > imp && moveimp > env->best_imp) {
1377 * If the improvement from just moving env->p direction is
1378 * better than swapping tasks around, check if a move is
1379 * possible. Store a slightly smaller score than moveimp,
1380 * so an actually idle CPU will win.
1382 if (!load_too_imbalanced(src_load, dst_load, env)) {
1384 put_task_struct(cur);
1390 if (imp <= env->best_imp)
1394 load = task_h_load(cur);
1399 if (load_too_imbalanced(src_load, dst_load, env))
1403 * One idle CPU per node is evaluated for a task numa move.
1404 * Call select_idle_sibling to maybe find a better one.
1407 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1411 task_numa_assign(env, cur, imp);
1415 * The dst_rq->curr isn't assigned. The protection for task_struct is
1418 if (cur && !assigned)
1419 put_task_struct(cur);
1422 static void task_numa_find_cpu(struct task_numa_env *env,
1423 long taskimp, long groupimp)
1427 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1428 /* Skip this CPU if the source task cannot migrate */
1429 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1433 task_numa_compare(env, taskimp, groupimp);
1437 /* Only move tasks to a NUMA node less busy than the current node. */
1438 static bool numa_has_capacity(struct task_numa_env *env)
1440 struct numa_stats *src = &env->src_stats;
1441 struct numa_stats *dst = &env->dst_stats;
1443 if (src->has_free_capacity && !dst->has_free_capacity)
1447 * Only consider a task move if the source has a higher load
1448 * than the destination, corrected for CPU capacity on each node.
1450 * src->load dst->load
1451 * --------------------- vs ---------------------
1452 * src->compute_capacity dst->compute_capacity
1454 if (src->load * dst->compute_capacity * env->imbalance_pct >
1456 dst->load * src->compute_capacity * 100)
1462 static int task_numa_migrate(struct task_struct *p)
1464 struct task_numa_env env = {
1467 .src_cpu = task_cpu(p),
1468 .src_nid = task_node(p),
1470 .imbalance_pct = 112,
1476 struct sched_domain *sd;
1477 unsigned long taskweight, groupweight;
1479 long taskimp, groupimp;
1482 * Pick the lowest SD_NUMA domain, as that would have the smallest
1483 * imbalance and would be the first to start moving tasks about.
1485 * And we want to avoid any moving of tasks about, as that would create
1486 * random movement of tasks -- counter the numa conditions we're trying
1490 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1492 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1496 * Cpusets can break the scheduler domain tree into smaller
1497 * balance domains, some of which do not cross NUMA boundaries.
1498 * Tasks that are "trapped" in such domains cannot be migrated
1499 * elsewhere, so there is no point in (re)trying.
1501 if (unlikely(!sd)) {
1502 p->numa_preferred_nid = task_node(p);
1506 env.dst_nid = p->numa_preferred_nid;
1507 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1508 taskweight = task_weight(p, env.src_nid, dist);
1509 groupweight = group_weight(p, env.src_nid, dist);
1510 update_numa_stats(&env.src_stats, env.src_nid);
1511 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1512 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1513 update_numa_stats(&env.dst_stats, env.dst_nid);
1515 /* Try to find a spot on the preferred nid. */
1516 if (numa_has_capacity(&env))
1517 task_numa_find_cpu(&env, taskimp, groupimp);
1520 * Look at other nodes in these cases:
1521 * - there is no space available on the preferred_nid
1522 * - the task is part of a numa_group that is interleaved across
1523 * multiple NUMA nodes; in order to better consolidate the group,
1524 * we need to check other locations.
1526 if (env.best_cpu == -1 || (p->numa_group &&
1527 nodes_weight(p->numa_group->active_nodes) > 1)) {
1528 for_each_online_node(nid) {
1529 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1532 dist = node_distance(env.src_nid, env.dst_nid);
1533 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1535 taskweight = task_weight(p, env.src_nid, dist);
1536 groupweight = group_weight(p, env.src_nid, dist);
1539 /* Only consider nodes where both task and groups benefit */
1540 taskimp = task_weight(p, nid, dist) - taskweight;
1541 groupimp = group_weight(p, nid, dist) - groupweight;
1542 if (taskimp < 0 && groupimp < 0)
1547 update_numa_stats(&env.dst_stats, env.dst_nid);
1548 if (numa_has_capacity(&env))
1549 task_numa_find_cpu(&env, taskimp, groupimp);
1554 * If the task is part of a workload that spans multiple NUMA nodes,
1555 * and is migrating into one of the workload's active nodes, remember
1556 * this node as the task's preferred numa node, so the workload can
1558 * A task that migrated to a second choice node will be better off
1559 * trying for a better one later. Do not set the preferred node here.
1561 if (p->numa_group) {
1562 if (env.best_cpu == -1)
1567 if (node_isset(nid, p->numa_group->active_nodes))
1568 sched_setnuma(p, env.dst_nid);
1571 /* No better CPU than the current one was found. */
1572 if (env.best_cpu == -1)
1576 * Reset the scan period if the task is being rescheduled on an
1577 * alternative node to recheck if the tasks is now properly placed.
1579 p->numa_scan_period = task_scan_min(p);
1581 if (env.best_task == NULL) {
1582 ret = migrate_task_to(p, env.best_cpu);
1584 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1588 ret = migrate_swap(p, env.best_task);
1590 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1591 put_task_struct(env.best_task);
1595 /* Attempt to migrate a task to a CPU on the preferred node. */
1596 static void numa_migrate_preferred(struct task_struct *p)
1598 unsigned long interval = HZ;
1600 /* This task has no NUMA fault statistics yet */
1601 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1604 /* Periodically retry migrating the task to the preferred node */
1605 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1606 p->numa_migrate_retry = jiffies + interval;
1608 /* Success if task is already running on preferred CPU */
1609 if (task_node(p) == p->numa_preferred_nid)
1612 /* Otherwise, try migrate to a CPU on the preferred node */
1613 task_numa_migrate(p);
1617 * Find the nodes on which the workload is actively running. We do this by
1618 * tracking the nodes from which NUMA hinting faults are triggered. This can
1619 * be different from the set of nodes where the workload's memory is currently
1622 * The bitmask is used to make smarter decisions on when to do NUMA page
1623 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1624 * are added when they cause over 6/16 of the maximum number of faults, but
1625 * only removed when they drop below 3/16.
1627 static void update_numa_active_node_mask(struct numa_group *numa_group)
1629 unsigned long faults, max_faults = 0;
1632 for_each_online_node(nid) {
1633 faults = group_faults_cpu(numa_group, nid);
1634 if (faults > max_faults)
1635 max_faults = faults;
1638 for_each_online_node(nid) {
1639 faults = group_faults_cpu(numa_group, nid);
1640 if (!node_isset(nid, numa_group->active_nodes)) {
1641 if (faults > max_faults * 6 / 16)
1642 node_set(nid, numa_group->active_nodes);
1643 } else if (faults < max_faults * 3 / 16)
1644 node_clear(nid, numa_group->active_nodes);
1649 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1650 * increments. The more local the fault statistics are, the higher the scan
1651 * period will be for the next scan window. If local/(local+remote) ratio is
1652 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1653 * the scan period will decrease. Aim for 70% local accesses.
1655 #define NUMA_PERIOD_SLOTS 10
1656 #define NUMA_PERIOD_THRESHOLD 7
1659 * Increase the scan period (slow down scanning) if the majority of
1660 * our memory is already on our local node, or if the majority of
1661 * the page accesses are shared with other processes.
1662 * Otherwise, decrease the scan period.
1664 static void update_task_scan_period(struct task_struct *p,
1665 unsigned long shared, unsigned long private)
1667 unsigned int period_slot;
1671 unsigned long remote = p->numa_faults_locality[0];
1672 unsigned long local = p->numa_faults_locality[1];
1675 * If there were no record hinting faults then either the task is
1676 * completely idle or all activity is areas that are not of interest
1677 * to automatic numa balancing. Related to that, if there were failed
1678 * migration then it implies we are migrating too quickly or the local
1679 * node is overloaded. In either case, scan slower
1681 if (local + shared == 0 || p->numa_faults_locality[2]) {
1682 p->numa_scan_period = min(p->numa_scan_period_max,
1683 p->numa_scan_period << 1);
1685 p->mm->numa_next_scan = jiffies +
1686 msecs_to_jiffies(p->numa_scan_period);
1692 * Prepare to scale scan period relative to the current period.
1693 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1694 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1695 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1697 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1698 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1699 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1700 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1703 diff = slot * period_slot;
1705 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1708 * Scale scan rate increases based on sharing. There is an
1709 * inverse relationship between the degree of sharing and
1710 * the adjustment made to the scanning period. Broadly
1711 * speaking the intent is that there is little point
1712 * scanning faster if shared accesses dominate as it may
1713 * simply bounce migrations uselessly
1715 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1716 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1719 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1720 task_scan_min(p), task_scan_max(p));
1721 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1725 * Get the fraction of time the task has been running since the last
1726 * NUMA placement cycle. The scheduler keeps similar statistics, but
1727 * decays those on a 32ms period, which is orders of magnitude off
1728 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1729 * stats only if the task is so new there are no NUMA statistics yet.
1731 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1733 u64 runtime, delta, now;
1734 /* Use the start of this time slice to avoid calculations. */
1735 now = p->se.exec_start;
1736 runtime = p->se.sum_exec_runtime;
1738 if (p->last_task_numa_placement) {
1739 delta = runtime - p->last_sum_exec_runtime;
1740 *period = now - p->last_task_numa_placement;
1742 delta = p->se.avg.load_sum / p->se.load.weight;
1743 *period = LOAD_AVG_MAX;
1746 p->last_sum_exec_runtime = runtime;
1747 p->last_task_numa_placement = now;
1753 * Determine the preferred nid for a task in a numa_group. This needs to
1754 * be done in a way that produces consistent results with group_weight,
1755 * otherwise workloads might not converge.
1757 static int preferred_group_nid(struct task_struct *p, int nid)
1762 /* Direct connections between all NUMA nodes. */
1763 if (sched_numa_topology_type == NUMA_DIRECT)
1767 * On a system with glueless mesh NUMA topology, group_weight
1768 * scores nodes according to the number of NUMA hinting faults on
1769 * both the node itself, and on nearby nodes.
1771 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1772 unsigned long score, max_score = 0;
1773 int node, max_node = nid;
1775 dist = sched_max_numa_distance;
1777 for_each_online_node(node) {
1778 score = group_weight(p, node, dist);
1779 if (score > max_score) {
1788 * Finding the preferred nid in a system with NUMA backplane
1789 * interconnect topology is more involved. The goal is to locate
1790 * tasks from numa_groups near each other in the system, and
1791 * untangle workloads from different sides of the system. This requires
1792 * searching down the hierarchy of node groups, recursively searching
1793 * inside the highest scoring group of nodes. The nodemask tricks
1794 * keep the complexity of the search down.
1796 nodes = node_online_map;
1797 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1798 unsigned long max_faults = 0;
1799 nodemask_t max_group = NODE_MASK_NONE;
1802 /* Are there nodes at this distance from each other? */
1803 if (!find_numa_distance(dist))
1806 for_each_node_mask(a, nodes) {
1807 unsigned long faults = 0;
1808 nodemask_t this_group;
1809 nodes_clear(this_group);
1811 /* Sum group's NUMA faults; includes a==b case. */
1812 for_each_node_mask(b, nodes) {
1813 if (node_distance(a, b) < dist) {
1814 faults += group_faults(p, b);
1815 node_set(b, this_group);
1816 node_clear(b, nodes);
1820 /* Remember the top group. */
1821 if (faults > max_faults) {
1822 max_faults = faults;
1823 max_group = this_group;
1825 * subtle: at the smallest distance there is
1826 * just one node left in each "group", the
1827 * winner is the preferred nid.
1832 /* Next round, evaluate the nodes within max_group. */
1840 static void task_numa_placement(struct task_struct *p)
1842 int seq, nid, max_nid = -1, max_group_nid = -1;
1843 unsigned long max_faults = 0, max_group_faults = 0;
1844 unsigned long fault_types[2] = { 0, 0 };
1845 unsigned long total_faults;
1846 u64 runtime, period;
1847 spinlock_t *group_lock = NULL;
1850 * The p->mm->numa_scan_seq field gets updated without
1851 * exclusive access. Use READ_ONCE() here to ensure
1852 * that the field is read in a single access:
1854 seq = READ_ONCE(p->mm->numa_scan_seq);
1855 if (p->numa_scan_seq == seq)
1857 p->numa_scan_seq = seq;
1858 p->numa_scan_period_max = task_scan_max(p);
1860 total_faults = p->numa_faults_locality[0] +
1861 p->numa_faults_locality[1];
1862 runtime = numa_get_avg_runtime(p, &period);
1864 /* If the task is part of a group prevent parallel updates to group stats */
1865 if (p->numa_group) {
1866 group_lock = &p->numa_group->lock;
1867 spin_lock_irq(group_lock);
1870 /* Find the node with the highest number of faults */
1871 for_each_online_node(nid) {
1872 /* Keep track of the offsets in numa_faults array */
1873 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1874 unsigned long faults = 0, group_faults = 0;
1877 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1878 long diff, f_diff, f_weight;
1880 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1881 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1882 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1883 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1885 /* Decay existing window, copy faults since last scan */
1886 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1887 fault_types[priv] += p->numa_faults[membuf_idx];
1888 p->numa_faults[membuf_idx] = 0;
1891 * Normalize the faults_from, so all tasks in a group
1892 * count according to CPU use, instead of by the raw
1893 * number of faults. Tasks with little runtime have
1894 * little over-all impact on throughput, and thus their
1895 * faults are less important.
1897 f_weight = div64_u64(runtime << 16, period + 1);
1898 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1900 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1901 p->numa_faults[cpubuf_idx] = 0;
1903 p->numa_faults[mem_idx] += diff;
1904 p->numa_faults[cpu_idx] += f_diff;
1905 faults += p->numa_faults[mem_idx];
1906 p->total_numa_faults += diff;
1907 if (p->numa_group) {
1909 * safe because we can only change our own group
1911 * mem_idx represents the offset for a given
1912 * nid and priv in a specific region because it
1913 * is at the beginning of the numa_faults array.
1915 p->numa_group->faults[mem_idx] += diff;
1916 p->numa_group->faults_cpu[mem_idx] += f_diff;
1917 p->numa_group->total_faults += diff;
1918 group_faults += p->numa_group->faults[mem_idx];
1922 if (faults > max_faults) {
1923 max_faults = faults;
1927 if (group_faults > max_group_faults) {
1928 max_group_faults = group_faults;
1929 max_group_nid = nid;
1933 update_task_scan_period(p, fault_types[0], fault_types[1]);
1935 if (p->numa_group) {
1936 update_numa_active_node_mask(p->numa_group);
1937 spin_unlock_irq(group_lock);
1938 max_nid = preferred_group_nid(p, max_group_nid);
1942 /* Set the new preferred node */
1943 if (max_nid != p->numa_preferred_nid)
1944 sched_setnuma(p, max_nid);
1946 if (task_node(p) != p->numa_preferred_nid)
1947 numa_migrate_preferred(p);
1951 static inline int get_numa_group(struct numa_group *grp)
1953 return atomic_inc_not_zero(&grp->refcount);
1956 static inline void put_numa_group(struct numa_group *grp)
1958 if (atomic_dec_and_test(&grp->refcount))
1959 kfree_rcu(grp, rcu);
1962 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1965 struct numa_group *grp, *my_grp;
1966 struct task_struct *tsk;
1968 int cpu = cpupid_to_cpu(cpupid);
1971 if (unlikely(!p->numa_group)) {
1972 unsigned int size = sizeof(struct numa_group) +
1973 4*nr_node_ids*sizeof(unsigned long);
1975 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1979 atomic_set(&grp->refcount, 1);
1980 spin_lock_init(&grp->lock);
1982 /* Second half of the array tracks nids where faults happen */
1983 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1986 node_set(task_node(current), grp->active_nodes);
1988 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1989 grp->faults[i] = p->numa_faults[i];
1991 grp->total_faults = p->total_numa_faults;
1994 rcu_assign_pointer(p->numa_group, grp);
1998 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2000 if (!cpupid_match_pid(tsk, cpupid))
2003 grp = rcu_dereference(tsk->numa_group);
2007 my_grp = p->numa_group;
2012 * Only join the other group if its bigger; if we're the bigger group,
2013 * the other task will join us.
2015 if (my_grp->nr_tasks > grp->nr_tasks)
2019 * Tie-break on the grp address.
2021 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2024 /* Always join threads in the same process. */
2025 if (tsk->mm == current->mm)
2028 /* Simple filter to avoid false positives due to PID collisions */
2029 if (flags & TNF_SHARED)
2032 /* Update priv based on whether false sharing was detected */
2035 if (join && !get_numa_group(grp))
2043 BUG_ON(irqs_disabled());
2044 double_lock_irq(&my_grp->lock, &grp->lock);
2046 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2047 my_grp->faults[i] -= p->numa_faults[i];
2048 grp->faults[i] += p->numa_faults[i];
2050 my_grp->total_faults -= p->total_numa_faults;
2051 grp->total_faults += p->total_numa_faults;
2056 spin_unlock(&my_grp->lock);
2057 spin_unlock_irq(&grp->lock);
2059 rcu_assign_pointer(p->numa_group, grp);
2061 put_numa_group(my_grp);
2069 void task_numa_free(struct task_struct *p)
2071 struct numa_group *grp = p->numa_group;
2072 void *numa_faults = p->numa_faults;
2073 unsigned long flags;
2077 spin_lock_irqsave(&grp->lock, flags);
2078 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2079 grp->faults[i] -= p->numa_faults[i];
2080 grp->total_faults -= p->total_numa_faults;
2083 spin_unlock_irqrestore(&grp->lock, flags);
2084 RCU_INIT_POINTER(p->numa_group, NULL);
2085 put_numa_group(grp);
2088 p->numa_faults = NULL;
2093 * Got a PROT_NONE fault for a page on @node.
2095 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2097 struct task_struct *p = current;
2098 bool migrated = flags & TNF_MIGRATED;
2099 int cpu_node = task_node(current);
2100 int local = !!(flags & TNF_FAULT_LOCAL);
2103 if (!static_branch_likely(&sched_numa_balancing))
2106 /* for example, ksmd faulting in a user's mm */
2110 /* Allocate buffer to track faults on a per-node basis */
2111 if (unlikely(!p->numa_faults)) {
2112 int size = sizeof(*p->numa_faults) *
2113 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2115 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2116 if (!p->numa_faults)
2119 p->total_numa_faults = 0;
2120 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2124 * First accesses are treated as private, otherwise consider accesses
2125 * to be private if the accessing pid has not changed
2127 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2130 priv = cpupid_match_pid(p, last_cpupid);
2131 if (!priv && !(flags & TNF_NO_GROUP))
2132 task_numa_group(p, last_cpupid, flags, &priv);
2136 * If a workload spans multiple NUMA nodes, a shared fault that
2137 * occurs wholly within the set of nodes that the workload is
2138 * actively using should be counted as local. This allows the
2139 * scan rate to slow down when a workload has settled down.
2141 if (!priv && !local && p->numa_group &&
2142 node_isset(cpu_node, p->numa_group->active_nodes) &&
2143 node_isset(mem_node, p->numa_group->active_nodes))
2146 task_numa_placement(p);
2149 * Retry task to preferred node migration periodically, in case it
2150 * case it previously failed, or the scheduler moved us.
2152 if (time_after(jiffies, p->numa_migrate_retry))
2153 numa_migrate_preferred(p);
2156 p->numa_pages_migrated += pages;
2157 if (flags & TNF_MIGRATE_FAIL)
2158 p->numa_faults_locality[2] += pages;
2160 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2161 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2162 p->numa_faults_locality[local] += pages;
2165 static void reset_ptenuma_scan(struct task_struct *p)
2168 * We only did a read acquisition of the mmap sem, so
2169 * p->mm->numa_scan_seq is written to without exclusive access
2170 * and the update is not guaranteed to be atomic. That's not
2171 * much of an issue though, since this is just used for
2172 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2173 * expensive, to avoid any form of compiler optimizations:
2175 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2176 p->mm->numa_scan_offset = 0;
2180 * The expensive part of numa migration is done from task_work context.
2181 * Triggered from task_tick_numa().
2183 void task_numa_work(struct callback_head *work)
2185 unsigned long migrate, next_scan, now = jiffies;
2186 struct task_struct *p = current;
2187 struct mm_struct *mm = p->mm;
2188 struct vm_area_struct *vma;
2189 unsigned long start, end;
2190 unsigned long nr_pte_updates = 0;
2191 long pages, virtpages;
2193 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2195 work->next = work; /* protect against double add */
2197 * Who cares about NUMA placement when they're dying.
2199 * NOTE: make sure not to dereference p->mm before this check,
2200 * exit_task_work() happens _after_ exit_mm() so we could be called
2201 * without p->mm even though we still had it when we enqueued this
2204 if (p->flags & PF_EXITING)
2207 if (!mm->numa_next_scan) {
2208 mm->numa_next_scan = now +
2209 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2213 * Enforce maximal scan/migration frequency..
2215 migrate = mm->numa_next_scan;
2216 if (time_before(now, migrate))
2219 if (p->numa_scan_period == 0) {
2220 p->numa_scan_period_max = task_scan_max(p);
2221 p->numa_scan_period = task_scan_min(p);
2224 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2225 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2229 * Delay this task enough that another task of this mm will likely win
2230 * the next time around.
2232 p->node_stamp += 2 * TICK_NSEC;
2234 start = mm->numa_scan_offset;
2235 pages = sysctl_numa_balancing_scan_size;
2236 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2237 virtpages = pages * 8; /* Scan up to this much virtual space */
2242 down_read(&mm->mmap_sem);
2243 vma = find_vma(mm, start);
2245 reset_ptenuma_scan(p);
2249 for (; vma; vma = vma->vm_next) {
2250 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2251 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2256 * Shared library pages mapped by multiple processes are not
2257 * migrated as it is expected they are cache replicated. Avoid
2258 * hinting faults in read-only file-backed mappings or the vdso
2259 * as migrating the pages will be of marginal benefit.
2262 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2266 * Skip inaccessible VMAs to avoid any confusion between
2267 * PROT_NONE and NUMA hinting ptes
2269 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2273 start = max(start, vma->vm_start);
2274 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2275 end = min(end, vma->vm_end);
2276 nr_pte_updates = change_prot_numa(vma, start, end);
2279 * Try to scan sysctl_numa_balancing_size worth of
2280 * hpages that have at least one present PTE that
2281 * is not already pte-numa. If the VMA contains
2282 * areas that are unused or already full of prot_numa
2283 * PTEs, scan up to virtpages, to skip through those
2287 pages -= (end - start) >> PAGE_SHIFT;
2288 virtpages -= (end - start) >> PAGE_SHIFT;
2291 if (pages <= 0 || virtpages <= 0)
2295 } while (end != vma->vm_end);
2300 * It is possible to reach the end of the VMA list but the last few
2301 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2302 * would find the !migratable VMA on the next scan but not reset the
2303 * scanner to the start so check it now.
2306 mm->numa_scan_offset = start;
2308 reset_ptenuma_scan(p);
2309 up_read(&mm->mmap_sem);
2313 * Drive the periodic memory faults..
2315 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2317 struct callback_head *work = &curr->numa_work;
2321 * We don't care about NUMA placement if we don't have memory.
2323 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2327 * Using runtime rather than walltime has the dual advantage that
2328 * we (mostly) drive the selection from busy threads and that the
2329 * task needs to have done some actual work before we bother with
2332 now = curr->se.sum_exec_runtime;
2333 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2335 if (now > curr->node_stamp + period) {
2336 if (!curr->node_stamp)
2337 curr->numa_scan_period = task_scan_min(curr);
2338 curr->node_stamp += period;
2340 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2341 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2342 task_work_add(curr, work, true);
2347 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2351 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2355 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2358 #endif /* CONFIG_NUMA_BALANCING */
2361 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2363 update_load_add(&cfs_rq->load, se->load.weight);
2364 if (!parent_entity(se))
2365 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2367 if (entity_is_task(se)) {
2368 struct rq *rq = rq_of(cfs_rq);
2370 account_numa_enqueue(rq, task_of(se));
2371 list_add(&se->group_node, &rq->cfs_tasks);
2374 cfs_rq->nr_running++;
2378 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2380 update_load_sub(&cfs_rq->load, se->load.weight);
2381 if (!parent_entity(se))
2382 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2383 if (entity_is_task(se)) {
2384 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2385 list_del_init(&se->group_node);
2387 cfs_rq->nr_running--;
2390 #ifdef CONFIG_FAIR_GROUP_SCHED
2392 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2397 * Use this CPU's real-time load instead of the last load contribution
2398 * as the updating of the contribution is delayed, and we will use the
2399 * the real-time load to calc the share. See update_tg_load_avg().
2401 tg_weight = atomic_long_read(&tg->load_avg);
2402 tg_weight -= cfs_rq->tg_load_avg_contrib;
2403 tg_weight += cfs_rq->load.weight;
2408 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2410 long tg_weight, load, shares;
2412 tg_weight = calc_tg_weight(tg, cfs_rq);
2413 load = cfs_rq->load.weight;
2415 shares = (tg->shares * load);
2417 shares /= tg_weight;
2419 if (shares < MIN_SHARES)
2420 shares = MIN_SHARES;
2421 if (shares > tg->shares)
2422 shares = tg->shares;
2426 # else /* CONFIG_SMP */
2427 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2431 # endif /* CONFIG_SMP */
2432 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2433 unsigned long weight)
2436 /* commit outstanding execution time */
2437 if (cfs_rq->curr == se)
2438 update_curr(cfs_rq);
2439 account_entity_dequeue(cfs_rq, se);
2442 update_load_set(&se->load, weight);
2445 account_entity_enqueue(cfs_rq, se);
2448 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2450 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2452 struct task_group *tg;
2453 struct sched_entity *se;
2457 se = tg->se[cpu_of(rq_of(cfs_rq))];
2458 if (!se || throttled_hierarchy(cfs_rq))
2461 if (likely(se->load.weight == tg->shares))
2464 shares = calc_cfs_shares(cfs_rq, tg);
2466 reweight_entity(cfs_rq_of(se), se, shares);
2468 #else /* CONFIG_FAIR_GROUP_SCHED */
2469 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2472 #endif /* CONFIG_FAIR_GROUP_SCHED */
2475 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2476 static const u32 runnable_avg_yN_inv[] = {
2477 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2478 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2479 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2480 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2481 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2482 0x85aac367, 0x82cd8698,
2486 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2487 * over-estimates when re-combining.
2489 static const u32 runnable_avg_yN_sum[] = {
2490 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2491 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2492 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2497 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2499 static __always_inline u64 decay_load(u64 val, u64 n)
2501 unsigned int local_n;
2505 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2508 /* after bounds checking we can collapse to 32-bit */
2512 * As y^PERIOD = 1/2, we can combine
2513 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2514 * With a look-up table which covers y^n (n<PERIOD)
2516 * To achieve constant time decay_load.
2518 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2519 val >>= local_n / LOAD_AVG_PERIOD;
2520 local_n %= LOAD_AVG_PERIOD;
2523 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2528 * For updates fully spanning n periods, the contribution to runnable
2529 * average will be: \Sum 1024*y^n
2531 * We can compute this reasonably efficiently by combining:
2532 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2534 static u32 __compute_runnable_contrib(u64 n)
2538 if (likely(n <= LOAD_AVG_PERIOD))
2539 return runnable_avg_yN_sum[n];
2540 else if (unlikely(n >= LOAD_AVG_MAX_N))
2541 return LOAD_AVG_MAX;
2543 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2545 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2546 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2548 n -= LOAD_AVG_PERIOD;
2549 } while (n > LOAD_AVG_PERIOD);
2551 contrib = decay_load(contrib, n);
2552 return contrib + runnable_avg_yN_sum[n];
2555 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2556 #error "load tracking assumes 2^10 as unit"
2559 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2562 * We can represent the historical contribution to runnable average as the
2563 * coefficients of a geometric series. To do this we sub-divide our runnable
2564 * history into segments of approximately 1ms (1024us); label the segment that
2565 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2567 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2569 * (now) (~1ms ago) (~2ms ago)
2571 * Let u_i denote the fraction of p_i that the entity was runnable.
2573 * We then designate the fractions u_i as our co-efficients, yielding the
2574 * following representation of historical load:
2575 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2577 * We choose y based on the with of a reasonably scheduling period, fixing:
2580 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2581 * approximately half as much as the contribution to load within the last ms
2584 * When a period "rolls over" and we have new u_0`, multiplying the previous
2585 * sum again by y is sufficient to update:
2586 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2587 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2589 static __always_inline int
2590 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2591 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2593 u64 delta, scaled_delta, periods;
2595 unsigned int delta_w, scaled_delta_w, decayed = 0;
2596 unsigned long scale_freq, scale_cpu;
2598 delta = now - sa->last_update_time;
2600 * This should only happen when time goes backwards, which it
2601 * unfortunately does during sched clock init when we swap over to TSC.
2603 if ((s64)delta < 0) {
2604 sa->last_update_time = now;
2609 * Use 1024ns as the unit of measurement since it's a reasonable
2610 * approximation of 1us and fast to compute.
2615 sa->last_update_time = now;
2617 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2618 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2619 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2621 /* delta_w is the amount already accumulated against our next period */
2622 delta_w = sa->period_contrib;
2623 if (delta + delta_w >= 1024) {
2626 /* how much left for next period will start over, we don't know yet */
2627 sa->period_contrib = 0;
2630 * Now that we know we're crossing a period boundary, figure
2631 * out how much from delta we need to complete the current
2632 * period and accrue it.
2634 delta_w = 1024 - delta_w;
2635 scaled_delta_w = cap_scale(delta_w, scale_freq);
2637 sa->load_sum += weight * scaled_delta_w;
2639 cfs_rq->runnable_load_sum +=
2640 weight * scaled_delta_w;
2644 sa->util_sum += scaled_delta_w * scale_cpu;
2648 /* Figure out how many additional periods this update spans */
2649 periods = delta / 1024;
2652 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2654 cfs_rq->runnable_load_sum =
2655 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2657 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2659 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2660 contrib = __compute_runnable_contrib(periods);
2661 contrib = cap_scale(contrib, scale_freq);
2663 sa->load_sum += weight * contrib;
2665 cfs_rq->runnable_load_sum += weight * contrib;
2668 sa->util_sum += contrib * scale_cpu;
2671 /* Remainder of delta accrued against u_0` */
2672 scaled_delta = cap_scale(delta, scale_freq);
2674 sa->load_sum += weight * scaled_delta;
2676 cfs_rq->runnable_load_sum += weight * scaled_delta;
2679 sa->util_sum += scaled_delta * scale_cpu;
2681 sa->period_contrib += delta;
2684 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2686 cfs_rq->runnable_load_avg =
2687 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2689 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2695 #ifdef CONFIG_FAIR_GROUP_SCHED
2697 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2698 * and effective_load (which is not done because it is too costly).
2700 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2702 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2704 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2705 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2706 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2710 #else /* CONFIG_FAIR_GROUP_SCHED */
2711 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2712 #endif /* CONFIG_FAIR_GROUP_SCHED */
2714 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2717 * Unsigned subtract and clamp on underflow.
2719 * Explicitly do a load-store to ensure the intermediate value never hits
2720 * memory. This allows lockless observations without ever seeing the negative
2723 #define sub_positive(_ptr, _val) do { \
2724 typeof(_ptr) ptr = (_ptr); \
2725 typeof(*ptr) val = (_val); \
2726 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2730 WRITE_ONCE(*ptr, res); \
2733 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2734 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2736 struct sched_avg *sa = &cfs_rq->avg;
2737 int decayed, removed = 0;
2739 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2740 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2741 sub_positive(&sa->load_avg, r);
2742 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2746 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2747 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2748 sub_positive(&sa->util_avg, r);
2749 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2752 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2753 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2755 #ifndef CONFIG_64BIT
2757 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2760 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
2762 return decayed || removed;
2765 /* Update task and its cfs_rq load average */
2766 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2768 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2769 u64 now = cfs_rq_clock_task(cfs_rq);
2770 int cpu = cpu_of(rq_of(cfs_rq));
2773 * Track task load average for carrying it to new CPU after migrated, and
2774 * track group sched_entity load average for task_h_load calc in migration
2776 __update_load_avg(now, cpu, &se->avg,
2777 se->on_rq * scale_load_down(se->load.weight),
2778 cfs_rq->curr == se, NULL);
2780 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2781 update_tg_load_avg(cfs_rq, 0);
2783 if (entity_is_task(se))
2784 trace_sched_load_avg_task(task_of(se), &se->avg);
2787 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2789 if (!sched_feat(ATTACH_AGE_LOAD))
2793 * If we got migrated (either between CPUs or between cgroups) we'll
2794 * have aged the average right before clearing @last_update_time.
2796 if (se->avg.last_update_time) {
2797 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2798 &se->avg, 0, 0, NULL);
2801 * XXX: we could have just aged the entire load away if we've been
2802 * absent from the fair class for too long.
2807 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2808 cfs_rq->avg.load_avg += se->avg.load_avg;
2809 cfs_rq->avg.load_sum += se->avg.load_sum;
2810 cfs_rq->avg.util_avg += se->avg.util_avg;
2811 cfs_rq->avg.util_sum += se->avg.util_sum;
2814 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2816 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2817 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2818 cfs_rq->curr == se, NULL);
2820 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2821 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2822 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2823 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2826 /* Add the load generated by se into cfs_rq's load average */
2828 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2830 struct sched_avg *sa = &se->avg;
2831 u64 now = cfs_rq_clock_task(cfs_rq);
2832 int migrated, decayed;
2834 migrated = !sa->last_update_time;
2836 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2837 se->on_rq * scale_load_down(se->load.weight),
2838 cfs_rq->curr == se, NULL);
2841 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2843 cfs_rq->runnable_load_avg += sa->load_avg;
2844 cfs_rq->runnable_load_sum += sa->load_sum;
2847 attach_entity_load_avg(cfs_rq, se);
2849 if (decayed || migrated)
2850 update_tg_load_avg(cfs_rq, 0);
2853 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2855 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2857 update_load_avg(se, 1);
2859 cfs_rq->runnable_load_avg =
2860 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2861 cfs_rq->runnable_load_sum =
2862 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2865 #ifndef CONFIG_64BIT
2866 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2868 u64 last_update_time_copy;
2869 u64 last_update_time;
2872 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2874 last_update_time = cfs_rq->avg.last_update_time;
2875 } while (last_update_time != last_update_time_copy);
2877 return last_update_time;
2880 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2882 return cfs_rq->avg.last_update_time;
2887 * Task first catches up with cfs_rq, and then subtract
2888 * itself from the cfs_rq (task must be off the queue now).
2890 void remove_entity_load_avg(struct sched_entity *se)
2892 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2893 u64 last_update_time;
2896 * Newly created task or never used group entity should not be removed
2897 * from its (source) cfs_rq
2899 if (se->avg.last_update_time == 0)
2902 last_update_time = cfs_rq_last_update_time(cfs_rq);
2904 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2905 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2906 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2910 * Update the rq's load with the elapsed running time before entering
2911 * idle. if the last scheduled task is not a CFS task, idle_enter will
2912 * be the only way to update the runnable statistic.
2914 void idle_enter_fair(struct rq *this_rq)
2919 * Update the rq's load with the elapsed idle time before a task is
2920 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2921 * be the only way to update the runnable statistic.
2923 void idle_exit_fair(struct rq *this_rq)
2927 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2929 return cfs_rq->runnable_load_avg;
2932 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2934 return cfs_rq->avg.load_avg;
2937 static int idle_balance(struct rq *this_rq);
2939 #else /* CONFIG_SMP */
2941 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2943 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2945 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2946 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2949 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2951 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2953 static inline int idle_balance(struct rq *rq)
2958 #endif /* CONFIG_SMP */
2960 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2962 #ifdef CONFIG_SCHEDSTATS
2963 struct task_struct *tsk = NULL;
2965 if (entity_is_task(se))
2968 if (se->statistics.sleep_start) {
2969 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2974 if (unlikely(delta > se->statistics.sleep_max))
2975 se->statistics.sleep_max = delta;
2977 se->statistics.sleep_start = 0;
2978 se->statistics.sum_sleep_runtime += delta;
2981 account_scheduler_latency(tsk, delta >> 10, 1);
2982 trace_sched_stat_sleep(tsk, delta);
2985 if (se->statistics.block_start) {
2986 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2991 if (unlikely(delta > se->statistics.block_max))
2992 se->statistics.block_max = delta;
2994 se->statistics.block_start = 0;
2995 se->statistics.sum_sleep_runtime += delta;
2998 if (tsk->in_iowait) {
2999 se->statistics.iowait_sum += delta;
3000 se->statistics.iowait_count++;
3001 trace_sched_stat_iowait(tsk, delta);
3004 trace_sched_stat_blocked(tsk, delta);
3005 trace_sched_blocked_reason(tsk);
3008 * Blocking time is in units of nanosecs, so shift by
3009 * 20 to get a milliseconds-range estimation of the
3010 * amount of time that the task spent sleeping:
3012 if (unlikely(prof_on == SLEEP_PROFILING)) {
3013 profile_hits(SLEEP_PROFILING,
3014 (void *)get_wchan(tsk),
3017 account_scheduler_latency(tsk, delta >> 10, 0);
3023 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3025 #ifdef CONFIG_SCHED_DEBUG
3026 s64 d = se->vruntime - cfs_rq->min_vruntime;
3031 if (d > 3*sysctl_sched_latency)
3032 schedstat_inc(cfs_rq, nr_spread_over);
3037 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3039 u64 vruntime = cfs_rq->min_vruntime;
3042 * The 'current' period is already promised to the current tasks,
3043 * however the extra weight of the new task will slow them down a
3044 * little, place the new task so that it fits in the slot that
3045 * stays open at the end.
3047 if (initial && sched_feat(START_DEBIT))
3048 vruntime += sched_vslice(cfs_rq, se);
3050 /* sleeps up to a single latency don't count. */
3052 unsigned long thresh = sysctl_sched_latency;
3055 * Halve their sleep time's effect, to allow
3056 * for a gentler effect of sleepers:
3058 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3064 /* ensure we never gain time by being placed backwards. */
3065 se->vruntime = max_vruntime(se->vruntime, vruntime);
3068 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3071 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3074 * Update the normalized vruntime before updating min_vruntime
3075 * through calling update_curr().
3077 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3078 se->vruntime += cfs_rq->min_vruntime;
3081 * Update run-time statistics of the 'current'.
3083 update_curr(cfs_rq);
3084 enqueue_entity_load_avg(cfs_rq, se);
3085 account_entity_enqueue(cfs_rq, se);
3086 update_cfs_shares(cfs_rq);
3088 if (flags & ENQUEUE_WAKEUP) {
3089 place_entity(cfs_rq, se, 0);
3090 enqueue_sleeper(cfs_rq, se);
3093 update_stats_enqueue(cfs_rq, se);
3094 check_spread(cfs_rq, se);
3095 if (se != cfs_rq->curr)
3096 __enqueue_entity(cfs_rq, se);
3099 if (cfs_rq->nr_running == 1) {
3100 list_add_leaf_cfs_rq(cfs_rq);
3101 check_enqueue_throttle(cfs_rq);
3105 static void __clear_buddies_last(struct sched_entity *se)
3107 for_each_sched_entity(se) {
3108 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3109 if (cfs_rq->last != se)
3112 cfs_rq->last = NULL;
3116 static void __clear_buddies_next(struct sched_entity *se)
3118 for_each_sched_entity(se) {
3119 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3120 if (cfs_rq->next != se)
3123 cfs_rq->next = NULL;
3127 static void __clear_buddies_skip(struct sched_entity *se)
3129 for_each_sched_entity(se) {
3130 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3131 if (cfs_rq->skip != se)
3134 cfs_rq->skip = NULL;
3138 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3140 if (cfs_rq->last == se)
3141 __clear_buddies_last(se);
3143 if (cfs_rq->next == se)
3144 __clear_buddies_next(se);
3146 if (cfs_rq->skip == se)
3147 __clear_buddies_skip(se);
3150 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3153 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3156 * Update run-time statistics of the 'current'.
3158 update_curr(cfs_rq);
3159 dequeue_entity_load_avg(cfs_rq, se);
3161 update_stats_dequeue(cfs_rq, se);
3162 if (flags & DEQUEUE_SLEEP) {
3163 #ifdef CONFIG_SCHEDSTATS
3164 if (entity_is_task(se)) {
3165 struct task_struct *tsk = task_of(se);
3167 if (tsk->state & TASK_INTERRUPTIBLE)
3168 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3169 if (tsk->state & TASK_UNINTERRUPTIBLE)
3170 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3175 clear_buddies(cfs_rq, se);
3177 if (se != cfs_rq->curr)
3178 __dequeue_entity(cfs_rq, se);
3180 account_entity_dequeue(cfs_rq, se);
3183 * Normalize the entity after updating the min_vruntime because the
3184 * update can refer to the ->curr item and we need to reflect this
3185 * movement in our normalized position.
3187 if (!(flags & DEQUEUE_SLEEP))
3188 se->vruntime -= cfs_rq->min_vruntime;
3190 /* return excess runtime on last dequeue */
3191 return_cfs_rq_runtime(cfs_rq);
3193 update_min_vruntime(cfs_rq);
3194 update_cfs_shares(cfs_rq);
3198 * Preempt the current task with a newly woken task if needed:
3201 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3203 unsigned long ideal_runtime, delta_exec;
3204 struct sched_entity *se;
3207 ideal_runtime = sched_slice(cfs_rq, curr);
3208 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3209 if (delta_exec > ideal_runtime) {
3210 resched_curr(rq_of(cfs_rq));
3212 * The current task ran long enough, ensure it doesn't get
3213 * re-elected due to buddy favours.
3215 clear_buddies(cfs_rq, curr);
3220 * Ensure that a task that missed wakeup preemption by a
3221 * narrow margin doesn't have to wait for a full slice.
3222 * This also mitigates buddy induced latencies under load.
3224 if (delta_exec < sysctl_sched_min_granularity)
3227 se = __pick_first_entity(cfs_rq);
3228 delta = curr->vruntime - se->vruntime;
3233 if (delta > ideal_runtime)
3234 resched_curr(rq_of(cfs_rq));
3238 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3240 /* 'current' is not kept within the tree. */
3243 * Any task has to be enqueued before it get to execute on
3244 * a CPU. So account for the time it spent waiting on the
3247 update_stats_wait_end(cfs_rq, se);
3248 __dequeue_entity(cfs_rq, se);
3249 update_load_avg(se, 1);
3252 update_stats_curr_start(cfs_rq, se);
3254 #ifdef CONFIG_SCHEDSTATS
3256 * Track our maximum slice length, if the CPU's load is at
3257 * least twice that of our own weight (i.e. dont track it
3258 * when there are only lesser-weight tasks around):
3260 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3261 se->statistics.slice_max = max(se->statistics.slice_max,
3262 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3265 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3269 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3272 * Pick the next process, keeping these things in mind, in this order:
3273 * 1) keep things fair between processes/task groups
3274 * 2) pick the "next" process, since someone really wants that to run
3275 * 3) pick the "last" process, for cache locality
3276 * 4) do not run the "skip" process, if something else is available
3278 static struct sched_entity *
3279 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3281 struct sched_entity *left = __pick_first_entity(cfs_rq);
3282 struct sched_entity *se;
3285 * If curr is set we have to see if its left of the leftmost entity
3286 * still in the tree, provided there was anything in the tree at all.
3288 if (!left || (curr && entity_before(curr, left)))
3291 se = left; /* ideally we run the leftmost entity */
3294 * Avoid running the skip buddy, if running something else can
3295 * be done without getting too unfair.
3297 if (cfs_rq->skip == se) {
3298 struct sched_entity *second;
3301 second = __pick_first_entity(cfs_rq);
3303 second = __pick_next_entity(se);
3304 if (!second || (curr && entity_before(curr, second)))
3308 if (second && wakeup_preempt_entity(second, left) < 1)
3313 * Prefer last buddy, try to return the CPU to a preempted task.
3315 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3319 * Someone really wants this to run. If it's not unfair, run it.
3321 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3324 clear_buddies(cfs_rq, se);
3329 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3331 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3334 * If still on the runqueue then deactivate_task()
3335 * was not called and update_curr() has to be done:
3338 update_curr(cfs_rq);
3340 /* throttle cfs_rqs exceeding runtime */
3341 check_cfs_rq_runtime(cfs_rq);
3343 check_spread(cfs_rq, prev);
3345 update_stats_wait_start(cfs_rq, prev);
3346 /* Put 'current' back into the tree. */
3347 __enqueue_entity(cfs_rq, prev);
3348 /* in !on_rq case, update occurred at dequeue */
3349 update_load_avg(prev, 0);
3351 cfs_rq->curr = NULL;
3355 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3358 * Update run-time statistics of the 'current'.
3360 update_curr(cfs_rq);
3363 * Ensure that runnable average is periodically updated.
3365 update_load_avg(curr, 1);
3366 update_cfs_shares(cfs_rq);
3368 #ifdef CONFIG_SCHED_HRTICK
3370 * queued ticks are scheduled to match the slice, so don't bother
3371 * validating it and just reschedule.
3374 resched_curr(rq_of(cfs_rq));
3378 * don't let the period tick interfere with the hrtick preemption
3380 if (!sched_feat(DOUBLE_TICK) &&
3381 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3385 if (cfs_rq->nr_running > 1)
3386 check_preempt_tick(cfs_rq, curr);
3390 /**************************************************
3391 * CFS bandwidth control machinery
3394 #ifdef CONFIG_CFS_BANDWIDTH
3396 #ifdef HAVE_JUMP_LABEL
3397 static struct static_key __cfs_bandwidth_used;
3399 static inline bool cfs_bandwidth_used(void)
3401 return static_key_false(&__cfs_bandwidth_used);
3404 void cfs_bandwidth_usage_inc(void)
3406 static_key_slow_inc(&__cfs_bandwidth_used);
3409 void cfs_bandwidth_usage_dec(void)
3411 static_key_slow_dec(&__cfs_bandwidth_used);
3413 #else /* HAVE_JUMP_LABEL */
3414 static bool cfs_bandwidth_used(void)
3419 void cfs_bandwidth_usage_inc(void) {}
3420 void cfs_bandwidth_usage_dec(void) {}
3421 #endif /* HAVE_JUMP_LABEL */
3424 * default period for cfs group bandwidth.
3425 * default: 0.1s, units: nanoseconds
3427 static inline u64 default_cfs_period(void)
3429 return 100000000ULL;
3432 static inline u64 sched_cfs_bandwidth_slice(void)
3434 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3438 * Replenish runtime according to assigned quota and update expiration time.
3439 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3440 * additional synchronization around rq->lock.
3442 * requires cfs_b->lock
3444 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3448 if (cfs_b->quota == RUNTIME_INF)
3451 now = sched_clock_cpu(smp_processor_id());
3452 cfs_b->runtime = cfs_b->quota;
3453 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3456 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3458 return &tg->cfs_bandwidth;
3461 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3462 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3464 if (unlikely(cfs_rq->throttle_count))
3465 return cfs_rq->throttled_clock_task;
3467 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3470 /* returns 0 on failure to allocate runtime */
3471 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3473 struct task_group *tg = cfs_rq->tg;
3474 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3475 u64 amount = 0, min_amount, expires;
3477 /* note: this is a positive sum as runtime_remaining <= 0 */
3478 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3480 raw_spin_lock(&cfs_b->lock);
3481 if (cfs_b->quota == RUNTIME_INF)
3482 amount = min_amount;
3484 start_cfs_bandwidth(cfs_b);
3486 if (cfs_b->runtime > 0) {
3487 amount = min(cfs_b->runtime, min_amount);
3488 cfs_b->runtime -= amount;
3492 expires = cfs_b->runtime_expires;
3493 raw_spin_unlock(&cfs_b->lock);
3495 cfs_rq->runtime_remaining += amount;
3497 * we may have advanced our local expiration to account for allowed
3498 * spread between our sched_clock and the one on which runtime was
3501 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3502 cfs_rq->runtime_expires = expires;
3504 return cfs_rq->runtime_remaining > 0;
3508 * Note: This depends on the synchronization provided by sched_clock and the
3509 * fact that rq->clock snapshots this value.
3511 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3513 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3515 /* if the deadline is ahead of our clock, nothing to do */
3516 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3519 if (cfs_rq->runtime_remaining < 0)
3523 * If the local deadline has passed we have to consider the
3524 * possibility that our sched_clock is 'fast' and the global deadline
3525 * has not truly expired.
3527 * Fortunately we can check determine whether this the case by checking
3528 * whether the global deadline has advanced. It is valid to compare
3529 * cfs_b->runtime_expires without any locks since we only care about
3530 * exact equality, so a partial write will still work.
3533 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3534 /* extend local deadline, drift is bounded above by 2 ticks */
3535 cfs_rq->runtime_expires += TICK_NSEC;
3537 /* global deadline is ahead, expiration has passed */
3538 cfs_rq->runtime_remaining = 0;
3542 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3544 /* dock delta_exec before expiring quota (as it could span periods) */
3545 cfs_rq->runtime_remaining -= delta_exec;
3546 expire_cfs_rq_runtime(cfs_rq);
3548 if (likely(cfs_rq->runtime_remaining > 0))
3552 * if we're unable to extend our runtime we resched so that the active
3553 * hierarchy can be throttled
3555 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3556 resched_curr(rq_of(cfs_rq));
3559 static __always_inline
3560 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3562 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3565 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3568 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3570 return cfs_bandwidth_used() && cfs_rq->throttled;
3573 /* check whether cfs_rq, or any parent, is throttled */
3574 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3576 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3580 * Ensure that neither of the group entities corresponding to src_cpu or
3581 * dest_cpu are members of a throttled hierarchy when performing group
3582 * load-balance operations.
3584 static inline int throttled_lb_pair(struct task_group *tg,
3585 int src_cpu, int dest_cpu)
3587 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3589 src_cfs_rq = tg->cfs_rq[src_cpu];
3590 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3592 return throttled_hierarchy(src_cfs_rq) ||
3593 throttled_hierarchy(dest_cfs_rq);
3596 /* updated child weight may affect parent so we have to do this bottom up */
3597 static int tg_unthrottle_up(struct task_group *tg, void *data)
3599 struct rq *rq = data;
3600 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3602 cfs_rq->throttle_count--;
3604 if (!cfs_rq->throttle_count) {
3605 /* adjust cfs_rq_clock_task() */
3606 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3607 cfs_rq->throttled_clock_task;
3614 static int tg_throttle_down(struct task_group *tg, void *data)
3616 struct rq *rq = data;
3617 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3619 /* group is entering throttled state, stop time */
3620 if (!cfs_rq->throttle_count)
3621 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3622 cfs_rq->throttle_count++;
3627 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3629 struct rq *rq = rq_of(cfs_rq);
3630 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3631 struct sched_entity *se;
3632 long task_delta, dequeue = 1;
3635 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3637 /* freeze hierarchy runnable averages while throttled */
3639 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3642 task_delta = cfs_rq->h_nr_running;
3643 for_each_sched_entity(se) {
3644 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3645 /* throttled entity or throttle-on-deactivate */
3650 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3651 qcfs_rq->h_nr_running -= task_delta;
3653 if (qcfs_rq->load.weight)
3658 sub_nr_running(rq, task_delta);
3660 cfs_rq->throttled = 1;
3661 cfs_rq->throttled_clock = rq_clock(rq);
3662 raw_spin_lock(&cfs_b->lock);
3663 empty = list_empty(&cfs_b->throttled_cfs_rq);
3666 * Add to the _head_ of the list, so that an already-started
3667 * distribute_cfs_runtime will not see us
3669 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3672 * If we're the first throttled task, make sure the bandwidth
3676 start_cfs_bandwidth(cfs_b);
3678 raw_spin_unlock(&cfs_b->lock);
3681 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3683 struct rq *rq = rq_of(cfs_rq);
3684 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3685 struct sched_entity *se;
3689 se = cfs_rq->tg->se[cpu_of(rq)];
3691 cfs_rq->throttled = 0;
3693 update_rq_clock(rq);
3695 raw_spin_lock(&cfs_b->lock);
3696 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3697 list_del_rcu(&cfs_rq->throttled_list);
3698 raw_spin_unlock(&cfs_b->lock);
3700 /* update hierarchical throttle state */
3701 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3703 if (!cfs_rq->load.weight)
3706 task_delta = cfs_rq->h_nr_running;
3707 for_each_sched_entity(se) {
3711 cfs_rq = cfs_rq_of(se);
3713 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3714 cfs_rq->h_nr_running += task_delta;
3716 if (cfs_rq_throttled(cfs_rq))
3721 add_nr_running(rq, task_delta);
3723 /* determine whether we need to wake up potentially idle cpu */
3724 if (rq->curr == rq->idle && rq->cfs.nr_running)
3728 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3729 u64 remaining, u64 expires)
3731 struct cfs_rq *cfs_rq;
3733 u64 starting_runtime = remaining;
3736 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3738 struct rq *rq = rq_of(cfs_rq);
3740 raw_spin_lock(&rq->lock);
3741 if (!cfs_rq_throttled(cfs_rq))
3744 runtime = -cfs_rq->runtime_remaining + 1;
3745 if (runtime > remaining)
3746 runtime = remaining;
3747 remaining -= runtime;
3749 cfs_rq->runtime_remaining += runtime;
3750 cfs_rq->runtime_expires = expires;
3752 /* we check whether we're throttled above */
3753 if (cfs_rq->runtime_remaining > 0)
3754 unthrottle_cfs_rq(cfs_rq);
3757 raw_spin_unlock(&rq->lock);
3764 return starting_runtime - remaining;
3768 * Responsible for refilling a task_group's bandwidth and unthrottling its
3769 * cfs_rqs as appropriate. If there has been no activity within the last
3770 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3771 * used to track this state.
3773 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3775 u64 runtime, runtime_expires;
3778 /* no need to continue the timer with no bandwidth constraint */
3779 if (cfs_b->quota == RUNTIME_INF)
3780 goto out_deactivate;
3782 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3783 cfs_b->nr_periods += overrun;
3786 * idle depends on !throttled (for the case of a large deficit), and if
3787 * we're going inactive then everything else can be deferred
3789 if (cfs_b->idle && !throttled)
3790 goto out_deactivate;
3792 __refill_cfs_bandwidth_runtime(cfs_b);
3795 /* mark as potentially idle for the upcoming period */
3800 /* account preceding periods in which throttling occurred */
3801 cfs_b->nr_throttled += overrun;
3803 runtime_expires = cfs_b->runtime_expires;
3806 * This check is repeated as we are holding onto the new bandwidth while
3807 * we unthrottle. This can potentially race with an unthrottled group
3808 * trying to acquire new bandwidth from the global pool. This can result
3809 * in us over-using our runtime if it is all used during this loop, but
3810 * only by limited amounts in that extreme case.
3812 while (throttled && cfs_b->runtime > 0) {
3813 runtime = cfs_b->runtime;
3814 raw_spin_unlock(&cfs_b->lock);
3815 /* we can't nest cfs_b->lock while distributing bandwidth */
3816 runtime = distribute_cfs_runtime(cfs_b, runtime,
3818 raw_spin_lock(&cfs_b->lock);
3820 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3822 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3826 * While we are ensured activity in the period following an
3827 * unthrottle, this also covers the case in which the new bandwidth is
3828 * insufficient to cover the existing bandwidth deficit. (Forcing the
3829 * timer to remain active while there are any throttled entities.)
3839 /* a cfs_rq won't donate quota below this amount */
3840 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3841 /* minimum remaining period time to redistribute slack quota */
3842 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3843 /* how long we wait to gather additional slack before distributing */
3844 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3847 * Are we near the end of the current quota period?
3849 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3850 * hrtimer base being cleared by hrtimer_start. In the case of
3851 * migrate_hrtimers, base is never cleared, so we are fine.
3853 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3855 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3858 /* if the call-back is running a quota refresh is already occurring */
3859 if (hrtimer_callback_running(refresh_timer))
3862 /* is a quota refresh about to occur? */
3863 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3864 if (remaining < min_expire)
3870 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3872 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3874 /* if there's a quota refresh soon don't bother with slack */
3875 if (runtime_refresh_within(cfs_b, min_left))
3878 hrtimer_start(&cfs_b->slack_timer,
3879 ns_to_ktime(cfs_bandwidth_slack_period),
3883 /* we know any runtime found here is valid as update_curr() precedes return */
3884 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3886 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3887 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3889 if (slack_runtime <= 0)
3892 raw_spin_lock(&cfs_b->lock);
3893 if (cfs_b->quota != RUNTIME_INF &&
3894 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3895 cfs_b->runtime += slack_runtime;
3897 /* we are under rq->lock, defer unthrottling using a timer */
3898 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3899 !list_empty(&cfs_b->throttled_cfs_rq))
3900 start_cfs_slack_bandwidth(cfs_b);
3902 raw_spin_unlock(&cfs_b->lock);
3904 /* even if it's not valid for return we don't want to try again */
3905 cfs_rq->runtime_remaining -= slack_runtime;
3908 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3910 if (!cfs_bandwidth_used())
3913 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3916 __return_cfs_rq_runtime(cfs_rq);
3920 * This is done with a timer (instead of inline with bandwidth return) since
3921 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3923 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3925 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3928 /* confirm we're still not at a refresh boundary */
3929 raw_spin_lock(&cfs_b->lock);
3930 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3931 raw_spin_unlock(&cfs_b->lock);
3935 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3936 runtime = cfs_b->runtime;
3938 expires = cfs_b->runtime_expires;
3939 raw_spin_unlock(&cfs_b->lock);
3944 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3946 raw_spin_lock(&cfs_b->lock);
3947 if (expires == cfs_b->runtime_expires)
3948 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3949 raw_spin_unlock(&cfs_b->lock);
3953 * When a group wakes up we want to make sure that its quota is not already
3954 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3955 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3957 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3959 if (!cfs_bandwidth_used())
3962 /* an active group must be handled by the update_curr()->put() path */
3963 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3966 /* ensure the group is not already throttled */
3967 if (cfs_rq_throttled(cfs_rq))
3970 /* update runtime allocation */
3971 account_cfs_rq_runtime(cfs_rq, 0);
3972 if (cfs_rq->runtime_remaining <= 0)
3973 throttle_cfs_rq(cfs_rq);
3976 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3977 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3979 if (!cfs_bandwidth_used())
3982 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3986 * it's possible for a throttled entity to be forced into a running
3987 * state (e.g. set_curr_task), in this case we're finished.
3989 if (cfs_rq_throttled(cfs_rq))
3992 throttle_cfs_rq(cfs_rq);
3996 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3998 struct cfs_bandwidth *cfs_b =
3999 container_of(timer, struct cfs_bandwidth, slack_timer);
4001 do_sched_cfs_slack_timer(cfs_b);
4003 return HRTIMER_NORESTART;
4006 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4008 struct cfs_bandwidth *cfs_b =
4009 container_of(timer, struct cfs_bandwidth, period_timer);
4013 raw_spin_lock(&cfs_b->lock);
4015 overrun = hrtimer_forward_now(timer, cfs_b->period);
4019 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4022 cfs_b->period_active = 0;
4023 raw_spin_unlock(&cfs_b->lock);
4025 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4028 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4030 raw_spin_lock_init(&cfs_b->lock);
4032 cfs_b->quota = RUNTIME_INF;
4033 cfs_b->period = ns_to_ktime(default_cfs_period());
4035 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4036 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4037 cfs_b->period_timer.function = sched_cfs_period_timer;
4038 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4039 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4042 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4044 cfs_rq->runtime_enabled = 0;
4045 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4048 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4050 lockdep_assert_held(&cfs_b->lock);
4052 if (!cfs_b->period_active) {
4053 cfs_b->period_active = 1;
4054 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4055 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4059 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4061 /* init_cfs_bandwidth() was not called */
4062 if (!cfs_b->throttled_cfs_rq.next)
4065 hrtimer_cancel(&cfs_b->period_timer);
4066 hrtimer_cancel(&cfs_b->slack_timer);
4069 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4071 struct cfs_rq *cfs_rq;
4073 for_each_leaf_cfs_rq(rq, cfs_rq) {
4074 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4076 raw_spin_lock(&cfs_b->lock);
4077 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4078 raw_spin_unlock(&cfs_b->lock);
4082 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4084 struct cfs_rq *cfs_rq;
4086 for_each_leaf_cfs_rq(rq, cfs_rq) {
4087 if (!cfs_rq->runtime_enabled)
4091 * clock_task is not advancing so we just need to make sure
4092 * there's some valid quota amount
4094 cfs_rq->runtime_remaining = 1;
4096 * Offline rq is schedulable till cpu is completely disabled
4097 * in take_cpu_down(), so we prevent new cfs throttling here.
4099 cfs_rq->runtime_enabled = 0;
4101 if (cfs_rq_throttled(cfs_rq))
4102 unthrottle_cfs_rq(cfs_rq);
4106 #else /* CONFIG_CFS_BANDWIDTH */
4107 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4109 return rq_clock_task(rq_of(cfs_rq));
4112 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4113 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4114 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4115 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4117 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4122 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4127 static inline int throttled_lb_pair(struct task_group *tg,
4128 int src_cpu, int dest_cpu)
4133 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4135 #ifdef CONFIG_FAIR_GROUP_SCHED
4136 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4139 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4143 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4144 static inline void update_runtime_enabled(struct rq *rq) {}
4145 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4147 #endif /* CONFIG_CFS_BANDWIDTH */
4149 /**************************************************
4150 * CFS operations on tasks:
4153 #ifdef CONFIG_SCHED_HRTICK
4154 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4156 struct sched_entity *se = &p->se;
4157 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4159 WARN_ON(task_rq(p) != rq);
4161 if (cfs_rq->nr_running > 1) {
4162 u64 slice = sched_slice(cfs_rq, se);
4163 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4164 s64 delta = slice - ran;
4171 hrtick_start(rq, delta);
4176 * called from enqueue/dequeue and updates the hrtick when the
4177 * current task is from our class and nr_running is low enough
4180 static void hrtick_update(struct rq *rq)
4182 struct task_struct *curr = rq->curr;
4184 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4187 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4188 hrtick_start_fair(rq, curr);
4190 #else /* !CONFIG_SCHED_HRTICK */
4192 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4196 static inline void hrtick_update(struct rq *rq)
4202 static bool cpu_overutilized(int cpu);
4203 static inline unsigned long boosted_cpu_util(int cpu);
4205 #define boosted_cpu_util(cpu) cpu_util(cpu)
4209 static void update_capacity_of(int cpu)
4211 unsigned long req_cap;
4216 /* Convert scale-invariant capacity to cpu. */
4217 req_cap = boosted_cpu_util(cpu);
4218 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4219 set_cfs_cpu_capacity(cpu, true, req_cap);
4224 * The enqueue_task method is called before nr_running is
4225 * increased. Here we update the fair scheduling stats and
4226 * then put the task into the rbtree:
4229 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4231 struct cfs_rq *cfs_rq;
4232 struct sched_entity *se = &p->se;
4234 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4235 int task_wakeup = flags & ENQUEUE_WAKEUP;
4238 for_each_sched_entity(se) {
4241 cfs_rq = cfs_rq_of(se);
4242 enqueue_entity(cfs_rq, se, flags);
4245 * end evaluation on encountering a throttled cfs_rq
4247 * note: in the case of encountering a throttled cfs_rq we will
4248 * post the final h_nr_running increment below.
4250 if (cfs_rq_throttled(cfs_rq))
4252 cfs_rq->h_nr_running++;
4253 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4255 flags = ENQUEUE_WAKEUP;
4258 for_each_sched_entity(se) {
4259 cfs_rq = cfs_rq_of(se);
4260 cfs_rq->h_nr_running++;
4261 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4263 if (cfs_rq_throttled(cfs_rq))
4266 update_load_avg(se, 1);
4267 update_cfs_shares(cfs_rq);
4271 add_nr_running(rq, 1);
4276 * Update SchedTune accounting.
4278 * We do it before updating the CPU capacity to ensure the
4279 * boost value of the current task is accounted for in the
4280 * selection of the OPP.
4282 * We do it also in the case where we enqueue a throttled task;
4283 * we could argue that a throttled task should not boost a CPU,
4285 * a) properly implementing CPU boosting considering throttled
4286 * tasks will increase a lot the complexity of the solution
4287 * b) it's not easy to quantify the benefits introduced by
4288 * such a more complex solution.
4289 * Thus, for the time being we go for the simple solution and boost
4290 * also for throttled RQs.
4292 schedtune_enqueue_task(p, cpu_of(rq));
4295 walt_inc_cumulative_runnable_avg(rq, p);
4296 if (!task_new && !rq->rd->overutilized &&
4297 cpu_overutilized(rq->cpu)) {
4298 rq->rd->overutilized = true;
4299 trace_sched_overutilized(true);
4303 * We want to potentially trigger a freq switch
4304 * request only for tasks that are waking up; this is
4305 * because we get here also during load balancing, but
4306 * in these cases it seems wise to trigger as single
4307 * request after load balancing is done.
4309 if (task_new || task_wakeup)
4310 update_capacity_of(cpu_of(rq));
4313 #endif /* CONFIG_SMP */
4317 static void set_next_buddy(struct sched_entity *se);
4320 * The dequeue_task method is called before nr_running is
4321 * decreased. We remove the task from the rbtree and
4322 * update the fair scheduling stats:
4324 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4326 struct cfs_rq *cfs_rq;
4327 struct sched_entity *se = &p->se;
4328 int task_sleep = flags & DEQUEUE_SLEEP;
4330 for_each_sched_entity(se) {
4331 cfs_rq = cfs_rq_of(se);
4332 dequeue_entity(cfs_rq, se, flags);
4335 * end evaluation on encountering a throttled cfs_rq
4337 * note: in the case of encountering a throttled cfs_rq we will
4338 * post the final h_nr_running decrement below.
4340 if (cfs_rq_throttled(cfs_rq))
4342 cfs_rq->h_nr_running--;
4343 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4345 /* Don't dequeue parent if it has other entities besides us */
4346 if (cfs_rq->load.weight) {
4348 * Bias pick_next to pick a task from this cfs_rq, as
4349 * p is sleeping when it is within its sched_slice.
4351 if (task_sleep && parent_entity(se))
4352 set_next_buddy(parent_entity(se));
4354 /* avoid re-evaluating load for this entity */
4355 se = parent_entity(se);
4358 flags |= DEQUEUE_SLEEP;
4361 for_each_sched_entity(se) {
4362 cfs_rq = cfs_rq_of(se);
4363 cfs_rq->h_nr_running--;
4364 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4366 if (cfs_rq_throttled(cfs_rq))
4369 update_load_avg(se, 1);
4370 update_cfs_shares(cfs_rq);
4374 sub_nr_running(rq, 1);
4379 * Update SchedTune accounting
4381 * We do it before updating the CPU capacity to ensure the
4382 * boost value of the current task is accounted for in the
4383 * selection of the OPP.
4385 schedtune_dequeue_task(p, cpu_of(rq));
4388 walt_dec_cumulative_runnable_avg(rq, p);
4391 * We want to potentially trigger a freq switch
4392 * request only for tasks that are going to sleep;
4393 * this is because we get here also during load
4394 * balancing, but in these cases it seems wise to
4395 * trigger as single request after load balancing is
4399 if (rq->cfs.nr_running)
4400 update_capacity_of(cpu_of(rq));
4401 else if (sched_freq())
4402 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4406 #endif /* CONFIG_SMP */
4414 * per rq 'load' arrray crap; XXX kill this.
4418 * The exact cpuload at various idx values, calculated at every tick would be
4419 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4421 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4422 * on nth tick when cpu may be busy, then we have:
4423 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4424 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4426 * decay_load_missed() below does efficient calculation of
4427 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4428 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4430 * The calculation is approximated on a 128 point scale.
4431 * degrade_zero_ticks is the number of ticks after which load at any
4432 * particular idx is approximated to be zero.
4433 * degrade_factor is a precomputed table, a row for each load idx.
4434 * Each column corresponds to degradation factor for a power of two ticks,
4435 * based on 128 point scale.
4437 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4438 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4440 * With this power of 2 load factors, we can degrade the load n times
4441 * by looking at 1 bits in n and doing as many mult/shift instead of
4442 * n mult/shifts needed by the exact degradation.
4444 #define DEGRADE_SHIFT 7
4445 static const unsigned char
4446 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4447 static const unsigned char
4448 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4449 {0, 0, 0, 0, 0, 0, 0, 0},
4450 {64, 32, 8, 0, 0, 0, 0, 0},
4451 {96, 72, 40, 12, 1, 0, 0},
4452 {112, 98, 75, 43, 15, 1, 0},
4453 {120, 112, 98, 76, 45, 16, 2} };
4456 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4457 * would be when CPU is idle and so we just decay the old load without
4458 * adding any new load.
4460 static unsigned long
4461 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4465 if (!missed_updates)
4468 if (missed_updates >= degrade_zero_ticks[idx])
4472 return load >> missed_updates;
4474 while (missed_updates) {
4475 if (missed_updates % 2)
4476 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4478 missed_updates >>= 1;
4485 * Update rq->cpu_load[] statistics. This function is usually called every
4486 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4487 * every tick. We fix it up based on jiffies.
4489 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4490 unsigned long pending_updates)
4494 this_rq->nr_load_updates++;
4496 /* Update our load: */
4497 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4498 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4499 unsigned long old_load, new_load;
4501 /* scale is effectively 1 << i now, and >> i divides by scale */
4503 old_load = this_rq->cpu_load[i];
4504 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4505 new_load = this_load;
4507 * Round up the averaging division if load is increasing. This
4508 * prevents us from getting stuck on 9 if the load is 10, for
4511 if (new_load > old_load)
4512 new_load += scale - 1;
4514 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4517 sched_avg_update(this_rq);
4520 /* Used instead of source_load when we know the type == 0 */
4521 static unsigned long weighted_cpuload(const int cpu)
4523 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4526 #ifdef CONFIG_NO_HZ_COMMON
4528 * There is no sane way to deal with nohz on smp when using jiffies because the
4529 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4530 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4532 * Therefore we cannot use the delta approach from the regular tick since that
4533 * would seriously skew the load calculation. However we'll make do for those
4534 * updates happening while idle (nohz_idle_balance) or coming out of idle
4535 * (tick_nohz_idle_exit).
4537 * This means we might still be one tick off for nohz periods.
4541 * Called from nohz_idle_balance() to update the load ratings before doing the
4544 static void update_idle_cpu_load(struct rq *this_rq)
4546 unsigned long curr_jiffies = READ_ONCE(jiffies);
4547 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4548 unsigned long pending_updates;
4551 * bail if there's load or we're actually up-to-date.
4553 if (load || curr_jiffies == this_rq->last_load_update_tick)
4556 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4557 this_rq->last_load_update_tick = curr_jiffies;
4559 __update_cpu_load(this_rq, load, pending_updates);
4563 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4565 void update_cpu_load_nohz(void)
4567 struct rq *this_rq = this_rq();
4568 unsigned long curr_jiffies = READ_ONCE(jiffies);
4569 unsigned long pending_updates;
4571 if (curr_jiffies == this_rq->last_load_update_tick)
4574 raw_spin_lock(&this_rq->lock);
4575 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4576 if (pending_updates) {
4577 this_rq->last_load_update_tick = curr_jiffies;
4579 * We were idle, this means load 0, the current load might be
4580 * !0 due to remote wakeups and the sort.
4582 __update_cpu_load(this_rq, 0, pending_updates);
4584 raw_spin_unlock(&this_rq->lock);
4586 #endif /* CONFIG_NO_HZ */
4589 * Called from scheduler_tick()
4591 void update_cpu_load_active(struct rq *this_rq)
4593 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4595 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4597 this_rq->last_load_update_tick = jiffies;
4598 __update_cpu_load(this_rq, load, 1);
4602 * Return a low guess at the load of a migration-source cpu weighted
4603 * according to the scheduling class and "nice" value.
4605 * We want to under-estimate the load of migration sources, to
4606 * balance conservatively.
4608 static unsigned long source_load(int cpu, int type)
4610 struct rq *rq = cpu_rq(cpu);
4611 unsigned long total = weighted_cpuload(cpu);
4613 if (type == 0 || !sched_feat(LB_BIAS))
4616 return min(rq->cpu_load[type-1], total);
4620 * Return a high guess at the load of a migration-target cpu weighted
4621 * according to the scheduling class and "nice" value.
4623 static unsigned long target_load(int cpu, int type)
4625 struct rq *rq = cpu_rq(cpu);
4626 unsigned long total = weighted_cpuload(cpu);
4628 if (type == 0 || !sched_feat(LB_BIAS))
4631 return max(rq->cpu_load[type-1], total);
4635 static unsigned long cpu_avg_load_per_task(int cpu)
4637 struct rq *rq = cpu_rq(cpu);
4638 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4639 unsigned long load_avg = weighted_cpuload(cpu);
4642 return load_avg / nr_running;
4647 static void record_wakee(struct task_struct *p)
4650 * Rough decay (wiping) for cost saving, don't worry
4651 * about the boundary, really active task won't care
4654 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4655 current->wakee_flips >>= 1;
4656 current->wakee_flip_decay_ts = jiffies;
4659 if (current->last_wakee != p) {
4660 current->last_wakee = p;
4661 current->wakee_flips++;
4665 static void task_waking_fair(struct task_struct *p)
4667 struct sched_entity *se = &p->se;
4668 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4671 #ifndef CONFIG_64BIT
4672 u64 min_vruntime_copy;
4675 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4677 min_vruntime = cfs_rq->min_vruntime;
4678 } while (min_vruntime != min_vruntime_copy);
4680 min_vruntime = cfs_rq->min_vruntime;
4683 se->vruntime -= min_vruntime;
4687 #ifdef CONFIG_FAIR_GROUP_SCHED
4689 * effective_load() calculates the load change as seen from the root_task_group
4691 * Adding load to a group doesn't make a group heavier, but can cause movement
4692 * of group shares between cpus. Assuming the shares were perfectly aligned one
4693 * can calculate the shift in shares.
4695 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4696 * on this @cpu and results in a total addition (subtraction) of @wg to the
4697 * total group weight.
4699 * Given a runqueue weight distribution (rw_i) we can compute a shares
4700 * distribution (s_i) using:
4702 * s_i = rw_i / \Sum rw_j (1)
4704 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4705 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4706 * shares distribution (s_i):
4708 * rw_i = { 2, 4, 1, 0 }
4709 * s_i = { 2/7, 4/7, 1/7, 0 }
4711 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4712 * task used to run on and the CPU the waker is running on), we need to
4713 * compute the effect of waking a task on either CPU and, in case of a sync
4714 * wakeup, compute the effect of the current task going to sleep.
4716 * So for a change of @wl to the local @cpu with an overall group weight change
4717 * of @wl we can compute the new shares distribution (s'_i) using:
4719 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4721 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4722 * differences in waking a task to CPU 0. The additional task changes the
4723 * weight and shares distributions like:
4725 * rw'_i = { 3, 4, 1, 0 }
4726 * s'_i = { 3/8, 4/8, 1/8, 0 }
4728 * We can then compute the difference in effective weight by using:
4730 * dw_i = S * (s'_i - s_i) (3)
4732 * Where 'S' is the group weight as seen by its parent.
4734 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4735 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4736 * 4/7) times the weight of the group.
4738 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4740 struct sched_entity *se = tg->se[cpu];
4742 if (!tg->parent) /* the trivial, non-cgroup case */
4745 for_each_sched_entity(se) {
4746 struct cfs_rq *cfs_rq = se->my_q;
4747 long W, w = cfs_rq_load_avg(cfs_rq);
4752 * W = @wg + \Sum rw_j
4754 W = wg + atomic_long_read(&tg->load_avg);
4756 /* Ensure \Sum rw_j >= rw_i */
4757 W -= cfs_rq->tg_load_avg_contrib;
4766 * wl = S * s'_i; see (2)
4769 wl = (w * (long)tg->shares) / W;
4774 * Per the above, wl is the new se->load.weight value; since
4775 * those are clipped to [MIN_SHARES, ...) do so now. See
4776 * calc_cfs_shares().
4778 if (wl < MIN_SHARES)
4782 * wl = dw_i = S * (s'_i - s_i); see (3)
4784 wl -= se->avg.load_avg;
4787 * Recursively apply this logic to all parent groups to compute
4788 * the final effective load change on the root group. Since
4789 * only the @tg group gets extra weight, all parent groups can
4790 * only redistribute existing shares. @wl is the shift in shares
4791 * resulting from this level per the above.
4800 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4808 * Returns the current capacity of cpu after applying both
4809 * cpu and freq scaling.
4811 unsigned long capacity_curr_of(int cpu)
4813 return cpu_rq(cpu)->cpu_capacity_orig *
4814 arch_scale_freq_capacity(NULL, cpu)
4815 >> SCHED_CAPACITY_SHIFT;
4818 static inline bool energy_aware(void)
4820 return sched_feat(ENERGY_AWARE);
4824 struct sched_group *sg_top;
4825 struct sched_group *sg_cap;
4832 struct task_struct *task;
4847 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4848 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4849 * energy calculations. Using the scale-invariant util returned by
4850 * cpu_util() and approximating scale-invariant util by:
4852 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4854 * the normalized util can be found using the specific capacity.
4856 * capacity = capacity_orig * curr_freq/max_freq
4858 * norm_util = running_time/time ~ util/capacity
4860 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4862 int util = __cpu_util(cpu, delta);
4864 if (util >= capacity)
4865 return SCHED_CAPACITY_SCALE;
4867 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4870 static int calc_util_delta(struct energy_env *eenv, int cpu)
4872 if (cpu == eenv->src_cpu)
4873 return -eenv->util_delta;
4874 if (cpu == eenv->dst_cpu)
4875 return eenv->util_delta;
4880 unsigned long group_max_util(struct energy_env *eenv)
4883 unsigned long max_util = 0;
4885 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4886 delta = calc_util_delta(eenv, i);
4887 max_util = max(max_util, __cpu_util(i, delta));
4894 * group_norm_util() returns the approximated group util relative to it's
4895 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4896 * energy calculations. Since task executions may or may not overlap in time in
4897 * the group the true normalized util is between max(cpu_norm_util(i)) and
4898 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4899 * latter is used as the estimate as it leads to a more pessimistic energy
4900 * estimate (more busy).
4903 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4906 unsigned long util_sum = 0;
4907 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4909 for_each_cpu(i, sched_group_cpus(sg)) {
4910 delta = calc_util_delta(eenv, i);
4911 util_sum += __cpu_norm_util(i, capacity, delta);
4914 if (util_sum > SCHED_CAPACITY_SCALE)
4915 return SCHED_CAPACITY_SCALE;
4919 static int find_new_capacity(struct energy_env *eenv,
4920 const struct sched_group_energy const *sge)
4923 unsigned long util = group_max_util(eenv);
4925 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4926 if (sge->cap_states[idx].cap >= util)
4930 eenv->cap_idx = idx;
4935 static int group_idle_state(struct sched_group *sg)
4937 int i, state = INT_MAX;
4939 /* Find the shallowest idle state in the sched group. */
4940 for_each_cpu(i, sched_group_cpus(sg))
4941 state = min(state, idle_get_state_idx(cpu_rq(i)));
4943 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4950 * sched_group_energy(): Computes the absolute energy consumption of cpus
4951 * belonging to the sched_group including shared resources shared only by
4952 * members of the group. Iterates over all cpus in the hierarchy below the
4953 * sched_group starting from the bottom working it's way up before going to
4954 * the next cpu until all cpus are covered at all levels. The current
4955 * implementation is likely to gather the same util statistics multiple times.
4956 * This can probably be done in a faster but more complex way.
4957 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4959 static int sched_group_energy(struct energy_env *eenv)
4961 struct sched_domain *sd;
4962 int cpu, total_energy = 0;
4963 struct cpumask visit_cpus;
4964 struct sched_group *sg;
4966 WARN_ON(!eenv->sg_top->sge);
4968 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4970 while (!cpumask_empty(&visit_cpus)) {
4971 struct sched_group *sg_shared_cap = NULL;
4973 cpu = cpumask_first(&visit_cpus);
4976 * Is the group utilization affected by cpus outside this
4979 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4983 * We most probably raced with hotplug; returning a
4984 * wrong energy estimation is better than entering an
4990 sg_shared_cap = sd->parent->groups;
4992 for_each_domain(cpu, sd) {
4995 /* Has this sched_domain already been visited? */
4996 if (sd->child && group_first_cpu(sg) != cpu)
5000 unsigned long group_util;
5001 int sg_busy_energy, sg_idle_energy;
5002 int cap_idx, idle_idx;
5004 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5005 eenv->sg_cap = sg_shared_cap;
5009 cap_idx = find_new_capacity(eenv, sg->sge);
5011 if (sg->group_weight == 1) {
5012 /* Remove capacity of src CPU (before task move) */
5013 if (eenv->util_delta == 0 &&
5014 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5015 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5016 eenv->cap.delta -= eenv->cap.before;
5018 /* Add capacity of dst CPU (after task move) */
5019 if (eenv->util_delta != 0 &&
5020 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5021 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5022 eenv->cap.delta += eenv->cap.after;
5026 idle_idx = group_idle_state(sg);
5027 group_util = group_norm_util(eenv, sg);
5028 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5029 >> SCHED_CAPACITY_SHIFT;
5030 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5031 * sg->sge->idle_states[idle_idx].power)
5032 >> SCHED_CAPACITY_SHIFT;
5034 total_energy += sg_busy_energy + sg_idle_energy;
5037 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5039 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5042 } while (sg = sg->next, sg != sd->groups);
5045 cpumask_clear_cpu(cpu, &visit_cpus);
5049 eenv->energy = total_energy;
5053 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5055 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5059 * energy_diff(): Estimate the energy impact of changing the utilization
5060 * distribution. eenv specifies the change: utilisation amount, source, and
5061 * destination cpu. Source or destination cpu may be -1 in which case the
5062 * utilization is removed from or added to the system (e.g. task wake-up). If
5063 * both are specified, the utilization is migrated.
5065 static inline int __energy_diff(struct energy_env *eenv)
5067 struct sched_domain *sd;
5068 struct sched_group *sg;
5069 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5071 struct energy_env eenv_before = {
5073 .src_cpu = eenv->src_cpu,
5074 .dst_cpu = eenv->dst_cpu,
5075 .nrg = { 0, 0, 0, 0},
5079 if (eenv->src_cpu == eenv->dst_cpu)
5082 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5083 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5086 return 0; /* Error */
5091 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5092 eenv_before.sg_top = eenv->sg_top = sg;
5094 if (sched_group_energy(&eenv_before))
5095 return 0; /* Invalid result abort */
5096 energy_before += eenv_before.energy;
5098 /* Keep track of SRC cpu (before) capacity */
5099 eenv->cap.before = eenv_before.cap.before;
5100 eenv->cap.delta = eenv_before.cap.delta;
5102 if (sched_group_energy(eenv))
5103 return 0; /* Invalid result abort */
5104 energy_after += eenv->energy;
5106 } while (sg = sg->next, sg != sd->groups);
5108 eenv->nrg.before = energy_before;
5109 eenv->nrg.after = energy_after;
5110 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5113 trace_sched_energy_diff(eenv->task,
5114 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5115 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5116 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5117 eenv->nrg.delta, eenv->payoff);
5119 return eenv->nrg.diff;
5122 #ifdef CONFIG_SCHED_TUNE
5124 struct target_nrg schedtune_target_nrg;
5127 * System energy normalization
5128 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5129 * corresponding to the specified energy variation.
5132 normalize_energy(int energy_diff)
5135 #ifdef CONFIG_SCHED_DEBUG
5138 /* Check for boundaries */
5139 max_delta = schedtune_target_nrg.max_power;
5140 max_delta -= schedtune_target_nrg.min_power;
5141 WARN_ON(abs(energy_diff) >= max_delta);
5144 /* Do scaling using positive numbers to increase the range */
5145 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5147 /* Scale by energy magnitude */
5148 normalized_nrg <<= SCHED_LOAD_SHIFT;
5150 /* Normalize on max energy for target platform */
5151 normalized_nrg = reciprocal_divide(
5152 normalized_nrg, schedtune_target_nrg.rdiv);
5154 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5158 energy_diff(struct energy_env *eenv)
5160 int boost = schedtune_task_boost(eenv->task);
5163 /* Conpute "absolute" energy diff */
5164 __energy_diff(eenv);
5166 /* Return energy diff when boost margin is 0 */
5168 return eenv->nrg.diff;
5170 /* Compute normalized energy diff */
5171 nrg_delta = normalize_energy(eenv->nrg.diff);
5172 eenv->nrg.delta = nrg_delta;
5174 eenv->payoff = schedtune_accept_deltas(
5180 * When SchedTune is enabled, the energy_diff() function will return
5181 * the computed energy payoff value. Since the energy_diff() return
5182 * value is expected to be negative by its callers, this evaluation
5183 * function return a negative value each time the evaluation return a
5184 * positive payoff, which is the condition for the acceptance of
5185 * a scheduling decision
5187 return -eenv->payoff;
5189 #else /* CONFIG_SCHED_TUNE */
5190 #define energy_diff(eenv) __energy_diff(eenv)
5194 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5195 * A waker of many should wake a different task than the one last awakened
5196 * at a frequency roughly N times higher than one of its wakees. In order
5197 * to determine whether we should let the load spread vs consolodating to
5198 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5199 * partner, and a factor of lls_size higher frequency in the other. With
5200 * both conditions met, we can be relatively sure that the relationship is
5201 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5202 * being client/server, worker/dispatcher, interrupt source or whatever is
5203 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5205 static int wake_wide(struct task_struct *p)
5207 unsigned int master = current->wakee_flips;
5208 unsigned int slave = p->wakee_flips;
5209 int factor = this_cpu_read(sd_llc_size);
5212 swap(master, slave);
5213 if (slave < factor || master < slave * factor)
5218 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5220 s64 this_load, load;
5221 s64 this_eff_load, prev_eff_load;
5222 int idx, this_cpu, prev_cpu;
5223 struct task_group *tg;
5224 unsigned long weight;
5228 this_cpu = smp_processor_id();
5229 prev_cpu = task_cpu(p);
5230 load = source_load(prev_cpu, idx);
5231 this_load = target_load(this_cpu, idx);
5234 * If sync wakeup then subtract the (maximum possible)
5235 * effect of the currently running task from the load
5236 * of the current CPU:
5239 tg = task_group(current);
5240 weight = current->se.avg.load_avg;
5242 this_load += effective_load(tg, this_cpu, -weight, -weight);
5243 load += effective_load(tg, prev_cpu, 0, -weight);
5247 weight = p->se.avg.load_avg;
5250 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5251 * due to the sync cause above having dropped this_load to 0, we'll
5252 * always have an imbalance, but there's really nothing you can do
5253 * about that, so that's good too.
5255 * Otherwise check if either cpus are near enough in load to allow this
5256 * task to be woken on this_cpu.
5258 this_eff_load = 100;
5259 this_eff_load *= capacity_of(prev_cpu);
5261 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5262 prev_eff_load *= capacity_of(this_cpu);
5264 if (this_load > 0) {
5265 this_eff_load *= this_load +
5266 effective_load(tg, this_cpu, weight, weight);
5268 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5271 balanced = this_eff_load <= prev_eff_load;
5273 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5278 schedstat_inc(sd, ttwu_move_affine);
5279 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5284 static inline unsigned long task_util(struct task_struct *p)
5286 #ifdef CONFIG_SCHED_WALT
5287 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5288 unsigned long demand = p->ravg.demand;
5289 return (demand << 10) / walt_ravg_window;
5292 return p->se.avg.util_avg;
5295 unsigned int capacity_margin = 1280; /* ~20% margin */
5297 static inline unsigned long boosted_task_util(struct task_struct *task);
5299 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5301 unsigned long capacity = capacity_of(cpu);
5303 util += boosted_task_util(p);
5305 return (capacity * 1024) > (util * capacity_margin);
5308 static inline bool task_fits_max(struct task_struct *p, int cpu)
5310 unsigned long capacity = capacity_of(cpu);
5311 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5313 if (capacity == max_capacity)
5316 if (capacity * capacity_margin > max_capacity * 1024)
5319 return __task_fits(p, cpu, 0);
5322 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5324 return __task_fits(p, cpu, cpu_util(cpu));
5327 static bool cpu_overutilized(int cpu)
5329 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5332 #ifdef CONFIG_SCHED_TUNE
5335 schedtune_margin(unsigned long signal, long boost)
5337 long long margin = 0;
5340 * Signal proportional compensation (SPC)
5342 * The Boost (B) value is used to compute a Margin (M) which is
5343 * proportional to the complement of the original Signal (S):
5344 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5345 * M = B * S, if B is negative
5346 * The obtained M could be used by the caller to "boost" S.
5349 margin = SCHED_LOAD_SCALE - signal;
5352 margin = -signal * boost;
5354 * Fast integer division by constant:
5355 * Constant : (C) = 100
5356 * Precision : 0.1% (P) = 0.1
5357 * Reference : C * 100 / P (R) = 100000
5360 * Shift bits : ceil(log(R,2)) (S) = 17
5361 * Mult const : round(2^S/C) (M) = 1311
5374 schedtune_cpu_margin(unsigned long util, int cpu)
5376 int boost = schedtune_cpu_boost(cpu);
5381 return schedtune_margin(util, boost);
5385 schedtune_task_margin(struct task_struct *task)
5387 int boost = schedtune_task_boost(task);
5394 util = task_util(task);
5395 margin = schedtune_margin(util, boost);
5400 #else /* CONFIG_SCHED_TUNE */
5403 schedtune_cpu_margin(unsigned long util, int cpu)
5409 schedtune_task_margin(struct task_struct *task)
5414 #endif /* CONFIG_SCHED_TUNE */
5416 static inline unsigned long
5417 boosted_cpu_util(int cpu)
5419 unsigned long util = cpu_util(cpu);
5420 long margin = schedtune_cpu_margin(util, cpu);
5422 trace_sched_boost_cpu(cpu, util, margin);
5424 return util + margin;
5427 static inline unsigned long
5428 boosted_task_util(struct task_struct *task)
5430 unsigned long util = task_util(task);
5431 long margin = schedtune_task_margin(task);
5433 trace_sched_boost_task(task, util, margin);
5435 return util + margin;
5439 * find_idlest_group finds and returns the least busy CPU group within the
5442 static struct sched_group *
5443 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5444 int this_cpu, int sd_flag)
5446 struct sched_group *idlest = NULL, *group = sd->groups;
5447 struct sched_group *fit_group = NULL, *spare_group = NULL;
5448 unsigned long min_load = ULONG_MAX, this_load = 0;
5449 unsigned long fit_capacity = ULONG_MAX;
5450 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5451 int load_idx = sd->forkexec_idx;
5452 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5454 if (sd_flag & SD_BALANCE_WAKE)
5455 load_idx = sd->wake_idx;
5458 unsigned long load, avg_load, spare_capacity;
5462 /* Skip over this group if it has no CPUs allowed */
5463 if (!cpumask_intersects(sched_group_cpus(group),
5464 tsk_cpus_allowed(p)))
5467 local_group = cpumask_test_cpu(this_cpu,
5468 sched_group_cpus(group));
5470 /* Tally up the load of all CPUs in the group */
5473 for_each_cpu(i, sched_group_cpus(group)) {
5474 /* Bias balancing toward cpus of our domain */
5476 load = source_load(i, load_idx);
5478 load = target_load(i, load_idx);
5483 * Look for most energy-efficient group that can fit
5484 * that can fit the task.
5486 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5487 fit_capacity = capacity_of(i);
5492 * Look for group which has most spare capacity on a
5495 spare_capacity = capacity_of(i) - cpu_util(i);
5496 if (spare_capacity > max_spare_capacity) {
5497 max_spare_capacity = spare_capacity;
5498 spare_group = group;
5502 /* Adjust by relative CPU capacity of the group */
5503 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5506 this_load = avg_load;
5507 } else if (avg_load < min_load) {
5508 min_load = avg_load;
5511 } while (group = group->next, group != sd->groups);
5519 if (!idlest || 100*this_load < imbalance*min_load)
5525 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5528 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5530 unsigned long load, min_load = ULONG_MAX;
5531 unsigned int min_exit_latency = UINT_MAX;
5532 u64 latest_idle_timestamp = 0;
5533 int least_loaded_cpu = this_cpu;
5534 int shallowest_idle_cpu = -1;
5537 /* Traverse only the allowed CPUs */
5538 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5539 if (task_fits_spare(p, i)) {
5540 struct rq *rq = cpu_rq(i);
5541 struct cpuidle_state *idle = idle_get_state(rq);
5542 if (idle && idle->exit_latency < min_exit_latency) {
5544 * We give priority to a CPU whose idle state
5545 * has the smallest exit latency irrespective
5546 * of any idle timestamp.
5548 min_exit_latency = idle->exit_latency;
5549 latest_idle_timestamp = rq->idle_stamp;
5550 shallowest_idle_cpu = i;
5551 } else if (idle_cpu(i) &&
5552 (!idle || idle->exit_latency == min_exit_latency) &&
5553 rq->idle_stamp > latest_idle_timestamp) {
5555 * If equal or no active idle state, then
5556 * the most recently idled CPU might have
5559 latest_idle_timestamp = rq->idle_stamp;
5560 shallowest_idle_cpu = i;
5561 } else if (shallowest_idle_cpu == -1) {
5563 * If we haven't found an idle CPU yet
5564 * pick a non-idle one that can fit the task as
5567 shallowest_idle_cpu = i;
5569 } else if (shallowest_idle_cpu == -1) {
5570 load = weighted_cpuload(i);
5571 if (load < min_load || (load == min_load && i == this_cpu)) {
5573 least_loaded_cpu = i;
5578 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5582 * Try and locate an idle CPU in the sched_domain.
5584 static int select_idle_sibling(struct task_struct *p, int target)
5586 struct sched_domain *sd;
5587 struct sched_group *sg;
5588 int i = task_cpu(p);
5590 int best_idle_cstate = -1;
5591 int best_idle_capacity = INT_MAX;
5593 if (!sysctl_sched_cstate_aware) {
5594 if (idle_cpu(target))
5598 * If the prevous cpu is cache affine and idle, don't be stupid.
5600 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5605 * Otherwise, iterate the domains and find an elegible idle cpu.
5607 sd = rcu_dereference(per_cpu(sd_llc, target));
5608 for_each_lower_domain(sd) {
5611 if (!cpumask_intersects(sched_group_cpus(sg),
5612 tsk_cpus_allowed(p)))
5615 if (sysctl_sched_cstate_aware) {
5616 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5617 struct rq *rq = cpu_rq(i);
5618 int idle_idx = idle_get_state_idx(rq);
5619 unsigned long new_usage = boosted_task_util(p);
5620 unsigned long capacity_orig = capacity_orig_of(i);
5621 if (new_usage > capacity_orig || !idle_cpu(i))
5624 if (i == target && new_usage <= capacity_curr_of(target))
5627 if (best_idle < 0 || (idle_idx < best_idle_cstate && capacity_orig <= best_idle_capacity)) {
5629 best_idle_cstate = idle_idx;
5630 best_idle_capacity = capacity_orig;
5634 for_each_cpu(i, sched_group_cpus(sg)) {
5635 if (i == target || !idle_cpu(i))
5639 target = cpumask_first_and(sched_group_cpus(sg),
5640 tsk_cpus_allowed(p));
5645 } while (sg != sd->groups);
5654 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
5657 int target_cpu = -1;
5658 int target_util = 0;
5659 int backup_capacity = 0;
5660 int best_idle_cpu = -1;
5661 int best_idle_cstate = INT_MAX;
5662 int backup_cpu = -1;
5663 unsigned long task_util_boosted, new_util;
5665 task_util_boosted = boosted_task_util(p);
5666 for (iter_cpu = 0; iter_cpu < NR_CPUS; iter_cpu++) {
5672 * Iterate from higher cpus for boosted tasks.
5674 int i = boosted ? NR_CPUS-iter_cpu-1 : iter_cpu;
5676 if (!cpu_online(i) || !cpumask_test_cpu(i, tsk_cpus_allowed(p)))
5680 * p's blocked utilization is still accounted for on prev_cpu
5681 * so prev_cpu will receive a negative bias due to the double
5682 * accounting. However, the blocked utilization may be zero.
5684 new_util = cpu_util(i) + task_util_boosted;
5687 * Ensure minimum capacity to grant the required boost.
5688 * The target CPU can be already at a capacity level higher
5689 * than the one required to boost the task.
5691 if (new_util > capacity_orig_of(i))
5694 #ifdef CONFIG_SCHED_WALT
5695 if (walt_cpu_high_irqload(i))
5699 * Unconditionally favoring tasks that prefer idle cpus to
5702 if (idle_cpu(i) && prefer_idle) {
5703 if (best_idle_cpu < 0)
5708 cur_capacity = capacity_curr_of(i);
5710 idle_idx = idle_get_state_idx(rq);
5712 if (new_util < cur_capacity) {
5713 if (cpu_rq(i)->nr_running) {
5715 /* Find a target cpu with highest
5718 if (target_util == 0 ||
5719 target_util < new_util) {
5721 target_util = new_util;
5724 /* Find a target cpu with lowest
5727 if (target_util == 0 ||
5728 target_util > new_util) {
5730 target_util = new_util;
5733 } else if (!prefer_idle) {
5734 if (best_idle_cpu < 0 ||
5735 (sysctl_sched_cstate_aware &&
5736 best_idle_cstate > idle_idx)) {
5737 best_idle_cstate = idle_idx;
5741 } else if (backup_capacity == 0 ||
5742 backup_capacity > cur_capacity) {
5743 // Find a backup cpu with least capacity.
5744 backup_capacity = cur_capacity;
5749 if (prefer_idle && best_idle_cpu >= 0)
5750 target_cpu = best_idle_cpu;
5751 else if (target_cpu < 0)
5752 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5757 static int energy_aware_wake_cpu(struct task_struct *p, int target, int sync)
5759 struct sched_domain *sd;
5760 struct sched_group *sg, *sg_target;
5761 int target_max_cap = INT_MAX;
5762 int target_cpu = task_cpu(p);
5763 unsigned long task_util_boosted, new_util;
5766 if (sysctl_sched_sync_hint_enable && sync) {
5767 int cpu = smp_processor_id();
5768 cpumask_t search_cpus;
5769 cpumask_and(&search_cpus, tsk_cpus_allowed(p), cpu_online_mask);
5770 if (cpumask_test_cpu(cpu, &search_cpus))
5774 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5782 if (sysctl_sched_is_big_little) {
5785 * Find group with sufficient capacity. We only get here if no cpu is
5786 * overutilized. We may end up overutilizing a cpu by adding the task,
5787 * but that should not be any worse than select_idle_sibling().
5788 * load_balance() should sort it out later as we get above the tipping
5792 /* Assuming all cpus are the same in group */
5793 int max_cap_cpu = group_first_cpu(sg);
5796 * Assume smaller max capacity means more energy-efficient.
5797 * Ideally we should query the energy model for the right
5798 * answer but it easily ends up in an exhaustive search.
5800 if (capacity_of(max_cap_cpu) < target_max_cap &&
5801 task_fits_max(p, max_cap_cpu)) {
5803 target_max_cap = capacity_of(max_cap_cpu);
5805 } while (sg = sg->next, sg != sd->groups);
5807 task_util_boosted = boosted_task_util(p);
5808 /* Find cpu with sufficient capacity */
5809 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5811 * p's blocked utilization is still accounted for on prev_cpu
5812 * so prev_cpu will receive a negative bias due to the double
5813 * accounting. However, the blocked utilization may be zero.
5815 new_util = cpu_util(i) + task_util_boosted;
5818 * Ensure minimum capacity to grant the required boost.
5819 * The target CPU can be already at a capacity level higher
5820 * than the one required to boost the task.
5822 if (new_util > capacity_orig_of(i))
5825 if (new_util < capacity_curr_of(i)) {
5827 if (cpu_rq(i)->nr_running)
5831 /* cpu has capacity at higher OPP, keep it as fallback */
5832 if (target_cpu == task_cpu(p))
5837 * Find a cpu with sufficient capacity
5839 #ifdef CONFIG_CGROUP_SCHEDTUNE
5840 bool boosted = schedtune_task_boost(p) > 0;
5841 bool prefer_idle = schedtune_prefer_idle(p) > 0;
5844 bool prefer_idle = 0;
5846 int tmp_target = find_best_target(p, boosted, prefer_idle);
5847 if (tmp_target >= 0) {
5848 target_cpu = tmp_target;
5849 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
5854 if (target_cpu != task_cpu(p)) {
5855 struct energy_env eenv = {
5856 .util_delta = task_util(p),
5857 .src_cpu = task_cpu(p),
5858 .dst_cpu = target_cpu,
5862 /* Not enough spare capacity on previous cpu */
5863 if (cpu_overutilized(task_cpu(p)))
5866 if (energy_diff(&eenv) >= 0)
5874 * select_task_rq_fair: Select target runqueue for the waking task in domains
5875 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5876 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5878 * Balances load by selecting the idlest cpu in the idlest group, or under
5879 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5881 * Returns the target cpu number.
5883 * preempt must be disabled.
5886 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5888 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5889 int cpu = smp_processor_id();
5890 int new_cpu = prev_cpu;
5891 int want_affine = 0;
5892 int sync = wake_flags & WF_SYNC;
5894 if (sd_flag & SD_BALANCE_WAKE)
5895 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5896 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5900 for_each_domain(cpu, tmp) {
5901 if (!(tmp->flags & SD_LOAD_BALANCE))
5905 * If both cpu and prev_cpu are part of this domain,
5906 * cpu is a valid SD_WAKE_AFFINE target.
5908 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5909 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5914 if (tmp->flags & sd_flag)
5916 else if (!want_affine)
5921 sd = NULL; /* Prefer wake_affine over balance flags */
5922 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5927 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5928 new_cpu = energy_aware_wake_cpu(p, prev_cpu, sync);
5929 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5930 new_cpu = select_idle_sibling(p, new_cpu);
5933 struct sched_group *group;
5936 if (!(sd->flags & sd_flag)) {
5941 group = find_idlest_group(sd, p, cpu, sd_flag);
5947 new_cpu = find_idlest_cpu(group, p, cpu);
5948 if (new_cpu == -1 || new_cpu == cpu) {
5949 /* Now try balancing at a lower domain level of cpu */
5954 /* Now try balancing at a lower domain level of new_cpu */
5956 weight = sd->span_weight;
5958 for_each_domain(cpu, tmp) {
5959 if (weight <= tmp->span_weight)
5961 if (tmp->flags & sd_flag)
5964 /* while loop will break here if sd == NULL */
5972 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5973 * cfs_rq_of(p) references at time of call are still valid and identify the
5974 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5975 * other assumptions, including the state of rq->lock, should be made.
5977 static void migrate_task_rq_fair(struct task_struct *p)
5980 * We are supposed to update the task to "current" time, then its up to date
5981 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5982 * what current time is, so simply throw away the out-of-date time. This
5983 * will result in the wakee task is less decayed, but giving the wakee more
5984 * load sounds not bad.
5986 remove_entity_load_avg(&p->se);
5988 /* Tell new CPU we are migrated */
5989 p->se.avg.last_update_time = 0;
5991 /* We have migrated, no longer consider this task hot */
5992 p->se.exec_start = 0;
5995 static void task_dead_fair(struct task_struct *p)
5997 remove_entity_load_avg(&p->se);
6000 #define task_fits_max(p, cpu) true
6001 #endif /* CONFIG_SMP */
6003 static unsigned long
6004 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6006 unsigned long gran = sysctl_sched_wakeup_granularity;
6009 * Since its curr running now, convert the gran from real-time
6010 * to virtual-time in his units.
6012 * By using 'se' instead of 'curr' we penalize light tasks, so
6013 * they get preempted easier. That is, if 'se' < 'curr' then
6014 * the resulting gran will be larger, therefore penalizing the
6015 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6016 * be smaller, again penalizing the lighter task.
6018 * This is especially important for buddies when the leftmost
6019 * task is higher priority than the buddy.
6021 return calc_delta_fair(gran, se);
6025 * Should 'se' preempt 'curr'.
6039 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6041 s64 gran, vdiff = curr->vruntime - se->vruntime;
6046 gran = wakeup_gran(curr, se);
6053 static void set_last_buddy(struct sched_entity *se)
6055 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6058 for_each_sched_entity(se)
6059 cfs_rq_of(se)->last = se;
6062 static void set_next_buddy(struct sched_entity *se)
6064 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6067 for_each_sched_entity(se)
6068 cfs_rq_of(se)->next = se;
6071 static void set_skip_buddy(struct sched_entity *se)
6073 for_each_sched_entity(se)
6074 cfs_rq_of(se)->skip = se;
6078 * Preempt the current task with a newly woken task if needed:
6080 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6082 struct task_struct *curr = rq->curr;
6083 struct sched_entity *se = &curr->se, *pse = &p->se;
6084 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6085 int scale = cfs_rq->nr_running >= sched_nr_latency;
6086 int next_buddy_marked = 0;
6088 if (unlikely(se == pse))
6092 * This is possible from callers such as attach_tasks(), in which we
6093 * unconditionally check_prempt_curr() after an enqueue (which may have
6094 * lead to a throttle). This both saves work and prevents false
6095 * next-buddy nomination below.
6097 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6100 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6101 set_next_buddy(pse);
6102 next_buddy_marked = 1;
6106 * We can come here with TIF_NEED_RESCHED already set from new task
6109 * Note: this also catches the edge-case of curr being in a throttled
6110 * group (e.g. via set_curr_task), since update_curr() (in the
6111 * enqueue of curr) will have resulted in resched being set. This
6112 * prevents us from potentially nominating it as a false LAST_BUDDY
6115 if (test_tsk_need_resched(curr))
6118 /* Idle tasks are by definition preempted by non-idle tasks. */
6119 if (unlikely(curr->policy == SCHED_IDLE) &&
6120 likely(p->policy != SCHED_IDLE))
6124 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6125 * is driven by the tick):
6127 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6130 find_matching_se(&se, &pse);
6131 update_curr(cfs_rq_of(se));
6133 if (wakeup_preempt_entity(se, pse) == 1) {
6135 * Bias pick_next to pick the sched entity that is
6136 * triggering this preemption.
6138 if (!next_buddy_marked)
6139 set_next_buddy(pse);
6148 * Only set the backward buddy when the current task is still
6149 * on the rq. This can happen when a wakeup gets interleaved
6150 * with schedule on the ->pre_schedule() or idle_balance()
6151 * point, either of which can * drop the rq lock.
6153 * Also, during early boot the idle thread is in the fair class,
6154 * for obvious reasons its a bad idea to schedule back to it.
6156 if (unlikely(!se->on_rq || curr == rq->idle))
6159 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6163 static struct task_struct *
6164 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6166 struct cfs_rq *cfs_rq = &rq->cfs;
6167 struct sched_entity *se;
6168 struct task_struct *p;
6172 #ifdef CONFIG_FAIR_GROUP_SCHED
6173 if (!cfs_rq->nr_running)
6176 if (prev->sched_class != &fair_sched_class)
6180 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6181 * likely that a next task is from the same cgroup as the current.
6183 * Therefore attempt to avoid putting and setting the entire cgroup
6184 * hierarchy, only change the part that actually changes.
6188 struct sched_entity *curr = cfs_rq->curr;
6191 * Since we got here without doing put_prev_entity() we also
6192 * have to consider cfs_rq->curr. If it is still a runnable
6193 * entity, update_curr() will update its vruntime, otherwise
6194 * forget we've ever seen it.
6198 update_curr(cfs_rq);
6203 * This call to check_cfs_rq_runtime() will do the
6204 * throttle and dequeue its entity in the parent(s).
6205 * Therefore the 'simple' nr_running test will indeed
6208 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6212 se = pick_next_entity(cfs_rq, curr);
6213 cfs_rq = group_cfs_rq(se);
6219 * Since we haven't yet done put_prev_entity and if the selected task
6220 * is a different task than we started out with, try and touch the
6221 * least amount of cfs_rqs.
6224 struct sched_entity *pse = &prev->se;
6226 while (!(cfs_rq = is_same_group(se, pse))) {
6227 int se_depth = se->depth;
6228 int pse_depth = pse->depth;
6230 if (se_depth <= pse_depth) {
6231 put_prev_entity(cfs_rq_of(pse), pse);
6232 pse = parent_entity(pse);
6234 if (se_depth >= pse_depth) {
6235 set_next_entity(cfs_rq_of(se), se);
6236 se = parent_entity(se);
6240 put_prev_entity(cfs_rq, pse);
6241 set_next_entity(cfs_rq, se);
6244 if (hrtick_enabled(rq))
6245 hrtick_start_fair(rq, p);
6247 rq->misfit_task = !task_fits_max(p, rq->cpu);
6254 if (!cfs_rq->nr_running)
6257 put_prev_task(rq, prev);
6260 se = pick_next_entity(cfs_rq, NULL);
6261 set_next_entity(cfs_rq, se);
6262 cfs_rq = group_cfs_rq(se);
6267 if (hrtick_enabled(rq))
6268 hrtick_start_fair(rq, p);
6270 rq->misfit_task = !task_fits_max(p, rq->cpu);
6275 rq->misfit_task = 0;
6277 * This is OK, because current is on_cpu, which avoids it being picked
6278 * for load-balance and preemption/IRQs are still disabled avoiding
6279 * further scheduler activity on it and we're being very careful to
6280 * re-start the picking loop.
6282 lockdep_unpin_lock(&rq->lock);
6283 new_tasks = idle_balance(rq);
6284 lockdep_pin_lock(&rq->lock);
6286 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6287 * possible for any higher priority task to appear. In that case we
6288 * must re-start the pick_next_entity() loop.
6300 * Account for a descheduled task:
6302 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6304 struct sched_entity *se = &prev->se;
6305 struct cfs_rq *cfs_rq;
6307 for_each_sched_entity(se) {
6308 cfs_rq = cfs_rq_of(se);
6309 put_prev_entity(cfs_rq, se);
6314 * sched_yield() is very simple
6316 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6318 static void yield_task_fair(struct rq *rq)
6320 struct task_struct *curr = rq->curr;
6321 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6322 struct sched_entity *se = &curr->se;
6325 * Are we the only task in the tree?
6327 if (unlikely(rq->nr_running == 1))
6330 clear_buddies(cfs_rq, se);
6332 if (curr->policy != SCHED_BATCH) {
6333 update_rq_clock(rq);
6335 * Update run-time statistics of the 'current'.
6337 update_curr(cfs_rq);
6339 * Tell update_rq_clock() that we've just updated,
6340 * so we don't do microscopic update in schedule()
6341 * and double the fastpath cost.
6343 rq_clock_skip_update(rq, true);
6349 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6351 struct sched_entity *se = &p->se;
6353 /* throttled hierarchies are not runnable */
6354 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6357 /* Tell the scheduler that we'd really like pse to run next. */
6360 yield_task_fair(rq);
6366 /**************************************************
6367 * Fair scheduling class load-balancing methods.
6371 * The purpose of load-balancing is to achieve the same basic fairness the
6372 * per-cpu scheduler provides, namely provide a proportional amount of compute
6373 * time to each task. This is expressed in the following equation:
6375 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6377 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6378 * W_i,0 is defined as:
6380 * W_i,0 = \Sum_j w_i,j (2)
6382 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6383 * is derived from the nice value as per prio_to_weight[].
6385 * The weight average is an exponential decay average of the instantaneous
6388 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6390 * C_i is the compute capacity of cpu i, typically it is the
6391 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6392 * can also include other factors [XXX].
6394 * To achieve this balance we define a measure of imbalance which follows
6395 * directly from (1):
6397 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6399 * We them move tasks around to minimize the imbalance. In the continuous
6400 * function space it is obvious this converges, in the discrete case we get
6401 * a few fun cases generally called infeasible weight scenarios.
6404 * - infeasible weights;
6405 * - local vs global optima in the discrete case. ]
6410 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6411 * for all i,j solution, we create a tree of cpus that follows the hardware
6412 * topology where each level pairs two lower groups (or better). This results
6413 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6414 * tree to only the first of the previous level and we decrease the frequency
6415 * of load-balance at each level inv. proportional to the number of cpus in
6421 * \Sum { --- * --- * 2^i } = O(n) (5)
6423 * `- size of each group
6424 * | | `- number of cpus doing load-balance
6426 * `- sum over all levels
6428 * Coupled with a limit on how many tasks we can migrate every balance pass,
6429 * this makes (5) the runtime complexity of the balancer.
6431 * An important property here is that each CPU is still (indirectly) connected
6432 * to every other cpu in at most O(log n) steps:
6434 * The adjacency matrix of the resulting graph is given by:
6437 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6440 * And you'll find that:
6442 * A^(log_2 n)_i,j != 0 for all i,j (7)
6444 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6445 * The task movement gives a factor of O(m), giving a convergence complexity
6448 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6453 * In order to avoid CPUs going idle while there's still work to do, new idle
6454 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6455 * tree itself instead of relying on other CPUs to bring it work.
6457 * This adds some complexity to both (5) and (8) but it reduces the total idle
6465 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6468 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6473 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6475 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6477 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6480 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6481 * rewrite all of this once again.]
6484 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6486 enum fbq_type { regular, remote, all };
6495 #define LBF_ALL_PINNED 0x01
6496 #define LBF_NEED_BREAK 0x02
6497 #define LBF_DST_PINNED 0x04
6498 #define LBF_SOME_PINNED 0x08
6501 struct sched_domain *sd;
6509 struct cpumask *dst_grpmask;
6511 enum cpu_idle_type idle;
6513 unsigned int src_grp_nr_running;
6514 /* The set of CPUs under consideration for load-balancing */
6515 struct cpumask *cpus;
6520 unsigned int loop_break;
6521 unsigned int loop_max;
6523 enum fbq_type fbq_type;
6524 enum group_type busiest_group_type;
6525 struct list_head tasks;
6529 * Is this task likely cache-hot:
6531 static int task_hot(struct task_struct *p, struct lb_env *env)
6535 lockdep_assert_held(&env->src_rq->lock);
6537 if (p->sched_class != &fair_sched_class)
6540 if (unlikely(p->policy == SCHED_IDLE))
6544 * Buddy candidates are cache hot:
6546 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6547 (&p->se == cfs_rq_of(&p->se)->next ||
6548 &p->se == cfs_rq_of(&p->se)->last))
6551 if (sysctl_sched_migration_cost == -1)
6553 if (sysctl_sched_migration_cost == 0)
6556 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6558 return delta < (s64)sysctl_sched_migration_cost;
6561 #ifdef CONFIG_NUMA_BALANCING
6563 * Returns 1, if task migration degrades locality
6564 * Returns 0, if task migration improves locality i.e migration preferred.
6565 * Returns -1, if task migration is not affected by locality.
6567 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6569 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6570 unsigned long src_faults, dst_faults;
6571 int src_nid, dst_nid;
6573 if (!static_branch_likely(&sched_numa_balancing))
6576 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6579 src_nid = cpu_to_node(env->src_cpu);
6580 dst_nid = cpu_to_node(env->dst_cpu);
6582 if (src_nid == dst_nid)
6585 /* Migrating away from the preferred node is always bad. */
6586 if (src_nid == p->numa_preferred_nid) {
6587 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6593 /* Encourage migration to the preferred node. */
6594 if (dst_nid == p->numa_preferred_nid)
6598 src_faults = group_faults(p, src_nid);
6599 dst_faults = group_faults(p, dst_nid);
6601 src_faults = task_faults(p, src_nid);
6602 dst_faults = task_faults(p, dst_nid);
6605 return dst_faults < src_faults;
6609 static inline int migrate_degrades_locality(struct task_struct *p,
6617 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6620 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6624 lockdep_assert_held(&env->src_rq->lock);
6627 * We do not migrate tasks that are:
6628 * 1) throttled_lb_pair, or
6629 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6630 * 3) running (obviously), or
6631 * 4) are cache-hot on their current CPU.
6633 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6636 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6639 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6641 env->flags |= LBF_SOME_PINNED;
6644 * Remember if this task can be migrated to any other cpu in
6645 * our sched_group. We may want to revisit it if we couldn't
6646 * meet load balance goals by pulling other tasks on src_cpu.
6648 * Also avoid computing new_dst_cpu if we have already computed
6649 * one in current iteration.
6651 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6654 /* Prevent to re-select dst_cpu via env's cpus */
6655 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6656 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6657 env->flags |= LBF_DST_PINNED;
6658 env->new_dst_cpu = cpu;
6666 /* Record that we found atleast one task that could run on dst_cpu */
6667 env->flags &= ~LBF_ALL_PINNED;
6669 if (task_running(env->src_rq, p)) {
6670 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6675 * Aggressive migration if:
6676 * 1) destination numa is preferred
6677 * 2) task is cache cold, or
6678 * 3) too many balance attempts have failed.
6680 tsk_cache_hot = migrate_degrades_locality(p, env);
6681 if (tsk_cache_hot == -1)
6682 tsk_cache_hot = task_hot(p, env);
6684 if (tsk_cache_hot <= 0 ||
6685 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6686 if (tsk_cache_hot == 1) {
6687 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6688 schedstat_inc(p, se.statistics.nr_forced_migrations);
6693 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6698 * detach_task() -- detach the task for the migration specified in env
6700 static void detach_task(struct task_struct *p, struct lb_env *env)
6702 lockdep_assert_held(&env->src_rq->lock);
6704 deactivate_task(env->src_rq, p, 0);
6705 p->on_rq = TASK_ON_RQ_MIGRATING;
6706 double_lock_balance(env->src_rq, env->dst_rq);
6707 set_task_cpu(p, env->dst_cpu);
6708 double_unlock_balance(env->src_rq, env->dst_rq);
6712 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6713 * part of active balancing operations within "domain".
6715 * Returns a task if successful and NULL otherwise.
6717 static struct task_struct *detach_one_task(struct lb_env *env)
6719 struct task_struct *p, *n;
6721 lockdep_assert_held(&env->src_rq->lock);
6723 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6724 if (!can_migrate_task(p, env))
6727 detach_task(p, env);
6730 * Right now, this is only the second place where
6731 * lb_gained[env->idle] is updated (other is detach_tasks)
6732 * so we can safely collect stats here rather than
6733 * inside detach_tasks().
6735 schedstat_inc(env->sd, lb_gained[env->idle]);
6741 static const unsigned int sched_nr_migrate_break = 32;
6744 * detach_tasks() -- tries to detach up to imbalance weighted load from
6745 * busiest_rq, as part of a balancing operation within domain "sd".
6747 * Returns number of detached tasks if successful and 0 otherwise.
6749 static int detach_tasks(struct lb_env *env)
6751 struct list_head *tasks = &env->src_rq->cfs_tasks;
6752 struct task_struct *p;
6756 lockdep_assert_held(&env->src_rq->lock);
6758 if (env->imbalance <= 0)
6761 while (!list_empty(tasks)) {
6763 * We don't want to steal all, otherwise we may be treated likewise,
6764 * which could at worst lead to a livelock crash.
6766 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6769 p = list_first_entry(tasks, struct task_struct, se.group_node);
6772 /* We've more or less seen every task there is, call it quits */
6773 if (env->loop > env->loop_max)
6776 /* take a breather every nr_migrate tasks */
6777 if (env->loop > env->loop_break) {
6778 env->loop_break += sched_nr_migrate_break;
6779 env->flags |= LBF_NEED_BREAK;
6783 if (!can_migrate_task(p, env))
6786 load = task_h_load(p);
6788 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6791 if ((load / 2) > env->imbalance)
6794 detach_task(p, env);
6795 list_add(&p->se.group_node, &env->tasks);
6798 env->imbalance -= load;
6800 #ifdef CONFIG_PREEMPT
6802 * NEWIDLE balancing is a source of latency, so preemptible
6803 * kernels will stop after the first task is detached to minimize
6804 * the critical section.
6806 if (env->idle == CPU_NEWLY_IDLE)
6811 * We only want to steal up to the prescribed amount of
6814 if (env->imbalance <= 0)
6819 list_move_tail(&p->se.group_node, tasks);
6823 * Right now, this is one of only two places we collect this stat
6824 * so we can safely collect detach_one_task() stats here rather
6825 * than inside detach_one_task().
6827 schedstat_add(env->sd, lb_gained[env->idle], detached);
6833 * attach_task() -- attach the task detached by detach_task() to its new rq.
6835 static void attach_task(struct rq *rq, struct task_struct *p)
6837 lockdep_assert_held(&rq->lock);
6839 BUG_ON(task_rq(p) != rq);
6840 p->on_rq = TASK_ON_RQ_QUEUED;
6841 activate_task(rq, p, 0);
6842 check_preempt_curr(rq, p, 0);
6846 * attach_one_task() -- attaches the task returned from detach_one_task() to
6849 static void attach_one_task(struct rq *rq, struct task_struct *p)
6851 raw_spin_lock(&rq->lock);
6854 * We want to potentially raise target_cpu's OPP.
6856 update_capacity_of(cpu_of(rq));
6857 raw_spin_unlock(&rq->lock);
6861 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6864 static void attach_tasks(struct lb_env *env)
6866 struct list_head *tasks = &env->tasks;
6867 struct task_struct *p;
6869 raw_spin_lock(&env->dst_rq->lock);
6871 while (!list_empty(tasks)) {
6872 p = list_first_entry(tasks, struct task_struct, se.group_node);
6873 list_del_init(&p->se.group_node);
6875 attach_task(env->dst_rq, p);
6879 * We want to potentially raise env.dst_cpu's OPP.
6881 update_capacity_of(env->dst_cpu);
6883 raw_spin_unlock(&env->dst_rq->lock);
6886 #ifdef CONFIG_FAIR_GROUP_SCHED
6887 static void update_blocked_averages(int cpu)
6889 struct rq *rq = cpu_rq(cpu);
6890 struct cfs_rq *cfs_rq;
6891 unsigned long flags;
6893 raw_spin_lock_irqsave(&rq->lock, flags);
6894 update_rq_clock(rq);
6897 * Iterates the task_group tree in a bottom up fashion, see
6898 * list_add_leaf_cfs_rq() for details.
6900 for_each_leaf_cfs_rq(rq, cfs_rq) {
6901 /* throttled entities do not contribute to load */
6902 if (throttled_hierarchy(cfs_rq))
6905 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6906 update_tg_load_avg(cfs_rq, 0);
6908 raw_spin_unlock_irqrestore(&rq->lock, flags);
6912 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6913 * This needs to be done in a top-down fashion because the load of a child
6914 * group is a fraction of its parents load.
6916 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6918 struct rq *rq = rq_of(cfs_rq);
6919 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6920 unsigned long now = jiffies;
6923 if (cfs_rq->last_h_load_update == now)
6926 cfs_rq->h_load_next = NULL;
6927 for_each_sched_entity(se) {
6928 cfs_rq = cfs_rq_of(se);
6929 cfs_rq->h_load_next = se;
6930 if (cfs_rq->last_h_load_update == now)
6935 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6936 cfs_rq->last_h_load_update = now;
6939 while ((se = cfs_rq->h_load_next) != NULL) {
6940 load = cfs_rq->h_load;
6941 load = div64_ul(load * se->avg.load_avg,
6942 cfs_rq_load_avg(cfs_rq) + 1);
6943 cfs_rq = group_cfs_rq(se);
6944 cfs_rq->h_load = load;
6945 cfs_rq->last_h_load_update = now;
6949 static unsigned long task_h_load(struct task_struct *p)
6951 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6953 update_cfs_rq_h_load(cfs_rq);
6954 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6955 cfs_rq_load_avg(cfs_rq) + 1);
6958 static inline void update_blocked_averages(int cpu)
6960 struct rq *rq = cpu_rq(cpu);
6961 struct cfs_rq *cfs_rq = &rq->cfs;
6962 unsigned long flags;
6964 raw_spin_lock_irqsave(&rq->lock, flags);
6965 update_rq_clock(rq);
6966 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6967 raw_spin_unlock_irqrestore(&rq->lock, flags);
6970 static unsigned long task_h_load(struct task_struct *p)
6972 return p->se.avg.load_avg;
6976 /********** Helpers for find_busiest_group ************************/
6979 * sg_lb_stats - stats of a sched_group required for load_balancing
6981 struct sg_lb_stats {
6982 unsigned long avg_load; /*Avg load across the CPUs of the group */
6983 unsigned long group_load; /* Total load over the CPUs of the group */
6984 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6985 unsigned long load_per_task;
6986 unsigned long group_capacity;
6987 unsigned long group_util; /* Total utilization of the group */
6988 unsigned int sum_nr_running; /* Nr tasks running in the group */
6989 unsigned int idle_cpus;
6990 unsigned int group_weight;
6991 enum group_type group_type;
6992 int group_no_capacity;
6993 int group_misfit_task; /* A cpu has a task too big for its capacity */
6994 #ifdef CONFIG_NUMA_BALANCING
6995 unsigned int nr_numa_running;
6996 unsigned int nr_preferred_running;
7001 * sd_lb_stats - Structure to store the statistics of a sched_domain
7002 * during load balancing.
7004 struct sd_lb_stats {
7005 struct sched_group *busiest; /* Busiest group in this sd */
7006 struct sched_group *local; /* Local group in this sd */
7007 unsigned long total_load; /* Total load of all groups in sd */
7008 unsigned long total_capacity; /* Total capacity of all groups in sd */
7009 unsigned long avg_load; /* Average load across all groups in sd */
7011 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7012 struct sg_lb_stats local_stat; /* Statistics of the local group */
7015 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7018 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7019 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7020 * We must however clear busiest_stat::avg_load because
7021 * update_sd_pick_busiest() reads this before assignment.
7023 *sds = (struct sd_lb_stats){
7027 .total_capacity = 0UL,
7030 .sum_nr_running = 0,
7031 .group_type = group_other,
7037 * get_sd_load_idx - Obtain the load index for a given sched domain.
7038 * @sd: The sched_domain whose load_idx is to be obtained.
7039 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7041 * Return: The load index.
7043 static inline int get_sd_load_idx(struct sched_domain *sd,
7044 enum cpu_idle_type idle)
7050 load_idx = sd->busy_idx;
7053 case CPU_NEWLY_IDLE:
7054 load_idx = sd->newidle_idx;
7057 load_idx = sd->idle_idx;
7064 static unsigned long scale_rt_capacity(int cpu)
7066 struct rq *rq = cpu_rq(cpu);
7067 u64 total, used, age_stamp, avg;
7071 * Since we're reading these variables without serialization make sure
7072 * we read them once before doing sanity checks on them.
7074 age_stamp = READ_ONCE(rq->age_stamp);
7075 avg = READ_ONCE(rq->rt_avg);
7076 delta = __rq_clock_broken(rq) - age_stamp;
7078 if (unlikely(delta < 0))
7081 total = sched_avg_period() + delta;
7083 used = div_u64(avg, total);
7086 * deadline bandwidth is defined at system level so we must
7087 * weight this bandwidth with the max capacity of the system.
7088 * As a reminder, avg_bw is 20bits width and
7089 * scale_cpu_capacity is 10 bits width
7091 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7093 if (likely(used < SCHED_CAPACITY_SCALE))
7094 return SCHED_CAPACITY_SCALE - used;
7099 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7101 raw_spin_lock_init(&mcc->lock);
7106 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7108 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7109 struct sched_group *sdg = sd->groups;
7110 struct max_cpu_capacity *mcc;
7111 unsigned long max_capacity;
7113 unsigned long flags;
7115 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7117 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7119 raw_spin_lock_irqsave(&mcc->lock, flags);
7120 max_capacity = mcc->val;
7121 max_cap_cpu = mcc->cpu;
7123 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7124 (max_capacity < capacity)) {
7125 mcc->val = capacity;
7127 #ifdef CONFIG_SCHED_DEBUG
7128 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7129 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7134 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7136 skip_unlock: __attribute__ ((unused));
7137 capacity *= scale_rt_capacity(cpu);
7138 capacity >>= SCHED_CAPACITY_SHIFT;
7143 cpu_rq(cpu)->cpu_capacity = capacity;
7144 sdg->sgc->capacity = capacity;
7145 sdg->sgc->max_capacity = capacity;
7148 void update_group_capacity(struct sched_domain *sd, int cpu)
7150 struct sched_domain *child = sd->child;
7151 struct sched_group *group, *sdg = sd->groups;
7152 unsigned long capacity, max_capacity;
7153 unsigned long interval;
7155 interval = msecs_to_jiffies(sd->balance_interval);
7156 interval = clamp(interval, 1UL, max_load_balance_interval);
7157 sdg->sgc->next_update = jiffies + interval;
7160 update_cpu_capacity(sd, cpu);
7167 if (child->flags & SD_OVERLAP) {
7169 * SD_OVERLAP domains cannot assume that child groups
7170 * span the current group.
7173 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7174 struct sched_group_capacity *sgc;
7175 struct rq *rq = cpu_rq(cpu);
7178 * build_sched_domains() -> init_sched_groups_capacity()
7179 * gets here before we've attached the domains to the
7182 * Use capacity_of(), which is set irrespective of domains
7183 * in update_cpu_capacity().
7185 * This avoids capacity from being 0 and
7186 * causing divide-by-zero issues on boot.
7188 if (unlikely(!rq->sd)) {
7189 capacity += capacity_of(cpu);
7191 sgc = rq->sd->groups->sgc;
7192 capacity += sgc->capacity;
7195 max_capacity = max(capacity, max_capacity);
7199 * !SD_OVERLAP domains can assume that child groups
7200 * span the current group.
7203 group = child->groups;
7205 struct sched_group_capacity *sgc = group->sgc;
7207 capacity += sgc->capacity;
7208 max_capacity = max(sgc->max_capacity, max_capacity);
7209 group = group->next;
7210 } while (group != child->groups);
7213 sdg->sgc->capacity = capacity;
7214 sdg->sgc->max_capacity = max_capacity;
7218 * Check whether the capacity of the rq has been noticeably reduced by side
7219 * activity. The imbalance_pct is used for the threshold.
7220 * Return true is the capacity is reduced
7223 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7225 return ((rq->cpu_capacity * sd->imbalance_pct) <
7226 (rq->cpu_capacity_orig * 100));
7230 * Group imbalance indicates (and tries to solve) the problem where balancing
7231 * groups is inadequate due to tsk_cpus_allowed() constraints.
7233 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7234 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7237 * { 0 1 2 3 } { 4 5 6 7 }
7240 * If we were to balance group-wise we'd place two tasks in the first group and
7241 * two tasks in the second group. Clearly this is undesired as it will overload
7242 * cpu 3 and leave one of the cpus in the second group unused.
7244 * The current solution to this issue is detecting the skew in the first group
7245 * by noticing the lower domain failed to reach balance and had difficulty
7246 * moving tasks due to affinity constraints.
7248 * When this is so detected; this group becomes a candidate for busiest; see
7249 * update_sd_pick_busiest(). And calculate_imbalance() and
7250 * find_busiest_group() avoid some of the usual balance conditions to allow it
7251 * to create an effective group imbalance.
7253 * This is a somewhat tricky proposition since the next run might not find the
7254 * group imbalance and decide the groups need to be balanced again. A most
7255 * subtle and fragile situation.
7258 static inline int sg_imbalanced(struct sched_group *group)
7260 return group->sgc->imbalance;
7264 * group_has_capacity returns true if the group has spare capacity that could
7265 * be used by some tasks.
7266 * We consider that a group has spare capacity if the * number of task is
7267 * smaller than the number of CPUs or if the utilization is lower than the
7268 * available capacity for CFS tasks.
7269 * For the latter, we use a threshold to stabilize the state, to take into
7270 * account the variance of the tasks' load and to return true if the available
7271 * capacity in meaningful for the load balancer.
7272 * As an example, an available capacity of 1% can appear but it doesn't make
7273 * any benefit for the load balance.
7276 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7278 if (sgs->sum_nr_running < sgs->group_weight)
7281 if ((sgs->group_capacity * 100) >
7282 (sgs->group_util * env->sd->imbalance_pct))
7289 * group_is_overloaded returns true if the group has more tasks than it can
7291 * group_is_overloaded is not equals to !group_has_capacity because a group
7292 * with the exact right number of tasks, has no more spare capacity but is not
7293 * overloaded so both group_has_capacity and group_is_overloaded return
7297 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7299 if (sgs->sum_nr_running <= sgs->group_weight)
7302 if ((sgs->group_capacity * 100) <
7303 (sgs->group_util * env->sd->imbalance_pct))
7311 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7312 * per-cpu capacity than sched_group ref.
7315 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7317 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7318 ref->sgc->max_capacity;
7322 group_type group_classify(struct sched_group *group,
7323 struct sg_lb_stats *sgs)
7325 if (sgs->group_no_capacity)
7326 return group_overloaded;
7328 if (sg_imbalanced(group))
7329 return group_imbalanced;
7331 if (sgs->group_misfit_task)
7332 return group_misfit_task;
7338 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7339 * @env: The load balancing environment.
7340 * @group: sched_group whose statistics are to be updated.
7341 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7342 * @local_group: Does group contain this_cpu.
7343 * @sgs: variable to hold the statistics for this group.
7344 * @overload: Indicate more than one runnable task for any CPU.
7345 * @overutilized: Indicate overutilization for any CPU.
7347 static inline void update_sg_lb_stats(struct lb_env *env,
7348 struct sched_group *group, int load_idx,
7349 int local_group, struct sg_lb_stats *sgs,
7350 bool *overload, bool *overutilized)
7355 memset(sgs, 0, sizeof(*sgs));
7357 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7358 struct rq *rq = cpu_rq(i);
7360 /* Bias balancing toward cpus of our domain */
7362 load = target_load(i, load_idx);
7364 load = source_load(i, load_idx);
7366 sgs->group_load += load;
7367 sgs->group_util += cpu_util(i);
7368 sgs->sum_nr_running += rq->cfs.h_nr_running;
7370 nr_running = rq->nr_running;
7374 #ifdef CONFIG_NUMA_BALANCING
7375 sgs->nr_numa_running += rq->nr_numa_running;
7376 sgs->nr_preferred_running += rq->nr_preferred_running;
7378 sgs->sum_weighted_load += weighted_cpuload(i);
7380 * No need to call idle_cpu() if nr_running is not 0
7382 if (!nr_running && idle_cpu(i))
7385 if (cpu_overutilized(i)) {
7386 *overutilized = true;
7387 if (!sgs->group_misfit_task && rq->misfit_task)
7388 sgs->group_misfit_task = capacity_of(i);
7392 /* Adjust by relative CPU capacity of the group */
7393 sgs->group_capacity = group->sgc->capacity;
7394 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7396 if (sgs->sum_nr_running)
7397 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7399 sgs->group_weight = group->group_weight;
7401 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7402 sgs->group_type = group_classify(group, sgs);
7406 * update_sd_pick_busiest - return 1 on busiest group
7407 * @env: The load balancing environment.
7408 * @sds: sched_domain statistics
7409 * @sg: sched_group candidate to be checked for being the busiest
7410 * @sgs: sched_group statistics
7412 * Determine if @sg is a busier group than the previously selected
7415 * Return: %true if @sg is a busier group than the previously selected
7416 * busiest group. %false otherwise.
7418 static bool update_sd_pick_busiest(struct lb_env *env,
7419 struct sd_lb_stats *sds,
7420 struct sched_group *sg,
7421 struct sg_lb_stats *sgs)
7423 struct sg_lb_stats *busiest = &sds->busiest_stat;
7425 if (sgs->group_type > busiest->group_type)
7428 if (sgs->group_type < busiest->group_type)
7432 * Candidate sg doesn't face any serious load-balance problems
7433 * so don't pick it if the local sg is already filled up.
7435 if (sgs->group_type == group_other &&
7436 !group_has_capacity(env, &sds->local_stat))
7439 if (sgs->avg_load <= busiest->avg_load)
7443 * Candiate sg has no more than one task per cpu and has higher
7444 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7446 if (sgs->sum_nr_running <= sgs->group_weight &&
7447 group_smaller_cpu_capacity(sds->local, sg))
7450 /* This is the busiest node in its class. */
7451 if (!(env->sd->flags & SD_ASYM_PACKING))
7455 * ASYM_PACKING needs to move all the work to the lowest
7456 * numbered CPUs in the group, therefore mark all groups
7457 * higher than ourself as busy.
7459 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7463 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7470 #ifdef CONFIG_NUMA_BALANCING
7471 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7473 if (sgs->sum_nr_running > sgs->nr_numa_running)
7475 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7480 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7482 if (rq->nr_running > rq->nr_numa_running)
7484 if (rq->nr_running > rq->nr_preferred_running)
7489 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7494 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7498 #endif /* CONFIG_NUMA_BALANCING */
7501 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7502 * @env: The load balancing environment.
7503 * @sds: variable to hold the statistics for this sched_domain.
7505 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7507 struct sched_domain *child = env->sd->child;
7508 struct sched_group *sg = env->sd->groups;
7509 struct sg_lb_stats tmp_sgs;
7510 int load_idx, prefer_sibling = 0;
7511 bool overload = false, overutilized = false;
7513 if (child && child->flags & SD_PREFER_SIBLING)
7516 load_idx = get_sd_load_idx(env->sd, env->idle);
7519 struct sg_lb_stats *sgs = &tmp_sgs;
7522 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7525 sgs = &sds->local_stat;
7527 if (env->idle != CPU_NEWLY_IDLE ||
7528 time_after_eq(jiffies, sg->sgc->next_update))
7529 update_group_capacity(env->sd, env->dst_cpu);
7532 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7533 &overload, &overutilized);
7539 * In case the child domain prefers tasks go to siblings
7540 * first, lower the sg capacity so that we'll try
7541 * and move all the excess tasks away. We lower the capacity
7542 * of a group only if the local group has the capacity to fit
7543 * these excess tasks. The extra check prevents the case where
7544 * you always pull from the heaviest group when it is already
7545 * under-utilized (possible with a large weight task outweighs
7546 * the tasks on the system).
7548 if (prefer_sibling && sds->local &&
7549 group_has_capacity(env, &sds->local_stat) &&
7550 (sgs->sum_nr_running > 1)) {
7551 sgs->group_no_capacity = 1;
7552 sgs->group_type = group_classify(sg, sgs);
7556 * Ignore task groups with misfit tasks if local group has no
7557 * capacity or if per-cpu capacity isn't higher.
7559 if (sgs->group_type == group_misfit_task &&
7560 (!group_has_capacity(env, &sds->local_stat) ||
7561 !group_smaller_cpu_capacity(sg, sds->local)))
7562 sgs->group_type = group_other;
7564 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7566 sds->busiest_stat = *sgs;
7570 /* Now, start updating sd_lb_stats */
7571 sds->total_load += sgs->group_load;
7572 sds->total_capacity += sgs->group_capacity;
7575 } while (sg != env->sd->groups);
7577 if (env->sd->flags & SD_NUMA)
7578 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7580 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7582 if (!env->sd->parent) {
7583 /* update overload indicator if we are at root domain */
7584 if (env->dst_rq->rd->overload != overload)
7585 env->dst_rq->rd->overload = overload;
7587 /* Update over-utilization (tipping point, U >= 0) indicator */
7588 if (env->dst_rq->rd->overutilized != overutilized) {
7589 env->dst_rq->rd->overutilized = overutilized;
7590 trace_sched_overutilized(overutilized);
7593 if (!env->dst_rq->rd->overutilized && overutilized) {
7594 env->dst_rq->rd->overutilized = true;
7595 trace_sched_overutilized(true);
7602 * check_asym_packing - Check to see if the group is packed into the
7605 * This is primarily intended to used at the sibling level. Some
7606 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7607 * case of POWER7, it can move to lower SMT modes only when higher
7608 * threads are idle. When in lower SMT modes, the threads will
7609 * perform better since they share less core resources. Hence when we
7610 * have idle threads, we want them to be the higher ones.
7612 * This packing function is run on idle threads. It checks to see if
7613 * the busiest CPU in this domain (core in the P7 case) has a higher
7614 * CPU number than the packing function is being run on. Here we are
7615 * assuming lower CPU number will be equivalent to lower a SMT thread
7618 * Return: 1 when packing is required and a task should be moved to
7619 * this CPU. The amount of the imbalance is returned in *imbalance.
7621 * @env: The load balancing environment.
7622 * @sds: Statistics of the sched_domain which is to be packed
7624 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7628 if (!(env->sd->flags & SD_ASYM_PACKING))
7634 busiest_cpu = group_first_cpu(sds->busiest);
7635 if (env->dst_cpu > busiest_cpu)
7638 env->imbalance = DIV_ROUND_CLOSEST(
7639 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7640 SCHED_CAPACITY_SCALE);
7646 * fix_small_imbalance - Calculate the minor imbalance that exists
7647 * amongst the groups of a sched_domain, during
7649 * @env: The load balancing environment.
7650 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7653 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7655 unsigned long tmp, capa_now = 0, capa_move = 0;
7656 unsigned int imbn = 2;
7657 unsigned long scaled_busy_load_per_task;
7658 struct sg_lb_stats *local, *busiest;
7660 local = &sds->local_stat;
7661 busiest = &sds->busiest_stat;
7663 if (!local->sum_nr_running)
7664 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7665 else if (busiest->load_per_task > local->load_per_task)
7668 scaled_busy_load_per_task =
7669 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7670 busiest->group_capacity;
7672 if (busiest->avg_load + scaled_busy_load_per_task >=
7673 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7674 env->imbalance = busiest->load_per_task;
7679 * OK, we don't have enough imbalance to justify moving tasks,
7680 * however we may be able to increase total CPU capacity used by
7684 capa_now += busiest->group_capacity *
7685 min(busiest->load_per_task, busiest->avg_load);
7686 capa_now += local->group_capacity *
7687 min(local->load_per_task, local->avg_load);
7688 capa_now /= SCHED_CAPACITY_SCALE;
7690 /* Amount of load we'd subtract */
7691 if (busiest->avg_load > scaled_busy_load_per_task) {
7692 capa_move += busiest->group_capacity *
7693 min(busiest->load_per_task,
7694 busiest->avg_load - scaled_busy_load_per_task);
7697 /* Amount of load we'd add */
7698 if (busiest->avg_load * busiest->group_capacity <
7699 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7700 tmp = (busiest->avg_load * busiest->group_capacity) /
7701 local->group_capacity;
7703 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7704 local->group_capacity;
7706 capa_move += local->group_capacity *
7707 min(local->load_per_task, local->avg_load + tmp);
7708 capa_move /= SCHED_CAPACITY_SCALE;
7710 /* Move if we gain throughput */
7711 if (capa_move > capa_now)
7712 env->imbalance = busiest->load_per_task;
7716 * calculate_imbalance - Calculate the amount of imbalance present within the
7717 * groups of a given sched_domain during load balance.
7718 * @env: load balance environment
7719 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7721 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7723 unsigned long max_pull, load_above_capacity = ~0UL;
7724 struct sg_lb_stats *local, *busiest;
7726 local = &sds->local_stat;
7727 busiest = &sds->busiest_stat;
7729 if (busiest->group_type == group_imbalanced) {
7731 * In the group_imb case we cannot rely on group-wide averages
7732 * to ensure cpu-load equilibrium, look at wider averages. XXX
7734 busiest->load_per_task =
7735 min(busiest->load_per_task, sds->avg_load);
7739 * In the presence of smp nice balancing, certain scenarios can have
7740 * max load less than avg load(as we skip the groups at or below
7741 * its cpu_capacity, while calculating max_load..)
7743 if (busiest->avg_load <= sds->avg_load ||
7744 local->avg_load >= sds->avg_load) {
7745 /* Misfitting tasks should be migrated in any case */
7746 if (busiest->group_type == group_misfit_task) {
7747 env->imbalance = busiest->group_misfit_task;
7752 * Busiest group is overloaded, local is not, use the spare
7753 * cycles to maximize throughput
7755 if (busiest->group_type == group_overloaded &&
7756 local->group_type <= group_misfit_task) {
7757 env->imbalance = busiest->load_per_task;
7762 return fix_small_imbalance(env, sds);
7766 * If there aren't any idle cpus, avoid creating some.
7768 if (busiest->group_type == group_overloaded &&
7769 local->group_type == group_overloaded) {
7770 load_above_capacity = busiest->sum_nr_running *
7772 if (load_above_capacity > busiest->group_capacity)
7773 load_above_capacity -= busiest->group_capacity;
7775 load_above_capacity = ~0UL;
7779 * We're trying to get all the cpus to the average_load, so we don't
7780 * want to push ourselves above the average load, nor do we wish to
7781 * reduce the max loaded cpu below the average load. At the same time,
7782 * we also don't want to reduce the group load below the group capacity
7783 * (so that we can implement power-savings policies etc). Thus we look
7784 * for the minimum possible imbalance.
7786 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7788 /* How much load to actually move to equalise the imbalance */
7789 env->imbalance = min(
7790 max_pull * busiest->group_capacity,
7791 (sds->avg_load - local->avg_load) * local->group_capacity
7792 ) / SCHED_CAPACITY_SCALE;
7794 /* Boost imbalance to allow misfit task to be balanced. */
7795 if (busiest->group_type == group_misfit_task)
7796 env->imbalance = max_t(long, env->imbalance,
7797 busiest->group_misfit_task);
7800 * if *imbalance is less than the average load per runnable task
7801 * there is no guarantee that any tasks will be moved so we'll have
7802 * a think about bumping its value to force at least one task to be
7805 if (env->imbalance < busiest->load_per_task)
7806 return fix_small_imbalance(env, sds);
7809 /******* find_busiest_group() helpers end here *********************/
7812 * find_busiest_group - Returns the busiest group within the sched_domain
7813 * if there is an imbalance. If there isn't an imbalance, and
7814 * the user has opted for power-savings, it returns a group whose
7815 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7816 * such a group exists.
7818 * Also calculates the amount of weighted load which should be moved
7819 * to restore balance.
7821 * @env: The load balancing environment.
7823 * Return: - The busiest group if imbalance exists.
7824 * - If no imbalance and user has opted for power-savings balance,
7825 * return the least loaded group whose CPUs can be
7826 * put to idle by rebalancing its tasks onto our group.
7828 static struct sched_group *find_busiest_group(struct lb_env *env)
7830 struct sg_lb_stats *local, *busiest;
7831 struct sd_lb_stats sds;
7833 init_sd_lb_stats(&sds);
7836 * Compute the various statistics relavent for load balancing at
7839 update_sd_lb_stats(env, &sds);
7841 if (energy_aware() && !env->dst_rq->rd->overutilized)
7844 local = &sds.local_stat;
7845 busiest = &sds.busiest_stat;
7847 /* ASYM feature bypasses nice load balance check */
7848 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7849 check_asym_packing(env, &sds))
7852 /* There is no busy sibling group to pull tasks from */
7853 if (!sds.busiest || busiest->sum_nr_running == 0)
7856 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7857 / sds.total_capacity;
7860 * If the busiest group is imbalanced the below checks don't
7861 * work because they assume all things are equal, which typically
7862 * isn't true due to cpus_allowed constraints and the like.
7864 if (busiest->group_type == group_imbalanced)
7867 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7868 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7869 busiest->group_no_capacity)
7872 /* Misfitting tasks should be dealt with regardless of the avg load */
7873 if (busiest->group_type == group_misfit_task) {
7878 * If the local group is busier than the selected busiest group
7879 * don't try and pull any tasks.
7881 if (local->avg_load >= busiest->avg_load)
7885 * Don't pull any tasks if this group is already above the domain
7888 if (local->avg_load >= sds.avg_load)
7891 if (env->idle == CPU_IDLE) {
7893 * This cpu is idle. If the busiest group is not overloaded
7894 * and there is no imbalance between this and busiest group
7895 * wrt idle cpus, it is balanced. The imbalance becomes
7896 * significant if the diff is greater than 1 otherwise we
7897 * might end up to just move the imbalance on another group
7899 if ((busiest->group_type != group_overloaded) &&
7900 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7901 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7905 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7906 * imbalance_pct to be conservative.
7908 if (100 * busiest->avg_load <=
7909 env->sd->imbalance_pct * local->avg_load)
7914 env->busiest_group_type = busiest->group_type;
7915 /* Looks like there is an imbalance. Compute it */
7916 calculate_imbalance(env, &sds);
7925 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7927 static struct rq *find_busiest_queue(struct lb_env *env,
7928 struct sched_group *group)
7930 struct rq *busiest = NULL, *rq;
7931 unsigned long busiest_load = 0, busiest_capacity = 1;
7934 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7935 unsigned long capacity, wl;
7939 rt = fbq_classify_rq(rq);
7942 * We classify groups/runqueues into three groups:
7943 * - regular: there are !numa tasks
7944 * - remote: there are numa tasks that run on the 'wrong' node
7945 * - all: there is no distinction
7947 * In order to avoid migrating ideally placed numa tasks,
7948 * ignore those when there's better options.
7950 * If we ignore the actual busiest queue to migrate another
7951 * task, the next balance pass can still reduce the busiest
7952 * queue by moving tasks around inside the node.
7954 * If we cannot move enough load due to this classification
7955 * the next pass will adjust the group classification and
7956 * allow migration of more tasks.
7958 * Both cases only affect the total convergence complexity.
7960 if (rt > env->fbq_type)
7963 capacity = capacity_of(i);
7965 wl = weighted_cpuload(i);
7968 * When comparing with imbalance, use weighted_cpuload()
7969 * which is not scaled with the cpu capacity.
7972 if (rq->nr_running == 1 && wl > env->imbalance &&
7973 !check_cpu_capacity(rq, env->sd) &&
7974 env->busiest_group_type != group_misfit_task)
7978 * For the load comparisons with the other cpu's, consider
7979 * the weighted_cpuload() scaled with the cpu capacity, so
7980 * that the load can be moved away from the cpu that is
7981 * potentially running at a lower capacity.
7983 * Thus we're looking for max(wl_i / capacity_i), crosswise
7984 * multiplication to rid ourselves of the division works out
7985 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7986 * our previous maximum.
7988 if (wl * busiest_capacity > busiest_load * capacity) {
7990 busiest_capacity = capacity;
7999 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8000 * so long as it is large enough.
8002 #define MAX_PINNED_INTERVAL 512
8004 /* Working cpumask for load_balance and load_balance_newidle. */
8005 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8007 static int need_active_balance(struct lb_env *env)
8009 struct sched_domain *sd = env->sd;
8011 if (env->idle == CPU_NEWLY_IDLE) {
8014 * ASYM_PACKING needs to force migrate tasks from busy but
8015 * higher numbered CPUs in order to pack all tasks in the
8016 * lowest numbered CPUs.
8018 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8023 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8024 * It's worth migrating the task if the src_cpu's capacity is reduced
8025 * because of other sched_class or IRQs if more capacity stays
8026 * available on dst_cpu.
8028 if ((env->idle != CPU_NOT_IDLE) &&
8029 (env->src_rq->cfs.h_nr_running == 1)) {
8030 if ((check_cpu_capacity(env->src_rq, sd)) &&
8031 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8035 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8036 env->src_rq->cfs.h_nr_running == 1 &&
8037 cpu_overutilized(env->src_cpu) &&
8038 !cpu_overutilized(env->dst_cpu)) {
8042 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8045 static int active_load_balance_cpu_stop(void *data);
8047 static int should_we_balance(struct lb_env *env)
8049 struct sched_group *sg = env->sd->groups;
8050 struct cpumask *sg_cpus, *sg_mask;
8051 int cpu, balance_cpu = -1;
8054 * In the newly idle case, we will allow all the cpu's
8055 * to do the newly idle load balance.
8057 if (env->idle == CPU_NEWLY_IDLE)
8060 sg_cpus = sched_group_cpus(sg);
8061 sg_mask = sched_group_mask(sg);
8062 /* Try to find first idle cpu */
8063 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8064 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8071 if (balance_cpu == -1)
8072 balance_cpu = group_balance_cpu(sg);
8075 * First idle cpu or the first cpu(busiest) in this sched group
8076 * is eligible for doing load balancing at this and above domains.
8078 return balance_cpu == env->dst_cpu;
8082 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8083 * tasks if there is an imbalance.
8085 static int load_balance(int this_cpu, struct rq *this_rq,
8086 struct sched_domain *sd, enum cpu_idle_type idle,
8087 int *continue_balancing)
8089 int ld_moved, cur_ld_moved, active_balance = 0;
8090 struct sched_domain *sd_parent = sd->parent;
8091 struct sched_group *group;
8093 unsigned long flags;
8094 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8096 struct lb_env env = {
8098 .dst_cpu = this_cpu,
8100 .dst_grpmask = sched_group_cpus(sd->groups),
8102 .loop_break = sched_nr_migrate_break,
8105 .tasks = LIST_HEAD_INIT(env.tasks),
8109 * For NEWLY_IDLE load_balancing, we don't need to consider
8110 * other cpus in our group
8112 if (idle == CPU_NEWLY_IDLE)
8113 env.dst_grpmask = NULL;
8115 cpumask_copy(cpus, cpu_active_mask);
8117 schedstat_inc(sd, lb_count[idle]);
8120 if (!should_we_balance(&env)) {
8121 *continue_balancing = 0;
8125 group = find_busiest_group(&env);
8127 schedstat_inc(sd, lb_nobusyg[idle]);
8131 busiest = find_busiest_queue(&env, group);
8133 schedstat_inc(sd, lb_nobusyq[idle]);
8137 BUG_ON(busiest == env.dst_rq);
8139 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8141 env.src_cpu = busiest->cpu;
8142 env.src_rq = busiest;
8145 if (busiest->nr_running > 1) {
8147 * Attempt to move tasks. If find_busiest_group has found
8148 * an imbalance but busiest->nr_running <= 1, the group is
8149 * still unbalanced. ld_moved simply stays zero, so it is
8150 * correctly treated as an imbalance.
8152 env.flags |= LBF_ALL_PINNED;
8153 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8156 raw_spin_lock_irqsave(&busiest->lock, flags);
8159 * cur_ld_moved - load moved in current iteration
8160 * ld_moved - cumulative load moved across iterations
8162 cur_ld_moved = detach_tasks(&env);
8164 * We want to potentially lower env.src_cpu's OPP.
8167 update_capacity_of(env.src_cpu);
8170 * We've detached some tasks from busiest_rq. Every
8171 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8172 * unlock busiest->lock, and we are able to be sure
8173 * that nobody can manipulate the tasks in parallel.
8174 * See task_rq_lock() family for the details.
8177 raw_spin_unlock(&busiest->lock);
8181 ld_moved += cur_ld_moved;
8184 local_irq_restore(flags);
8186 if (env.flags & LBF_NEED_BREAK) {
8187 env.flags &= ~LBF_NEED_BREAK;
8192 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8193 * us and move them to an alternate dst_cpu in our sched_group
8194 * where they can run. The upper limit on how many times we
8195 * iterate on same src_cpu is dependent on number of cpus in our
8198 * This changes load balance semantics a bit on who can move
8199 * load to a given_cpu. In addition to the given_cpu itself
8200 * (or a ilb_cpu acting on its behalf where given_cpu is
8201 * nohz-idle), we now have balance_cpu in a position to move
8202 * load to given_cpu. In rare situations, this may cause
8203 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8204 * _independently_ and at _same_ time to move some load to
8205 * given_cpu) causing exceess load to be moved to given_cpu.
8206 * This however should not happen so much in practice and
8207 * moreover subsequent load balance cycles should correct the
8208 * excess load moved.
8210 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8212 /* Prevent to re-select dst_cpu via env's cpus */
8213 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8215 env.dst_rq = cpu_rq(env.new_dst_cpu);
8216 env.dst_cpu = env.new_dst_cpu;
8217 env.flags &= ~LBF_DST_PINNED;
8219 env.loop_break = sched_nr_migrate_break;
8222 * Go back to "more_balance" rather than "redo" since we
8223 * need to continue with same src_cpu.
8229 * We failed to reach balance because of affinity.
8232 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8234 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8235 *group_imbalance = 1;
8238 /* All tasks on this runqueue were pinned by CPU affinity */
8239 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8240 cpumask_clear_cpu(cpu_of(busiest), cpus);
8241 if (!cpumask_empty(cpus)) {
8243 env.loop_break = sched_nr_migrate_break;
8246 goto out_all_pinned;
8251 schedstat_inc(sd, lb_failed[idle]);
8253 * Increment the failure counter only on periodic balance.
8254 * We do not want newidle balance, which can be very
8255 * frequent, pollute the failure counter causing
8256 * excessive cache_hot migrations and active balances.
8258 if (idle != CPU_NEWLY_IDLE)
8259 if (env.src_grp_nr_running > 1)
8260 sd->nr_balance_failed++;
8262 if (need_active_balance(&env)) {
8263 raw_spin_lock_irqsave(&busiest->lock, flags);
8265 /* don't kick the active_load_balance_cpu_stop,
8266 * if the curr task on busiest cpu can't be
8269 if (!cpumask_test_cpu(this_cpu,
8270 tsk_cpus_allowed(busiest->curr))) {
8271 raw_spin_unlock_irqrestore(&busiest->lock,
8273 env.flags |= LBF_ALL_PINNED;
8274 goto out_one_pinned;
8278 * ->active_balance synchronizes accesses to
8279 * ->active_balance_work. Once set, it's cleared
8280 * only after active load balance is finished.
8282 if (!busiest->active_balance) {
8283 busiest->active_balance = 1;
8284 busiest->push_cpu = this_cpu;
8287 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8289 if (active_balance) {
8290 stop_one_cpu_nowait(cpu_of(busiest),
8291 active_load_balance_cpu_stop, busiest,
8292 &busiest->active_balance_work);
8296 * We've kicked active balancing, reset the failure
8299 sd->nr_balance_failed = sd->cache_nice_tries+1;
8302 sd->nr_balance_failed = 0;
8304 if (likely(!active_balance)) {
8305 /* We were unbalanced, so reset the balancing interval */
8306 sd->balance_interval = sd->min_interval;
8309 * If we've begun active balancing, start to back off. This
8310 * case may not be covered by the all_pinned logic if there
8311 * is only 1 task on the busy runqueue (because we don't call
8314 if (sd->balance_interval < sd->max_interval)
8315 sd->balance_interval *= 2;
8322 * We reach balance although we may have faced some affinity
8323 * constraints. Clear the imbalance flag if it was set.
8326 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8328 if (*group_imbalance)
8329 *group_imbalance = 0;
8334 * We reach balance because all tasks are pinned at this level so
8335 * we can't migrate them. Let the imbalance flag set so parent level
8336 * can try to migrate them.
8338 schedstat_inc(sd, lb_balanced[idle]);
8340 sd->nr_balance_failed = 0;
8343 /* tune up the balancing interval */
8344 if (((env.flags & LBF_ALL_PINNED) &&
8345 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8346 (sd->balance_interval < sd->max_interval))
8347 sd->balance_interval *= 2;
8354 static inline unsigned long
8355 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8357 unsigned long interval = sd->balance_interval;
8360 interval *= sd->busy_factor;
8362 /* scale ms to jiffies */
8363 interval = msecs_to_jiffies(interval);
8364 interval = clamp(interval, 1UL, max_load_balance_interval);
8370 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8372 unsigned long interval, next;
8374 interval = get_sd_balance_interval(sd, cpu_busy);
8375 next = sd->last_balance + interval;
8377 if (time_after(*next_balance, next))
8378 *next_balance = next;
8382 * idle_balance is called by schedule() if this_cpu is about to become
8383 * idle. Attempts to pull tasks from other CPUs.
8385 static int idle_balance(struct rq *this_rq)
8387 unsigned long next_balance = jiffies + HZ;
8388 int this_cpu = this_rq->cpu;
8389 struct sched_domain *sd;
8390 int pulled_task = 0;
8392 long removed_util=0;
8394 idle_enter_fair(this_rq);
8397 * We must set idle_stamp _before_ calling idle_balance(), such that we
8398 * measure the duration of idle_balance() as idle time.
8400 this_rq->idle_stamp = rq_clock(this_rq);
8402 if (!energy_aware() &&
8403 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8404 !this_rq->rd->overload)) {
8406 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8408 update_next_balance(sd, 0, &next_balance);
8414 raw_spin_unlock(&this_rq->lock);
8417 * If removed_util_avg is !0 we most probably migrated some task away
8418 * from this_cpu. In this case we might be willing to trigger an OPP
8419 * update, but we want to do so if we don't find anybody else to pull
8420 * here (we will trigger an OPP update with the pulled task's enqueue
8423 * Record removed_util before calling update_blocked_averages, and use
8424 * it below (before returning) to see if an OPP update is required.
8426 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8427 update_blocked_averages(this_cpu);
8429 for_each_domain(this_cpu, sd) {
8430 int continue_balancing = 1;
8431 u64 t0, domain_cost;
8433 if (!(sd->flags & SD_LOAD_BALANCE))
8436 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8437 update_next_balance(sd, 0, &next_balance);
8441 if (sd->flags & SD_BALANCE_NEWIDLE) {
8442 t0 = sched_clock_cpu(this_cpu);
8444 pulled_task = load_balance(this_cpu, this_rq,
8446 &continue_balancing);
8448 domain_cost = sched_clock_cpu(this_cpu) - t0;
8449 if (domain_cost > sd->max_newidle_lb_cost)
8450 sd->max_newidle_lb_cost = domain_cost;
8452 curr_cost += domain_cost;
8455 update_next_balance(sd, 0, &next_balance);
8458 * Stop searching for tasks to pull if there are
8459 * now runnable tasks on this rq.
8461 if (pulled_task || this_rq->nr_running > 0)
8466 raw_spin_lock(&this_rq->lock);
8468 if (curr_cost > this_rq->max_idle_balance_cost)
8469 this_rq->max_idle_balance_cost = curr_cost;
8472 * While browsing the domains, we released the rq lock, a task could
8473 * have been enqueued in the meantime. Since we're not going idle,
8474 * pretend we pulled a task.
8476 if (this_rq->cfs.h_nr_running && !pulled_task)
8480 /* Move the next balance forward */
8481 if (time_after(this_rq->next_balance, next_balance))
8482 this_rq->next_balance = next_balance;
8484 /* Is there a task of a high priority class? */
8485 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8489 idle_exit_fair(this_rq);
8490 this_rq->idle_stamp = 0;
8491 } else if (removed_util) {
8493 * No task pulled and someone has been migrated away.
8494 * Good case to trigger an OPP update.
8496 update_capacity_of(this_cpu);
8503 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8504 * running tasks off the busiest CPU onto idle CPUs. It requires at
8505 * least 1 task to be running on each physical CPU where possible, and
8506 * avoids physical / logical imbalances.
8508 static int active_load_balance_cpu_stop(void *data)
8510 struct rq *busiest_rq = data;
8511 int busiest_cpu = cpu_of(busiest_rq);
8512 int target_cpu = busiest_rq->push_cpu;
8513 struct rq *target_rq = cpu_rq(target_cpu);
8514 struct sched_domain *sd;
8515 struct task_struct *p = NULL;
8517 raw_spin_lock_irq(&busiest_rq->lock);
8519 /* make sure the requested cpu hasn't gone down in the meantime */
8520 if (unlikely(busiest_cpu != smp_processor_id() ||
8521 !busiest_rq->active_balance))
8524 /* Is there any task to move? */
8525 if (busiest_rq->nr_running <= 1)
8529 * This condition is "impossible", if it occurs
8530 * we need to fix it. Originally reported by
8531 * Bjorn Helgaas on a 128-cpu setup.
8533 BUG_ON(busiest_rq == target_rq);
8535 /* Search for an sd spanning us and the target CPU. */
8537 for_each_domain(target_cpu, sd) {
8538 if ((sd->flags & SD_LOAD_BALANCE) &&
8539 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8544 struct lb_env env = {
8546 .dst_cpu = target_cpu,
8547 .dst_rq = target_rq,
8548 .src_cpu = busiest_rq->cpu,
8549 .src_rq = busiest_rq,
8553 schedstat_inc(sd, alb_count);
8555 p = detach_one_task(&env);
8557 schedstat_inc(sd, alb_pushed);
8559 * We want to potentially lower env.src_cpu's OPP.
8561 update_capacity_of(env.src_cpu);
8564 schedstat_inc(sd, alb_failed);
8568 busiest_rq->active_balance = 0;
8569 raw_spin_unlock(&busiest_rq->lock);
8572 attach_one_task(target_rq, p);
8579 static inline int on_null_domain(struct rq *rq)
8581 return unlikely(!rcu_dereference_sched(rq->sd));
8584 #ifdef CONFIG_NO_HZ_COMMON
8586 * idle load balancing details
8587 * - When one of the busy CPUs notice that there may be an idle rebalancing
8588 * needed, they will kick the idle load balancer, which then does idle
8589 * load balancing for all the idle CPUs.
8592 cpumask_var_t idle_cpus_mask;
8594 unsigned long next_balance; /* in jiffy units */
8595 } nohz ____cacheline_aligned;
8597 static inline int find_new_ilb(void)
8599 int ilb = cpumask_first(nohz.idle_cpus_mask);
8601 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8608 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8609 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8610 * CPU (if there is one).
8612 static void nohz_balancer_kick(void)
8616 nohz.next_balance++;
8618 ilb_cpu = find_new_ilb();
8620 if (ilb_cpu >= nr_cpu_ids)
8623 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8626 * Use smp_send_reschedule() instead of resched_cpu().
8627 * This way we generate a sched IPI on the target cpu which
8628 * is idle. And the softirq performing nohz idle load balance
8629 * will be run before returning from the IPI.
8631 smp_send_reschedule(ilb_cpu);
8635 static inline void nohz_balance_exit_idle(int cpu)
8637 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8639 * Completely isolated CPUs don't ever set, so we must test.
8641 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8642 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8643 atomic_dec(&nohz.nr_cpus);
8645 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8649 static inline void set_cpu_sd_state_busy(void)
8651 struct sched_domain *sd;
8652 int cpu = smp_processor_id();
8655 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8657 if (!sd || !sd->nohz_idle)
8661 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8666 void set_cpu_sd_state_idle(void)
8668 struct sched_domain *sd;
8669 int cpu = smp_processor_id();
8672 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8674 if (!sd || sd->nohz_idle)
8678 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8684 * This routine will record that the cpu is going idle with tick stopped.
8685 * This info will be used in performing idle load balancing in the future.
8687 void nohz_balance_enter_idle(int cpu)
8690 * If this cpu is going down, then nothing needs to be done.
8692 if (!cpu_active(cpu))
8695 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8699 * If we're a completely isolated CPU, we don't play.
8701 if (on_null_domain(cpu_rq(cpu)))
8704 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8705 atomic_inc(&nohz.nr_cpus);
8706 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8709 static int sched_ilb_notifier(struct notifier_block *nfb,
8710 unsigned long action, void *hcpu)
8712 switch (action & ~CPU_TASKS_FROZEN) {
8714 nohz_balance_exit_idle(smp_processor_id());
8722 static DEFINE_SPINLOCK(balancing);
8725 * Scale the max load_balance interval with the number of CPUs in the system.
8726 * This trades load-balance latency on larger machines for less cross talk.
8728 void update_max_interval(void)
8730 max_load_balance_interval = HZ*num_online_cpus()/10;
8734 * It checks each scheduling domain to see if it is due to be balanced,
8735 * and initiates a balancing operation if so.
8737 * Balancing parameters are set up in init_sched_domains.
8739 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8741 int continue_balancing = 1;
8743 unsigned long interval;
8744 struct sched_domain *sd;
8745 /* Earliest time when we have to do rebalance again */
8746 unsigned long next_balance = jiffies + 60*HZ;
8747 int update_next_balance = 0;
8748 int need_serialize, need_decay = 0;
8751 update_blocked_averages(cpu);
8754 for_each_domain(cpu, sd) {
8756 * Decay the newidle max times here because this is a regular
8757 * visit to all the domains. Decay ~1% per second.
8759 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8760 sd->max_newidle_lb_cost =
8761 (sd->max_newidle_lb_cost * 253) / 256;
8762 sd->next_decay_max_lb_cost = jiffies + HZ;
8765 max_cost += sd->max_newidle_lb_cost;
8767 if (!(sd->flags & SD_LOAD_BALANCE))
8771 * Stop the load balance at this level. There is another
8772 * CPU in our sched group which is doing load balancing more
8775 if (!continue_balancing) {
8781 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8783 need_serialize = sd->flags & SD_SERIALIZE;
8784 if (need_serialize) {
8785 if (!spin_trylock(&balancing))
8789 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8790 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8792 * The LBF_DST_PINNED logic could have changed
8793 * env->dst_cpu, so we can't know our idle
8794 * state even if we migrated tasks. Update it.
8796 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8798 sd->last_balance = jiffies;
8799 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8802 spin_unlock(&balancing);
8804 if (time_after(next_balance, sd->last_balance + interval)) {
8805 next_balance = sd->last_balance + interval;
8806 update_next_balance = 1;
8811 * Ensure the rq-wide value also decays but keep it at a
8812 * reasonable floor to avoid funnies with rq->avg_idle.
8814 rq->max_idle_balance_cost =
8815 max((u64)sysctl_sched_migration_cost, max_cost);
8820 * next_balance will be updated only when there is a need.
8821 * When the cpu is attached to null domain for ex, it will not be
8824 if (likely(update_next_balance)) {
8825 rq->next_balance = next_balance;
8827 #ifdef CONFIG_NO_HZ_COMMON
8829 * If this CPU has been elected to perform the nohz idle
8830 * balance. Other idle CPUs have already rebalanced with
8831 * nohz_idle_balance() and nohz.next_balance has been
8832 * updated accordingly. This CPU is now running the idle load
8833 * balance for itself and we need to update the
8834 * nohz.next_balance accordingly.
8836 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8837 nohz.next_balance = rq->next_balance;
8842 #ifdef CONFIG_NO_HZ_COMMON
8844 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8845 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8847 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8849 int this_cpu = this_rq->cpu;
8852 /* Earliest time when we have to do rebalance again */
8853 unsigned long next_balance = jiffies + 60*HZ;
8854 int update_next_balance = 0;
8856 if (idle != CPU_IDLE ||
8857 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8860 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8861 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8865 * If this cpu gets work to do, stop the load balancing
8866 * work being done for other cpus. Next load
8867 * balancing owner will pick it up.
8872 rq = cpu_rq(balance_cpu);
8875 * If time for next balance is due,
8878 if (time_after_eq(jiffies, rq->next_balance)) {
8879 raw_spin_lock_irq(&rq->lock);
8880 update_rq_clock(rq);
8881 update_idle_cpu_load(rq);
8882 raw_spin_unlock_irq(&rq->lock);
8883 rebalance_domains(rq, CPU_IDLE);
8886 if (time_after(next_balance, rq->next_balance)) {
8887 next_balance = rq->next_balance;
8888 update_next_balance = 1;
8893 * next_balance will be updated only when there is a need.
8894 * When the CPU is attached to null domain for ex, it will not be
8897 if (likely(update_next_balance))
8898 nohz.next_balance = next_balance;
8900 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8904 * Current heuristic for kicking the idle load balancer in the presence
8905 * of an idle cpu in the system.
8906 * - This rq has more than one task.
8907 * - This rq has at least one CFS task and the capacity of the CPU is
8908 * significantly reduced because of RT tasks or IRQs.
8909 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8910 * multiple busy cpu.
8911 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8912 * domain span are idle.
8914 static inline bool nohz_kick_needed(struct rq *rq)
8916 unsigned long now = jiffies;
8917 struct sched_domain *sd;
8918 struct sched_group_capacity *sgc;
8919 int nr_busy, cpu = rq->cpu;
8922 if (unlikely(rq->idle_balance))
8926 * We may be recently in ticked or tickless idle mode. At the first
8927 * busy tick after returning from idle, we will update the busy stats.
8929 set_cpu_sd_state_busy();
8930 nohz_balance_exit_idle(cpu);
8933 * None are in tickless mode and hence no need for NOHZ idle load
8936 if (likely(!atomic_read(&nohz.nr_cpus)))
8939 if (time_before(now, nohz.next_balance))
8942 if (rq->nr_running >= 2 &&
8943 (!energy_aware() || cpu_overutilized(cpu)))
8947 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8948 if (sd && !energy_aware()) {
8949 sgc = sd->groups->sgc;
8950 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8959 sd = rcu_dereference(rq->sd);
8961 if ((rq->cfs.h_nr_running >= 1) &&
8962 check_cpu_capacity(rq, sd)) {
8968 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8969 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8970 sched_domain_span(sd)) < cpu)) {
8980 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8984 * run_rebalance_domains is triggered when needed from the scheduler tick.
8985 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8987 static void run_rebalance_domains(struct softirq_action *h)
8989 struct rq *this_rq = this_rq();
8990 enum cpu_idle_type idle = this_rq->idle_balance ?
8991 CPU_IDLE : CPU_NOT_IDLE;
8994 * If this cpu has a pending nohz_balance_kick, then do the
8995 * balancing on behalf of the other idle cpus whose ticks are
8996 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8997 * give the idle cpus a chance to load balance. Else we may
8998 * load balance only within the local sched_domain hierarchy
8999 * and abort nohz_idle_balance altogether if we pull some load.
9001 nohz_idle_balance(this_rq, idle);
9002 rebalance_domains(this_rq, idle);
9006 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9008 void trigger_load_balance(struct rq *rq)
9010 /* Don't need to rebalance while attached to NULL domain */
9011 if (unlikely(on_null_domain(rq)))
9014 if (time_after_eq(jiffies, rq->next_balance))
9015 raise_softirq(SCHED_SOFTIRQ);
9016 #ifdef CONFIG_NO_HZ_COMMON
9017 if (nohz_kick_needed(rq))
9018 nohz_balancer_kick();
9022 static void rq_online_fair(struct rq *rq)
9026 update_runtime_enabled(rq);
9029 static void rq_offline_fair(struct rq *rq)
9033 /* Ensure any throttled groups are reachable by pick_next_task */
9034 unthrottle_offline_cfs_rqs(rq);
9037 #endif /* CONFIG_SMP */
9040 * scheduler tick hitting a task of our scheduling class:
9042 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9044 struct cfs_rq *cfs_rq;
9045 struct sched_entity *se = &curr->se;
9047 for_each_sched_entity(se) {
9048 cfs_rq = cfs_rq_of(se);
9049 entity_tick(cfs_rq, se, queued);
9052 if (static_branch_unlikely(&sched_numa_balancing))
9053 task_tick_numa(rq, curr);
9056 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9057 rq->rd->overutilized = true;
9058 trace_sched_overutilized(true);
9061 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9067 * called on fork with the child task as argument from the parent's context
9068 * - child not yet on the tasklist
9069 * - preemption disabled
9071 static void task_fork_fair(struct task_struct *p)
9073 struct cfs_rq *cfs_rq;
9074 struct sched_entity *se = &p->se, *curr;
9075 int this_cpu = smp_processor_id();
9076 struct rq *rq = this_rq();
9077 unsigned long flags;
9079 raw_spin_lock_irqsave(&rq->lock, flags);
9081 update_rq_clock(rq);
9083 cfs_rq = task_cfs_rq(current);
9084 curr = cfs_rq->curr;
9087 * Not only the cpu but also the task_group of the parent might have
9088 * been changed after parent->se.parent,cfs_rq were copied to
9089 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9090 * of child point to valid ones.
9093 __set_task_cpu(p, this_cpu);
9096 update_curr(cfs_rq);
9099 se->vruntime = curr->vruntime;
9100 place_entity(cfs_rq, se, 1);
9102 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9104 * Upon rescheduling, sched_class::put_prev_task() will place
9105 * 'current' within the tree based on its new key value.
9107 swap(curr->vruntime, se->vruntime);
9111 se->vruntime -= cfs_rq->min_vruntime;
9113 raw_spin_unlock_irqrestore(&rq->lock, flags);
9117 * Priority of the task has changed. Check to see if we preempt
9121 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9123 if (!task_on_rq_queued(p))
9127 * Reschedule if we are currently running on this runqueue and
9128 * our priority decreased, or if we are not currently running on
9129 * this runqueue and our priority is higher than the current's
9131 if (rq->curr == p) {
9132 if (p->prio > oldprio)
9135 check_preempt_curr(rq, p, 0);
9138 static inline bool vruntime_normalized(struct task_struct *p)
9140 struct sched_entity *se = &p->se;
9143 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9144 * the dequeue_entity(.flags=0) will already have normalized the
9151 * When !on_rq, vruntime of the task has usually NOT been normalized.
9152 * But there are some cases where it has already been normalized:
9154 * - A forked child which is waiting for being woken up by
9155 * wake_up_new_task().
9156 * - A task which has been woken up by try_to_wake_up() and
9157 * waiting for actually being woken up by sched_ttwu_pending().
9159 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9165 static void detach_task_cfs_rq(struct task_struct *p)
9167 struct sched_entity *se = &p->se;
9168 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9170 if (!vruntime_normalized(p)) {
9172 * Fix up our vruntime so that the current sleep doesn't
9173 * cause 'unlimited' sleep bonus.
9175 place_entity(cfs_rq, se, 0);
9176 se->vruntime -= cfs_rq->min_vruntime;
9179 /* Catch up with the cfs_rq and remove our load when we leave */
9180 detach_entity_load_avg(cfs_rq, se);
9183 static void attach_task_cfs_rq(struct task_struct *p)
9185 struct sched_entity *se = &p->se;
9186 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9188 #ifdef CONFIG_FAIR_GROUP_SCHED
9190 * Since the real-depth could have been changed (only FAIR
9191 * class maintain depth value), reset depth properly.
9193 se->depth = se->parent ? se->parent->depth + 1 : 0;
9196 /* Synchronize task with its cfs_rq */
9197 attach_entity_load_avg(cfs_rq, se);
9199 if (!vruntime_normalized(p))
9200 se->vruntime += cfs_rq->min_vruntime;
9203 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9205 detach_task_cfs_rq(p);
9208 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9210 attach_task_cfs_rq(p);
9212 if (task_on_rq_queued(p)) {
9214 * We were most likely switched from sched_rt, so
9215 * kick off the schedule if running, otherwise just see
9216 * if we can still preempt the current task.
9221 check_preempt_curr(rq, p, 0);
9225 /* Account for a task changing its policy or group.
9227 * This routine is mostly called to set cfs_rq->curr field when a task
9228 * migrates between groups/classes.
9230 static void set_curr_task_fair(struct rq *rq)
9232 struct sched_entity *se = &rq->curr->se;
9234 for_each_sched_entity(se) {
9235 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9237 set_next_entity(cfs_rq, se);
9238 /* ensure bandwidth has been allocated on our new cfs_rq */
9239 account_cfs_rq_runtime(cfs_rq, 0);
9243 void init_cfs_rq(struct cfs_rq *cfs_rq)
9245 cfs_rq->tasks_timeline = RB_ROOT;
9246 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9247 #ifndef CONFIG_64BIT
9248 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9251 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9252 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9256 #ifdef CONFIG_FAIR_GROUP_SCHED
9257 static void task_move_group_fair(struct task_struct *p)
9259 detach_task_cfs_rq(p);
9260 set_task_rq(p, task_cpu(p));
9263 /* Tell se's cfs_rq has been changed -- migrated */
9264 p->se.avg.last_update_time = 0;
9266 attach_task_cfs_rq(p);
9269 void free_fair_sched_group(struct task_group *tg)
9273 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9275 for_each_possible_cpu(i) {
9277 kfree(tg->cfs_rq[i]);
9280 remove_entity_load_avg(tg->se[i]);
9289 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9291 struct cfs_rq *cfs_rq;
9292 struct sched_entity *se;
9295 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9298 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9302 tg->shares = NICE_0_LOAD;
9304 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9306 for_each_possible_cpu(i) {
9307 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9308 GFP_KERNEL, cpu_to_node(i));
9312 se = kzalloc_node(sizeof(struct sched_entity),
9313 GFP_KERNEL, cpu_to_node(i));
9317 init_cfs_rq(cfs_rq);
9318 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9319 init_entity_runnable_average(se);
9330 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9332 struct rq *rq = cpu_rq(cpu);
9333 unsigned long flags;
9336 * Only empty task groups can be destroyed; so we can speculatively
9337 * check on_list without danger of it being re-added.
9339 if (!tg->cfs_rq[cpu]->on_list)
9342 raw_spin_lock_irqsave(&rq->lock, flags);
9343 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9344 raw_spin_unlock_irqrestore(&rq->lock, flags);
9347 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9348 struct sched_entity *se, int cpu,
9349 struct sched_entity *parent)
9351 struct rq *rq = cpu_rq(cpu);
9355 init_cfs_rq_runtime(cfs_rq);
9357 tg->cfs_rq[cpu] = cfs_rq;
9360 /* se could be NULL for root_task_group */
9365 se->cfs_rq = &rq->cfs;
9368 se->cfs_rq = parent->my_q;
9369 se->depth = parent->depth + 1;
9373 /* guarantee group entities always have weight */
9374 update_load_set(&se->load, NICE_0_LOAD);
9375 se->parent = parent;
9378 static DEFINE_MUTEX(shares_mutex);
9380 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9383 unsigned long flags;
9386 * We can't change the weight of the root cgroup.
9391 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9393 mutex_lock(&shares_mutex);
9394 if (tg->shares == shares)
9397 tg->shares = shares;
9398 for_each_possible_cpu(i) {
9399 struct rq *rq = cpu_rq(i);
9400 struct sched_entity *se;
9403 /* Propagate contribution to hierarchy */
9404 raw_spin_lock_irqsave(&rq->lock, flags);
9406 /* Possible calls to update_curr() need rq clock */
9407 update_rq_clock(rq);
9408 for_each_sched_entity(se)
9409 update_cfs_shares(group_cfs_rq(se));
9410 raw_spin_unlock_irqrestore(&rq->lock, flags);
9414 mutex_unlock(&shares_mutex);
9417 #else /* CONFIG_FAIR_GROUP_SCHED */
9419 void free_fair_sched_group(struct task_group *tg) { }
9421 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9426 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9428 #endif /* CONFIG_FAIR_GROUP_SCHED */
9431 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9433 struct sched_entity *se = &task->se;
9434 unsigned int rr_interval = 0;
9437 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9440 if (rq->cfs.load.weight)
9441 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9447 * All the scheduling class methods:
9449 const struct sched_class fair_sched_class = {
9450 .next = &idle_sched_class,
9451 .enqueue_task = enqueue_task_fair,
9452 .dequeue_task = dequeue_task_fair,
9453 .yield_task = yield_task_fair,
9454 .yield_to_task = yield_to_task_fair,
9456 .check_preempt_curr = check_preempt_wakeup,
9458 .pick_next_task = pick_next_task_fair,
9459 .put_prev_task = put_prev_task_fair,
9462 .select_task_rq = select_task_rq_fair,
9463 .migrate_task_rq = migrate_task_rq_fair,
9465 .rq_online = rq_online_fair,
9466 .rq_offline = rq_offline_fair,
9468 .task_waking = task_waking_fair,
9469 .task_dead = task_dead_fair,
9470 .set_cpus_allowed = set_cpus_allowed_common,
9473 .set_curr_task = set_curr_task_fair,
9474 .task_tick = task_tick_fair,
9475 .task_fork = task_fork_fair,
9477 .prio_changed = prio_changed_fair,
9478 .switched_from = switched_from_fair,
9479 .switched_to = switched_to_fair,
9481 .get_rr_interval = get_rr_interval_fair,
9483 .update_curr = update_curr_fair,
9485 #ifdef CONFIG_FAIR_GROUP_SCHED
9486 .task_move_group = task_move_group_fair,
9490 #ifdef CONFIG_SCHED_DEBUG
9491 void print_cfs_stats(struct seq_file *m, int cpu)
9493 struct cfs_rq *cfs_rq;
9496 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9497 print_cfs_rq(m, cpu, cfs_rq);
9501 #ifdef CONFIG_NUMA_BALANCING
9502 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9505 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9507 for_each_online_node(node) {
9508 if (p->numa_faults) {
9509 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9510 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9512 if (p->numa_group) {
9513 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9514 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9516 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9519 #endif /* CONFIG_NUMA_BALANCING */
9520 #endif /* CONFIG_SCHED_DEBUG */
9522 __init void init_sched_fair_class(void)
9525 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9527 #ifdef CONFIG_NO_HZ_COMMON
9528 nohz.next_balance = jiffies;
9529 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9530 cpu_notifier(sched_ilb_notifier, 0);